Improve the error message returned on failed perf_event_open() on AMD
systems when using IBS (Instruction-Based Sampling).
Output of executing 'perf record -e ibs_op// true' as a non root user
BEFORE this patch (perf will add the 'u' modifier at the end to exclude
kernel/hypervisor sampling):
The sys_perf_event_open() syscall returned with 22 (Invalid argument)for event (ibs_op//u).
/bin/dmesg | grep -i perf may provide additional information.
Output after:
AMD IBS can't exclude kernel events. Try running at a higher privilege level.
Output of executing 'sudo perf record -e ibs_op// true' BEFORE this patch:
Error:
The sys_perf_event_open() syscall returned with 22 (Invalid argument) for event (ibs_op//).
/bin/dmesg | grep -i perf may provide additional information.
Output after:
Error:
Invalid event (ibs_op//) in per-thread mode, enable system wide with '-a'.
Folowing the suggestion:
$ sudo perf record -a -e ibs_op// true
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 1.664 MB perf.data (194 samples) ]
$
Signed-off-by: Kim Phillips <kim.phillips@amd.com>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com>
Cc: Ian Rogers <irogers@google.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: João Martins <joao.m.martins@oracle.com>
Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Michael Petlan <mpetlan@redhat.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Rafael J. Wysocki <rafael@kernel.org>
Cc: Ravi Bangoria <ravi.bangoria@amd.com>
Cc: Robert Richter <robert.richter@amd.com>
Cc: Sandipan Das <sandipan.das@amd.com>
Cc: Song Liu <songliubraving@fb.com>
Cc: Stephane Eranian <eranian@google.com>
Link: http://lore.kernel.org/lkml/20220322221517.2510440-12-eranian@google.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
EOPNOTSUPP is a possible return value when branch stacks are requested
but they aren't enabled in the kernel or hardware. It's also returned if
they aren't supported on the specific event type. The currently printed
error message about sampling/overflow-interrupts is not correct in this
case.
Add a check for branch stacks before sample_period is checked because
sample_period is also set (to the default value) when using branch
stacks.
Before this change (when branch stacks aren't supported):
perf record -j any
Error:
cycles: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'
After this change:
perf record -j any
Error:
cycles: PMU Hardware or event type doesn't support branch stack sampling.
Signed-off-by: James Clark <james.clark@arm.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Anshuman Khandual <anshuman.khandual@arm.com>
Cc: German Gomez <german.gomez@arm.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Leo Yan <leo.yan@linaro.org>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Link: https://lore.kernel.org/r/20220307171917.2555829-2-james.clark@arm.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Pull perf tool updates from Arnaldo Carvalho de Melo:
"New features:
- Add 'trace' subcommand for 'perf ftrace', setting the stage for
more 'perf ftrace' subcommands. Not using a subcommand yields the
previous behaviour of 'perf ftrace'.
- Add 'latency' subcommand to 'perf ftrace', that can use the
function graph tracer or a BPF optimized one, via the -b/--use-bpf
option.
E.g.:
$ sudo perf ftrace latency -a -T mutex_lock sleep 1
# DURATION | COUNT | GRAPH |
0 - 1 us | 4596 | ######################## |
1 - 2 us | 1680 | ######### |
2 - 4 us | 1106 | ##### |
4 - 8 us | 546 | ## |
8 - 16 us | 562 | ### |
16 - 32 us | 1 | |
32 - 64 us | 0 | |
64 - 128 us | 0 | |
128 - 256 us | 0 | |
256 - 512 us | 0 | |
512 - 1024 us | 0 | |
1 - 2 ms | 0 | |
2 - 4 ms | 0 | |
4 - 8 ms | 0 | |
8 - 16 ms | 0 | |
16 - 32 ms | 0 | |
32 - 64 ms | 0 | |
64 - 128 ms | 0 | |
128 - 256 ms | 0 | |
256 - 512 ms | 0 | |
512 - 1024 ms | 0 | |
1 - ... s | 0 | |
The original implementation of this command was in the bcc tool.
- Support --cputype option for hybrid events in 'perf stat'.
