struct rq isn't visible outside of sched.o so its near useless to
expose the pointer, also there are no users of it, so remove it.
Acked-by: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <1272997616.1642.207.camel@laptop>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
When sleep_time is off the function profiler ignores the time that a task
is scheduled out. When the task is scheduled out a timestamp is taken.
When the task is scheduled back in, the timestamp is compared to the
current time and the saved calltimes are adjusted accordingly.
But when stopping the function profiler, the sched switch hook that
does this adjustment was stopped before shutting down the tracer.
This allowed some tasks to not get their timestamps set when they
scheduled out. When the function profiler started again, this would
skew the times of the scheduler functions.
This patch moves the stopping of the sched switch to after the function
profiler is stopped. It also ignores zero set calltimes, which may
happen on start up.
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
When combined with function graph tracing the ftrace function profiler
also prints the average run time of functions. While this gives us some
good information, it doesn't tell us anything about the variance of the
run times of the function. This change prints out the s^2 sample
standard deviation alongside the average.
This change adds one entry to the profile record structure. This
increases the memory footprint of the function profiler by 1/3 on a
32-bit system, and by 1/5 on a 64-bit system when function graphing is
enabled, though the memory is only allocated when the profiler is turned
on. During the profiling, one extra line of code adds the squared
calltime to the new record entry, so this should not adversly affect
performance.
Note that the square of the sample standard deviation is printed because
there is no sqrt implementation for unsigned long long in the kernel.
Signed-off-by: Chase Douglas <chase.douglas@canonical.com>
LKML-Reference: <1272304925-2436-1-git-send-email-chase.douglas@canonical.com>
[ fixed comment about ns^2 -> us^2 conversion ]
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
With the addition of the "missed events" flags that is stored in the
commit field of the ring buffer page, the ring_buffer_benchmark
was not updated to handle this. If events are missed, then the
missed events flag is set in the ring buffer page, the benchmark
will count that flag as part of the size of the page and will hit the BUG()
when it tries to read beyond the page.
The solution is simply to have the ring buffer benchmark mask off
the extra bits.
Reported-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
When performing a non-consuming read, a synchronize_sched() is
performed once for every cpu which is actively tracing.
This is very expensive, and can make it take several seconds to open
up the 'trace' file with lots of cpus.
Only one synchronize_sched() call is actually necessary. What is
desired is for all cpus to see the disabling state change. So we
transform the existing sequence:
for_each_cpu() {
ring_buffer_read_start();
}
where each ring_buffer_start() call performs a synchronize_sched(),
into the following:
for_each_cpu() {
ring_buffer_read_prepare();
}
ring_buffer_read_prepare_sync();
for_each_cpu() {
ring_buffer_read_start();
}
wherein only the single ring_buffer_read_prepare_sync() call needs to
do the synchronize_sched().
The first phase, via ring_buffer_read_prepare(), allocates the 'iter'
memory and increments ->record_disabled.
In the second phase, ring_buffer_read_prepare_sync() makes sure this
->record_disabled state is visible fully to all cpus.
And in the final third phase, the ring_buffer_read_start() calls reset
the 'iter' objects allocated in the first phase since we now know that
none of the cpus are adding trace entries any more.
This makes openning the 'trace' file nearly instantaneous on a
sparc64 Niagara2 box with 128 cpus tracing.
Signed-off-by: David S. Miller <davem@davemloft.net>
LKML-Reference: <20100420.154711.11246950.davem@davemloft.net>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Add function graph output to irqsoff tracer.
The graph output is enabled by setting new 'display-graph' trace option.
Signed-off-by: Jiri Olsa <jolsa@redhat.com>
LKML-Reference: <1270227683-14631-4-git-send-email-jolsa@redhat.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Let the function graph tracer have custom flags passed to its
output functions.
Signed-off-by: Jiri Olsa <jolsa@redhat.com>
LKML-Reference: <1270227683-14631-3-git-send-email-jolsa@redhat.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Add ftrace events for graph tracer, so the graph output could be shared
with other tracers.
Signed-off-by: Jiri Olsa <jolsa@redhat.com>
LKML-Reference: <1270227683-14631-2-git-send-email-jolsa@redhat.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
The ftrace_dump_on_oops kernel parameter, sysctl and sysrq let one
dump every cpu buffers when an oops or panic happens.
It's nice when you have few cpus but it may take ages if have many,
plus you miss the real origin of the problem in all the cpu traces.
Sometimes, all you need is to dump the cpu buffer that triggered the
opps, most of the time it is our main interest.
