As the "perf_" prefix should be restricted to functions and types in
tools/lib/perf/, aka libperf, this way we reduce a bit the confusion for
types only in libperf or the ones in the more contained tools/perf/
project.
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Song Liu <songliubraving@fb.com>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Change Intel PT's branch stack support to use thread stacks. The
advantages of using branch stack support from the thread-stack are:
1. the branches are accumulated separately for each thread
2. the branch stack is cleared only in between continuous traces
This helps pave the way for adding branch stacks to regular events, not
just synthesized events as at present.
While the 2 approaches are not identical, in simple cases the results
can be identical e.g.
Before:
# perf record --kcore -e intel_pt// uname
# perf script --itrace=i10usl -F+brstacksym,+addr,+flags > cmp1.txt
After:
# perf script --itrace=i10usl -F+brstacksym,+addr,+flags > cmp2.txt
# diff -s cmp1.txt cmp2.txt
Files cmp1.txt and cmp2.txt are identical
Signed-off-by: Adrian Hunter <adrian.hunter@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Link: http://lore.kernel.org/lkml/20200429150751.12570-4-adrian.hunter@intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Intel PT already has support for creating branch stacks for each context
(per-cpu or per-thread). In the more common per-cpu case, the branch stack
is not separated for different threads, instead being cleared in between
each sample.
That approach will not work very well for adding branch stacks to
regular events. The branch stacks really need to be accumulated
separately for each thread.
As a start to accomplishing that, this patch adds support for putting
branch stack support into the thread-stack. The advantages are:
1. the branches are accumulated separately for each thread
2. the branch stack is cleared only in between continuous traces
This helps pave the way for adding branch stacks to regular events, not
just synthesized events as at present.
Signed-off-by: Adrian Hunter <adrian.hunter@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Link: http://lore.kernel.org/lkml/20200429150751.12570-2-adrian.hunter@intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Trying to disentangle this a bit further, unfortunately it uses
parse_events(), its interesting to have it separated anyway, so do it.
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Patch enhances current metric infrastructure to handle "?" in the metric
expression. The "?" can be use for parameters whose value not known
while creating metric events and which can be replace later at runtime
to the proper value. It also add flexibility to create multiple events
out of single metric event added in JSON file.
Patch adds function 'arch_get_runtimeparam' which is a arch specific
function, returns the count of metric events need to be created. By
default it return 1.
This infrastructure needed for hv_24x7 socket/chip level events.
"hv_24x7" chip level events needs specific chip-id to which the data is
requested. Function 'arch_get_runtimeparam' implemented in header.c
which extract number of sockets from sysfs file "sockets" under
"/sys/devices/hv_24x7/interface/".
With this patch basically we are trying to create as many metric events
as define by runtime_param.
For that one loop is added in function 'metricgroup__add_metric', which
create multiple events at run time depend on return value of
'arch_get_runtimeparam' and merge that event in 'group_list'.
To achieve that we are actually passing this parameter value as part of
`expr__find_other` function and changing "?" present in metric
expression with this value.
As in our JSON file, there gonna be single metric event, and out of
which we are creating multiple events.
To understand which data count belongs to which parameter value,
we also printing param value in generic_metric function.
For example,
command:# ./perf stat -M PowerBUS_Frequency -C 0 -I 1000
1.000101867 9,356,933 hv_24x7/pm_pb_cyc,chip=0/ # 2.3 GHz PowerBUS_Frequency_0
1.000101867 9,366,134 hv_24x7/pm_pb_cyc,chip=1/ # 2.3 GHz PowerBUS_Frequency_1
2.000314878 9,365,868 hv_24x7/pm_pb_cyc,chip=0/ # 2.3 GHz PowerBUS_Frequency_0
2.000314878 9,366,092 hv_24x7/pm_pb_cyc,chip=1/ # 2.3 GHz PowerBUS_Frequency_1
So, here _0 and _1 after PowerBUS_Frequency specify parameter value.
