26 Nov, 2010
2 commits
-
Stephane noticed that because the perf_sw_event() call is inside the
perf_event_task_sched_out() call it won't get called unless we
have a per-task counter.Reported-by: Stephane Eranian
Signed-off-by: Peter Zijlstra
LKML-Reference:
Signed-off-by: Ingo Molnar -
It was found that sometimes children of tasks with inherited events had
one extra event. Eventually it turned out to be due to the list rotation
no being exclusive with the list iteration in the inheritance code.Cure this by temporarily disabling the rotation while we inherit the events.
Signed-off-by: Thomas Gleixner
Signed-off-by: Peter Zijlstra
LKML-Reference:
Cc:
Signed-off-by: Ingo Molnar
18 Nov, 2010
2 commits
-
Oleg noticed that a perf-fd keeping a reference on the creating task
leads to a few funny side effects.There's two different aspects to this:
- kernel based perf-events, these should not take out
a reference on the creating task and appear on the task's
event list since they're not bound to fds nor visible
to userspace.- fork() and pthread_create(), these can lead to the creating
task dying (and thus the task's event-list becomming useless)
but keeping the list and ref alive until the event is closed.Combined they lead to malfunction of the ptrace hw_tracepoints.
Cure this by not considering kernel based perf_events for the
owner-list and destroying the owner-list when the owner dies.Reported-by: Oleg Nesterov
Signed-off-by: Peter Zijlstra
Acked-by: Oleg Nesterov
LKML-Reference:
Signed-off-by: Ingo Molnar -
…eric/random-tracing into perf/urgent
12 Nov, 2010
1 commit
-
When using early debugging, the kernel does not initialize the
hw_breakpoint API early enough and causes the late initialization of
the kernel debugger to fail. The boot arguments are:earlyprintk=vga ekgdboc=kbd kgdbwait
Then simply type "go" at the kdb prompt and boot. The kernel will
later emit the message:kgdb: Could not allocate hwbreakpoints
And at that point the kernel debugger will cease to work correctly.
The solution is to initialize the hw_breakpoint at the same time that
all the other perf call backs are initialized instead of using a
core_initcall() initialization which happens well after the kernel
debugger can make use of hardware breakpoints.Signed-off-by: Jason Wessel
CC: Frederic Weisbecker
CC: Ingo Molnar
CC: Peter Zijlstra
LKML-Reference:
Signed-off-by: Frederic Weisbecker
11 Nov, 2010
1 commit
-
This patch corrects time tracking in samples. Without this patch
both time_enabled and time_running are bogus when user asks for
PERF_SAMPLE_READ.One uses PERF_SAMPLE_READ to sample the values of other counters
in each sample. Because of multiplexing, it is necessary to know
both time_enabled, time_running to be able to scale counts correctly.In this second version of the patch, we maintain a shadow
copy of ctx->time which allows us to compute ctx->time without
calling update_context_time() from NMI context. We avoid the
issue that update_context_time() must always be called with
ctx->lock held.We do not keep shadow copies of the other event timings
because if the lead event is overflowing then it is active
and thus it's been scheduled in via event_sched_in() in
which case neither tstamp_stopped, tstamp_running can be modified.This timing logic only applies to samples when PERF_SAMPLE_READ
is used.Note that this patch does not address timing issues related
to sampling inheritance between tasks. This will be addressed
in a future patch.With this patch, the libpfm4 example task_smpl now reports
correct counts (shown on 2.4GHz Core 2):$ task_smpl -p 2400000000 -e unhalted_core_cycles:u,instructions_retired:u,baclears noploop 5
noploop for 5 seconds
IIP:0x000000004006d6 PID:5596 TID:5596 TIME:466,210,211,430 STREAM_ID:33 PERIOD:2,400,000,000 ENA=1,010,157,814 RUN=1,010,157,814 NR=3
2,400,000,254 unhalted_core_cycles:u (33)
2,399,273,744 instructions_retired:u (34)
53,340 baclears (35)Signed-off-by: Stephane Eranian
Signed-off-by: Peter Zijlstra
LKML-Reference:
Signed-off-by: Ingo Molnar
22 Oct, 2010
3 commits
-
This new version (see commit 8e5fc1a) is much simpler and ensures that
in case of error in group_sched_in() during event_sched_in(), the
events up to the failed event go through regular event_sched_out().
