Before this commit, DNS timeouts always used CLOCK_REALTIME, which
could produce spurious timeouts or delays if wall time changed for
whatever reason.
Now we try CLOCK_MONOTONIC and only fall back to CLOCK_REALTIME when
it is unavailable.
As a result of using simple subtraction to implement the return values
for wcscmp and wcsncmp, integer overflow can occur (producing
undefined behavior, and in practice, a wrong comparison result). This
does not occur for meaningful character values (21-bit range) but the
functions are specified to work on arbitrary wchar_t arrays.
This patch replaces the subtraction with a little bit of code that
orders the characters correctly, returning -1 if the character from
the first string is smaller than the one from the second, 0 if they
are equal and 1 if the character from the first string is larger than
the one from the second.
A signed int shift overflowed when computing a constant mask, use hex
literal instead. This is unlikely to cause actual issues unless the
code was compiled with ubsan or similar instrumentation specifically
to catch this. The stripped libc.so is unchanged on x86_64.
Reported by q66 on irc.
When a dot is encountered, the loop counter is incremented before
exiting the loop, but the corresponding ip array element is left
uninitialized, so the subsequent memmove (if "::" was seen) and the
loop copying ip to the output buffer will operate on an uninitialized
uint16_t.
The uninitialized data never directly influences the control flow and
is overwritten on successful return by the second half of the parsed
IPv4 address. But it's better to fix this to avoid unexpected
transformations by a sufficiently smart compiler and reports from
UB-detection tools.
The kernel defines a limit on the number of fds that can be passed
through an SCM_RIGHTS ancillary message as SCM_MAX_FD. The value was
255 before kernel 2.6.38 (after that it is 253), and an SCM_RIGHTS
ancillary message with 255 fds requires 1040 bytes, slightly more than
the current 1024 byte internal buffer in sendmsg. 1024 is an arbitrary
size, so increase it to match the the arbitrary size limit in the
kernel. This fixes tests that are verifying they support up to
SCM_MAX_FD fds.
until the mq notification event arrives, it is mandatory that signals
be blocked. otherwise, a signal can be received, and its handler
executed, in a thread which does not yet exist on the abstract
machine.
after the point of the event arriving, having signals blocked is not a
conformance requirement but a QoI requirement. while the application
can unblock any signals it wants unblocked in the event handler
thread, if they did not start out blocked, it could not block them
without a race window where they are momentarily unblocked, and this
would preclude controlled delivery or other forms of acceptance
(sigwait, etc.) anywhere in the application.
in the error path where the mq_notify syscall fails, the initiating
thread may have closed the socket before the worker thread calls recv
on it. even in the absence of such a race, if the recv call failed,
e.g. due to seccomp policy blocking it, the worker thread could
proceed to close, producing a double-close condition.
this can all be simplified by moving the mq_notify syscall into the
new thread, so that the error case does not require pthread_cancel.
now, the initiating thread only needs to read back the error status
after waiting for the worker thread to consume its arguments.
disabling cancellation around the pthread_join call seems to be the
safest and logically simplest fix. i believe it would also be possible
to just perform the unmap directly here after __tl_sync, removing the
dependency on pthread_join, but such an approach duplicately encodes a
lot more implementation assumptions.
the logic to check hwcap for SPE register file inadvertently clobbered
the val argument before use. switch to a different work register so
this doesn't happen.
we wrongly defined a dummy SA_RESTORER flag on these archs, despite
the kernel interface not actually having such a feature. on archs
which lack SA_RESTORER, the kernel sigaction structure also lacks the
restorer function pointer member, which means the signal mask appears
at a different offset. the kernel was thereby interpreting the bits of
the code address as part of the signal set to be masked while handling
the signal.
this patch removes the erroneous SA_RESTORER definitions from archs
which do not have it, makes access to the member conditional on
whether SA_RESTORER is defined for the arch, and removes the
now-unused asm for the affected archs.
because there are reportedly versions of qemu-user which also use the
wrong ABI here, the old ksigaction struct size is preserved with an
unused member at the end. this is harmless and mitigates the risk of
such a bug turning into a buffer overflow onto the sigaction
function's stack.
