We detected a serious issue with PERF_SAMPLE_READ and
timing information when events were being multiplexing.
Samples would have time_running > time_enabled. That
was easy to reproduce with a libpfm4 example (ran 3
times to cause multiplexing on Core 2):
$ syst_smpl -e uops_retired:freq=1 &
$ syst_smpl -e uops_retired:freq=1 &
$ syst_smpl -e uops_retired:freq=1 &
IIP:0x0000000040062d ... PERIOD:2355332948 ENA=40144625315 RUN=60014875184
syst_smpl: WARNING: time_running > time_enabled
63277537998 uops_retired:freq=1 , scaled
The bug was not present in kernel up to (and including) 3.0. It turns
out the bug was introduced by the following commit:
commit c479429591
events: Move lockless timer calculation into helper function
The parameters of the function got reversed yet the call sites
were not updated to reflect the change. That lead to time_running
and time_enabled being swapped. That had no effect when there was
no multiplexing because in that case time_running = time_enabled
but it would show up in any other scenario.
Signed-off-by: Stephane Eranian <eranian@google.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/20110829124112.GA4828@quad
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The current cgroup context switch code was incorrect leading
to bogus counts. Furthermore, as soon as there was an active
cgroup event on a CPU, the context switch cost on that CPU
would increase by a significant amount as demonstrated by a
simple ping/pong example:
$ ./pong
Both processes pinned to CPU1, running for 10s
10684.51 ctxsw/s
Now start a cgroup perf stat:
$ perf stat -e cycles,cycles -A -a -G test -C 1 -- sleep 100
$ ./pong
Both processes pinned to CPU1, running for 10s
6674.61 ctxsw/s
That's a 37% penalty.
Note that pong is not even in the monitored cgroup.
The results shown by perf stat are bogus:
$ perf stat -e cycles,cycles -A -a -G test -C 1 -- sleep 100
Performance counter stats for 'sleep 100':
CPU1 <not counted> cycles test
CPU1 16,984,189,138 cycles # 0.000 GHz
The second 'cycles' event should report a count @ CPU clock
(here 2.4GHz) as it is counting across all cgroups.
The patch below fixes the bogus accounting and bypasses any
cgroup switches in case the outgoing and incoming tasks are
in the same cgroup.
With this patch the same test now yields:
$ ./pong
Both processes pinned to CPU1, running for 10s
10775.30 ctxsw/s
Start perf stat with cgroup:
$ perf stat -e cycles,cycles -A -a -G test -C 1 -- sleep 10
Run pong outside the cgroup:
$ /pong
Both processes pinned to CPU1, running for 10s
10687.80 ctxsw/s
The penalty is now less than 2%.
And the results for perf stat are correct:
$ perf stat -e cycles,cycles -A -a -G test -C 1 -- sleep 10
Performance counter stats for 'sleep 10':
CPU1 <not counted> cycles test # 0.000 GHz
CPU1 23,933,981,448 cycles # 0.000 GHz
Now perf stat reports the correct counts for
for the non cgroup event.
If we run pong inside the cgroup, then we also get the
correct counts:
$ perf stat -e cycles,cycles -A -a -G test -C 1 -- sleep 10
Performance counter stats for 'sleep 10':
CPU1 22,297,726,205 cycles test # 0.000 GHz
CPU1 23,933,981,448 cycles # 0.000 GHz
10.001457237 seconds time elapsed
Signed-off-by: Stephane Eranian <eranian@google.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/20110825135803.GA4697@quad
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Currently, an event's 'pmu' field is set after pmu::event_init() is
called. This means that pmu::event_init() must figure out which struct
pmu the event was initialised from. This makes it difficult to
consolidate common event initialisation code for similar PMUs, and
very difficult to implement drivers for PMUs which can have multiple
instances (e.g. a USB controller PMU, a GPU PMU, etc).
This patch sets the 'pmu' field before initialising the event, allowing
event init code to identify the struct pmu instance easily. In the
event of failure to initialise an event, the event is destroyed via
kfree() without calling perf_event::destroy(), so this shouldn't
result in bad behaviour even if the destroy field was set before
failure to initialise was noted.
Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Reviewed-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/1313062280-19123-1-git-send-email-mark.rutland@arm.com
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Francis reports that s2r gets him spurious NMIs, this is because the
suspend code leaves the boot cpu up and running.
Cure this by adding a suspend notifier. The problem is that hotplug
and suspend are completely un-serialized and the PM notifiers run
before the suspend cpu unplug of all but the boot cpu.
This leaves a window where the user can initialize another hotplug
operation (either remove or add a cpu) resulting in either one too
many or one too few hotplug ops. Thus we cannot use the hotplug code
for the suspend case.
There's another reason to not use the hotplug code, which is that the
hotplug code totally destroys the perf state, we can do better for
suspend and simply remove all counters from the PMU so that we can
re-instate them on resume.