Improvements:
- Call chain improvements for ARM64.
- No need to do any affinity setup when profiling pids.
- Reduce multiplexing with duration_time in 'perf stat' metrics.
- Improve error message for uncore events, stating that some event
groups are can only be used in system wide (-a) mode.
- perf stat metric group leader fixes/improvements, including arch
specific changes to better support Intel topdown events.
- Probe non-deprecated sysfs path first, i.e. try the path
/sys/devices/system/cpu/cpuN/topology/thread_siblings first, then
the old /sys/devices/system/cpu/cpuN/topology/core_cpus.
- Disable debuginfod by default in 'perf record', to avoid stalls on
distros such as Fedora 35.
- Use unbuffered output in 'perf bench' when pipe/tee'ing to a file.
- Enable ignore_missing_thread in 'perf trace'
Fixes:
- Avoid TUI crash when navigating in the annotation of recursive
functions.
- Fix hex dump character output in 'perf script'.
- Fix JSON indentation to 4 spaces standard in the ARM vendor event
files.
- Fix use after free in metric__new().
- Fix IS_ERR_OR_NULL() usage in the perf BPF loader.
- Fix up cross-arch register support, i.e. when printing register
names take into account the architecture where the perf.data file
was collected.
- Fix SMT fallback with large core counts.
- Don't lower case MetricExpr when parsing JSON files so as not to
lose info such as the ":G" event modifier in metrics.
perf test:
- Add basic stress test for sigtrap handling to 'perf test'.
- Fix 'perf test' failures on s/390
- Enable system wide for metricgroups test in 'perf test´.
- Use 3 digits for test numbering now we can have more tests.
Arch specific:
- Add events for Arm Neoverse N2 in the ARM JSON vendor event files
- Support PERF_MEM_LVLNUM encodings in powerpc, that came from a
single patch series, where I incorrectly merged the kernel bits,
that were then reverted after coordination with Michael Ellerman
and Stephen Rothwell.
- Add ARM SPE total latency as PERF_SAMPLE_WEIGHT.
- Update AMD documentation, with info on raw event encoding.
- Add support for global and local variants of the "p_stage_cyc" sort
key, applicable to perf.data files collected on powerpc.
- Remove duplicate and incorrect aux size checks in the ARM CoreSight
ETM code.
Refactorings:
- Add a perf_cpu abstraction to disambiguate CPUs and CPU map
indexes, fixing problems along the way.
- Document CPU map methods.
UAPI sync:
- Update arch/x86/lib/mem{cpy,set}_64.S copies used in 'perf bench
mem memcpy'
- Sync UAPI files with the kernel sources: drm, msr-index,
cpufeatures.
Build system
- Enable warnings through HOSTCFLAGS.
- Drop requirement for libstdc++.so for libopencsd check
libperf:
- Make libperf adopt perf_counts_values__scale() from tools/perf/util/.
- Add a stat multiplexing test to libperf"
* tag 'perf-tools-for-v5.17-2022-01-16' of git://git.kernel.org/pub/scm/linux/kernel/git/acme/linux: (115 commits)
perf record: Disable debuginfod by default
perf evlist: No need to do any affinity setup when profiling pids
perf cpumap: Add is_dummy() method
perf metric: Fix metric_leader
perf cputopo: Fix CPU topology reading on s/390
perf metricgroup: Fix use after free in metric__new()
libperf tests: Update a use of the new cpumap API
perf arm: Fix off-by-one directory path
tools arch x86: Sync the msr-index.h copy with the kernel sources
tools headers cpufeatures: Sync with the kernel sources
tools headers UAPI: Update tools's copy of drm.h header
tools arch: Update arch/x86/lib/mem{cpy,set}_64.S copies used in 'perf bench mem memcpy'
perf pmu-events: Don't lower case MetricExpr
perf expr: Add debug logging for literals
perf tools: Probe non-deprecated sysfs path 1st
perf tools: Fix SMT fallback with large core counts
perf cpumap: Give CPUs their own type
perf stat: Correct first_shadow_cpu to return index
perf script: Fix flipped index and cpu
perf c2c: Use more intention revealing iterator
...
