| Age | Commit message (Collapse) | Author |
|---|
|
Add the evsel from evsel__parse_sample into the struct
perf_sample. Sometimes we want to alter the evsel associated with a
sample, such as with off-cpu bpf-output events. In general the evsel
and perf_sample are passed as a pair, but this makes an altered evsel
something of a chore to keep checking for and setting up. Later
patches will remove passing an evsel with the perf_sample and switch
to just using the perf_sample's value.
Signed-off-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
|
|
The deferred stack trace code wasn't using perf_sample__init/exit. Add
the deferred stack trace clean up to perf_sample__exit which requires
proper NULL initialization in perf_sample__init. Make the
perf_sample__exit robust to being called more than once by using
zfree. Make the error paths in evsel__parse_sample exit the
sample. Add a merged_callchain boolean to capture that callchain is
allocated, deferred_callchain doen't suffice for this. Pack the struct
variables to avoid padding bytes for this.
Similiarly powerpc_vpadtl_sample wasn't using perf_sample__init/exit,
use it for consistency and potential issues with uninitialized
variables.
Similarly guest_session__inject_events in builtin-inject wasn't using
perf_sample_init/exit. The lifetime management for fetched events is
somewhat complex there, but when an event is fetched the sample should
be initialized and needs exiting on error. The sample may be left in
place so that future injects have access to it.
Signed-off-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
|
|
Add kernel-doc for struct perf_sample capturing the somewhat unusual
population of fields and lifetime relationships.
Signed-off-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
|
|
Store cacheline size during perf record in header, so that cacheline
size can be used for other features, like sort keys for perf report.
Testing example with feat enabled:
$ perf record ./Example
$ perf report --header-only | grep -C 3 cacheline
CPU_DOMAIN_INFO info available, use -I to display
e_machine : 62
e_flags : 0
cacheline size: 64
missing features: TRACING_DATA BUILD_ID BRANCH_STACK GROUP_DESC AUXTRACE \
STAT CLOCKID DIR_FORMAT COMPRESSED CLOCK_DATA
========
[namhyung: Update the commit message and remove blank lines]
Signed-off-by: Ricky Ringler <ricky.ringler@proton.me>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
|
|
Writing to the perf.data file can fail in various contexts such as
continual test. Other tests write to a mktemp-ed file, make the "perf
sched stats tests" follow this convention.
Signed-off-by: Ian Rogers <irogers@google.com>
Tested-by: Swapnil Sapkal <swapnil.sapkal@amd.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
|
|
Doing a `perf sched record` then `perf sched stats report` crashes as
the tp_handler isn't set. Add a dummy tp_handler for it rather than
adding an extra check.
Reported-by: Ian Rogers <irogers@google.com>
Reviewed-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
|
|
Per Linus' comments requesting the replacement of "INDIR_BR_LP" in the
indirect branch tracking prctl()s with something more readable, and
suggesting the use of the speculation control prctl()s as an exemplar,
reimplement the prctl()s and related constants that control per-task
forward-edge control flow integrity.
This primarily involves two changes. First, the prctls are
restructured to resemble the style of the speculative execution
workaround control prctls PR_{GET,SET}_SPECULATION_CTRL, to make them
easier to extend in the future. Second, the "indir_br_lp" abbrevation
is expanded to "branch_landing_pads" to be less telegraphic. The
kselftest and documentation is adjusted accordingly.
Link: https://lore.kernel.org/linux-riscv/CAHk-=whhSLGZAx3N5jJpb4GLFDqH_QvS07D+6BnkPWmCEzTAgw@mail.gmail.com/
Cc: Deepak Gupta <debug@rivosinc.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Mark Brown <broonie@kernel.org>
Signed-off-by: Paul Walmsley <pjw@kernel.org>
|
|
The metric code uses the event parsing code but it generally assumes
all events are supported. Arnaldo reported AMD supporting
stalled-cycles-frontend but not stalled-cycles-backend [1]. An issue
with this is that before parsing happens the metric code tries to
share events within groups to reduce the number of events and
multiplexing. If the group has some supported and not supported
events, the whole group will become broken. To avoid this situation
add has_event tests to the metrics for stalled-cycles-frontend and
stalled-cycles-backend. has_events is evaluated when parsing the
metric and its result constant propagated (with if-elses) to reduce
the number of events. This means when the metric code considers
sharing the events, only supported events will be shared.
