Gleb writes:
> Currently pmu is disabled and re-enabled on each timer interrupt even
> when no rotation or frequency adjustment is needed. On Intel CPU this
> results in two writes into PERF_GLOBAL_CTRL MSR per tick. On bare metal
> it does not cause significant slowdown, but when running perf in a virtual
> machine it leads to 20% slowdown on my machine.
Cure this by keeping a perf_event_context::nr_freq counter that counts the
number of active events that require frequency adjustments and use this in a
similar fashion to the already existing nr_events != nr_active test in
perf_rotate_context().
By being able to exclude both rotation and frequency adjustments a-priory for
the common case we can avoid the otherwise superfluous PMU disable.
Suggested-by: Gleb Natapov <gleb@redhat.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/n/tip-515yhoatehd3gza7we9fapaa@git.kernel.org
Signed-off-by: Ingo Molnar <mingo@elte.hu>
When you do:
$ perf record -e cycles,cycles,cycles noploop 10
You expect about 10,000 samples for each event, i.e., 10s at
1000samples/sec. However, this is not what's happening. You
get much fewer samples, maybe 3700 samples/event:
$ perf report -D | tail -15
Aggregated stats:
TOTAL events: 10998
MMAP events: 66
COMM events: 2
SAMPLE events: 10930
cycles stats:
TOTAL events: 3644
SAMPLE events: 3644
cycles stats:
TOTAL events: 3642
SAMPLE events: 3642
cycles stats:
TOTAL events: 3644
SAMPLE events: 3644
On a Intel Nehalem or even AMD64, there are 4 counters capable
of measuring cycles, so there is plenty of space to measure those
events without multiplexing (even with the NMI watchdog active).
And even with multiplexing, we'd expect roughly the same number
of samples per event.
The root of the problem was that when the event that caused the buffer
to become full was not the first event passed on the cmdline, the user
notification would get lost. The notification was sent to the file
descriptor of the overflowed event but the perf tool was not polling
on it. The perf tool aggregates all samples into a single buffer,
i.e., the buffer of the first event. Consequently, it assumes
notifications for any event will come via that descriptor.
The seemingly straight forward solution of moving the waitq into the
ringbuffer object doesn't work because of life-time issues. One could
perf_event_set_output() on a fd that you're also blocking on and cause
the old rb object to be freed while its waitq would still be
referenced by the blocked thread -> FAIL.
Therefore link all events to the ringbuffer and broadcast the wakeup
from the ringbuffer object to all possible events that could be waited
upon. This is rather ugly, and we're open to better solutions but it
works for now.
Reported-by: Stephane Eranian <eranian@google.com>
Finished-by: Stephane Eranian <eranian@google.com>
Reviewed-by: Stephane Eranian <eranian@google.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/20111126014731.GA7030@quad
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The two new attributes exclude_guest and exclude_host can
bes used by user-space to tell the kernel to setup
performance counter to either only count while the CPU is in
guest or in host mode.
An additional check is also introduced to make sure
user-space does not try to exclude guest and host mode from
counting.
Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
Signed-off-by: Gleb Natapov <gleb@redhat.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/1317816084-18026-2-git-send-email-gleb@redhat.com
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>
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>
Add a NODE level to the generic cache events which is used to measure
local vs remote memory accesses. Like all other cache events, an
ACCESS is HIT+MISS, if there is no way to distinguish between reads
and writes do reads only etc..
The below needs filling out for !x86 (which I filled out with
unsupported events).
I'm fairly sure ARM can leave it like that since it doesn't strike me as
an architecture that even has NUMA support. SH might have something since
it does appear to have some NUMA bits.
Sparc64, PowerPC and MIPS certainly want a good look there since they
clearly are NUMA capable.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: David Miller <davem@davemloft.net>
Cc: Anton Blanchard <anton@samba.org>
Cc: David Daney <ddaney@caviumnetworks.com>
Cc: Deng-Cheng Zhu <dengcheng.zhu@gmail.com>
Cc: Paul Mundt <lethal@linux-sh.org>
Cc: Will Deacon <will.deacon@arm.com>
Cc: Robert Richter <robert.richter@amd.com>
Cc: Stephane Eranian <eranian@google.com>
Link: http://lkml.kernel.org/r/1303508226.4865.8.camel@laptop
Signed-off-by: Ingo Molnar <mingo@elte.hu>
This patch improves the code managing the extra shared registers
used for offcore_response events on Intel Nehalem/Westmere. The
idea is to use static allocation instead of dynamic allocation.
This simplifies greatly the get and put constraint routines for
those events.
The patch also renames per_core to shared_regs because the same
data structure gets used whether or not HT is on. When HT is
off, those events still need to coordination because they use
a extra MSR that has to be shared within an event group.
Signed-off-by: Stephane Eranian <eranian@google.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/20110606145703.GA7258@quad
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>
Add two generic hardware events: front-end and back-end stalled cycles.