This patch modifies ftrace_dump_on_oops to handle this choice.
The ftrace_dump_on_oops kernel parameter, when it comes alone, has
the same behaviour than before. But ftrace_dump_on_oops=orig_cpu
will only dump the buffer of the cpu that oops'ed.
Similarly, sysctl kernel.ftrace_dump_on_oops=1 and
echo 1 > /proc/sys/kernel/ftrace_dump_on_oops keep their previous
behaviour. But setting 2 jumps into cpu origin dump mode.
v2: Fix double setup
v3: Fix spelling issues reported by Randy Dunlap
v4: Also update __ftrace_dump in the selftests
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Acked-by: David S. Miller <davem@davemloft.net>
Acked-by: Steven Rostedt <rostedt@goodmis.org>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Li Zefan <lizf@cn.fujitsu.com>
Cc: Lai Jiangshan <laijs@cn.fujitsu.com>
Because a local variable is not initialized, I got these
when I did 'cat tracing/trace'. (not trace_pipe):
CPU:0 [LOST 18446744071579453134 EVENTS]
ps-3099 [000] 560.770221: lock_acquire: ffff880030865010 &(&dentry->d_lock)->rlock
CPU:0 [LOST 18446744071579453134 EVENTS]
ps-3099 [000] 560.770221: lock_release: ffff880030865010 &(&dentry->d_lock)->rlock
CPU:0 [LOST 18446612133255294080 EVENTS]
ps-3099 [000] 560.770221: lock_acquire: ffff880030865010 &(&dentry->d_lock)->rlock
CPU:0 [LOST 18446744071579453134 EVENTS]
ps-3099 [000] 560.770222: lock_release: ffff880030865010 &(&dentry->d_lock)->rlock
CPU:0 [LOST 18446744071579453134 EVENTS]
ps-3099 [000] 560.770222: lock_release: ffffffff816cfb98 dcache_lock
See peek_next_entry(), it does not set *lost_events when we 'cat tracing/trace'
Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
LKML-Reference: <4BB9A929.2000303@cn.fujitsu.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
* 'perf-fixes-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/linux-2.6-tip:
perf: Always build the powerpc perf_arch_fetch_caller_regs version
perf: Always build the stub perf_arch_fetch_caller_regs version
perf, probe-finder: Build fix on Debian
perf/scripts: Tuple was set from long in both branches in python_process_event()
perf: Fix 'perf sched record' deadlock
perf, x86: Fix callgraphs of 32-bit processes on 64-bit kernels
perf, x86: Fix AMD hotplug & constraint initialization
x86: Move notify_cpu_starting() callback to a later stage
x86,kgdb: Always initialize the hw breakpoint attribute
perf: Use hot regs with software sched switch/migrate events
perf: Correctly align perf event tracing buffer
The trace event buffer used by perf to record raw sample events
is typed as an array of char and may then not be aligned to 8
by alloc_percpu().
But we need it to be aligned to 8 in sparc64 because we cast
this buffer into a random structure type built by the TRACE_EVENT()
macro to store the traces. So if a random 64 bits field is accessed
inside, it may be not under an expected good alignment.
Use an array of long instead to force the appropriate alignment, and
perform a compile time check to ensure the size in byte of the buffer
is a multiple of sizeof(long) so that its actual size doesn't get
shrinked under us.
This fixes unaligned accesses reported while using perf lock
in sparc 64.
Suggested-by: David Miller <davem@davemloft.net>
Suggested-by: Tejun Heo <htejun@gmail.com>
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: David Miller <davem@davemloft.net>
Cc: Steven Rostedt <rostedt@goodmis.org>
Currently, binary readers of the ring buffer only know where events were
lost, but not how many events were lost at that location.
This information is available, but it would require adding another
field to the sub buffer header to include it.
But when a event can not fit at the end of a sub buffer, it is written
to the next sub buffer. This means there is a good chance that the
buffer may have room to hold this counter. If it does, write
the counter at the end of the sub buffer and set another flag
in the data size field that states that this information exists.
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Now that the ring buffer can keep track of where events are lost.