Note for backporting. This change updates
tools/perf/pmu-events/empty-pmu-events.c a convenience file for builds
on systems without python present. While the metrics.json code should
backport easily there can be conflicts on empty-pmu-events.c. In this
case the build will have left a file test-empty-pmu-events.c that can
be copied over empty-pmu-events.c to resolve issues and make an
appropriate empty-pmu-events.c for the json in the source tree at the
time of the build.
[1] https://lore.kernel.org/lkml/abm1nR-2xjOUBroD@x1/
Reported-by: Arnaldo Carvalho de Melo <acme@kernel.org>
Closes: https://lore.kernel.org/lkml/abm1nR-2xjOUBroD@x1/
Fixes: c7adeb0974f1 ("perf jevents: Add set of common metrics based on default ones")
Signed-off-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
|
|
Add basic kwork coverage tests for record, report, latency, timehist
and top.
Signed-off-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
|
|
Handle the finished_round event. Set up the CTF events when the
feature event desc is read. In pipe mode the attr events will create
the evsels and the feature event desc events will name the evsels. The
CTF events need the evsel name, so wait until feature event descs are
read (in pipe mode) before setting up the events except for tracepoint
events. Handle the tracing_data event so that tracepoint information
is available when setting up tracepoint events.
Signed-off-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
|
|
In situations like the perf data converter the evsel__name will be
used to create babeltrace events. If the events have the same name
then creation can fail. Avoid these failures by including more
information into the unknown event names.
Signed-off-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
|
|
Some event processing functions like perf_event__process_tracing_data
return a zero or positive value on success. Ordered event processing
handles any non-zero value as an error, which is inconsistent with
reader__process_events and reader__read_event that only treat negative
values as errors. Make the ordered events error handling consistent
with that of the events reader.
Signed-off-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
|
|
In non-pipe/data mode the header has a 256-bit bitmap representing
whether a feature is enabled or not. In pipe mode features are written
out in perf_event__synthesize_features as PERF_RECORD_HEADER_FEATURE
events with a special zero sized marker for the last feature. If a new
feature is added the last feature marker event appears as that feature
from old pipe mode perf data. As the event is zero sized it will fail
to be processed and generally terminate perf.
Add a last_feat variable to the header that in non-pipe/data mode is
just HEADER_LAST_FEATURE. In pipe mode compute the last_feat by
handling zero sized feature events, assuming they are the marker and
updating last_feat accordingly. Potentially a feature event could be
zero sized and so still process the feature event, just ignore the
error if it fails.
As perf_event__process_feature can properly handle pipe mode data,
migrate users to it except for report that still wants to group events
and stop header printing with the last feature marker. Make
perf_event__process_feature non-fatal in the case of a newer feature
than this version of perf's HEADER_LAST_FEATURE, which was the
behavior all users wanted.
Signed-off-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
|
|
Print log information in ordered event processing so that the cause of
finished round failing is clearer. Print the event name along with its
number when an event isn't processed. Add extra detail about where the
failure happened.
The following log lines come from running `perf data convert`. Before:
0xa250 [0x10]: failed to process type: 80
After:
0xa250 [0x10]: piped event processing failed for event of type: FEATURE (80)
Signed-off-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
|
|
By removing the features from feat_ops with ifdefs the previous logic
would print "# (null)" when perf processed a feature that lacked
builtin support. Remove the ifdefs from feat_ops and in the relevant
functions print errors/messages about the lack of support.
Signed-off-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
|
|
For logging and debug messages it can be convenient to convert a
feature number to a name. Add header_feat__name for this and reuse the
data already within the feat_ops struct.
Signed-off-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
|
|
clockid_t is declared in time.h but the include is missing. Reordering
header files may result in build breakages. Add the include to avoid
this.