But the failed event and the remaining events in the group have their
timings adjusted as if they had also gone through event_sched_in() and
event_sched_out(). This ensures timing uniformity across all events in
a group. This also takes care of the tstamp_stopped problem in case
the group could never be scheduled. The tstamp_stopped is updated as
if the event had actually run.With this patch, the following now reports correct time_enabled,
in case the NMI watchdog is active:$ task -e unhalted_core_cycles,instructions_retired,baclears,baclears
noploop 1
noploop for 1 seconds0 unhalted_core_cycles (100.00% scaling, ena=997,552,872, run=0)
0 instructions_retired (100.00% scaling, ena=997,552,872, run=0)
0 baclears (100.00% scaling, ena=997,552,872, run=0)
0 baclears (100.00% scaling, ena=997,552,872, run=0)And the older test case also works:
$ task -einstructions_retired,baclears,baclears -e
unhalted_core_cycles,baclears,baclears sleep 51680885 instructions_retired (69.39% scaling, ena=950756, run=291006)
10735 baclears (69.39% scaling, ena=950756, run=291006)
10735 baclears (69.39% scaling, ena=950756, run=291006)0 unhalted_core_cycles (100.00% scaling, ena=817932, run=0)
0 baclears (100.00% scaling, ena=817932, run=0)
0 baclears (100.00% scaling, ena=817932, run=0)Signed-off-by: Stephane Eranian
Signed-off-by: Peter Zijlstra
LKML-Reference:
Signed-off-by: Ingo Molnar -
This patch reverts commit 8e5fc1a (perf_events: Fix transaction
recovery in group_sched_in()) because it had one flaw in case the
group could never be scheduled. It would cause time_enabled to get
negative.Signed-off-by: Stephane Eranian
Signed-off-by: Peter Zijlstra
LKML-Reference:
Signed-off-by: Ingo Molnar -
…git/tip/linux-2.6-tip
* 'perf-core-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/linux-2.6-tip: (163 commits)
tracing: Fix compile issue for trace_sched_wakeup.c
[S390] hardirq: remove pointless header file includes
[IA64] Move local_softirq_pending() definition
perf, powerpc: Fix power_pmu_event_init to not use event->ctx
ftrace: Remove recursion between recordmcount and scripts/mod/empty
jump_label: Add COND_STMT(), reducer wrappery
perf: Optimize sw events
perf: Use jump_labels to optimize the scheduler hooks
jump_label: Add atomic_t interface
jump_label: Use more consistent naming
perf, hw_breakpoint: Fix crash in hw_breakpoint creation
perf: Find task before event alloc
perf: Fix task refcount bugs
perf: Fix group moving
irq_work: Add generic hardirq context callbacks
perf_events: Fix transaction recovery in group_sched_in()
perf_events: Fix bogus AMD64 generic TLB events
perf_events: Fix bogus context time tracking
tracing: Remove parent recording in latency tracer graph options
tracing: Use one prologue for the preempt irqs off tracer function tracers
...
19 Oct, 2010
9 commits
-
Acked-by: Frederic Weisbecker
Signed-off-by: Peter Zijlstra
LKML-Reference:
Signed-off-by: Ingo Molnar -
Trades a call + conditional + ret for an unconditional jmp.
Acked-by: Frederic Weisbecker
Signed-off-by: Peter Zijlstra
LKML-Reference:
Signed-off-by: Ingo Molnar -
hw_breakpoint creation needs to account stuff per-task to ensure there
is always sufficient hardware resources to back these things due to
ptrace.With the perf per pmu context changes the event initialization no
longer has access to the event context, for the simple reason that we
need to first find the pmu (result of initialization) before we can
find the context.This makes hw_breakpoints unhappy, because it can no longer do per
task accounting, cure this by frobbing a task pointer in the event::hw
bits for now...Signed-off-by: Peter Zijlstra
Cc: Frederic Weisbecker
LKML-Reference:
Signed-off-by: Ingo Molnar -
So that we can pass the task pointer to the event allocation, so that
we can use task associated data during event initialization.Signed-off-by: Peter Zijlstra
LKML-Reference:
Signed-off-by: Ingo Molnar -
Currently it looks like find_lively_task_by_vpid() takes a task ref
and relies on find_get_context() to drop it.The problem is that perf_event_create_kernel_counter() shouldn't be
dropping task refs.Signed-off-by: Peter Zijlstra
Acked-by: Frederic Weisbecker
Acked-by: Matt Helsley
LKML-Reference:
Signed-off-by: Ingo Molnar -
Matt found we trigger the WARN_ON_ONCE() in perf_group_attach() when we take
the move_group path in perf_event_open().Since we cannot de-construct the group (we rely on it to move the events), we
have to simply ignore the double attach. The group state is context invariant
and doesn't need changing.Reported-by: Matt Fleming
Signed-off-by: Peter Zijlstra
LKML-Reference:
Signed-off-by: Ingo Molnar -
Provide a mechanism that allows running code in IRQ context. It is
most useful for NMI code that needs to interact with the rest of the
system -- like wakeup a task to drain buffers.Perf currently has such a mechanism, so extract that and provide it as
a generic feature, independent of perf so that others may also
benefit.The IRQ context callback is generated through self-IPIs where
possible, or on architectures like powerpc the decrementer (the
built-in timer facility) is set to generate an interrupt immediately.Architectures that don't have anything like this get to do with a
callback from the timer tick. These architectures can call
irq_work_run() at the tail of any IRQ handlers that might enqueue such
work (like the perf IRQ handler) to avoid undue latencies in
processing the work.Signed-off-by: Peter Zijlstra
Acked-by: Kyle McMartin
Acked-by: Martin Schwidefsky
[ various fixes ]
Signed-off-by: Huang Ying
LKML-Reference:
Signed-off-by: Ingo Molnar -
The group_sched_in() function uses a transactional approach to schedule
a group of events. In a group, either all events can be scheduled or
none are. To schedule each event in, the function calls event_sched_in().