When a group has multiple events and the leader fails it can yield
errors like:
$ perf stat -e '{uncore_imc/cas_count_read/},instructions' /bin/true
Error:
The sys_perf_event_open() syscall returned with 22 (Invalid argument) for event (uncore_imc/cas_count_read/).
/bin/dmesg | grep -i perf may provide additional information.
However, when not the group leader <not supported> is given:
$ perf stat -e '{instructions,uncore_imc/cas_count_read/}' /bin/true
...
1,619,057 instructions
<not supported> MiB uncore_imc/cas_count_read/
This is necessary because get_group_fd will fail if the leader fails and
is the direct result of the check on line 750 of builtin-stat.c in
stat_handle_error that returns COUNTER_SKIP for the latter case.
This patch improves the error message to:
$ perf stat -e '{uncore_imc/cas_count_read/},instructions' /bin/true
Error:
Invalid event (uncore_imc/cas_count_read/) in per-thread mode, enable system wide with '-a'.
v2. Changed the test to use !target__has_cpu as suggested by Namhyung Kim.
Signed-off-by: Ian Rogers <irogers@google.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Ingo Molnar <mingo@redhat.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>
Cc: Stephane Eranian <eranian@google.com>
Link: https://lore.kernel.org/r/20211223183948.3423989-2-irogers@google.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Perf tool sets exclude_guest by default while calling perf_event_open().
Because IBS does not have filtering capability, it always gets rejected
by IBS PMU driver and thus perf falls back to non-precise sampling. Fix
it by not setting exclude_guest by default on AMD.
Before:
$ sudo ./perf record -C 0 -vvv true |& grep precise
precise_ip 3
decreasing precise_ip by one (2)
precise_ip 2
decreasing precise_ip by one (1)
precise_ip 1
decreasing precise_ip by one (0)
After:
$ sudo ./perf record -C 0 -vvv true |& grep precise
precise_ip 3
decreasing precise_ip by one (2)
precise_ip 2
Committer notes:
Fixup init to zero for perf_env in older compilers:
arch/x86/util/evsel.c:15:26: error: missing field 'os_release' initializer [-Werror,-Wmissing-field-initializers]
struct perf_env env = {0};
^
Committer notes:
Namhyung remarked:
It'd be nice if it can cover explicit "-e cycles:pp" as well.
Ravi clarified:
For explicit :pp modifier, evsel->precise_max does not get set and thus perf
does not try with different attr->precise_ip values while exclude_guest set.
So no issue with explicit :pp:
$ sudo ./perf record -C 0 -e cycles:pp -vvv |& grep "precise_ip\|exclude_guest"
precise_ip 2
exclude_guest 1
precise_ip 2
exclude_guest 1
switching off exclude_guest, exclude_host
precise_ip 2
^C
Also, with :P modifier, evsel->precise_max gets set but exclude_guest does
not and thus :P also works fine:
$ sudo ./perf record -C 0 -e cycles:P -vvv |& grep "precise_ip\|exclude_guest"
precise_ip 3
decreasing precise_ip by one (2)
precise_ip 2
^C
Reported-by: Kim Phillips <kim.phillips@amd.com>
Signed-off-by: Ravi Bangoria <ravi.bangoria@amd.com>
Acked-by: Namhyung Kim <namhyung@kernel.org>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Ian Rogers <irogers@google.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@google.com>
Link: http://lore.kernel.org/lkml/20211103072112.32312-1-ravi.bangoria@amd.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
The current logic for the perf missing feature has a bug that it can
wrongly clear some modifiers like G or H. Actually some PMUs don't
support any filtering or exclusion while others do. But we check it as
a global feature.
For example, the cycles event can have 'G' modifier to enable it only in
the guest mode on x86. When you don't run any VMs it'll return 0.
# perf stat -a -e cycles:G sleep 1
Performance counter stats for 'system wide':
0 cycles:G
1.000721670 seconds time elapsed
But when it's used with other pmu events that don't support G modifier,
it'll be reset and return non-zero values.