the result of the 0xffff mask with the exit status could have bit 15
set, in which case multiplying by 0x10001 overflows 32-bit signed int.
making the multiply unsigned avoids the overflow. it also changes the
sign extension behavior of the subsequent >> operation, but the
affected bits are all unwanted anyway and all discarded by the cast to
short.
mips has its own mechanisms for DT_DEBUG because it makes _DYNAMIC
read-only, and the original mechanism, DT_MIPS_RLD_MAP, was
PIE-incompatible. DT_MIPS_RLD_MAP_REL was added to remedy this, but we
never implemented support for it. add it now using the same idioms for
mips-specific ldso logic.
memmem has been adopted for the next issue of POSIX (outcome of
tracker item 1061). since mem* is in the reserved namespace for
string.h it's already fully conforming to expose it by default, so
just do so.
while no lock is held here making it a lock-order issue, replacement
malloc is likely to want to use pthread_atfork, possibly making the
call to malloc infinitely recursive.
even if not, there is no reason to prefer an application-provided
malloc here.
printf_core() runs twice, and during its first run, nl_arg is
uninitialized and must not be read. It gets initialized at the end of
the first run. Conversely, nl_type does not need to be set during the
second run, as its useful life has ended at that point, since the only
time it is read is during that exact same initialization. Therefore we
can simply alternate the assignments.
p and w do still need to get values assigned to them, since at least one
line in the same if-statement depends on that, but they can be dummy
values. arg does not need to be assigned, since in the first run, we
encounter a continue statement before using the argument.
because the has-waiters state in the semaphore value futex word is
only representable when the value is zero (the special value -1
represents "0 with potential new waiters"), it's lost if intervening
operations make the semaphore value positive again. this creates an
ABA issue in sem_post, whereby the post uses a stale waiters count
rather than re-evaluating it, skipping the futex wake if the stale
count was zero.
the fix here is based on a proposal by Alexey Izbyshev, with minor
changes to eliminate costly new spurious wake syscalls.
the basic idea is to replace the special value -1 with a sticky
waiters bit (repurposing the sign bit) preserved under both wait and
post. any post that takes place with the waiters bit set will perform
a futex wake.
to be useful, the waiters bit needs to be removable, and to remove it
safely, we perform a broadcast wake instead of a normal single-task
wake whenever removing the bit. this lets any un-accounted-for waiters
wake and re-add the waiters bit if they still need it.
there are multiple possible choices for when to perform this
broadcast, but the optimal choice seems to be doing it whenever the
observed waiters count is less than two (semantically, this means
exactly one, but we might see a stale count of zero). in this case,
the expected number of threads to be woken is one, with exactly the
same cost as a non-broadcast wake.
when PAGE_SIZE is not constant, internal/libc.h defines it to expand
to libc.page_size. however, kernel_mapped_dso, reachable from stage 2
of the dynamic linker bootstrap (__dls2), needs PAGE_SIZE to interpret
the relro range. at this point the libc object is both uninitialized
and invalid to access according to our model for bootstrapping, which
does not assume any external-linkage objects are accessible until
stages 2b/3. in practice it likely worked because hidden visibility
tends to behave like internal linkage, but this is not a property that
the dynamic linker was designed to rely upon.
this bug likely manifested as relro malfunction on archs with variable
page size, due to incorrect mask when aligning the relro bounds to
page boundaries.
while there are certainly more direct ways to fix the known problem
point here, a maximally future-proof way is to just bypass the libc.h
PAGE_SIZE definition in the dynamic linker and instead have dynlink.c
define its own internal-linkage object for variable page size. then,
if anything else in stage 2 ever ends up referencing PAGE_SIZE, it
will just automatically work right.
this is analogous to skip_relative logic in do_relocs -- because
relative relocations for the dynamic linker itself were already
performed at entry (stage 1), they must not be applied again.
