Counting events system-wide with a specified CPU prior to this change
worked:
```
$ perf stat -e 'msr/tsc/,msr/tsc,cpu=cpu_core/,msr/tsc,cpu=cpu_atom/' -a sleep 1
Performance counter stats for 'system wide':
59,393,419,099 msr/tsc/
33,927,965,927 msr/tsc,cpu=cpu_core/
25,465,608,044 msr/tsc,cpu=cpu_atom/
```
However, when counting with process the counts became system wide:
```
$ perf stat -e 'msr/tsc/,msr/tsc,cpu=cpu_core/,msr/tsc,cpu=cpu_atom/' perf test -F 10
10.1: Basic parsing test : Ok
10.2: Parsing without PMU name : Ok
10.3: Parsing with PMU name : Ok
Performance counter stats for 'perf test -F 10':
59,233,549 msr/tsc/
59,227,556 msr/tsc,cpu=cpu_core/
59,224,053 msr/tsc,cpu=cpu_atom/
```
Make the handling of CPU maps with event parsing clearer. When an
event is parsed creating an evsel the cpus should be either the PMU's
cpumask or user specified CPUs.
Update perf_evlist__propagate_maps so that it doesn't clobber the user
specified CPUs. Try to make the behavior clearer, firstly fix up
missing cpumasks. Next, perform sanity checks and adjustments from the
global evlist CPU requests and for the PMU including simplifying to
the "any CPU"(-1) value. Finally remove the event if the cpumask is
empty.
So that events are opened with a CPU and a thread change stat's
create_perf_stat_counter to give both.
With the change things are fixed:
```
$ perf stat --no-scale -e 'msr/tsc/,msr/tsc,cpu=cpu_core/,msr/tsc,cpu=cpu_atom/' perf test -F 10
10.1: Basic parsing test : Ok
10.2: Parsing without PMU name : Ok
10.3: Parsing with PMU name : Ok
Performance counter stats for 'perf test -F 10':
63,704,975 msr/tsc/
47,060,704 msr/tsc,cpu=cpu_core/ (4.62%)
16,640,591 msr/tsc,cpu=cpu_atom/ (2.18%)
```
However, note the "--no-scale" option is used. This is necessary as
the running time for the event on the counter isn't the same as the
enabled time because the thread doesn't necessarily run on the CPUs
specified for the counter. All counter values are scaled with:
scaled_value = value * time_enabled / time_running
and so without --no-scale the scaled_value becomes very large. This
problem already exists on hybrid systems for the same reason. Here are
2 runs of the same code with an instructions event that counts the
same on both types of core, there is no real multiplexing happening on
the event:
```
$ perf stat -e instructions perf test -F 10
...
Performance counter stats for 'perf test -F 10':
87,896,447 cpu_atom/instructions/ (14.37%)
98,171,964 cpu_core/instructions/ (85.63%)
...
$ perf stat --no-scale -e instructions perf test -F 10
...
Performance counter stats for 'perf test -F 10':
13,069,890 cpu_atom/instructions/ (19.32%)
83,460,274 cpu_core/instructions/ (80.68%)
...
```
The scaling has inflated per-PMU instruction counts and the overall
count by 2x.
To fix this the kernel needs changing when a task+CPU event (or just
task event on hybrid) is scheduled out. A fix could be that the state
isn't inactive but off for such events, so that time_enabled counts
don't accumulate on them.
Reviewed-by: Thomas Falcon <thomas.falcon@intel.com>
Signed-off-by: Ian Rogers <irogers@google.com>
Link: https://lore.kernel.org/r/20250719030517.1990983-13-irogers@google.com
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Look for .rodata maps, find ones with 'bpf_metadata_' variables, extract
their values as strings, and create a new PERF_RECORD_BPF_METADATA
synthetic event using that data. The code gets invoked from the existing
routine perf_event__synthesize_one_bpf_prog().