Signed-off-by: Kajol Jain <kjain@linux.ibm.com>
Acked-by: Jiri Olsa <jolsa@redhat.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Anju T Sudhakar <anju@linux.vnet.ibm.com>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Jin Yao <yao.jin@linux.intel.com>
Cc: Joe Mario <jmario@redhat.com>
Cc: Kan Liang <kan.liang@linux.intel.com>
Cc: Madhavan Srinivasan <maddy@linux.vnet.ibm.com>
Cc: Mamatha Inamdar <mamatha4@linux.vnet.ibm.com>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Michael Petlan <mpetlan@redhat.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Paul Mackerras <paulus@ozlabs.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Ravi Bangoria <ravi.bangoria@linux.ibm.com>
Cc: Sukadev Bhattiprolu <sukadev@linux.vnet.ibm.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: linuxppc-dev@lists.ozlabs.org
Link: http://lore.kernel.org/lkml/20200401203340.31402-5-kjain@linux.ibm.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
The synthesize benchmark, run on a single process and thread, shows
perf_event__synthesize_mmap_events as the hottest function with fgets
and sscanf taking the majority of execution time.
fscanf performs similarly well. Replace the scanf call with manual
reading of each field of the /proc/pid/maps line, and remove some
unnecessary buffering.
This change also addresses potential, but unlikely, buffer overruns for
the string values read by scanf.
Performance before is:
$ sudo perf bench internals synthesize -m 16 -M 16 -s -t
\# Running 'internals/synthesize' benchmark:
Computing performance of single threaded perf event synthesis by
synthesizing events on the perf process itself:
Average synthesis took: 102.810 usec (+- 0.027 usec)
Average num. events: 17.000 (+- 0.000)
Average time per event 6.048 usec
Average data synthesis took: 106.325 usec (+- 0.018 usec)
Average num. events: 89.000 (+- 0.000)
Average time per event 1.195 usec
Computing performance of multi threaded perf event synthesis by
synthesizing events on CPU 0:
Number of synthesis threads: 16
Average synthesis took: 68103.100 usec (+- 441.234 usec)
Average num. events: 30703.000 (+- 0.730)
Average time per event 2.218 usec
And after is:
$ sudo perf bench internals synthesize -m 16 -M 16 -s -t
\# Running 'internals/synthesize' benchmark:
Computing performance of single threaded perf event synthesis by
synthesizing events on the perf process itself:
Average synthesis took: 50.388 usec (+- 0.031 usec)
Average num. events: 17.000 (+- 0.000)
Average time per event 2.964 usec
Average data synthesis took: 52.693 usec (+- 0.020 usec)
Average num. events: 89.000 (+- 0.000)
Average time per event 0.592 usec
Computing performance of multi threaded perf event synthesis by
synthesizing events on CPU 0:
Number of synthesis threads: 16
Average synthesis took: 45022.400 usec (+- 552.740 usec)
Average num. events: 30624.200 (+- 10.037)
Average time per event 1.470 usec
On a Intel Xeon 6154 compiling with Debian gcc 9.2.1.
Committer testing:
On a AMD Ryzen 5 3600X 6-Core Processor:
Before:
# perf bench internals synthesize --min-threads 12 --max-threads 12 --st --mt
# Running 'internals/synthesize' benchmark:
Computing performance of single threaded perf event synthesis by
synthesizing events on the perf process itself:
Average synthesis took: 267.491 usec (+- 0.176 usec)
Average num. events: 56.000 (+- 0.000)
Average time per event 4.777 usec
Average data synthesis took: 277.257 usec (+- 0.169 usec)
Average num. events: 287.000 (+- 0.000)
Average time per event 0.966 usec
Computing performance of multi threaded perf event synthesis by
synthesizing events on CPU 0:
Number of synthesis threads: 12
Average synthesis took: 81599.500 usec (+- 346.315 usec)
Average num. events: 36096.100 (+- 2.523)
Average time per event 2.261 usec
#
After:
# perf bench internals synthesize --min-threads 12 --max-threads 12 --st --mt
# Running 'internals/synthesize' benchmark:
Computing performance of single threaded perf event synthesis by
synthesizing events on the perf process itself:
Average synthesis took: 110.125 usec (+- 0.080 usec)
Average num. events: 56.000 (+- 0.000)
Average time per event 1.967 usec
Average data synthesis took: 118.518 usec (+- 0.057 usec)
Average num. events: 287.000 (+- 0.000)
Average time per event 0.413 usec
Computing performance of multi threaded perf event synthesis by
synthesizing events on CPU 0:
Number of synthesis threads: 12
Average synthesis took: 43490.700 usec (+- 284.527 usec)
Average num. events: 37028.500 (+- 0.563)
Average time per event 1.175 usec
#
Signed-off-by: Ian Rogers <irogers@google.com>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Acked-by: Jiri Olsa <jolsa@redhat.com>
Acked-by: Namhyung Kim <namhyung@kernel.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Andrey Zhizhikin <andrey.z@gmail.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Kan Liang <kan.liang@linux.intel.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Petr Mladek <pmladek@suse.com>
Cc: Stephane Eranian <eranian@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lore.kernel.org/lkml/20200415054050.31645-4-irogers@google.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
To control degree of parallelism of the synthesize_mmap() code which
is scanning /proc/PID/task/PID/maps and can be time consuming.
