git/run-command.c

1983 lines
44 KiB
C
Raw Normal View History

global: introduce `USE_THE_REPOSITORY_VARIABLE` macro Use of the `the_repository` variable is deprecated nowadays, and we slowly but steadily convert the codebase to not use it anymore. Instead, callers should be passing down the repository to work on via parameters. It is hard though to prove that a given code unit does not use this variable anymore. The most trivial case, merely demonstrating that there is no direct use of `the_repository`, is already a bit of a pain during code reviews as the reviewer needs to manually verify claims made by the patch author. The bigger problem though is that we have many interfaces that implicitly rely on `the_repository`. Introduce a new `USE_THE_REPOSITORY_VARIABLE` macro that allows code units to opt into usage of `the_repository`. The intent of this macro is to demonstrate that a certain code unit does not use this variable anymore, and to keep it from new dependencies on it in future changes, be it explicit or implicit For now, the macro only guards `the_repository` itself as well as `the_hash_algo`. There are many more known interfaces where we have an implicit dependency on `the_repository`, but those are not guarded at the current point in time. Over time though, we should start to add guards as required (or even better, just remove them). Define the macro as required in our code units. As expected, most of our code still relies on the global variable. Nearly all of our builtins rely on the variable as there is no way yet to pass `the_repository` to their entry point. For now, declare the macro in "biultin.h" to keep the required changes at least a little bit more contained. Signed-off-by: Patrick Steinhardt <ps@pks.im> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2024-06-14 14:50:23 +08:00
#define USE_THE_REPOSITORY_VARIABLE
#include "git-compat-util.h"
#include "run-command.h"
#include "environment.h"
#include "exec-cmd.h"
#include "gettext.h"
#include "sigchain.h"
#include "strvec.h"
#include "symlinks.h"
run-command: add an asynchronous parallel child processor This allows to run external commands in parallel with ordered output on stderr. If we run external commands in parallel we cannot pipe the output directly to the our stdout/err as it would mix up. So each process's output will flow through a pipe, which we buffer. One subprocess can be directly piped to out stdout/err for a low latency feedback to the user. Example: Let's assume we have 5 submodules A,B,C,D,E and each fetch takes a different amount of time as the different submodules vary in size, then the output of fetches in sequential order might look like this: time --> output: |---A---| |-B-| |-------C-------| |-D-| |-E-| When we schedule these submodules into maximal two parallel processes, a schedule and sample output over time may look like this: process 1: |---A---| |-D-| |-E-| process 2: |-B-| |-------C-------| output: |---A---|B|---C-------|DE So A will be perceived as it would run normally in the single child version. As B has finished by the time A is done, we can dump its whole progress buffer on stderr, such that it looks like it finished in no time. Once that is done, C is determined to be the visible child and its progress will be reported in real time. So this way of output is really good for human consumption, as it only changes the timing, not the actual output. For machine consumption the output needs to be prepared in the tasks, by either having a prefix per line or per block to indicate whose tasks output is displayed, because the output order may not follow the original sequential ordering: |----A----| |--B--| |-C-| will be scheduled to be all parallel: process 1: |----A----| process 2: |--B--| process 3: |-C-| output: |----A----|CB This happens because C finished before B did, so it will be queued for output before B. To detect when a child has finished executing, we check interleaved with other actions (such as checking the liveliness of children or starting new processes) whether the stderr pipe still exists. Once a child closed its stderr stream, we assume it is terminating very soon, and use `finish_command()` from the single external process execution interface to collect the exit status. By maintaining the strong assumption of stderr being open until the very end of a child process, we can avoid other hassle such as an implementation using `waitpid(-1)`, which is not implemented in Windows. Signed-off-by: Stefan Beller <sbeller@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-12-16 08:04:10 +08:00
#include "thread-utils.h"
#include "strbuf.h"
#include "string-list.h"
#include "trace.h"
#include "trace2.h"
#include "quote.h"
#include "config.h"
#include "packfile.h"
pipe_command(): mark stdin descriptor as non-blocking Our pipe_command() helper lets you both write to and read from a child process on its stdin/stdout. It's supposed to work without deadlocks because we use poll() to check when descriptors are ready for reading or writing. But there's a bug: if both the data to be written and the data to be read back exceed the pipe buffer, we'll deadlock. The issue is that the code assumes that if you have, say, a 2MB buffer to write and poll() tells you that the pipe descriptor is ready for writing, that calling: write(cmd->in, buf, 2*1024*1024); will do a partial write, filling the pipe buffer and then returning what it did write. And that is what it would do on a socket, but not for a pipe. When writing to a pipe, at least on Linux, it will block waiting for the child process to read() more. And now we have a potential deadlock, because the child may be writing back to us, waiting for us to read() ourselves. An easy way to trigger this is: git -c add.interactive.useBuiltin=true \ -c interactive.diffFilter=cat \ checkout -p HEAD~200 The diff against HEAD~200 will be big, and the filter wants to write all of it back to us (obviously this is a dummy filter, but in the real world something like diff-highlight would similarly stream back a big output). If you set add.interactive.useBuiltin to false, the problem goes away, because now we're not using pipe_command() anymore (instead, that part happens in perl). But this isn't a bug in the interactive code at all. It's the underlying pipe_command() code which is broken, and has been all along. We presumably didn't notice because most calls only do input _or_ output, not both. And the few that do both, like gpg calls, may have large inputs or outputs, but never both at the same time (e.g., consider signing, which has a large payload but a small signature comes back). The obvious fix is to put the descriptor into non-blocking mode, and indeed, that makes the problem go away. Callers shouldn't need to care, because they never see the descriptor (they hand us a buffer to feed into it). The included test fails reliably on Linux without this patch. Curiously, it doesn't fail in our Windows CI environment, but has been reported to do so for individual developers. It should pass in any environment after this patch (courtesy of the compat/ layers added in the last few commits). Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2022-08-17 14:10:22 +08:00
#include "compat/nonblock.h"
void child_process_init(struct child_process *child)
{
struct child_process blank = CHILD_PROCESS_INIT;
memcpy(child, &blank, sizeof(*child));
}
void child_process_clear(struct child_process *child)
{
strvec_clear(&child->args);
strvec_clear(&child->env);
}
struct child_to_clean {
pid_t pid;
struct child_process *process;
struct child_to_clean *next;
};
static struct child_to_clean *children_to_clean;
static int installed_child_cleanup_handler;
pager: don't use unsafe functions in signal handlers Since the commit a3da8821208d (pager: do wait_for_pager on signal death), we call wait_for_pager() in the pager's signal handler. The recent bug report revealed that this causes a deadlock in glibc at aborting "git log" [*1*]. When this happens, git process is left unterminated, and it can't be killed by SIGTERM but only by SIGKILL. The problem is that wait_for_pager() function does more than waiting for pager process's termination, but it does cleanups and printing errors. Unfortunately, the functions that may be used in a signal handler are very limited [*2*]. Particularly, malloc(), free() and the variants can't be used in a signal handler because they take a mutex internally in glibc. This was the cause of the deadlock above. Other than the direct calls of malloc/free, many functions calling malloc/free can't be used. strerror() is such one, either. Also the usage of fflush() and printf() in a signal handler is bad, although it seems working so far. In a safer side, we should avoid them, too. This patch tries to reduce the calls of such functions in signal handlers. wait_for_signal() takes a flag and avoids the unsafe calls. Also, finish_command_in_signal() is introduced for the same reason. There the free() calls are removed, and only waits for the children without whining at errors. [*1*] https://bugzilla.opensuse.org/show_bug.cgi?id=942297 [*2*] http://pubs.opengroup.org/onlinepubs/9699919799/functions/V2_chap02.html#tag_15_04_03 Signed-off-by: Takashi Iwai <tiwai@suse.de> Reviewed-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-09-04 17:35:57 +08:00
static void cleanup_children(int sig, int in_signal)
{
execv_dashed_external: wait for child on signal death When you hit ^C to interrupt a git command going to a pager, this usually leaves the pager running. But when a dashed external is in use, the pager ends up in a funny state and quits (but only after eating one more character from the terminal!). This fixes it. Explaining the reason will require a little background. When git runs a pager, it's important for the git process to hang around and wait for the pager to finish, even though it has no more data to feed it. This is because git spawns the pager as a child, and thus the git process is the session leader on the terminal. After it dies, the pager will finish its current read from the terminal (eating the one character), and then get EIO trying to read again. When you hit ^C, that sends SIGINT to git and to the pager, and it's a similar situation. The pager ignores it, but the git process needs to hang around until the pager is done. We addressed that long ago in a3da882120 (pager: do wait_for_pager on signal death, 2009-01-22). But when you have a dashed external (or an alias pointing to a builtin, which will re-exec git for the builtin), there's an extra process in the mix. For instance, running: $ git -c alias.l=log l will end up with a process tree like: git (parent) \ git-log (child) \ less (pager) If you hit ^C, SIGINT goes to all of them. The pager ignores it, and the child git process will end up in wait_for_pager(). But the parent git process will die, and the usual EIO trouble happens. So we really want the parent git process to wait_for_pager(), but of course it doesn't know anything about the pager at all, since it was started by the child. However, we can have it wait on the git-log child, which in turn is waiting on the pager. And that's what this patch does. There are a few design decisions here worth explaining: 1. The new feature is attached to run-command's clean_on_exit feature. Partly this is convenience, since that feature already has a signal handler that deals with child cleanup. But it's also a meaningful connection. The main reason that dashed externals use clean_on_exit is to bind the two processes together. If somebody kills the parent with a signal, we propagate that to the child (in this instance with SIGINT, we do propagate but it doesn't matter because the original signal went to the whole process group). Likewise, we do not want the parent to go away until the child has done so. In a traditional Unix world, we'd probably accomplish this binding by just having the parent execve() the child directly. But since that doesn't work on Windows, everything goes through run_command's more spawn-like interface. 2. We do _not_ automatically waitpid() on any clean_on_exit children. For dashed externals this makes sense; we know that the parent is doing nothing but waiting for the child to exit anyway. But with other children, it's possible that the child, after getting the signal, could be waiting on the parent to do something (like closing a descriptor). If we were to wait on such a child, we'd end up in a deadlock. So this errs on the side of caution, and lets callers enable the feature explicitly. 3. When we send children the cleanup signal, we send all the signals first, before waiting on any children. This is to avoid the case where one child might be waiting on another one to exit, causing a deadlock. We inform all of them that it's time to die before reaping any. In practice, there is only ever one dashed external run from a given process, so this doesn't matter much now. But it future-proofs us if other callers start using the wait_after_clean mechanism. There's no automated test here, because it would end up racy and unportable. But it's easy to reproduce the situation by running the log command given above and hitting ^C. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2017-01-07 09:22:23 +08:00
struct child_to_clean *children_to_wait_for = NULL;
while (children_to_clean) {
struct child_to_clean *p = children_to_clean;
children_to_clean = p->next;
if (p->process && !in_signal) {
struct child_process *process = p->process;
if (process->clean_on_exit_handler) {
trace_printf(
"trace: run_command: running exit handler for pid %"
PRIuMAX, (uintmax_t)p->pid
);
process->clean_on_exit_handler(process);
}
}
kill(p->pid, sig);
execv_dashed_external: wait for child on signal death When you hit ^C to interrupt a git command going to a pager, this usually leaves the pager running. But when a dashed external is in use, the pager ends up in a funny state and quits (but only after eating one more character from the terminal!). This fixes it. Explaining the reason will require a little background. When git runs a pager, it's important for the git process to hang around and wait for the pager to finish, even though it has no more data to feed it. This is because git spawns the pager as a child, and thus the git process is the session leader on the terminal. After it dies, the pager will finish its current read from the terminal (eating the one character), and then get EIO trying to read again. When you hit ^C, that sends SIGINT to git and to the pager, and it's a similar situation. The pager ignores it, but the git process needs to hang around until the pager is done. We addressed that long ago in a3da882120 (pager: do wait_for_pager on signal death, 2009-01-22). But when you have a dashed external (or an alias pointing to a builtin, which will re-exec git for the builtin), there's an extra process in the mix. For instance, running: $ git -c alias.l=log l will end up with a process tree like: git (parent) \ git-log (child) \ less (pager) If you hit ^C, SIGINT goes to all of them. The pager ignores it, and the child git process will end up in wait_for_pager(). But the parent git process will die, and the usual EIO trouble happens. So we really want the parent git process to wait_for_pager(), but of course it doesn't know anything about the pager at all, since it was started by the child. However, we can have it wait on the git-log child, which in turn is waiting on the pager. And that's what this patch does. There are a few design decisions here worth explaining: 1. The new feature is attached to run-command's clean_on_exit feature. Partly this is convenience, since that feature already has a signal handler that deals with child cleanup. But it's also a meaningful connection. The main reason that dashed externals use clean_on_exit is to bind the two processes together. If somebody kills the parent with a signal, we propagate that to the child (in this instance with SIGINT, we do propagate but it doesn't matter because the original signal went to the whole process group). Likewise, we do not want the parent to go away until the child has done so. In a traditional Unix world, we'd probably accomplish this binding by just having the parent execve() the child directly. But since that doesn't work on Windows, everything goes through run_command's more spawn-like interface. 2. We do _not_ automatically waitpid() on any clean_on_exit children. For dashed externals this makes sense; we know that the parent is doing nothing but waiting for the child to exit anyway. But with other children, it's possible that the child, after getting the signal, could be waiting on the parent to do something (like closing a descriptor). If we were to wait on such a child, we'd end up in a deadlock. So this errs on the side of caution, and lets callers enable the feature explicitly. 3. When we send children the cleanup signal, we send all the signals first, before waiting on any children. This is to avoid the case where one child might be waiting on another one to exit, causing a deadlock. We inform all of them that it's time to die before reaping any. In practice, there is only ever one dashed external run from a given process, so this doesn't matter much now. But it future-proofs us if other callers start using the wait_after_clean mechanism. There's no automated test here, because it would end up racy and unportable. But it's easy to reproduce the situation by running the log command given above and hitting ^C. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2017-01-07 09:22:23 +08:00
if (p->process && p->process->wait_after_clean) {
execv_dashed_external: wait for child on signal death When you hit ^C to interrupt a git command going to a pager, this usually leaves the pager running. But when a dashed external is in use, the pager ends up in a funny state and quits (but only after eating one more character from the terminal!). This fixes it. Explaining the reason will require a little background. When git runs a pager, it's important for the git process to hang around and wait for the pager to finish, even though it has no more data to feed it. This is because git spawns the pager as a child, and thus the git process is the session leader on the terminal. After it dies, the pager will finish its current read from the terminal (eating the one character), and then get EIO trying to read again. When you hit ^C, that sends SIGINT to git and to the pager, and it's a similar situation. The pager ignores it, but the git process needs to hang around until the pager is done. We addressed that long ago in a3da882120 (pager: do wait_for_pager on signal death, 2009-01-22). But when you have a dashed external (or an alias pointing to a builtin, which will re-exec git for the builtin), there's an extra process in the mix. For instance, running: $ git -c alias.l=log l will end up with a process tree like: git (parent) \ git-log (child) \ less (pager) If you hit ^C, SIGINT goes to all of them. The pager ignores it, and the child git process will end up in wait_for_pager(). But the parent git process will die, and the usual EIO trouble happens. So we really want the parent git process to wait_for_pager(), but of course it doesn't know anything about the pager at all, since it was started by the child. However, we can have it wait on the git-log child, which in turn is waiting on the pager. And that's what this patch does. There are a few design decisions here worth explaining: 1. The new feature is attached to run-command's clean_on_exit feature. Partly this is convenience, since that feature already has a signal handler that deals with child cleanup. But it's also a meaningful connection. The main reason that dashed externals use clean_on_exit is to bind the two processes together. If somebody kills the parent with a signal, we propagate that to the child (in this instance with SIGINT, we do propagate but it doesn't matter because the original signal went to the whole process group). Likewise, we do not want the parent to go away until the child has done so. In a traditional Unix world, we'd probably accomplish this binding by just having the parent execve() the child directly. But since that doesn't work on Windows, everything goes through run_command's more spawn-like interface. 2. We do _not_ automatically waitpid() on any clean_on_exit children. For dashed externals this makes sense; we know that the parent is doing nothing but waiting for the child to exit anyway. But with other children, it's possible that the child, after getting the signal, could be waiting on the parent to do something (like closing a descriptor). If we were to wait on such a child, we'd end up in a deadlock. So this errs on the side of caution, and lets callers enable the feature explicitly. 3. When we send children the cleanup signal, we send all the signals first, before waiting on any children. This is to avoid the case where one child might be waiting on another one to exit, causing a deadlock. We inform all of them that it's time to die before reaping any. In practice, there is only ever one dashed external run from a given process, so this doesn't matter much now. But it future-proofs us if other callers start using the wait_after_clean mechanism. There's no automated test here, because it would end up racy and unportable. But it's easy to reproduce the situation by running the log command given above and hitting ^C. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2017-01-07 09:22:23 +08:00
p->next = children_to_wait_for;
children_to_wait_for = p;
} else {
if (!in_signal)
free(p);
}
}
while (children_to_wait_for) {
struct child_to_clean *p = children_to_wait_for;
children_to_wait_for = p->next;
while (waitpid(p->pid, NULL, 0) < 0 && errno == EINTR)
; /* spin waiting for process exit or error */
pager: don't use unsafe functions in signal handlers Since the commit a3da8821208d (pager: do wait_for_pager on signal death), we call wait_for_pager() in the pager's signal handler. The recent bug report revealed that this causes a deadlock in glibc at aborting "git log" [*1*]. When this happens, git process is left unterminated, and it can't be killed by SIGTERM but only by SIGKILL. The problem is that wait_for_pager() function does more than waiting for pager process's termination, but it does cleanups and printing errors. Unfortunately, the functions that may be used in a signal handler are very limited [*2*]. Particularly, malloc(), free() and the variants can't be used in a signal handler because they take a mutex internally in glibc. This was the cause of the deadlock above. Other than the direct calls of malloc/free, many functions calling malloc/free can't be used. strerror() is such one, either. Also the usage of fflush() and printf() in a signal handler is bad, although it seems working so far. In a safer side, we should avoid them, too. This patch tries to reduce the calls of such functions in signal handlers. wait_for_signal() takes a flag and avoids the unsafe calls. Also, finish_command_in_signal() is introduced for the same reason. There the free() calls are removed, and only waits for the children without whining at errors. [*1*] https://bugzilla.opensuse.org/show_bug.cgi?id=942297 [*2*] http://pubs.opengroup.org/onlinepubs/9699919799/functions/V2_chap02.html#tag_15_04_03 Signed-off-by: Takashi Iwai <tiwai@suse.de> Reviewed-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-09-04 17:35:57 +08:00
if (!in_signal)
free(p);
}
}
static void cleanup_children_on_signal(int sig)
{
pager: don't use unsafe functions in signal handlers Since the commit a3da8821208d (pager: do wait_for_pager on signal death), we call wait_for_pager() in the pager's signal handler. The recent bug report revealed that this causes a deadlock in glibc at aborting "git log" [*1*]. When this happens, git process is left unterminated, and it can't be killed by SIGTERM but only by SIGKILL. The problem is that wait_for_pager() function does more than waiting for pager process's termination, but it does cleanups and printing errors. Unfortunately, the functions that may be used in a signal handler are very limited [*2*]. Particularly, malloc(), free() and the variants can't be used in a signal handler because they take a mutex internally in glibc. This was the cause of the deadlock above. Other than the direct calls of malloc/free, many functions calling malloc/free can't be used. strerror() is such one, either. Also the usage of fflush() and printf() in a signal handler is bad, although it seems working so far. In a safer side, we should avoid them, too. This patch tries to reduce the calls of such functions in signal handlers. wait_for_signal() takes a flag and avoids the unsafe calls. Also, finish_command_in_signal() is introduced for the same reason. There the free() calls are removed, and only waits for the children without whining at errors. [*1*] https://bugzilla.opensuse.org/show_bug.cgi?id=942297 [*2*] http://pubs.opengroup.org/onlinepubs/9699919799/functions/V2_chap02.html#tag_15_04_03 Signed-off-by: Takashi Iwai <tiwai@suse.de> Reviewed-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-09-04 17:35:57 +08:00
cleanup_children(sig, 1);
sigchain_pop(sig);
raise(sig);
}
static void cleanup_children_on_exit(void)
{
pager: don't use unsafe functions in signal handlers Since the commit a3da8821208d (pager: do wait_for_pager on signal death), we call wait_for_pager() in the pager's signal handler. The recent bug report revealed that this causes a deadlock in glibc at aborting "git log" [*1*]. When this happens, git process is left unterminated, and it can't be killed by SIGTERM but only by SIGKILL. The problem is that wait_for_pager() function does more than waiting for pager process's termination, but it does cleanups and printing errors. Unfortunately, the functions that may be used in a signal handler are very limited [*2*]. Particularly, malloc(), free() and the variants can't be used in a signal handler because they take a mutex internally in glibc. This was the cause of the deadlock above. Other than the direct calls of malloc/free, many functions calling malloc/free can't be used. strerror() is such one, either. Also the usage of fflush() and printf() in a signal handler is bad, although it seems working so far. In a safer side, we should avoid them, too. This patch tries to reduce the calls of such functions in signal handlers. wait_for_signal() takes a flag and avoids the unsafe calls. Also, finish_command_in_signal() is introduced for the same reason. There the free() calls are removed, and only waits for the children without whining at errors. [*1*] https://bugzilla.opensuse.org/show_bug.cgi?id=942297 [*2*] http://pubs.opengroup.org/onlinepubs/9699919799/functions/V2_chap02.html#tag_15_04_03 Signed-off-by: Takashi Iwai <tiwai@suse.de> Reviewed-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-09-04 17:35:57 +08:00
cleanup_children(SIGTERM, 0);
}
static void mark_child_for_cleanup(pid_t pid, struct child_process *process)
{
struct child_to_clean *p = xmalloc(sizeof(*p));
p->pid = pid;
p->process = process;
p->next = children_to_clean;
children_to_clean = p;
if (!installed_child_cleanup_handler) {
atexit(cleanup_children_on_exit);
sigchain_push_common(cleanup_children_on_signal);
installed_child_cleanup_handler = 1;
}
}
static void clear_child_for_cleanup(pid_t pid)
{
struct child_to_clean **pp;
for (pp = &children_to_clean; *pp; pp = &(*pp)->next) {
struct child_to_clean *clean_me = *pp;
if (clean_me->pid == pid) {
*pp = clean_me->next;
free(clean_me);
return;
}
}
}
static inline void close_pair(int fd[2])
{
close(fd[0]);
close(fd[1]);
}
int is_executable(const char *name)
{
struct stat st;
if (stat(name, &st) || /* stat, not lstat */
!S_ISREG(st.st_mode))
return 0;
#if defined(GIT_WINDOWS_NATIVE)
/*
* On Windows there is no executable bit. The file extension
* indicates whether it can be run as an executable, and Git
* has special-handling to detect scripts and launch them
* through the indicated script interpreter. We test for the
* file extension first because virus scanners may make
* it quite expensive to open many files.
*/
if (ends_with(name, ".exe"))
return S_IXUSR;
{
/*
* Now that we know it does not have an executable extension,
* peek into the file instead.
*/
char buf[3] = { 0 };
int n;
int fd = open(name, O_RDONLY);
st.st_mode &= ~S_IXUSR;
if (fd >= 0) {
n = read(fd, buf, 2);
if (n == 2)
/* look for a she-bang */
if (!strcmp(buf, "#!"))
st.st_mode |= S_IXUSR;
close(fd);
}
}
#endif
return st.st_mode & S_IXUSR;
}
#ifndef locate_in_PATH
/*
* Search $PATH for a command. This emulates the path search that
* execvp would perform, without actually executing the command so it
* can be used before fork() to prepare to run a command using
* execve() or after execvp() to diagnose why it failed.
*
* The caller should ensure that file contains no directory
* separators.
*
* Returns the path to the command, as found in $PATH or NULL if the
* command could not be found. The caller inherits ownership of the memory
* used to store the resultant path.
*
* This should not be used on Windows, where the $PATH search rules
* are more complicated (e.g., a search for "foo" should find
* "foo.exe").
*/
static char *locate_in_PATH(const char *file)
{
const char *p = getenv("PATH");
struct strbuf buf = STRBUF_INIT;
if (!p || !*p)
return NULL;
while (1) {
const char *end = strchrnul(p, ':');
strbuf_reset(&buf);
/* POSIX specifies an empty entry as the current directory. */
if (end != p) {
strbuf_add(&buf, p, end - p);
strbuf_addch(&buf, '/');
}
strbuf_addstr(&buf, file);
if (is_executable(buf.buf))
return strbuf_detach(&buf, NULL);
if (!*end)
break;
p = end + 1;
}
strbuf_release(&buf);
return NULL;
}
#endif
int exists_in_PATH(const char *command)
{
char *r = locate_in_PATH(command);
int found = r != NULL;
free(r);
return found;
}
int sane_execvp(const char *file, char * const argv[])
{
#ifndef GIT_WINDOWS_NATIVE
/*
* execvp() doesn't return, so we all we can do is tell trace2
* what we are about to do and let it leave a hint in the log
* (unless of course the execvp() fails).
*
* we skip this for Windows because the compat layer already
* has to emulate the execvp() call anyway.
*/
int exec_id = trace2_exec(file, (const char **)argv);
#endif
if (!execvp(file, argv))
return 0; /* cannot happen ;-) */
#ifndef GIT_WINDOWS_NATIVE
{
int ec = errno;
trace2_exec_result(exec_id, ec);
errno = ec;
}
#endif
/*
* When a command can't be found because one of the directories
* listed in $PATH is unsearchable, execvp reports EACCES, but
* careful usability testing (read: analysis of occasional bug
* reports) reveals that "No such file or directory" is more
* intuitive.
*
* We avoid commands with "/", because execvp will not do $PATH
* lookups in that case.
*
* The reassignment of EACCES to errno looks like a no-op below,
* but we need to protect against exists_in_PATH overwriting errno.
*/
if (errno == EACCES && !strchr(file, '/'))
errno = exists_in_PATH(file) ? EACCES : ENOENT;
else if (errno == ENOTDIR && !strchr(file, '/'))
errno = ENOENT;
return -1;
}
char *git_shell_path(void)
{
#ifndef GIT_WINDOWS_NATIVE
return xstrdup(SHELL_PATH);
#else
char *p = locate_in_PATH("sh");
convert_slashes(p);
return p;
#endif
}
static const char **prepare_shell_cmd(struct strvec *out, const char **argv)
{
if (!argv[0])
BUG("shell command is empty");
if (strcspn(argv[0], "|&;<>()$`\\\"' \t\n*?[#~=%") != strlen(argv[0])) {
strvec_push_nodup(out, git_shell_path());
strvec_push(out, "-c");
/*
* If we have no extra arguments, we do not even need to
* bother with the "$@" magic.
*/
if (!argv[1])
strvec_push(out, argv[0]);
else
strvec_pushf(out, "%s \"$@\"", argv[0]);
}
strvec_pushv(out, argv);
return out->v;
}
#ifndef GIT_WINDOWS_NATIVE
static int child_notifier = -1;
enum child_errcode {
CHILD_ERR_CHDIR,
CHILD_ERR_DUP2,
CHILD_ERR_CLOSE,
CHILD_ERR_SIGPROCMASK,
CHILD_ERR_SILENT,
CHILD_ERR_ERRNO
};
struct child_err {
enum child_errcode err;
int syserr; /* errno */
};
static void child_die(enum child_errcode err)
{
struct child_err buf;
buf.err = err;
buf.syserr = errno;
/* write(2) on buf smaller than PIPE_BUF (min 512) is atomic: */
xwrite(child_notifier, &buf, sizeof(buf));
_exit(1);
}
static void child_dup2(int fd, int to)
{
if (dup2(fd, to) < 0)
child_die(CHILD_ERR_DUP2);
}
static void child_close(int fd)
{
if (close(fd))
child_die(CHILD_ERR_CLOSE);
}
static void child_close_pair(int fd[2])
{
child_close(fd[0]);
child_close(fd[1]);
}
static void child_error_fn(const char *err UNUSED, va_list params UNUSED)
{
const char msg[] = "error() should not be called in child\n";
xwrite(2, msg, sizeof(msg) - 1);
}
static void child_warn_fn(const char *err UNUSED, va_list params UNUSED)
{
const char msg[] = "warn() should not be called in child\n";
xwrite(2, msg, sizeof(msg) - 1);
}
static void NORETURN child_die_fn(const char *err UNUSED, va_list params UNUSED)
{
const char msg[] = "die() should not be called in child\n";
xwrite(2, msg, sizeof(msg) - 1);
_exit(2);
}
/* this runs in the parent process */
static void child_err_spew(struct child_process *cmd, struct child_err *cerr)
{
static void (*old_errfn)(const char *err, va_list params);
report_fn die_message_routine = get_die_message_routine();
old_errfn = get_error_routine();
set_error_routine(die_message_routine);
errno = cerr->syserr;
switch (cerr->err) {
case CHILD_ERR_CHDIR:
error_errno("exec '%s': cd to '%s' failed",
run-command API: remove "argv" member, always use "args" Remove the "argv" member from the run-command API, ever since "args" was added in c460c0ecdca (run-command: store an optional argv_array, 2014-05-15) being able to provide either "argv" or "args" has led to some confusion and bugs. If we hadn't gone in that direction and only had an "argv" our problems wouldn't have been solved either, as noted in [1] (and in the documentation amended here) it comes with inherent memory management issues: The caller would have to hang on to the "argv" until the run-command API was finished. If the "argv" was an argument to main() this wasn't an issue, but if it it was manually constructed using the API might be painful. We also have a recent report[2] of a user of the API segfaulting, which is a direct result of it being complex to use. This commit addresses the root cause of that bug. This change is larger than I'd like, but there's no easy way to avoid it that wouldn't involve even more verbose intermediate steps. We use the "argv" as the source of truth over the "args", so we need to change all parts of run-command.[ch] itself, as well as the trace2 logging at the same time. The resulting Windows-specific code in start_command() is a bit nasty, as we're now assigning to a strvec's "v" member, instead of to our own "argv". There was a suggestion of some alternate approaches in reply to an earlier version of this commit[3], but let's leave larger a larger and needless refactoring of this code for now. 1. http://lore.kernel.org/git/YT6BnnXeAWn8BycF@coredump.intra.peff.net 2. https://lore.kernel.org/git/20211120194048.12125-1-ematsumiya@suse.de/ 3. https://lore.kernel.org/git/patch-5.5-ea1011f7473-20211122T153605Z-avarab@gmail.com/ Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-11-26 06:52:22 +08:00
cmd->args.v[0], cmd->dir);
break;
case CHILD_ERR_DUP2:
error_errno("dup2() in child failed");
break;
case CHILD_ERR_CLOSE:
error_errno("close() in child failed");
break;
case CHILD_ERR_SIGPROCMASK:
error_errno("sigprocmask failed restoring signals");
break;
case CHILD_ERR_SILENT:
break;
case CHILD_ERR_ERRNO:
run-command API: remove "argv" member, always use "args" Remove the "argv" member from the run-command API, ever since "args" was added in c460c0ecdca (run-command: store an optional argv_array, 2014-05-15) being able to provide either "argv" or "args" has led to some confusion and bugs. If we hadn't gone in that direction and only had an "argv" our problems wouldn't have been solved either, as noted in [1] (and in the documentation amended here) it comes with inherent memory management issues: The caller would have to hang on to the "argv" until the run-command API was finished. If the "argv" was an argument to main() this wasn't an issue, but if it it was manually constructed using the API might be painful. We also have a recent report[2] of a user of the API segfaulting, which is a direct result of it being complex to use. This commit addresses the root cause of that bug. This change is larger than I'd like, but there's no easy way to avoid it that wouldn't involve even more verbose intermediate steps. We use the "argv" as the source of truth over the "args", so we need to change all parts of run-command.[ch] itself, as well as the trace2 logging at the same time. The resulting Windows-specific code in start_command() is a bit nasty, as we're now assigning to a strvec's "v" member, instead of to our own "argv". There was a suggestion of some alternate approaches in reply to an earlier version of this commit[3], but let's leave larger a larger and needless refactoring of this code for now. 1. http://lore.kernel.org/git/YT6BnnXeAWn8BycF@coredump.intra.peff.net 2. https://lore.kernel.org/git/20211120194048.12125-1-ematsumiya@suse.de/ 3. https://lore.kernel.org/git/patch-5.5-ea1011f7473-20211122T153605Z-avarab@gmail.com/ Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-11-26 06:52:22 +08:00
error_errno("cannot exec '%s'", cmd->args.v[0]);
break;
}
set_error_routine(old_errfn);
}
static int prepare_cmd(struct strvec *out, const struct child_process *cmd)
{
run-command API: remove "argv" member, always use "args" Remove the "argv" member from the run-command API, ever since "args" was added in c460c0ecdca (run-command: store an optional argv_array, 2014-05-15) being able to provide either "argv" or "args" has led to some confusion and bugs. If we hadn't gone in that direction and only had an "argv" our problems wouldn't have been solved either, as noted in [1] (and in the documentation amended here) it comes with inherent memory management issues: The caller would have to hang on to the "argv" until the run-command API was finished. If the "argv" was an argument to main() this wasn't an issue, but if it it was manually constructed using the API might be painful. We also have a recent report[2] of a user of the API segfaulting, which is a direct result of it being complex to use. This commit addresses the root cause of that bug. This change is larger than I'd like, but there's no easy way to avoid it that wouldn't involve even more verbose intermediate steps. We use the "argv" as the source of truth over the "args", so we need to change all parts of run-command.[ch] itself, as well as the trace2 logging at the same time. The resulting Windows-specific code in start_command() is a bit nasty, as we're now assigning to a strvec's "v" member, instead of to our own "argv". There was a suggestion of some alternate approaches in reply to an earlier version of this commit[3], but let's leave larger a larger and needless refactoring of this code for now. 1. http://lore.kernel.org/git/YT6BnnXeAWn8BycF@coredump.intra.peff.net 2. https://lore.kernel.org/git/20211120194048.12125-1-ematsumiya@suse.de/ 3. https://lore.kernel.org/git/patch-5.5-ea1011f7473-20211122T153605Z-avarab@gmail.com/ Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-11-26 06:52:22 +08:00
if (!cmd->args.v[0])
BUG("command is empty");
/*
* Add SHELL_PATH so in the event exec fails with ENOEXEC we can
* attempt to interpret the command with 'sh'.
*/
strvec_push(out, SHELL_PATH);
if (cmd->git_cmd) {
run-command API: remove "argv" member, always use "args" Remove the "argv" member from the run-command API, ever since "args" was added in c460c0ecdca (run-command: store an optional argv_array, 2014-05-15) being able to provide either "argv" or "args" has led to some confusion and bugs. If we hadn't gone in that direction and only had an "argv" our problems wouldn't have been solved either, as noted in [1] (and in the documentation amended here) it comes with inherent memory management issues: The caller would have to hang on to the "argv" until the run-command API was finished. If the "argv" was an argument to main() this wasn't an issue, but if it it was manually constructed using the API might be painful. We also have a recent report[2] of a user of the API segfaulting, which is a direct result of it being complex to use. This commit addresses the root cause of that bug. This change is larger than I'd like, but there's no easy way to avoid it that wouldn't involve even more verbose intermediate steps. We use the "argv" as the source of truth over the "args", so we need to change all parts of run-command.[ch] itself, as well as the trace2 logging at the same time. The resulting Windows-specific code in start_command() is a bit nasty, as we're now assigning to a strvec's "v" member, instead of to our own "argv". There was a suggestion of some alternate approaches in reply to an earlier version of this commit[3], but let's leave larger a larger and needless refactoring of this code for now. 1. http://lore.kernel.org/git/YT6BnnXeAWn8BycF@coredump.intra.peff.net 2. https://lore.kernel.org/git/20211120194048.12125-1-ematsumiya@suse.de/ 3. https://lore.kernel.org/git/patch-5.5-ea1011f7473-20211122T153605Z-avarab@gmail.com/ Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-11-26 06:52:22 +08:00
prepare_git_cmd(out, cmd->args.v);
} else if (cmd->use_shell) {
run-command API: remove "argv" member, always use "args" Remove the "argv" member from the run-command API, ever since "args" was added in c460c0ecdca (run-command: store an optional argv_array, 2014-05-15) being able to provide either "argv" or "args" has led to some confusion and bugs. If we hadn't gone in that direction and only had an "argv" our problems wouldn't have been solved either, as noted in [1] (and in the documentation amended here) it comes with inherent memory management issues: The caller would have to hang on to the "argv" until the run-command API was finished. If the "argv" was an argument to main() this wasn't an issue, but if it it was manually constructed using the API might be painful. We also have a recent report[2] of a user of the API segfaulting, which is a direct result of it being complex to use. This commit addresses the root cause of that bug. This change is larger than I'd like, but there's no easy way to avoid it that wouldn't involve even more verbose intermediate steps. We use the "argv" as the source of truth over the "args", so we need to change all parts of run-command.[ch] itself, as well as the trace2 logging at the same time. The resulting Windows-specific code in start_command() is a bit nasty, as we're now assigning to a strvec's "v" member, instead of to our own "argv". There was a suggestion of some alternate approaches in reply to an earlier version of this commit[3], but let's leave larger a larger and needless refactoring of this code for now. 1. http://lore.kernel.org/git/YT6BnnXeAWn8BycF@coredump.intra.peff.net 2. https://lore.kernel.org/git/20211120194048.12125-1-ematsumiya@suse.de/ 3. https://lore.kernel.org/git/patch-5.5-ea1011f7473-20211122T153605Z-avarab@gmail.com/ Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-11-26 06:52:22 +08:00
prepare_shell_cmd(out, cmd->args.v);
} else {
run-command API: remove "argv" member, always use "args" Remove the "argv" member from the run-command API, ever since "args" was added in c460c0ecdca (run-command: store an optional argv_array, 2014-05-15) being able to provide either "argv" or "args" has led to some confusion and bugs. If we hadn't gone in that direction and only had an "argv" our problems wouldn't have been solved either, as noted in [1] (and in the documentation amended here) it comes with inherent memory management issues: The caller would have to hang on to the "argv" until the run-command API was finished. If the "argv" was an argument to main() this wasn't an issue, but if it it was manually constructed using the API might be painful. We also have a recent report[2] of a user of the API segfaulting, which is a direct result of it being complex to use. This commit addresses the root cause of that bug. This change is larger than I'd like, but there's no easy way to avoid it that wouldn't involve even more verbose intermediate steps. We use the "argv" as the source of truth over the "args", so we need to change all parts of run-command.[ch] itself, as well as the trace2 logging at the same time. The resulting Windows-specific code in start_command() is a bit nasty, as we're now assigning to a strvec's "v" member, instead of to our own "argv". There was a suggestion of some alternate approaches in reply to an earlier version of this commit[3], but let's leave larger a larger and needless refactoring of this code for now. 1. http://lore.kernel.org/git/YT6BnnXeAWn8BycF@coredump.intra.peff.net 2. https://lore.kernel.org/git/20211120194048.12125-1-ematsumiya@suse.de/ 3. https://lore.kernel.org/git/patch-5.5-ea1011f7473-20211122T153605Z-avarab@gmail.com/ Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-11-26 06:52:22 +08:00
strvec_pushv(out, cmd->args.v);
}
/*
* If there are no dir separator characters in the command then perform
* a path lookup and use the resolved path as the command to exec. If
* there are dir separator characters, we have exec attempt to invoke
* the command directly.