Use this information to the output of trace_pipe:
hackbench-3588 [001] 1326.701660: lock_acquire: ffffffff816591e0 read rcu_read_lock
hackbench-3588 [001] 1326.701661: lock_acquire: ffff88003f4091f0 &(&dentry->d_lock)->rlock
hackbench-3588 [001] 1326.701664: lock_release: ffff88003f4091f0 &(&dentry->d_lock)->rlock
CPU:1 [LOST 673 EVENTS]
hackbench-3588 [001] 1326.702711: kmem_cache_free: call_site=ffffffff81102b85 ptr=ffff880026d96738
hackbench-3588 [001] 1326.702712: lock_release: ffff88003e1480a8 &mm->mmap_sem
hackbench-3588 [001] 1326.702713: lock_acquire: ffff88003e1480a8 &mm->mmap_sem
Even works with the function graph tracer:
2) ! 170.098 us | }
2) 4.036 us | rcu_irq_exit();
2) 3.657 us | idle_cpu();
2) ! 190.301 us | }
CPU:2 [LOST 2196 EVENTS]
2) 0.853 us | } /* cancel_dirty_page */
2) | remove_from_page_cache() {
2) 1.578 us | _raw_spin_lock_irq();
2) | __remove_from_page_cache() {
Note, it does not work with the iterator "trace" file, since it requires
the use of consuming the page from the ring buffer to determine how many
events were lost, which the iterator does not do.
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Currently, when the ring buffer drops events, it does not record
the fact that it did so. It does inform the writer that the event
was dropped by returning a NULL event, but it does not put in any
place holder where the event was dropped.
This is not a trivial thing to add because the ring buffer mostly
runs in overwrite (flight recorder) mode. That is, when the ring
buffer is full, new data will overwrite old data.
In a produce/consumer mode, where new data is simply dropped when
the ring buffer is full, it is trivial to add the placeholder
for dropped events. When there's more room to write new data, then
a special event can be added to notify the reader about the dropped
events.
But in overwrite mode, any new write can overwrite events. A place
holder can not be inserted into the ring buffer since there never
may be room. A reader could also come in at anytime and miss the
placeholder.
Luckily, the way the ring buffer works, the read side can find out
if events were lost or not, and how many events. Everytime a write
takes place, if it overwrites the header page (the next read) it
updates a "overrun" variable that keeps track of the number of
lost events. When a reader swaps out a page from the ring buffer,
it can record this number, perfom the swap, and then check to
see if the number changed, and take the diff if it has, which would be
the number of events dropped. This can be stored by the reader
and returned to callers of the reader.
Since the reader page swap will fail if the writer moved the head
page since the time the reader page set up the swap, this gives room
to record the overruns without worrying about races. If the reader
sets up the pages, records the overrun, than performs the swap,
if the swap succeeds, then the overrun variable has not been
updated since the setup before the swap.
For binary readers of the ring buffer, a flag is set in the header
of each sub page (sub buffer) of the ring buffer. This flag is embedded
in the size field of the data on the sub buffer, in the 31st bit (the size
can be 32 or 64 bits depending on the architecture), but only 27
bits needs to be used for the actual size (less actually).
We could add a new field in the sub buffer header to also record the
number of events dropped since the last read, but this will change the
format of the binary ring buffer a bit too much. Perhaps this change can
be made if the information on the number of events dropped is considered
important enough.
Note, the notification of dropped events is only used by consuming reads
or peeking at the ring buffer. Iterating over the ring buffer does not
keep this information because the necessary data is only available when
a page swap is made, and the iterator does not swap out pages.
Cc: Robert Richter <robert.richter@amd.com>
Cc: Andi Kleen <andi@firstfloor.org>
Cc: Li Zefan <lizf@cn.fujitsu.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: "Luis Claudio R. Goncalves" <lclaudio@uudg.org>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
percpu.h is included by sched.h and module.h and thus ends up being
included when building most .c files. percpu.h includes slab.h which
in turn includes gfp.h making everything defined by the two files
universally available and complicating inclusion dependencies.
percpu.h -> slab.h dependency is about to be removed. Prepare for
this change by updating users of gfp and slab facilities include those
headers directly instead of assuming availability. As this conversion
needs to touch large number of source files, the following script is
used as the basis of conversion.
http://userweb.kernel.org/~tj/misc/slabh-sweep.py
The script does the followings.
* Scan files for gfp and slab usages and update includes such that
only the necessary includes are there. ie. if only gfp is used,
gfp.h, if slab is used, slab.h.
* When the script inserts a new include, it looks at the include
blocks and try to put the new include such that its order conforms
to its surrounding. It's put in the include block which contains
core kernel includes, in the same order that the rest are ordered -
alphabetical, Christmas tree, rev-Xmas-tree or at the end if there
doesn't seem to be any matching order.