Reported-by: Francis Moreau <francis.moro@gmail.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/n/tip-1cvevybkgmv4s6v5y37t4847@git.kernel.org
Signed-off-by: Ingo Molnar <mingo@elte.hu>
PMU type id can be allocated dynamically, so perf_event_attr::type check
when copying attribute from userspace to kernel is not valid.
Signed-off-by: Lin Ming <ming.m.lin@intel.com>
Cc: Robert Richter <robert.richter@amd.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/1309421396-17438-4-git-send-email-ming.m.lin@intel.com
Signed-off-by: Ingo Molnar <mingo@elte.hu>
KVM needs one-shot samples, since a PMC programmed to -X will fire after X
events and then again after 2^40 events (i.e. variable period).
Signed-off-by: Avi Kivity <avi@redhat.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/1309362157-6596-4-git-send-email-avi@redhat.com
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The perf_event overflow handler does not receive any caller-derived
argument, so many callers need to resort to looking up the perf_event
in their local data structure. This is ugly and doesn't scale if a
single callback services many perf_events.
Fix by adding a context parameter to perf_event_create_kernel_counter()
(and derived hardware breakpoints APIs) and storing it in the perf_event.
The field can be accessed from the callback as event->overflow_handler_context.
All callers are updated.
Signed-off-by: Avi Kivity <avi@redhat.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/1309362157-6596-2-git-send-email-avi@redhat.com
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Since only samples call perf_output_sample() its much saner (and more
correct) to put the sample logic in there than in the
perf_output_begin()/perf_output_end() pair.
Saves a useless argument, reduces conditionals and shrinks
struct perf_output_handle, win!
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/n/tip-2crpvsx3cqu67q3zqjbnlpsc@git.kernel.org
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The nmi parameter indicated if we could do wakeups from the current
context, if not, we would set some state and self-IPI and let the
resulting interrupt do the wakeup.
For the various event classes:
- hardware: nmi=0; PMI is in fact an NMI or we run irq_work_run from
the PMI-tail (ARM etc.)
- tracepoint: nmi=0; since tracepoint could be from NMI context.
- software: nmi=[0,1]; some, like the schedule thing cannot
perform wakeups, and hence need 0.
As one can see, there is very little nmi=1 usage, and the down-side of
not using it is that on some platforms some software events can have a
jiffy delay in wakeup (when arch_irq_work_raise isn't implemented).
The up-side however is that we can remove the nmi parameter and save a
bunch of conditionals in fast paths.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Michael Cree <mcree@orcon.net.nz>
Cc: Will Deacon <will.deacon@arm.com>
Cc: Deng-Cheng Zhu <dengcheng.zhu@gmail.com>
Cc: Anton Blanchard <anton@samba.org>
Cc: Eric B Munson <emunson@mgebm.net>
Cc: Heiko Carstens <heiko.carstens@de.ibm.com>
Cc: Paul Mundt <lethal@linux-sh.org>
Cc: David S. Miller <davem@davemloft.net>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Jason Wessel <jason.wessel@windriver.com>
Cc: Don Zickus <dzickus@redhat.com>
Link: http://lkml.kernel.org/n/tip-agjev8eu666tvknpb3iaj0fg@git.kernel.org
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The event tracing infrastructure exposes two timers which should be updated
each time the value of the counter is updated. Currently, these counters are
only updated when userspace calls read() on the fd associated with an event.
This means that counters which are read via the mmap'd page exclusively never
have their timers updated. This patch adds ensures that the timers are updated
each time the values in the mmap'd page are updated.
Signed-off-by: Eric B Munson <emunson@mgebm.net>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/1308932786-5111-1-git-send-email-emunson@mgebm.net
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Take the timer calculation from perf_output_read and move it to a helper
function for any place that needs timer values but cannot take the ctx->lock.
Signed-off-by: Eric B Munson <emunson@mgebm.net>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/1308861279-15216-2-git-send-email-emunson@mgebm.net
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Since 2.6.36 (specifically commit d57e34fdd6 ("perf: Simplify the
ring-buffer logic: make perf_buffer_alloc() do everything needed"),
the perf_buffer_init_code() has been mis-setting the buffer watermark
if perf_event_attr.wakeup_events has a non-zero value.
This is because perf_event_attr.wakeup_events is a union with
perf_event_attr.wakeup_watermark.
This commit re-enables the check for perf_event_attr.watermark being
set before continuing with setting a non-default watermark.
This bug is most noticable when you are trying to use PERF_IOC_REFRESH
with a value larger than one and perf_event_attr.wakeup_events is set to
one. In this case the buffer watermark will be set to 1 and you will
get extraneous POLL_IN overflows rather than POLL_HUP as expected.