*/
if (!has_dir_sep(out->v[1])) {
char *program = locate_in_PATH(out->v[1]);
if (program) {
free((char *)out->v[1]);
out->v[1] = program;
} else {
strvec_clear(out);
errno = ENOENT;
return -1;
}
}
return 0;
}
static char **prep_childenv(const char *const *deltaenv)
{
extern char **environ;
char **childenv;
struct string_list env = STRING_LIST_INIT_DUP;
struct strbuf key = STRBUF_INIT;
const char *const *p;
int i;
/* Construct a sorted string list consisting of the current environ */
for (p = (const char *const *) environ; p && *p; p++) {
const char *equals = strchr(*p, '=');
if (equals) {
strbuf_reset(&key);
strbuf_add(&key, *p, equals - *p);
string_list_append(&env, key.buf)->util = (void *) *p;
} else {
string_list_append(&env, *p)->util = (void *) *p;
}
}
string_list_sort(&env);
/* Merge in 'deltaenv' with the current environ */
for (p = deltaenv; p && *p; p++) {
const char *equals = strchr(*p, '=');
if (equals) {
/* ('key=value'), insert or replace entry */
strbuf_reset(&key);
strbuf_add(&key, *p, equals - *p);
string_list_insert(&env, key.buf)->util = (void *) *p;
} else {
/* otherwise ('key') remove existing entry */
string_list_remove(&env, *p, 0);
}
}
/* Create an array of 'char *' to be used as the childenv */
ALLOC_ARRAY(childenv, env.nr + 1);
for (i = 0; i < env.nr; i++)
childenv[i] = env.items[i].util;
childenv[env.nr] = NULL;
string_list_clear(&env, 0);
strbuf_release(&key);
return childenv;
}
struct atfork_state {
#ifndef NO_PTHREADS
int cs;
#endif
sigset_t old;
};
#define CHECK_BUG(err, msg) \
do { \
int e = (err); \
if (e) \
BUG("%s: %s", msg, strerror(e)); \
} while(0)
static void atfork_prepare(struct atfork_state *as)
{
sigset_t all;
if (sigfillset(&all))
die_errno("sigfillset");
#ifdef NO_PTHREADS
if (sigprocmask(SIG_SETMASK, &all, &as->old))
die_errno("sigprocmask");
#else
CHECK_BUG(pthread_sigmask(SIG_SETMASK, &all, &as->old),
"blocking all signals");
CHECK_BUG(pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, &as->cs),
"disabling cancellation");
#endif
}
static void atfork_parent(struct atfork_state *as)
{
#ifdef NO_PTHREADS
if (sigprocmask(SIG_SETMASK, &as->old, NULL))
die_errno("sigprocmask");
#else
CHECK_BUG(pthread_setcancelstate(as->cs, NULL),
"re-enabling cancellation");
CHECK_BUG(pthread_sigmask(SIG_SETMASK, &as->old, NULL),
"restoring signal mask");
#endif
}
#endif /* GIT_WINDOWS_NATIVE */
static inline void set_cloexec(int fd)
{
int flags = fcntl(fd, F_GETFD);
if (flags >= 0)
fcntl(fd, F_SETFD, flags | FD_CLOEXEC);
}
pager: don't use unsafe functions in signal handlers Since the commit a3da8821208d (pager: do wait_for_pager on signal death), we call wait_for_pager() in the pager's signal handler. The recent bug report revealed that this causes a deadlock in glibc at aborting "git log" [*1*]. When this happens, git process is left unterminated, and it can't be killed by SIGTERM but only by SIGKILL. The problem is that wait_for_pager() function does more than waiting for pager process's termination, but it does cleanups and printing errors. Unfortunately, the functions that may be used in a signal handler are very limited [*2*]. Particularly, malloc(), free() and the variants can't be used in a signal handler because they take a mutex internally in glibc. This was the cause of the deadlock above. Other than the direct calls of malloc/free, many functions calling malloc/free can't be used. strerror() is such one, either. Also the usage of fflush() and printf() in a signal handler is bad, although it seems working so far. In a safer side, we should avoid them, too. This patch tries to reduce the calls of such functions in signal handlers. wait_for_signal() takes a flag and avoids the unsafe calls. Also, finish_command_in_signal() is introduced for the same reason. There the free() calls are removed, and only waits for the children without whining at errors. [*1*] https://bugzilla.opensuse.org/show_bug.cgi?id=942297 [*2*] http://pubs.opengroup.org/onlinepubs/9699919799/functions/V2_chap02.html#tag_15_04_03 Signed-off-by: Takashi Iwai <tiwai@suse.de> Reviewed-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-09-04 17:35:57 +08:00
static int wait_or_whine(pid_t pid, const char *argv0, int in_signal)
{
int status, code = -1;
pid_t waiting;
int failed_errno = 0;
while ((waiting = waitpid(pid, &status, 0)) < 0 && errno == EINTR)
; /* nothing */
if (waiting < 0) {
failed_errno = errno;
run-command: unify signal and regular logic for wait_or_whine() Since 507d7804c0 (pager: don't use unsafe functions in signal handlers, 2015-09-04), we have a separate code path in wait_or_whine() for the case that we're in a signal handler. But that code path misses some of the cases handled by the main logic. This was improved in be8fc53e36 (pager: properly log pager exit code when signalled, 2021-02-02), but that covered only case: actually returning the correct error code. But there are some other cases: - if waitpid() returns failure, we wouldn't notice and would look at uninitialized garbage in the status variable; it's not clear if it's possible to trigger this or not - if the process exited by signal, then we would still report "-1" rather than the correct signal code This latter case even had a test added in be8fc53e36, but it doesn't work reliably. It sets the pager command to: >pager-used; test-tool sigchain The latter command will die by signal, but because there are multiple commands, there will be a shell in between. And it's the shell whose waitpid() call will see the signal death, and it will then exit with code 143, which is what Git will see. To make matters even more confusing, some shells (such as bash) will realize that there's nothing for the shell to do after test-tool finishes, and will turn it into an exec. So the test was only checking what it thought when /bin/sh points to a shell like bash (we're relying on the shell used internally by Git to spawn sub-commands here, so even running the test under bash would not be enough). This patch adjusts the tests to explicitly call "exec" in the pager command, which produces a consistent outcome regardless of shell. Note that without the code change in this patch it _should_ fail reliably, but doesn't. That test, like its siblings, tries to trigger SIGPIPE in the git-log process writing to the pager, but only do so racily. That will be fixed in a follow-on patch. For the code change here, we have two options: - we can teach the in_signal code to handle WIFSIGNALED() - we can stop returning early when in_signal is set, and instead annotate individual calls that we need to skip in this case The former is a simpler patch, but means we're essentially duplicating all of the logic. So instead I went with the latter. The result is a bigger patch, and we do run the risk of new code being added but forgetting to handle in_signal. But in the long run it seems more maintainable. I've skipped any non-trivial calls for the in_signal case, like calling error(). We'll also skip the call to clear_child_for_cleanup(), as we were before. This is arguably the wrong thing to do, since we wouldn't want to try to clean it up again. But: - we can't call it as-is, because it calls free(), which we must avoid in a signal handler (we'd have to pass in_signal so it can skip the free() call) - we'll only go through the list of children to clean once, since our cleanup_children_on_signal() handler pops itself after running (and then re-raises, so eventually we'd just exit). So this cleanup only matters if a process is on the cleanup list _and_ it has a separate handler to clean itself up. Which is questionable in the first place (and AFAIK we do not do). - double-cleanup isn't actually that bad anyway. waitpid() will just return an error, which we won't even report because of in_signal. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-11-23 05:28:13 +08:00
if (!in_signal)
error_errno("waitpid for %s failed", argv0);
} else if (waiting != pid) {
run-command: unify signal and regular logic for wait_or_whine() Since 507d7804c0 (pager: don't use unsafe functions in signal handlers, 2015-09-04), we have a separate code path in wait_or_whine() for the case that we're in a signal handler. But that code path misses some of the cases handled by the main logic. This was improved in be8fc53e36 (pager: properly log pager exit code when signalled, 2021-02-02), but that covered only case: actually returning the correct error code. But there are some other cases: - if waitpid() returns failure, we wouldn't notice and would look at uninitialized garbage in the status variable; it's not clear if it's possible to trigger this or not - if the process exited by signal, then we would still report "-1" rather than the correct signal code This latter case even had a test added in be8fc53e36, but it doesn't work reliably. It sets the pager command to: >pager-used; test-tool sigchain The latter command will die by signal, but because there are multiple commands, there will be a shell in between. And it's the shell whose waitpid() call will see the signal death, and it will then exit with code 143, which is what Git will see. To make matters even more confusing, some shells (such as bash) will realize that there's nothing for the shell to do after test-tool finishes, and will turn it into an exec. So the test was only checking what it thought when /bin/sh points to a shell like bash (we're relying on the shell used internally by Git to spawn sub-commands here, so even running the test under bash would not be enough). This patch adjusts the tests to explicitly call "exec" in the pager command, which produces a consistent outcome regardless of shell. Note that without the code change in this patch it _should_ fail reliably, but doesn't. That test, like its siblings, tries to trigger SIGPIPE in the git-log process writing to the pager, but only do so racily. That will be fixed in a follow-on patch. For the code change here, we have two options: - we can teach the in_signal code to handle WIFSIGNALED() - we can stop returning early when in_signal is set, and instead annotate individual calls that we need to skip in this case The former is a simpler patch, but means we're essentially duplicating all of the logic. So instead I went with the latter. The result is a bigger patch, and we do run the risk of new code being added but forgetting to handle in_signal. But in the long run it seems more maintainable. I've skipped any non-trivial calls for the in_signal case, like calling error(). We'll also skip the call to clear_child_for_cleanup(), as we were before. This is arguably the wrong thing to do, since we wouldn't want to try to clean it up again. But: - we can't call it as-is, because it calls free(), which we must avoid in a signal handler (we'd have to pass in_signal so it can skip the free() call) - we'll only go through the list of children to clean once, since our cleanup_children_on_signal() handler pops itself after running (and then re-raises, so eventually we'd just exit). So this cleanup only matters if a process is on the cleanup list _and_ it has a separate handler to clean itself up. Which is questionable in the first place (and AFAIK we do not do). - double-cleanup isn't actually that bad anyway. waitpid() will just return an error, which we won't even report because of in_signal. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-11-23 05:28:13 +08:00
if (!in_signal)
error("waitpid is confused (%s)", argv0);
} else if (WIFSIGNALED(status)) {
code = WTERMSIG(status);
run-command: unify signal and regular logic for wait_or_whine() Since 507d7804c0 (pager: don't use unsafe functions in signal handlers, 2015-09-04), we have a separate code path in wait_or_whine() for the case that we're in a signal handler. But that code path misses some of the cases handled by the main logic. This was improved in be8fc53e36 (pager: properly log pager exit code when signalled, 2021-02-02), but that covered only case: actually returning the correct error code. But there are some other cases: - if waitpid() returns failure, we wouldn't notice and would look at uninitialized garbage in the status variable; it's not clear if it's possible to trigger this or not - if the process exited by signal, then we would still report "-1" rather than the correct signal code This latter case even had a test added in be8fc53e36, but it doesn't work reliably. It sets the pager command to: >pager-used; test-tool sigchain The latter command will die by signal, but because there are multiple commands, there will be a shell in between. And it's the shell whose waitpid() call will see the signal death, and it will then exit with code 143, which is what Git will see. To make matters even more confusing, some shells (such as bash) will realize that there's nothing for the shell to do after test-tool finishes, and will turn it into an exec. So the test was only checking what it thought when /bin/sh points to a shell like bash (we're relying on the shell used internally by Git to spawn sub-commands here, so even running the test under bash would not be enough). This patch adjusts the tests to explicitly call "exec" in the pager command, which produces a consistent outcome regardless of shell. Note that without the code change in this patch it _should_ fail reliably, but doesn't. That test, like its siblings, tries to trigger SIGPIPE in the git-log process writing to the pager, but only do so racily. That will be fixed in a follow-on patch. For the code change here, we have two options: - we can teach the in_signal code to handle WIFSIGNALED() - we can stop returning early when in_signal is set, and instead annotate individual calls that we need to skip in this case The former is a simpler patch, but means we're essentially duplicating all of the logic. So instead I went with the latter. The result is a bigger patch, and we do run the risk of new code being added but forgetting to handle in_signal. But in the long run it seems more maintainable. I've skipped any non-trivial calls for the in_signal case, like calling error(). We'll also skip the call to clear_child_for_cleanup(), as we were before. This is arguably the wrong thing to do, since we wouldn't want to try to clean it up again. But: - we can't call it as-is, because it calls free(), which we must avoid in a signal handler (we'd have to pass in_signal so it can skip the free() call) - we'll only go through the list of children to clean once, since our cleanup_children_on_signal() handler pops itself after running (and then re-raises, so eventually we'd just exit). So this cleanup only matters if a process is on the cleanup list _and_ it has a separate handler to clean itself up. Which is questionable in the first place (and AFAIK we do not do). - double-cleanup isn't actually that bad anyway. waitpid() will just return an error, which we won't even report because of in_signal. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-11-23 05:28:13 +08:00
if (!in_signal && code != SIGINT && code != SIGQUIT && code != SIGPIPE)
error("%s died of signal %d", argv0, code);
/*
* This return value is chosen so that code & 0xff
* mimics the exit code that a POSIX shell would report for
* a program that died from this signal.
*/
run-command: encode signal death as a positive integer When a sub-command dies due to a signal, we encode the signal number into the numeric exit status as "signal - 128". This is easy to identify (versus a regular positive error code), and when cast to an unsigned integer (e.g., by feeding it to exit), matches what a POSIX shell would return when reporting a signal death in $? or through its own exit code. So we have a negative value inside the code, but once it passes across an exit() barrier, it looks positive (and any code we receive from a sub-shell will have the positive form). E.g., death by SIGPIPE (signal 13) will look like -115 to us in inside git, but will end up as 141 when we call exit() with it. And a program killed by SIGPIPE but run via the shell will come to us with an exit code of 141. Unfortunately, this means that when the "use_shell" option is set, we need to be on the lookout for _both_ forms. We might or might not have actually invoked the shell (because we optimize out some useless shell calls). If we didn't invoke the shell, we will will see the sub-process's signal death directly, and run-command converts it into a negative value. But if we did invoke the shell, we will see the shell's 128+signal exit status. To be thorough, we would need to check both, or cast the value to an unsigned char (after checking that it is not -1, which is a magic error value). Fortunately, most callsites do not care at all whether the exit was from a code or from a signal; they merely check for a non-zero status, and sometimes propagate the error via exit(). But for the callers that do care, we can make life slightly easier by just using the consistent positive form. This actually fixes two minor bugs: 1. In launch_editor, we check whether the editor died from SIGINT or SIGQUIT. But we checked only the negative form, meaning that we would fail to notice a signal death exit code which was propagated through the shell. 2. In handle_alias, we assume that a negative return value from run_command means that errno tells us something interesting (like a fork failure, or ENOENT). Otherwise, we simply propagate the exit code. Negative signal death codes confuse us, and we print a useless "unable to run alias 'foo': Success" message. By encoding signal deaths using the positive form, the existing code just propagates it as it would a normal non-zero exit code. The downside is that callers of run_command can no longer differentiate between a signal received directly by the sub-process, and one propagated. However, no caller currently cares, and since we already optimize out some calls to the shell under the hood, that distinction is not something that should be relied upon by callers. Fix the same logic in t/test-terminal.perl for consistency [jc: raised by Jonathan in the discussion]. Signed-off-by: Jeff King <peff@peff.net> Acked-by: Johannes Sixt <j6t@kdbg.org> Reviewed-by: Jonathan Nieder <jrnieder@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-01-05 22:49:49 +08:00
code += 128;
} else if (WIFEXITED(status)) {
code = WEXITSTATUS(status);
} else {
run-command: unify signal and regular logic for wait_or_whine() Since 507d7804c0 (pager: don't use unsafe functions in signal handlers, 2015-09-04), we have a separate code path in wait_or_whine() for the case that we're in a signal handler. But that code path misses some of the cases handled by the main logic. This was improved in be8fc53e36 (pager: properly log pager exit code when signalled, 2021-02-02), but that covered only case: actually returning the correct error code. But there are some other cases: - if waitpid() returns failure, we wouldn't notice and would look at uninitialized garbage in the status variable; it's not clear if it's possible to trigger this or not - if the process exited by signal, then we would still report "-1" rather than the correct signal code This latter case even had a test added in be8fc53e36, but it doesn't work reliably. It sets the pager command to: >pager-used; test-tool sigchain The latter command will die by signal, but because there are multiple commands, there will be a shell in between. And it's the shell whose waitpid() call will see the signal death, and it will then exit with code 143, which is what Git will see. To make matters even more confusing, some shells (such as bash) will realize that there's nothing for the shell to do after test-tool finishes, and will turn it into an exec. So the test was only checking what it thought when /bin/sh points to a shell like bash (we're relying on the shell used internally by Git to spawn sub-commands here, so even running the test under bash would not be enough). This patch adjusts the tests to explicitly call "exec" in the pager command, which produces a consistent outcome regardless of shell. Note that without the code change in this patch it _should_ fail reliably, but doesn't. That test, like its siblings, tries to trigger SIGPIPE in the git-log process writing to the pager, but only do so racily. That will be fixed in a follow-on patch. For the code change here, we have two options: - we can teach the in_signal code to handle WIFSIGNALED() - we can stop returning early when in_signal is set, and instead annotate individual calls that we need to skip in this case The former is a simpler patch, but means we're essentially duplicating all of the logic. So instead I went with the latter. The result is a bigger patch, and we do run the risk of new code being added but forgetting to handle in_signal. But in the long run it seems more maintainable. I've skipped any non-trivial calls for the in_signal case, like calling error(). We'll also skip the call to clear_child_for_cleanup(), as we were before. This is arguably the wrong thing to do, since we wouldn't want to try to clean it up again. But: - we can't call it as-is, because it calls free(), which we must avoid in a signal handler (we'd have to pass in_signal so it can skip the free() call) - we'll only go through the list of children to clean once, since our cleanup_children_on_signal() handler pops itself after running (and then re-raises, so eventually we'd just exit). So this cleanup only matters if a process is on the cleanup list _and_ it has a separate handler to clean itself up. Which is questionable in the first place (and AFAIK we do not do). - double-cleanup isn't actually that bad anyway. waitpid() will just return an error, which we won't even report because of in_signal. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-11-23 05:28:13 +08:00
if (!in_signal)
error("waitpid is confused (%s)", argv0);
}
run-command: unify signal and regular logic for wait_or_whine() Since 507d7804c0 (pager: don't use unsafe functions in signal handlers, 2015-09-04), we have a separate code path in wait_or_whine() for the case that we're in a signal handler. But that code path misses some of the cases handled by the main logic. This was improved in be8fc53e36 (pager: properly log pager exit code when signalled, 2021-02-02), but that covered only case: actually returning the correct error code. But there are some other cases: - if waitpid() returns failure, we wouldn't notice and would look at uninitialized garbage in the status variable; it's not clear if it's possible to trigger this or not - if the process exited by signal, then we would still report "-1" rather than the correct signal code This latter case even had a test added in be8fc53e36, but it doesn't work reliably. It sets the pager command to: >pager-used; test-tool sigchain The latter command will die by signal, but because there are multiple commands, there will be a shell in between. And it's the shell whose waitpid() call will see the signal death, and it will then exit with code 143, which is what Git will see. To make matters even more confusing, some shells (such as bash) will realize that there's nothing for the shell to do after test-tool finishes, and will turn it into an exec. So the test was only checking what it thought when /bin/sh points to a shell like bash (we're relying on the shell used internally by Git to spawn sub-commands here, so even running the test under bash would not be enough). This patch adjusts the tests to explicitly call "exec" in the pager command, which produces a consistent outcome regardless of shell. Note that without the code change in this patch it _should_ fail reliably, but doesn't. That test, like its siblings, tries to trigger SIGPIPE in the git-log process writing to the pager, but only do so racily. That will be fixed in a follow-on patch. For the code change here, we have two options: - we can teach the in_signal code to handle WIFSIGNALED() - we can stop returning early when in_signal is set, and instead annotate individual calls that we need to skip in this case The former is a simpler patch, but means we're essentially duplicating all of the logic. So instead I went with the latter. The result is a bigger patch, and we do run the risk of new code being added but forgetting to handle in_signal. But in the long run it seems more maintainable. I've skipped any non-trivial calls for the in_signal case, like calling error(). We'll also skip the call to clear_child_for_cleanup(), as we were before. This is arguably the wrong thing to do, since we wouldn't want to try to clean it up again. But: - we can't call it as-is, because it calls free(), which we must avoid in a signal handler (we'd have to pass in_signal so it can skip the free() call) - we'll only go through the list of children to clean once, since our cleanup_children_on_signal() handler pops itself after running (and then re-raises, so eventually we'd just exit). So this cleanup only matters if a process is on the cleanup list _and_ it has a separate handler to clean itself up. Which is questionable in the first place (and AFAIK we do not do). - double-cleanup isn't actually that bad anyway. waitpid() will just return an error, which we won't even report because of in_signal. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-11-23 05:28:13 +08:00
if (!in_signal)
clear_child_for_cleanup(pid);
errno = failed_errno;
return code;
}
static void trace_add_env(struct strbuf *dst, const char *const *deltaenv)
{
struct string_list envs = STRING_LIST_INIT_DUP;
const char *const *e;
int i;
int printed_unset = 0;
/* Last one wins, see run-command.c:prep_childenv() for context */
for (e = deltaenv; e && *e; e++) {
struct strbuf key = STRBUF_INIT;
char *equals = strchr(*e, '=');
if (equals) {
strbuf_add(&key, *e, equals - *e);
string_list_insert(&envs, key.buf)->util = equals + 1;
} else {
string_list_insert(&envs, *e)->util = NULL;
}
strbuf_release(&key);
}
/* "unset X Y...;" */
for (i = 0; i < envs.nr; i++) {
const char *var = envs.items[i].string;
const char *val = envs.items[i].util;
if (val || !getenv(var))
continue;
if (!printed_unset) {
strbuf_addstr(dst, " unset");
printed_unset = 1;
}
strbuf_addf(dst, " %s", var);
}
if (printed_unset)
strbuf_addch(dst, ';');
/* ... followed by "A=B C=D ..." */
for (i = 0; i < envs.nr; i++) {
const char *var = envs.items[i].string;
const char *val = envs.items[i].util;
const char *oldval;
if (!val)
continue;
oldval = getenv(var);
if (oldval && !strcmp(val, oldval))
continue;
strbuf_addf(dst, " %s=", var);
sq_quote_buf_pretty(dst, val);
}
string_list_clear(&envs, 0);
}
static void trace_run_command(const struct child_process *cp)
{
struct strbuf buf = STRBUF_INIT;
if (!trace_want(&trace_default_key))
return;
strbuf_addstr(&buf, "trace: run_command:");
if (cp->dir) {
strbuf_addstr(&buf, " cd ");
sq_quote_buf_pretty(&buf, cp->dir);
strbuf_addch(&buf, ';');
}
trace_add_env(&buf, cp->env.v);
if (cp->git_cmd)
strbuf_addstr(&buf, " git");
run-command API: remove "argv" member, always use "args" Remove the "argv" member from the run-command API, ever since "args" was added in c460c0ecdca (run-command: store an optional argv_array, 2014-05-15) being able to provide either "argv" or "args" has led to some confusion and bugs. If we hadn't gone in that direction and only had an "argv" our problems wouldn't have been solved either, as noted in [1] (and in the documentation amended here) it comes with inherent memory management issues: The caller would have to hang on to the "argv" until the run-command API was finished. If the "argv" was an argument to main() this wasn't an issue, but if it it was manually constructed using the API might be painful. We also have a recent report[2] of a user of the API segfaulting, which is a direct result of it being complex to use. This commit addresses the root cause of that bug. This change is larger than I'd like, but there's no easy way to avoid it that wouldn't involve even more verbose intermediate steps. We use the "argv" as the source of truth over the "args", so we need to change all parts of run-command.[ch] itself, as well as the trace2 logging at the same time. The resulting Windows-specific code in start_command() is a bit nasty, as we're now assigning to a strvec's "v" member, instead of to our own "argv". There was a suggestion of some alternate approaches in reply to an earlier version of this commit[3], but let's leave larger a larger and needless refactoring of this code for now. 1. http://lore.kernel.org/git/YT6BnnXeAWn8BycF@coredump.intra.peff.net 2. https://lore.kernel.org/git/20211120194048.12125-1-ematsumiya@suse.de/ 3. https://lore.kernel.org/git/patch-5.5-ea1011f7473-20211122T153605Z-avarab@gmail.com/ Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-11-26 06:52:22 +08:00
sq_quote_argv_pretty(&buf, cp->args.v);
trace_printf("%s", buf.buf);
strbuf_release(&buf);
}
int start_command(struct child_process *cmd)
{
int need_in, need_out, need_err;
int fdin[2], fdout[2], fderr[2];
int failed_errno;
const char *str;
/*
* In case of errors we must keep the promise to close FDs
* that have been passed in via ->in and ->out.
*/
need_in = !cmd->no_stdin && cmd->in < 0;
if (need_in) {
if (pipe(fdin) < 0) {
run_command: report system call errors instead of returning error codes The motivation for this change is that system call failures are serious errors that should be reported to the user, but only few callers took the burden to decode the error codes that the functions returned into error messages. If at all, then only an unspecific error message was given. A prominent example is this: $ git upload-pack . | : fatal: unable to run 'git-upload-pack' In this example, git-upload-pack, the external command invoked through the git wrapper, dies due to SIGPIPE, but the git wrapper does not bother to report the real cause. In fact, this very error message is copied to the syslog if git-daemon's client aborts the connection early. With this change, system call failures are reported immediately after the failure and only a generic failure code is returned to the caller. In the above example the error is now to the point: $ git upload-pack . | : error: git-upload-pack died of signal Note that there is no error report if the invoked program terminated with a non-zero exit code, because it is reasonable to expect that the invoked program has already reported an error. (But many run_command call sites nevertheless write a generic error message.) There was one special return code that was used to identify the case where run_command failed because the requested program could not be exec'd. This special case is now treated like a system call failure with errno set to ENOENT. No error is reported in this case, because the call site in git.c expects this as a normal result. Therefore, the callers that carefully decoded the return value still check for this condition. Signed-off-by: Johannes Sixt <j6t@kdbg.org> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2009-07-05 03:26:40 +08:00
failed_errno = errno;
if (cmd->out > 0)
close(cmd->out);
str = "standard input";
run_command: report system call errors instead of returning error codes The motivation for this change is that system call failures are serious errors that should be reported to the user, but only few callers took the burden to decode the error codes that the functions returned into error messages. If at all, then only an unspecific error message was given. A prominent example is this: $ git upload-pack . | : fatal: unable to run 'git-upload-pack' In this example, git-upload-pack, the external command invoked through the git wrapper, dies due to SIGPIPE, but the git wrapper does not bother to report the real cause. In fact, this very error message is copied to the syslog if git-daemon's client aborts the connection early. With this change, system call failures are reported immediately after the failure and only a generic failure code is returned to the caller. In the above example the error is now to the point: $ git upload-pack . | : error: git-upload-pack died of signal Note that there is no error report if the invoked program terminated with a non-zero exit code, because it is reasonable to expect that the invoked program has already reported an error. (But many run_command call sites nevertheless write a generic error message.) There was one special return code that was used to identify the case where run_command failed because the requested program could not be exec'd. This special case is now treated like a system call failure with errno set to ENOENT. No error is reported in this case, because the call site in git.c expects this as a normal result. Therefore, the callers that carefully decoded the return value still check for this condition. Signed-off-by: Johannes Sixt <j6t@kdbg.org> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2009-07-05 03:26:40 +08:00
goto fail_pipe;
}
cmd->in = fdin[1];
}
need_out = !cmd->no_stdout
&& !cmd->stdout_to_stderr
&& cmd->out < 0;
if (need_out) {
if (pipe(fdout) < 0) {
run_command: report system call errors instead of returning error codes The motivation for this change is that system call failures are serious errors that should be reported to the user, but only few callers took the burden to decode the error codes that the functions returned into error messages. If at all, then only an unspecific error message was given. A prominent example is this: $ git upload-pack . | : fatal: unable to run 'git-upload-pack' In this example, git-upload-pack, the external command invoked through the git wrapper, dies due to SIGPIPE, but the git wrapper does not bother to report the real cause. In fact, this very error message is copied to the syslog if git-daemon's client aborts the connection early. With this change, system call failures are reported immediately after the failure and only a generic failure code is returned to the caller. In the above example the error is now to the point: $ git upload-pack . | : error: git-upload-pack died of signal Note that there is no error report if the invoked program terminated with a non-zero exit code, because it is reasonable to expect that the invoked program has already reported an error. (But many run_command call sites nevertheless write a generic error message.) There was one special return code that was used to identify the case where run_command failed because the requested program could not be exec'd. This special case is now treated like a system call failure with errno set to ENOENT. No error is reported in this case, because the call site in git.c expects this as a normal result. Therefore, the callers that carefully decoded the return value still check for this condition. Signed-off-by: Johannes Sixt <j6t@kdbg.org> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2009-07-05 03:26:40 +08:00
failed_errno = errno;
if (need_in)
close_pair(fdin);
else if (cmd->in)
close(cmd->in);
str = "standard output";
run_command: report system call errors instead of returning error codes The motivation for this change is that system call failures are serious errors that should be reported to the user, but only few callers took the burden to decode the error codes that the functions returned into error messages. If at all, then only an unspecific error message was given. A prominent example is this: $ git upload-pack . | : fatal: unable to run 'git-upload-pack' In this example, git-upload-pack, the external command invoked through the git wrapper, dies due to SIGPIPE, but the git wrapper does not bother to report the real cause. In fact, this very error message is copied to the syslog if git-daemon's client aborts the connection early. With this change, system call failures are reported immediately after the failure and only a generic failure code is returned to the caller. In the above example the error is now to the point: $ git upload-pack . | : error: git-upload-pack died of signal Note that there is no error report if the invoked program terminated with a non-zero exit code, because it is reasonable to expect that the invoked program has already reported an error. (But many run_command call sites nevertheless write a generic error message.) There was one special return code that was used to identify the case where run_command failed because the requested program could not be exec'd. This special case is now treated like a system call failure with errno set to ENOENT. No error is reported in this case, because the call site in git.c expects this as a normal result. Therefore, the callers that carefully decoded the return value still check for this condition. Signed-off-by: Johannes Sixt <j6t@kdbg.org> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2009-07-05 03:26:40 +08:00
goto fail_pipe;
}
cmd->out = fdout[0];
}
need_err = !cmd->no_stderr && cmd->err < 0;
if (need_err) {
if (pipe(fderr) < 0) {
run_command: report system call errors instead of returning error codes The motivation for this change is that system call failures are serious errors that should be reported to the user, but only few callers took the burden to decode the error codes that the functions returned into error messages. If at all, then only an unspecific error message was given. A prominent example is this: $ git upload-pack . | : fatal: unable to run 'git-upload-pack' In this example, git-upload-pack, the external command invoked through the git wrapper, dies due to SIGPIPE, but the git wrapper does not bother to report the real cause. In fact, this very error message is copied to the syslog if git-daemon's client aborts the connection early. With this change, system call failures are reported immediately after the failure and only a generic failure code is returned to the caller. In the above example the error is now to the point: $ git upload-pack . | : error: git-upload-pack died of signal Note that there is no error report if the invoked program terminated with a non-zero exit code, because it is reasonable to expect that the invoked program has already reported an error. (But many run_command call sites nevertheless write a generic error message.) There was one special return code that was used to identify the case where run_command failed because the requested program could not be exec'd. This special case is now treated like a system call failure with errno set to ENOENT. No error is reported in this case, because the call site in git.c expects this as a normal result. Therefore, the callers that carefully decoded the return value still check for this condition. Signed-off-by: Johannes Sixt <j6t@kdbg.org> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2009-07-05 03:26:40 +08:00
failed_errno = errno;
if (need_in)
close_pair(fdin);
else if (cmd->in)
close(cmd->in);
if (need_out)
close_pair(fdout);
else if (cmd->out)
close(cmd->out);
str = "standard error";
run_command: report system call errors instead of returning error codes The motivation for this change is that system call failures are serious errors that should be reported to the user, but only few callers took the burden to decode the error codes that the functions returned into error messages. If at all, then only an unspecific error message was given. A prominent example is this: $ git upload-pack . | : fatal: unable to run 'git-upload-pack' In this example, git-upload-pack, the external command invoked through the git wrapper, dies due to SIGPIPE, but the git wrapper does not bother to report the real cause. In fact, this very error message is copied to the syslog if git-daemon's client aborts the connection early. With this change, system call failures are reported immediately after the failure and only a generic failure code is returned to the caller. In the above example the error is now to the point: $ git upload-pack . | : error: git-upload-pack died of signal Note that there is no error report if the invoked program terminated with a non-zero exit code, because it is reasonable to expect that the invoked program has already reported an error. (But many run_command call sites nevertheless write a generic error message.) There was one special return code that was used to identify the case where run_command failed because the requested program could not be exec'd. This special case is now treated like a system call failure with errno set to ENOENT. No error is reported in this case, because the call site in git.c expects this as a normal result. Therefore, the callers that carefully decoded the return value still check for this condition. Signed-off-by: Johannes Sixt <j6t@kdbg.org> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2009-07-05 03:26:40 +08:00
fail_pipe:
error("cannot create %s pipe for %s: %s",
run-command API: remove "argv" member, always use "args" Remove the "argv" member from the run-command API, ever since "args" was added in c460c0ecdca (run-command: store an optional argv_array, 2014-05-15) being able to provide either "argv" or "args" has led to some confusion and bugs. If we hadn't gone in that direction and only had an "argv" our problems wouldn't have been solved either, as noted in [1] (and in the documentation amended here) it comes with inherent memory management issues: The caller would have to hang on to the "argv" until the run-command API was finished. If the "argv" was an argument to main() this wasn't an issue, but if it it was manually constructed using the API might be painful. We also have a recent report[2] of a user of the API segfaulting, which is a direct result of it being complex to use. This commit addresses the root cause of that bug. This change is larger than I'd like, but there's no easy way to avoid it that wouldn't involve even more verbose intermediate steps. We use the "argv" as the source of truth over the "args", so we need to change all parts of run-command.[ch] itself, as well as the trace2 logging at the same time. The resulting Windows-specific code in start_command() is a bit nasty, as we're now assigning to a strvec's "v" member, instead of to our own "argv". There was a suggestion of some alternate approaches in reply to an earlier version of this commit[3], but let's leave larger a larger and needless refactoring of this code for now. 1. http://lore.kernel.org/git/YT6BnnXeAWn8BycF@coredump.intra.peff.net 2. https://lore.kernel.org/git/20211120194048.12125-1-ematsumiya@suse.de/ 3. https://lore.kernel.org/git/patch-5.5-ea1011f7473-20211122T153605Z-avarab@gmail.com/ Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-11-26 06:52:22 +08:00
str, cmd->args.v[0], strerror(failed_errno));
child_process_clear(cmd);
run_command: report system call errors instead of returning error codes The motivation for this change is that system call failures are serious errors that should be reported to the user, but only few callers took the burden to decode the error codes that the functions returned into error messages. If at all, then only an unspecific error message was given. A prominent example is this: $ git upload-pack . | : fatal: unable to run 'git-upload-pack' In this example, git-upload-pack, the external command invoked through the git wrapper, dies due to SIGPIPE, but the git wrapper does not bother to report the real cause. In fact, this very error message is copied to the syslog if git-daemon's client aborts the connection early. With this change, system call failures are reported immediately after the failure and only a generic failure code is returned to the caller. In the above example the error is now to the point: $ git upload-pack . | : error: git-upload-pack died of signal Note that there is no error report if the invoked program terminated with a non-zero exit code, because it is reasonable to expect that the invoked program has already reported an error. (But many run_command call sites nevertheless write a generic error message.) There was one special return code that was used to identify the case where run_command failed because the requested program could not be exec'd. This special case is now treated like a system call failure with errno set to ENOENT. No error is reported in this case, because the call site in git.c expects this as a normal result. Therefore, the callers that carefully decoded the return value still check for this condition. Signed-off-by: Johannes Sixt <j6t@kdbg.org> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2009-07-05 03:26:40 +08:00
errno = failed_errno;
return -1;
}
cmd->err = fderr[0];
}
trace2_child_start(cmd);
trace_run_command(cmd);
fflush(NULL);
if (cmd->close_object_store)
close_object_store(the_repository->objects);
#ifndef GIT_WINDOWS_NATIVE
{
int notify_pipe[2];
int null_fd = -1;
char **childenv;
struct strvec argv = STRVEC_INIT;
struct child_err cerr;
struct atfork_state as;
if (prepare_cmd(&argv, cmd) < 0) {
failed_errno = errno;
cmd->pid = -1;
if (!cmd->silent_exec_failure)
run-command API: remove "argv" member, always use "args" Remove the "argv" member from the run-command API, ever since "args" was added in c460c0ecdca (run-command: store an optional argv_array, 2014-05-15) being able to provide either "argv" or "args" has led to some confusion and bugs. If we hadn't gone in that direction and only had an "argv" our problems wouldn't have been solved either, as noted in [1] (and in the documentation amended here) it comes with inherent memory management issues: The caller would have to hang on to the "argv" until the run-command API was finished. If the "argv" was an argument to main() this wasn't an issue, but if it it was manually constructed using the API might be painful. We also have a recent report[2] of a user of the API segfaulting, which is a direct result of it being complex to use. This commit addresses the root cause of that bug. This change is larger than I'd like, but there's no easy way to avoid it that wouldn't involve even more verbose intermediate steps. We use the "argv" as the source of truth over the "args", so we need to change all parts of run-command.[ch] itself, as well as the trace2 logging at the same time. The resulting Windows-specific code in start_command() is a bit nasty, as we're now assigning to a strvec's "v" member, instead of to our own "argv". There was a suggestion of some alternate approaches in reply to an earlier version of this commit[3], but let's leave larger a larger and needless refactoring of this code for now. 1. http://lore.kernel.org/git/YT6BnnXeAWn8BycF@coredump.intra.peff.net 2. https://lore.kernel.org/git/20211120194048.12125-1-ematsumiya@suse.de/ 3. https://lore.kernel.org/git/patch-5.5-ea1011f7473-20211122T153605Z-avarab@gmail.com/ Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-11-26 06:52:22 +08:00
error_errno("cannot run %s", cmd->args.v[0]);
goto end_of_spawn;
}
trace_argv_printf(&argv.v[1], "trace: start_command:");
if (pipe(notify_pipe))
notify_pipe[0] = notify_pipe[1] = -1;
if (cmd->no_stdin || cmd->no_stdout || cmd->no_stderr) {
null_fd = xopen("/dev/null", O_RDWR | O_CLOEXEC);
set_cloexec(null_fd);
}
childenv = prep_childenv(cmd->env.v);
atfork_prepare(&as);
/*
* NOTE: In order to prevent deadlocking when using threads special
* care should be taken with the function calls made in between the
* fork() and exec() calls. No calls should be made to functions which
* require acquiring a lock (e.g. malloc) as the lock could have been
* held by another thread at the time of forking, causing the lock to
* never be released in the child process. This means only
* Async-Signal-Safe functions are permitted in the child.