# perf stat -a -e cycles:G,msr/tsc/ sleep 1
Performance counter stats for 'system wide':
538,029,960 cycles:G
16,924,010,738 msr/tsc/
1.001815327 seconds time elapsed
This is because of the missing feature detection logic being global.
Add a hashmap to set pmu-specific exclude_host/guest features.
Committer notes:
Fix 'perf test python' by adding a stub for evsel__find_pmu() in
tools/perf/util/python.c, document that it is used so far only for the
above reasons so that if anybody needs this in the python binding
usecases, we can revisit this.
Reported-by: Stephane Eranian <eranian@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Ian Rogers <irogers@google.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Ravi Bangoria <ravi.bangoria@amd.com>
Link: http://lore.kernel.org/lkml/20211105205847.120950-1-namhyung@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Move evsel::leader to perf_evsel::leader, so we can move the group
interface to libperf.
Also add several evsel helpers to ease up the transition:
struct evsel *evsel__leader(struct evsel *evsel);
- get leader evsel
bool evsel__has_leader(struct evsel *evsel, struct evsel *leader);
- true if evsel has leader as leader
bool evsel__is_leader(struct evsel *evsel);
- true if evsel is itw own leader
void evsel__set_leader(struct evsel *evsel, struct evsel *leader);
- set leader for evsel
Committer notes:
Fix this when building with 'make BUILD_BPF_SKEL=1'
tools/perf/util/bpf_counter.c
- if (evsel->leader->core.nr_members > 1) {
+ if (evsel->core.leader->nr_members > 1) {
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
Requested-by: Shunsuke Nakamura <nakamura.shun@fujitsu.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Ian Rogers <irogers@google.com>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Michael Petlan <mpetlan@redhat.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lore.kernel.org/lkml/20210706151704.73662-4-jolsa@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
If a group has events which are from different hybrid PMUs,
shows a warning:
"WARNING: events in group from different hybrid PMUs!"
This is to remind the user not to put the core event and atom
event into one group.
Next, just disable grouping.
# perf stat -e "{cpu_core/cycles/,cpu_atom/cycles/}" -a -- sleep 1
WARNING: events in group from different hybrid PMUs!
WARNING: grouped events cpus do not match, disabling group:
anon group { cpu_core/cycles/, cpu_atom/cycles/ }
Performance counter stats for 'system wide':
5,438,125 cpu_core/cycles/
3,914,586 cpu_atom/cycles/
1.004250966 seconds time elapsed
Signed-off-by: Jin Yao <yao.jin@linux.intel.com>
Reviewed-by: Jiri Olsa <jolsa@kernel.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Kan Liang <kan.liang@intel.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: https://lore.kernel.org/r/20210427070139.25256-17-yao.jin@linux.intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Currently, to use BPF to aggregate perf event counters, the user uses
--bpf-counters option. Enable "use bpf by default" events with a config
option, stat.bpf-counter-events. Events with name in the option will use
BPF.
This also enables mixed BPF event and regular event in the same sesssion.
For example:
perf config stat.bpf-counter-events=instructions
perf stat -e instructions,cs
The second command will use BPF for "instructions" but not "cs".
Signed-off-by: Song Liu <song@kernel.org>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Song Liu <songliubraving@fb.com>
Link: https://lore.kernel.org/r/20210425214333.1090950-4-song@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
For X86, the var2_w field of PERF_SAMPLE_WEIGHT_STRUCT stands for the
instruction latency. Current perf forces the var2_w to the data->ins_lat
in the generic code. It works well for now because X86 is the only
architecture that supports the PERF_SAMPLE_WEIGHT_STRUCT, but it may
bring problems once other architectures support the sample type. For
example, the var2_w may be used to capture something else on PowerPC.
Create two architecture specific functions to parse and synthesize the
weight related samples. Move the X86 specific codes to the X86 version
functions. Other architectures can implement their own functions later
separately.