Signed-off-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
|
|
Although it happens very rarely, in case of out-of-order events (i.e.
due to CPU migration when a syscall is executed), the calculation of
event duration might underflow and thus a bogus value is printed:
2.804 ( 0.001 ms): :49553/49553 rt_sigaction(sig: QUIT, act: 0x7fff403ed6e0, oact: 0x7fff403ed780, sigsetsize: 8) = 0
2.807 ( 0.001 ms): :49553/49553 rt_sigaction(sig: CHLD, act: 0x7fff403ed6e0, oact: 0x7fff403ed780, sigsetsize: 8) = 0
2.815 (18446744073709.438 ms): :49553/49553 execve(filename: 0xbb173a30, argv: 0x55aabb171930, envp: 0x55aabb171120) = 0
2.815 ( 0.534 ms): pwd/49553 ... [continued]: execve()) = 0
Check for possible underflow first and in case of a bogus value, do
not print it.
2.804 ( 0.001 ms): :49553/49553 rt_sigaction(sig: QUIT, act: 0x7fff403ed6e0, oact: 0x7fff403ed780, sigsetsize: 8) = 0
2.807 ( 0.001 ms): :49553/49553 rt_sigaction(sig: CHLD, act: 0x7fff403ed6e0, oact: 0x7fff403ed780, sigsetsize: 8) = 0
2.815 ( ): :49553/49553 execve(filename: 0xbb173a30, argv: 0x55aabb171930, envp: 0x55aabb171120) = 0
2.815 ( 0.534 ms): pwd/49553 ... [continued]: execve()) = 0
Signed-off-by: Michael Petlan <mpetlan@redhat.com>
Reviewed-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
|
|
The build_id parsing functions calculate a filename length from the
event header size and read directly into a stack buffer of PATH_MAX
bytes without bounds checking. A malformed perf.data file with a
crafted header.size can cause the length to be negative or exceed
PATH_MAX, resulting in a stack buffer overflow.
Add bounds checking for the filename length in both
perf_header__read_build_ids() and the ABI quirk variant. Print a
warning message when invalid length is detected.
Signed-off-by: SeungJu Cheon <suunj1331@gmail.com>
Reviewed-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
|
|
Return -ENOENT when no metric/group matches, and directly use the return
value from expr__find_ids(), so -EINVAL is reserved for parse failures.
Print separate logs to make it clear.
Before:
perf stat -C 5 -vvv
Using CPUID 0x00000000410fd490
metric expr 100 * (STALL_SLOT_BACKEND / (CPU_CYCLES * #slots) - BR_MIS_PRED * 3 / CPU_CYCLES) for backend_bound
parsing metric: 100 * (STALL_SLOT_BACKEND / (CPU_CYCLES * #slots) - BR_MIS_PRED * 3 / CPU_CYCLES)
Failure to read '#slots'
literal: #slots = nan
syntax error
Cannot find metric or group `Default'
After:
perf stat -C 5 -vvv
Using CPUID 0x00000000410fd490
metric expr 100 * (STALL_SLOT_BACKEND / (CPU_CYCLES * #slots) - BR_MIS_PRED * 3 / CPU_CYCLES) for backend_bound
parsing metric: 100 * (STALL_SLOT_BACKEND / (CPU_CYCLES * #slots) - BR_MIS_PRED * 3 / CPU_CYCLES)
Failure to read '#slots'
literal: #slots = nan
syntax error
Fail to parse metric or group `Default'
Signed-off-by: Leo Yan <leo.yan@arm.com>
Reviewed-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
|
|
Add a trailing newline for logs.
Before:
perf stat -C 5
Failure to read '#slots'Cannot find metric or group `Default'
After:
perf stat -C 5
Failure to read '#slots'
Cannot find metric or group `Default'
Signed-off-by: Leo Yan <leo.yan@arm.com>
Reviewed-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
|
|
expr__find_ids() propagates the parser return value directly. For syntax
errors, the parser can return a positive value, but callers treat it as
success, e.g., for below case on Arm64 platform:
metric expr 100 * (STALL_SLOT_BACKEND / (CPU_CYCLES * #slots) - BR_MIS_PRED * 3 / CPU_CYCLES) for backend_bound
parsing metric: 100 * (STALL_SLOT_BACKEND / (CPU_CYCLES * #slots) - BR_MIS_PRED * 3 / CPU_CYCLES)
Failure to read '#slots' literal: #slots = nan
syntax error
Convert positive parser returns in expr__find_ids() to -EINVAL, as a
result, the error value will be respected by callers.