*/
cmd->pid = fork();
failed_errno = errno;
if (!cmd->pid) {
int sig;
/*
* Ensure the default die/error/warn routines do not get
* called, they can take stdio locks and malloc.
*/
set_die_routine(child_die_fn);
set_error_routine(child_error_fn);
set_warn_routine(child_warn_fn);
close(notify_pipe[0]);
set_cloexec(notify_pipe[1]);
child_notifier = notify_pipe[1];
if (cmd->no_stdin)
child_dup2(null_fd, 0);
else if (need_in) {
child_dup2(fdin[0], 0);
child_close_pair(fdin);
} else if (cmd->in) {
child_dup2(cmd->in, 0);
child_close(cmd->in);
}
if (cmd->no_stderr)
child_dup2(null_fd, 2);
else if (need_err) {
child_dup2(fderr[1], 2);
child_close_pair(fderr);
} else if (cmd->err > 1) {
child_dup2(cmd->err, 2);
child_close(cmd->err);
}
if (cmd->no_stdout)
child_dup2(null_fd, 1);
else if (cmd->stdout_to_stderr)
child_dup2(2, 1);
else if (need_out) {
child_dup2(fdout[1], 1);
child_close_pair(fdout);
} else if (cmd->out > 1) {
child_dup2(cmd->out, 1);
child_close(cmd->out);
}
if (cmd->dir && chdir(cmd->dir))
child_die(CHILD_ERR_CHDIR);
/*
* restore default signal handlers here, in case
* we catch a signal right before execve below
*/
for (sig = 1; sig < NSIG; sig++) {
/* ignored signals get reset to SIG_DFL on execve */
if (signal(sig, SIG_DFL) == SIG_IGN)
signal(sig, SIG_IGN);
}
if (sigprocmask(SIG_SETMASK, &as.old, NULL) != 0)
child_die(CHILD_ERR_SIGPROCMASK);
/*
* Attempt to exec using the command and arguments starting at
* argv.argv[1]. argv.argv[0] contains SHELL_PATH which will
* be used in the event exec failed with ENOEXEC at which point
* we will try to interpret the command using 'sh'.
*/
execve(argv.v[1], (char *const *) argv.v + 1,
(char *const *) childenv);
if (errno == ENOEXEC)
execve(argv.v[0], (char *const *) argv.v,
(char *const *) childenv);
if (cmd->silent_exec_failure && errno == ENOENT)
child_die(CHILD_ERR_SILENT);
child_die(CHILD_ERR_ERRNO);
}
atfork_parent(&as);
run_command: report system call errors instead of returning error codes The motivation for this change is that system call failures are serious errors that should be reported to the user, but only few callers took the burden to decode the error codes that the functions returned into error messages. If at all, then only an unspecific error message was given. A prominent example is this: $ git upload-pack . | : fatal: unable to run 'git-upload-pack' In this example, git-upload-pack, the external command invoked through the git wrapper, dies due to SIGPIPE, but the git wrapper does not bother to report the real cause. In fact, this very error message is copied to the syslog if git-daemon's client aborts the connection early. With this change, system call failures are reported immediately after the failure and only a generic failure code is returned to the caller. In the above example the error is now to the point: $ git upload-pack . | : error: git-upload-pack died of signal Note that there is no error report if the invoked program terminated with a non-zero exit code, because it is reasonable to expect that the invoked program has already reported an error. (But many run_command call sites nevertheless write a generic error message.) There was one special return code that was used to identify the case where run_command failed because the requested program could not be exec'd. This special case is now treated like a system call failure with errno set to ENOENT. No error is reported in this case, because the call site in git.c expects this as a normal result. Therefore, the callers that carefully decoded the return value still check for this condition. Signed-off-by: Johannes Sixt <j6t@kdbg.org> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2009-07-05 03:26:40 +08:00
if (cmd->pid < 0)
run-command API: remove "argv" member, always use "args" Remove the "argv" member from the run-command API, ever since "args" was added in c460c0ecdca (run-command: store an optional argv_array, 2014-05-15) being able to provide either "argv" or "args" has led to some confusion and bugs. If we hadn't gone in that direction and only had an "argv" our problems wouldn't have been solved either, as noted in [1] (and in the documentation amended here) it comes with inherent memory management issues: The caller would have to hang on to the "argv" until the run-command API was finished. If the "argv" was an argument to main() this wasn't an issue, but if it it was manually constructed using the API might be painful. We also have a recent report[2] of a user of the API segfaulting, which is a direct result of it being complex to use. This commit addresses the root cause of that bug. This change is larger than I'd like, but there's no easy way to avoid it that wouldn't involve even more verbose intermediate steps. We use the "argv" as the source of truth over the "args", so we need to change all parts of run-command.[ch] itself, as well as the trace2 logging at the same time. The resulting Windows-specific code in start_command() is a bit nasty, as we're now assigning to a strvec's "v" member, instead of to our own "argv". There was a suggestion of some alternate approaches in reply to an earlier version of this commit[3], but let's leave larger a larger and needless refactoring of this code for now. 1. http://lore.kernel.org/git/YT6BnnXeAWn8BycF@coredump.intra.peff.net 2. https://lore.kernel.org/git/20211120194048.12125-1-ematsumiya@suse.de/ 3. https://lore.kernel.org/git/patch-5.5-ea1011f7473-20211122T153605Z-avarab@gmail.com/ Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-11-26 06:52:22 +08:00
error_errno("cannot fork() for %s", cmd->args.v[0]);
else if (cmd->clean_on_exit)
mark_child_for_cleanup(cmd->pid, cmd);
/*
* Wait for child's exec. If the exec succeeds (or if fork()
* failed), EOF is seen immediately by the parent. Otherwise, the
* child process sends a child_err struct.
* Note that use of this infrastructure is completely advisory,
* therefore, we keep error checks minimal.
*/
close(notify_pipe[1]);
if (xread(notify_pipe[0], &cerr, sizeof(cerr)) == sizeof(cerr)) {
/*
* At this point we know that fork() succeeded, but exec()
* failed. Errors have been reported to our stderr.
*/
run-command API: remove "argv" member, always use "args" Remove the "argv" member from the run-command API, ever since "args" was added in c460c0ecdca (run-command: store an optional argv_array, 2014-05-15) being able to provide either "argv" or "args" has led to some confusion and bugs. If we hadn't gone in that direction and only had an "argv" our problems wouldn't have been solved either, as noted in [1] (and in the documentation amended here) it comes with inherent memory management issues: The caller would have to hang on to the "argv" until the run-command API was finished. If the "argv" was an argument to main() this wasn't an issue, but if it it was manually constructed using the API might be painful. We also have a recent report[2] of a user of the API segfaulting, which is a direct result of it being complex to use. This commit addresses the root cause of that bug. This change is larger than I'd like, but there's no easy way to avoid it that wouldn't involve even more verbose intermediate steps. We use the "argv" as the source of truth over the "args", so we need to change all parts of run-command.[ch] itself, as well as the trace2 logging at the same time. The resulting Windows-specific code in start_command() is a bit nasty, as we're now assigning to a strvec's "v" member, instead of to our own "argv". There was a suggestion of some alternate approaches in reply to an earlier version of this commit[3], but let's leave larger a larger and needless refactoring of this code for now. 1. http://lore.kernel.org/git/YT6BnnXeAWn8BycF@coredump.intra.peff.net 2. https://lore.kernel.org/git/20211120194048.12125-1-ematsumiya@suse.de/ 3. https://lore.kernel.org/git/patch-5.5-ea1011f7473-20211122T153605Z-avarab@gmail.com/ Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-11-26 06:52:22 +08:00
wait_or_whine(cmd->pid, cmd->args.v[0], 0);
child_err_spew(cmd, &cerr);
failed_errno = errno;
cmd->pid = -1;
}
close(notify_pipe[0]);
if (null_fd >= 0)
close(null_fd);
strvec_clear(&argv);
free(childenv);
}
end_of_spawn:
#else
{
Windows: avoid the "dup dance" when spawning a child process When stdin, stdout, or stderr must be redirected for a child process that on Windows is spawned using one of the spawn() functions of Microsoft's C runtime, then there is no choice other than to 1. make a backup copy of fd 0,1,2 with dup 2. dup2 the redirection source fd into 0,1,2 3. spawn 4. dup2 the backup back into 0,1,2 5. close the backup copy and the redirection source We used this idiom as well -- but we are not using the spawn() functions anymore! Instead, we have our own implementation. We had hardcoded that stdin, stdout, and stderr of the child process were inherited from the parent's fds 0, 1, and 2. But we can actually specify any fd. With this patch, the fds to inherit are passed from start_command()'s WIN32 section to our spawn implementation. This way, we can avoid the backup copies of the fds. The backup copies were a bug waiting to surface: The OS handles underlying the dup()ed fds were inherited by the child process (but were not associated with a file descriptor in the child). Consequently, the file or pipe represented by the OS handle remained open even after the backup copy was closed in the parent process until the child exited. Since our implementation of pipe() creates non-inheritable OS handles, we still dup() file descriptors in start_command() because dup() happens to create inheritable duplicates. (A nice side effect is that the fd cleanup in start_command is the same for Windows and Unix and remains unchanged.) Signed-off-by: Johannes Sixt <j6t@kdbg.org> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2010-01-16 04:12:18 +08:00
int fhin = 0, fhout = 1, fherr = 2;
run-command API: remove "argv" member, always use "args" Remove the "argv" member from the run-command API, ever since "args" was added in c460c0ecdca (run-command: store an optional argv_array, 2014-05-15) being able to provide either "argv" or "args" has led to some confusion and bugs. If we hadn't gone in that direction and only had an "argv" our problems wouldn't have been solved either, as noted in [1] (and in the documentation amended here) it comes with inherent memory management issues: The caller would have to hang on to the "argv" until the run-command API was finished. If the "argv" was an argument to main() this wasn't an issue, but if it it was manually constructed using the API might be painful. We also have a recent report[2] of a user of the API segfaulting, which is a direct result of it being complex to use. This commit addresses the root cause of that bug. This change is larger than I'd like, but there's no easy way to avoid it that wouldn't involve even more verbose intermediate steps. We use the "argv" as the source of truth over the "args", so we need to change all parts of run-command.[ch] itself, as well as the trace2 logging at the same time. The resulting Windows-specific code in start_command() is a bit nasty, as we're now assigning to a strvec's "v" member, instead of to our own "argv". There was a suggestion of some alternate approaches in reply to an earlier version of this commit[3], but let's leave larger a larger and needless refactoring of this code for now. 1. http://lore.kernel.org/git/YT6BnnXeAWn8BycF@coredump.intra.peff.net 2. https://lore.kernel.org/git/20211120194048.12125-1-ematsumiya@suse.de/ 3. https://lore.kernel.org/git/patch-5.5-ea1011f7473-20211122T153605Z-avarab@gmail.com/ Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-11-26 06:52:22 +08:00
const char **sargv = cmd->args.v;
struct strvec nargv = STRVEC_INIT;
Windows: avoid the "dup dance" when spawning a child process When stdin, stdout, or stderr must be redirected for a child process that on Windows is spawned using one of the spawn() functions of Microsoft's C runtime, then there is no choice other than to 1. make a backup copy of fd 0,1,2 with dup 2. dup2 the redirection source fd into 0,1,2 3. spawn 4. dup2 the backup back into 0,1,2 5. close the backup copy and the redirection source We used this idiom as well -- but we are not using the spawn() functions anymore! Instead, we have our own implementation. We had hardcoded that stdin, stdout, and stderr of the child process were inherited from the parent's fds 0, 1, and 2. But we can actually specify any fd. With this patch, the fds to inherit are passed from start_command()'s WIN32 section to our spawn implementation. This way, we can avoid the backup copies of the fds. The backup copies were a bug waiting to surface: The OS handles underlying the dup()ed fds were inherited by the child process (but were not associated with a file descriptor in the child). Consequently, the file or pipe represented by the OS handle remained open even after the backup copy was closed in the parent process until the child exited. Since our implementation of pipe() creates non-inheritable OS handles, we still dup() file descriptors in start_command() because dup() happens to create inheritable duplicates. (A nice side effect is that the fd cleanup in start_command is the same for Windows and Unix and remains unchanged.) Signed-off-by: Johannes Sixt <j6t@kdbg.org> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2010-01-16 04:12:18 +08:00
if (cmd->no_stdin)
fhin = open("/dev/null", O_RDWR);
else if (need_in)
fhin = dup(fdin[0]);
else if (cmd->in)
fhin = dup(cmd->in);
if (cmd->no_stderr)
fherr = open("/dev/null", O_RDWR);
else if (need_err)
fherr = dup(fderr[1]);
else if (cmd->err > 2)
fherr = dup(cmd->err);
Windows: avoid the "dup dance" when spawning a child process When stdin, stdout, or stderr must be redirected for a child process that on Windows is spawned using one of the spawn() functions of Microsoft's C runtime, then there is no choice other than to 1. make a backup copy of fd 0,1,2 with dup 2. dup2 the redirection source fd into 0,1,2 3. spawn 4. dup2 the backup back into 0,1,2 5. close the backup copy and the redirection source We used this idiom as well -- but we are not using the spawn() functions anymore! Instead, we have our own implementation. We had hardcoded that stdin, stdout, and stderr of the child process were inherited from the parent's fds 0, 1, and 2. But we can actually specify any fd. With this patch, the fds to inherit are passed from start_command()'s WIN32 section to our spawn implementation. This way, we can avoid the backup copies of the fds. The backup copies were a bug waiting to surface: The OS handles underlying the dup()ed fds were inherited by the child process (but were not associated with a file descriptor in the child). Consequently, the file or pipe represented by the OS handle remained open even after the backup copy was closed in the parent process until the child exited. Since our implementation of pipe() creates non-inheritable OS handles, we still dup() file descriptors in start_command() because dup() happens to create inheritable duplicates. (A nice side effect is that the fd cleanup in start_command is the same for Windows and Unix and remains unchanged.) Signed-off-by: Johannes Sixt <j6t@kdbg.org> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2010-01-16 04:12:18 +08:00
if (cmd->no_stdout)
fhout = open("/dev/null", O_RDWR);
else if (cmd->stdout_to_stderr)
fhout = dup(fherr);
else if (need_out)
fhout = dup(fdout[1]);
else if (cmd->out > 1)
fhout = dup(cmd->out);
if (cmd->git_cmd)
run-command API: remove "argv" member, always use "args" Remove the "argv" member from the run-command API, ever since "args" was added in c460c0ecdca (run-command: store an optional argv_array, 2014-05-15) being able to provide either "argv" or "args" has led to some confusion and bugs. If we hadn't gone in that direction and only had an "argv" our problems wouldn't have been solved either, as noted in [1] (and in the documentation amended here) it comes with inherent memory management issues: The caller would have to hang on to the "argv" until the run-command API was finished. If the "argv" was an argument to main() this wasn't an issue, but if it it was manually constructed using the API might be painful. We also have a recent report[2] of a user of the API segfaulting, which is a direct result of it being complex to use. This commit addresses the root cause of that bug. This change is larger than I'd like, but there's no easy way to avoid it that wouldn't involve even more verbose intermediate steps. We use the "argv" as the source of truth over the "args", so we need to change all parts of run-command.[ch] itself, as well as the trace2 logging at the same time. The resulting Windows-specific code in start_command() is a bit nasty, as we're now assigning to a strvec's "v" member, instead of to our own "argv". There was a suggestion of some alternate approaches in reply to an earlier version of this commit[3], but let's leave larger a larger and needless refactoring of this code for now. 1. http://lore.kernel.org/git/YT6BnnXeAWn8BycF@coredump.intra.peff.net 2. https://lore.kernel.org/git/20211120194048.12125-1-ematsumiya@suse.de/ 3. https://lore.kernel.org/git/patch-5.5-ea1011f7473-20211122T153605Z-avarab@gmail.com/ Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-11-26 06:52:22 +08:00
cmd->args.v = prepare_git_cmd(&nargv, sargv);
else if (cmd->use_shell)
run-command API: remove "argv" member, always use "args" Remove the "argv" member from the run-command API, ever since "args" was added in c460c0ecdca (run-command: store an optional argv_array, 2014-05-15) being able to provide either "argv" or "args" has led to some confusion and bugs. If we hadn't gone in that direction and only had an "argv" our problems wouldn't have been solved either, as noted in [1] (and in the documentation amended here) it comes with inherent memory management issues: The caller would have to hang on to the "argv" until the run-command API was finished. If the "argv" was an argument to main() this wasn't an issue, but if it it was manually constructed using the API might be painful. We also have a recent report[2] of a user of the API segfaulting, which is a direct result of it being complex to use. This commit addresses the root cause of that bug. This change is larger than I'd like, but there's no easy way to avoid it that wouldn't involve even more verbose intermediate steps. We use the "argv" as the source of truth over the "args", so we need to change all parts of run-command.[ch] itself, as well as the trace2 logging at the same time. The resulting Windows-specific code in start_command() is a bit nasty, as we're now assigning to a strvec's "v" member, instead of to our own "argv". There was a suggestion of some alternate approaches in reply to an earlier version of this commit[3], but let's leave larger a larger and needless refactoring of this code for now. 1. http://lore.kernel.org/git/YT6BnnXeAWn8BycF@coredump.intra.peff.net 2. https://lore.kernel.org/git/20211120194048.12125-1-ematsumiya@suse.de/ 3. https://lore.kernel.org/git/patch-5.5-ea1011f7473-20211122T153605Z-avarab@gmail.com/ Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-11-26 06:52:22 +08:00
cmd->args.v = prepare_shell_cmd(&nargv, sargv);
trace_argv_printf(cmd->args.v, "trace: start_command:");
cmd->pid = mingw_spawnvpe(cmd->args.v[0], cmd->args.v,
(char**) cmd->env.v,
cmd->dir, fhin, fhout, fherr);
run_command: report system call errors instead of returning error codes The motivation for this change is that system call failures are serious errors that should be reported to the user, but only few callers took the burden to decode the error codes that the functions returned into error messages. If at all, then only an unspecific error message was given. A prominent example is this: $ git upload-pack . | : fatal: unable to run 'git-upload-pack' In this example, git-upload-pack, the external command invoked through the git wrapper, dies due to SIGPIPE, but the git wrapper does not bother to report the real cause. In fact, this very error message is copied to the syslog if git-daemon's client aborts the connection early. With this change, system call failures are reported immediately after the failure and only a generic failure code is returned to the caller. In the above example the error is now to the point: $ git upload-pack . | : error: git-upload-pack died of signal Note that there is no error report if the invoked program terminated with a non-zero exit code, because it is reasonable to expect that the invoked program has already reported an error. (But many run_command call sites nevertheless write a generic error message.) There was one special return code that was used to identify the case where run_command failed because the requested program could not be exec'd. This special case is now treated like a system call failure with errno set to ENOENT. No error is reported in this case, because the call site in git.c expects this as a normal result. Therefore, the callers that carefully decoded the return value still check for this condition. Signed-off-by: Johannes Sixt <j6t@kdbg.org> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2009-07-05 03:26:40 +08:00
failed_errno = errno;
if (cmd->pid < 0 && (!cmd->silent_exec_failure || errno != ENOENT))
run-command API: remove "argv" member, always use "args" Remove the "argv" member from the run-command API, ever since "args" was added in c460c0ecdca (run-command: store an optional argv_array, 2014-05-15) being able to provide either "argv" or "args" has led to some confusion and bugs. If we hadn't gone in that direction and only had an "argv" our problems wouldn't have been solved either, as noted in [1] (and in the documentation amended here) it comes with inherent memory management issues: The caller would have to hang on to the "argv" until the run-command API was finished. If the "argv" was an argument to main() this wasn't an issue, but if it it was manually constructed using the API might be painful. We also have a recent report[2] of a user of the API segfaulting, which is a direct result of it being complex to use. This commit addresses the root cause of that bug. This change is larger than I'd like, but there's no easy way to avoid it that wouldn't involve even more verbose intermediate steps. We use the "argv" as the source of truth over the "args", so we need to change all parts of run-command.[ch] itself, as well as the trace2 logging at the same time. The resulting Windows-specific code in start_command() is a bit nasty, as we're now assigning to a strvec's "v" member, instead of to our own "argv". There was a suggestion of some alternate approaches in reply to an earlier version of this commit[3], but let's leave larger a larger and needless refactoring of this code for now. 1. http://lore.kernel.org/git/YT6BnnXeAWn8BycF@coredump.intra.peff.net 2. https://lore.kernel.org/git/20211120194048.12125-1-ematsumiya@suse.de/ 3. https://lore.kernel.org/git/patch-5.5-ea1011f7473-20211122T153605Z-avarab@gmail.com/ Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-11-26 06:52:22 +08:00
error_errno("cannot spawn %s", cmd->args.v[0]);
if (cmd->clean_on_exit && cmd->pid >= 0)
mark_child_for_cleanup(cmd->pid, cmd);
strvec_clear(&nargv);
run-command API: remove "argv" member, always use "args" Remove the "argv" member from the run-command API, ever since "args" was added in c460c0ecdca (run-command: store an optional argv_array, 2014-05-15) being able to provide either "argv" or "args" has led to some confusion and bugs. If we hadn't gone in that direction and only had an "argv" our problems wouldn't have been solved either, as noted in [1] (and in the documentation amended here) it comes with inherent memory management issues: The caller would have to hang on to the "argv" until the run-command API was finished. If the "argv" was an argument to main() this wasn't an issue, but if it it was manually constructed using the API might be painful. We also have a recent report[2] of a user of the API segfaulting, which is a direct result of it being complex to use. This commit addresses the root cause of that bug. This change is larger than I'd like, but there's no easy way to avoid it that wouldn't involve even more verbose intermediate steps. We use the "argv" as the source of truth over the "args", so we need to change all parts of run-command.[ch] itself, as well as the trace2 logging at the same time. The resulting Windows-specific code in start_command() is a bit nasty, as we're now assigning to a strvec's "v" member, instead of to our own "argv". There was a suggestion of some alternate approaches in reply to an earlier version of this commit[3], but let's leave larger a larger and needless refactoring of this code for now. 1. http://lore.kernel.org/git/YT6BnnXeAWn8BycF@coredump.intra.peff.net 2. https://lore.kernel.org/git/20211120194048.12125-1-ematsumiya@suse.de/ 3. https://lore.kernel.org/git/patch-5.5-ea1011f7473-20211122T153605Z-avarab@gmail.com/ Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-11-26 06:52:22 +08:00
cmd->args.v = sargv;
Windows: avoid the "dup dance" when spawning a child process When stdin, stdout, or stderr must be redirected for a child process that on Windows is spawned using one of the spawn() functions of Microsoft's C runtime, then there is no choice other than to 1. make a backup copy of fd 0,1,2 with dup 2. dup2 the redirection source fd into 0,1,2 3. spawn 4. dup2 the backup back into 0,1,2 5. close the backup copy and the redirection source We used this idiom as well -- but we are not using the spawn() functions anymore! Instead, we have our own implementation. We had hardcoded that stdin, stdout, and stderr of the child process were inherited from the parent's fds 0, 1, and 2. But we can actually specify any fd. With this patch, the fds to inherit are passed from start_command()'s WIN32 section to our spawn implementation. This way, we can avoid the backup copies of the fds. The backup copies were a bug waiting to surface: The OS handles underlying the dup()ed fds were inherited by the child process (but were not associated with a file descriptor in the child). Consequently, the file or pipe represented by the OS handle remained open even after the backup copy was closed in the parent process until the child exited. Since our implementation of pipe() creates non-inheritable OS handles, we still dup() file descriptors in start_command() because dup() happens to create inheritable duplicates. (A nice side effect is that the fd cleanup in start_command is the same for Windows and Unix and remains unchanged.) Signed-off-by: Johannes Sixt <j6t@kdbg.org> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2010-01-16 04:12:18 +08:00
if (fhin != 0)
close(fhin);
if (fhout != 1)
close(fhout);
if (fherr != 2)
close(fherr);
}
#endif
if (cmd->pid < 0) {
trace2_child_exit(cmd, -1);
if (need_in)
close_pair(fdin);
else if (cmd->in)
close(cmd->in);
if (need_out)
close_pair(fdout);
else if (cmd->out)
close(cmd->out);
if (need_err)
close_pair(fderr);
else if (cmd->err)
close(cmd->err);
child_process_clear(cmd);
run_command: report system call errors instead of returning error codes The motivation for this change is that system call failures are serious errors that should be reported to the user, but only few callers took the burden to decode the error codes that the functions returned into error messages. If at all, then only an unspecific error message was given. A prominent example is this: $ git upload-pack . | : fatal: unable to run 'git-upload-pack' In this example, git-upload-pack, the external command invoked through the git wrapper, dies due to SIGPIPE, but the git wrapper does not bother to report the real cause. In fact, this very error message is copied to the syslog if git-daemon's client aborts the connection early. With this change, system call failures are reported immediately after the failure and only a generic failure code is returned to the caller. In the above example the error is now to the point: $ git upload-pack . | : error: git-upload-pack died of signal Note that there is no error report if the invoked program terminated with a non-zero exit code, because it is reasonable to expect that the invoked program has already reported an error. (But many run_command call sites nevertheless write a generic error message.) There was one special return code that was used to identify the case where run_command failed because the requested program could not be exec'd. This special case is now treated like a system call failure with errno set to ENOENT. No error is reported in this case, because the call site in git.c expects this as a normal result. Therefore, the callers that carefully decoded the return value still check for this condition. Signed-off-by: Johannes Sixt <j6t@kdbg.org> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2009-07-05 03:26:40 +08:00
errno = failed_errno;
return -1;
}
if (need_in)
close(fdin[0]);
else if (cmd->in)
close(cmd->in);
if (need_out)
close(fdout[1]);
else if (cmd->out)
close(cmd->out);
if (need_err)
close(fderr[1]);
else if (cmd->err)
close(cmd->err);
return 0;
}
int finish_command(struct child_process *cmd)
{
run-command API: remove "argv" member, always use "args" Remove the "argv" member from the run-command API, ever since "args" was added in c460c0ecdca (run-command: store an optional argv_array, 2014-05-15) being able to provide either "argv" or "args" has led to some confusion and bugs. If we hadn't gone in that direction and only had an "argv" our problems wouldn't have been solved either, as noted in [1] (and in the documentation amended here) it comes with inherent memory management issues: The caller would have to hang on to the "argv" until the run-command API was finished. If the "argv" was an argument to main() this wasn't an issue, but if it it was manually constructed using the API might be painful. We also have a recent report[2] of a user of the API segfaulting, which is a direct result of it being complex to use. This commit addresses the root cause of that bug. This change is larger than I'd like, but there's no easy way to avoid it that wouldn't involve even more verbose intermediate steps. We use the "argv" as the source of truth over the "args", so we need to change all parts of run-command.[ch] itself, as well as the trace2 logging at the same time. The resulting Windows-specific code in start_command() is a bit nasty, as we're now assigning to a strvec's "v" member, instead of to our own "argv". There was a suggestion of some alternate approaches in reply to an earlier version of this commit[3], but let's leave larger a larger and needless refactoring of this code for now. 1. http://lore.kernel.org/git/YT6BnnXeAWn8BycF@coredump.intra.peff.net 2. https://lore.kernel.org/git/20211120194048.12125-1-ematsumiya@suse.de/ 3. https://lore.kernel.org/git/patch-5.5-ea1011f7473-20211122T153605Z-avarab@gmail.com/ Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-11-26 06:52:22 +08:00
int ret = wait_or_whine(cmd->pid, cmd->args.v[0], 0);
trace2_child_exit(cmd, ret);
child_process_clear(cmd);
invalidate_lstat_cache();
return ret;
}
pager: don't use unsafe functions in signal handlers Since the commit a3da8821208d (pager: do wait_for_pager on signal death), we call wait_for_pager() in the pager's signal handler. The recent bug report revealed that this causes a deadlock in glibc at aborting "git log" [*1*]. When this happens, git process is left unterminated, and it can't be killed by SIGTERM but only by SIGKILL. The problem is that wait_for_pager() function does more than waiting for pager process's termination, but it does cleanups and printing errors. Unfortunately, the functions that may be used in a signal handler are very limited [*2*]. Particularly, malloc(), free() and the variants can't be used in a signal handler because they take a mutex internally in glibc. This was the cause of the deadlock above. Other than the direct calls of malloc/free, many functions calling malloc/free can't be used. strerror() is such one, either. Also the usage of fflush() and printf() in a signal handler is bad, although it seems working so far. In a safer side, we should avoid them, too. This patch tries to reduce the calls of such functions in signal handlers. wait_for_signal() takes a flag and avoids the unsafe calls. Also, finish_command_in_signal() is introduced for the same reason. There the free() calls are removed, and only waits for the children without whining at errors. [*1*] https://bugzilla.opensuse.org/show_bug.cgi?id=942297 [*2*] http://pubs.opengroup.org/onlinepubs/9699919799/functions/V2_chap02.html#tag_15_04_03 Signed-off-by: Takashi Iwai <tiwai@suse.de> Reviewed-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-09-04 17:35:57 +08:00
int finish_command_in_signal(struct child_process *cmd)
{
run-command API: remove "argv" member, always use "args" Remove the "argv" member from the run-command API, ever since "args" was added in c460c0ecdca (run-command: store an optional argv_array, 2014-05-15) being able to provide either "argv" or "args" has led to some confusion and bugs. If we hadn't gone in that direction and only had an "argv" our problems wouldn't have been solved either, as noted in [1] (and in the documentation amended here) it comes with inherent memory management issues: The caller would have to hang on to the "argv" until the run-command API was finished. If the "argv" was an argument to main() this wasn't an issue, but if it it was manually constructed using the API might be painful. We also have a recent report[2] of a user of the API segfaulting, which is a direct result of it being complex to use. This commit addresses the root cause of that bug. This change is larger than I'd like, but there's no easy way to avoid it that wouldn't involve even more verbose intermediate steps. We use the "argv" as the source of truth over the "args", so we need to change all parts of run-command.[ch] itself, as well as the trace2 logging at the same time. The resulting Windows-specific code in start_command() is a bit nasty, as we're now assigning to a strvec's "v" member, instead of to our own "argv". There was a suggestion of some alternate approaches in reply to an earlier version of this commit[3], but let's leave larger a larger and needless refactoring of this code for now. 1. http://lore.kernel.org/git/YT6BnnXeAWn8BycF@coredump.intra.peff.net 2. https://lore.kernel.org/git/20211120194048.12125-1-ematsumiya@suse.de/ 3. https://lore.kernel.org/git/patch-5.5-ea1011f7473-20211122T153605Z-avarab@gmail.com/ Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-11-26 06:52:22 +08:00
int ret = wait_or_whine(cmd->pid, cmd->args.v[0], 1);
if (ret != -1)
trace2_child_exit(cmd, ret);
return ret;
pager: don't use unsafe functions in signal handlers Since the commit a3da8821208d (pager: do wait_for_pager on signal death), we call wait_for_pager() in the pager's signal handler. The recent bug report revealed that this causes a deadlock in glibc at aborting "git log" [*1*]. When this happens, git process is left unterminated, and it can't be killed by SIGTERM but only by SIGKILL. The problem is that wait_for_pager() function does more than waiting for pager process's termination, but it does cleanups and printing errors. Unfortunately, the functions that may be used in a signal handler are very limited [*2*]. Particularly, malloc(), free() and the variants can't be used in a signal handler because they take a mutex internally in glibc. This was the cause of the deadlock above. Other than the direct calls of malloc/free, many functions calling malloc/free can't be used. strerror() is such one, either. Also the usage of fflush() and printf() in a signal handler is bad, although it seems working so far. In a safer side, we should avoid them, too. This patch tries to reduce the calls of such functions in signal handlers. wait_for_signal() takes a flag and avoids the unsafe calls. Also, finish_command_in_signal() is introduced for the same reason. There the free() calls are removed, and only waits for the children without whining at errors. [*1*] https://bugzilla.opensuse.org/show_bug.cgi?id=942297 [*2*] http://pubs.opengroup.org/onlinepubs/9699919799/functions/V2_chap02.html#tag_15_04_03 Signed-off-by: Takashi Iwai <tiwai@suse.de> Reviewed-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-09-04 17:35:57 +08:00
}
int run_command(struct child_process *cmd)
{
int code;
if (cmd->out < 0 || cmd->err < 0)
BUG("run_command with a pipe can cause deadlock");
code = start_command(cmd);
if (code)
return code;
return finish_command(cmd);
}
#ifndef NO_PTHREADS
static pthread_t main_thread;
static int main_thread_set;
static pthread_key_t async_key;
run-command: use thread-aware die_is_recursing routine If we die from an async thread, we do not actually exit the program, but just kill the thread. This confuses the static counter in usage.c's default die_is_recursing function; it updates the counter once for the thread death, and then when the main program calls die() itself, it erroneously thinks we are recursing. The end result is that we print "recursion detected in die handler" instead of the real error in such a case (the easiest way to trigger this is having a remote connection hang up while running a sideband demultiplexer). This patch solves it by using a per-thread counter when the async_die function is installed; we detect recursion in each thread (including the main one), but they do not step on each other's toes. Other threaded code does not need to worry about this, as they do not install specialized die handlers; they just let a die() from a sub-thread take down the whole program. Since we are overriding the default recursion-check function, there is an interesting corner case that is not a problem, but bears some explanation. Imagine the main thread calls die(), and then in the die_routine starts an async call. We will switch to using thread-local storage, which starts at 0, for the main thread's counter, even though the original counter was actually at 1. That's OK, though, for two reasons: 1. It would miss only the first level of recursion, and would still find recursive failures inside the async helper. 2. We do not currently and are not likely to start doing anything as heavyweight as starting an async routine from within a die routine or helper function. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-04-17 03:50:07 +08:00
static pthread_key_t async_die_counter;
static void *run_thread(void *data)
{
struct async *async = data;
intptr_t ret;
run-command: teach async threads to ignore SIGPIPE Async processes can be implemented as separate forked processes, or as threads (depending on the NO_PTHREADS setting). In the latter case, if an async thread gets SIGPIPE, it takes down the whole process. This is obviously bad if the main process was not otherwise going to die, but even if we were going to die, it means the main process does not have a chance to report a useful error message. There's also the small matter that forked async processes will not take the main process down on a signal, meaning git will behave differently depending on the NO_PTHREADS setting. This patch fixes it by adding a new flag to "struct async" to block SIGPIPE just in the async thread. In theory, this should always be on (which makes async threads behave more like async processes), but we would first want to make sure that each async process we spawn is careful about checking return codes from write() and would not spew endlessly into a dead pipe. So let's start with it as optional, and we can enable it for specific sites in future patches. The natural name for this option would be "ignore_sigpipe", since that's what it does for the threaded case. But since that name might imply that we are ignoring it in all cases (including the separate-process one), let's call it "isolate_sigpipe". What we are really asking for is isolation. I.e., not to have our main process taken down by signals spawned by the async process. How that is implemented is up to the run-command code. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2016-04-20 06:49:41 +08:00
if (async->isolate_sigpipe) {
sigset_t mask;
sigemptyset(&mask);
sigaddset(&mask, SIGPIPE);
if (pthread_sigmask(SIG_BLOCK, &mask, NULL)) {
run-command: teach async threads to ignore SIGPIPE Async processes can be implemented as separate forked processes, or as threads (depending on the NO_PTHREADS setting). In the latter case, if an async thread gets SIGPIPE, it takes down the whole process. This is obviously bad if the main process was not otherwise going to die, but even if we were going to die, it means the main process does not have a chance to report a useful error message. There's also the small matter that forked async processes will not take the main process down on a signal, meaning git will behave differently depending on the NO_PTHREADS setting. This patch fixes it by adding a new flag to "struct async" to block SIGPIPE just in the async thread. In theory, this should always be on (which makes async threads behave more like async processes), but we would first want to make sure that each async process we spawn is careful about checking return codes from write() and would not spew endlessly into a dead pipe. So let's start with it as optional, and we can enable it for specific sites in future patches. The natural name for this option would be "ignore_sigpipe", since that's what it does for the threaded case. But since that name might imply that we are ignoring it in all cases (including the separate-process one), let's call it "isolate_sigpipe". What we are really asking for is isolation. I.e., not to have our main process taken down by signals spawned by the async process. How that is implemented is up to the run-command code. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2016-04-20 06:49:41 +08:00
ret = error("unable to block SIGPIPE in async thread");
return (void *)ret;
}
}
pthread_setspecific(async_key, async);
ret = async->proc(async->proc_in, async->proc_out, async->data);
return (void *)ret;
}
static NORETURN void die_async(const char *err, va_list params)
{
report_fn die_message_fn = get_die_message_routine();
die_message_fn(err, params);
if (in_async()) {
struct async *async = pthread_getspecific(async_key);
if (async->proc_in >= 0)
close(async->proc_in);
if (async->proc_out >= 0)
close(async->proc_out);
pthread_exit((void *)128);
}
exit(128);
}
run-command: use thread-aware die_is_recursing routine If we die from an async thread, we do not actually exit the program, but just kill the thread. This confuses the static counter in usage.c's default die_is_recursing function; it updates the counter once for the thread death, and then when the main program calls die() itself, it erroneously thinks we are recursing. The end result is that we print "recursion detected in die handler" instead of the real error in such a case (the easiest way to trigger this is having a remote connection hang up while running a sideband demultiplexer). This patch solves it by using a per-thread counter when the async_die function is installed; we detect recursion in each thread (including the main one), but they do not step on each other's toes. Other threaded code does not need to worry about this, as they do not install specialized die handlers; they just let a die() from a sub-thread take down the whole program. Since we are overriding the default recursion-check function, there is an interesting corner case that is not a problem, but bears some explanation. Imagine the main thread calls die(), and then in the die_routine starts an async call. We will switch to using thread-local storage, which starts at 0, for the main thread's counter, even though the original counter was actually at 1. That's OK, though, for two reasons: 1. It would miss only the first level of recursion, and would still find recursive failures inside the async helper. 2. We do not currently and are not likely to start doing anything as heavyweight as starting an async routine from within a die routine or helper function. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-04-17 03:50:07 +08:00
static int async_die_is_recursing(void)
{
void *ret = pthread_getspecific(async_die_counter);
pthread_setspecific(async_die_counter, &async_die_counter); /* set to any non-NULL valid pointer */
run-command: use thread-aware die_is_recursing routine If we die from an async thread, we do not actually exit the program, but just kill the thread. This confuses the static counter in usage.c's default die_is_recursing function; it updates the counter once for the thread death, and then when the main program calls die() itself, it erroneously thinks we are recursing. The end result is that we print "recursion detected in die handler" instead of the real error in such a case (the easiest way to trigger this is having a remote connection hang up while running a sideband demultiplexer). This patch solves it by using a per-thread counter when the async_die function is installed; we detect recursion in each thread (including the main one), but they do not step on each other's toes. Other threaded code does not need to worry about this, as they do not install specialized die handlers; they just let a die() from a sub-thread take down the whole program. Since we are overriding the default recursion-check function, there is an interesting corner case that is not a problem, but bears some explanation. Imagine the main thread calls die(), and then in the die_routine starts an async call. We will switch to using thread-local storage, which starts at 0, for the main thread's counter, even though the original counter was actually at 1. That's OK, though, for two reasons: 1. It would miss only the first level of recursion, and would still find recursive failures inside the async helper. 2. We do not currently and are not likely to start doing anything as heavyweight as starting an async routine from within a die routine or helper function. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-04-17 03:50:07 +08:00
return ret != NULL;
}
int in_async(void)
{
if (!main_thread_set)
return 0; /* no asyncs started yet */
return !pthread_equal(main_thread, pthread_self());
}
static void NORETURN async_exit(int code)
write_or_die: handle EPIPE in async threads When write_or_die() sees EPIPE, it treats it specially by converting it into a SIGPIPE death. We obviously cannot ignore it, as the write has failed and the caller expects us to die. But likewise, we cannot just call die(), because printing any message at all would be a nuisance during normal operations. However, this is a problem if write_or_die() is called from a thread. Our raised signal ends up killing the whole process, when logically we just need to kill the thread (after all, if we are ignoring SIGPIPE, there is good reason to think that the main thread is expecting to handle it). Inside an async thread, the die() code already does the right thing, because we use our custom die_async() routine, which calls pthread_join(). So ideally we would piggy-back on that, and simply call: die_quietly_with_code(141); or similar. But refactoring the die code to do this is surprisingly non-trivial. The die_routines themselves handle both printing and the decision of the exit code. Every one of them would have to be modified to take new parameters for the code, and to tell us to be quiet. Instead, we can just teach write_or_die() to check for the async case and handle it specially. We do have to build an interface to abstract the async exit, but it's simple and self-contained. If we had many call-sites that wanted to do this die_quietly_with_code(), this approach wouldn't scale as well, but we don't. This is the only place where do this weird exit trick. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2016-02-24 15:40:16 +08:00
{
pthread_exit((void *)(intptr_t)code);
}
#else
static struct {
void (**handlers)(void);
size_t nr;
size_t alloc;
} git_atexit_hdlrs;
static int git_atexit_installed;
static void git_atexit_dispatch(void)
{
size_t i;
for (i=git_atexit_hdlrs.nr ; i ; i--)
git_atexit_hdlrs.handlers[i-1]();
}
static void git_atexit_clear(void)
{
free(git_atexit_hdlrs.handlers);
memset(&git_atexit_hdlrs, 0, sizeof(git_atexit_hdlrs));
git_atexit_installed = 0;
}
#undef atexit
int git_atexit(void (*handler)(void))
{
ALLOC_GROW(git_atexit_hdlrs.handlers, git_atexit_hdlrs.nr + 1, git_atexit_hdlrs.alloc);
git_atexit_hdlrs.handlers[git_atexit_hdlrs.nr++] = handler;
if (!git_atexit_installed) {
if (atexit(&git_atexit_dispatch))
return -1;
git_atexit_installed = 1;
}
return 0;
}
#define atexit git_atexit
static int process_is_async;
int in_async(void)
{
return process_is_async;
}
static void NORETURN async_exit(int code)
write_or_die: handle EPIPE in async threads When write_or_die() sees EPIPE, it treats it specially by converting it into a SIGPIPE death. We obviously cannot ignore it, as the write has failed and the caller expects us to die. But likewise, we cannot just call die(), because printing any message at all would be a nuisance during normal operations. However, this is a problem if write_or_die() is called from a thread. Our raised signal ends up killing the whole process, when logically we just need to kill the thread (after all, if we are ignoring SIGPIPE, there is good reason to think that the main thread is expecting to handle it). Inside an async thread, the die() code already does the right thing, because we use our custom die_async() routine, which calls pthread_join(). So ideally we would piggy-back on that, and simply call: die_quietly_with_code(141); or similar. But refactoring the die code to do this is surprisingly non-trivial. The die_routines themselves handle both printing and the decision of the exit code. Every one of them would have to be modified to take new parameters for the code, and to tell us to be quiet. Instead, we can just teach write_or_die() to check for the async case and handle it specially. We do have to build an interface to abstract the async exit, but it's simple and self-contained. If we had many call-sites that wanted to do this die_quietly_with_code(), this approach wouldn't scale as well, but we don't. This is the only place where do this weird exit trick. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2016-02-24 15:40:16 +08:00
{
exit(code);
}
#endif
void check_pipe(int err)
{
if (err == EPIPE) {
if (in_async())
async_exit(141);
signal(SIGPIPE, SIG_DFL);
raise(SIGPIPE);
/* Should never happen, but just in case... */
exit(141);
}
}
int start_async(struct async *async)
{
int need_in, need_out;
int fdin[2], fdout[2];
int proc_in, proc_out;
need_in = async->in < 0;
if (need_in) {
if (pipe(fdin) < 0) {
if (async->out > 0)
close(async->out);
return error_errno("cannot create pipe");
}
async->in = fdin[1];
}
need_out = async->out < 0;
if (need_out) {
if (pipe(fdout) < 0) {
if (need_in)
close_pair(fdin);
else if (async->in)
close(async->in);
return error_errno("cannot create pipe");
}
async->out = fdout[0];
}
if (need_in)
proc_in = fdin[0];
else if (async->in)
proc_in = async->in;
else
proc_in = -1;
if (need_out)
proc_out = fdout[1];
else if (async->out)
proc_out = async->out;
else
proc_out = -1;
#ifdef NO_PTHREADS
/* Flush stdio before fork() to avoid cloning buffers */
fflush(NULL);
async->pid = fork();
if (async->pid < 0) {
error_errno("fork (async) failed");
goto error;
}
if (!async->pid) {
if (need_in)
close(fdin[1]);
if (need_out)
close(fdout[0]);
git_atexit_clear();
process_is_async = 1;
exit(!!async->proc(proc_in, proc_out, async->data));
}
mark_child_for_cleanup(async->pid, NULL);
if (need_in)
close(fdin[0]);
else if (async->in)
close(async->in);
if (need_out)
close(fdout[1]);
else if (async->out)
close(async->out);
#else
if (!main_thread_set) {
/*
* We assume that the first time that start_async is called
* it is from the main thread.