For example, a BPF program with the following variables:
const char bpf_metadata_version[] SEC(".rodata") = "3.14159";
int bpf_metadata_value[] SEC(".rodata") = 42;
would generate a PERF_RECORD_BPF_METADATA record with:
.prog_name = <BPF program name, e.g. "bpf_prog_a1b2c3_foo">
.nr_entries = 2
.entries[0].key = "version"
.entries[0].value = "3.14159"
.entries[1].key = "value"
.entries[1].value = "42"
Each of the BPF programs and subprograms that share those variables would
get a distinct PERF_RECORD_BPF_METADATA record, with the ".prog_name"
showing the name of each program or subprogram. The prog_name is
deliberately the same as the ".name" field in the corresponding
PERF_RECORD_KSYMBOL record.
This code only gets invoked if support for displaying BTF char arrays
as strings is detected.
Signed-off-by: Blake Jones <blakejones@google.com>
Link: https://lore.kernel.org/r/20250612194939.162730-3-blakejones@google.com
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
The original PERF_RECORD_COMPRESS is not 8-byte aligned, which can cause
asan runtime error:
# Build with asan
$ make -C tools/perf O=/tmp/perf DEBUG=1 EXTRA_CFLAGS="-O0 -g -fno-omit-frame-pointer -fsanitize=undefined"
# Test success with many asan runtime errors:
$ /tmp/perf/perf test "Zstd perf.data compression/decompression" -vv
83: Zstd perf.data compression/decompression:
...
util/session.c:1959:13: runtime error: member access within misaligned address 0x7f69e3f99653 for type 'union perf_event', which requires 13 byte alignment
0x7f69e3f99653: note: pointer points here
d0 3a 50 69 44 00 00 00 00 00 08 00 bb 07 00 00 00 00 00 00 44 00 00 00 00 00 00 00 ff 07 00 00
^
util/session.c:2163:22: runtime error: member access within misaligned address 0x7f69e3f99653 for type 'union perf_event', which requires 8 byte alignment
0x7f69e3f99653: note: pointer points here
d0 3a 50 69 44 00 00 00 00 00 08 00 bb 07 00 00 00 00 00 00 44 00 00 00 00 00 00 00 ff 07 00 00
^
...
Since there is no way to align compressed data in zstd compression, this
patch add a new event type `PERF_RECORD_COMPRESSED2`, which adds a field
`data_size` to specify the actual compressed data size.
The `header.size` contains the total record size, including the padding
at the end to make it 8-byte aligned.
Tested with `Zstd perf.data compression/decompression`
Signed-off-by: Chun-Tse Shao <ctshao@google.com>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Ben Gainey <ben.gainey@arm.com>
Cc: Christophe Leroy <christophe.leroy@csgroup.eu>
Cc: Ian Rogers <irogers@google.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: James Clark <james.clark@linaro.org>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Kan Liang <kan.liang@linux.intel.com>
Cc: Leo Yan <leo.yan@arm.com>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Nick Terrell <terrelln@fb.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: https://lore.kernel.org/r/20250303183646.327510-1-ctshao@google.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
When libperf is built alone in-source, $(OUTPUT) isn't set. This causes
the generated uapi path to resolve to '/../arch' which results in a
permissions error:
mkdir: cannot create directory '/../arch': Permission denied
Fix it by removing the preceding '/..' which means that it gets
generated either in the tools/lib/perf part of the tree or the OUTPUT
folder. Some other rules that rely on OUTPUT further refine this
conditionally depending on whether it's an in-source or out-of-source
build, but I don't think we need the extra complexity here. And this
rule is slightly different to others because the header is needed by
both libperf and Perf. This is further complicated by the fact that Perf
always passes O=... to libperf even for in source builds, meaning that
OUTPUT isn't set consistently between projects.
Because we're no longer going one level up to try to generate the file
in the tools/ folder, Perf's include rule needs to descend into libperf.
Also fix the clean rule while we're here.