Before:
perf stat -C 5
Failure to read '#slots'Failure to read '#slots'Failure to read '#slots'Failure to read '#slots'Segmentation fault
After:
perf stat -C 5
Failure to read '#slots'Cannot find metric or group `Default'
Fixes: ded80bda8bc9 ("perf expr: Migrate expr ids table to a hashmap")
Signed-off-by: Leo Yan <leo.yan@arm.com>
Reviewed-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
|
|
Test case 'perf data type profiling tests' fails on s390 with this
error:
# ./perf mem record -- ./perf test -w code_with_type
failed: no PMU supports the memory events
# echo $?
255
#
because s390 does not support memory events at all. According to the
man page, perf annotate --code-with-type only works with memory
instructions only. As command 'perf mem record ...' is not supported
on s390, skip this test for s390.
Output before:
# ./perf test 'perf data type profiling tests'
77: perf data type profiling tests : FAILED!
Output after:
# ./perf test 'perf data type profiling tests'
77: perf data type profiling tests : Skip
Fixes: f60a5c22967b8 ("perf tests: Test annotate with data type profiling and rust")
Signed-off-by: Thomas Richter <tmricht@linux.ibm.com>
Reviewed-by: Ian Rogers <irogers@google.com>
Cc: Dmitrii Dolgov <9erthalion6@gmail.com>
Suggested-by: Namhyung Kim <namhyung@kernel.org>
Suggested-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
|
|
When sorting the dso array we sometimes get a crash due to null
comparisons in comparator functions. So prevent __dsos__add from
adding null to the dso array to avoid out-of-memory related errors.
Signed-off-by: Anubhav Shelat <ashelat@redhat.com>
Reviewed-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
|
|
test__ratio_to_prev() assumed the first event in a group is the leader,
which is not the case when the event is expanded into two event groups
on hybrid PMU's with auto counter reload support. Instead, iterate over the
event group generated for each core PMU. Also update "wrong leader" test to
check that the subordinate event has the correct leader instead of checking
that it is not the group leader. Finally, do not exit immediately if a PMU
without auto counter reload support is found.
Signed-off-by: Thomas Falcon <thomas.falcon@intel.com>
Reviewed-by: Dapeng Mi <dapeng1.mi@linux.intel.com>
Reviewed-by: Ian Rogers <irogers@google.com>
Fixes: 56be0fe5f62c ("perf record: Add auto counter reload parse and regression tests")
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
|
|
When perf resolves symbols from kernel module ELF files (ET_REL),
it converts symbol addresses to file offsets so that sample IPs
can be matched to the correct symbol. The conversion adjusts each
symbol's st_value:
sym->st_value -= shdr->sh_addr - shdr->sh_offset;
For vmlinux (ET_EXEC), st_value is a virtual address and sh_addr
is the section's virtual base, so subtracting sh_addr and adding
sh_offset correctly yields a file offset.
For kernel modules (ET_REL), st_value is a section-relative
offset. The module loader ignores sh_addr entirely and places
symbols at module_base + st_value. Converting to file offset
requires only adding sh_offset; subtracting sh_addr introduces an
error equal to sh_addr bytes.
When .text has sh_addr == 0 -- the historical norm for simple
modules -- both formulas produce the same result and the bug is
latent. As modules gain more metadata sections before .text (.note,
.static_call.text, etc.), the linker assigns .text a non-zero
sh_addr, exposing the defect. For example, nfsd.ko on this kernel
has sh_addr=0xa80, kvm-intel.ko has sh_addr=0x1e90.
The effect is that all .text symbols in affected modules
shift by sh_addr bytes relative to sample IPs, causing perf
report to attribute samples to incorrect, nearby symbols. This
was observed as 13% of LLC-load-miss samples misattributed
to nfsd_file_get_dio_attrs when the actual hot function was
nfsd_cache_lookup, approximately 0xa80 bytes away in the symbol
table.