*/
main_thread_set = 1;
main_thread = pthread_self();
pthread_key_create(&async_key, NULL);
run-command: use thread-aware die_is_recursing routine If we die from an async thread, we do not actually exit the program, but just kill the thread. This confuses the static counter in usage.c's default die_is_recursing function; it updates the counter once for the thread death, and then when the main program calls die() itself, it erroneously thinks we are recursing. The end result is that we print "recursion detected in die handler" instead of the real error in such a case (the easiest way to trigger this is having a remote connection hang up while running a sideband demultiplexer). This patch solves it by using a per-thread counter when the async_die function is installed; we detect recursion in each thread (including the main one), but they do not step on each other's toes. Other threaded code does not need to worry about this, as they do not install specialized die handlers; they just let a die() from a sub-thread take down the whole program. Since we are overriding the default recursion-check function, there is an interesting corner case that is not a problem, but bears some explanation. Imagine the main thread calls die(), and then in the die_routine starts an async call. We will switch to using thread-local storage, which starts at 0, for the main thread's counter, even though the original counter was actually at 1. That's OK, though, for two reasons: 1. It would miss only the first level of recursion, and would still find recursive failures inside the async helper. 2. We do not currently and are not likely to start doing anything as heavyweight as starting an async routine from within a die routine or helper function. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-04-17 03:50:07 +08:00
pthread_key_create(&async_die_counter, NULL);
set_die_routine(die_async);
run-command: use thread-aware die_is_recursing routine If we die from an async thread, we do not actually exit the program, but just kill the thread. This confuses the static counter in usage.c's default die_is_recursing function; it updates the counter once for the thread death, and then when the main program calls die() itself, it erroneously thinks we are recursing. The end result is that we print "recursion detected in die handler" instead of the real error in such a case (the easiest way to trigger this is having a remote connection hang up while running a sideband demultiplexer). This patch solves it by using a per-thread counter when the async_die function is installed; we detect recursion in each thread (including the main one), but they do not step on each other's toes. Other threaded code does not need to worry about this, as they do not install specialized die handlers; they just let a die() from a sub-thread take down the whole program. Since we are overriding the default recursion-check function, there is an interesting corner case that is not a problem, but bears some explanation. Imagine the main thread calls die(), and then in the die_routine starts an async call. We will switch to using thread-local storage, which starts at 0, for the main thread's counter, even though the original counter was actually at 1. That's OK, though, for two reasons: 1. It would miss only the first level of recursion, and would still find recursive failures inside the async helper. 2. We do not currently and are not likely to start doing anything as heavyweight as starting an async routine from within a die routine or helper function. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-04-17 03:50:07 +08:00
set_die_is_recursing_routine(async_die_is_recursing);
}
if (proc_in >= 0)
set_cloexec(proc_in);
if (proc_out >= 0)
set_cloexec(proc_out);
async->proc_in = proc_in;
async->proc_out = proc_out;
{
int err = pthread_create(&async->tid, NULL, run_thread, async);
if (err) {
error(_("cannot create async thread: %s"), strerror(err));
goto error;
}
}
#endif
return 0;
error:
if (need_in)
close_pair(fdin);
else if (async->in)
close(async->in);
if (need_out)
close_pair(fdout);
else if (async->out)
close(async->out);
return -1;
}
int finish_async(struct async *async)
{
#ifdef NO_PTHREADS
int ret = wait_or_whine(async->pid, "child process", 0);
invalidate_lstat_cache();
return ret;
#else
void *ret = (void *)(intptr_t)(-1);
if (pthread_join(async->tid, &ret))
error("pthread_join failed");
invalidate_lstat_cache();
return (int)(intptr_t)ret;
#endif
}
int async_with_fork(void)
{
#ifdef NO_PTHREADS
return 1;
#else
return 0;
#endif
}
struct io_pump {
/* initialized by caller */
int fd;
int type; /* POLLOUT or POLLIN */
union {
struct {
const char *buf;
size_t len;
} out;
struct {
struct strbuf *buf;
size_t hint;
} in;
} u;
/* returned by pump_io */
int error; /* 0 for success, otherwise errno */
/* internal use */
struct pollfd *pfd;
};
static int pump_io_round(struct io_pump *slots, int nr, struct pollfd *pfd)
{
int pollsize = 0;
int i;
for (i = 0; i < nr; i++) {
struct io_pump *io = &slots[i];
if (io->fd < 0)
continue;
pfd[pollsize].fd = io->fd;
pfd[pollsize].events = io->type;
io->pfd = &pfd[pollsize++];
}
if (!pollsize)
return 0;
if (poll(pfd, pollsize, -1) < 0) {
if (errno == EINTR)
return 1;
die_errno("poll failed");
}
for (i = 0; i < nr; i++) {
struct io_pump *io = &slots[i];
if (io->fd < 0)
continue;
if (!(io->pfd->revents & (POLLOUT|POLLIN|POLLHUP|POLLERR|POLLNVAL)))
continue;
if (io->type == POLLOUT) {
pipe_command(): avoid xwrite() for writing to pipe If xwrite() sees an EAGAIN response, it will loop forever until the write succeeds (or encounters a real error). This is due to ef1cf0167a (xwrite: poll on non-blocking FDs, 2016-06-26), with the idea that we won't be surprised by a descriptor unexpectedly set as non-blocking. But that will make things awkward when we do want a non-blocking descriptor, and a future patch will switch pipe_command() to using one. In that case, looping on EAGAIN is bad, because the process on the other end of the pipe may be waiting on us before doing another read() on the pipe, which would mean we deadlock. In practice we're not supposed to ever see EAGAIN here, since poll() will have just told us the descriptor is ready for writing. But our Windows emulation of poll() will always return "ready" for writing to a pipe descriptor! This is due to 94f4d01932 (mingw: workaround for hangs when sending STDIN, 2020-02-17). Our best bet in that case is to keep handling other descriptors, as any read() we do may allow the child command to make forward progress (i.e., its write() finishes, and then it read()s from its stdin, freeing up space in the pipe buffer). This means we might busy-loop between poll() and write() on Windows if the child command is slow to read our input, but it's much better than the alternative of deadlocking. In practice, this busy-looping should be rare: - for small inputs, we'll just write the whole thing in a single write() anyway, non-blocking or not - for larger inputs where the child reads input and then processes it before writing (e.g., gpg verifying a signature), we may make a few extra write() calls that get EAGAIN during the initial write, but once it has taken in the whole input, we'll correctly block waiting to read back the data. - for larger inputs where the child process is streaming output back (like a diff filter), we'll likewise see some extra EAGAINs, but most of them will be followed immediately by a read(), which will let the child command make forward progress. Of course it won't happen at all for now, since we don't yet use a non-blocking pipe. This is just preparation for when we do. Helped-by: René Scharfe <l.s.r@web.de> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2022-08-17 14:08:06 +08:00
ssize_t len;
/*
* Don't use xwrite() here. It loops forever on EAGAIN,
* and we're in our own poll() loop here.
*
* Note that we lose xwrite()'s handling of MAX_IO_SIZE
* and EINTR, so we have to implement those ourselves.
*/
len = write(io->fd, io->u.out.buf,
io->u.out.len <= MAX_IO_SIZE ?
io->u.out.len : MAX_IO_SIZE);
if (len < 0) {
if (errno != EINTR && errno != EAGAIN &&
errno != ENOSPC) {
pipe_command(): avoid xwrite() for writing to pipe If xwrite() sees an EAGAIN response, it will loop forever until the write succeeds (or encounters a real error). This is due to ef1cf0167a (xwrite: poll on non-blocking FDs, 2016-06-26), with the idea that we won't be surprised by a descriptor unexpectedly set as non-blocking. But that will make things awkward when we do want a non-blocking descriptor, and a future patch will switch pipe_command() to using one. In that case, looping on EAGAIN is bad, because the process on the other end of the pipe may be waiting on us before doing another read() on the pipe, which would mean we deadlock. In practice we're not supposed to ever see EAGAIN here, since poll() will have just told us the descriptor is ready for writing. But our Windows emulation of poll() will always return "ready" for writing to a pipe descriptor! This is due to 94f4d01932 (mingw: workaround for hangs when sending STDIN, 2020-02-17). Our best bet in that case is to keep handling other descriptors, as any read() we do may allow the child command to make forward progress (i.e., its write() finishes, and then it read()s from its stdin, freeing up space in the pipe buffer). This means we might busy-loop between poll() and write() on Windows if the child command is slow to read our input, but it's much better than the alternative of deadlocking. In practice, this busy-looping should be rare: - for small inputs, we'll just write the whole thing in a single write() anyway, non-blocking or not - for larger inputs where the child reads input and then processes it before writing (e.g., gpg verifying a signature), we may make a few extra write() calls that get EAGAIN during the initial write, but once it has taken in the whole input, we'll correctly block waiting to read back the data. - for larger inputs where the child process is streaming output back (like a diff filter), we'll likewise see some extra EAGAINs, but most of them will be followed immediately by a read(), which will let the child command make forward progress. Of course it won't happen at all for now, since we don't yet use a non-blocking pipe. This is just preparation for when we do. Helped-by: René Scharfe <l.s.r@web.de> Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2022-08-17 14:08:06 +08:00
io->error = errno;
close(io->fd);
io->fd = -1;
}
} else {
io->u.out.buf += len;
io->u.out.len -= len;
if (!io->u.out.len) {
close(io->fd);
io->fd = -1;
}
}
}
if (io->type == POLLIN) {
ssize_t len = strbuf_read_once(io->u.in.buf,
io->fd, io->u.in.hint);
if (len < 0)
io->error = errno;
if (len <= 0) {
close(io->fd);
io->fd = -1;
}
}
}
return 1;
}
static int pump_io(struct io_pump *slots, int nr)
{
struct pollfd *pfd;
int i;
for (i = 0; i < nr; i++)
slots[i].error = 0;
ALLOC_ARRAY(pfd, nr);
while (pump_io_round(slots, nr, pfd))
; /* nothing */
free(pfd);
/* There may be multiple errno values, so just pick the first. */
for (i = 0; i < nr; i++) {
if (slots[i].error) {
errno = slots[i].error;
return -1;
}
}
return 0;
}
int pipe_command(struct child_process *cmd,
const char *in, size_t in_len,
struct strbuf *out, size_t out_hint,
struct strbuf *err, size_t err_hint)
{
struct io_pump io[3];
int nr = 0;
if (in)
cmd->in = -1;
if (out)
cmd->out = -1;
if (err)
cmd->err = -1;
if (start_command(cmd) < 0)
return -1;
if (in) {
pipe_command(): mark stdin descriptor as non-blocking Our pipe_command() helper lets you both write to and read from a child process on its stdin/stdout. It's supposed to work without deadlocks because we use poll() to check when descriptors are ready for reading or writing. But there's a bug: if both the data to be written and the data to be read back exceed the pipe buffer, we'll deadlock. The issue is that the code assumes that if you have, say, a 2MB buffer to write and poll() tells you that the pipe descriptor is ready for writing, that calling: write(cmd->in, buf, 2*1024*1024); will do a partial write, filling the pipe buffer and then returning what it did write. And that is what it would do on a socket, but not for a pipe. When writing to a pipe, at least on Linux, it will block waiting for the child process to read() more. And now we have a potential deadlock, because the child may be writing back to us, waiting for us to read() ourselves. An easy way to trigger this is: git -c add.interactive.useBuiltin=true \ -c interactive.diffFilter=cat \ checkout -p HEAD~200 The diff against HEAD~200 will be big, and the filter wants to write all of it back to us (obviously this is a dummy filter, but in the real world something like diff-highlight would similarly stream back a big output). If you set add.interactive.useBuiltin to false, the problem goes away, because now we're not using pipe_command() anymore (instead, that part happens in perl). But this isn't a bug in the interactive code at all. It's the underlying pipe_command() code which is broken, and has been all along. We presumably didn't notice because most calls only do input _or_ output, not both. And the few that do both, like gpg calls, may have large inputs or outputs, but never both at the same time (e.g., consider signing, which has a large payload but a small signature comes back). The obvious fix is to put the descriptor into non-blocking mode, and indeed, that makes the problem go away. Callers shouldn't need to care, because they never see the descriptor (they hand us a buffer to feed into it). The included test fails reliably on Linux without this patch. Curiously, it doesn't fail in our Windows CI environment, but has been reported to do so for individual developers. It should pass in any environment after this patch (courtesy of the compat/ layers added in the last few commits). Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2022-08-17 14:10:22 +08:00
if (enable_pipe_nonblock(cmd->in) < 0) {
error_errno("unable to make pipe non-blocking");
close(cmd->in);
if (out)
close(cmd->out);
if (err)
close(cmd->err);
return -1;
}
io[nr].fd = cmd->in;
io[nr].type = POLLOUT;
io[nr].u.out.buf = in;
io[nr].u.out.len = in_len;
nr++;
}
if (out) {
io[nr].fd = cmd->out;
io[nr].type = POLLIN;
io[nr].u.in.buf = out;
io[nr].u.in.hint = out_hint;
nr++;
}
if (err) {
io[nr].fd = cmd->err;
io[nr].type = POLLIN;
io[nr].u.in.buf = err;
io[nr].u.in.hint = err_hint;
nr++;
}
if (pump_io(io, nr) < 0) {
finish_command(cmd); /* throw away exit code */
return -1;
}
return finish_command(cmd);
}
run-command: add an asynchronous parallel child processor This allows to run external commands in parallel with ordered output on stderr. If we run external commands in parallel we cannot pipe the output directly to the our stdout/err as it would mix up. So each process's output will flow through a pipe, which we buffer. One subprocess can be directly piped to out stdout/err for a low latency feedback to the user. Example: Let's assume we have 5 submodules A,B,C,D,E and each fetch takes a different amount of time as the different submodules vary in size, then the output of fetches in sequential order might look like this: time --> output: |---A---| |-B-| |-------C-------| |-D-| |-E-| When we schedule these submodules into maximal two parallel processes, a schedule and sample output over time may look like this: process 1: |---A---| |-D-| |-E-| process 2: |-B-| |-------C-------| output: |---A---|B|---C-------|DE So A will be perceived as it would run normally in the single child version. As B has finished by the time A is done, we can dump its whole progress buffer on stderr, such that it looks like it finished in no time. Once that is done, C is determined to be the visible child and its progress will be reported in real time. So this way of output is really good for human consumption, as it only changes the timing, not the actual output. For machine consumption the output needs to be prepared in the tasks, by either having a prefix per line or per block to indicate whose tasks output is displayed, because the output order may not follow the original sequential ordering: |----A----| |--B--| |-C-| will be scheduled to be all parallel: process 1: |----A----| process 2: |--B--| process 3: |-C-| output: |----A----|CB This happens because C finished before B did, so it will be queued for output before B. To detect when a child has finished executing, we check interleaved with other actions (such as checking the liveliness of children or starting new processes) whether the stderr pipe still exists. Once a child closed its stderr stream, we assume it is terminating very soon, and use `finish_command()` from the single external process execution interface to collect the exit status. By maintaining the strong assumption of stderr being open until the very end of a child process, we can avoid other hassle such as an implementation using `waitpid(-1)`, which is not implemented in Windows. Signed-off-by: Stefan Beller <sbeller@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-12-16 08:04:10 +08:00
enum child_state {
GIT_CP_FREE,
GIT_CP_WORKING,
GIT_CP_WAIT_CLEANUP,
};
struct parallel_processes {
run-command API: make "n" parameter a "size_t" Make the "n" variable added in c553c72eed6 (run-command: add an asynchronous parallel child processor, 2015-12-15) a "size_t". As we'll see in a subsequent commit we do pass "0" here, but never "jobs < 0". We could have made it an "unsigned int", but as we're having to change this let's not leave another case in the codebase where a size_t and "unsigned int" size differ on some platforms. In this case it's likely to never matter, but it's easier to not need to worry about it. After this and preceding changes: make run-command.o DEVOPTS=extra-all CFLAGS=-Wno-unused-parameter Only has one (and new) -Wsigned-compare warning relevant to a comparison about our "n" or "{nr,max}_processes": About using our "n" (size_t) in the same expression as online_cpus() (int). A subsequent commit will adjust & deal with online_cpus() and that warning. The only users of the "n" parameter are: * builtin/fetch.c: defaults to 1, reads from the "fetch.parallel" config. As seen in the code that parses the config added in d54dea77dba (fetch: let --jobs=<n> parallelize --multiple, too, 2019-10-05) will die if the git_config_int() return value is < 0. It will however pass us n = 0, as we'll see in a subsequent commit. * submodule.c: defaults to 1, reads from "submodule.fetchJobs" config. Read via code originally added in a028a1930c6 (fetching submodules: respect `submodule.fetchJobs` config option, 2016-02-29). It now piggy-backs on the the submodule.fetchJobs code and validation added in f20e7c1ea24 (submodule: remove submodule.fetchjobs from submodule-config parsing, 2017-08-02). Like builtin/fetch.c it will die if the git_config_int() return value is < 0, but like builtin/fetch.c it will pass us n = 0. * builtin/submodule--helper.c: defaults to 1. Read via code originally added in 2335b870fa7 (submodule update: expose parallelism to the user, 2016-02-29). Since f20e7c1ea24 (submodule: remove submodule.fetchjobs from submodule-config parsing, 2017-08-02) it shares a config parser and semantics with the submodule.c caller. * hook.c: hardcoded to 1, see 96e7225b310 (hook: add 'run' subcommand, 2021-12-22). * t/helper/test-run-command.c: can be -1 after parsing the arguments, but will then be overridden to online_cpus() before passing it to this API. See be5d88e1128 (test-tool run-command: learn to run (parts of) the testsuite, 2019-10-04). Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2022-10-13 05:02:23 +08:00
size_t nr_processes;
run-command: add an asynchronous parallel child processor This allows to run external commands in parallel with ordered output on stderr. If we run external commands in parallel we cannot pipe the output directly to the our stdout/err as it would mix up. So each process's output will flow through a pipe, which we buffer. One subprocess can be directly piped to out stdout/err for a low latency feedback to the user. Example: Let's assume we have 5 submodules A,B,C,D,E and each fetch takes a different amount of time as the different submodules vary in size, then the output of fetches in sequential order might look like this: time --> output: |---A---| |-B-| |-------C-------| |-D-| |-E-| When we schedule these submodules into maximal two parallel processes, a schedule and sample output over time may look like this: process 1: |---A---| |-D-| |-E-| process 2: |-B-| |-------C-------| output: |---A---|B|---C-------|DE So A will be perceived as it would run normally in the single child version. As B has finished by the time A is done, we can dump its whole progress buffer on stderr, such that it looks like it finished in no time. Once that is done, C is determined to be the visible child and its progress will be reported in real time. So this way of output is really good for human consumption, as it only changes the timing, not the actual output. For machine consumption the output needs to be prepared in the tasks, by either having a prefix per line or per block to indicate whose tasks output is displayed, because the output order may not follow the original sequential ordering: |----A----| |--B--| |-C-| will be scheduled to be all parallel: process 1: |----A----| process 2: |--B--| process 3: |-C-| output: |----A----|CB This happens because C finished before B did, so it will be queued for output before B. To detect when a child has finished executing, we check interleaved with other actions (such as checking the liveliness of children or starting new processes) whether the stderr pipe still exists. Once a child closed its stderr stream, we assume it is terminating very soon, and use `finish_command()` from the single external process execution interface to collect the exit status. By maintaining the strong assumption of stderr being open until the very end of a child process, we can avoid other hassle such as an implementation using `waitpid(-1)`, which is not implemented in Windows. Signed-off-by: Stefan Beller <sbeller@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-12-16 08:04:10 +08:00
struct {
enum child_state state;
struct child_process process;
struct strbuf err;
void *data;
} *children;
/*
* The struct pollfd is logically part of *children,
* but the system call expects it as its own array.
*/
struct pollfd *pfd;
unsigned shutdown : 1;
run-command API: make "n" parameter a "size_t" Make the "n" variable added in c553c72eed6 (run-command: add an asynchronous parallel child processor, 2015-12-15) a "size_t". As we'll see in a subsequent commit we do pass "0" here, but never "jobs < 0". We could have made it an "unsigned int", but as we're having to change this let's not leave another case in the codebase where a size_t and "unsigned int" size differ on some platforms. In this case it's likely to never matter, but it's easier to not need to worry about it. After this and preceding changes: make run-command.o DEVOPTS=extra-all CFLAGS=-Wno-unused-parameter Only has one (and new) -Wsigned-compare warning relevant to a comparison about our "n" or "{nr,max}_processes": About using our "n" (size_t) in the same expression as online_cpus() (int). A subsequent commit will adjust & deal with online_cpus() and that warning. The only users of the "n" parameter are: * builtin/fetch.c: defaults to 1, reads from the "fetch.parallel" config. As seen in the code that parses the config added in d54dea77dba (fetch: let --jobs=<n> parallelize --multiple, too, 2019-10-05) will die if the git_config_int() return value is < 0. It will however pass us n = 0, as we'll see in a subsequent commit. * submodule.c: defaults to 1, reads from "submodule.fetchJobs" config. Read via code originally added in a028a1930c6 (fetching submodules: respect `submodule.fetchJobs` config option, 2016-02-29). It now piggy-backs on the the submodule.fetchJobs code and validation added in f20e7c1ea24 (submodule: remove submodule.fetchjobs from submodule-config parsing, 2017-08-02). Like builtin/fetch.c it will die if the git_config_int() return value is < 0, but like builtin/fetch.c it will pass us n = 0. * builtin/submodule--helper.c: defaults to 1. Read via code originally added in 2335b870fa7 (submodule update: expose parallelism to the user, 2016-02-29). Since f20e7c1ea24 (submodule: remove submodule.fetchjobs from submodule-config parsing, 2017-08-02) it shares a config parser and semantics with the submodule.c caller. * hook.c: hardcoded to 1, see 96e7225b310 (hook: add 'run' subcommand, 2021-12-22). * t/helper/test-run-command.c: can be -1 after parsing the arguments, but will then be overridden to online_cpus() before passing it to this API. See be5d88e1128 (test-tool run-command: learn to run (parts of) the testsuite, 2019-10-04). Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2022-10-13 05:02:23 +08:00
size_t output_owner;
run-command: add an asynchronous parallel child processor This allows to run external commands in parallel with ordered output on stderr. If we run external commands in parallel we cannot pipe the output directly to the our stdout/err as it would mix up. So each process's output will flow through a pipe, which we buffer. One subprocess can be directly piped to out stdout/err for a low latency feedback to the user. Example: Let's assume we have 5 submodules A,B,C,D,E and each fetch takes a different amount of time as the different submodules vary in size, then the output of fetches in sequential order might look like this: time --> output: |---A---| |-B-| |-------C-------| |-D-| |-E-| When we schedule these submodules into maximal two parallel processes, a schedule and sample output over time may look like this: process 1: |---A---| |-D-| |-E-| process 2: |-B-| |-------C-------| output: |---A---|B|---C-------|DE So A will be perceived as it would run normally in the single child version. As B has finished by the time A is done, we can dump its whole progress buffer on stderr, such that it looks like it finished in no time. Once that is done, C is determined to be the visible child and its progress will be reported in real time. So this way of output is really good for human consumption, as it only changes the timing, not the actual output. For machine consumption the output needs to be prepared in the tasks, by either having a prefix per line or per block to indicate whose tasks output is displayed, because the output order may not follow the original sequential ordering: |----A----| |--B--| |-C-| will be scheduled to be all parallel: process 1: |----A----| process 2: |--B--| process 3: |-C-| output: |----A----|CB This happens because C finished before B did, so it will be queued for output before B. To detect when a child has finished executing, we check interleaved with other actions (such as checking the liveliness of children or starting new processes) whether the stderr pipe still exists. Once a child closed its stderr stream, we assume it is terminating very soon, and use `finish_command()` from the single external process execution interface to collect the exit status. By maintaining the strong assumption of stderr being open until the very end of a child process, we can avoid other hassle such as an implementation using `waitpid(-1)`, which is not implemented in Windows. Signed-off-by: Stefan Beller <sbeller@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-12-16 08:04:10 +08:00
struct strbuf buffered_output; /* of finished children */
};
struct parallel_processes_for_signal {
const struct run_process_parallel_opts *opts;
const struct parallel_processes *pp;
};
static void kill_children(const struct parallel_processes *pp,
const struct run_process_parallel_opts *opts,
int signo)
run-command: add an asynchronous parallel child processor This allows to run external commands in parallel with ordered output on stderr. If we run external commands in parallel we cannot pipe the output directly to the our stdout/err as it would mix up. So each process's output will flow through a pipe, which we buffer. One subprocess can be directly piped to out stdout/err for a low latency feedback to the user. Example: Let's assume we have 5 submodules A,B,C,D,E and each fetch takes a different amount of time as the different submodules vary in size, then the output of fetches in sequential order might look like this: time --> output: |---A---| |-B-| |-------C-------| |-D-| |-E-| When we schedule these submodules into maximal two parallel processes, a schedule and sample output over time may look like this: process 1: |---A---| |-D-| |-E-| process 2: |-B-| |-------C-------| output: |---A---|B|---C-------|DE So A will be perceived as it would run normally in the single child version. As B has finished by the time A is done, we can dump its whole progress buffer on stderr, such that it looks like it finished in no time. Once that is done, C is determined to be the visible child and its progress will be reported in real time. So this way of output is really good for human consumption, as it only changes the timing, not the actual output. For machine consumption the output needs to be prepared in the tasks, by either having a prefix per line or per block to indicate whose tasks output is displayed, because the output order may not follow the original sequential ordering: |----A----| |--B--| |-C-| will be scheduled to be all parallel: process 1: |----A----| process 2: |--B--| process 3: |-C-| output: |----A----|CB This happens because C finished before B did, so it will be queued for output before B. To detect when a child has finished executing, we check interleaved with other actions (such as checking the liveliness of children or starting new processes) whether the stderr pipe still exists. Once a child closed its stderr stream, we assume it is terminating very soon, and use `finish_command()` from the single external process execution interface to collect the exit status. By maintaining the strong assumption of stderr being open until the very end of a child process, we can avoid other hassle such as an implementation using `waitpid(-1)`, which is not implemented in Windows. Signed-off-by: Stefan Beller <sbeller@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-12-16 08:04:10 +08:00
{
for (size_t i = 0; i < opts->processes; i++)
run-command: add an asynchronous parallel child processor This allows to run external commands in parallel with ordered output on stderr. If we run external commands in parallel we cannot pipe the output directly to the our stdout/err as it would mix up. So each process's output will flow through a pipe, which we buffer. One subprocess can be directly piped to out stdout/err for a low latency feedback to the user. Example: Let's assume we have 5 submodules A,B,C,D,E and each fetch takes a different amount of time as the different submodules vary in size, then the output of fetches in sequential order might look like this: time --> output: |---A---| |-B-| |-------C-------| |-D-| |-E-| When we schedule these submodules into maximal two parallel processes, a schedule and sample output over time may look like this: process 1: |---A---| |-D-| |-E-| process 2: |-B-| |-------C-------| output: |---A---|B|---C-------|DE So A will be perceived as it would run normally in the single child version. As B has finished by the time A is done, we can dump its whole progress buffer on stderr, such that it looks like it finished in no time. Once that is done, C is determined to be the visible child and its progress will be reported in real time. So this way of output is really good for human consumption, as it only changes the timing, not the actual output. For machine consumption the output needs to be prepared in the tasks, by either having a prefix per line or per block to indicate whose tasks output is displayed, because the output order may not follow the original sequential ordering: |----A----| |--B--| |-C-| will be scheduled to be all parallel: process 1: |----A----| process 2: |--B--| process 3: |-C-| output: |----A----|CB This happens because C finished before B did, so it will be queued for output before B. To detect when a child has finished executing, we check interleaved with other actions (such as checking the liveliness of children or starting new processes) whether the stderr pipe still exists. Once a child closed its stderr stream, we assume it is terminating very soon, and use `finish_command()` from the single external process execution interface to collect the exit status. By maintaining the strong assumption of stderr being open until the very end of a child process, we can avoid other hassle such as an implementation using `waitpid(-1)`, which is not implemented in Windows. Signed-off-by: Stefan Beller <sbeller@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-12-16 08:04:10 +08:00
if (pp->children[i].state == GIT_CP_WORKING)
kill(pp->children[i].process.pid, signo);
}
static void kill_children_signal(const struct parallel_processes_for_signal *pp_sig,
int signo)
{
kill_children(pp_sig->pp, pp_sig->opts, signo);
}
static struct parallel_processes_for_signal *pp_for_signal;
run-command: add an asynchronous parallel child processor This allows to run external commands in parallel with ordered output on stderr. If we run external commands in parallel we cannot pipe the output directly to the our stdout/err as it would mix up. So each process's output will flow through a pipe, which we buffer. One subprocess can be directly piped to out stdout/err for a low latency feedback to the user. Example: Let's assume we have 5 submodules A,B,C,D,E and each fetch takes a different amount of time as the different submodules vary in size, then the output of fetches in sequential order might look like this: time --> output: |---A---| |-B-| |-------C-------| |-D-| |-E-| When we schedule these submodules into maximal two parallel processes, a schedule and sample output over time may look like this: process 1: |---A---| |-D-| |-E-| process 2: |-B-| |-------C-------| output: |---A---|B|---C-------|DE So A will be perceived as it would run normally in the single child version. As B has finished by the time A is done, we can dump its whole progress buffer on stderr, such that it looks like it finished in no time. Once that is done, C is determined to be the visible child and its progress will be reported in real time. So this way of output is really good for human consumption, as it only changes the timing, not the actual output. For machine consumption the output needs to be prepared in the tasks, by either having a prefix per line or per block to indicate whose tasks output is displayed, because the output order may not follow the original sequential ordering: |----A----| |--B--| |-C-| will be scheduled to be all parallel: process 1: |----A----| process 2: |--B--| process 3: |-C-| output: |----A----|CB This happens because C finished before B did, so it will be queued for output before B. To detect when a child has finished executing, we check interleaved with other actions (such as checking the liveliness of children or starting new processes) whether the stderr pipe still exists. Once a child closed its stderr stream, we assume it is terminating very soon, and use `finish_command()` from the single external process execution interface to collect the exit status. By maintaining the strong assumption of stderr being open until the very end of a child process, we can avoid other hassle such as an implementation using `waitpid(-1)`, which is not implemented in Windows. Signed-off-by: Stefan Beller <sbeller@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-12-16 08:04:10 +08:00
static void handle_children_on_signal(int signo)
{
kill_children_signal(pp_for_signal, signo);
run-command: add an asynchronous parallel child processor This allows to run external commands in parallel with ordered output on stderr. If we run external commands in parallel we cannot pipe the output directly to the our stdout/err as it would mix up. So each process's output will flow through a pipe, which we buffer. One subprocess can be directly piped to out stdout/err for a low latency feedback to the user. Example: Let's assume we have 5 submodules A,B,C,D,E and each fetch takes a different amount of time as the different submodules vary in size, then the output of fetches in sequential order might look like this: time --> output: |---A---| |-B-| |-------C-------| |-D-| |-E-| When we schedule these submodules into maximal two parallel processes, a schedule and sample output over time may look like this: process 1: |---A---| |-D-| |-E-| process 2: |-B-| |-------C-------| output: |---A---|B|---C-------|DE So A will be perceived as it would run normally in the single child version. As B has finished by the time A is done, we can dump its whole progress buffer on stderr, such that it looks like it finished in no time. Once that is done, C is determined to be the visible child and its progress will be reported in real time. So this way of output is really good for human consumption, as it only changes the timing, not the actual output. For machine consumption the output needs to be prepared in the tasks, by either having a prefix per line or per block to indicate whose tasks output is displayed, because the output order may not follow the original sequential ordering: |----A----| |--B--| |-C-| will be scheduled to be all parallel: process 1: |----A----| process 2: |--B--| process 3: |-C-| output: |----A----|CB This happens because C finished before B did, so it will be queued for output before B. To detect when a child has finished executing, we check interleaved with other actions (such as checking the liveliness of children or starting new processes) whether the stderr pipe still exists. Once a child closed its stderr stream, we assume it is terminating very soon, and use `finish_command()` from the single external process execution interface to collect the exit status. By maintaining the strong assumption of stderr being open until the very end of a child process, we can avoid other hassle such as an implementation using `waitpid(-1)`, which is not implemented in Windows. Signed-off-by: Stefan Beller <sbeller@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-12-16 08:04:10 +08:00
sigchain_pop(signo);
raise(signo);
}
static void pp_init(struct parallel_processes *pp,
const struct run_process_parallel_opts *opts,
struct parallel_processes_for_signal *pp_sig)
run-command: add an asynchronous parallel child processor This allows to run external commands in parallel with ordered output on stderr. If we run external commands in parallel we cannot pipe the output directly to the our stdout/err as it would mix up. So each process's output will flow through a pipe, which we buffer. One subprocess can be directly piped to out stdout/err for a low latency feedback to the user. Example: Let's assume we have 5 submodules A,B,C,D,E and each fetch takes a different amount of time as the different submodules vary in size, then the output of fetches in sequential order might look like this: time --> output: |---A---| |-B-| |-------C-------| |-D-| |-E-| When we schedule these submodules into maximal two parallel processes, a schedule and sample output over time may look like this: process 1: |---A---| |-D-| |-E-| process 2: |-B-| |-------C-------| output: |---A---|B|---C-------|DE So A will be perceived as it would run normally in the single child version. As B has finished by the time A is done, we can dump its whole progress buffer on stderr, such that it looks like it finished in no time. Once that is done, C is determined to be the visible child and its progress will be reported in real time. So this way of output is really good for human consumption, as it only changes the timing, not the actual output. For machine consumption the output needs to be prepared in the tasks, by either having a prefix per line or per block to indicate whose tasks output is displayed, because the output order may not follow the original sequential ordering: |----A----| |--B--| |-C-| will be scheduled to be all parallel: process 1: |----A----| process 2: |--B--| process 3: |-C-| output: |----A----|CB This happens because C finished before B did, so it will be queued for output before B. To detect when a child has finished executing, we check interleaved with other actions (such as checking the liveliness of children or starting new processes) whether the stderr pipe still exists. Once a child closed its stderr stream, we assume it is terminating very soon, and use `finish_command()` from the single external process execution interface to collect the exit status. By maintaining the strong assumption of stderr being open until the very end of a child process, we can avoid other hassle such as an implementation using `waitpid(-1)`, which is not implemented in Windows. Signed-off-by: Stefan Beller <sbeller@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-12-16 08:04:10 +08:00