* If the script can't find a place to put a new include (mostly
because the file doesn't have fitting include block), it prints out
an error message indicating which .h file needs to be added to the
file.
The conversion was done in the following steps.
1. The initial automatic conversion of all .c files updated slightly
over 4000 files, deleting around 700 includes and adding ~480 gfp.h
and ~3000 slab.h inclusions. The script emitted errors for ~400
files.
2. Each error was manually checked. Some didn't need the inclusion,
some needed manual addition while adding it to implementation .h or
embedding .c file was more appropriate for others. This step added
inclusions to around 150 files.
3. The script was run again and the output was compared to the edits
from #2 to make sure no file was left behind.
4. Several build tests were done and a couple of problems were fixed.
e.g. lib/decompress_*.c used malloc/free() wrappers around slab
APIs requiring slab.h to be added manually.
5. The script was run on all .h files but without automatically
editing them as sprinkling gfp.h and slab.h inclusions around .h
files could easily lead to inclusion dependency hell. Most gfp.h
inclusion directives were ignored as stuff from gfp.h was usually
wildly available and often used in preprocessor macros. Each
slab.h inclusion directive was examined and added manually as
necessary.
6. percpu.h was updated not to include slab.h.
7. Build test were done on the following configurations and failures
were fixed. CONFIG_GCOV_KERNEL was turned off for all tests (as my
distributed build env didn't work with gcov compiles) and a few
more options had to be turned off depending on archs to make things
build (like ipr on powerpc/64 which failed due to missing writeq).
* x86 and x86_64 UP and SMP allmodconfig and a custom test config.
* powerpc and powerpc64 SMP allmodconfig
* sparc and sparc64 SMP allmodconfig
* ia64 SMP allmodconfig
* s390 SMP allmodconfig
* alpha SMP allmodconfig
* um on x86_64 SMP allmodconfig
8. percpu.h modifications were reverted so that it could be applied as
a separate patch and serve as bisection point.
Given the fact that I had only a couple of failures from tests on step
6, I'm fairly confident about the coverage of this conversion patch.
If there is a breakage, it's likely to be something in one of the arch
headers which should be easily discoverable easily on most builds of
the specific arch.
Signed-off-by: Tejun Heo <tj@kernel.org>
Guess-its-ok-by: Christoph Lameter <cl@linux-foundation.org>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
In some error handling cases the lock is not unlocked. The return is
converted to a goto, to share the unlock at the end of the function.
A simplified version of the semantic patch that finds this problem is as
follows: (http://coccinelle.lip6.fr/)
// <smpl>
@r exists@
expression E1;
identifier f;
@@
f (...) { <+...
* spin_lock_irq (E1,...);
... when != E1
* return ...;
...+> }
// </smpl>
Signed-off-by: Julia Lawall <julia@diku.dk>
LKML-Reference: <Pine.LNX.4.64.1003291736440.21896@ask.diku.dk>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
# echo 1 > events/enable
# echo global > trace_clock
------------[ cut here ]------------
WARNING: at kernel/lockdep.c:3162 check_flags+0xb2/0x190()
...
---[ end trace 3f86734a89416623 ]---
possible reason: unannotated irqs-on.
...
There's no reason to use the raw_local_irq_save() in trace_clock_global.
The local_irq_save() version is fine, and does not cause the bug in lockdep.
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Li Zefan <lizf@cn.fujitsu.com>
LKML-Reference: <4BA97FA1.7030606@cn.fujitsu.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
* 'tracing-fixes-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/linux-2.6-tip:
ring-buffer: Do 8 byte alignment for 64 bit that can not handle 4 byte align
The ring buffer uses 4 byte alignment while recording events into the
buffer, even on 64bit machines. This saves space when there are lots
of events being recorded at 4 byte boundaries.
The ring buffer has a zero copy method to write into the buffer, with
the reserving of space and then committing it. This may cause problems
when writing an 8 byte word into a 4 byte alignment (not 8). For x86 and
PPC this is not an issue, but on some architectures this would cause an
out-of-alignment exception.
This patch uses CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS to determine
if it is OK to use 4 byte alignments on 64 bit machines. If it is not,
it forces the ring buffer event header to be 8 bytes and not 4,
and will align the length of the data to be 8 byte aligned.
This keeps the data payload at 8 byte alignments and will allow these
machines to run without issue.
The trick to this is that the header can be either 4 bytes or 8 bytes
depending on the length of the data payload. The 4 byte header
has a length field that supports up to 112 bytes. If the length of
the data is more than 112, the length field is set to zero, and the actual
length is stored in the next 4 bytes after the header.
When CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS is not set, the code forces
zero in the 4 byte header forcing the length to be stored in the 4 byte
array, even with a small data load. It also forces the length of the
data load to be 8 byte aligned. The combination of these two guarantee
that the data is always at 8 byte alignment.
Tested-by: Frederic Weisbecker <fweisbec@gmail.com>
(on sparc64)
Reported-by: Frederic Weisbecker <fweisbec@gmail.com>
Acked-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
* 'perf-fixes-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/linux-2.6-tip: (35 commits)
perf: Fix unexported generic perf_arch_fetch_caller_regs
perf record: Don't try to find buildids in a zero sized file
perf: export perf_trace_regs and perf_arch_fetch_caller_regs
perf, x86: Fix hw_perf_enable() event assignment
perf, ppc: Fix compile error due to new cpu notifiers
perf: Make the install relative to DESTDIR if specified
kprobes: Calculate the index correctly when freeing the out-of-line execution slot
perf tools: Fix sparse CPU numbering related bugs
perf_event: Fix oops triggered by cpu offline/online
perf: Drop the obsolete profile naming for trace events
perf: Take a hot regs snapshot for trace events
perf: Introduce new perf_fetch_caller_regs() for hot regs snapshot
perf/x86-64: Use frame pointer to walk on irq and process stacks
lockdep: Move lock events under lockdep recursion protection
perf report: Print the map table just after samples for which no map was found
perf report: Add multiple event support
perf session: Change perf_session post processing functions to take histogram tree
perf session: Add storage for seperating event types in report
perf session: Change add_hist_entry to take the tree root instead of session
perf record: Add ID and to recorded event data when recording multiple events
...
perf_arch_fetch_caller_regs() is exported for the overriden x86
version, but not for the generic weak version.
As a general rule, weak functions should not have their symbol
exported in the same file they are defined.
So let's export it on trace_event_perf.c as it is used by trace
events only.
This fixes:
ERROR: ".perf_arch_fetch_caller_regs" [fs/xfs/xfs.ko] undefined!
ERROR: ".perf_arch_fetch_caller_regs" [arch/powerpc/platforms/cell/spufs/spufs.ko] undefined!
-v2: And also only build it if trace events are enabled.
-v3: Fix changelog mistake
Reported-by: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Xiao Guangrong <xiaoguangrong@cn.fujitsu.com>
Cc: Paul Mackerras <paulus@samba.org>
LKML-Reference: <1268697902-9518-1-git-send-regression-fweisbec@gmail.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
* 'core-fixes-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/linux-2.6-tip:
locking: Make sparse work with inline spinlocks and rwlocks
x86/mce: Fix RCU lockdep splats
rcu: Increase RCU CPU stall timeouts if PROVE_RCU
ftrace: Replace read_barrier_depends() with rcu_dereference_raw()
rcu: Suppress RCU lockdep warnings during early boot
rcu, ftrace: Fix RCU lockdep splat in ftrace_perf_buf_prepare()
rcu: Suppress __mpol_dup() false positive from RCU lockdep
rcu: Make rcu_read_lock_sched_held() handle !PREEMPT
rcu: Add control variables to lockdep_rcu_dereference() diagnostics
rcu, cgroup: Relax the check in task_subsys_state() as early boot is now handled by lockdep-RCU
rcu: Use wrapper function instead of exporting tasklist_lock
sched, rcu: Fix rcu_dereference() for RCU-lockdep
rcu: Make task_subsys_state() RCU-lockdep checks handle boot-time use
rcu: Fix holdoff for accelerated GPs for last non-dynticked CPU
x86/gart: Unexport gart_iommu_aperture
Fix trivial conflicts in kernel/trace/ftrace.c
* 'tracing-fixes-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/linux-2.6-tip:
tracing: Do not record user stack trace from NMI context
tracing: Disable buffer switching when starting or stopping trace
tracing: Use same local variable when resetting the ring buffer
function-graph: Init curr_ret_stack with ret_stack
ring-buffer: Move disabled check into preempt disable section
function-graph: Add tracing_thresh support to function_graph tracer
tracing: Update the comm field in the right variable in update_max_tr
function-graph: Use comment notation for func names of dangling '}'
function-graph: Fix unused reference to ftrace_set_func()
tracing: Fix warning in s_next of trace file ops
tracing: Include irqflags headers from trace clock
A bug was found with Li Zefan's ftrace_stress_test that caused applications
to segfault during the test.