Mimic perf top way of handling the option.
If not specified will default to 1 thread, i.e. default behavior before
this option.
On a desktop computer the processing of /proc/PID/task/PID/maps isn't
slow enough to warrant parallel processing and the thread creation has
some cost - hence the default of 1. On a loaded server with
>100 cores it is possible to see synthesis times in the order of
seconds and in this case having the option is desirable.
As the processing is a synchronization point, it is legitimate to worry if
Amdahl's law will apply to this patch. Profiling with this patch in
place:
https://lore.kernel.org/lkml/20200415054050.31645-4-irogers@google.com/
shows:
...
- 32.59% __perf_event__synthesize_threads
- 32.54% __event__synthesize_thread
+ 22.13% perf_event__synthesize_mmap_events
+ 6.68% perf_event__get_comm_ids.constprop.0
+ 1.49% process_synthesized_event
+ 1.29% __GI___readdir64
+ 0.60% __opendir
...
That is the processing is 1.49% of execution time and there is plenty to
make parallel. This is shown in the benchmark in this patch:
https://lore.kernel.org/lkml/20200415054050.31645-2-irogers@google.com/
Computing performance of multi threaded perf event synthesis by
synthesizing events on CPU 0:
Number of synthesis threads: 1
Average synthesis took: 127729.000 usec (+- 3372.880 usec)
Average num. events: 21548.600 (+- 0.306)
Average time per event 5.927 usec
Number of synthesis threads: 2
Average synthesis took: 88863.500 usec (+- 385.168 usec)
Average num. events: 21552.800 (+- 0.327)
Average time per event 4.123 usec
Number of synthesis threads: 3
Average synthesis took: 83257.400 usec (+- 348.617 usec)
Average num. events: 21553.200 (+- 0.327)
Average time per event 3.863 usec
Number of synthesis threads: 4
Average synthesis took: 75093.000 usec (+- 422.978 usec)
Average num. events: 21554.200 (+- 0.200)
Average time per event 3.484 usec
Number of synthesis threads: 5
Average synthesis took: 64896.600 usec (+- 353.348 usec)
Average num. events: 21558.000 (+- 0.000)
Average time per event 3.010 usec
Number of synthesis threads: 6
Average synthesis took: 59210.200 usec (+- 342.890 usec)
Average num. events: 21560.000 (+- 0.000)
Average time per event 2.746 usec
Number of synthesis threads: 7
Average synthesis took: 54093.900 usec (+- 306.247 usec)
Average num. events: 21562.000 (+- 0.000)
Average time per event 2.509 usec
Number of synthesis threads: 8
Average synthesis took: 48938.700 usec (+- 341.732 usec)
Average num. events: 21564.000 (+- 0.000)
Average time per event 2.269 usec
Where average time per synthesized event goes from 5.927 usec with 1
thread to 2.269 usec with 8. This isn't a linear speed up as not all of
synthesize code has been made parallel. If the synthesis time was about
10 seconds then using 8 threads may bring this down to less than 4.
Signed-off-by: Stephane Eranian <eranian@google.com>
Reviewed-by: Ian Rogers <irogers@google.com>
Acked-by: Jiri Olsa <jolsa@redhat.com>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Alexey Budankov <alexey.budankov@linux.intel.com>
Cc: Kan Liang <kan.liang@linux.intel.com>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Tony Jones <tonyj@suse.de>
Cc: yuzhoujian <yuzhoujian@didichuxing.com>
Link: http://lore.kernel.org/lkml/20200422155038.9380-1-irogers@google.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Do not bother with close() if fd is not valid, just to silence valgrind:
$ valgrind ./perf script
==59169== Memcheck, a memory error detector
==59169== Copyright (C) 2002-2017, and GNU GPL'd, by Julian Seward et al.