Reported-by: Thorsten Leemhuis <linux@leemhuis.info>
Closes: https://lore.kernel.org/linux-perf-users/7703f88e-ccb7-4c98-9da4-8aad224e780f@leemhuis.info/
Fixes: bfb713ea53 ("perf tools: Fix arm64 build by generating unistd_64.h")
Signed-off-by: James Clark <james.clark@linaro.org>
Tested-by: Thorsten Leemhuis <linux@leemhuis.info>
Link: https://lore.kernel.org/r/20250429-james-perf-fix-libperf-in-source-build-v1-1-a1a827ac15e5@linaro.org
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Since pulling in the kernel changes in commit 22f72088ff ("tools
headers: Update the syscall table with the kernel sources"), arm64 is
no longer using a generic syscall header and generates one from the
syscall table. Therefore we must also generate the syscall header for
arm64 before building Perf.
Add it as a dependency to libperf which uses one syscall number. Perf
uses more, but as libperf is a dependency of Perf it will be generated
for both.
Future platforms that need this will have to add their own syscall-y
targets in libperf manually. Unfortunately the arch specific files that
do this (e.g. arch/arm64/include/asm/Kbuild) can't easily be imported
into the Perf build. But Perf only needs a subset of the generated files
anyway, so redefining them is probably the correct thing to do.
Fixes: 22f72088ff ("tools headers: Update the syscall table with the kernel sources")
Signed-off-by: James Clark <james.clark@linaro.org>
Tested-by: Harshit Mogalapalli <harshit.m.mogalapalli@oracle.com>
Link: https://lore.kernel.org/r/20250417-james-perf-fix-gen-syscall-v1-1-1d268c923901@linaro.org
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Pull perf tools updates from Namhyung Kim:
"perf record:
- Introduce latency profiling using scheduler information.
The latency profiling is to show impacts on wall-time rather than
cpu-time. By tracking context switches, it can weight samples and
find which part of the code contributed more to the execution
latency.
The value (period) of the sample is weighted by dividing it by the
number of parallel execution at the moment. The parallelism is
tracked in perf report with sched-switch records. This will reduce
the portion that are run in parallel and in turn increase the
portion of serial executions.
For now, it's limited to profile processes, IOW system-wide
profiling is not supported. You can add --latency option to enable
this.
$ perf record --latency -- make -C tools/perf
I've run the above command for perf build which adds -j option to
make with the number of CPUs in the system internally. Normally
it'd show something like below:
$ perf report -F overhead,comm
...
#
# Overhead Command
# ........ ...............
#
78.97% cc1
6.54% python3
4.21% shellcheck
3.28% ld
1.80% as
1.37% cc1plus
0.80% sh
0.62% clang
0.56% gcc
0.44% perl
0.39% make
...
The cc1 takes around 80% of the overhead as it's the actual
compiler. However it runs in parallel so its contribution to
latency may be less than that. Now, perf report will show both
overhead and latency (if --latency was given at record time) like
below:
$ perf report -s comm
...
#
# Overhead Latency Command
# ........ ........ ...............
#
78.97% 48.66% cc1
6.54% 25.68% python3
4.21% 0.39% shellcheck
3.28% 13.70% ld
1.80% 2.56% as
1.37% 3.08% cc1plus
0.80% 0.98% sh
0.62% 0.61% clang
0.56% 0.33% gcc
0.44% 1.71% perl
0.39% 0.83% make
...
You can see latency of cc1 goes down to around 50% and python3 and
ld contribute a lot more than their overhead. You can use --latency
option in perf report to get the same result but ordered by
latency.
$ perf report --latency -s comm
perf report:
- As a side effect of the latency profiling work, it adds a new
output field 'latency' and a sort key 'parallelism'. The below is a
result from my system with 64 CPUs. The build was well-parallelized
but contained some serial portions.
$ perf report -s parallelism
...
#
# Overhead Latency Parallelism
# ........ ........ ...........
#
16.95% 1.54% 62
13.38% 1.24% 61
12.50% 70.47% 1
11.81% 1.06% 63
7.59% 0.71% 60
4.33% 12.20% 2
3.41% 0.33% 59
2.05% 0.18% 64
1.75% 1.09% 9
1.64% 1.85% 5
...