the rule that longest digit sequence not beginning with a zero is
greater only applies when both sequences being compared are
non-degenerate. this is spelled out explicitly in the man page, which
may be deemed authoritative for this nonstandard function: "If one or
both of these is empty, then return what strcmp(3) would have
returned..."
we were wrongly treating any sequence of digits not beginning with a
zero as greater than a non-digit in the other string.
if async cancellation is enabled and acted upon, the stack pointer is
not necessarily pointing to a __syscall_cp_asm stack frame. the
contents of the stack being wrong don't really matter, but if the
stack pointer is not suitably aligned, the procedure call ABI is
violated when calling back into C code via __cancel, and pthread_exit,
cancellation cleanup handlers, TSD destructors, etc. may malfunction
or crash.
for the async cancel case, just call __cancel directly like we did
prior to commit 102f6a01e2. restore the
signal mask prior to doing this since the cancellation handler runs
with all signals blocked.
commit f081d5336a fixed
gethostbyname[2]_r to treat negative results as a non-error, leaving
gethostbyname[2] wrongly returning a pointer to the unfilled result
buffer rather than a null pointer. since, as documented with commit
fe82bb9b92, the caller of
gethostby{name[2],addr}_r can always rely on the result pointer being
set, use that consistently rather than trying to duplicate logic about
whether we have a result or not in gethostby{name[2],addr}.
the only functional change here should be that MAXADDRS is only
checked for RRs that provide address results, so that a CNAME which
appears after an excessive number of address RRs does not get ignored.
I'm not aware of any servers that order the RRs this way, and it may
even be forbidden to do so, but I prefer having the callback logic not
be order dependent.
other than that, the motivation for this change is that the A and AAAA
cases were mostly duplicate code that could be combined as a single
code path.
returning -1 rather than 0 from the parse function causes __dns_parse
to bail out and return an error. presently, name_from_dns does not
check the return value anyway, so this does not matter, but if it ever
started treating this as an error, lookups with large numbers of
addresses would break. this is a consequence of adding TCP support and
extending the buffer size used in name_from_dns.
reportedly there is nameserver software with question-rewriting
"functionality" which gives A answers when AAAA is queried. since we
made no effort to validate that the answer RR type actually
corresponds to the question asked, it was possible (depending on
flags, etc.) for these answers to leak through, which the caller might
not be prepared for. indeed, our implementation of gethostbyname2_r
makes an assumption that the resulting addresses are in the family
requested, and will misinterpret the results if they don't.
commit 45ca5d3fcb already noted in
fixing CVE-2017-15650 that this could happen, but did nothing to
validate that the RR type of the answer matches the question; it just
enforced the limit on number of results to preclude overflow.
presently, name_from_dns ignores the return value of __dns_parse, so
it doesn't really matter whether we return 0 (ignoring the RR) or -1
(parse-ending error) upon encountering the mismatched RR. if that ever
changes, though, ignoring irrelevant answer RRs sounds like the
semantically correct thing to do, so for now let's return 0 from the
callback when this happens.
commit 167390f055 seems to have
overlooked the presence of a lock here, probably because it was one of
the exceptions not using LOCK() but a rwlock.
as such, it can't be added to the generic table of locks to take, so
add an explicit atfork function for the pthread keys table. the order
it is called does not particularly matter since nothing else in libc
but pthread_exit interacts with keys.
performing n-- is not a safe operation for arbitrary signed input n.
only perform the decrement in the code path where the initial n is
greater than 1, and adjust the condition in the n<=1 code path to
compensate for it not having been decremented.
the aio operations that lead to calling __aio_get_queue with the
possibility to expand the fd map are not AS-safe, but if they are
interrupted by a signal handler, the signal handler may call close,
which is required to be AS-safe. due to __aio_get_queue taking the
write lock without blocking signals, such a call to close from a
signal handler could deadlock.
change __aio_get_queue to block signals if it needs to obtain a write
lock, and restore when finished.
aio_suspend waits on a dummy futex in the corner case when the array of
requests contains NULL pointers only. But the value of this futex was
left uninitialized, so if it happens to be non-zero, aio_suspend
degrades to spinning instead of blocking.
