We were using the frame pointer based stack walker on every
contexts in x86-32, but not in x86-64 where we only use the
seven-league boots on the exception stacks.
Use it also on irq and process stacks. This utterly accelerate
the captures.
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Callers of a stacktrace might pass bad frame pointers. Those
are usually checked for safety in stack walking helpers before
any dereferencing, but this is not the case when we need to go
through one more frame pointer that backlinks the irq stack to
the previous one, as we don't have any reliable address boudaries
to compare this frame pointer against.
This raises crashes when we record callchains for ftrace events
with perf because we don't use the right helpers to capture
registers there. We get wrong frame pointers as we call
task_pt_regs() even on kernel threads, which is a wrong thing
as it gives us the initial state of any kernel threads freshly
created. This is even not what we want for user tasks. What we want
is a hot snapshot of registers when the ftrace event triggers, not
the state before a task entered the kernel.
This requires more thoughts to do it correctly though.
So first put a guardian to ensure the given frame pointer
can be dereferenced to avoid crashes. We'll think about how to fix
the callers in a subsequent patch.
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: 2.6.33.x <stable@kernel.org>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Print modules list during kernel BUG.
Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The check that ignores the debug and nmi stack frames is useless
now that we have a frame pointer that makes us start at the
right place. We don't anymore have to deal with these.
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Paul Mackerras <paulus@samba.org>
LKML-Reference: <1262235183-5320-2-git-send-regression-fweisbec@gmail.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The current print_context_stack helper that does the stack
walking job is good for usual stacktraces as it walks through
all the stack and reports even addresses that look unreliable,
which is nice when we don't have frame pointers for example.
But we have users like perf that only require reliable
stacktraces, and those may want a more adapted stack walker, so
lets make this function a callback in stacktrace_ops that users
can tune for their needs.
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Paul Mackerras <paulus@samba.org>
LKML-Reference: <1261024834-5336-1-git-send-regression-fweisbec@gmail.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
* 'perf-fixes-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/linux-2.6-tip: (57 commits)
x86, perf events: Check if we have APIC enabled
perf_event: Fix variable initialization in other codepaths
perf kmem: Fix unused argument build warning
perf symbols: perf_header__read_build_ids() offset'n'size should be u64
perf symbols: dsos__read_build_ids() should read both user and kernel buildids
perf tools: Align long options which have no short forms
perf kmem: Show usage if no option is specified
sched: Mark sched_clock() as notrace
perf sched: Add max delay time snapshot
perf tools: Correct size given to memset
perf_event: Fix perf_swevent_hrtimer() variable initialization
perf sched: Fix for getting task's execution time
tracing/kprobes: Fix field creation's bad error handling
perf_event: Cleanup for cpu_clock_perf_event_update()
perf_event: Allocate children's perf_event_ctxp at the right time
perf_event: Clean up __perf_event_init_context()
hw-breakpoints: Modify breakpoints without unregistering them
perf probe: Update perf-probe document
perf probe: Support --del option
trace-kprobe: Support delete probe syntax
...
When we enter in irq, two things can happen to preserve the link
to the previous frame pointer:
- If we were in an irq already, we don't switch to the irq stack
as we are inside. We just need to save the previous frame
pointer and to link the new one to the previous.
- Otherwise we need another level of indirection. We enter the irq with
the previous stack. We save the previous bp inside and make bp
pointing to its saved address. Then we switch to the irq stack and
push bp another time but to the new stack. This makes two levels to
dereference instead of one.
In the second case, the current stacktrace code omits the second level
and loses the frame pointer accuracy. The stack that follows will then
be considered as unreliable.
Handling that makes the perf callchain happier.
Before:
43.94% [k] _raw_read_lock
|
--- _read_lock
|
|--60.53%-- send_sigio
| __kill_fasync
| kill_fasync
| evdev_pass_event
| evdev_event
| input_pass_event
| input_handle_event
| input_event
| synaptics_process_byte
| psmouse_handle_byte
| psmouse_interrupt
| serio_interrupt
| i8042_interrupt
| handle_IRQ_event
| handle_edge_irq
| handle_irq
| __irqentry_text_start
| ret_from_intr
| |
| |--30.43%-- __select
| |
| |--17.39%-- 0x454f15
| |
| |--13.04%-- __read
| |
| |--13.04%-- vread_hpet
| |
| |--13.04%-- _xcb_lock_io
| |
| --13.04%-- 0x7f630878ce8
After:
50.00% [k] _raw_read_lock
|
--- _read_lock
|
|--98.97%-- send_sigio
| __kill_fasync
| kill_fasync
| evdev_pass_event
| evdev_event
| input_pass_event
| input_handle_event
| input_event
| |
| |--96.88%-- synaptics_process_byte
| | psmouse_handle_byte
| | psmouse_interrupt
| | serio_interrupt
| | i8042_interrupt
| | handle_IRQ_event
| | handle_edge_irq
| | handle_irq
| | __irqentry_text_start
| | ret_from_intr
| | |
| | |--39.78%-- __const_udelay
| | | |
| | | |--91.89%-- ath5k_hw_register_timeout
| | | | ath5k_hw_noise_floor_calibration
| | | | ath5k_hw_reset
| | | | ath5k_reset
| | | | ath5k_config
| | | | ieee80211_hw_config
| | | | |
| | | | |--88.24%-- ieee80211_scan_work
| | | | | worker_thread
| | | | | kthread
| | | | | child_rip
| | | | |
| | | | --11.76%-- ieee80211_scan_completed
| | | | ieee80211_scan_work
| | | | worker_thread
| | | | kthread
| | | | child_rip
| | | |
| | | --8.11%-- ath5k_hw_noise_floor_calibration
| | | ath5k_hw_reset
| | | ath5k_reset
| | | ath5k_config
Note: This does not only affect perf events but also x86-64
stacktraces. They were considered as unreliable once we quit
the irq stack frame.