- Support Feodra mini-debuginfo which is a LZMA compressed symbol
table inside ".gnu_debugdata" ELF section.
perf annotate:
- Add --code-with-type option to enable data-type profiling with the
usual annotate output.
Instead of focusing on data structure, it shows code annotation
together with data type it accesses in case the instruction refers
to a memory location (and it was able to resolve the target data
type). Currently it only works with --stdio.
$ perf annotate --stdio --code-with-type
...
Percent | Source code & Disassembly of vmlinux for cpu/mem-loads,ldlat=30/pp (18 samples, percent: local period)
----------------------------------------------------------------------------------------------------------------------
: 0 0xffffffff81050610 <__fdget>:
0.00 : ffffffff81050610: callq 0xffffffff81c01b80 <__fentry__> # data-type: (stack operation)
0.00 : ffffffff81050615: pushq %rbp # data-type: (stack operation)
0.00 : ffffffff81050616: movq %rsp, %rbp
0.00 : ffffffff81050619: pushq %r15 # data-type: (stack operation)
0.00 : ffffffff8105061b: pushq %r14 # data-type: (stack operation)
0.00 : ffffffff8105061d: pushq %rbx # data-type: (stack operation)
0.00 : ffffffff8105061e: subq $0x10, %rsp
0.00 : ffffffff81050622: movl %edi, %ebx
0.00 : ffffffff81050624: movq %gs:0x7efc4814(%rip), %rax # 0x14e40 <current_task> # data-type: struct task_struct* +0
0.00 : ffffffff8105062c: movq 0x8d0(%rax), %r14 # data-type: struct task_struct +0x8d0 (files)
0.00 : ffffffff81050633: movl (%r14), %eax # data-type: struct files_struct +0 (count.counter)
0.00 : ffffffff81050636: cmpl $0x1, %eax
0.00 : ffffffff81050639: je 0xffffffff810506a9 <__fdget+0x99>
0.00 : ffffffff8105063b: movq 0x20(%r14), %rcx # data-type: struct files_struct +0x20 (fdt)
0.00 : ffffffff8105063f: movl (%rcx), %eax # data-type: struct fdtable +0 (max_fds)
0.00 : ffffffff81050641: cmpl %ebx, %eax
0.00 : ffffffff81050643: jbe 0xffffffff810506ef <__fdget+0xdf>
0.00 : ffffffff81050649: movl %ebx, %r15d
5.56 : ffffffff8105064c: movq 0x8(%rcx), %rdx # data-type: struct fdtable +0x8 (fd)
...
The "# data-type:" part was added with this change. The first few
entries are not very interesting. But later you can it accesses a
couple of fields in the task_struct, files_struct and fdtable.
perf trace:
- Support syscall tracing for different ABI. For example it can trace
system calls for 32-bit applications on 64-bit kernel
transparently.
- Add --summary-mode=total option to show global syscall summary. The
default is 'thread' to show per-thread syscall summary.
Python support:
- Add more interfaces to 'perf' module to parse events, and config,
enable or disable the event list properly so that it can implement
basic functionalities purely in Python. There is an example code
for these new interfaces in python/tracepoint.py.
- Add mypy and pylint support to enable build time checking. Fix some
code based on the findings from these tools.
Internals:
- Introduce io_dir__readdir() API to make directory traveral (usually
for proc or sysfs) efficient with less memory footprint.