as reported by Alexey Izbyshev, there is a lock order inversion
deadlock between the malloc lock and aio maplock at MT-fork time:
_Fork attempts to take the aio maplock while fork already has the
malloc lock, but a concurrent aio operation holding the maplock may
attempt to allocate memory.
move the __aio_atfork calls in the parent from _Fork to fork, and
reorder the lock before most other locks, since nothing else depends
on aio(*). this leaves us with the possibility that the child will not
be able to obtain the read lock, if _Fork is used directly and happens
concurrent with an aio operation. however, in that case, the child
context is an async signal context that cannot call any further aio
functions, so all we need is to ensure that close does not attempt to
perform any aio cancellation. this can be achieved just by nulling out
the map pointer.
(*) even if other functions call close, they will only need a read
lock, not a write lock, and read locks being recursive ensures they
can obtain it. moreover, the number of read references held is bounded
by something like twice the number of live threads, meaning that the
read lock count cannot saturate.
as reported by Alexey Izbyshev, when the second-to-last thread exits
causing a return to single-threaded (no locks needed) state, it
creates a situation where the last remaining thread may obtain the
killlock that's already held by the exiting thread. this means it may
erroneously use the tid of the exiting thread, and may corrupt the
lock state due to double-unlock.
commit 8d81ba8c0b, which (re)introduced
the switch back to single-threaded state, documents the intent that
the first lock after switching back should provide the necessary
synchronization. this is correct, but only works if the switch back is
made after there is no further need for synchronization with locks
(other than the thread list lock, which can't be bypassed) held by the
exiting thread.
in order to hit the bug, the remaining thread must first take a
different lock, causing it to perform an actual lock one last time,
consume the need_locks==-1 state, and transition to need_locks==0.
after that, the next attempt to lock the exiting thread's killlock
will bypass locking.
fix this by reordering the unlocking of killlock at thread exit time,
along with changes to the state protected by it, to occur earlier,
before the switch to single-threaded state. there are really no
constraints on where it's done, except that it occur after there is no
longer any possibility of application code executing in the exiting
thread, so do it as early as possible.
ever since commit 8f11e6127f introduced
the thread list lock, this has been wrong. initially, it was wrong via
calling free from the context with the thread list lock held. commit
aa5a9d15e0 deferred the unsafe free but
added a lock, which was also unsafe. in particular, it could deadlock
if code holding freebuf_queue_lock was interrupted by a signal handler
that takes the thread list lock.
commit 4d5aa20a94 observed that there
was a lock here but failed to notice that it's invalid.
there is no easy solution to this problem with locks; any attempt at
solving it while still using locks would require the lock to be an
AS-safe one (blocking signals on each access to the dlerror buffer
list to check if there's deferred free work to be done) which would be
excessively costly, and there are also lock order considerations with
respect to how the lock would be handled at fork.
instead, just use an atomic list.
unlike most projects that use -fno-strict-aliasing, we aim to have all
sources respect the C language rules for effective type that make
type-based alias analysis optimizations possible. unfortunately, it
turns out that there are deep, and likely very difficult to fix, flaws
in the TBAA performed by GCC and likely other compilers, whereby this
kind of optimization can transform code that follows the rules
strictly in ways that will make it malfunction. see for example GCC
bugs 107107 and 107115, the latter of which also affects clang.
there are not presently any known instances of breakage due to wrong
type-based aliasing optimizations in our codebase. nonetheless, since
the transformations are unsound and could introduce breakage,
configure CFLAGS to build with -fno-strict-aliasing.
some casual analysis of the effects on codegen suggest that this is
unlikely to affect performance except possibly in the regex engine. in
general, we should probably prefer making better use of the restrict
keyword over relying on types to imply non-aliasing for optimization
purposes; doing so should be able to get back any performance that was
lost and more, should it turn out to matter (unlikely).