Placing a tracing_off() in the segfault code, and examining several
traces, I found that the following was always the case. The lock tracer
was enabled (lockdep being required) and userstack was enabled. Testing
this out, I just enabled the two, but that was not good enough. I needed
to run something else that could trigger it. Running a load like hackbench
did not work, but executing a new program would. The following would
trigger the segfault within seconds:
# echo 1 > /debug/tracing/options/userstacktrace
# echo 1 > /debug/tracing/events/lock/enable
# while :; do ls > /dev/null ; done
Enabling the function graph tracer and looking at what was happening
I finally noticed that all cashes happened just after an NMI.
1) | copy_user_handle_tail() {
1) | bad_area_nosemaphore() {
1) | __bad_area_nosemaphore() {
1) | no_context() {
1) | fixup_exception() {
1) 0.319 us | search_exception_tables();
1) 0.873 us | }
[...]
1) 0.314 us | __rcu_read_unlock();
1) 0.325 us | native_apic_mem_write();
1) 0.943 us | }
1) 0.304 us | rcu_nmi_exit();
[...]
1) 0.479 us | find_vma();
1) | bad_area() {
1) | __bad_area() {
After capturing several traces of failures, all of them happened
after an NMI. Curious about this, I added a trace_printk() to the NMI
handler to read the regs->ip to see where the NMI happened. In which I
found out it was here:
ffffffff8135b660 <page_fault>:
ffffffff8135b660: 48 83 ec 78 sub $0x78,%rsp
ffffffff8135b664: e8 97 01 00 00 callq ffffffff8135b800 <error_entry>
What was happening is that the NMI would happen at the place that a page
fault occurred. It would call rcu_read_lock() which was traced by
the lock events, and the user_stack_trace would run. This would trigger
a page fault inside the NMI. I do not see where the CR2 register is
saved or restored in NMI handling. This means that it would corrupt
the page fault handling that the NMI interrupted.
The reason the while loop of ls helped trigger the bug, was that
each execution of ls would cause lots of pages to be faulted in, and
increase the chances of the race happening.
The simple solution is to not allow user stack traces in NMI context.
After this patch, I ran the above "ls" test for a couple of hours
without any issues. Without this patch, the bug would trigger in less
than a minute.
Cc: stable@kernel.org
Reported-by: Li Zefan <lizf@cn.fujitsu.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
When the trace iterator is read, tracing_start() and tracing_stop()
is called to stop tracing while the iterator is processing the trace
output.
These functions disable both the standard buffer and the max latency
buffer. But if the wakeup tracer is running, it can switch these
buffers between the two disables:
buffer = global_trace.buffer;
if (buffer)
ring_buffer_record_disable(buffer);
<<<--------- swap happens here
buffer = max_tr.buffer;
if (buffer)
ring_buffer_record_disable(buffer);
What happens is that we disabled the same buffer twice. On tracing_start()
we can enable the same buffer twice. All ring_buffer_record_disable()
must be matched with a ring_buffer_record_enable() or the buffer
can be disable permanently, or enable prematurely, and cause a bug
where a reset happens while a trace is commiting.
This patch protects these two by taking the ftrace_max_lock to prevent
a switch from occurring.
Found with Li Zefan's ftrace_stress_test.
Cc: stable@kernel.org
Reported-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
In the ftrace code that resets the ring buffer it references the
buffer with a local variable, but then uses the tr->buffer as the
parameter to reset. If the wakeup tracer is running, which can
switch the tr->buffer with the max saved buffer, this can break
the requirement of disabling the buffer before the reset.
buffer = tr->buffer;
ring_buffer_record_disable(buffer);
synchronize_sched();
__tracing_reset(tr->buffer, cpu);
If the tr->buffer is swapped, then the reset is not happening to the
buffer that was disabled. This will cause the ring buffer to fail.
Found with Li Zefan's ftrace_stress_test.
Cc: stable@kernel.org
Reported-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
If the graph tracer is active, and a task is forked but the allocating of
the processes graph stack fails, it can cause crash later on.
This is due to the temporary stack being NULL, but the curr_ret_stack
variable is copied from the parent. If it is not -1, then in
ftrace_graph_probe_sched_switch() the following:
for (index = next->curr_ret_stack; index >= 0; index--)
next->ret_stack[index].calltime += timestamp;
Will cause a kernel OOPS.
Found with Li Zefan's ftrace_stress_test.
Cc: stable@kernel.org
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
The ring buffer resizing and resetting relies on a schedule RCU
action. The buffers are disabled, a synchronize_sched() is called
and then the resize or reset takes place.