==59169== Using Valgrind-3.14.0 and LibVEX; rerun with -h for copyright info
==59169== Command: ./perf script
==59169==
==59169== Warning: invalid file descriptor -1 in syscall close()
==59169== Warning: invalid file descriptor -1 in syscall close()
==59169== Warning: invalid file descriptor -1 in syscall close()
==59169== Warning: invalid file descriptor -1 in syscall close()
==59169== Warning: invalid file descriptor -1 in syscall close()
==59169== Warning: invalid file descriptor -1 in syscall close()
==59169== Warning: invalid file descriptor -1 in syscall close()
==59169== Warning: invalid file descriptor -1 in syscall close()
Signed-off-by: Tommi Rantala <tommi.t.rantala@nokia.com>
Acked-by: Jiri Olsa <jolsa@redhat.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lore.kernel.org/lkml/20200417132330.119407-1-tommi.t.rantala@nokia.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Perf checks the duplicate entries in a callchain before adding an entry.
However the check is very slow especially with deeper call stack.
Almost ~50% elapsed time of perf report is spent on the check when the
call stack is always depth of 32.
The hist_entry__cmp() is used to compare the new entry with the old
entries. It will go through all the available sorts in the sort_list,
and call the specific cmp of each sort, which is very slow.
Actually, for most cases, there are no duplicate entries in callchain.
The symbols are usually different. It's much faster to do a quick check
for symbols first. Only do the full cmp when the symbols are exactly the
same.
The quick check is only to check symbols, not dso. Export
_sort__sym_cmp.
$ perf record --call-graph lbr ./tchain_edit_64
Without the patch
$time perf report --stdio
real 0m21.142s
user 0m21.110s
sys 0m0.033s
With the patch
$time perf report --stdio
real 0m10.977s
user 0m10.948s
sys 0m0.027s
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@redhat.com>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Alexey Budankov <alexey.budankov@linux.intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Mathieu Poirier <mathieu.poirier@linaro.org>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Pavel Gerasimov <pavel.gerasimov@intel.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Ravi Bangoria <ravi.bangoria@linux.ibm.com>
Cc: Stephane Eranian <eranian@google.com>
Cc: Vitaly Slobodskoy <vitaly.slobodskoy@intel.com>
Link: http://lore.kernel.org/lkml/20200319202517.23423-18-kan.liang@linux.intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
In LBR call stack mode, the depth of reconstructed LBR call stack limits
to the number of LBR registers.
For example, on skylake, the depth of reconstructed LBR call stack is
always <= 32.
# To display the perf.data header info, please use
# --header/--header-only options.
#
#
# Total Lost Samples: 0
#
# Samples: 6K of event 'cycles'
# Event count (approx.): 6487119731
#
# Children Self Command Shared Object Symbol
# ........ ........ ............... ..................
# ................................
99.97% 99.97% tchain_edit tchain_edit [.] f43
|
--99.64%--f11
f12
f13
f14
f15
f16
f17
f18
f19
f20
f21
f22
f23
f24
f25
f26
f27
f28
f29
f30
f31
f32
f33
f34
f35
f36
f37
f38
f39
f40
f41
f42
f43
For a call stack which is deeper than LBR limit, HW will overwrite the
LBR register with oldest branch. Only partial call stacks can be
reconstructed.
However, the overwritten LBRs may still be retrieved from previous
sample. At that moment, HW hasn't overwritten the LBR registers yet.
Perf tools can stitch those overwritten LBRs on current call stacks to
get a more complete call stack.
To determine if LBRs can be stitched, perf tools need to compare current
sample with previous sample.
- They should have identical LBR records (Same from, to and flags
values, and the same physical index of LBR registers).
- The searching starts from the base-of-stack of current sample.
Once perf determines to stitch the previous LBRs, the corresponding LBR
cursor nodes will be copied to 'lists'. The 'lists' is to track the LBR
cursor nodes which are going to be stitched.
When the stitching is over, the nodes will not be freed immediately.
They will be moved to 'free_lists'. Next stitching may reuse the space.
Both 'lists' and 'free_lists' will be freed when all samples are
processed.