the entire intent of using madvise/MADV_FREE on freed slots is to
improve system performance by avoiding evicting cache of useful data,
or swapping useless data to disk, by marking any whole pages in the
freed slot as discardable by the kernel. in particular, unlike
unmapping the memory or replacing it with a PROT_NONE region, use of
MADV_FREE does not make any difference to memory accounting for commit
charge purposes, and so does not increase the memory available to
other processes in a non-overcommitted environment.
however, various measurements have shown that inordinate amounts of
time are spent performing madvise syscalls in processes which
frequently allocate and free medium sized objects in the size range
roughly between PAGESIZE and MMAP_THRESHOLD, to the point that the net
effect is almost surely significant performance degredation. so, turn
it off.
the code, which has some nontrivial logic for efficiently determining
whether there is a whole-page range to apply madvise to, is left in
place so that it can easily be re-enabled if desired, or later tuned
to only apply to certain sizes or to use additional heuristics.
these badly pollute the namespace with macros whenever _GNU_SOURCE is
defined, which is always the case with g++, and especially tends to
interfere with C++ constructs.
as our implementation of these was macro-only, their removal cannot
affect any existing binaries. at the source level, portable software
should be prepared for them not to exist.
for now, they are left in place with explicit _LARGEFILE64_SOURCE.
this provides an easy temporary path for integrators/distributions to
get packages building again right away if they break while working on
a proper, upstreamable fix. the intent is that this be a very
short-term measure and that the macros be removed entirely in the next
release cycle.
originally the namespace-infringing "large file support" interfaces
were included as part of glibc-ABI-compat, with the intent that they
not be used for linking, since our off_t is and always has been
unconditionally 64-bit and since we usually do not aim to support
nonstandard interfaces when there is an equivalent standard interface.
unfortunately, having the symbols present and available for linking
caused configure scripts to detect them and attempt to use them
without declarations, producing all the expected ill effects that
entails.
as a result, commit 2dd8d5e1b8 was made
to prevent this, using macros to redirect the LFS64 names to the
standard names, conditional on _GNU_SOURCE or _LARGEFILE64_SOURCE.
however, this has turned out to be a source of further problems,
especially since g++ defines _GNU_SOURCE by default. in particular,
the presence of these names as macros breaks a lot of valid code.
this commit removes all the LFS64 symbols and replaces them with a
mechanism in the dynamic linker symbol lookup failure path to retry
with the spurious "64" removed from the symbol name. in the future,
if/when the rest of glibc-ABI-compat is moved out of libc, this can be
removed.
we already attempt to preclude this case by having res_send use a
sufficiently large temporary buffer even if the caller did not provide
one as large as or larger than the udp dns max of 512 bytes. however,
it's possible that the caller passed a custom-crafted query packet
using EDNS0, e.g. to get detailed DNSSEC results, with a larger udp
size allowance.
I have also seen claims that there are some broken nameservers in the
wild that do not honor the dns udp limit of 512 and send large answers
without the TC bit set, when the query was not using EDNS.
we generally don't aim to support broken nameservers, but in this case
both problems, if the latter is even real, have a common solution:
using recvmsg instead of recvfrom so we can examine the MSG_TRUNC
flag.
the size of 512 is not sufficient to get at least one address in the
worst case where the name is at or near max length and resolves to a
CNAME at or near max length. prior to tcp fallback, there was nothing
we could do about this case anyway, but now it's fixable.
the new limit 768 is chosen so as to admit roughly the number of
addresses with a worst-case CNAME as could fit for a worst-case name
that's not a CNAME in the old 512-byte limit. outside of this
worst-case, the number of addresses that might be obtained is
increased.
MAXADDRS (48) was originally chosen as an upper bound on the combined
number of A and AAAA records that could fit in 512-byte packets (31
and 17, respectively). it is not increased at this time.
so as to prevent a situation where the A records consume almost all of
these slots (at 768 bytes, a "best-case" name can fit almost 47 A
records), the order of parsing is swapped to process AAAA first. this
ensures roughly half of the slots are available to each address
family.
tcp fallback was originally deemed unwanted and unnecessary, since we
aim to return a bounded-size result from getaddrinfo anyway and
normally plenty of address records fit in the 512-byte udp dns limit.
however, this turned out to have several problems:
- some recursive nameservers truncate by omitting all the answers,
rather than sending as many as can fit.