Use the existing dso__rel() flag (already set for ET_REL modules)
to select the correct adjustment: add sh_offset for ET_REL,
subtract (sh_addr - sh_offset) for ET_EXEC/ET_DYN.
Fixes: 0131c4ec794a ("perf tools: Make it possible to read object code from kernel modules")
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Reviewed-by: Ian Rogers <irogers@google.com>
Tested-by: Thomas Richter <tmricht@linux.ibm.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
|
|
It needs to synthesize task info for the comm name. The mmap
information is only needed for callchain symbolization which is not used
by the summary mode. Also total or cgroup summary mode don't require
the task info. Let's skip the processing if possible.
Reviewed-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
|
|
Perf stat is crashing on arm64 hosts with the following issue:
# make -C tools/perf DEBUG=1
# perf stat sleep 1
perf: util/evsel.c:2034: get_group_fd: Assertion `!(!leader->core.fd)' failed.
[1] 1220794 IOT instruction (core dumped) ./perf stat
The sorting function introduced by commit a745c0831c15c ("perf stat:
Sort default events/metrics") compares events based on their individual
properties. This can cause events from different groups to be
interleaved, resulting in group members appearing before their leaders
in the sorted evlist.
When the iterator opens events in list order, a group member may be
processed before its leader has been opened.
For example, CPU_CYCLES (idx=32) with leader STALL_SLOT_BACKEND (idx=37)
could be sorted before its leader, causing the crash when CPU_CYCLES
tries to get its group fd from the not-yet-opened leader.
Fix this by comparing events based on their leader's attributes instead
of their own attributes when the events are in different groups. This
ensures all members of a group share the same sort key as their leader,
keeping groups together and guaranteeing leaders are opened before their
members.
Fixes: a745c0831c15c ("perf stat: Sort default events/metrics")
Reported-by: Denis Yaroshevskiy <dyaroshev@meta.com>
Tested-by: Dmitry Ilvokhin <d@ilvokhin.com>
Tested-by: Ian Rogers <irogers@google.com>
Signed-off-by: Breno Leitao <leitao@debian.org>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
|
|
The test constantly fails on my Intel hybrid machine. The issue was it
has two events in the output even if I only gave it one event.
$ perf stat -e instructions -- perf test -w sqrtloop
Performance counter stats for 'perf test -w sqrtloop':
910,856,421 cpu_atom/instructions/ (28.05%)
14,852,865,997 cpu_core/instructions/ (96.79%)
1.014313341 seconds time elapsed
1.004114000 seconds user
0.008174000 seconds sys
Let's modify the awk script to add the values for each line and print
the total. The variable 'i' has a number of input lines that have valid
output and variable 'c' has the sum of actual counter values. That way
it should work on any platforms.
Reviewed-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
|
|
Writing to the test output files in the current working directory can
fail in various contexts such as continual test. Other tests write to
a mktemp-ed file, make the "perf script task-analyszer tests" follow
this convention too. Currently this isn't possible for the perf.data
file due to a lack of perf script support, add a variable for when
this support is available.
Signed-off-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
|
|
The index into the cpumap array and the number of entries within the
array can never be negative, so let's make them unsigned. This is
prompted by reports that gcc 13 with -O6 is giving a
alloc-size-larger-than errors. The change makes the cpumap changes and
then updates the declaration of index variables throughout perf and
libperf to be unsigned. The two things are hard to separate as
compiler warnings about mixing signed and unsigned types breaks the
build.
Reported-by: Chingbin Li <liqb365@163.com>
Closes: https://lore.kernel.org/lkml/20260212025127.841090-1-liqb365@163.com/
Tested-by: Chingbin Li <liqb365@163.com>
Signed-off-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
|
|
tools/perf/trace/beauty/include/
As it is not really used when compiling anything, just being parsed to
collect number->string tables for 'perf trace'.
$ git grep fadvise.h tools/
tools/perf/Makefile.perf:$(fadvise_advice_array): $(beauty_uapi_linux_dir)/fadvise.h $(fadvise_advice_tbl)
tools/perf/check-headers.sh: "include/uapi/linux/fadvise.h"
tools/perf/trace/beauty/fadvise.sh:grep -E $regex ${header_dir}/fadvise.h | \
tools/perf/trace/beauty/fadvise.sh:# tools/include/uapi/linux/fadvise.h for details.