Committer notes:
Fix the intel-pt.c initialization of the union with 'struct
branch_flags', that breaks the build with its unnamed union on older gcc
versions.
Uninline thread__free_stitch_list(), as it grew big and started dragging
includes to thread.h, so move it to thread.c where what it needs in
terms of headers are already there.
This fixes the build in several systems such as debian:experimental when
cross building to the MIPS32 architecture, i.e. in the other cases what
was needed was being included by sheer luck.
In file included from builtin-sched.c:11:
util/thread.h: In function 'thread__free_stitch_list':
util/thread.h:169:3: error: implicit declaration of function 'free' [-Werror=implicit-function-declaration]
169 | free(pos);
| ^~~~
util/thread.h:169:3: error: incompatible implicit declaration of built-in function 'free' [-Werror]
util/thread.h:19:1: note: include '<stdlib.h>' or provide a declaration of 'free'
18 | #include "callchain.h"
+++ |+#include <stdlib.h>
19 |
util/thread.h:174:3: error: incompatible implicit declaration of built-in function 'free' [-Werror]
174 | free(pos);
| ^~~~
util/thread.h:174:3: note: include '<stdlib.h>' or provide a declaration of 'free'
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Reviewed-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@redhat.com>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Alexey Budankov <alexey.budankov@linux.intel.com>
Cc: Mathieu Poirier <mathieu.poirier@linaro.org>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Pavel Gerasimov <pavel.gerasimov@intel.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Ravi Bangoria <ravi.bangoria@linux.ibm.com>
Cc: Stephane Eranian <eranian@google.com>
Cc: Vitaly Slobodskoy <vitaly.slobodskoy@intel.com>
Link: http://lore.kernel.org/lkml/20200319202517.23423-13-kan.liang@linux.intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
LBR only collect the user call stack. To reconstruct a call stack, both
kernel call stack and user call stack are required. The function
resolve_lbr_callchain_sample() mix the kernel call stack and user call
stack.
Now, with the help of HW idx, perf tool can reconstruct a more complete
call stack by adding some user call stack from previous sample. However,
current implementation is hard to be extended to support it.
Current code path for resolve_lbr_callchain_sample()
for (j = 0; j < mix_chain_nr; j++) {
if (ORDER_CALLEE) {
if (kernel callchain)
Fill callchain info
else if (LBR callchain)
Fill callchain info
} else {
if (LBR callchain)
Fill callchain info
else if (kernel callchain)
Fill callchain info
}
add_callchain_ip();
}
With the patch,
if (ORDER_CALLEE) {
for (j = 0; j < NUM of kernel callchain) {
Fill callchain info
add_callchain_ip();
}
for (; j < mix_chain_nr) {
Fill callchain info
add_callchain_ip();
}
} else {
for (; j < NUM of LBR callchain) {
Fill callchain info
add_callchain_ip();
}
for (j = 0; j < mix_chain_nr) {
Fill callchain info
add_callchain_ip();
}
}
No functional changes.
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Reviewed-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@redhat.com>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Alexey Budankov <alexey.budankov@linux.intel.com>
Cc: Mathieu Poirier <mathieu.poirier@linaro.org>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Pavel Gerasimov <pavel.gerasimov@intel.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Ravi Bangoria <ravi.bangoria@linux.ibm.com>
Cc: Stephane Eranian <eranian@google.com>
Cc: Vitaly Slobodskoy <vitaly.slobodskoy@intel.com>
Link: http://lore.kernel.org/lkml/20200319202517.23423-7-kan.liang@linux.intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
The current rXXXX event specification creates event under PERF_TYPE_RAW
pmu type. This change allows to use rXXXX within pmu syntax, so it's
type is used via the following syntax:
-e 'cpu/r3c/'
-e 'cpum_cf/r0/'
The XXXX number goes directly to perf_event_attr::config the same way as
in '-e rXXXX' event. The perf_event_attr::type is filled with pmu type.