{
const size_t n = opts->processes;
if (!n)
BUG("you must provide a non-zero number of processes!");
run-command: add an asynchronous parallel child processor This allows to run external commands in parallel with ordered output on stderr. If we run external commands in parallel we cannot pipe the output directly to the our stdout/err as it would mix up. So each process's output will flow through a pipe, which we buffer. One subprocess can be directly piped to out stdout/err for a low latency feedback to the user. Example: Let's assume we have 5 submodules A,B,C,D,E and each fetch takes a different amount of time as the different submodules vary in size, then the output of fetches in sequential order might look like this: time --> output: |---A---| |-B-| |-------C-------| |-D-| |-E-| When we schedule these submodules into maximal two parallel processes, a schedule and sample output over time may look like this: process 1: |---A---| |-D-| |-E-| process 2: |-B-| |-------C-------| output: |---A---|B|---C-------|DE So A will be perceived as it would run normally in the single child version. As B has finished by the time A is done, we can dump its whole progress buffer on stderr, such that it looks like it finished in no time. Once that is done, C is determined to be the visible child and its progress will be reported in real time. So this way of output is really good for human consumption, as it only changes the timing, not the actual output. For machine consumption the output needs to be prepared in the tasks, by either having a prefix per line or per block to indicate whose tasks output is displayed, because the output order may not follow the original sequential ordering: |----A----| |--B--| |-C-| will be scheduled to be all parallel: process 1: |----A----| process 2: |--B--| process 3: |-C-| output: |----A----|CB This happens because C finished before B did, so it will be queued for output before B. To detect when a child has finished executing, we check interleaved with other actions (such as checking the liveliness of children or starting new processes) whether the stderr pipe still exists. Once a child closed its stderr stream, we assume it is terminating very soon, and use `finish_command()` from the single external process execution interface to collect the exit status. By maintaining the strong assumption of stderr being open until the very end of a child process, we can avoid other hassle such as an implementation using `waitpid(-1)`, which is not implemented in Windows. Signed-off-by: Stefan Beller <sbeller@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-12-16 08:04:10 +08:00
run-command API: make "n" parameter a "size_t" Make the "n" variable added in c553c72eed6 (run-command: add an asynchronous parallel child processor, 2015-12-15) a "size_t". As we'll see in a subsequent commit we do pass "0" here, but never "jobs < 0". We could have made it an "unsigned int", but as we're having to change this let's not leave another case in the codebase where a size_t and "unsigned int" size differ on some platforms. In this case it's likely to never matter, but it's easier to not need to worry about it. After this and preceding changes: make run-command.o DEVOPTS=extra-all CFLAGS=-Wno-unused-parameter Only has one (and new) -Wsigned-compare warning relevant to a comparison about our "n" or "{nr,max}_processes": About using our "n" (size_t) in the same expression as online_cpus() (int). A subsequent commit will adjust & deal with online_cpus() and that warning. The only users of the "n" parameter are: * builtin/fetch.c: defaults to 1, reads from the "fetch.parallel" config. As seen in the code that parses the config added in d54dea77dba (fetch: let --jobs=<n> parallelize --multiple, too, 2019-10-05) will die if the git_config_int() return value is < 0. It will however pass us n = 0, as we'll see in a subsequent commit. * submodule.c: defaults to 1, reads from "submodule.fetchJobs" config. Read via code originally added in a028a1930c6 (fetching submodules: respect `submodule.fetchJobs` config option, 2016-02-29). It now piggy-backs on the the submodule.fetchJobs code and validation added in f20e7c1ea24 (submodule: remove submodule.fetchjobs from submodule-config parsing, 2017-08-02). Like builtin/fetch.c it will die if the git_config_int() return value is < 0, but like builtin/fetch.c it will pass us n = 0. * builtin/submodule--helper.c: defaults to 1. Read via code originally added in 2335b870fa7 (submodule update: expose parallelism to the user, 2016-02-29). Since f20e7c1ea24 (submodule: remove submodule.fetchjobs from submodule-config parsing, 2017-08-02) it shares a config parser and semantics with the submodule.c caller. * hook.c: hardcoded to 1, see 96e7225b310 (hook: add 'run' subcommand, 2021-12-22). * t/helper/test-run-command.c: can be -1 after parsing the arguments, but will then be overridden to online_cpus() before passing it to this API. See be5d88e1128 (test-tool run-command: learn to run (parts of) the testsuite, 2019-10-04). Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2022-10-13 05:02:23 +08:00
trace_printf("run_processes_parallel: preparing to run up to %"PRIuMAX" tasks",
(uintmax_t)n);
run-command: add an asynchronous parallel child processor This allows to run external commands in parallel with ordered output on stderr. If we run external commands in parallel we cannot pipe the output directly to the our stdout/err as it would mix up. So each process's output will flow through a pipe, which we buffer. One subprocess can be directly piped to out stdout/err for a low latency feedback to the user. Example: Let's assume we have 5 submodules A,B,C,D,E and each fetch takes a different amount of time as the different submodules vary in size, then the output of fetches in sequential order might look like this: time --> output: |---A---| |-B-| |-------C-------| |-D-| |-E-| When we schedule these submodules into maximal two parallel processes, a schedule and sample output over time may look like this: process 1: |---A---| |-D-| |-E-| process 2: |-B-| |-------C-------| output: |---A---|B|---C-------|DE So A will be perceived as it would run normally in the single child version. As B has finished by the time A is done, we can dump its whole progress buffer on stderr, such that it looks like it finished in no time. Once that is done, C is determined to be the visible child and its progress will be reported in real time. So this way of output is really good for human consumption, as it only changes the timing, not the actual output. For machine consumption the output needs to be prepared in the tasks, by either having a prefix per line or per block to indicate whose tasks output is displayed, because the output order may not follow the original sequential ordering: |----A----| |--B--| |-C-| will be scheduled to be all parallel: process 1: |----A----| process 2: |--B--| process 3: |-C-| output: |----A----|CB This happens because C finished before B did, so it will be queued for output before B. To detect when a child has finished executing, we check interleaved with other actions (such as checking the liveliness of children or starting new processes) whether the stderr pipe still exists. Once a child closed its stderr stream, we assume it is terminating very soon, and use `finish_command()` from the single external process execution interface to collect the exit status. By maintaining the strong assumption of stderr being open until the very end of a child process, we can avoid other hassle such as an implementation using `waitpid(-1)`, which is not implemented in Windows. Signed-off-by: Stefan Beller <sbeller@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-12-16 08:04:10 +08:00
if (!opts->get_next_task)
BUG("you need to specify a get_next_task function");
run-command: add an asynchronous parallel child processor This allows to run external commands in parallel with ordered output on stderr. If we run external commands in parallel we cannot pipe the output directly to the our stdout/err as it would mix up. So each process's output will flow through a pipe, which we buffer. One subprocess can be directly piped to out stdout/err for a low latency feedback to the user. Example: Let's assume we have 5 submodules A,B,C,D,E and each fetch takes a different amount of time as the different submodules vary in size, then the output of fetches in sequential order might look like this: time --> output: |---A---| |-B-| |-------C-------| |-D-| |-E-| When we schedule these submodules into maximal two parallel processes, a schedule and sample output over time may look like this: process 1: |---A---| |-D-| |-E-| process 2: |-B-| |-------C-------| output: |---A---|B|---C-------|DE So A will be perceived as it would run normally in the single child version. As B has finished by the time A is done, we can dump its whole progress buffer on stderr, such that it looks like it finished in no time. Once that is done, C is determined to be the visible child and its progress will be reported in real time. So this way of output is really good for human consumption, as it only changes the timing, not the actual output. For machine consumption the output needs to be prepared in the tasks, by either having a prefix per line or per block to indicate whose tasks output is displayed, because the output order may not follow the original sequential ordering: |----A----| |--B--| |-C-| will be scheduled to be all parallel: process 1: |----A----| process 2: |--B--| process 3: |-C-| output: |----A----|CB This happens because C finished before B did, so it will be queued for output before B. To detect when a child has finished executing, we check interleaved with other actions (such as checking the liveliness of children or starting new processes) whether the stderr pipe still exists. Once a child closed its stderr stream, we assume it is terminating very soon, and use `finish_command()` from the single external process execution interface to collect the exit status. By maintaining the strong assumption of stderr being open until the very end of a child process, we can avoid other hassle such as an implementation using `waitpid(-1)`, which is not implemented in Windows. Signed-off-by: Stefan Beller <sbeller@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-12-16 08:04:10 +08:00
CALLOC_ARRAY(pp->children, n);
if (!opts->ungroup)
run-command: add an "ungroup" option to run_process_parallel() Extend the parallel execution API added in c553c72eed6 (run-command: add an asynchronous parallel child processor, 2015-12-15) to support a mode where the stdout and stderr of the processes isn't captured and output in a deterministic order, instead we'll leave it to the kernel and stdio to sort it out. This gives the API same functionality as GNU parallel's --ungroup option. As we'll see in a subsequent commit the main reason to want this is to support stdout and stderr being connected to the TTY in the case of jobs=1, demonstrated here with GNU parallel: $ parallel --ungroup 'test -t {} && echo TTY || echo NTTY' ::: 1 2 TTY TTY $ parallel 'test -t {} && echo TTY || echo NTTY' ::: 1 2 NTTY NTTY Another is as GNU parallel's documentation notes a potential for optimization. As demonstrated in next commit our results with "git hook run" will be similar, but generally speaking this shows that if you want to run processes in parallel where the exact order isn't important this can be a lot faster: $ hyperfine -r 3 -L o ,--ungroup 'parallel {o} seq ::: 10000000 >/dev/null ' Benchmark 1: parallel seq ::: 10000000 >/dev/null Time (mean ± σ): 220.2 ms ± 9.3 ms [User: 124.9 ms, System: 96.1 ms] Range (min … max): 212.3 ms … 230.5 ms 3 runs Benchmark 2: parallel --ungroup seq ::: 10000000 >/dev/null Time (mean ± σ): 154.7 ms ± 0.9 ms [User: 136.2 ms, System: 25.1 ms] Range (min … max): 153.9 ms … 155.7 ms 3 runs Summary 'parallel --ungroup seq ::: 10000000 >/dev/null ' ran 1.42 ± 0.06 times faster than 'parallel seq ::: 10000000 >/dev/null ' A large part of the juggling in the API is to make the API safer for its maintenance and consumers alike. For the maintenance of the API we e.g. avoid malloc()-ing the "pp->pfd", ensuring that SANITIZE=address and other similar tools will catch any unexpected misuse. For API consumers we take pains to never pass the non-NULL "out" buffer to an API user that provided the "ungroup" option. The resulting code in t/helper/test-run-command.c isn't typical of such a user, i.e. they'd typically use one mode or the other, and would know whether they'd provided "ungroup" or not. We could also avoid the strbuf_init() for "buffered_output" by having "struct parallel_processes" use a static PARALLEL_PROCESSES_INIT initializer, but let's leave that cleanup for later. Using a global "run_processes_parallel_ungroup" variable to enable this option is rather nasty, but is being done here to produce as minimal of a change as possible for a subsequent regression fix. This change is extracted from a larger initial version[1] which ends up with a better end-state for the API, but in doing so needed to modify all existing callers of the API. Let's defer that for now, and narrowly focus on what we need for fixing the regression in the subsequent commit. It's safe to do this with a global variable because: A) hook.c is the only user of it that sets it to non-zero, and before we'll get any other API users we'll refactor away this method of passing in the option, i.e. re-roll [1]. B) Even if hook.c wasn't the only user we don't have callers of this API that concurrently invoke this parallel process starting API itself in parallel. As noted above "A" && "B" are rather nasty, and we don't want to live with those caveats long-term, but for now they should be an acceptable compromise. 1. https://lore.kernel.org/git/cover-v2-0.8-00000000000-20220518T195858Z-avarab@gmail.com/ Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2022-06-07 16:48:19 +08:00
CALLOC_ARRAY(pp->pfd, n);
run-command: add an asynchronous parallel child processor This allows to run external commands in parallel with ordered output on stderr. If we run external commands in parallel we cannot pipe the output directly to the our stdout/err as it would mix up. So each process's output will flow through a pipe, which we buffer. One subprocess can be directly piped to out stdout/err for a low latency feedback to the user. Example: Let's assume we have 5 submodules A,B,C,D,E and each fetch takes a different amount of time as the different submodules vary in size, then the output of fetches in sequential order might look like this: time --> output: |---A---| |-B-| |-------C-------| |-D-| |-E-| When we schedule these submodules into maximal two parallel processes, a schedule and sample output over time may look like this: process 1: |---A---| |-D-| |-E-| process 2: |-B-| |-------C-------| output: |---A---|B|---C-------|DE So A will be perceived as it would run normally in the single child version. As B has finished by the time A is done, we can dump its whole progress buffer on stderr, such that it looks like it finished in no time. Once that is done, C is determined to be the visible child and its progress will be reported in real time. So this way of output is really good for human consumption, as it only changes the timing, not the actual output. For machine consumption the output needs to be prepared in the tasks, by either having a prefix per line or per block to indicate whose tasks output is displayed, because the output order may not follow the original sequential ordering: |----A----| |--B--| |-C-| will be scheduled to be all parallel: process 1: |----A----| process 2: |--B--| process 3: |-C-| output: |----A----|CB This happens because C finished before B did, so it will be queued for output before B. To detect when a child has finished executing, we check interleaved with other actions (such as checking the liveliness of children or starting new processes) whether the stderr pipe still exists. Once a child closed its stderr stream, we assume it is terminating very soon, and use `finish_command()` from the single external process execution interface to collect the exit status. By maintaining the strong assumption of stderr being open until the very end of a child process, we can avoid other hassle such as an implementation using `waitpid(-1)`, which is not implemented in Windows. Signed-off-by: Stefan Beller <sbeller@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-12-16 08:04:10 +08:00
run-command API: make "n" parameter a "size_t" Make the "n" variable added in c553c72eed6 (run-command: add an asynchronous parallel child processor, 2015-12-15) a "size_t". As we'll see in a subsequent commit we do pass "0" here, but never "jobs < 0". We could have made it an "unsigned int", but as we're having to change this let's not leave another case in the codebase where a size_t and "unsigned int" size differ on some platforms. In this case it's likely to never matter, but it's easier to not need to worry about it. After this and preceding changes: make run-command.o DEVOPTS=extra-all CFLAGS=-Wno-unused-parameter Only has one (and new) -Wsigned-compare warning relevant to a comparison about our "n" or "{nr,max}_processes": About using our "n" (size_t) in the same expression as online_cpus() (int). A subsequent commit will adjust & deal with online_cpus() and that warning. The only users of the "n" parameter are: * builtin/fetch.c: defaults to 1, reads from the "fetch.parallel" config. As seen in the code that parses the config added in d54dea77dba (fetch: let --jobs=<n> parallelize --multiple, too, 2019-10-05) will die if the git_config_int() return value is < 0. It will however pass us n = 0, as we'll see in a subsequent commit. * submodule.c: defaults to 1, reads from "submodule.fetchJobs" config. Read via code originally added in a028a1930c6 (fetching submodules: respect `submodule.fetchJobs` config option, 2016-02-29). It now piggy-backs on the the submodule.fetchJobs code and validation added in f20e7c1ea24 (submodule: remove submodule.fetchjobs from submodule-config parsing, 2017-08-02). Like builtin/fetch.c it will die if the git_config_int() return value is < 0, but like builtin/fetch.c it will pass us n = 0. * builtin/submodule--helper.c: defaults to 1. Read via code originally added in 2335b870fa7 (submodule update: expose parallelism to the user, 2016-02-29). Since f20e7c1ea24 (submodule: remove submodule.fetchjobs from submodule-config parsing, 2017-08-02) it shares a config parser and semantics with the submodule.c caller. * hook.c: hardcoded to 1, see 96e7225b310 (hook: add 'run' subcommand, 2021-12-22). * t/helper/test-run-command.c: can be -1 after parsing the arguments, but will then be overridden to online_cpus() before passing it to this API. See be5d88e1128 (test-tool run-command: learn to run (parts of) the testsuite, 2019-10-04). Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2022-10-13 05:02:23 +08:00
for (size_t i = 0; i < n; i++) {
run-command: add an asynchronous parallel child processor This allows to run external commands in parallel with ordered output on stderr. If we run external commands in parallel we cannot pipe the output directly to the our stdout/err as it would mix up. So each process's output will flow through a pipe, which we buffer. One subprocess can be directly piped to out stdout/err for a low latency feedback to the user. Example: Let's assume we have 5 submodules A,B,C,D,E and each fetch takes a different amount of time as the different submodules vary in size, then the output of fetches in sequential order might look like this: time --> output: |---A---| |-B-| |-------C-------| |-D-| |-E-| When we schedule these submodules into maximal two parallel processes, a schedule and sample output over time may look like this: process 1: |---A---| |-D-| |-E-| process 2: |-B-| |-------C-------| output: |---A---|B|---C-------|DE So A will be perceived as it would run normally in the single child version. As B has finished by the time A is done, we can dump its whole progress buffer on stderr, such that it looks like it finished in no time. Once that is done, C is determined to be the visible child and its progress will be reported in real time. So this way of output is really good for human consumption, as it only changes the timing, not the actual output. For machine consumption the output needs to be prepared in the tasks, by either having a prefix per line or per block to indicate whose tasks output is displayed, because the output order may not follow the original sequential ordering: |----A----| |--B--| |-C-| will be scheduled to be all parallel: process 1: |----A----| process 2: |--B--| process 3: |-C-| output: |----A----|CB This happens because C finished before B did, so it will be queued for output before B. To detect when a child has finished executing, we check interleaved with other actions (such as checking the liveliness of children or starting new processes) whether the stderr pipe still exists. Once a child closed its stderr stream, we assume it is terminating very soon, and use `finish_command()` from the single external process execution interface to collect the exit status. By maintaining the strong assumption of stderr being open until the very end of a child process, we can avoid other hassle such as an implementation using `waitpid(-1)`, which is not implemented in Windows. Signed-off-by: Stefan Beller <sbeller@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-12-16 08:04:10 +08:00
strbuf_init(&pp->children[i].err, 0);
child_process_init(&pp->children[i].process);
run-command: add an "ungroup" option to run_process_parallel() Extend the parallel execution API added in c553c72eed6 (run-command: add an asynchronous parallel child processor, 2015-12-15) to support a mode where the stdout and stderr of the processes isn't captured and output in a deterministic order, instead we'll leave it to the kernel and stdio to sort it out. This gives the API same functionality as GNU parallel's --ungroup option. As we'll see in a subsequent commit the main reason to want this is to support stdout and stderr being connected to the TTY in the case of jobs=1, demonstrated here with GNU parallel: $ parallel --ungroup 'test -t {} && echo TTY || echo NTTY' ::: 1 2 TTY TTY $ parallel 'test -t {} && echo TTY || echo NTTY' ::: 1 2 NTTY NTTY Another is as GNU parallel's documentation notes a potential for optimization. As demonstrated in next commit our results with "git hook run" will be similar, but generally speaking this shows that if you want to run processes in parallel where the exact order isn't important this can be a lot faster: $ hyperfine -r 3 -L o ,--ungroup 'parallel {o} seq ::: 10000000 >/dev/null ' Benchmark 1: parallel seq ::: 10000000 >/dev/null Time (mean ± σ): 220.2 ms ± 9.3 ms [User: 124.9 ms, System: 96.1 ms] Range (min … max): 212.3 ms … 230.5 ms 3 runs Benchmark 2: parallel --ungroup seq ::: 10000000 >/dev/null Time (mean ± σ): 154.7 ms ± 0.9 ms [User: 136.2 ms, System: 25.1 ms] Range (min … max): 153.9 ms … 155.7 ms 3 runs Summary 'parallel --ungroup seq ::: 10000000 >/dev/null ' ran 1.42 ± 0.06 times faster than 'parallel seq ::: 10000000 >/dev/null ' A large part of the juggling in the API is to make the API safer for its maintenance and consumers alike. For the maintenance of the API we e.g. avoid malloc()-ing the "pp->pfd", ensuring that SANITIZE=address and other similar tools will catch any unexpected misuse. For API consumers we take pains to never pass the non-NULL "out" buffer to an API user that provided the "ungroup" option. The resulting code in t/helper/test-run-command.c isn't typical of such a user, i.e. they'd typically use one mode or the other, and would know whether they'd provided "ungroup" or not. We could also avoid the strbuf_init() for "buffered_output" by having "struct parallel_processes" use a static PARALLEL_PROCESSES_INIT initializer, but let's leave that cleanup for later. Using a global "run_processes_parallel_ungroup" variable to enable this option is rather nasty, but is being done here to produce as minimal of a change as possible for a subsequent regression fix. This change is extracted from a larger initial version[1] which ends up with a better end-state for the API, but in doing so needed to modify all existing callers of the API. Let's defer that for now, and narrowly focus on what we need for fixing the regression in the subsequent commit. It's safe to do this with a global variable because: A) hook.c is the only user of it that sets it to non-zero, and before we'll get any other API users we'll refactor away this method of passing in the option, i.e. re-roll [1]. B) Even if hook.c wasn't the only user we don't have callers of this API that concurrently invoke this parallel process starting API itself in parallel. As noted above "A" && "B" are rather nasty, and we don't want to live with those caveats long-term, but for now they should be an acceptable compromise. 1. https://lore.kernel.org/git/cover-v2-0.8-00000000000-20220518T195858Z-avarab@gmail.com/ Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2022-06-07 16:48:19 +08:00
if (pp->pfd) {
pp->pfd[i].events = POLLIN | POLLHUP;
pp->pfd[i].fd = -1;
}
run-command: add an asynchronous parallel child processor This allows to run external commands in parallel with ordered output on stderr. If we run external commands in parallel we cannot pipe the output directly to the our stdout/err as it would mix up. So each process's output will flow through a pipe, which we buffer. One subprocess can be directly piped to out stdout/err for a low latency feedback to the user. Example: Let's assume we have 5 submodules A,B,C,D,E and each fetch takes a different amount of time as the different submodules vary in size, then the output of fetches in sequential order might look like this: time --> output: |---A---| |-B-| |-------C-------| |-D-| |-E-| When we schedule these submodules into maximal two parallel processes, a schedule and sample output over time may look like this: process 1: |---A---| |-D-| |-E-| process 2: |-B-| |-------C-------| output: |---A---|B|---C-------|DE So A will be perceived as it would run normally in the single child version. As B has finished by the time A is done, we can dump its whole progress buffer on stderr, such that it looks like it finished in no time. Once that is done, C is determined to be the visible child and its progress will be reported in real time. So this way of output is really good for human consumption, as it only changes the timing, not the actual output. For machine consumption the output needs to be prepared in the tasks, by either having a prefix per line or per block to indicate whose tasks output is displayed, because the output order may not follow the original sequential ordering: |----A----| |--B--| |-C-| will be scheduled to be all parallel: process 1: |----A----| process 2: |--B--| process 3: |-C-| output: |----A----|CB This happens because C finished before B did, so it will be queued for output before B. To detect when a child has finished executing, we check interleaved with other actions (such as checking the liveliness of children or starting new processes) whether the stderr pipe still exists. Once a child closed its stderr stream, we assume it is terminating very soon, and use `finish_command()` from the single external process execution interface to collect the exit status. By maintaining the strong assumption of stderr being open until the very end of a child process, we can avoid other hassle such as an implementation using `waitpid(-1)`, which is not implemented in Windows. Signed-off-by: Stefan Beller <sbeller@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-12-16 08:04:10 +08:00
}
pp_sig->pp = pp;
pp_sig->opts = opts;
pp_for_signal = pp_sig;
run-command: add an asynchronous parallel child processor This allows to run external commands in parallel with ordered output on stderr. If we run external commands in parallel we cannot pipe the output directly to the our stdout/err as it would mix up. So each process's output will flow through a pipe, which we buffer. One subprocess can be directly piped to out stdout/err for a low latency feedback to the user. Example: Let's assume we have 5 submodules A,B,C,D,E and each fetch takes a different amount of time as the different submodules vary in size, then the output of fetches in sequential order might look like this: time --> output: |---A---| |-B-| |-------C-------| |-D-| |-E-| When we schedule these submodules into maximal two parallel processes, a schedule and sample output over time may look like this: process 1: |---A---| |-D-| |-E-| process 2: |-B-| |-------C-------| output: |---A---|B|---C-------|DE So A will be perceived as it would run normally in the single child version. As B has finished by the time A is done, we can dump its whole progress buffer on stderr, such that it looks like it finished in no time. Once that is done, C is determined to be the visible child and its progress will be reported in real time. So this way of output is really good for human consumption, as it only changes the timing, not the actual output. For machine consumption the output needs to be prepared in the tasks, by either having a prefix per line or per block to indicate whose tasks output is displayed, because the output order may not follow the original sequential ordering: |----A----| |--B--| |-C-| will be scheduled to be all parallel: process 1: |----A----| process 2: |--B--| process 3: |-C-| output: |----A----|CB This happens because C finished before B did, so it will be queued for output before B. To detect when a child has finished executing, we check interleaved with other actions (such as checking the liveliness of children or starting new processes) whether the stderr pipe still exists. Once a child closed its stderr stream, we assume it is terminating very soon, and use `finish_command()` from the single external process execution interface to collect the exit status. By maintaining the strong assumption of stderr being open until the very end of a child process, we can avoid other hassle such as an implementation using `waitpid(-1)`, which is not implemented in Windows. Signed-off-by: Stefan Beller <sbeller@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-12-16 08:04:10 +08:00
sigchain_push_common(handle_children_on_signal);
}
static void pp_cleanup(struct parallel_processes *pp,
const struct run_process_parallel_opts *opts)
run-command: add an asynchronous parallel child processor This allows to run external commands in parallel with ordered output on stderr. If we run external commands in parallel we cannot pipe the output directly to the our stdout/err as it would mix up. So each process's output will flow through a pipe, which we buffer. One subprocess can be directly piped to out stdout/err for a low latency feedback to the user. Example: Let's assume we have 5 submodules A,B,C,D,E and each fetch takes a different amount of time as the different submodules vary in size, then the output of fetches in sequential order might look like this: time --> output: |---A---| |-B-| |-------C-------| |-D-| |-E-| When we schedule these submodules into maximal two parallel processes, a schedule and sample output over time may look like this: process 1: |---A---| |-D-| |-E-| process 2: |-B-| |-------C-------| output: |---A---|B|---C-------|DE So A will be perceived as it would run normally in the single child version. As B has finished by the time A is done, we can dump its whole progress buffer on stderr, such that it looks like it finished in no time. Once that is done, C is determined to be the visible child and its progress will be reported in real time. So this way of output is really good for human consumption, as it only changes the timing, not the actual output. For machine consumption the output needs to be prepared in the tasks, by either having a prefix per line or per block to indicate whose tasks output is displayed, because the output order may not follow the original sequential ordering: |----A----| |--B--| |-C-| will be scheduled to be all parallel: process 1: |----A----| process 2: |--B--| process 3: |-C-| output: |----A----|CB This happens because C finished before B did, so it will be queued for output before B. To detect when a child has finished executing, we check interleaved with other actions (such as checking the liveliness of children or starting new processes) whether the stderr pipe still exists. Once a child closed its stderr stream, we assume it is terminating very soon, and use `finish_command()` from the single external process execution interface to collect the exit status. By maintaining the strong assumption of stderr being open until the very end of a child process, we can avoid other hassle such as an implementation using `waitpid(-1)`, which is not implemented in Windows. Signed-off-by: Stefan Beller <sbeller@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-12-16 08:04:10 +08:00
{
trace_printf("run_processes_parallel: done");
for (size_t i = 0; i < opts->processes; i++) {
run-command: add an asynchronous parallel child processor This allows to run external commands in parallel with ordered output on stderr. If we run external commands in parallel we cannot pipe the output directly to the our stdout/err as it would mix up. So each process's output will flow through a pipe, which we buffer. One subprocess can be directly piped to out stdout/err for a low latency feedback to the user. Example: Let's assume we have 5 submodules A,B,C,D,E and each fetch takes a different amount of time as the different submodules vary in size, then the output of fetches in sequential order might look like this: time --> output: |---A---| |-B-| |-------C-------| |-D-| |-E-| When we schedule these submodules into maximal two parallel processes, a schedule and sample output over time may look like this: process 1: |---A---| |-D-| |-E-| process 2: |-B-| |-------C-------| output: |---A---|B|---C-------|DE So A will be perceived as it would run normally in the single child version. As B has finished by the time A is done, we can dump its whole progress buffer on stderr, such that it looks like it finished in no time. Once that is done, C is determined to be the visible child and its progress will be reported in real time. So this way of output is really good for human consumption, as it only changes the timing, not the actual output. For machine consumption the output needs to be prepared in the tasks, by either having a prefix per line or per block to indicate whose tasks output is displayed, because the output order may not follow the original sequential ordering: |----A----| |--B--| |-C-| will be scheduled to be all parallel: process 1: |----A----| process 2: |--B--| process 3: |-C-| output: |----A----|CB This happens because C finished before B did, so it will be queued for output before B. To detect when a child has finished executing, we check interleaved with other actions (such as checking the liveliness of children or starting new processes) whether the stderr pipe still exists. Once a child closed its stderr stream, we assume it is terminating very soon, and use `finish_command()` from the single external process execution interface to collect the exit status. By maintaining the strong assumption of stderr being open until the very end of a child process, we can avoid other hassle such as an implementation using `waitpid(-1)`, which is not implemented in Windows. Signed-off-by: Stefan Beller <sbeller@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-12-16 08:04:10 +08:00
strbuf_release(&pp->children[i].err);
child_process_clear(&pp->children[i].process);
}
free(pp->children);
free(pp->pfd);
/*
* When get_next_task added messages to the buffer in its last
* iteration, the buffered output is non empty.
*/
strbuf_write(&pp->buffered_output, stderr);
run-command: add an asynchronous parallel child processor This allows to run external commands in parallel with ordered output on stderr. If we run external commands in parallel we cannot pipe the output directly to the our stdout/err as it would mix up. So each process's output will flow through a pipe, which we buffer. One subprocess can be directly piped to out stdout/err for a low latency feedback to the user. Example: Let's assume we have 5 submodules A,B,C,D,E and each fetch takes a different amount of time as the different submodules vary in size, then the output of fetches in sequential order might look like this: time --> output: |---A---| |-B-| |-------C-------| |-D-| |-E-| When we schedule these submodules into maximal two parallel processes, a schedule and sample output over time may look like this: process 1: |---A---| |-D-| |-E-| process 2: |-B-| |-------C-------| output: |---A---|B|---C-------|DE So A will be perceived as it would run normally in the single child version. As B has finished by the time A is done, we can dump its whole progress buffer on stderr, such that it looks like it finished in no time. Once that is done, C is determined to be the visible child and its progress will be reported in real time. So this way of output is really good for human consumption, as it only changes the timing, not the actual output. For machine consumption the output needs to be prepared in the tasks, by either having a prefix per line or per block to indicate whose tasks output is displayed, because the output order may not follow the original sequential ordering: |----A----| |--B--| |-C-| will be scheduled to be all parallel: process 1: |----A----| process 2: |--B--| process 3: |-C-| output: |----A----|CB This happens because C finished before B did, so it will be queued for output before B. To detect when a child has finished executing, we check interleaved with other actions (such as checking the liveliness of children or starting new processes) whether the stderr pipe still exists. Once a child closed its stderr stream, we assume it is terminating very soon, and use `finish_command()` from the single external process execution interface to collect the exit status. By maintaining the strong assumption of stderr being open until the very end of a child process, we can avoid other hassle such as an implementation using `waitpid(-1)`, which is not implemented in Windows. Signed-off-by: Stefan Beller <sbeller@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-12-16 08:04:10 +08:00
strbuf_release(&pp->buffered_output);
sigchain_pop_common();
}
/* returns
* 0 if a new task was started.
* 1 if no new jobs was started (get_next_task ran out of work, non critical
* problem with starting a new command)
* <0 no new job was started, user wishes to shutdown early. Use negative code
* to signal the children.