In case of error, event_sched_out() is called on each event in the group.The problem is that event_sched_out() does not completely cancel the
effects of event_sched_in(). Furthermore event_sched_out() changes the
state of the event as if it had run which is not true is this particular
case.Those inconsistencies impact time tracking fields and may lead to events
in a group not all reporting the same time_enabled and time_running values.
This is demonstrated with the example below:$ task -eunhalted_core_cycles,baclears,baclears -e unhalted_core_cycles,baclears,baclears sleep 5
1946101 unhalted_core_cycles (32.85% scaling, ena=829181, run=556827)
11423 baclears (32.85% scaling, ena=829181, run=556827)
7671 baclears (0.00% scaling, ena=556827, run=556827)2250443 unhalted_core_cycles (57.83% scaling, ena=962822, run=405995)
11705 baclears (57.83% scaling, ena=962822, run=405995)
11705 baclears (57.83% scaling, ena=962822, run=405995)Notice that in the first group, the last baclears event does not
report the same timings as its siblings.This issue comes from the fact that tstamp_stopped is updated
by event_sched_out() as if the event had actually run.To solve the issue, we must ensure that, in case of error, there is
no change in the event state whatsoever. That means timings must
remain as they were when entering group_sched_in().To do this we defer updating tstamp_running until we know the
transaction succeeded. Therefore, we have split event_sched_in()
in two parts separating the update to tstamp_running.Similarly, in case of error, we do not want to update tstamp_stopped.
Therefore, we have split event_sched_out() in two parts separating
the update to tstamp_stopped.With this patch, we now get the following output:
$ task -eunhalted_core_cycles,baclears,baclears -e unhalted_core_cycles,baclears,baclears sleep 5
2492050 unhalted_core_cycles (71.75% scaling, ena=1093330, run=308841)
11243 baclears (71.75% scaling, ena=1093330, run=308841)
11243 baclears (71.75% scaling, ena=1093330, run=308841)1852746 unhalted_core_cycles (0.00% scaling, ena=784489, run=784489)
9253 baclears (0.00% scaling, ena=784489, run=784489)
9253 baclears (0.00% scaling, ena=784489, run=784489)Note that the uneven timing between groups is a side effect of
the process spending most of its time sleeping, i.e., not enough
event rotations (but that's a separate issue).Signed-off-by: Stephane Eranian
Signed-off-by: Peter Zijlstra
LKML-Reference:
Signed-off-by: Ingo Molnar -
You can only call update_context_time() when the context
is active, i.e., the thread it is attached to is still running.However, perf_event_read() can be called even when the context
is inactive, e.g., user read() the counters. The call to
update_context_time() must be conditioned on the status of
the context, otherwise, bogus time_enabled, time_running may
be returned. Here is an example on AMD64. The task program
is an example from libpfm4. The -p prints deltas every 1s.$ task -p -e cpu_clk_unhalted sleep 5
2,266,610 cpu_clk_unhalted (0.00% scaling, ena=2,158,982, run=2,158,982)
0 cpu_clk_unhalted (0.00% scaling, ena=2,158,982, run=2,158,982)
0 cpu_clk_unhalted (0.00% scaling, ena=2,158,982, run=2,158,982)
0 cpu_clk_unhalted (0.00% scaling, ena=2,158,982, run=2,158,982)
0 cpu_clk_unhalted (0.00% scaling, ena=2,158,982, run=2,158,982)
5,242,358,071 cpu_clk_unhalted (99.95% scaling, ena=5,000,359,984, run=2,319,270)Whereas if you don't read deltas, e.g., no call to perf_event_read() until
the process terminates:$ task -e cpu_clk_unhalted sleep 5
2,497,783 cpu_clk_unhalted (0.00% scaling, ena=2,376,899, run=2,376,899)Notice that time_enable, time_running are bogus in the first example
causing bogus scaling.This patch fixes the problem, by conditionally calling update_context_time()
in perf_event_read().Signed-off-by: Stephane Eranian