But this only works if the disabling of the buffers are within the
preempt disabled section, otherwise a window exists that the buffers
can be written to while a reset or resize takes place.
Cc: stable@kernel.org
Reported-by: Li Zefan <lizf@cn.fujitsu.com>
Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
LKML-Reference: <4B949E43.2010906@cn.fujitsu.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Export perf_trace_regs and perf_arch_fetch_caller_regs since module will
use these.
Signed-off-by: Xiao Guangrong <xiaoguangrong@cn.fujitsu.com>
[ use EXPORT_PER_CPU_SYMBOL_GPL() ]
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <4B989C1B.2090407@cn.fujitsu.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Replace the calls to read_barrier_depends() in
ftrace_list_func() with rcu_dereference_raw() to improve
readability. The reason that we use rcu_dereference_raw() here
is that removed entries are never freed, instead they are simply
leaked. This is one of a very few cases where use of
rcu_dereference_raw() is the long-term right answer. And I
don't yet know of any others. ;-)
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Acked-by: Steven Rostedt <rostedt@goodmis.org>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: laijs@cn.fujitsu.com
Cc: dipankar@in.ibm.com
Cc: mathieu.desnoyers@polymtl.ca
Cc: josh@joshtriplett.org
Cc: dvhltc@us.ibm.com
Cc: niv@us.ibm.com
Cc: peterz@infradead.org
Cc: Valdis.Kletnieks@vt.edu
Cc: dhowells@redhat.com
LKML-Reference: <1267830207-9474-1-git-send-email-paulmck@linux.vnet.ibm.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Drop the obsolete "profile" naming used by perf for trace events.
Perf can now do more than simple events counting, so generalize
the API naming.
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: Steven Rostedt <rostedt@goodmis.org>
Cc: Masami Hiramatsu <mhiramat@redhat.com>
Cc: Jason Baron <jbaron@redhat.com>
We are taking a wrong regs snapshot when a trace event triggers.
Either we use get_irq_regs(), which gives us the interrupted
registers if we are in an interrupt, or we use task_pt_regs()
which gives us the state before we entered the kernel, assuming
we are lucky enough to be no kernel thread, in which case
task_pt_regs() returns the initial set of regs when the kernel
thread was started.
What we want is different. We need a hot snapshot of the regs,
so that we can get the instruction pointer to record in the
sample, the frame pointer for the callchain, and some other
things.
Let's use the new perf_fetch_caller_regs() for that.
Comparison with perf record -e lock: -R -a -f -g
Before:
perf [kernel] [k] __do_softirq
|
--- __do_softirq
|
|--55.16%-- __open
|
--44.84%-- __write_nocancel
After:
perf [kernel] [k] perf_tp_event
|
--- perf_tp_event
|
|--41.07%-- lock_acquire
| |
| |--39.36%-- _raw_spin_lock
| | |
| | |--7.81%-- hrtimer_interrupt
| | | smp_apic_timer_interrupt
| | | apic_timer_interrupt
The old case was producing unreliable callchains. Now having
right frame and instruction pointers, we have the trace we
want.
Also syscalls and kprobe events already have the right regs,
let's use them instead of wasting a retrieval.
v2: Follow the rename perf_save_regs() -> perf_fetch_caller_regs()
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: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Masami Hiramatsu <mhiramat@redhat.com>
Cc: Jason Baron <jbaron@redhat.com>
Cc: Archs <linux-arch@vger.kernel.org>
Add support for tracing_thresh to the function_graph tracer. This
version of this feature isolates the checks into new entry and
return functions, to avoid adding more conditional code into the
main function_graph paths.
When the tracing_thresh is set and the function graph tracer is
enabled, only the functions that took longer than the time in
microseconds that was set in tracing_thresh are recorded. To do this
efficiently, only the function exits are recorded:
[tracing]# echo 100 > tracing_thresh
[tracing]# echo function_graph > current_tracer
[tracing]# cat trace
# tracer: function_graph
#
# CPU DURATION FUNCTION CALLS
# | | | | | | |
1) ! 119.214 us | } /* smp_apic_timer_interrupt */
1) <========== |
0) ! 101.527 us | } /* __rcu_process_callbacks */
0) ! 126.461 us | } /* rcu_process_callbacks */
0) ! 145.111 us | } /* __do_softirq */
0) ! 149.667 us | } /* do_softirq */
0) ! 168.817 us | } /* irq_exit */
0) ! 248.254 us | } /* smp_apic_timer_interrupt */
Also, add support for specifying tracing_thresh on the kernel
command line. When used like so: "tracing_thresh=200 ftrace=function_graph"
this can be used to analyse system startup. It is important to disable
tracing soon after boot, in order to avoid losing the trace data.