[ avoid using attr.wakeup_events when attr.watermark is set ]
Signed-off-by: Vince Weaver <vweaver1@eecs.utk.edu>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: <stable@kernel.org>
Link: http://lkml.kernel.org/r/alpine.DEB.2.00.1106011506390.5384@cl320.eecs.utk.edu
Signed-off-by: Ingo Molnar <mingo@elte.hu>
A lost Quilt refresh of 2c29ef0fef (perf: Simplify and fix
__perf_install_in_context()) is causing grief and lockups,
reported by Jiri Olsa.
When installing an event in a task context, there's a number of
issues:
- there might not be an existing task context, in which case
we should install the now current context;
- there might already be a context, not the current one, in
which case we should de-schedule the old and install the new;
these cases were dealt with in the lost refresh, however there is one
further case that was found in testing:
- there might already be a context, the current one, in which
case we should still de-schedule, and should take care
to re-install it (note that task_ctx_sched_out() clears
cpuctx->task_ctx).
Reported-by: Jiri Olsa <jolsa@redhat.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/1307399008.2497.971.camel@laptop
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Ben changed the cgroup API in commit f780bdb7c1 (cgroups: add
per-thread subsystem callbacks) in an incompatible way, but
forgot to convert the perf cgroup bits.
Avoid compile warnings and runtime splats and convert perf too ;-)
Acked-by: Ben Blum <bblum@andrew.cmu.edu>
Cc: Stephane Eranian <eranian@google.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/1306767651.1200.2990.camel@twins
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Since perf_install_in_context() will now install a context when we
add the first event, we can de-schedule the context when the last
event is removed.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/20110409192142.090431763@chello.nl
Signed-off-by: Ingo Molnar <mingo@elte.hu>
In order to always call list_del_event() on the correct cpu if the
event is part of an active context and avoid having to do two IPIs,
change the close() semantics slightly.
The current perf_event_disable() call would disable a whole group if
the event that's being closed is the group leader, whereas the new
code keeps the group siblings enabled.
People should not rely on this behaviour and I don't think they do,
but in case we find they do, the fix is easy and we have to take the
double IPI cost.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Vince Weaver <vweaver1@eecs.utk.edu>
Link: http://lkml.kernel.org/r/20110409192142.038377551@chello.nl
Signed-off-by: Ingo Molnar <mingo@elte.hu>
This was scattered out - refactor it into a single function.
No change in functionality.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/20110409192141.979862055@chello.nl
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Instead of tracking if a context is active or not, track which events
of the context are active. By making it a bitmask of
EVENT_PINNED|EVENT_FLEXIBLE we can simplify some of the scheduling
routines since it can avoid adding events that are already active.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/20110409192141.930282378@chello.nl
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Currently __perf_install_in_context() will try and schedule in the
event irrespective of our event scheduling rules, that is, we try to
schedule CPU-pinned, TASK-pinned, CPU-flexible, TASK-flexible, but
when creating a new event we simply try and schedule it on top of
whatever is already on the PMU, this can lead to errors for pinned
events.
Therefore, simplify things and simply schedule everything out, add the
event to the corresponding context and schedule everything back in.
This also nicely handles the case where with
__ARCH_WANT_INTERRUPTS_ON_CTXSW the IPI can come right in the middle
of schedule, before we managed to call perf_event_task_sched_in().
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/20110409192141.870894224@chello.nl
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Make task_ctx_sched_*() imply EVENT_ALL, since anything less will not
actually have scheduled the task in/out at all.
Since there's no site that schedules all of a task in (due to the
interleave with flexible cpuctx) we can remove this function.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/20110409192141.817893268@chello.nl
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Currently we only hold one ctx->lock at a time, which results in us
flipping back and forth between cpuctx->ctx.lock and task_ctx->lock.
Avoid this and gain large atomic regions by holding both locks. We
nest the task lock inside the cpu lock, since with task scheduling we
might have to change task ctx while holding the cpu ctx lock.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/20110409192141.769881865@chello.nl
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Small cleanup to how we refcount in find_get_context(), this also
allows us to use put_ctx() to free things instead of using kfree().
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/20110409192141.719340481@chello.nl
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Oleg noted that ctx_sched_out() disables the PMU even though it might
not actually do something, avoid needless PMU-disabling.
Reported-by: Oleg Nesterov <oleg@redhat.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/20110409192141.665385503@chello.nl
Signed-off-by: Ingo Molnar <mingo@elte.hu>
mv kernel/perf_event.c -> kernel/events/core.c. From there, all further
sensible splitting can happen. The idea is that due to perf_event.c
becoming pretty sizable and with the advent of the marriage with ftrace,
splitting functionality into its logical parts should help speeding up
the unification and to manage the complexity of the subsystem.
Signed-off-by: Borislav Petkov <borislav.petkov@amd.com>
Ingo Molnar
Eric B Munson
Stephane Eranian
Mark Rutland
Peter Zijlstra
Lin Ming
Avi Kivity
Vince Weaver
Frederic Weisbecker
Peter Zijlstra
Linus Torvalds
Borislav Petkov