Signed-off-by: Peter Zijlstra
Cc: stable@kernel.org
LKML-Reference:
Signed-off-by: Ingo Molnar
15 Oct, 2010
1 commit
-
Conflicts:
arch/arm/oprofile/common.c
kernel/perf_event.c
12 Oct, 2010
2 commits
-
perf events: repair incorrect use of copy_from_user
This makes the perf_event_period() return 0 instead of
-EFAULT on success.Signed-off-by: John Blackwood
Signed-off-by: Joe Korty
Acked-by: Peter Zijlstra
LKML-Reference:
Signed-off-by: Ingo Molnar
11 Oct, 2010
1 commit
-
Introduce perf_pmu_name() helper function that returns the name of the
pmu. This gives us a generic way to get the name of a pmu regardless of
how an architecture identifies it internally.Signed-off-by: Matt Fleming
Acked-by: Peter Zijlstra
Acked-by: Paul Mundt
Signed-off-by: Robert Richter
04 Oct, 2010
1 commit
-
This patch fixes an error in perf_event_open() when the pid
provided by the user is invalid. find_lively_task_by_vpid()
does not return NULL on error but an error code. Without the
fix the error code was silently passed to find_get_context()
which would eventually cause a invalid pointer dereference.Signed-off-by: Stephane Eranian
Cc: peterz@infradead.org
Cc: paulus@samba.org
Cc: davem@davemloft.net
Cc: fweisbec@gmail.com
Cc: perfmon2-devel@lists.sf.net
Cc: eranian@gmail.com
Cc: robert.richter@amd.com
LKML-Reference:
Signed-off-by: Ingo Molnar
21 Sep, 2010
1 commit
-
The per-pmu per-cpu context patch converted things from
get_cpu_var() to this_cpu_ptr(), but that only works if
rcu_read_lock() actually disables preemption, and since
there is no such guarantee, we need to fix that.Use the newly introduced {get,put}_cpu_ptr().
Signed-off-by: Peter Zijlstra
Cc: Tejun Heo
LKML-Reference:
Signed-off-by: Ingo Molnar
17 Sep, 2010
4 commits
-
Revert the timer per cpu-context timers because of unfortunate
nohz interaction. Fixing that would have been somewhat ugly, so
go back to driving things from the regular tick. Provide a
jiffies interval feature for people who want slower rotations.Signed-off-by: Peter Zijlstra
Cc: Stephane Eranian
Cc: Robert Richter
Cc: Yinghai Lu
LKML-Reference:
Signed-off-by: Ingo Molnar -
Use the right cpu-context.. spotted by preempt warning on
hot-unplugSigned-off-by: Peter Zijlstra
Cc: Stephane Eranian
Cc: Robert Richter
LKML-Reference:
Signed-off-by: Ingo Molnar -
Aside from allowing software events into a !software group,
allow adding !software events to pure software groups.Once we've moved the software group and attached the first
!software event, the group will no longer be a pure software
group and hence no longer be eligible for movement, at which
point the straight ctx comparison is correct again.Signed-off-by: Peter Zijlstra
Cc: Stephane Eranian
Cc: Robert Richter
Cc: Paul Mackerras
LKML-Reference:
Signed-off-by: Ingo Molnar -
Events were not grouped anymore. The reason was that in
perf_event_open(), the field event->group_leader was
initialized before the function looked up the group_fd
to find the event leader. This patch fixes this by
reordering the code correctly.Signed-off-by: Stephane Eranian
Signed-off-by: Peter Zijlstra
Cc: Robert Richter
LKML-Reference:
Signed-off-by: Ingo Molnar
15 Sep, 2010
2 commits
-
The kernel perf event creation path shouldn't use find_task_by_vpid()
because a vpid exists in a specific namespace. find_task_by_vpid() uses
current's pid namespace which isn't always the correct namespace to use
for the vpid in all the places perf_event_create_kernel_counter() (and
thus find_get_context()) is called.The goal is to clean up pid namespace handling and prevent bugs like:
https://bugzilla.kernel.org/show_bug.cgi?id=17281
Instead of using pids switch find_get_context() to use task struct
pointers directly. The syscall is responsible for resolving the pid to
a task struct. This moves the pid namespace resolution into the syscall