*/
static int pp_start_one(struct parallel_processes *pp,
const struct run_process_parallel_opts *opts)
run-command: add an asynchronous parallel child processor This allows to run external commands in parallel with ordered output on stderr. If we run external commands in parallel we cannot pipe the output directly to the our stdout/err as it would mix up. So each process's output will flow through a pipe, which we buffer. One subprocess can be directly piped to out stdout/err for a low latency feedback to the user. Example: Let's assume we have 5 submodules A,B,C,D,E and each fetch takes a different amount of time as the different submodules vary in size, then the output of fetches in sequential order might look like this: time --> output: |---A---| |-B-| |-------C-------| |-D-| |-E-| When we schedule these submodules into maximal two parallel processes, a schedule and sample output over time may look like this: process 1: |---A---| |-D-| |-E-| process 2: |-B-| |-------C-------| output: |---A---|B|---C-------|DE So A will be perceived as it would run normally in the single child version. As B has finished by the time A is done, we can dump its whole progress buffer on stderr, such that it looks like it finished in no time. Once that is done, C is determined to be the visible child and its progress will be reported in real time. So this way of output is really good for human consumption, as it only changes the timing, not the actual output. For machine consumption the output needs to be prepared in the tasks, by either having a prefix per line or per block to indicate whose tasks output is displayed, because the output order may not follow the original sequential ordering: |----A----| |--B--| |-C-| will be scheduled to be all parallel: process 1: |----A----| process 2: |--B--| process 3: |-C-| output: |----A----|CB This happens because C finished before B did, so it will be queued for output before B. To detect when a child has finished executing, we check interleaved with other actions (such as checking the liveliness of children or starting new processes) whether the stderr pipe still exists. Once a child closed its stderr stream, we assume it is terminating very soon, and use `finish_command()` from the single external process execution interface to collect the exit status. By maintaining the strong assumption of stderr being open until the very end of a child process, we can avoid other hassle such as an implementation using `waitpid(-1)`, which is not implemented in Windows. Signed-off-by: Stefan Beller <sbeller@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-12-16 08:04:10 +08:00
{
run-command API: make "n" parameter a "size_t" Make the "n" variable added in c553c72eed6 (run-command: add an asynchronous parallel child processor, 2015-12-15) a "size_t". As we'll see in a subsequent commit we do pass "0" here, but never "jobs < 0". We could have made it an "unsigned int", but as we're having to change this let's not leave another case in the codebase where a size_t and "unsigned int" size differ on some platforms. In this case it's likely to never matter, but it's easier to not need to worry about it. After this and preceding changes: make run-command.o DEVOPTS=extra-all CFLAGS=-Wno-unused-parameter Only has one (and new) -Wsigned-compare warning relevant to a comparison about our "n" or "{nr,max}_processes": About using our "n" (size_t) in the same expression as online_cpus() (int). A subsequent commit will adjust & deal with online_cpus() and that warning. The only users of the "n" parameter are: * builtin/fetch.c: defaults to 1, reads from the "fetch.parallel" config. As seen in the code that parses the config added in d54dea77dba (fetch: let --jobs=<n> parallelize --multiple, too, 2019-10-05) will die if the git_config_int() return value is < 0. It will however pass us n = 0, as we'll see in a subsequent commit. * submodule.c: defaults to 1, reads from "submodule.fetchJobs" config. Read via code originally added in a028a1930c6 (fetching submodules: respect `submodule.fetchJobs` config option, 2016-02-29). It now piggy-backs on the the submodule.fetchJobs code and validation added in f20e7c1ea24 (submodule: remove submodule.fetchjobs from submodule-config parsing, 2017-08-02). Like builtin/fetch.c it will die if the git_config_int() return value is < 0, but like builtin/fetch.c it will pass us n = 0. * builtin/submodule--helper.c: defaults to 1. Read via code originally added in 2335b870fa7 (submodule update: expose parallelism to the user, 2016-02-29). Since f20e7c1ea24 (submodule: remove submodule.fetchjobs from submodule-config parsing, 2017-08-02) it shares a config parser and semantics with the submodule.c caller. * hook.c: hardcoded to 1, see 96e7225b310 (hook: add 'run' subcommand, 2021-12-22). * t/helper/test-run-command.c: can be -1 after parsing the arguments, but will then be overridden to online_cpus() before passing it to this API. See be5d88e1128 (test-tool run-command: learn to run (parts of) the testsuite, 2019-10-04). Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2022-10-13 05:02:23 +08:00
size_t i;
int code;
run-command: add an asynchronous parallel child processor This allows to run external commands in parallel with ordered output on stderr. If we run external commands in parallel we cannot pipe the output directly to the our stdout/err as it would mix up. So each process's output will flow through a pipe, which we buffer. One subprocess can be directly piped to out stdout/err for a low latency feedback to the user. Example: Let's assume we have 5 submodules A,B,C,D,E and each fetch takes a different amount of time as the different submodules vary in size, then the output of fetches in sequential order might look like this: time --> output: |---A---| |-B-| |-------C-------| |-D-| |-E-| When we schedule these submodules into maximal two parallel processes, a schedule and sample output over time may look like this: process 1: |---A---| |-D-| |-E-| process 2: |-B-| |-------C-------| output: |---A---|B|---C-------|DE So A will be perceived as it would run normally in the single child version. As B has finished by the time A is done, we can dump its whole progress buffer on stderr, such that it looks like it finished in no time. Once that is done, C is determined to be the visible child and its progress will be reported in real time. So this way of output is really good for human consumption, as it only changes the timing, not the actual output. For machine consumption the output needs to be prepared in the tasks, by either having a prefix per line or per block to indicate whose tasks output is displayed, because the output order may not follow the original sequential ordering: |----A----| |--B--| |-C-| will be scheduled to be all parallel: process 1: |----A----| process 2: |--B--| process 3: |-C-| output: |----A----|CB This happens because C finished before B did, so it will be queued for output before B. To detect when a child has finished executing, we check interleaved with other actions (such as checking the liveliness of children or starting new processes) whether the stderr pipe still exists. Once a child closed its stderr stream, we assume it is terminating very soon, and use `finish_command()` from the single external process execution interface to collect the exit status. By maintaining the strong assumption of stderr being open until the very end of a child process, we can avoid other hassle such as an implementation using `waitpid(-1)`, which is not implemented in Windows. Signed-off-by: Stefan Beller <sbeller@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-12-16 08:04:10 +08:00
for (i = 0; i < opts->processes; i++)
run-command: add an asynchronous parallel child processor This allows to run external commands in parallel with ordered output on stderr. If we run external commands in parallel we cannot pipe the output directly to the our stdout/err as it would mix up. So each process's output will flow through a pipe, which we buffer. One subprocess can be directly piped to out stdout/err for a low latency feedback to the user. Example: Let's assume we have 5 submodules A,B,C,D,E and each fetch takes a different amount of time as the different submodules vary in size, then the output of fetches in sequential order might look like this: time --> output: |---A---| |-B-| |-------C-------| |-D-| |-E-| When we schedule these submodules into maximal two parallel processes, a schedule and sample output over time may look like this: process 1: |---A---| |-D-| |-E-| process 2: |-B-| |-------C-------| output: |---A---|B|---C-------|DE So A will be perceived as it would run normally in the single child version. As B has finished by the time A is done, we can dump its whole progress buffer on stderr, such that it looks like it finished in no time. Once that is done, C is determined to be the visible child and its progress will be reported in real time. So this way of output is really good for human consumption, as it only changes the timing, not the actual output. For machine consumption the output needs to be prepared in the tasks, by either having a prefix per line or per block to indicate whose tasks output is displayed, because the output order may not follow the original sequential ordering: |----A----| |--B--| |-C-| will be scheduled to be all parallel: process 1: |----A----| process 2: |--B--| process 3: |-C-| output: |----A----|CB This happens because C finished before B did, so it will be queued for output before B. To detect when a child has finished executing, we check interleaved with other actions (such as checking the liveliness of children or starting new processes) whether the stderr pipe still exists. Once a child closed its stderr stream, we assume it is terminating very soon, and use `finish_command()` from the single external process execution interface to collect the exit status. By maintaining the strong assumption of stderr being open until the very end of a child process, we can avoid other hassle such as an implementation using `waitpid(-1)`, which is not implemented in Windows. Signed-off-by: Stefan Beller <sbeller@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-12-16 08:04:10 +08:00
if (pp->children[i].state == GIT_CP_FREE)
break;
if (i == opts->processes)
BUG("bookkeeping is hard");
run-command: add an asynchronous parallel child processor This allows to run external commands in parallel with ordered output on stderr. If we run external commands in parallel we cannot pipe the output directly to the our stdout/err as it would mix up. So each process's output will flow through a pipe, which we buffer. One subprocess can be directly piped to out stdout/err for a low latency feedback to the user. Example: Let's assume we have 5 submodules A,B,C,D,E and each fetch takes a different amount of time as the different submodules vary in size, then the output of fetches in sequential order might look like this: time --> output: |---A---| |-B-| |-------C-------| |-D-| |-E-| When we schedule these submodules into maximal two parallel processes, a schedule and sample output over time may look like this: process 1: |---A---| |-D-| |-E-| process 2: |-B-| |-------C-------| output: |---A---|B|---C-------|DE So A will be perceived as it would run normally in the single child version. As B has finished by the time A is done, we can dump its whole progress buffer on stderr, such that it looks like it finished in no time. Once that is done, C is determined to be the visible child and its progress will be reported in real time. So this way of output is really good for human consumption, as it only changes the timing, not the actual output. For machine consumption the output needs to be prepared in the tasks, by either having a prefix per line or per block to indicate whose tasks output is displayed, because the output order may not follow the original sequential ordering: |----A----| |--B--| |-C-| will be scheduled to be all parallel: process 1: |----A----| process 2: |--B--| process 3: |-C-| output: |----A----|CB This happens because C finished before B did, so it will be queued for output before B. To detect when a child has finished executing, we check interleaved with other actions (such as checking the liveliness of children or starting new processes) whether the stderr pipe still exists. Once a child closed its stderr stream, we assume it is terminating very soon, and use `finish_command()` from the single external process execution interface to collect the exit status. By maintaining the strong assumption of stderr being open until the very end of a child process, we can avoid other hassle such as an implementation using `waitpid(-1)`, which is not implemented in Windows. Signed-off-by: Stefan Beller <sbeller@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-12-16 08:04:10 +08:00
/*
* By default, do not inherit stdin from the parent process - otherwise,
* all children would share stdin! Users may overwrite this to provide
* something to the child's stdin by having their 'get_next_task'
* callback assign 0 to .no_stdin and an appropriate integer to .in.
*/
pp->children[i].process.no_stdin = 1;
code = opts->get_next_task(&pp->children[i].process,
opts->ungroup ? NULL : &pp->children[i].err,
opts->data,
&pp->children[i].data);
run-command: add an asynchronous parallel child processor This allows to run external commands in parallel with ordered output on stderr. If we run external commands in parallel we cannot pipe the output directly to the our stdout/err as it would mix up. So each process's output will flow through a pipe, which we buffer. One subprocess can be directly piped to out stdout/err for a low latency feedback to the user. Example: Let's assume we have 5 submodules A,B,C,D,E and each fetch takes a different amount of time as the different submodules vary in size, then the output of fetches in sequential order might look like this: time --> output: |---A---| |-B-| |-------C-------| |-D-| |-E-| When we schedule these submodules into maximal two parallel processes, a schedule and sample output over time may look like this: process 1: |---A---| |-D-| |-E-| process 2: |-B-| |-------C-------| output: |---A---|B|---C-------|DE So A will be perceived as it would run normally in the single child version. As B has finished by the time A is done, we can dump its whole progress buffer on stderr, such that it looks like it finished in no time. Once that is done, C is determined to be the visible child and its progress will be reported in real time. So this way of output is really good for human consumption, as it only changes the timing, not the actual output. For machine consumption the output needs to be prepared in the tasks, by either having a prefix per line or per block to indicate whose tasks output is displayed, because the output order may not follow the original sequential ordering: |----A----| |--B--| |-C-| will be scheduled to be all parallel: process 1: |----A----| process 2: |--B--| process 3: |-C-| output: |----A----|CB This happens because C finished before B did, so it will be queued for output before B. To detect when a child has finished executing, we check interleaved with other actions (such as checking the liveliness of children or starting new processes) whether the stderr pipe still exists. Once a child closed its stderr stream, we assume it is terminating very soon, and use `finish_command()` from the single external process execution interface to collect the exit status. By maintaining the strong assumption of stderr being open until the very end of a child process, we can avoid other hassle such as an implementation using `waitpid(-1)`, which is not implemented in Windows. Signed-off-by: Stefan Beller <sbeller@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-12-16 08:04:10 +08:00
if (!code) {
if (!opts->ungroup) {
run-command: add an "ungroup" option to run_process_parallel() Extend the parallel execution API added in c553c72eed6 (run-command: add an asynchronous parallel child processor, 2015-12-15) to support a mode where the stdout and stderr of the processes isn't captured and output in a deterministic order, instead we'll leave it to the kernel and stdio to sort it out. This gives the API same functionality as GNU parallel's --ungroup option. As we'll see in a subsequent commit the main reason to want this is to support stdout and stderr being connected to the TTY in the case of jobs=1, demonstrated here with GNU parallel: $ parallel --ungroup 'test -t {} && echo TTY || echo NTTY' ::: 1 2 TTY TTY $ parallel 'test -t {} && echo TTY || echo NTTY' ::: 1 2 NTTY NTTY Another is as GNU parallel's documentation notes a potential for optimization. As demonstrated in next commit our results with "git hook run" will be similar, but generally speaking this shows that if you want to run processes in parallel where the exact order isn't important this can be a lot faster: $ hyperfine -r 3 -L o ,--ungroup 'parallel {o} seq ::: 10000000 >/dev/null ' Benchmark 1: parallel seq ::: 10000000 >/dev/null Time (mean ± σ): 220.2 ms ± 9.3 ms [User: 124.9 ms, System: 96.1 ms] Range (min … max): 212.3 ms … 230.5 ms 3 runs Benchmark 2: parallel --ungroup seq ::: 10000000 >/dev/null Time (mean ± σ): 154.7 ms ± 0.9 ms [User: 136.2 ms, System: 25.1 ms] Range (min … max): 153.9 ms … 155.7 ms 3 runs Summary 'parallel --ungroup seq ::: 10000000 >/dev/null ' ran 1.42 ± 0.06 times faster than 'parallel seq ::: 10000000 >/dev/null ' A large part of the juggling in the API is to make the API safer for its maintenance and consumers alike. For the maintenance of the API we e.g. avoid malloc()-ing the "pp->pfd", ensuring that SANITIZE=address and other similar tools will catch any unexpected misuse. For API consumers we take pains to never pass the non-NULL "out" buffer to an API user that provided the "ungroup" option. The resulting code in t/helper/test-run-command.c isn't typical of such a user, i.e. they'd typically use one mode or the other, and would know whether they'd provided "ungroup" or not. We could also avoid the strbuf_init() for "buffered_output" by having "struct parallel_processes" use a static PARALLEL_PROCESSES_INIT initializer, but let's leave that cleanup for later. Using a global "run_processes_parallel_ungroup" variable to enable this option is rather nasty, but is being done here to produce as minimal of a change as possible for a subsequent regression fix. This change is extracted from a larger initial version[1] which ends up with a better end-state for the API, but in doing so needed to modify all existing callers of the API. Let's defer that for now, and narrowly focus on what we need for fixing the regression in the subsequent commit. It's safe to do this with a global variable because: A) hook.c is the only user of it that sets it to non-zero, and before we'll get any other API users we'll refactor away this method of passing in the option, i.e. re-roll [1]. B) Even if hook.c wasn't the only user we don't have callers of this API that concurrently invoke this parallel process starting API itself in parallel. As noted above "A" && "B" are rather nasty, and we don't want to live with those caveats long-term, but for now they should be an acceptable compromise. 1. https://lore.kernel.org/git/cover-v2-0.8-00000000000-20220518T195858Z-avarab@gmail.com/ Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2022-06-07 16:48:19 +08:00
strbuf_addbuf(&pp->buffered_output, &pp->children[i].err);
strbuf_reset(&pp->children[i].err);
}
run-command: add an asynchronous parallel child processor This allows to run external commands in parallel with ordered output on stderr. If we run external commands in parallel we cannot pipe the output directly to the our stdout/err as it would mix up. So each process's output will flow through a pipe, which we buffer. One subprocess can be directly piped to out stdout/err for a low latency feedback to the user. Example: Let's assume we have 5 submodules A,B,C,D,E and each fetch takes a different amount of time as the different submodules vary in size, then the output of fetches in sequential order might look like this: time --> output: |---A---| |-B-| |-------C-------| |-D-| |-E-| When we schedule these submodules into maximal two parallel processes, a schedule and sample output over time may look like this: process 1: |---A---| |-D-| |-E-| process 2: |-B-| |-------C-------| output: |---A---|B|---C-------|DE So A will be perceived as it would run normally in the single child version. As B has finished by the time A is done, we can dump its whole progress buffer on stderr, such that it looks like it finished in no time. Once that is done, C is determined to be the visible child and its progress will be reported in real time. So this way of output is really good for human consumption, as it only changes the timing, not the actual output. For machine consumption the output needs to be prepared in the tasks, by either having a prefix per line or per block to indicate whose tasks output is displayed, because the output order may not follow the original sequential ordering: |----A----| |--B--| |-C-| will be scheduled to be all parallel: process 1: |----A----| process 2: |--B--| process 3: |-C-| output: |----A----|CB This happens because C finished before B did, so it will be queued for output before B. To detect when a child has finished executing, we check interleaved with other actions (such as checking the liveliness of children or starting new processes) whether the stderr pipe still exists. Once a child closed its stderr stream, we assume it is terminating very soon, and use `finish_command()` from the single external process execution interface to collect the exit status. By maintaining the strong assumption of stderr being open until the very end of a child process, we can avoid other hassle such as an implementation using `waitpid(-1)`, which is not implemented in Windows. Signed-off-by: Stefan Beller <sbeller@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-12-16 08:04:10 +08:00
return 1;
}
if (!opts->ungroup) {
run-command: add an "ungroup" option to run_process_parallel() Extend the parallel execution API added in c553c72eed6 (run-command: add an asynchronous parallel child processor, 2015-12-15) to support a mode where the stdout and stderr of the processes isn't captured and output in a deterministic order, instead we'll leave it to the kernel and stdio to sort it out. This gives the API same functionality as GNU parallel's --ungroup option. As we'll see in a subsequent commit the main reason to want this is to support stdout and stderr being connected to the TTY in the case of jobs=1, demonstrated here with GNU parallel: $ parallel --ungroup 'test -t {} && echo TTY || echo NTTY' ::: 1 2 TTY TTY $ parallel 'test -t {} && echo TTY || echo NTTY' ::: 1 2 NTTY NTTY Another is as GNU parallel's documentation notes a potential for optimization. As demonstrated in next commit our results with "git hook run" will be similar, but generally speaking this shows that if you want to run processes in parallel where the exact order isn't important this can be a lot faster: $ hyperfine -r 3 -L o ,--ungroup 'parallel {o} seq ::: 10000000 >/dev/null ' Benchmark 1: parallel seq ::: 10000000 >/dev/null Time (mean ± σ): 220.2 ms ± 9.3 ms [User: 124.9 ms, System: 96.1 ms] Range (min … max): 212.3 ms … 230.5 ms 3 runs Benchmark 2: parallel --ungroup seq ::: 10000000 >/dev/null Time (mean ± σ): 154.7 ms ± 0.9 ms [User: 136.2 ms, System: 25.1 ms] Range (min … max): 153.9 ms … 155.7 ms 3 runs Summary 'parallel --ungroup seq ::: 10000000 >/dev/null ' ran 1.42 ± 0.06 times faster than 'parallel seq ::: 10000000 >/dev/null ' A large part of the juggling in the API is to make the API safer for its maintenance and consumers alike. For the maintenance of the API we e.g. avoid malloc()-ing the "pp->pfd", ensuring that SANITIZE=address and other similar tools will catch any unexpected misuse. For API consumers we take pains to never pass the non-NULL "out" buffer to an API user that provided the "ungroup" option. The resulting code in t/helper/test-run-command.c isn't typical of such a user, i.e. they'd typically use one mode or the other, and would know whether they'd provided "ungroup" or not. We could also avoid the strbuf_init() for "buffered_output" by having "struct parallel_processes" use a static PARALLEL_PROCESSES_INIT initializer, but let's leave that cleanup for later. Using a global "run_processes_parallel_ungroup" variable to enable this option is rather nasty, but is being done here to produce as minimal of a change as possible for a subsequent regression fix. This change is extracted from a larger initial version[1] which ends up with a better end-state for the API, but in doing so needed to modify all existing callers of the API. Let's defer that for now, and narrowly focus on what we need for fixing the regression in the subsequent commit. It's safe to do this with a global variable because: A) hook.c is the only user of it that sets it to non-zero, and before we'll get any other API users we'll refactor away this method of passing in the option, i.e. re-roll [1]. B) Even if hook.c wasn't the only user we don't have callers of this API that concurrently invoke this parallel process starting API itself in parallel. As noted above "A" && "B" are rather nasty, and we don't want to live with those caveats long-term, but for now they should be an acceptable compromise. 1. https://lore.kernel.org/git/cover-v2-0.8-00000000000-20220518T195858Z-avarab@gmail.com/ Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2022-06-07 16:48:19 +08:00
pp->children[i].process.err = -1;
pp->children[i].process.stdout_to_stderr = 1;
}
run-command: add an asynchronous parallel child processor This allows to run external commands in parallel with ordered output on stderr. If we run external commands in parallel we cannot pipe the output directly to the our stdout/err as it would mix up. So each process's output will flow through a pipe, which we buffer. One subprocess can be directly piped to out stdout/err for a low latency feedback to the user. Example: Let's assume we have 5 submodules A,B,C,D,E and each fetch takes a different amount of time as the different submodules vary in size, then the output of fetches in sequential order might look like this: time --> output: |---A---| |-B-| |-------C-------| |-D-| |-E-| When we schedule these submodules into maximal two parallel processes, a schedule and sample output over time may look like this: process 1: |---A---| |-D-| |-E-| process 2: |-B-| |-------C-------| output: |---A---|B|---C-------|DE So A will be perceived as it would run normally in the single child version. As B has finished by the time A is done, we can dump its whole progress buffer on stderr, such that it looks like it finished in no time. Once that is done, C is determined to be the visible child and its progress will be reported in real time. So this way of output is really good for human consumption, as it only changes the timing, not the actual output. For machine consumption the output needs to be prepared in the tasks, by either having a prefix per line or per block to indicate whose tasks output is displayed, because the output order may not follow the original sequential ordering: |----A----| |--B--| |-C-| will be scheduled to be all parallel: process 1: |----A----| process 2: |--B--| process 3: |-C-| output: |----A----|CB This happens because C finished before B did, so it will be queued for output before B. To detect when a child has finished executing, we check interleaved with other actions (such as checking the liveliness of children or starting new processes) whether the stderr pipe still exists. Once a child closed its stderr stream, we assume it is terminating very soon, and use `finish_command()` from the single external process execution interface to collect the exit status. By maintaining the strong assumption of stderr being open until the very end of a child process, we can avoid other hassle such as an implementation using `waitpid(-1)`, which is not implemented in Windows. Signed-off-by: Stefan Beller <sbeller@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-12-16 08:04:10 +08:00
if (start_command(&pp->children[i].process)) {
if (opts->start_failure)
code = opts->start_failure(opts->ungroup ? NULL :
&pp->children[i].err,
opts->data,
pp->children[i].data);
else
code = 0;
if (!opts->ungroup) {
run-command: add an "ungroup" option to run_process_parallel() Extend the parallel execution API added in c553c72eed6 (run-command: add an asynchronous parallel child processor, 2015-12-15) to support a mode where the stdout and stderr of the processes isn't captured and output in a deterministic order, instead we'll leave it to the kernel and stdio to sort it out. This gives the API same functionality as GNU parallel's --ungroup option. As we'll see in a subsequent commit the main reason to want this is to support stdout and stderr being connected to the TTY in the case of jobs=1, demonstrated here with GNU parallel: $ parallel --ungroup 'test -t {} && echo TTY || echo NTTY' ::: 1 2 TTY TTY $ parallel 'test -t {} && echo TTY || echo NTTY' ::: 1 2 NTTY NTTY Another is as GNU parallel's documentation notes a potential for optimization. As demonstrated in next commit our results with "git hook run" will be similar, but generally speaking this shows that if you want to run processes in parallel where the exact order isn't important this can be a lot faster: $ hyperfine -r 3 -L o ,--ungroup 'parallel {o} seq ::: 10000000 >/dev/null ' Benchmark 1: parallel seq ::: 10000000 >/dev/null Time (mean ± σ): 220.2 ms ± 9.3 ms [User: 124.9 ms, System: 96.1 ms] Range (min … max): 212.3 ms … 230.5 ms 3 runs Benchmark 2: parallel --ungroup seq ::: 10000000 >/dev/null Time (mean ± σ): 154.7 ms ± 0.9 ms [User: 136.2 ms, System: 25.1 ms] Range (min … max): 153.9 ms … 155.7 ms 3 runs Summary 'parallel --ungroup seq ::: 10000000 >/dev/null ' ran 1.42 ± 0.06 times faster than 'parallel seq ::: 10000000 >/dev/null ' A large part of the juggling in the API is to make the API safer for its maintenance and consumers alike. For the maintenance of the API we e.g. avoid malloc()-ing the "pp->pfd", ensuring that SANITIZE=address and other similar tools will catch any unexpected misuse. For API consumers we take pains to never pass the non-NULL "out" buffer to an API user that provided the "ungroup" option. The resulting code in t/helper/test-run-command.c isn't typical of such a user, i.e. they'd typically use one mode or the other, and would know whether they'd provided "ungroup" or not. We could also avoid the strbuf_init() for "buffered_output" by having "struct parallel_processes" use a static PARALLEL_PROCESSES_INIT initializer, but let's leave that cleanup for later. Using a global "run_processes_parallel_ungroup" variable to enable this option is rather nasty, but is being done here to produce as minimal of a change as possible for a subsequent regression fix. This change is extracted from a larger initial version[1] which ends up with a better end-state for the API, but in doing so needed to modify all existing callers of the API. Let's defer that for now, and narrowly focus on what we need for fixing the regression in the subsequent commit. It's safe to do this with a global variable because: A) hook.c is the only user of it that sets it to non-zero, and before we'll get any other API users we'll refactor away this method of passing in the option, i.e. re-roll [1]. B) Even if hook.c wasn't the only user we don't have callers of this API that concurrently invoke this parallel process starting API itself in parallel. As noted above "A" && "B" are rather nasty, and we don't want to live with those caveats long-term, but for now they should be an acceptable compromise. 1. https://lore.kernel.org/git/cover-v2-0.8-00000000000-20220518T195858Z-avarab@gmail.com/ Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2022-06-07 16:48:19 +08:00
strbuf_addbuf(&pp->buffered_output, &pp->children[i].err);
strbuf_reset(&pp->children[i].err);
}
run-command: add an asynchronous parallel child processor This allows to run external commands in parallel with ordered output on stderr. If we run external commands in parallel we cannot pipe the output directly to the our stdout/err as it would mix up. So each process's output will flow through a pipe, which we buffer. One subprocess can be directly piped to out stdout/err for a low latency feedback to the user. Example: Let's assume we have 5 submodules A,B,C,D,E and each fetch takes a different amount of time as the different submodules vary in size, then the output of fetches in sequential order might look like this: time --> output: |---A---| |-B-| |-------C-------| |-D-| |-E-| When we schedule these submodules into maximal two parallel processes, a schedule and sample output over time may look like this: process 1: |---A---| |-D-| |-E-| process 2: |-B-| |-------C-------| output: |---A---|B|---C-------|DE So A will be perceived as it would run normally in the single child version. As B has finished by the time A is done, we can dump its whole progress buffer on stderr, such that it looks like it finished in no time. Once that is done, C is determined to be the visible child and its progress will be reported in real time. So this way of output is really good for human consumption, as it only changes the timing, not the actual output. For machine consumption the output needs to be prepared in the tasks, by either having a prefix per line or per block to indicate whose tasks output is displayed, because the output order may not follow the original sequential ordering: |----A----| |--B--| |-C-| will be scheduled to be all parallel: process 1: |----A----| process 2: |--B--| process 3: |-C-| output: |----A----|CB This happens because C finished before B did, so it will be queued for output before B. To detect when a child has finished executing, we check interleaved with other actions (such as checking the liveliness of children or starting new processes) whether the stderr pipe still exists. Once a child closed its stderr stream, we assume it is terminating very soon, and use `finish_command()` from the single external process execution interface to collect the exit status. By maintaining the strong assumption of stderr being open until the very end of a child process, we can avoid other hassle such as an implementation using `waitpid(-1)`, which is not implemented in Windows. Signed-off-by: Stefan Beller <sbeller@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-12-16 08:04:10 +08:00
if (code)
pp->shutdown = 1;
return code;
}
pp->nr_processes++;
pp->children[i].state = GIT_CP_WORKING;
run-command: add an "ungroup" option to run_process_parallel() Extend the parallel execution API added in c553c72eed6 (run-command: add an asynchronous parallel child processor, 2015-12-15) to support a mode where the stdout and stderr of the processes isn't captured and output in a deterministic order, instead we'll leave it to the kernel and stdio to sort it out. This gives the API same functionality as GNU parallel's --ungroup option. As we'll see in a subsequent commit the main reason to want this is to support stdout and stderr being connected to the TTY in the case of jobs=1, demonstrated here with GNU parallel: $ parallel --ungroup 'test -t {} && echo TTY || echo NTTY' ::: 1 2 TTY TTY $ parallel 'test -t {} && echo TTY || echo NTTY' ::: 1 2 NTTY NTTY Another is as GNU parallel's documentation notes a potential for optimization. As demonstrated in next commit our results with "git hook run" will be similar, but generally speaking this shows that if you want to run processes in parallel where the exact order isn't important this can be a lot faster: $ hyperfine -r 3 -L o ,--ungroup 'parallel {o} seq ::: 10000000 >/dev/null ' Benchmark 1: parallel seq ::: 10000000 >/dev/null Time (mean ± σ): 220.2 ms ± 9.3 ms [User: 124.9 ms, System: 96.1 ms] Range (min … max): 212.3 ms … 230.5 ms 3 runs Benchmark 2: parallel --ungroup seq ::: 10000000 >/dev/null Time (mean ± σ): 154.7 ms ± 0.9 ms [User: 136.2 ms, System: 25.1 ms] Range (min … max): 153.9 ms … 155.7 ms 3 runs Summary 'parallel --ungroup seq ::: 10000000 >/dev/null ' ran 1.42 ± 0.06 times faster than 'parallel seq ::: 10000000 >/dev/null ' A large part of the juggling in the API is to make the API safer for its maintenance and consumers alike. For the maintenance of the API we e.g. avoid malloc()-ing the "pp->pfd", ensuring that SANITIZE=address and other similar tools will catch any unexpected misuse. For API consumers we take pains to never pass the non-NULL "out" buffer to an API user that provided the "ungroup" option. The resulting code in t/helper/test-run-command.c isn't typical of such a user, i.e. they'd typically use one mode or the other, and would know whether they'd provided "ungroup" or not. We could also avoid the strbuf_init() for "buffered_output" by having "struct parallel_processes" use a static PARALLEL_PROCESSES_INIT initializer, but let's leave that cleanup for later. Using a global "run_processes_parallel_ungroup" variable to enable this option is rather nasty, but is being done here to produce as minimal of a change as possible for a subsequent regression fix. This change is extracted from a larger initial version[1] which ends up with a better end-state for the API, but in doing so needed to modify all existing callers of the API. Let's defer that for now, and narrowly focus on what we need for fixing the regression in the subsequent commit. It's safe to do this with a global variable because: A) hook.c is the only user of it that sets it to non-zero, and before we'll get any other API users we'll refactor away this method of passing in the option, i.e. re-roll [1]. B) Even if hook.c wasn't the only user we don't have callers of this API that concurrently invoke this parallel process starting API itself in parallel. As noted above "A" && "B" are rather nasty, and we don't want to live with those caveats long-term, but for now they should be an acceptable compromise. 1. https://lore.kernel.org/git/cover-v2-0.8-00000000000-20220518T195858Z-avarab@gmail.com/ Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2022-06-07 16:48:19 +08:00
if (pp->pfd)
pp->pfd[i].fd = pp->children[i].process.err;
run-command: add an asynchronous parallel child processor This allows to run external commands in parallel with ordered output on stderr. If we run external commands in parallel we cannot pipe the output directly to the our stdout/err as it would mix up. So each process's output will flow through a pipe, which we buffer. One subprocess can be directly piped to out stdout/err for a low latency feedback to the user. Example: Let's assume we have 5 submodules A,B,C,D,E and each fetch takes a different amount of time as the different submodules vary in size, then the output of fetches in sequential order might look like this: time --> output: |---A---| |-B-| |-------C-------| |-D-| |-E-| When we schedule these submodules into maximal two parallel processes, a schedule and sample output over time may look like this: process 1: |---A---| |-D-| |-E-| process 2: |-B-| |-------C-------| output: |---A---|B|---C-------|DE So A will be perceived as it would run normally in the single child version. As B has finished by the time A is done, we can dump its whole progress buffer on stderr, such that it looks like it finished in no time. Once that is done, C is determined to be the visible child and its progress will be reported in real time. So this way of output is really good for human consumption, as it only changes the timing, not the actual output. For machine consumption the output needs to be prepared in the tasks, by either having a prefix per line or per block to indicate whose tasks output is displayed, because the output order may not follow the original sequential ordering: |----A----| |--B--| |-C-| will be scheduled to be all parallel: process 1: |----A----| process 2: |--B--| process 3: |-C-| output: |----A----|CB This happens because C finished before B did, so it will be queued for output before B. To detect when a child has finished executing, we check interleaved with other actions (such as checking the liveliness of children or starting new processes) whether the stderr pipe still exists. Once a child closed its stderr stream, we assume it is terminating very soon, and use `finish_command()` from the single external process execution interface to collect the exit status. By maintaining the strong assumption of stderr being open until the very end of a child process, we can avoid other hassle such as an implementation using `waitpid(-1)`, which is not implemented in Windows. Signed-off-by: Stefan Beller <sbeller@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-12-16 08:04:10 +08:00
return 0;
}
static void pp_buffer_stderr(struct parallel_processes *pp,
const struct run_process_parallel_opts *opts,
int output_timeout)
run-command: add an asynchronous parallel child processor This allows to run external commands in parallel with ordered output on stderr. If we run external commands in parallel we cannot pipe the output directly to the our stdout/err as it would mix up. So each process's output will flow through a pipe, which we buffer. One subprocess can be directly piped to out stdout/err for a low latency feedback to the user. Example: Let's assume we have 5 submodules A,B,C,D,E and each fetch takes a different amount of time as the different submodules vary in size, then the output of fetches in sequential order might look like this: time --> output: |---A---| |-B-| |-------C-------| |-D-| |-E-| When we schedule these submodules into maximal two parallel processes, a schedule and sample output over time may look like this: process 1: |---A---| |-D-| |-E-| process 2: |-B-| |-------C-------| output: |---A---|B|---C-------|DE So A will be perceived as it would run normally in the single child version. As B has finished by the time A is done, we can dump its whole progress buffer on stderr, such that it looks like it finished in no time. Once that is done, C is determined to be the visible child and its progress will be reported in real time. So this way of output is really good for human consumption, as it only changes the timing, not the actual output. For machine consumption the output needs to be prepared in the tasks, by either having a prefix per line or per block to indicate whose tasks output is displayed, because the output order may not follow the original sequential ordering: |----A----| |--B--| |-C-| will be scheduled to be all parallel: process 1: |----A----| process 2: |--B--| process 3: |-C-| output: |----A----|CB This happens because C finished before B did, so it will be queued for output before B. To detect when a child has finished executing, we check interleaved with other actions (such as checking the liveliness of children or starting new processes) whether the stderr pipe still exists. Once a child closed its stderr stream, we assume it is terminating very soon, and use `finish_command()` from the single external process execution interface to collect the exit status. By maintaining the strong assumption of stderr being open until the very end of a child process, we can avoid other hassle such as an implementation using `waitpid(-1)`, which is not implemented in Windows. Signed-off-by: Stefan Beller <sbeller@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-12-16 08:04:10 +08:00
{
while (poll(pp->pfd, opts->processes, output_timeout) < 0) {
run-command: add an asynchronous parallel child processor This allows to run external commands in parallel with ordered output on stderr. If we run external commands in parallel we cannot pipe the output directly to the our stdout/err as it would mix up. So each process's output will flow through a pipe, which we buffer. One subprocess can be directly piped to out stdout/err for a low latency feedback to the user. Example: Let's assume we have 5 submodules A,B,C,D,E and each fetch takes a different amount of time as the different submodules vary in size, then the output of fetches in sequential order might look like this: time --> output: |---A---| |-B-| |-------C-------| |-D-| |-E-| When we schedule these submodules into maximal two parallel processes, a schedule and sample output over time may look like this: process 1: |---A---| |-D-| |-E-| process 2: |-B-| |-------C-------| output: |---A---|B|---C-------|DE So A will be perceived as it would run normally in the single child version. As B has finished by the time A is done, we can dump its whole progress buffer on stderr, such that it looks like it finished in no time. Once that is done, C is determined to be the visible child and its progress will be reported in real time. So this way of output is really good for human consumption, as it only changes the timing, not the actual output. For machine consumption the output needs to be prepared in the tasks, by either having a prefix per line or per block to indicate whose tasks output is displayed, because the output order may not follow the original sequential ordering: |----A----| |--B--| |-C-| will be scheduled to be all parallel: process 1: |----A----| process 2: |--B--| process 3: |-C-| output: |----A----|CB This happens because C finished before B did, so it will be queued for output before B. To detect when a child has finished executing, we check interleaved with other actions (such as checking the liveliness of children or starting new processes) whether the stderr pipe still exists. Once a child closed its stderr stream, we assume it is terminating very soon, and use `finish_command()` from the single external process execution interface to collect the exit status. By maintaining the strong assumption of stderr being open until the very end of a child process, we can avoid other hassle such as an implementation using `waitpid(-1)`, which is not implemented in Windows. Signed-off-by: Stefan Beller <sbeller@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-12-16 08:04:10 +08:00
if (errno == EINTR)
continue;
pp_cleanup(pp, opts);
run-command: add an asynchronous parallel child processor This allows to run external commands in parallel with ordered output on stderr. If we run external commands in parallel we cannot pipe the output directly to the our stdout/err as it would mix up. So each process's output will flow through a pipe, which we buffer. One subprocess can be directly piped to out stdout/err for a low latency feedback to the user. Example: Let's assume we have 5 submodules A,B,C,D,E and each fetch takes a different amount of time as the different submodules vary in size, then the output of fetches in sequential order might look like this: time --> output: |---A---| |-B-| |-------C-------| |-D-| |-E-| When we schedule these submodules into maximal two parallel processes, a schedule and sample output over time may look like this: process 1: |---A---| |-D-| |-E-| process 2: |-B-| |-------C-------| output: |---A---|B|---C-------|DE So A will be perceived as it would run normally in the single child version. As B has finished by the time A is done, we can dump its whole progress buffer on stderr, such that it looks like it finished in no time. Once that is done, C is determined to be the visible child and its progress will be reported in real time. So this way of output is really good for human consumption, as it only changes the timing, not the actual output. For machine consumption the output needs to be prepared in the tasks, by either having a prefix per line or per block to indicate whose tasks output is displayed, because the output order may not follow the original sequential ordering: |----A----| |--B--| |-C-| will be scheduled to be all parallel: process 1: |----A----| process 2: |--B--| process 3: |-C-| output: |----A----|CB This happens because C finished before B did, so it will be queued for output before B. To detect when a child has finished executing, we check interleaved with other actions (such as checking the liveliness of children or starting new processes) whether the stderr pipe still exists. Once a child closed its stderr stream, we assume it is terminating very soon, and use `finish_command()` from the single external process execution interface to collect the exit status. By maintaining the strong assumption of stderr being open until the very end of a child process, we can avoid other hassle such as an implementation using `waitpid(-1)`, which is not implemented in Windows. Signed-off-by: Stefan Beller <sbeller@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-12-16 08:04:10 +08:00
die_errno("poll");
}
/* Buffer output from all pipes. */
for (size_t i = 0; i < opts->processes; i++) {