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Athira Jajeev <atrajeev@linux.vnet.ibm.com>
Cc: Jin Yao <yao.jin@linux.intel.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Madhavan Srinivasan <maddy@linux.vnet.ibm.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Link: http://lore.kernel.org/lkml/1612540912-6562-1-git-send-email-kan.liang@linux.intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
The instruction latency information can be recorded on some platforms,
e.g., the Intel Sapphire Rapids server. With both memory latency
(weight) and the new instruction latency information, users can easily
locate the expensive load instructions, and also understand the time
spent in different stages. The users can optimize their applications in
different pipeline stages.
The 'weight' field is shared among different architectures. Reusing the
'weight' field may impacts other architectures. Add a new field to store
the instruction latency.
Like the 'weight' support, introduce a 'ins_lat' for the global
instruction latency, and a 'local_ins_lat' for the local instruction
latency version.
Add new sort functions, INSTR Latency and Local INSTR Latency,
accordingly.
Add local_ins_lat to the default_mem_sort_order[].
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Jin Yao <yao.jin@linux.intel.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Madhavan Srinivasan <maddy@linux.vnet.ibm.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@google.com>
Link: http://lore.kernel.org/lkml/1612296553-21962-7-git-send-email-kan.liang@linux.intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
The new sample type, PERF_SAMPLE_WEIGHT_STRUCT, is an alternative of the
PERF_SAMPLE_WEIGHT sample type. Users can apply either the
PERF_SAMPLE_WEIGHT sample type or the PERF_SAMPLE_WEIGHT_STRUCT sample
type to retrieve the sample weight, but they cannot apply both sample
types simultaneously.
The new sample type shares the same space as the PERF_SAMPLE_WEIGHT
sample type. The lower 32 bits are exactly the same for both sample
type. The higher 32 bits may be different for different architecture.
Add arch specific arch_evsel__set_sample_weight() to set the new sample
type for X86. Only store the lower 32 bits for the sample->weight if the
new sample type is applied. In practice, no memory access could last
than 4G cycles. No data will be lost.
If the kernel doesn't support the new sample type. Fall back to the
PERF_SAMPLE_WEIGHT sample type.
There is no impact for other architectures.
Committer notes:
Fixup related to PERF_SAMPLE_CODE_PAGE_SIZE, present in acme/perf/core
but not upstream yet.
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Jin Yao <yao.jin@linux.intel.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Madhavan Srinivasan <maddy@linux.vnet.ibm.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@google.com>
Link: http://lore.kernel.org/lkml/1612296553-21962-6-git-send-email-kan.liang@linux.intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
On the Intel Sapphire Rapids server, an auxiliary event has to be
enabled simultaneously with the load latency event to retrieve complete
Memory Info.
Add X86 specific perf_mem_events__name() to handle the auxiliary event.
- Users are only interested in the samples of the mem-loads event.
Sample read the auxiliary event.
- The auxiliary event must be in front of the load latency event in a
group. Assume the second event to sample if the auxiliary event is the
leader.
- Add a weak is_mem_loads_aux_event() to check the auxiliary event for
X86. For other ARCHs, it always return false.
Parse the unique event name, mem-loads-aux, for the auxiliary event.
Committer notes:
According to 61b985e3e7 ("perf/x86/intel: Add perf core PMU
support for Sapphire Rapids"), ENODATA is only returned by
sys_perf_event_open() when used with these auxiliary events, with this
in evsel__open_strerror():
case ENODATA:
return scnprintf(msg, size, "Cannot collect data source with the load latency event alone. "
"Please add an auxiliary event in front of the load latency event.");
This is Ok at this point in time, but fragile long term, I pointed this
out in the e-mail thread, requesting a follow up patch to check if
ENODATA is really for this specific case.
Fixed up sizeof(MEM_LOADS_AUX_NAME) bug pointed out by Namhyung.