- a pathological worst-case CNAME for a worst-case name can fill the
entire 512-byte space with just the two names, leaving no room for
any addresses.
- the res_* family of interfaces allow querying of non-address records
such as TLSA (DANE), TXT, etc. which can be very large. for many of
these, it's critical that the caller see the whole RRset. also,
res_send/res_query are specified to return the complete, untruncated
length so that the caller can retry with an appropriately-sized
buffer. determining this is not possible without tcp.
so, it's time to add tcp fallback.
the fallback strategy implemented here uses one tcp socket per
question (1 or 2 questions), initiated via tcp fastopen when possible.
the connection is made to the nameserver that issued the truncated
answer. right now, fallback happens unconditionally when truncation is
seen. this can, and may later be, relaxed for queries made by the
getaddrinfo system, since it will only use a bounded number of results
anyway.
retry is not attempted again after failure over tcp. the logic could
easily be adapted to do that, but it's of questionable value, since
the tcp stack automatically handles retransmission and the successs
answer with TC=1 over udp strongly suggests that the nameserver has
the full answer ready to give. further retry is likely just "take
longer to fail".
for extremely small buffer sizes, the DNS query core in __res_msend
may malfunction completely, being unable to get even the headers to
determine the response code. but there is also a problem for
reasonable sizes under 512 bytes: __res_msend is unable to determine
if the udp answer was truncated at the recv layer, in which case it
may be incomplete, and res_send is then unable to honor its contract
to return the length of the full, non-truncated answer.
at present, res_send does not honor that contract anyway when the full
answer would exceed 512 bytes, since there is no tcp fallback, but
this change at least makes it consistent in a context where this is
the only "full answer" to be had.
this was apparently omitted long ago out of a lack of understanding of
its importance and the fact that POSIX doesn't specify it. despite not
being officially standardized, however, it turns out that at least
AIX, glibc, NetBSD, OpenBSD, QNX, and Solaris document and support it.
in certain usage cases, such as implementing a DNS gateway on top of
the stub resolver interfaces, it's necessary to distinguish the case
where a name does not exit (NxDomain) from one where it exists but has
no addresses (or other records) of the requested type (NODATA). in
fact, even the legacy gethostbyname API had this distinction, which we
were previously unable to support correctly because the backend lacked
it.
apart from fixing an important functionality gap, adding this
distinction helps clarify to users how search domain fallback works
(falling back in cases corresponding to EAI_NONAME, not in ones
corresponding to EAI_NODATA), a topic that has been a source of
ongoing confusion and frustration.
as a result of this change, EAI_NONAME is no longer a valid universal
error code for getaddrinfo in the case where AI_ADDRCONFIG has
suppressed use of all address families. in order to return an accurate
result in this case, getaddrinfo is modified to still perform at least
one lookup. this will almost surely fail (with a network error, since
there is no v4 or v6 network to query DNS over) unless a result comes
from the hosts file or from ip literal parsing, but in case it does
succeed, the result is replaced by EAI_NODATA.
glibc has a related error code, EAI_ADDRFAMILY, that could be used for
the AI_ADDRCONFIG case and certain NODATA cases, but distinguishing
them properly in full generality seems to require additional DNS
queries that are otherwise not useful. on glibc, it is only used for
ip literals with mismatching family, not for DNS or hosts file results
where the name has addresses only in the opposite family. since this
seems misleading and inconsistent, and since EAI_NODATA already covers
the semantic case where the "name" exists but doesn't have any
addresses in the requested family, we do not adopt EAI_ADDRFAMILY at
this time. this could be changed at some point if desired, but the
logic for getting all the corner cases with AI_ADDRCONFIG right is
slightly nontrivial.
EAI_MEMORY is not possible (but would not provide errno if it were)
and EAI_FAIL does not provide errno. treat the latter as EBADMSG to
match how it's handled in gethostbyname2_r (it indicates erroneous or
failure response from the nameserver).