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: "K. Prasad" <prasad@linux.vnet.ibm.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Make the initialization more readable, plus tidy up a few small
visual details as well.
No change in functionality.
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
* remove asm/atomic.h inclusion from linux/utsname.h --
not needed after kref conversion
* remove linux/utsname.h inclusion from files which do not need it
NOTE: it looks like fs/binfmt_elf.c do not need utsname.h, however
due to some personality stuff it _is_ needed -- cowardly leave ELF-related
headers and files alone.
Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
About every callchains recorded with perf record are filled up
including the internal perfcounter nmi frame:
perf_callchain
perf_counter_overflow
intel_pmu_handle_irq
perf_counter_nmi_handler
notifier_call_chain
atomic_notifier_call_chain
notify_die
do_nmi
nmi
We want ignore this frame as it's not interesting for
instrumentation. To solve this, we simply ignore every frames
from nmi context.
New example of "perf report -s sym -c" after this patch:
9.59% [k] search_by_key
4.88%
search_by_key
reiserfs_read_locked_inode
reiserfs_iget
reiserfs_lookup
do_lookup
__link_path_walk
path_walk
do_path_lookup
user_path_at
vfs_fstatat
vfs_lstat
sys_newlstat
system_call_fastpath
__lxstat
0x406fb1
3.19%
search_by_key
search_by_entry_key
reiserfs_find_entry
reiserfs_lookup
do_lookup
__link_path_walk
path_walk
do_path_lookup
user_path_at
vfs_fstatat
vfs_lstat
sys_newlstat
system_call_fastpath
__lxstat
0x406fb1
[...]
For now this patch only solves the problem in x86-64.
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Anton Blanchard <anton@samba.org>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
LKML-Reference: <1246474930-6088-1-git-send-email-fweisbec@gmail.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Move the irqstackptr variable from the PDA to per-cpu. Make the
stacks themselves per-cpu, removing some specific allocation code.
Add a seperate flag (is_boot_cpu) to simplify the per-cpu boot
adjustments.
tj: * sprinkle some underbars around.
* irq_stack_ptr is not used till traps_init(), no reason to
initialize it early. On SMP, just leaving it NULL till proper
initialization in setup_per_cpu_areas() works. Dropped
is_boot_cpu and early irq_stack_ptr initialization.
* do DECLARE/DEFINE_PER_CPU(char[IRQ_STACK_SIZE], irq_stack)
instead of (char, irq_stack[IRQ_STACK_SIZE]).
Signed-off-by: Brian Gerst <brgerst@gmail.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Impact: better dumpstack output
I noticed in my crash dumps and even in the stack tracer that a
lot of functions listed in the stack trace are simply
return_to_handler which is ftrace graphs way to insert its own
call into the return of a function.
But we lose out where the actually function was called from.
This patch adds in hooks to the dumpstack mechanism that detects
this and finds the real function to print. Both are printed to
let the user know that a hook is still in place.
This does give a funny side effect in the stack tracer output:
Depth Size Location (80 entries)
----- ---- --------
0) 4144 48 save_stack_trace+0x2f/0x4d
1) 4096 128 ftrace_call+0x5/0x2b
2) 3968 16 mempool_alloc_slab+0x16/0x18
3) 3952 384 return_to_handler+0x0/0x73
4) 3568 -240 stack_trace_call+0x11d/0x209
5) 3808 144 return_to_handler+0x0/0x73
6) 3664 -128 mempool_alloc+0x4d/0xfe
7) 3792 128 return_to_handler+0x0/0x73
8) 3664 -32 scsi_sg_alloc+0x48/0x4a [scsi_mod]
As you can see, the real functions are now negative. This is due
to them not being found inside the stack.
Signed-off-by: Steven Rostedt <srostedt@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Impact: cleanup
As promised, now that dumpstack_32 and dumpstack_64 have so many bits
in common, we should merge the in-sync bits into a common file, to
prevent them from diverging again.
This patch removes bits which are common between dumpstack_32.c and
dumpstack_64.c and places them in a common dumpstack.c which is built
for both 32 and 64 bit arches.