JSON vendor events:
- Add events and metrics for ARM Neoverse N3 and V3
- Update events and metrics on various Intel CPUs
- Add/update events for a number of SiFive processors"
* tag 'perf-tools-for-v6.15-2025-03-27' of git://git.kernel.org/pub/scm/linux/kernel/git/perf/perf-tools: (229 commits)
perf bpf-filter: Fix a parsing error with comma
perf report: Fix a memory leak for perf_env on AMD
perf trace: Fix wrong size to bpf_map__update_elem call
perf tools: annotate asm_pure_loop.S
perf python: Fix setup.py mypy errors
perf test: Address attr.py mypy error
perf build: Add pylint build tests
perf build: Add mypy build tests
perf build: Rename TEST_LOGS to SHELL_TEST_LOGS
tools/build: Don't pass test log files to linker
perf bench sched pipe: fix enforced blocking reads in worker_thread
perf tools: Fix is_compat_mode build break in ppc64
perf build: filter all combinations of -flto for libperl
perf vendor events arm64 AmpereOneX: Fix frontend_bound calculation
perf vendor events arm64: AmpereOne/AmpereOneX: Mark LD_RETIRED impacted by errata
perf trace: Fix evlist memory leak
perf trace: Fix BTF memory leak
perf trace: Make syscall table stable
perf syscalltbl: Mask off ABI type for MIPS system calls
perf build: Remove Makefile.syscalls
...
When using EXTRA_CFLAGS, for example "EXTRA_CFLAGS=-DREFCNT_CHECKING=1",
this construct stops setting -g which you'd expect would not be affected
by adding extra flags. Additionally, EXTRA_CFLAGS should be the last
thing to be appended so that it can be used to undo any defaults. And no
condition is required, just += appends to any existing CFLAGS and also
appends or doesn't append EXTRA_CFLAGS if they are or aren't set.
It's not clear why DEBUG=1 is required for -g in Perf when in libperf
it's always on, but I don't think we need to change that behavior now
because someone may be depending on it.
Signed-off-by: James Clark <james.clark@linaro.org>
Reviewed-by: Ian Rogers <irogers@google.com>
Link: https://lore.kernel.org/r/20250319114009.417865-1-james.clark@linaro.org
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Fewer than 32k logical CPUs are currently supported by perf. A cpumap
is indexed by an integer (see perf_cpu_map__cpu) yielding a perf_cpu
that wraps a 4-byte int for the logical CPU - the wrapping is done
deliberately to avoid confusing a logical CPU with an index into a
cpumap. Using a 4-byte int within the perf_cpu is larger than required
so this patch reduces it to the 2-byte int16_t. For a cpumap
containing 16 entries this will reduce the array size from 64 to 32
bytes. For very large servers with lots of logical CPUs the size
savings will be greater.
Signed-off-by: Ian Rogers <irogers@google.com>
Reviewed-by: James Clark <james.clark@linaro.org>
Link: https://lore.kernel.org/r/20250210191231.156294-1-irogers@google.com
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Since the linked fixes: commit, specifying a CPU on hybrid platforms
results in an error because Perf tries to open an extended type event
on "any" CPU which isn't valid. Extended type events can only be opened
on CPUs that match the type.
Before (working):
$ perf record --cpu 1 -- true
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 2.385 MB perf.data (7 samples) ]
After (not working):
$ perf record -C 1 -- true
WARNING: A requested CPU in '1' is not supported by PMU 'cpu_atom' (CPUs 16-27) for event 'cycles:P'
Error:
The sys_perf_event_open() syscall returned with 22 (Invalid argument) for event (cpu_atom/cycles:P/).
/bin/dmesg | grep -i perf may provide additional information.
(Ignore the warning message, that's expected and not particularly
relevant to this issue).
This is because perf_cpu_map__intersect() of the user specified CPU (1)
and one of the PMU's CPUs (16-27) correctly results in an empty (NULL)
CPU map. However for the purposes of opening an event, libperf converts
empty CPU maps into an any CPU (-1) which the kernel rejects.
Fix it by deleting evsels with empty CPU maps in the specific case where
user requested CPU maps are evaluated.