Committer testing:
So, lets see what goes in perf_event_attr::config for, say, the
'instructions' PERF_TYPE_HARDWARE (0) event, first we should look at how
to encode this event as a PERF_TYPE_RAW event for this specific CPU, an
AMD Ryzen 5:
# cat /sys/devices/cpu/events/instructions
event=0xc0
#
Then try with it _and_ the instruction, just to see that they are close
enough:
# perf stat -e rc0,instructions sleep 1
Performance counter stats for 'sleep 1':
919,794 rc0
919,898 instructions
1.000754579 seconds time elapsed
0.000715000 seconds user
0.000000000 seconds sys
#
Now we should try, before this patch, the PMU event encoding:
# perf stat -e cpu/rc0/ sleep 1
event syntax error: 'cpu/rc0/'
\___ unknown term
valid terms: event,edge,inv,umask,cmask,config,config1,config2,name,period,percore
#
Now with this patch, the three ways of specifying the 'instructions' CPU
counter are accepted:
# perf stat -e cpu/rc0/,rc0,instructions sleep 1
Performance counter stats for 'sleep 1':
892,948 cpu/rc0/
893,052 rc0
893,156 instructions
1.000931819 seconds time elapsed
0.000916000 seconds user
0.000000000 seconds sys
#
Requested-by: Thomas Richter <tmricht@linux.ibm.com>
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Tested-by: Thomas Richter <tmricht@linux.ibm.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Michael Petlan <mpetlan@redhat.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Sumanth Korikkar <sumanthk@linux.ibm.com>
Cc: Vasily Gorbik <gor@linux.ibm.com>
Link: http://lore.kernel.org/lkml/20200416221405.437788-1-jolsa@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
When it is not possible for a non-privilege perf command to monitor at
the kernel level (:k), the fallback code forces a :u. That works if the
event was previously monitoring both levels. But if the event was
already constrained to kernel only, then it does not make sense to
restrict it to user only.
Given the code works by exclusion, a kernel only event would have:
attr->exclude_user = 1
The fallback code would add:
attr->exclude_kernel = 1
In the end the end would not monitor in either the user level or kernel
level. In other words, it would count nothing.
An event programmed to monitor kernel only cannot be switched to user
only without seriously warning the user.
This patch forces an error in this case to make it clear the request
cannot really be satisfied.
Behavior with paranoid 1:
$ sudo bash -c "echo 1 > /proc/sys/kernel/perf_event_paranoid"
$ perf stat -e cycles:k sleep 1
Performance counter stats for 'sleep 1':
1,520,413 cycles:k
1.002361664 seconds time elapsed
0.002480000 seconds user
0.000000000 seconds sys
Old behavior with paranoid 2:
$ sudo bash -c "echo 2 > /proc/sys/kernel/perf_event_paranoid"
$ perf stat -e cycles:k sleep 1
Performance counter stats for 'sleep 1':
0 cycles:ku
1.002358127 seconds time elapsed
0.002384000 seconds user
0.000000000 seconds sys
New behavior with paranoid 2:
$ sudo bash -c "echo 2 > /proc/sys/kernel/perf_event_paranoid"
$ perf stat -e cycles:k sleep 1
Error:
You may not have permission to collect stats.
Consider tweaking /proc/sys/kernel/perf_event_paranoid,
which controls use of the performance events system by
unprivileged users (without CAP_PERFMON or CAP_SYS_ADMIN).
The current value is 2:
-1: Allow use of (almost) all events by all users
Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK
>= 0: Disallow ftrace function tracepoint by users without CAP_PERFMON or CAP_SYS_ADMIN
Disallow raw tracepoint access by users without CAP_SYS_PERFMON or CAP_SYS_ADMIN
>= 1: Disallow CPU event access by users without CAP_PERFMON or CAP_SYS_ADMIN
>= 2: Disallow kernel profiling by users without CAP_PERFMON or CAP_SYS_ADMIN
To make this setting permanent, edit /etc/sysctl.conf too, e.g.:
kernel.perf_event_paranoid = -1
v2 of this patch addresses the review feedback from jolsa@redhat.com.
Signed-off-by: Stephane Eranian <eranian@google.com>
Reviewed-by: Ian Rogers <irogers@google.com>
Acked-by: Jiri Olsa <jolsa@redhat.com>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lore.kernel.org/lkml/20200414161550.225588-1-irogers@google.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
When AUX area events are used in sampling mode, they must be the group
leader, but the group leader is also used for leader-sampling. However,
it is not desirable to use an AUX area event as the leader for
leader-sampling, because it doesn't have any samples of its own. To support
leader-sampling with AUX area events, use the 2nd event of the group as the
"leader" for the purposes of leader-sampling.