Signed-off-by: Neil Horman <nhorman@tuxdriver.com>
Acked-by: Alexander van Heukelum <heukelum@fastmail.fm>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Makefile | 2
arch/x86/kernel/Makefile | 2
arch/x86/kernel/Makefile | 2
arch/x86/kernel/Makefile | 2
arch/x86/kernel/Makefile | 2
arch/x86/kernel/Makefile | 2
arch/x86/kernel/dumpstack.c | 319 +++++++++++++++++++++++++++++++++++++++++
arch/x86/kernel/dumpstack.h | 39 +++++
arch/x86/kernel/dumpstack_32.c | 294 -------------------------------------
arch/x86/kernel/dumpstack_64.c | 285 ------------------------------------
5 files changed, 363 insertions(+), 576 deletions(-)
Make i386's die() equal to x86_64's version.
Whitespace-only changes on x86_64, to make it equal to i386's
version. (user_mode and user_mode_vm are equal on x86_64.)
Signed-off-by: Alexander van Heukelum <heukelum@fastmail.fm>
Acked-by: Neil Horman <nhorman@tuxdriver.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Use oops_begin and oops_end in die_nmi.
Whitespace-only changes on x86_64, to make it equal to i386's
version.
Signed-off-by: Alexander van Heukelum <heukelum@fastmail.fm>
Acked-by: Neil Horman <nhorman@tuxdriver.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Always call oops_exit from oops_end, even if signr==0.
Also, move add_taint(TAINT_DIE) from __die to oops_end
on x86_64 and interchange two lines to make oops_end
more similar to the i386-version.
Signed-off-by: Alexander van Heukelum <heukelum@fastmail.fm>
Acked-by: Neil Horman <nhorman@tuxdriver.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
oops_end is preceded by either a call to __die, or a conditional
call to crash_kexec. Move the conditional call to crash_kexec
from the end of __die to the start of oops_end and remove
the superfluous call to crash_kexec in die_nmi.
Signed-off-by: Alexander van Heukelum <heukelum@fastmail.fm>
Acked-by: Neil Horman <nhorman@tuxdriver.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Change oops_end such that signr=0 signals that do_exit
is not to be called.
Currently, each use of __die is soon followed by a call
to oops_end and 'regs' is set to NULL if oops_end is expected
not to call do_exit. Change all such pairs to set signr=0
instead. On x86_64 oops_end is used 'bare' in die_nmi; use
signr=0 instead of regs=NULL there, too.
Signed-off-by: Alexander van Heukelum <heukelum@fastmail.fm>
Acked-by: Neil Horman <nhorman@tuxdriver.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Print the name of the last-accessed sysfs file when we oops, to help track
down oopses which occur in sysfs store/read handlers. Because these oopses
tend to not leave any trace of the offending code in the stack traces.
Cc: Kay Sievers <kay.sievers@vrfy.org>
Cc: Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>
After "dumpstack: x86: various small unification steps", the
assembler gives the following compile error. The error is in
dumpstack_64.c.
{standard input}: Assembler messages:
{standard input}:720: Error: Incorrect register `%rbx' used with `l' suffix
{standard input}:1340: Error: Incorrect register `%r12' used with `l' suffix
Indeed the suffix in get_bp() was wrong.
Signed-off-by: Alexander van Heukelum <heukelum@fastmail.fm>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
- define STACKSLOTS_PER_LINE and use it
- define get_bp macro to hide the %%ebp/%%rbp difference
- i386: check task==NULL in dump_trace, like x86_64
- i386: show_trace(NULL, ...) uses current automatically
- x86_64: use [#%d] for die_counter, like i386
- whitespace and comments
Signed-off-by: Alexander van Heukelum <heukelum@fastmail.fm>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
- x86: Write log_lvl strings if available
- start raw stack dumps on new line
- i386: Remove extra indentation for raw stack dumps
Signed-off-by: Alexander van Heukelum <heukelum@fastmail.fm>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
- i386 and x86_64: always printk the 'data' parameter
- i386: announce stack switch (irq -> normal)
- i386: check if there is a stack switch before announcing it
There is a warning that 'context' might come out corrupt in early
boot. If this is true it should be fixed, not worked around.
Signed-off-by: Alexander van Heukelum <heukelum@fastmail.fm>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
- Add "end" parameter to valid_stack_ptr and print_context_stack
- use sizeof(long) as the size of a word on the stack
Signed-off-by: Alexander van Heukelum <heukelum@fastmail.fm>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
- x86_64: use %p to print an address
- make i386-version the same as the above
The result should be the same on x86_64; on i386 the
output only changes if CONFIG_KALLSYMS is turned off,
in which case the address is printed twice.
Signed-off-by: Alexander van Heukelum <heukelum@fastmail.fm>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The dumpstack code is logically quite independent from the
hardware traps. Split it out into its own file.
Signed-off-by: Alexander van Heukelum <heukelum@fastmail.fm>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