These events measure conditions when the CPU is executing code but its
capabilities are not fully utilized. Understanding such situations and
analyzing them is an important sub-task of code optimization workflows.
Both events limit performance: most front end stalls tend to be caused
by branch misprediction or instruction fetch cachemisses, backend
stalls can be caused by various resource shortages or inefficient
instruction scheduling.
Front-end stalls are the more important ones: code cannot run fast
if the instruction stream is not being kept up.
An over-utilized back-end can cause front-end stalls and thus
has to be kept an eye on as well.
The exact composition is very program logic and instruction mix
dependent.
We use the terms 'stall', 'front-end' and 'back-end' loosely and
try to use the best available events from specific CPUs that
approximate these concepts.
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Link: http://lkml.kernel.org/n/tip-7y40wib8n000io7hjpn1dsrm@git.kernel.org
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Conflicts:
include/linux/perf_event.h
Merge reason: pick up the latest jump-label enhancements, they are cooked ready.
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Introduce:
static __always_inline bool static_branch(struct jump_label_key *key);
instead of the old JUMP_LABEL(key, label) macro.
In this way, jump labels become really easy to use:
Define:
struct jump_label_key jump_key;
Can be used as:
if (static_branch(&jump_key))
do unlikely code
enable/disale via:
jump_label_inc(&jump_key);
jump_label_dec(&jump_key);
that's it!
For the jump labels disabled case, the static_branch() becomes an
atomic_read(), and jump_label_inc()/dec() are simply atomic_inc(),
atomic_dec() operations. We show testing results for this change below.
Thanks to H. Peter Anvin for suggesting the 'static_branch()' construct.
Since we now require a 'struct jump_label_key *key', we can store a pointer into
the jump table addresses. In this way, we can enable/disable jump labels, in
basically constant time. This change allows us to completely remove the previous
hashtable scheme. Thanks to Peter Zijlstra for this re-write.
Testing:
I ran a series of 'tbench 20' runs 5 times (with reboots) for 3
configurations, where tracepoints were disabled.
jump label configured in
avg: 815.6
jump label *not* configured in (using atomic reads)
avg: 800.1
jump label *not* configured in (regular reads)
avg: 803.4
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <20110316212947.GA8792@redhat.com>
Signed-off-by: Jason Baron <jbaron@redhat.com>
Suggested-by: H. Peter Anvin <hpa@linux.intel.com>
Tested-by: David Daney <ddaney@caviumnetworks.com>
Acked-by: Ralf Baechle <ralf@linux-mips.org>
Acked-by: David S. Miller <davem@davemloft.net>
Acked-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Jiri reported:
|
| - once an event is created by sys_perf_event_open, task context
| is created and it stays even if the event is closed, until the
| task is finished ... thats what I see in code and I assume it's
| correct
|
| - when the task opens event, perf_sched_events jump label is
| incremented and following callbacks are started from scheduler
|
| __perf_event_task_sched_in
| __perf_event_task_sched_out
|
| These callback *in/out set/unset cpuctx->task_ctx value to the
| task context.
|
| - close is called on event on CPU 0:
| - the task is scheduled on CPU 0
| - __perf_event_task_sched_in is called
| - cpuctx->task_ctx is set
| - perf_sched_events jump label is decremented and == 0
| - __perf_event_task_sched_out is not called
| - cpuctx->task_ctx on CPU 0 stays set
|
| - exit is called on CPU 1:
| - the task is scheduled on CPU 1
| - perf_event_exit_task is called
| - task_ctx_sched_out unsets cpuctx->task_ctx on CPU 1
| - put_ctx destroys the context
|
| - another call of perf_rotate_context on CPU 0 will use invalid
| task_ctx pointer, and eventualy panic.
|
Cure this the simplest possibly way by partially reverting the
jump_label optimization for the sched_out case.
Reported-and-tested-by: Jiri Olsa <jolsa@redhat.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: <stable@kernel.org> # .37+
LKML-Reference: <1301520405.4859.213.camel@twins>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
This patch solves a stale pointer problem in
update_cgrp_time_from_cpuctx(). The cpuctx->cgrp
was not cleared on all possible event exit paths,
including:
close()
perf_release()
perf_release_kernel()
list_del_event()
This patch fixes list_del_event() to clear cpuctx->cgrp
when there are no cgroup events left in the context.
[ This second version makes the code compile when
CONFIG_CGROUP_PERF is not enabled. We unconditionally define
perf_cpu_context->cgrp. ]
Signed-off-by: Stephane Eranian <eranian@google.com>
Cc: peterz@infradead.org
Cc: perfmon2-devel@lists.sf.net
Cc: paulus@samba.org
Cc: davem@davemloft.net
LKML-Reference: <20110323150306.GA1580@quad>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Peter Zijlstra
Joerg Roedel
Stephane Eranian
Arun Sharma
Avi Kivity
Peter Zijlstra
Richard Kennedy
Ingo Molnar
Frederic Weisbecker
Vince Weaver
Linus Torvalds
Jason Baron
Lucas De Marchi