much like every other syscall that takes pid parameters.Signed-off-by: Matt Helsley
Signed-off-by: Peter Zijlstra
Cc: Robin Green
Cc: Prasad
Cc: Arnaldo Carvalho de Melo
Cc: Steven Rostedt
Cc: Will Deacon
Cc: Mahesh Salgaonkar
LKML-Reference:
Signed-off-by: Ingo Molnar -
Split out the code which searches for non-exiting tasks into its own
helper. Creating this helper not only makes the code slightly more
readable it prepares to move the search out of find_get_context() in
a subsequent commit.Signed-off-by: Matt Helsley
Signed-off-by: Peter Zijlstra
Cc: Robin Green
Cc: Prasad
Cc: Arnaldo Carvalho de Melo
Cc: Steven Rostedt
Cc: Will Deacon
Cc: Mahesh Salgaonkar
LKML-Reference:
Signed-off-by: Ingo Molnar
13 Sep, 2010
2 commits
-
With the context rework stuff we can actually end up freeing an event
before it gets attached to a context.Reported-by: Cyrill Gorcunov
Signed-off-by: Peter Zijlstra
LKML-Reference:
Signed-off-by: Ingo Molnar -
Simplify things and simply synchronize against two RCU variants for
PMU unregister -- we don't care about performance, its module unload
if anything.Reported-by: Frederic Weisbecker
Signed-off-by: Peter Zijlstra
Cc: Paul E. McKenney
LKML-Reference:
Signed-off-by: Ingo Molnar
10 Sep, 2010
8 commits
-
We ought to return -ENOENT when non of the registered PMUs
recognise the requested event.This fixes a boot crash that occurs if no PMU is available
but the NMI watchdog tries to register an event.Reported-by: Ingo Molnar
Signed-off-by: Peter Zijlstra
LKML-Reference:
Signed-off-by: Ingo Molnar -
Even though we call it from the inherit path, where the child is
not yet accessible, we need to hold ctx->lock, add_event_to_ctx()
assumes IRQs are disabled.Signed-off-by: Peter Zijlstra
LKML-Reference:
Signed-off-by: Ingo Molnar -
I missed a perf_event_ctxp user when converting it to an array. Pull this
last user into perf_event.c as well and fix it up.Signed-off-by: Peter Zijlstra
LKML-Reference:
Signed-off-by: Ingo Molnar -
Assuming we don't mix events of different pmus onto a single context
(with the exeption of software events inside a hardware group) we can
now assume that all events on a particular context belong to the same
pmu, hence we can disable the pmu for the entire context operations.This reduces the amount of hardware writes.
The exception for swevents comes from the fact that the sw pmu disable
is a nop.Signed-off-by: Peter Zijlstra
Cc: paulus
Cc: stephane eranian
Cc: Robert Richter
Cc: Frederic Weisbecker
Cc: Lin Ming
Cc: Yanmin
LKML-Reference:
Signed-off-by: Ingo Molnar -
Since software events are always schedulable, mixing them up with
hardware events (who are not) can lead to funny scheduling oddities.Giving them their own context solves this.
Signed-off-by: Peter Zijlstra
Cc: paulus
Cc: stephane eranian
Cc: Robert Richter
Cc: Frederic Weisbecker
Cc: Lin Ming
Cc: Yanmin
LKML-Reference:
Signed-off-by: Ingo Molnar -
Provide the infrastructure for multiple task contexts.
A more flexible approach would have resulted in more pointer chases
in the scheduling hot-paths. This approach has the limitation of a
static number of task contexts.Since I expect most external PMUs to be system wide, or at least node
wide (as per the intel uncore unit) they won't actually need a task
context.Signed-off-by: Peter Zijlstra
Cc: paulus
Cc: stephane eranian
Cc: Robert Richter
Cc: Frederic Weisbecker
Cc: Lin Ming
Cc: Yanmin
LKML-Reference:
Signed-off-by: Ingo Molnar -
Unify the two perf_event_context allocation sites.
Signed-off-by: Peter Zijlstra
Cc: paulus
Cc: stephane eranian
Cc: Robert Richter
Cc: Frederic Weisbecker
Cc: Lin Ming
Cc: Yanmin
LKML-Reference:
Signed-off-by: Ingo Molnar -
Move all inherit code near each other.
Signed-off-by: Peter Zijlstra
Cc: paulus
Cc: stephane eranian
Cc: Robert Richter
Cc: Frederic Weisbecker
Cc: Lin Ming
Cc: Yanmin
LKML-Reference:
Signed-off-by: Ingo Molnar