Example:
# perf record --kcore --aux-sample -e '{intel_pt//,cycles,instructions}:S' -c 10000 uname
[ perf record: Woken up 3 times to write data ]
[ perf record: Captured and wrote 0.786 MB perf.data ]
# perf report
Samples: 380 of events 'anon group { cycles, instructions }', Event count (approx.): 3026164
Children Self Command Shared Object Symbol
+ 38.76% 42.65% 0.00% 0.00% uname [kernel.kallsyms] [k] __x86_indirect_thunk_rax
+ 35.82% 31.33% 0.00% 0.00% uname ld-2.28.so [.] _dl_start_user
+ 34.29% 29.74% 0.55% 0.47% uname ld-2.28.so [.] _dl_start
+ 33.73% 28.62% 1.60% 0.97% uname ld-2.28.so [.] dl_main
+ 33.19% 29.04% 0.52% 0.32% uname ld-2.28.so [.] _dl_sysdep_start
+ 27.83% 33.74% 0.00% 0.00% uname [kernel.kallsyms] [k] do_syscall_64
+ 26.76% 33.29% 0.00% 0.00% uname [kernel.kallsyms] [k] entry_SYSCALL_64_after_hwframe
+ 23.78% 20.33% 5.97% 5.25% uname [kernel.kallsyms] [k] page_fault
+ 23.18% 24.60% 0.00% 0.00% uname libc-2.28.so [.] __libc_start_main
+ 22.64% 24.37% 0.00% 0.00% uname uname [.] _start
+ 21.04% 23.27% 0.00% 0.00% uname uname [.] main
+ 19.48% 18.08% 3.72% 3.64% uname ld-2.28.so [.] _dl_relocate_object
+ 19.47% 21.81% 0.00% 0.00% uname libc-2.28.so [.] setlocale
+ 19.44% 21.56% 0.52% 0.61% uname libc-2.28.so [.] _nl_find_locale
+ 17.87% 19.66% 0.00% 0.00% uname libc-2.28.so [.] _nl_load_locale_from_archive
+ 15.71% 13.73% 0.53% 0.52% uname [kernel.kallsyms] [k] do_page_fault
+ 15.18% 13.21% 1.03% 0.68% uname [kernel.kallsyms] [k] handle_mm_fault
+ 14.15% 12.53% 1.01% 1.12% uname [kernel.kallsyms] [k] __handle_mm_fault
+ 12.03% 9.67% 0.54% 0.32% uname ld-2.28.so [.] _dl_map_object
+ 10.55% 8.48% 0.00% 0.00% uname ld-2.28.so [.] openaux
+ 10.55% 20.20% 0.52% 0.61% uname libc-2.28.so [.] __run_exit_handlers
Comnmitter notes:
Fixed up this problem:
util/record.c: In function ‘perf_evlist__config’:
util/record.c:256:3: error: too few arguments to function ‘perf_evsel__config_leader_sampling’
256 | perf_evsel__config_leader_sampling(evsel);
| ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
util/record.c:190:13: note: declared here
190 | static void perf_evsel__config_leader_sampling(struct evsel *evsel,
| ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Signed-off-by: Adrian Hunter <adrian.hunter@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Link: http://lore.kernel.org/lkml/20200401101613.6201-17-adrian.hunter@intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Move leader-sampling configuration in preparation for adding support for
leader sampling with AUX area events.
Committer notes:
It only makes sense when configuring an evsel that is part of an evlist,
so the only case where it is called outside perf_evlist__config(), in
some 'perf test' entry, is safe, and even there we should just use
perf_evlist__config(), but since in that case we have just one evsel in
the evlist, it is equivalent.
Also fixed up this problem:
util/record.c: In function ‘perf_evlist__config’:
util/record.c:223:3: error: too many arguments to function ‘perf_evsel__config_leader_sampling’
223 | perf_evsel__config_leader_sampling(evsel, evlist);
| ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
util/record.c:170:13: note: declared here
170 | static void perf_evsel__config_leader_sampling(struct evsel *evsel)
| ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Signed-off-by: Adrian Hunter <adrian.hunter@intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Link: http://lore.kernel.org/lkml/20200401101613.6201-14-adrian.hunter@intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