run-command: add an asynchronous parallel child processor This allows to run external commands in parallel with ordered output on stderr. If we run external commands in parallel we cannot pipe the output directly to the our stdout/err as it would mix up. So each process's output will flow through a pipe, which we buffer. One subprocess can be directly piped to out stdout/err for a low latency feedback to the user. Example: Let's assume we have 5 submodules A,B,C,D,E and each fetch takes a different amount of time as the different submodules vary in size, then the output of fetches in sequential order might look like this: time --> output: |---A---| |-B-| |-------C-------| |-D-| |-E-| When we schedule these submodules into maximal two parallel processes, a schedule and sample output over time may look like this: process 1: |---A---| |-D-| |-E-| process 2: |-B-| |-------C-------| output: |---A---|B|---C-------|DE So A will be perceived as it would run normally in the single child version. As B has finished by the time A is done, we can dump its whole progress buffer on stderr, such that it looks like it finished in no time. Once that is done, C is determined to be the visible child and its progress will be reported in real time. So this way of output is really good for human consumption, as it only changes the timing, not the actual output. For machine consumption the output needs to be prepared in the tasks, by either having a prefix per line or per block to indicate whose tasks output is displayed, because the output order may not follow the original sequential ordering: |----A----| |--B--| |-C-| will be scheduled to be all parallel: process 1: |----A----| process 2: |--B--| process 3: |-C-| output: |----A----|CB This happens because C finished before B did, so it will be queued for output before B. To detect when a child has finished executing, we check interleaved with other actions (such as checking the liveliness of children or starting new processes) whether the stderr pipe still exists. Once a child closed its stderr stream, we assume it is terminating very soon, and use `finish_command()` from the single external process execution interface to collect the exit status. By maintaining the strong assumption of stderr being open until the very end of a child process, we can avoid other hassle such as an implementation using `waitpid(-1)`, which is not implemented in Windows. Signed-off-by: Stefan Beller <sbeller@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-12-16 08:04:10 +08:00
if (pp->children[i].state == GIT_CP_WORKING &&
pp->pfd[i].revents & (POLLIN | POLLHUP)) {
int n = strbuf_read_once(&pp->children[i].err,
pp->children[i].process.err, 0);
if (n == 0) {
close(pp->children[i].process.err);
pp->children[i].state = GIT_CP_WAIT_CLEANUP;
} else if (n < 0)
if (errno != EAGAIN)
die_errno("read");
}
}
}
static void pp_output(const struct parallel_processes *pp)
run-command: add an asynchronous parallel child processor This allows to run external commands in parallel with ordered output on stderr. If we run external commands in parallel we cannot pipe the output directly to the our stdout/err as it would mix up. So each process's output will flow through a pipe, which we buffer. One subprocess can be directly piped to out stdout/err for a low latency feedback to the user. Example: Let's assume we have 5 submodules A,B,C,D,E and each fetch takes a different amount of time as the different submodules vary in size, then the output of fetches in sequential order might look like this: time --> output: |---A---| |-B-| |-------C-------| |-D-| |-E-| When we schedule these submodules into maximal two parallel processes, a schedule and sample output over time may look like this: process 1: |---A---| |-D-| |-E-| process 2: |-B-| |-------C-------| output: |---A---|B|---C-------|DE So A will be perceived as it would run normally in the single child version. As B has finished by the time A is done, we can dump its whole progress buffer on stderr, such that it looks like it finished in no time. Once that is done, C is determined to be the visible child and its progress will be reported in real time. So this way of output is really good for human consumption, as it only changes the timing, not the actual output. For machine consumption the output needs to be prepared in the tasks, by either having a prefix per line or per block to indicate whose tasks output is displayed, because the output order may not follow the original sequential ordering: |----A----| |--B--| |-C-| will be scheduled to be all parallel: process 1: |----A----| process 2: |--B--| process 3: |-C-| output: |----A----|CB This happens because C finished before B did, so it will be queued for output before B. To detect when a child has finished executing, we check interleaved with other actions (such as checking the liveliness of children or starting new processes) whether the stderr pipe still exists. Once a child closed its stderr stream, we assume it is terminating very soon, and use `finish_command()` from the single external process execution interface to collect the exit status. By maintaining the strong assumption of stderr being open until the very end of a child process, we can avoid other hassle such as an implementation using `waitpid(-1)`, which is not implemented in Windows. Signed-off-by: Stefan Beller <sbeller@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-12-16 08:04:10 +08:00
{
run-command API: make "n" parameter a "size_t" Make the "n" variable added in c553c72eed6 (run-command: add an asynchronous parallel child processor, 2015-12-15) a "size_t". As we'll see in a subsequent commit we do pass "0" here, but never "jobs < 0". We could have made it an "unsigned int", but as we're having to change this let's not leave another case in the codebase where a size_t and "unsigned int" size differ on some platforms. In this case it's likely to never matter, but it's easier to not need to worry about it. After this and preceding changes: make run-command.o DEVOPTS=extra-all CFLAGS=-Wno-unused-parameter Only has one (and new) -Wsigned-compare warning relevant to a comparison about our "n" or "{nr,max}_processes": About using our "n" (size_t) in the same expression as online_cpus() (int). A subsequent commit will adjust & deal with online_cpus() and that warning. The only users of the "n" parameter are: * builtin/fetch.c: defaults to 1, reads from the "fetch.parallel" config. As seen in the code that parses the config added in d54dea77dba (fetch: let --jobs=<n> parallelize --multiple, too, 2019-10-05) will die if the git_config_int() return value is < 0. It will however pass us n = 0, as we'll see in a subsequent commit. * submodule.c: defaults to 1, reads from "submodule.fetchJobs" config. Read via code originally added in a028a1930c6 (fetching submodules: respect `submodule.fetchJobs` config option, 2016-02-29). It now piggy-backs on the the submodule.fetchJobs code and validation added in f20e7c1ea24 (submodule: remove submodule.fetchjobs from submodule-config parsing, 2017-08-02). Like builtin/fetch.c it will die if the git_config_int() return value is < 0, but like builtin/fetch.c it will pass us n = 0. * builtin/submodule--helper.c: defaults to 1. Read via code originally added in 2335b870fa7 (submodule update: expose parallelism to the user, 2016-02-29). Since f20e7c1ea24 (submodule: remove submodule.fetchjobs from submodule-config parsing, 2017-08-02) it shares a config parser and semantics with the submodule.c caller. * hook.c: hardcoded to 1, see 96e7225b310 (hook: add 'run' subcommand, 2021-12-22). * t/helper/test-run-command.c: can be -1 after parsing the arguments, but will then be overridden to online_cpus() before passing it to this API. See be5d88e1128 (test-tool run-command: learn to run (parts of) the testsuite, 2019-10-04). Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2022-10-13 05:02:23 +08:00
size_t i = pp->output_owner;
run-command: add an "ungroup" option to run_process_parallel() Extend the parallel execution API added in c553c72eed6 (run-command: add an asynchronous parallel child processor, 2015-12-15) to support a mode where the stdout and stderr of the processes isn't captured and output in a deterministic order, instead we'll leave it to the kernel and stdio to sort it out. This gives the API same functionality as GNU parallel's --ungroup option. As we'll see in a subsequent commit the main reason to want this is to support stdout and stderr being connected to the TTY in the case of jobs=1, demonstrated here with GNU parallel: $ parallel --ungroup 'test -t {} && echo TTY || echo NTTY' ::: 1 2 TTY TTY $ parallel 'test -t {} && echo TTY || echo NTTY' ::: 1 2 NTTY NTTY Another is as GNU parallel's documentation notes a potential for optimization. As demonstrated in next commit our results with "git hook run" will be similar, but generally speaking this shows that if you want to run processes in parallel where the exact order isn't important this can be a lot faster: $ hyperfine -r 3 -L o ,--ungroup 'parallel {o} seq ::: 10000000 >/dev/null ' Benchmark 1: parallel seq ::: 10000000 >/dev/null Time (mean ± σ): 220.2 ms ± 9.3 ms [User: 124.9 ms, System: 96.1 ms] Range (min … max): 212.3 ms … 230.5 ms 3 runs Benchmark 2: parallel --ungroup seq ::: 10000000 >/dev/null Time (mean ± σ): 154.7 ms ± 0.9 ms [User: 136.2 ms, System: 25.1 ms] Range (min … max): 153.9 ms … 155.7 ms 3 runs Summary 'parallel --ungroup seq ::: 10000000 >/dev/null ' ran 1.42 ± 0.06 times faster than 'parallel seq ::: 10000000 >/dev/null ' A large part of the juggling in the API is to make the API safer for its maintenance and consumers alike. For the maintenance of the API we e.g. avoid malloc()-ing the "pp->pfd", ensuring that SANITIZE=address and other similar tools will catch any unexpected misuse. For API consumers we take pains to never pass the non-NULL "out" buffer to an API user that provided the "ungroup" option. The resulting code in t/helper/test-run-command.c isn't typical of such a user, i.e. they'd typically use one mode or the other, and would know whether they'd provided "ungroup" or not. We could also avoid the strbuf_init() for "buffered_output" by having "struct parallel_processes" use a static PARALLEL_PROCESSES_INIT initializer, but let's leave that cleanup for later. Using a global "run_processes_parallel_ungroup" variable to enable this option is rather nasty, but is being done here to produce as minimal of a change as possible for a subsequent regression fix. This change is extracted from a larger initial version[1] which ends up with a better end-state for the API, but in doing so needed to modify all existing callers of the API. Let's defer that for now, and narrowly focus on what we need for fixing the regression in the subsequent commit. It's safe to do this with a global variable because: A) hook.c is the only user of it that sets it to non-zero, and before we'll get any other API users we'll refactor away this method of passing in the option, i.e. re-roll [1]. B) Even if hook.c wasn't the only user we don't have callers of this API that concurrently invoke this parallel process starting API itself in parallel. As noted above "A" && "B" are rather nasty, and we don't want to live with those caveats long-term, but for now they should be an acceptable compromise. 1. https://lore.kernel.org/git/cover-v2-0.8-00000000000-20220518T195858Z-avarab@gmail.com/ Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2022-06-07 16:48:19 +08:00
run-command: add an asynchronous parallel child processor This allows to run external commands in parallel with ordered output on stderr. If we run external commands in parallel we cannot pipe the output directly to the our stdout/err as it would mix up. So each process's output will flow through a pipe, which we buffer. One subprocess can be directly piped to out stdout/err for a low latency feedback to the user. Example: Let's assume we have 5 submodules A,B,C,D,E and each fetch takes a different amount of time as the different submodules vary in size, then the output of fetches in sequential order might look like this: time --> output: |---A---| |-B-| |-------C-------| |-D-| |-E-| When we schedule these submodules into maximal two parallel processes, a schedule and sample output over time may look like this: process 1: |---A---| |-D-| |-E-| process 2: |-B-| |-------C-------| output: |---A---|B|---C-------|DE So A will be perceived as it would run normally in the single child version. As B has finished by the time A is done, we can dump its whole progress buffer on stderr, such that it looks like it finished in no time. Once that is done, C is determined to be the visible child and its progress will be reported in real time. So this way of output is really good for human consumption, as it only changes the timing, not the actual output. For machine consumption the output needs to be prepared in the tasks, by either having a prefix per line or per block to indicate whose tasks output is displayed, because the output order may not follow the original sequential ordering: |----A----| |--B--| |-C-| will be scheduled to be all parallel: process 1: |----A----| process 2: |--B--| process 3: |-C-| output: |----A----|CB This happens because C finished before B did, so it will be queued for output before B. To detect when a child has finished executing, we check interleaved with other actions (such as checking the liveliness of children or starting new processes) whether the stderr pipe still exists. Once a child closed its stderr stream, we assume it is terminating very soon, and use `finish_command()` from the single external process execution interface to collect the exit status. By maintaining the strong assumption of stderr being open until the very end of a child process, we can avoid other hassle such as an implementation using `waitpid(-1)`, which is not implemented in Windows. Signed-off-by: Stefan Beller <sbeller@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-12-16 08:04:10 +08:00
if (pp->children[i].state == GIT_CP_WORKING &&
pp->children[i].err.len) {
strbuf_write(&pp->children[i].err, stderr);
run-command: add an asynchronous parallel child processor This allows to run external commands in parallel with ordered output on stderr. If we run external commands in parallel we cannot pipe the output directly to the our stdout/err as it would mix up. So each process's output will flow through a pipe, which we buffer. One subprocess can be directly piped to out stdout/err for a low latency feedback to the user. Example: Let's assume we have 5 submodules A,B,C,D,E and each fetch takes a different amount of time as the different submodules vary in size, then the output of fetches in sequential order might look like this: time --> output: |---A---| |-B-| |-------C-------| |-D-| |-E-| When we schedule these submodules into maximal two parallel processes, a schedule and sample output over time may look like this: process 1: |---A---| |-D-| |-E-| process 2: |-B-| |-------C-------| output: |---A---|B|---C-------|DE So A will be perceived as it would run normally in the single child version. As B has finished by the time A is done, we can dump its whole progress buffer on stderr, such that it looks like it finished in no time. Once that is done, C is determined to be the visible child and its progress will be reported in real time. So this way of output is really good for human consumption, as it only changes the timing, not the actual output. For machine consumption the output needs to be prepared in the tasks, by either having a prefix per line or per block to indicate whose tasks output is displayed, because the output order may not follow the original sequential ordering: |----A----| |--B--| |-C-| will be scheduled to be all parallel: process 1: |----A----| process 2: |--B--| process 3: |-C-| output: |----A----|CB This happens because C finished before B did, so it will be queued for output before B. To detect when a child has finished executing, we check interleaved with other actions (such as checking the liveliness of children or starting new processes) whether the stderr pipe still exists. Once a child closed its stderr stream, we assume it is terminating very soon, and use `finish_command()` from the single external process execution interface to collect the exit status. By maintaining the strong assumption of stderr being open until the very end of a child process, we can avoid other hassle such as an implementation using `waitpid(-1)`, which is not implemented in Windows. Signed-off-by: Stefan Beller <sbeller@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-12-16 08:04:10 +08:00
strbuf_reset(&pp->children[i].err);
}
}
static int pp_collect_finished(struct parallel_processes *pp,
const struct run_process_parallel_opts *opts)
run-command: add an asynchronous parallel child processor This allows to run external commands in parallel with ordered output on stderr. If we run external commands in parallel we cannot pipe the output directly to the our stdout/err as it would mix up. So each process's output will flow through a pipe, which we buffer. One subprocess can be directly piped to out stdout/err for a low latency feedback to the user. Example: Let's assume we have 5 submodules A,B,C,D,E and each fetch takes a different amount of time as the different submodules vary in size, then the output of fetches in sequential order might look like this: time --> output: |---A---| |-B-| |-------C-------| |-D-| |-E-| When we schedule these submodules into maximal two parallel processes, a schedule and sample output over time may look like this: process 1: |---A---| |-D-| |-E-| process 2: |-B-| |-------C-------| output: |---A---|B|---C-------|DE So A will be perceived as it would run normally in the single child version. As B has finished by the time A is done, we can dump its whole progress buffer on stderr, such that it looks like it finished in no time. Once that is done, C is determined to be the visible child and its progress will be reported in real time. So this way of output is really good for human consumption, as it only changes the timing, not the actual output. For machine consumption the output needs to be prepared in the tasks, by either having a prefix per line or per block to indicate whose tasks output is displayed, because the output order may not follow the original sequential ordering: |----A----| |--B--| |-C-| will be scheduled to be all parallel: process 1: |----A----| process 2: |--B--| process 3: |-C-| output: |----A----|CB This happens because C finished before B did, so it will be queued for output before B. To detect when a child has finished executing, we check interleaved with other actions (such as checking the liveliness of children or starting new processes) whether the stderr pipe still exists. Once a child closed its stderr stream, we assume it is terminating very soon, and use `finish_command()` from the single external process execution interface to collect the exit status. By maintaining the strong assumption of stderr being open until the very end of a child process, we can avoid other hassle such as an implementation using `waitpid(-1)`, which is not implemented in Windows. Signed-off-by: Stefan Beller <sbeller@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-12-16 08:04:10 +08:00
{
run-command API: make "n" parameter a "size_t" Make the "n" variable added in c553c72eed6 (run-command: add an asynchronous parallel child processor, 2015-12-15) a "size_t". As we'll see in a subsequent commit we do pass "0" here, but never "jobs < 0". We could have made it an "unsigned int", but as we're having to change this let's not leave another case in the codebase where a size_t and "unsigned int" size differ on some platforms. In this case it's likely to never matter, but it's easier to not need to worry about it. After this and preceding changes: make run-command.o DEVOPTS=extra-all CFLAGS=-Wno-unused-parameter Only has one (and new) -Wsigned-compare warning relevant to a comparison about our "n" or "{nr,max}_processes": About using our "n" (size_t) in the same expression as online_cpus() (int). A subsequent commit will adjust & deal with online_cpus() and that warning. The only users of the "n" parameter are: * builtin/fetch.c: defaults to 1, reads from the "fetch.parallel" config. As seen in the code that parses the config added in d54dea77dba (fetch: let --jobs=<n> parallelize --multiple, too, 2019-10-05) will die if the git_config_int() return value is < 0. It will however pass us n = 0, as we'll see in a subsequent commit. * submodule.c: defaults to 1, reads from "submodule.fetchJobs" config. Read via code originally added in a028a1930c6 (fetching submodules: respect `submodule.fetchJobs` config option, 2016-02-29). It now piggy-backs on the the submodule.fetchJobs code and validation added in f20e7c1ea24 (submodule: remove submodule.fetchjobs from submodule-config parsing, 2017-08-02). Like builtin/fetch.c it will die if the git_config_int() return value is < 0, but like builtin/fetch.c it will pass us n = 0. * builtin/submodule--helper.c: defaults to 1. Read via code originally added in 2335b870fa7 (submodule update: expose parallelism to the user, 2016-02-29). Since f20e7c1ea24 (submodule: remove submodule.fetchjobs from submodule-config parsing, 2017-08-02) it shares a config parser and semantics with the submodule.c caller. * hook.c: hardcoded to 1, see 96e7225b310 (hook: add 'run' subcommand, 2021-12-22). * t/helper/test-run-command.c: can be -1 after parsing the arguments, but will then be overridden to online_cpus() before passing it to this API. See be5d88e1128 (test-tool run-command: learn to run (parts of) the testsuite, 2019-10-04). Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2022-10-13 05:02:23 +08:00
int code;
size_t i;
run-command: add an asynchronous parallel child processor This allows to run external commands in parallel with ordered output on stderr. If we run external commands in parallel we cannot pipe the output directly to the our stdout/err as it would mix up. So each process's output will flow through a pipe, which we buffer. One subprocess can be directly piped to out stdout/err for a low latency feedback to the user. Example: Let's assume we have 5 submodules A,B,C,D,E and each fetch takes a different amount of time as the different submodules vary in size, then the output of fetches in sequential order might look like this: time --> output: |---A---| |-B-| |-------C-------| |-D-| |-E-| When we schedule these submodules into maximal two parallel processes, a schedule and sample output over time may look like this: process 1: |---A---| |-D-| |-E-| process 2: |-B-| |-------C-------| output: |---A---|B|---C-------|DE So A will be perceived as it would run normally in the single child version. As B has finished by the time A is done, we can dump its whole progress buffer on stderr, such that it looks like it finished in no time. Once that is done, C is determined to be the visible child and its progress will be reported in real time. So this way of output is really good for human consumption, as it only changes the timing, not the actual output. For machine consumption the output needs to be prepared in the tasks, by either having a prefix per line or per block to indicate whose tasks output is displayed, because the output order may not follow the original sequential ordering: |----A----| |--B--| |-C-| will be scheduled to be all parallel: process 1: |----A----| process 2: |--B--| process 3: |-C-| output: |----A----|CB This happens because C finished before B did, so it will be queued for output before B. To detect when a child has finished executing, we check interleaved with other actions (such as checking the liveliness of children or starting new processes) whether the stderr pipe still exists. Once a child closed its stderr stream, we assume it is terminating very soon, and use `finish_command()` from the single external process execution interface to collect the exit status. By maintaining the strong assumption of stderr being open until the very end of a child process, we can avoid other hassle such as an implementation using `waitpid(-1)`, which is not implemented in Windows. Signed-off-by: Stefan Beller <sbeller@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-12-16 08:04:10 +08:00
int result = 0;
while (pp->nr_processes > 0) {
for (i = 0; i < opts->processes; i++)
run-command: add an asynchronous parallel child processor This allows to run external commands in parallel with ordered output on stderr. If we run external commands in parallel we cannot pipe the output directly to the our stdout/err as it would mix up. So each process's output will flow through a pipe, which we buffer. One subprocess can be directly piped to out stdout/err for a low latency feedback to the user. Example: Let's assume we have 5 submodules A,B,C,D,E and each fetch takes a different amount of time as the different submodules vary in size, then the output of fetches in sequential order might look like this: time --> output: |---A---| |-B-| |-------C-------| |-D-| |-E-| When we schedule these submodules into maximal two parallel processes, a schedule and sample output over time may look like this: process 1: |---A---| |-D-| |-E-| process 2: |-B-| |-------C-------| output: |---A---|B|---C-------|DE So A will be perceived as it would run normally in the single child version. As B has finished by the time A is done, we can dump its whole progress buffer on stderr, such that it looks like it finished in no time. Once that is done, C is determined to be the visible child and its progress will be reported in real time. So this way of output is really good for human consumption, as it only changes the timing, not the actual output. For machine consumption the output needs to be prepared in the tasks, by either having a prefix per line or per block to indicate whose tasks output is displayed, because the output order may not follow the original sequential ordering: |----A----| |--B--| |-C-| will be scheduled to be all parallel: process 1: |----A----| process 2: |--B--| process 3: |-C-| output: |----A----|CB This happens because C finished before B did, so it will be queued for output before B. To detect when a child has finished executing, we check interleaved with other actions (such as checking the liveliness of children or starting new processes) whether the stderr pipe still exists. Once a child closed its stderr stream, we assume it is terminating very soon, and use `finish_command()` from the single external process execution interface to collect the exit status. By maintaining the strong assumption of stderr being open until the very end of a child process, we can avoid other hassle such as an implementation using `waitpid(-1)`, which is not implemented in Windows. Signed-off-by: Stefan Beller <sbeller@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-12-16 08:04:10 +08:00
if (pp->children[i].state == GIT_CP_WAIT_CLEANUP)
break;
if (i == opts->processes)
run-command: add an asynchronous parallel child processor This allows to run external commands in parallel with ordered output on stderr. If we run external commands in parallel we cannot pipe the output directly to the our stdout/err as it would mix up. So each process's output will flow through a pipe, which we buffer. One subprocess can be directly piped to out stdout/err for a low latency feedback to the user. Example: Let's assume we have 5 submodules A,B,C,D,E and each fetch takes a different amount of time as the different submodules vary in size, then the output of fetches in sequential order might look like this: time --> output: |---A---| |-B-| |-------C-------| |-D-| |-E-| When we schedule these submodules into maximal two parallel processes, a schedule and sample output over time may look like this: process 1: |---A---| |-D-| |-E-| process 2: |-B-| |-------C-------| output: |---A---|B|---C-------|DE So A will be perceived as it would run normally in the single child version. As B has finished by the time A is done, we can dump its whole progress buffer on stderr, such that it looks like it finished in no time. Once that is done, C is determined to be the visible child and its progress will be reported in real time. So this way of output is really good for human consumption, as it only changes the timing, not the actual output. For machine consumption the output needs to be prepared in the tasks, by either having a prefix per line or per block to indicate whose tasks output is displayed, because the output order may not follow the original sequential ordering: |----A----| |--B--| |-C-| will be scheduled to be all parallel: process 1: |----A----| process 2: |--B--| process 3: |-C-| output: |----A----|CB This happens because C finished before B did, so it will be queued for output before B. To detect when a child has finished executing, we check interleaved with other actions (such as checking the liveliness of children or starting new processes) whether the stderr pipe still exists. Once a child closed its stderr stream, we assume it is terminating very soon, and use `finish_command()` from the single external process execution interface to collect the exit status. By maintaining the strong assumption of stderr being open until the very end of a child process, we can avoid other hassle such as an implementation using `waitpid(-1)`, which is not implemented in Windows. Signed-off-by: Stefan Beller <sbeller@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-12-16 08:04:10 +08:00
break;
code = finish_command(&pp->children[i].process);
if (opts->task_finished)
code = opts->task_finished(code, opts->ungroup ? NULL :
&pp->children[i].err, opts->data,
pp->children[i].data);
else
code = 0;
run-command: add an asynchronous parallel child processor This allows to run external commands in parallel with ordered output on stderr. If we run external commands in parallel we cannot pipe the output directly to the our stdout/err as it would mix up. So each process's output will flow through a pipe, which we buffer. One subprocess can be directly piped to out stdout/err for a low latency feedback to the user. Example: Let's assume we have 5 submodules A,B,C,D,E and each fetch takes a different amount of time as the different submodules vary in size, then the output of fetches in sequential order might look like this: time --> output: |---A---| |-B-| |-------C-------| |-D-| |-E-| When we schedule these submodules into maximal two parallel processes, a schedule and sample output over time may look like this: process 1: |---A---| |-D-| |-E-| process 2: |-B-| |-------C-------| output: |---A---|B|---C-------|DE So A will be perceived as it would run normally in the single child version. As B has finished by the time A is done, we can dump its whole progress buffer on stderr, such that it looks like it finished in no time. Once that is done, C is determined to be the visible child and its progress will be reported in real time. So this way of output is really good for human consumption, as it only changes the timing, not the actual output. For machine consumption the output needs to be prepared in the tasks, by either having a prefix per line or per block to indicate whose tasks output is displayed, because the output order may not follow the original sequential ordering: |----A----| |--B--| |-C-| will be scheduled to be all parallel: process 1: |----A----| process 2: |--B--| process 3: |-C-| output: |----A----|CB This happens because C finished before B did, so it will be queued for output before B. To detect when a child has finished executing, we check interleaved with other actions (such as checking the liveliness of children or starting new processes) whether the stderr pipe still exists. Once a child closed its stderr stream, we assume it is terminating very soon, and use `finish_command()` from the single external process execution interface to collect the exit status. By maintaining the strong assumption of stderr being open until the very end of a child process, we can avoid other hassle such as an implementation using `waitpid(-1)`, which is not implemented in Windows. Signed-off-by: Stefan Beller <sbeller@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-12-16 08:04:10 +08:00
if (code)
result = code;
if (code < 0)
break;
pp->nr_processes--;
pp->children[i].state = GIT_CP_FREE;
run-command: add an "ungroup" option to run_process_parallel() Extend the parallel execution API added in c553c72eed6 (run-command: add an asynchronous parallel child processor, 2015-12-15) to support a mode where the stdout and stderr of the processes isn't captured and output in a deterministic order, instead we'll leave it to the kernel and stdio to sort it out. This gives the API same functionality as GNU parallel's --ungroup option. As we'll see in a subsequent commit the main reason to want this is to support stdout and stderr being connected to the TTY in the case of jobs=1, demonstrated here with GNU parallel: $ parallel --ungroup 'test -t {} && echo TTY || echo NTTY' ::: 1 2 TTY TTY $ parallel 'test -t {} && echo TTY || echo NTTY' ::: 1 2 NTTY NTTY Another is as GNU parallel's documentation notes a potential for optimization. As demonstrated in next commit our results with "git hook run" will be similar, but generally speaking this shows that if you want to run processes in parallel where the exact order isn't important this can be a lot faster: $ hyperfine -r 3 -L o ,--ungroup 'parallel {o} seq ::: 10000000 >/dev/null ' Benchmark 1: parallel seq ::: 10000000 >/dev/null Time (mean ± σ): 220.2 ms ± 9.3 ms [User: 124.9 ms, System: 96.1 ms] Range (min … max): 212.3 ms … 230.5 ms 3 runs Benchmark 2: parallel --ungroup seq ::: 10000000 >/dev/null Time (mean ± σ): 154.7 ms ± 0.9 ms [User: 136.2 ms, System: 25.1 ms] Range (min … max): 153.9 ms … 155.7 ms 3 runs Summary 'parallel --ungroup seq ::: 10000000 >/dev/null ' ran 1.42 ± 0.06 times faster than 'parallel seq ::: 10000000 >/dev/null ' A large part of the juggling in the API is to make the API safer for its maintenance and consumers alike. For the maintenance of the API we e.g. avoid malloc()-ing the "pp->pfd", ensuring that SANITIZE=address and other similar tools will catch any unexpected misuse. For API consumers we take pains to never pass the non-NULL "out" buffer to an API user that provided the "ungroup" option. The resulting code in t/helper/test-run-command.c isn't typical of such a user, i.e. they'd typically use one mode or the other, and would know whether they'd provided "ungroup" or not. We could also avoid the strbuf_init() for "buffered_output" by having "struct parallel_processes" use a static PARALLEL_PROCESSES_INIT initializer, but let's leave that cleanup for later. Using a global "run_processes_parallel_ungroup" variable to enable this option is rather nasty, but is being done here to produce as minimal of a change as possible for a subsequent regression fix. This change is extracted from a larger initial version[1] which ends up with a better end-state for the API, but in doing so needed to modify all existing callers of the API. Let's defer that for now, and narrowly focus on what we need for fixing the regression in the subsequent commit. It's safe to do this with a global variable because: A) hook.c is the only user of it that sets it to non-zero, and before we'll get any other API users we'll refactor away this method of passing in the option, i.e. re-roll [1]. B) Even if hook.c wasn't the only user we don't have callers of this API that concurrently invoke this parallel process starting API itself in parallel. As noted above "A" && "B" are rather nasty, and we don't want to live with those caveats long-term, but for now they should be an acceptable compromise. 1. https://lore.kernel.org/git/cover-v2-0.8-00000000000-20220518T195858Z-avarab@gmail.com/ Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2022-06-07 16:48:19 +08:00
if (pp->pfd)
pp->pfd[i].fd = -1;
run-command: add an asynchronous parallel child processor This allows to run external commands in parallel with ordered output on stderr. If we run external commands in parallel we cannot pipe the output directly to the our stdout/err as it would mix up. So each process's output will flow through a pipe, which we buffer. One subprocess can be directly piped to out stdout/err for a low latency feedback to the user. Example: Let's assume we have 5 submodules A,B,C,D,E and each fetch takes a different amount of time as the different submodules vary in size, then the output of fetches in sequential order might look like this: time --> output: |---A---| |-B-| |-------C-------| |-D-| |-E-| When we schedule these submodules into maximal two parallel processes, a schedule and sample output over time may look like this: process 1: |---A---| |-D-| |-E-| process 2: |-B-| |-------C-------| output: |---A---|B|---C-------|DE So A will be perceived as it would run normally in the single child version. As B has finished by the time A is done, we can dump its whole progress buffer on stderr, such that it looks like it finished in no time. Once that is done, C is determined to be the visible child and its progress will be reported in real time. So this way of output is really good for human consumption, as it only changes the timing, not the actual output. For machine consumption the output needs to be prepared in the tasks, by either having a prefix per line or per block to indicate whose tasks output is displayed, because the output order may not follow the original sequential ordering: |----A----| |--B--| |-C-| will be scheduled to be all parallel: process 1: |----A----| process 2: |--B--| process 3: |-C-| output: |----A----|CB This happens because C finished before B did, so it will be queued for output before B. To detect when a child has finished executing, we check interleaved with other actions (such as checking the liveliness of children or starting new processes) whether the stderr pipe still exists. Once a child closed its stderr stream, we assume it is terminating very soon, and use `finish_command()` from the single external process execution interface to collect the exit status. By maintaining the strong assumption of stderr being open until the very end of a child process, we can avoid other hassle such as an implementation using `waitpid(-1)`, which is not implemented in Windows. Signed-off-by: Stefan Beller <sbeller@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-12-16 08:04:10 +08:00
child_process_init(&pp->children[i].process);
if (opts->ungroup) {
run-command: add an "ungroup" option to run_process_parallel() Extend the parallel execution API added in c553c72eed6 (run-command: add an asynchronous parallel child processor, 2015-12-15) to support a mode where the stdout and stderr of the processes isn't captured and output in a deterministic order, instead we'll leave it to the kernel and stdio to sort it out. This gives the API same functionality as GNU parallel's --ungroup option. As we'll see in a subsequent commit the main reason to want this is to support stdout and stderr being connected to the TTY in the case of jobs=1, demonstrated here with GNU parallel: $ parallel --ungroup 'test -t {} && echo TTY || echo NTTY' ::: 1 2 TTY TTY $ parallel 'test -t {} && echo TTY || echo NTTY' ::: 1 2 NTTY NTTY Another is as GNU parallel's documentation notes a potential for optimization. As demonstrated in next commit our results with "git hook run" will be similar, but generally speaking this shows that if you want to run processes in parallel where the exact order isn't important this can be a lot faster: $ hyperfine -r 3 -L o ,--ungroup 'parallel {o} seq ::: 10000000 >/dev/null ' Benchmark 1: parallel seq ::: 10000000 >/dev/null Time (mean ± σ): 220.2 ms ± 9.3 ms [User: 124.9 ms, System: 96.1 ms] Range (min … max): 212.3 ms … 230.5 ms 3 runs Benchmark 2: parallel --ungroup seq ::: 10000000 >/dev/null Time (mean ± σ): 154.7 ms ± 0.9 ms [User: 136.2 ms, System: 25.1 ms] Range (min … max): 153.9 ms … 155.7 ms 3 runs Summary 'parallel --ungroup seq ::: 10000000 >/dev/null ' ran 1.42 ± 0.06 times faster than 'parallel seq ::: 10000000 >/dev/null ' A large part of the juggling in the API is to make the API safer for its maintenance and consumers alike. For the maintenance of the API we e.g. avoid malloc()-ing the "pp->pfd", ensuring that SANITIZE=address and other similar tools will catch any unexpected misuse. For API consumers we take pains to never pass the non-NULL "out" buffer to an API user that provided the "ungroup" option. The resulting code in t/helper/test-run-command.c isn't typical of such a user, i.e. they'd typically use one mode or the other, and would know whether they'd provided "ungroup" or not. We could also avoid the strbuf_init() for "buffered_output" by having "struct parallel_processes" use a static PARALLEL_PROCESSES_INIT initializer, but let's leave that cleanup for later. Using a global "run_processes_parallel_ungroup" variable to enable this option is rather nasty, but is being done here to produce as minimal of a change as possible for a subsequent regression fix. This change is extracted from a larger initial version[1] which ends up with a better end-state for the API, but in doing so needed to modify all existing callers of the API. Let's defer that for now, and narrowly focus on what we need for fixing the regression in the subsequent commit. It's safe to do this with a global variable because: A) hook.c is the only user of it that sets it to non-zero, and before we'll get any other API users we'll refactor away this method of passing in the option, i.e. re-roll [1]. B) Even if hook.c wasn't the only user we don't have callers of this API that concurrently invoke this parallel process starting API itself in parallel. As noted above "A" && "B" are rather nasty, and we don't want to live with those caveats long-term, but for now they should be an acceptable compromise. 1. https://lore.kernel.org/git/cover-v2-0.8-00000000000-20220518T195858Z-avarab@gmail.com/ Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2022-06-07 16:48:19 +08:00
; /* no strbuf_*() work to do here */
} else if (i != pp->output_owner) {
run-command: add an asynchronous parallel child processor This allows to run external commands in parallel with ordered output on stderr. If we run external commands in parallel we cannot pipe the output directly to the our stdout/err as it would mix up. So each process's output will flow through a pipe, which we buffer. One subprocess can be directly piped to out stdout/err for a low latency feedback to the user. Example: Let's assume we have 5 submodules A,B,C,D,E and each fetch takes a different amount of time as the different submodules vary in size, then the output of fetches in sequential order might look like this: time --> output: |---A---| |-B-| |-------C-------| |-D-| |-E-| When we schedule these submodules into maximal two parallel processes, a schedule and sample output over time may look like this: process 1: |---A---| |-D-| |-E-| process 2: |-B-| |-------C-------| output: |---A---|B|---C-------|DE So A will be perceived as it would run normally in the single child version. As B has finished by the time A is done, we can dump its whole progress buffer on stderr, such that it looks like it finished in no time. Once that is done, C is determined to be the visible child and its progress will be reported in real time. So this way of output is really good for human consumption, as it only changes the timing, not the actual output. For machine consumption the output needs to be prepared in the tasks, by either having a prefix per line or per block to indicate whose tasks output is displayed, because the output order may not follow the original sequential ordering: |----A----| |--B--| |-C-| will be scheduled to be all parallel: process 1: |----A----| process 2: |--B--| process 3: |-C-| output: |----A----|CB This happens because C finished before B did, so it will be queued for output before B. To detect when a child has finished executing, we check interleaved with other actions (such as checking the liveliness of children or starting new processes) whether the stderr pipe still exists. Once a child closed its stderr stream, we assume it is terminating very soon, and use `finish_command()` from the single external process execution interface to collect the exit status. By maintaining the strong assumption of stderr being open until the very end of a child process, we can avoid other hassle such as an implementation using `waitpid(-1)`, which is not implemented in Windows. Signed-off-by: Stefan Beller <sbeller@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-12-16 08:04:10 +08:00
strbuf_addbuf(&pp->buffered_output, &pp->children[i].err);
strbuf_reset(&pp->children[i].err);
} else {
const size_t n = opts->processes;
strbuf_write(&pp->children[i].err, stderr);
run-command: add an asynchronous parallel child processor This allows to run external commands in parallel with ordered output on stderr. If we run external commands in parallel we cannot pipe the output directly to the our stdout/err as it would mix up. So each process's output will flow through a pipe, which we buffer. One subprocess can be directly piped to out stdout/err for a low latency feedback to the user. Example: Let's assume we have 5 submodules A,B,C,D,E and each fetch takes a different amount of time as the different submodules vary in size, then the output of fetches in sequential order might look like this: time --> output: |---A---| |-B-| |-------C-------| |-D-| |-E-| When we schedule these submodules into maximal two parallel processes, a schedule and sample output over time may look like this: process 1: |---A---| |-D-| |-E-| process 2: |-B-| |-------C-------| output: |---A---|B|---C-------|DE So A will be perceived as it would run normally in the single child version. As B has finished by the time A is done, we can dump its whole progress buffer on stderr, such that it looks like it finished in no time. Once that is done, C is determined to be the visible child and its progress will be reported in real time. So this way of output is really good for human consumption, as it only changes the timing, not the actual output. For machine consumption the output needs to be prepared in the tasks, by either having a prefix per line or per block to indicate whose tasks output is displayed, because the output order may not follow the original sequential ordering: |----A----| |--B--| |-C-| will be scheduled to be all parallel: process 1: |----A----| process 2: |--B--| process 3: |-C-| output: |----A----|CB This happens because C finished before B did, so it will be queued for output before B. To detect when a child has finished executing, we check interleaved with other actions (such as checking the liveliness of children or starting new processes) whether the stderr pipe still exists. Once a child closed its stderr stream, we assume it is terminating very soon, and use `finish_command()` from the single external process execution interface to collect the exit status. By maintaining the strong assumption of stderr being open until the very end of a child process, we can avoid other hassle such as an implementation using `waitpid(-1)`, which is not implemented in Windows. Signed-off-by: Stefan Beller <sbeller@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-12-16 08:04:10 +08:00
strbuf_reset(&pp->children[i].err);
/* Output all other finished child processes */
strbuf_write(&pp->buffered_output, stderr);
run-command: add an asynchronous parallel child processor This allows to run external commands in parallel with ordered output on stderr. If we run external commands in parallel we cannot pipe the output directly to the our stdout/err as it would mix up. So each process's output will flow through a pipe, which we buffer. One subprocess can be directly piped to out stdout/err for a low latency feedback to the user. Example: Let's assume we have 5 submodules A,B,C,D,E and each fetch takes a different amount of time as the different submodules vary in size, then the output of fetches in sequential order might look like this: time --> output: |---A---| |-B-| |-------C-------| |-D-| |-E-| When we schedule these submodules into maximal two parallel processes, a schedule and sample output over time may look like this: process 1: |---A---| |-D-| |-E-| process 2: |-B-| |-------C-------| output: |---A---|B|---C-------|DE So A will be perceived as it would run normally in the single child version. As B has finished by the time A is done, we can dump its whole progress buffer on stderr, such that it looks like it finished in no time. Once that is done, C is determined to be the visible child and its progress will be reported in real time. So this way of output is really good for human consumption, as it only changes the timing, not the actual output. For machine consumption the output needs to be prepared in the tasks, by either having a prefix per line or per block to indicate whose tasks output is displayed, because the output order may not follow the original sequential ordering: |----A----| |--B--| |-C-| will be scheduled to be all parallel: process 1: |----A----| process 2: |--B--| process 3: |-C-| output: |----A----|CB This happens because C finished before B did, so it will be queued for output before B. To detect when a child has finished executing, we check interleaved with other actions (such as checking the liveliness of children or starting new processes) whether the stderr pipe still exists. Once a child closed its stderr stream, we assume it is terminating very soon, and use `finish_command()` from the single external process execution interface to collect the exit status. By maintaining the strong assumption of stderr being open until the very end of a child process, we can avoid other hassle such as an implementation using `waitpid(-1)`, which is not implemented in Windows. Signed-off-by: Stefan Beller <sbeller@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-12-16 08:04:10 +08:00
strbuf_reset(&pp->buffered_output);
/*
* Pick next process to output live.