Fixes: 251aa04024 ("perf parse-events: Wildcard most "numeric" events")
Reviewed-by: Ian Rogers <irogers@google.com>
Tested-by: Thomas Falcon <thomas.falcon@intel.com>
Signed-off-by: James Clark <james.clark@linaro.org>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Link: https://lore.kernel.org/r/20241114160450.295844-2-james.clark@linaro.org
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
The perf_cpu_map__merge() function has two arguments, 'orig' and
'other'. The function definition might cause confusion as it could give
the impression that the CPU maps in the two arguments are copied into a
new allocated structure, which is then returned as the result.
The purpose of the function is to merge the CPU map 'other' into the CPU
map 'orig'. This commit changes the 'orig' argument to a pointer to
pointer, so the new result will be updated into 'orig'.
The return value is changed to an int type, as an error number or 0 for
success.
Update callers and tests for the new function definition.
Reviewed-by: Adrian Hunter <adrian.hunter@intel.com>
Signed-off-by: Leo Yan <leo.yan@arm.com>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Ian Rogers <irogers@google.com>
Cc: James Clark <james.clark@linaro.org>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Kan Liang <kan.liang@linux.intel.com>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Link: https://lore.kernel.org/r/20241107125308.41226-2-leo.yan@arm.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Sample period calculation in deliver_sample_value is updated to
calculate the per-thread period delta for events that are inherit +
PERF_SAMPLE_READ. When the sampling event has this configuration, the
read_format.id is used with the tid from the sample to lookup the
storage of the previously accumulated counter total before calculating
the delta. All existing valid configurations where read_format.value
represents some global value continue to use just the read_format.id to
locate the storage of the previously accumulated total.
perf_sample_id is modified to support tracking per-thread
values, along with the existing global per-id values. In the
per-thread case, values are stored in a hash by tid within the
perf_sample_id, and are dynamically allocated as the number is not known
ahead of time.
Signed-off-by: Ben Gainey <ben.gainey@arm.com>
Cc: james.clark@arm.com
Link: https://lore.kernel.org/r/20241001121505.1009685-2-ben.gainey@arm.com
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
install_doc of tools/lib/perf/Makefile invokes install-man,
install-html, and install-examples of
tools/lib/perf/Documentation/Makefile at once. This invocation succeeds
when make runs in serial but can fail when make runs in parallel because
while install-man of tools/lib/perf/Documentation/Makefile depends on
all, install-html depends on nothing and can run ahead of all.
Explicitly specify the dependencies of install-html to ensure that
they are resolved before install-html.
Signed-off-by: Akihiko Odaki <akihiko.odaki@daynix.com>
Link: https://lore.kernel.org/r/20240915-perf-v1-1-cbfd9cd1d482@daynix.com
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Parallel testing appears to show a race between allocating and setting
evsel ids. As there is a bounds check on the xyarray it yields a segv
like:
```
AddressSanitizer:DEADLYSIGNAL
=================================================================
==484408==ERROR: AddressSanitizer: SEGV on unknown address 0x000000000010
==484408==The signal is caused by a WRITE memory access.
==484408==Hint: address points to the zero page.
#0 0x55cef5d4eff4 in perf_evlist__id_hash tools/lib/perf/evlist.c:256
#1 0x55cef5d4f132 in perf_evlist__id_add tools/lib/perf/evlist.c:274
#2 0x55cef5d4f545 in perf_evlist__id_add_fd tools/lib/perf/evlist.c:315
#3 0x55cef5a1923f in store_evsel_ids util/evsel.c:3130
#4 0x55cef5a19400 in evsel__store_ids util/evsel.c:3147
#5 0x55cef5888204 in __run_perf_stat tools/perf/builtin-stat.c:832
#6 0x55cef5888c06 in run_perf_stat tools/perf/builtin-stat.c:960
#7 0x55cef58932db in cmd_stat tools/perf/builtin-stat.c:2878
...
```
Avoid this crash by early exiting the perf_evlist__id_add_fd and
perf_evlist__id_add is the access is out-of-bounds.
Signed-off-by: Ian Rogers <irogers@google.com>
Cc: Yang Jihong <yangjihong1@huawei.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Link: https://lore.kernel.org/r/20240229070757.796244-1-irogers@google.com