Acked-by: Frederic Weisbecker <fweisbec@gmail.com>
Signed-off-by: Tim Bird <tim.bird@am.sony.com>
LKML-Reference: <4B87098B.4040308@am.sony.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
The latency output showed:
# | task: -3 (uid:0 nice:0 policy:1 rt_prio:99)
The comm is missing in the "task:" and it looks like a minus 3 is
the output. The correct display should be:
# | task: migration/0-3 (uid:0 nice:0 policy:1 rt_prio:99)
The problem is that the comm is being stored in the wrong data
structure. The max_tr.data[cpu] is what stores the comm, not the
tr->data[cpu].
Before this patch the max_tr.data[cpu]->comm was zeroed and the /debug/trace
ended up showing just the '-' sign followed by the pid.
Also remove a needless initialization of max_data.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
LKML-Reference: <1267824230-23861-1-git-send-email-acme@infradead.org>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
When a '}' does not have a matching function start, the name is printed
within parenthesis. But this makes it confusing between ending '}'
and function starts. This patch makes the function name appear in C comment
notation.
Old view:
3) 1.281 us | } (might_fault)
3) 3.620 us | } (filldir)
3) 5.251 us | } (call_filldir)
3) | call_filldir() {
3) | filldir() {
New view:
3) 1.281 us | } /* might_fault */
3) 3.620 us | } /* filldir */
3) 5.251 us | } /* call_filldir */
3) | call_filldir() {
3) | filldir() {
Requested-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
The declaration of ftrace_set_func() is at the start of the ftrace.c file
and wrapped with a #ifdef CONFIG_FUNCTION_GRAPH condition. If function
graph tracing is enabled but CONFIG_DYNAMIC_FTRACE is not, a warning
about that function being declared static and unused is given.
This really should have been placed within the CONFIG_FUNCTION_GRAPH
condition that uses ftrace_set_func().
Moving the declaration down fixes the warning and makes the code cleaner.
Reported-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
This warning in s_next() can be triggered by lseek():
[<c018b3f7>] ? s_next+0x77/0x80
[<c013e3c1>] warn_slowpath_common+0x81/0xa0
[<c018b3f7>] ? s_next+0x77/0x80
[<c013e3fa>] warn_slowpath_null+0x1a/0x20
[<c018b3f7>] s_next+0x77/0x80
[<c01efa77>] traverse+0x117/0x200
[<c01eff13>] seq_lseek+0xa3/0x120
[<c01efe70>] ? seq_lseek+0x0/0x120
[<c01d7081>] vfs_llseek+0x41/0x50
[<c01d8116>] sys_llseek+0x66/0xa0
[<c0102bd0>] sysenter_do_call+0x12/0x26
The iterator "leftover" variable is zeroed in the opening of the trace
file. But lseek can call s_start() which will call s_next() without
reseting the "leftover" variable back to zero, which might trigger
the WARN_ON_ONCE(iter->leftover) that is in s_next().
Cc: stable@kernel.org
Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
LKML-Reference: <4B8CE06A.9090207@cn.fujitsu.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
* 'for-2.6.34' of git://git.kernel.dk/linux-2.6-block: (38 commits)
block: don't access jiffies when initialising io_context
cfq: remove 8 bytes of padding from cfq_rb_root on 64 bit builds
block: fix for "Consolidate phys_segment and hw_segment limits"
cfq-iosched: quantum check tweak
blktrace: perform cleanup after setup error
blkdev: fix merge_bvec_fn return value checks
cfq-iosched: requests "in flight" vs "in driver" clarification
cciss: Fix problem with scatter gather elements in the scsi half of the driver
cciss: eliminate unnecessary pointer use in cciss scsi code
cciss: do not use void pointer for scsi hba data
cciss: factor out scatter gather chain block mapping code
cciss: fix scatter gather chain block dma direction kludge
cciss: simplify scatter gather code
cciss: factor out scatter gather chain block allocation and freeing
cciss: detect bad alignment of scsi commands at build time
cciss: clarify command list padding calculation
cfq-iosched: rethink seeky detection for SSDs
cfq-iosched: rework seeky detection
block: remove padding from io_context on 64bit builds
block: Consolidate phys_segment and hw_segment limits
...