* NEEDSWORK:
* For now we pick it randomly by doing a round
* robin. Later we may want to pick the one with
* the most output or the longest or shortest
* running process time.
*/
for (i = 0; i < n; i++)
if (pp->children[(pp->output_owner + i) % n].state == GIT_CP_WORKING)
break;
pp->output_owner = (pp->output_owner + i) % n;
}
}
return result;
}
void run_processes_parallel(const struct run_process_parallel_opts *opts)
run-command: add an asynchronous parallel child processor This allows to run external commands in parallel with ordered output on stderr. If we run external commands in parallel we cannot pipe the output directly to the our stdout/err as it would mix up. So each process's output will flow through a pipe, which we buffer. One subprocess can be directly piped to out stdout/err for a low latency feedback to the user. Example: Let's assume we have 5 submodules A,B,C,D,E and each fetch takes a different amount of time as the different submodules vary in size, then the output of fetches in sequential order might look like this: time --> output: |---A---| |-B-| |-------C-------| |-D-| |-E-| When we schedule these submodules into maximal two parallel processes, a schedule and sample output over time may look like this: process 1: |---A---| |-D-| |-E-| process 2: |-B-| |-------C-------| output: |---A---|B|---C-------|DE So A will be perceived as it would run normally in the single child version. As B has finished by the time A is done, we can dump its whole progress buffer on stderr, such that it looks like it finished in no time. Once that is done, C is determined to be the visible child and its progress will be reported in real time. So this way of output is really good for human consumption, as it only changes the timing, not the actual output. For machine consumption the output needs to be prepared in the tasks, by either having a prefix per line or per block to indicate whose tasks output is displayed, because the output order may not follow the original sequential ordering: |----A----| |--B--| |-C-| will be scheduled to be all parallel: process 1: |----A----| process 2: |--B--| process 3: |-C-| output: |----A----|CB This happens because C finished before B did, so it will be queued for output before B. To detect when a child has finished executing, we check interleaved with other actions (such as checking the liveliness of children or starting new processes) whether the stderr pipe still exists. Once a child closed its stderr stream, we assume it is terminating very soon, and use `finish_command()` from the single external process execution interface to collect the exit status. By maintaining the strong assumption of stderr being open until the very end of a child process, we can avoid other hassle such as an implementation using `waitpid(-1)`, which is not implemented in Windows. Signed-off-by: Stefan Beller <sbeller@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-12-16 08:04:10 +08:00
{
int i, code;
int output_timeout = 100;
int spawn_cap = 4;
struct parallel_processes_for_signal pp_sig;
struct parallel_processes pp = {
.buffered_output = STRBUF_INIT,
};
/* options */
const char *tr2_category = opts->tr2_category;
const char *tr2_label = opts->tr2_label;
const int do_trace2 = tr2_category && tr2_label;
run-command: add an asynchronous parallel child processor This allows to run external commands in parallel with ordered output on stderr. If we run external commands in parallel we cannot pipe the output directly to the our stdout/err as it would mix up. So each process's output will flow through a pipe, which we buffer. One subprocess can be directly piped to out stdout/err for a low latency feedback to the user. Example: Let's assume we have 5 submodules A,B,C,D,E and each fetch takes a different amount of time as the different submodules vary in size, then the output of fetches in sequential order might look like this: time --> output: |---A---| |-B-| |-------C-------| |-D-| |-E-| When we schedule these submodules into maximal two parallel processes, a schedule and sample output over time may look like this: process 1: |---A---| |-D-| |-E-| process 2: |-B-| |-------C-------| output: |---A---|B|---C-------|DE So A will be perceived as it would run normally in the single child version. As B has finished by the time A is done, we can dump its whole progress buffer on stderr, such that it looks like it finished in no time. Once that is done, C is determined to be the visible child and its progress will be reported in real time. So this way of output is really good for human consumption, as it only changes the timing, not the actual output. For machine consumption the output needs to be prepared in the tasks, by either having a prefix per line or per block to indicate whose tasks output is displayed, because the output order may not follow the original sequential ordering: |----A----| |--B--| |-C-| will be scheduled to be all parallel: process 1: |----A----| process 2: |--B--| process 3: |-C-| output: |----A----|CB This happens because C finished before B did, so it will be queued for output before B. To detect when a child has finished executing, we check interleaved with other actions (such as checking the liveliness of children or starting new processes) whether the stderr pipe still exists. Once a child closed its stderr stream, we assume it is terminating very soon, and use `finish_command()` from the single external process execution interface to collect the exit status. By maintaining the strong assumption of stderr being open until the very end of a child process, we can avoid other hassle such as an implementation using `waitpid(-1)`, which is not implemented in Windows. Signed-off-by: Stefan Beller <sbeller@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-12-16 08:04:10 +08:00
if (do_trace2)
trace2_region_enter_printf(tr2_category, tr2_label, NULL,
"max:%"PRIuMAX,
(uintmax_t)opts->processes);
run-command: add an "ungroup" option to run_process_parallel() Extend the parallel execution API added in c553c72eed6 (run-command: add an asynchronous parallel child processor, 2015-12-15) to support a mode where the stdout and stderr of the processes isn't captured and output in a deterministic order, instead we'll leave it to the kernel and stdio to sort it out. This gives the API same functionality as GNU parallel's --ungroup option. As we'll see in a subsequent commit the main reason to want this is to support stdout and stderr being connected to the TTY in the case of jobs=1, demonstrated here with GNU parallel: $ parallel --ungroup 'test -t {} && echo TTY || echo NTTY' ::: 1 2 TTY TTY $ parallel 'test -t {} && echo TTY || echo NTTY' ::: 1 2 NTTY NTTY Another is as GNU parallel's documentation notes a potential for optimization. As demonstrated in next commit our results with "git hook run" will be similar, but generally speaking this shows that if you want to run processes in parallel where the exact order isn't important this can be a lot faster: $ hyperfine -r 3 -L o ,--ungroup 'parallel {o} seq ::: 10000000 >/dev/null ' Benchmark 1: parallel seq ::: 10000000 >/dev/null Time (mean ± σ): 220.2 ms ± 9.3 ms [User: 124.9 ms, System: 96.1 ms] Range (min … max): 212.3 ms … 230.5 ms 3 runs Benchmark 2: parallel --ungroup seq ::: 10000000 >/dev/null Time (mean ± σ): 154.7 ms ± 0.9 ms [User: 136.2 ms, System: 25.1 ms] Range (min … max): 153.9 ms … 155.7 ms 3 runs Summary 'parallel --ungroup seq ::: 10000000 >/dev/null ' ran 1.42 ± 0.06 times faster than 'parallel seq ::: 10000000 >/dev/null ' A large part of the juggling in the API is to make the API safer for its maintenance and consumers alike. For the maintenance of the API we e.g. avoid malloc()-ing the "pp->pfd", ensuring that SANITIZE=address and other similar tools will catch any unexpected misuse. For API consumers we take pains to never pass the non-NULL "out" buffer to an API user that provided the "ungroup" option. The resulting code in t/helper/test-run-command.c isn't typical of such a user, i.e. they'd typically use one mode or the other, and would know whether they'd provided "ungroup" or not. We could also avoid the strbuf_init() for "buffered_output" by having "struct parallel_processes" use a static PARALLEL_PROCESSES_INIT initializer, but let's leave that cleanup for later. Using a global "run_processes_parallel_ungroup" variable to enable this option is rather nasty, but is being done here to produce as minimal of a change as possible for a subsequent regression fix. This change is extracted from a larger initial version[1] which ends up with a better end-state for the API, but in doing so needed to modify all existing callers of the API. Let's defer that for now, and narrowly focus on what we need for fixing the regression in the subsequent commit. It's safe to do this with a global variable because: A) hook.c is the only user of it that sets it to non-zero, and before we'll get any other API users we'll refactor away this method of passing in the option, i.e. re-roll [1]. B) Even if hook.c wasn't the only user we don't have callers of this API that concurrently invoke this parallel process starting API itself in parallel. As noted above "A" && "B" are rather nasty, and we don't want to live with those caveats long-term, but for now they should be an acceptable compromise. 1. https://lore.kernel.org/git/cover-v2-0.8-00000000000-20220518T195858Z-avarab@gmail.com/ Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2022-06-07 16:48:19 +08:00
pp_init(&pp, opts, &pp_sig);
run-command: add an asynchronous parallel child processor This allows to run external commands in parallel with ordered output on stderr. If we run external commands in parallel we cannot pipe the output directly to the our stdout/err as it would mix up. So each process's output will flow through a pipe, which we buffer. One subprocess can be directly piped to out stdout/err for a low latency feedback to the user. Example: Let's assume we have 5 submodules A,B,C,D,E and each fetch takes a different amount of time as the different submodules vary in size, then the output of fetches in sequential order might look like this: time --> output: |---A---| |-B-| |-------C-------| |-D-| |-E-| When we schedule these submodules into maximal two parallel processes, a schedule and sample output over time may look like this: process 1: |---A---| |-D-| |-E-| process 2: |-B-| |-------C-------| output: |---A---|B|---C-------|DE So A will be perceived as it would run normally in the single child version. As B has finished by the time A is done, we can dump its whole progress buffer on stderr, such that it looks like it finished in no time. Once that is done, C is determined to be the visible child and its progress will be reported in real time. So this way of output is really good for human consumption, as it only changes the timing, not the actual output. For machine consumption the output needs to be prepared in the tasks, by either having a prefix per line or per block to indicate whose tasks output is displayed, because the output order may not follow the original sequential ordering: |----A----| |--B--| |-C-| will be scheduled to be all parallel: process 1: |----A----| process 2: |--B--| process 3: |-C-| output: |----A----|CB This happens because C finished before B did, so it will be queued for output before B. To detect when a child has finished executing, we check interleaved with other actions (such as checking the liveliness of children or starting new processes) whether the stderr pipe still exists. Once a child closed its stderr stream, we assume it is terminating very soon, and use `finish_command()` from the single external process execution interface to collect the exit status. By maintaining the strong assumption of stderr being open until the very end of a child process, we can avoid other hassle such as an implementation using `waitpid(-1)`, which is not implemented in Windows. Signed-off-by: Stefan Beller <sbeller@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-12-16 08:04:10 +08:00
while (1) {
for (i = 0;
i < spawn_cap && !pp.shutdown &&
pp.nr_processes < opts->processes;
run-command: add an asynchronous parallel child processor This allows to run external commands in parallel with ordered output on stderr. If we run external commands in parallel we cannot pipe the output directly to the our stdout/err as it would mix up. So each process's output will flow through a pipe, which we buffer. One subprocess can be directly piped to out stdout/err for a low latency feedback to the user. Example: Let's assume we have 5 submodules A,B,C,D,E and each fetch takes a different amount of time as the different submodules vary in size, then the output of fetches in sequential order might look like this: time --> output: |---A---| |-B-| |-------C-------| |-D-| |-E-| When we schedule these submodules into maximal two parallel processes, a schedule and sample output over time may look like this: process 1: |---A---| |-D-| |-E-| process 2: |-B-| |-------C-------| output: |---A---|B|---C-------|DE So A will be perceived as it would run normally in the single child version. As B has finished by the time A is done, we can dump its whole progress buffer on stderr, such that it looks like it finished in no time. Once that is done, C is determined to be the visible child and its progress will be reported in real time. So this way of output is really good for human consumption, as it only changes the timing, not the actual output. For machine consumption the output needs to be prepared in the tasks, by either having a prefix per line or per block to indicate whose tasks output is displayed, because the output order may not follow the original sequential ordering: |----A----| |--B--| |-C-| will be scheduled to be all parallel: process 1: |----A----| process 2: |--B--| process 3: |-C-| output: |----A----|CB This happens because C finished before B did, so it will be queued for output before B. To detect when a child has finished executing, we check interleaved with other actions (such as checking the liveliness of children or starting new processes) whether the stderr pipe still exists. Once a child closed its stderr stream, we assume it is terminating very soon, and use `finish_command()` from the single external process execution interface to collect the exit status. By maintaining the strong assumption of stderr being open until the very end of a child process, we can avoid other hassle such as an implementation using `waitpid(-1)`, which is not implemented in Windows. Signed-off-by: Stefan Beller <sbeller@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-12-16 08:04:10 +08:00
i++) {
code = pp_start_one(&pp, opts);
run-command: add an asynchronous parallel child processor This allows to run external commands in parallel with ordered output on stderr. If we run external commands in parallel we cannot pipe the output directly to the our stdout/err as it would mix up. So each process's output will flow through a pipe, which we buffer. One subprocess can be directly piped to out stdout/err for a low latency feedback to the user. Example: Let's assume we have 5 submodules A,B,C,D,E and each fetch takes a different amount of time as the different submodules vary in size, then the output of fetches in sequential order might look like this: time --> output: |---A---| |-B-| |-------C-------| |-D-| |-E-| When we schedule these submodules into maximal two parallel processes, a schedule and sample output over time may look like this: process 1: |---A---| |-D-| |-E-| process 2: |-B-| |-------C-------| output: |---A---|B|---C-------|DE So A will be perceived as it would run normally in the single child version. As B has finished by the time A is done, we can dump its whole progress buffer on stderr, such that it looks like it finished in no time. Once that is done, C is determined to be the visible child and its progress will be reported in real time. So this way of output is really good for human consumption, as it only changes the timing, not the actual output. For machine consumption the output needs to be prepared in the tasks, by either having a prefix per line or per block to indicate whose tasks output is displayed, because the output order may not follow the original sequential ordering: |----A----| |--B--| |-C-| will be scheduled to be all parallel: process 1: |----A----| process 2: |--B--| process 3: |-C-| output: |----A----|CB This happens because C finished before B did, so it will be queued for output before B. To detect when a child has finished executing, we check interleaved with other actions (such as checking the liveliness of children or starting new processes) whether the stderr pipe still exists. Once a child closed its stderr stream, we assume it is terminating very soon, and use `finish_command()` from the single external process execution interface to collect the exit status. By maintaining the strong assumption of stderr being open until the very end of a child process, we can avoid other hassle such as an implementation using `waitpid(-1)`, which is not implemented in Windows. Signed-off-by: Stefan Beller <sbeller@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-12-16 08:04:10 +08:00
if (!code)
continue;
if (code < 0) {
pp.shutdown = 1;
kill_children(&pp, opts, -code);
run-command: add an asynchronous parallel child processor This allows to run external commands in parallel with ordered output on stderr. If we run external commands in parallel we cannot pipe the output directly to the our stdout/err as it would mix up. So each process's output will flow through a pipe, which we buffer. One subprocess can be directly piped to out stdout/err for a low latency feedback to the user. Example: Let's assume we have 5 submodules A,B,C,D,E and each fetch takes a different amount of time as the different submodules vary in size, then the output of fetches in sequential order might look like this: time --> output: |---A---| |-B-| |-------C-------| |-D-| |-E-| When we schedule these submodules into maximal two parallel processes, a schedule and sample output over time may look like this: process 1: |---A---| |-D-| |-E-| process 2: |-B-| |-------C-------| output: |---A---|B|---C-------|DE So A will be perceived as it would run normally in the single child version. As B has finished by the time A is done, we can dump its whole progress buffer on stderr, such that it looks like it finished in no time. Once that is done, C is determined to be the visible child and its progress will be reported in real time. So this way of output is really good for human consumption, as it only changes the timing, not the actual output. For machine consumption the output needs to be prepared in the tasks, by either having a prefix per line or per block to indicate whose tasks output is displayed, because the output order may not follow the original sequential ordering: |----A----| |--B--| |-C-| will be scheduled to be all parallel: process 1: |----A----| process 2: |--B--| process 3: |-C-| output: |----A----|CB This happens because C finished before B did, so it will be queued for output before B. To detect when a child has finished executing, we check interleaved with other actions (such as checking the liveliness of children or starting new processes) whether the stderr pipe still exists. Once a child closed its stderr stream, we assume it is terminating very soon, and use `finish_command()` from the single external process execution interface to collect the exit status. By maintaining the strong assumption of stderr being open until the very end of a child process, we can avoid other hassle such as an implementation using `waitpid(-1)`, which is not implemented in Windows. Signed-off-by: Stefan Beller <sbeller@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-12-16 08:04:10 +08:00
}
break;
}
if (!pp.nr_processes)
break;
if (opts->ungroup) {
for (size_t i = 0; i < opts->processes; i++)
run-command: add an "ungroup" option to run_process_parallel() Extend the parallel execution API added in c553c72eed6 (run-command: add an asynchronous parallel child processor, 2015-12-15) to support a mode where the stdout and stderr of the processes isn't captured and output in a deterministic order, instead we'll leave it to the kernel and stdio to sort it out. This gives the API same functionality as GNU parallel's --ungroup option. As we'll see in a subsequent commit the main reason to want this is to support stdout and stderr being connected to the TTY in the case of jobs=1, demonstrated here with GNU parallel: $ parallel --ungroup 'test -t {} && echo TTY || echo NTTY' ::: 1 2 TTY TTY $ parallel 'test -t {} && echo TTY || echo NTTY' ::: 1 2 NTTY NTTY Another is as GNU parallel's documentation notes a potential for optimization. As demonstrated in next commit our results with "git hook run" will be similar, but generally speaking this shows that if you want to run processes in parallel where the exact order isn't important this can be a lot faster: $ hyperfine -r 3 -L o ,--ungroup 'parallel {o} seq ::: 10000000 >/dev/null ' Benchmark 1: parallel seq ::: 10000000 >/dev/null Time (mean ± σ): 220.2 ms ± 9.3 ms [User: 124.9 ms, System: 96.1 ms] Range (min … max): 212.3 ms … 230.5 ms 3 runs Benchmark 2: parallel --ungroup seq ::: 10000000 >/dev/null Time (mean ± σ): 154.7 ms ± 0.9 ms [User: 136.2 ms, System: 25.1 ms] Range (min … max): 153.9 ms … 155.7 ms 3 runs Summary 'parallel --ungroup seq ::: 10000000 >/dev/null ' ran 1.42 ± 0.06 times faster than 'parallel seq ::: 10000000 >/dev/null ' A large part of the juggling in the API is to make the API safer for its maintenance and consumers alike. For the maintenance of the API we e.g. avoid malloc()-ing the "pp->pfd", ensuring that SANITIZE=address and other similar tools will catch any unexpected misuse. For API consumers we take pains to never pass the non-NULL "out" buffer to an API user that provided the "ungroup" option. The resulting code in t/helper/test-run-command.c isn't typical of such a user, i.e. they'd typically use one mode or the other, and would know whether they'd provided "ungroup" or not. We could also avoid the strbuf_init() for "buffered_output" by having "struct parallel_processes" use a static PARALLEL_PROCESSES_INIT initializer, but let's leave that cleanup for later. Using a global "run_processes_parallel_ungroup" variable to enable this option is rather nasty, but is being done here to produce as minimal of a change as possible for a subsequent regression fix. This change is extracted from a larger initial version[1] which ends up with a better end-state for the API, but in doing so needed to modify all existing callers of the API. Let's defer that for now, and narrowly focus on what we need for fixing the regression in the subsequent commit. It's safe to do this with a global variable because: A) hook.c is the only user of it that sets it to non-zero, and before we'll get any other API users we'll refactor away this method of passing in the option, i.e. re-roll [1]. B) Even if hook.c wasn't the only user we don't have callers of this API that concurrently invoke this parallel process starting API itself in parallel. As noted above "A" && "B" are rather nasty, and we don't want to live with those caveats long-term, but for now they should be an acceptable compromise. 1. https://lore.kernel.org/git/cover-v2-0.8-00000000000-20220518T195858Z-avarab@gmail.com/ Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2022-06-07 16:48:19 +08:00
pp.children[i].state = GIT_CP_WAIT_CLEANUP;
} else {
pp_buffer_stderr(&pp, opts, output_timeout);
run-command: add an "ungroup" option to run_process_parallel() Extend the parallel execution API added in c553c72eed6 (run-command: add an asynchronous parallel child processor, 2015-12-15) to support a mode where the stdout and stderr of the processes isn't captured and output in a deterministic order, instead we'll leave it to the kernel and stdio to sort it out. This gives the API same functionality as GNU parallel's --ungroup option. As we'll see in a subsequent commit the main reason to want this is to support stdout and stderr being connected to the TTY in the case of jobs=1, demonstrated here with GNU parallel: $ parallel --ungroup 'test -t {} && echo TTY || echo NTTY' ::: 1 2 TTY TTY $ parallel 'test -t {} && echo TTY || echo NTTY' ::: 1 2 NTTY NTTY Another is as GNU parallel's documentation notes a potential for optimization. As demonstrated in next commit our results with "git hook run" will be similar, but generally speaking this shows that if you want to run processes in parallel where the exact order isn't important this can be a lot faster: $ hyperfine -r 3 -L o ,--ungroup 'parallel {o} seq ::: 10000000 >/dev/null ' Benchmark 1: parallel seq ::: 10000000 >/dev/null Time (mean ± σ): 220.2 ms ± 9.3 ms [User: 124.9 ms, System: 96.1 ms] Range (min … max): 212.3 ms … 230.5 ms 3 runs Benchmark 2: parallel --ungroup seq ::: 10000000 >/dev/null Time (mean ± σ): 154.7 ms ± 0.9 ms [User: 136.2 ms, System: 25.1 ms] Range (min … max): 153.9 ms … 155.7 ms 3 runs Summary 'parallel --ungroup seq ::: 10000000 >/dev/null ' ran 1.42 ± 0.06 times faster than 'parallel seq ::: 10000000 >/dev/null ' A large part of the juggling in the API is to make the API safer for its maintenance and consumers alike. For the maintenance of the API we e.g. avoid malloc()-ing the "pp->pfd", ensuring that SANITIZE=address and other similar tools will catch any unexpected misuse. For API consumers we take pains to never pass the non-NULL "out" buffer to an API user that provided the "ungroup" option. The resulting code in t/helper/test-run-command.c isn't typical of such a user, i.e. they'd typically use one mode or the other, and would know whether they'd provided "ungroup" or not. We could also avoid the strbuf_init() for "buffered_output" by having "struct parallel_processes" use a static PARALLEL_PROCESSES_INIT initializer, but let's leave that cleanup for later. Using a global "run_processes_parallel_ungroup" variable to enable this option is rather nasty, but is being done here to produce as minimal of a change as possible for a subsequent regression fix. This change is extracted from a larger initial version[1] which ends up with a better end-state for the API, but in doing so needed to modify all existing callers of the API. Let's defer that for now, and narrowly focus on what we need for fixing the regression in the subsequent commit. It's safe to do this with a global variable because: A) hook.c is the only user of it that sets it to non-zero, and before we'll get any other API users we'll refactor away this method of passing in the option, i.e. re-roll [1]. B) Even if hook.c wasn't the only user we don't have callers of this API that concurrently invoke this parallel process starting API itself in parallel. As noted above "A" && "B" are rather nasty, and we don't want to live with those caveats long-term, but for now they should be an acceptable compromise. 1. https://lore.kernel.org/git/cover-v2-0.8-00000000000-20220518T195858Z-avarab@gmail.com/ Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2022-06-07 16:48:19 +08:00
pp_output(&pp);
}
code = pp_collect_finished(&pp, opts);
run-command: add an asynchronous parallel child processor This allows to run external commands in parallel with ordered output on stderr. If we run external commands in parallel we cannot pipe the output directly to the our stdout/err as it would mix up. So each process's output will flow through a pipe, which we buffer. One subprocess can be directly piped to out stdout/err for a low latency feedback to the user. Example: Let's assume we have 5 submodules A,B,C,D,E and each fetch takes a different amount of time as the different submodules vary in size, then the output of fetches in sequential order might look like this: time --> output: |---A---| |-B-| |-------C-------| |-D-| |-E-| When we schedule these submodules into maximal two parallel processes, a schedule and sample output over time may look like this: process 1: |---A---| |-D-| |-E-| process 2: |-B-| |-------C-------| output: |---A---|B|---C-------|DE So A will be perceived as it would run normally in the single child version. As B has finished by the time A is done, we can dump its whole progress buffer on stderr, such that it looks like it finished in no time. Once that is done, C is determined to be the visible child and its progress will be reported in real time. So this way of output is really good for human consumption, as it only changes the timing, not the actual output. For machine consumption the output needs to be prepared in the tasks, by either having a prefix per line or per block to indicate whose tasks output is displayed, because the output order may not follow the original sequential ordering: |----A----| |--B--| |-C-| will be scheduled to be all parallel: process 1: |----A----| process 2: |--B--| process 3: |-C-| output: |----A----|CB This happens because C finished before B did, so it will be queued for output before B. To detect when a child has finished executing, we check interleaved with other actions (such as checking the liveliness of children or starting new processes) whether the stderr pipe still exists. Once a child closed its stderr stream, we assume it is terminating very soon, and use `finish_command()` from the single external process execution interface to collect the exit status. By maintaining the strong assumption of stderr being open until the very end of a child process, we can avoid other hassle such as an implementation using `waitpid(-1)`, which is not implemented in Windows. Signed-off-by: Stefan Beller <sbeller@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-12-16 08:04:10 +08:00
if (code) {
pp.shutdown = 1;
if (code < 0)
kill_children(&pp, opts,-code);
run-command: add an asynchronous parallel child processor This allows to run external commands in parallel with ordered output on stderr. If we run external commands in parallel we cannot pipe the output directly to the our stdout/err as it would mix up. So each process's output will flow through a pipe, which we buffer. One subprocess can be directly piped to out stdout/err for a low latency feedback to the user. Example: Let's assume we have 5 submodules A,B,C,D,E and each fetch takes a different amount of time as the different submodules vary in size, then the output of fetches in sequential order might look like this: time --> output: |---A---| |-B-| |-------C-------| |-D-| |-E-| When we schedule these submodules into maximal two parallel processes, a schedule and sample output over time may look like this: process 1: |---A---| |-D-| |-E-| process 2: |-B-| |-------C-------| output: |---A---|B|---C-------|DE So A will be perceived as it would run normally in the single child version. As B has finished by the time A is done, we can dump its whole progress buffer on stderr, such that it looks like it finished in no time. Once that is done, C is determined to be the visible child and its progress will be reported in real time. So this way of output is really good for human consumption, as it only changes the timing, not the actual output. For machine consumption the output needs to be prepared in the tasks, by either having a prefix per line or per block to indicate whose tasks output is displayed, because the output order may not follow the original sequential ordering: |----A----| |--B--| |-C-| will be scheduled to be all parallel: process 1: |----A----| process 2: |--B--| process 3: |-C-| output: |----A----|CB This happens because C finished before B did, so it will be queued for output before B. To detect when a child has finished executing, we check interleaved with other actions (such as checking the liveliness of children or starting new processes) whether the stderr pipe still exists. Once a child closed its stderr stream, we assume it is terminating very soon, and use `finish_command()` from the single external process execution interface to collect the exit status. By maintaining the strong assumption of stderr being open until the very end of a child process, we can avoid other hassle such as an implementation using `waitpid(-1)`, which is not implemented in Windows. Signed-off-by: Stefan Beller <sbeller@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-12-16 08:04:10 +08:00
}
}
pp_cleanup(&pp, opts);
if (do_trace2)
trace2_region_leave(tr2_category, tr2_label, NULL);
run-command: add an asynchronous parallel child processor This allows to run external commands in parallel with ordered output on stderr. If we run external commands in parallel we cannot pipe the output directly to the our stdout/err as it would mix up. So each process's output will flow through a pipe, which we buffer. One subprocess can be directly piped to out stdout/err for a low latency feedback to the user. Example: Let's assume we have 5 submodules A,B,C,D,E and each fetch takes a different amount of time as the different submodules vary in size, then the output of fetches in sequential order might look like this: time --> output: |---A---| |-B-| |-------C-------| |-D-| |-E-| When we schedule these submodules into maximal two parallel processes, a schedule and sample output over time may look like this: process 1: |---A---| |-D-| |-E-| process 2: |-B-| |-------C-------| output: |---A---|B|---C-------|DE So A will be perceived as it would run normally in the single child version. As B has finished by the time A is done, we can dump its whole progress buffer on stderr, such that it looks like it finished in no time. Once that is done, C is determined to be the visible child and its progress will be reported in real time. So this way of output is really good for human consumption, as it only changes the timing, not the actual output. For machine consumption the output needs to be prepared in the tasks, by either having a prefix per line or per block to indicate whose tasks output is displayed, because the output order may not follow the original sequential ordering: |----A----| |--B--| |-C-| will be scheduled to be all parallel: process 1: |----A----| process 2: |--B--| process 3: |-C-| output: |----A----|CB This happens because C finished before B did, so it will be queued for output before B. To detect when a child has finished executing, we check interleaved with other actions (such as checking the liveliness of children or starting new processes) whether the stderr pipe still exists. Once a child closed its stderr stream, we assume it is terminating very soon, and use `finish_command()` from the single external process execution interface to collect the exit status. By maintaining the strong assumption of stderr being open until the very end of a child process, we can avoid other hassle such as an implementation using `waitpid(-1)`, which is not implemented in Windows. Signed-off-by: Stefan Beller <sbeller@google.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2015-12-16 08:04:10 +08:00
}
int prepare_auto_maintenance(int quiet, struct child_process *maint)
{
run-command: fix detaching when running auto maintenance In the past, we used to execute `git gc --auto` as part of our automatic housekeeping routines. As git-gc(1) may require quite some time to perform the housekeeping, it knows to detach itself and run in the background so that the user can continue their work. Eventually, we refactored our automatic housekeeping to instead use the more flexible git-maintenance(1) command. The upside of this new infra is that the user can configure which maintenance tasks are performed, at least to a certain degree. So while it continues to run git-gc(1) by default, it can also be adapted to e.g. use git-multi-pack-index(1) for maintenance of the object database. The auto-detach of the new infra is somewhat broken though once the user configures non-standard tasks. The problem is essentially that we detach at the wrong level in the process hierarchy: git-maintenance(1) never detaches itself, but instead it continues to be git-gc(1) which does. When configured to only run the git-gc(1) maintenance task, then the result is basically the same as before. But when configured to run other tasks, then git-maintenance(1) will wait for these to run to completion. Even worse, it may be that git-gc(1) runs concurrently with other housekeeping tasks, stomping on each others feet. Fix this bug by asking git-gc(1) to not detach when it is being invoked via git-maintenance(1). Instead, git-maintenance(1) now respects a new config "maintenance.autoDetach", the equivalent of "gc.autoDetach", and detaches itself into the background when running as part of our auto maintenance. This should continue to behave the same for all users which use the git-gc(1) task, only. For others though, it means that we now properly perform all tasks in the background. The default behaviour of git-maintenance(1) when executed by the user does not change, it will remain in the foreground unless they pass the `--detach` option. Signed-off-by: Patrick Steinhardt <ps@pks.im> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2024-08-16 18:45:17 +08:00
int enabled, auto_detach;
if (!git_config_get_bool("maintenance.auto", &enabled) &&
!enabled)
return 0;
run-command: fix detaching when running auto maintenance In the past, we used to execute `git gc --auto` as part of our automatic housekeeping routines. As git-gc(1) may require quite some time to perform the housekeeping, it knows to detach itself and run in the background so that the user can continue their work. Eventually, we refactored our automatic housekeeping to instead use the more flexible git-maintenance(1) command. The upside of this new infra is that the user can configure which maintenance tasks are performed, at least to a certain degree. So while it continues to run git-gc(1) by default, it can also be adapted to e.g. use git-multi-pack-index(1) for maintenance of the object database. The auto-detach of the new infra is somewhat broken though once the user configures non-standard tasks. The problem is essentially that we detach at the wrong level in the process hierarchy: git-maintenance(1) never detaches itself, but instead it continues to be git-gc(1) which does. When configured to only run the git-gc(1) maintenance task, then the result is basically the same as before. But when configured to run other tasks, then git-maintenance(1) will wait for these to run to completion. Even worse, it may be that git-gc(1) runs concurrently with other housekeeping tasks, stomping on each others feet. Fix this bug by asking git-gc(1) to not detach when it is being invoked via git-maintenance(1). Instead, git-maintenance(1) now respects a new config "maintenance.autoDetach", the equivalent of "gc.autoDetach", and detaches itself into the background when running as part of our auto maintenance. This should continue to behave the same for all users which use the git-gc(1) task, only. For others though, it means that we now properly perform all tasks in the background. The default behaviour of git-maintenance(1) when executed by the user does not change, it will remain in the foreground unless they pass the `--detach` option. Signed-off-by: Patrick Steinhardt <ps@pks.im> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2024-08-16 18:45:17 +08:00
/*
* When `maintenance.autoDetach` isn't set, then we fall back to
* honoring `gc.autoDetach`. This is somewhat weird, but required to
* retain behaviour from when we used to run git-gc(1) here.
*/
if (git_config_get_bool("maintenance.autodetach", &auto_detach) &&
git_config_get_bool("gc.autodetach", &auto_detach))
auto_detach = 1;
maint->git_cmd = 1;
maint->close_object_store = 1;
strvec_pushl(&maint->args, "maintenance", "run", "--auto", NULL);
strvec_push(&maint->args, quiet ? "--quiet" : "--no-quiet");
run-command: fix detaching when running auto maintenance In the past, we used to execute `git gc --auto` as part of our automatic housekeeping routines. As git-gc(1) may require quite some time to perform the housekeeping, it knows to detach itself and run in the background so that the user can continue their work. Eventually, we refactored our automatic housekeeping to instead use the more flexible git-maintenance(1) command. The upside of this new infra is that the user can configure which maintenance tasks are performed, at least to a certain degree. So while it continues to run git-gc(1) by default, it can also be adapted to e.g. use git-multi-pack-index(1) for maintenance of the object database. The auto-detach of the new infra is somewhat broken though once the user configures non-standard tasks. The problem is essentially that we detach at the wrong level in the process hierarchy: git-maintenance(1) never detaches itself, but instead it continues to be git-gc(1) which does. When configured to only run the git-gc(1) maintenance task, then the result is basically the same as before. But when configured to run other tasks, then git-maintenance(1) will wait for these to run to completion. Even worse, it may be that git-gc(1) runs concurrently with other housekeeping tasks, stomping on each others feet. Fix this bug by asking git-gc(1) to not detach when it is being invoked via git-maintenance(1). Instead, git-maintenance(1) now respects a new config "maintenance.autoDetach", the equivalent of "gc.autoDetach", and detaches itself into the background when running as part of our auto maintenance. This should continue to behave the same for all users which use the git-gc(1) task, only. For others though, it means that we now properly perform all tasks in the background. The default behaviour of git-maintenance(1) when executed by the user does not change, it will remain in the foreground unless they pass the `--detach` option. Signed-off-by: Patrick Steinhardt <ps@pks.im> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2024-08-16 18:45:17 +08:00
strvec_push(&maint->args, auto_detach ? "--detach" : "--no-detach");
return 1;
}
int run_auto_maintenance(int quiet)
{
struct child_process maint = CHILD_PROCESS_INIT;
if (!prepare_auto_maintenance(quiet, &maint))
return 0;
return run_command(&maint);
}
void prepare_other_repo_env(struct strvec *env, const char *new_git_dir)
{
const char * const *var;
for (var = local_repo_env; *var; var++) {
if (strcmp(*var, CONFIG_DATA_ENVIRONMENT) &&
strcmp(*var, CONFIG_COUNT_ENVIRONMENT))
strvec_push(env, *var);
}
strvec_pushf(env, "%s=%s", GIT_DIR_ENVIRONMENT, new_git_dir);
}
enum start_bg_result start_bg_command(struct child_process *cmd,
start_bg_wait_cb *wait_cb,
void *cb_data,
unsigned int timeout_sec)
{
enum start_bg_result sbgr = SBGR_ERROR;
int ret;
int wait_status;
pid_t pid_seen;
time_t time_limit;
/*
* We do not allow clean-on-exit because the child process
* should persist in the background and possibly/probably
* after this process exits. So we don't want to kill the
* child during our atexit routine.
*/
if (cmd->clean_on_exit)
BUG("start_bg_command() does not allow non-zero clean_on_exit");
if (!cmd->trace2_child_class)
cmd->trace2_child_class = "background";
ret = start_command(cmd);
if (ret) {
/*
* We assume that if `start_command()` fails, we
* either get a complete `trace2_child_start() /
* trace2_child_exit()` pair or it fails before the
* `trace2_child_start()` is emitted, so we do not
* need to worry about it here.
*
* We also assume that `start_command()` does not add
* us to the cleanup list. And that it calls
* `child_process_clear()`.
*/
sbgr = SBGR_ERROR;
goto done;
}
time(&time_limit);
time_limit += timeout_sec;
wait:
pid_seen = waitpid(cmd->pid, &wait_status, WNOHANG);
if (!pid_seen) {
/*
* The child is currently running. Ask the callback
* if the child is ready to do work or whether we
* should keep waiting for it to boot up.
*/
ret = (*wait_cb)(cmd, cb_data);
if (!ret) {
/*
* The child is running and "ready".
*/
trace2_child_ready(cmd, "ready");
sbgr = SBGR_READY;
goto done;
} else if (ret > 0) {
/*
* The callback said to give it more time to boot up
* (subject to our timeout limit).
*/
time_t now;
time(&now);
if (now < time_limit)
goto wait;
/*
* Our timeout has expired. We don't try to
* kill the child, but rather let it continue
* (hopefully) trying to startup.
*/
trace2_child_ready(cmd, "timeout");
sbgr = SBGR_TIMEOUT;
goto done;
} else {
/*
* The cb gave up on this child. It is still running,
* but our cb got an error trying to probe it.
*/
trace2_child_ready(cmd, "error");
sbgr = SBGR_CB_ERROR;
goto done;
}
}
else if (pid_seen == cmd->pid) {
int child_code = -1;
/*
* The child started, but exited or was terminated
* before becoming "ready".
*
* We try to match the behavior of `wait_or_whine()`
* WRT the handling of WIFSIGNALED() and WIFEXITED()
* and convert the child's status to a return code for
* tracing purposes and emit the `trace2_child_exit()`
* event.
*
* We do not want the wait_or_whine() error message
* because we will be called by client-side library
* routines.
*/
if (WIFEXITED(wait_status))
child_code = WEXITSTATUS(wait_status);
else if (WIFSIGNALED(wait_status))
child_code = WTERMSIG(wait_status) + 128;
trace2_child_exit(cmd, child_code);
sbgr = SBGR_DIED;
goto done;
}
else if (pid_seen < 0 && errno == EINTR)
goto wait;
trace2_child_exit(cmd, -1);
sbgr = SBGR_ERROR;
done:
child_process_clear(cmd);
invalidate_lstat_cache();
return sbgr;
}