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Jin Yao <yao.jin@linux.intel.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Madhavan Srinivasan <maddy@linux.vnet.ibm.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@google.com>
Link: http://lore.kernel.org/lkml/20210205152648.GC920417@kernel.org
Link: http://lore.kernel.org/lkml/1612296553-21962-3-git-send-email-kan.liang@linux.intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Introduce 'perf stat -b' option, which counts events for BPF programs, like:
[root@localhost ~]# ~/perf stat -e ref-cycles,cycles -b 254 -I 1000
1.487903822 115,200 ref-cycles
1.487903822 86,012 cycles
2.489147029 80,560 ref-cycles
2.489147029 73,784 cycles
3.490341825 60,720 ref-cycles
3.490341825 37,797 cycles
4.491540887 37,120 ref-cycles
4.491540887 31,963 cycles
The example above counts 'cycles' and 'ref-cycles' of BPF program of id
254. This is similar to bpftool-prog-profile command, but more
flexible.
'perf stat -b' creates per-cpu perf_event and loads fentry/fexit BPF
programs (monitor-progs) to the target BPF program (target-prog). The
monitor-progs read perf_event before and after the target-prog, and
aggregate the difference in a BPF map. Then the user space reads data
from these maps.
A new 'struct bpf_counter' is introduced to provide a common interface
that uses BPF programs/maps to count perf events.
Committer notes:
Removed all but bpf_counter.h includes from evsel.h, not needed at all.
Also BPF map lookups for PERCPU_ARRAYs need to have as its value receive
buffer passed to the kernel libbpf_num_possible_cpus() entries, not
evsel__nr_cpus(evsel), as the former uses
/sys/devices/system/cpu/possible while the later uses
/sys/devices/system/cpu/online, which may be less than the 'possible'
number making the bpf map lookup overwrite memory and cause hard to
debug memory corruption.
We need to continue using evsel__nr_cpus(evsel) when accessing the
perf_counts array tho, not to overwrite another are of memory :-)
Signed-off-by: Song Liu <songliubraving@fb.com>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Link: https://lore.kernel.org/lkml/20210120163031.GU12699@kernel.org/
Acked-by: Namhyung Kim <namhyung@kernel.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: kernel-team@fb.com
Link: http://lore.kernel.org/lkml/20201229214214.3413833-4-songliubraving@fb.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Add --buildid-mmap option to enable build id in PERF_RECORD_MMAP2 events.
It will only work if there's kernel support for that and it disables
build id cache (implies --no-buildid).
It's also possible to enable it permanently via config option in
~/.perfconfig file:
[record]
build-id=mmap
Also added build_id bit in the verbose output for perf_event_attr:
# perf record --buildid-mmap -vv
...
perf_event_attr:
type 1
size 120
...
build_id 1
Adding also missing text_poke bit.
Committer testing:
$ perf record -h build
Usage: perf record [<options>] [<command>]
or: perf record [<options>] -- <command> [<options>]
-B, --no-buildid do not collect buildids in perf.data
-N, --no-buildid-cache
do not update the buildid cache
--buildid-all Record build-id of all DSOs regardless of hits
--buildid-mmap Record build-id in map events
$
$ perf record --buildid-mmap sleep 1
Failed: no support to record build id in mmap events, update your kernel.
$
After adding the needed kernel bits in a test kernel:
$ perf record -vv --buildid-mmap sleep 1 |& grep -m1 build
Enabling build id in mmap2 events.
$ perf evlist -v
cycles:u: size: 120, { sample_period, sample_freq }: 4000, sample_type: IP|TID|TIME|PERIOD, read_format: ID, disabled: 1, inherit: 1, exclude_kernel: 1, mmap: 1, comm: 1, freq: 1, enable_on_exec: 1, task: 1, sample_id_all: 1, exclude_guest: 1, mmap2: 1, comm_exec: 1, ksymbol: 1, bpf_event: 1, build_id: 1
$
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Acked-by: Ian Rogers <irogers@google.com>
Acked-by: Namhyung Kim <namhyung@kernel.org>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Alexei Budankov <abudankov@huawei.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Michael Petlan <mpetlan@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Song Liu <songliubraving@fb.com>
Cc: Stephane Eranian <eranian@google.com>
Link: http://lore.kernel.org/lkml/20201214105457.543111-16-jolsa@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