$
Link: https://lore.kernel.org/r/CAP-5=fVBNQVF8k3JUQjH1nkP69ZVp8BqP+uwygcx=xO0zC4xrg@mail.gmail.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
|
|
As it is used only to parse ioctl numbers, not to build perf and so far
no other tools/ living tool uses it, so to clean up tools/include/ to be
used just for building tools, to have access to things available in the
kernel and not yet in the system headers, move it to the directory where
just the tools/perf/trace/beauty/ scripts can use to generate tables
used by perf.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
|
|
Commit 3bc753c06dd0 ("kbuild: treat char as always unsigned") made
chars unsigned by default in the Linux kernel. To avoid similar kinds
of bugs and warnings, make unsigned chars the default for the perf tool.
Signed-off-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
|
|
This patch adds a new --pmu-filter option to perf-stat command to allow
filtering events on specific PMUs. This is useful when there are
multiple PMUs with same type (e.g. hisi_sicl2_cpa0 and hisi_sicl0_cpa0).
[root@localhost tmp]# perf stat -M cpa_p0_avg_bw
Performance counter stats for 'system wide':
19,417,779,115 hisi_sicl0_cpa0/cpa_cycles/ # 0.00 cpa_p0_avg_bw
0 hisi_sicl0_cpa0/cpa_p0_wr_dat/
0 hisi_sicl0_cpa0/cpa_p0_rd_dat_64b/
0 hisi_sicl0_cpa0/cpa_p0_rd_dat_32b/
19,417,751,103 hisi_sicl10_cpa0/cpa_cycles/ # 0.00 cpa_p0_avg_bw
0 hisi_sicl10_cpa0/cpa_p0_wr_dat/
0 hisi_sicl10_cpa0/cpa_p0_rd_dat_64b/
0 hisi_sicl10_cpa0/cpa_p0_rd_dat_32b/
19,417,730,679 hisi_sicl2_cpa0/cpa_cycles/ # 0.31 cpa_p0_avg_bw
75,635,749 hisi_sicl2_cpa0/cpa_p0_wr_dat/
18,520,640 hisi_sicl2_cpa0/cpa_p0_rd_dat_64b/
0 hisi_sicl2_cpa0/cpa_p0_rd_dat_32b/
19,417,674,227 hisi_sicl8_cpa0/cpa_cycles/ # 0.00 cpa_p0_avg_bw
0 hisi_sicl8_cpa0/cpa_p0_wr_dat/
0 hisi_sicl8_cpa0/cpa_p0_rd_dat_64b/
0 hisi_sicl8_cpa0/cpa_p0_rd_dat_32b/
19.417734480 seconds time elapsed
[root@localhost tmp]# perf stat --pmu-filter hisi_sicl2_cpa0 -M cpa_p0_avg_bw
Performance counter stats for 'system wide':
6,234,093,559 cpa_cycles # 0.60 cpa_p0_avg_bw
50,548,465 cpa_p0_wr_dat
7,552,182 cpa_p0_rd_dat_64b
0 cpa_p0_rd_dat_32b
6.234139320 seconds time elapsed
Signed-off-by: Qinxin Xia <xiaqinxin@huawei.com>
Reviewed-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
|
|
The "comm" column allows grouping events by the process command. It is
intended to group like programs, despite having different PIDs. But some
workloads may adjust their own command, so that a unique identifier
(e.g. a PID or some other numeric value) is part of the command name.
This destroys the utility of "comm", forcing perf to place each unique
process name into its own bucket, which can contribute to a
combinatorial explosion of memory use in perf report.
Create a less strict version of this column, which ignores digits when
comparing command names. Commands whose names are the same (ignoring
digits) are sorted into the same histogram buckets, and displayed with
the placeholder value "<N>" in the place of digits. For example,
hypothetical command names "kworker/1" "kworker/2" "kworker/3" would
sort into the same bucket and be represented as "kworker/<N>".
Committer testing:
$ perf report -s comm,comm_nodigit | grep -F "<N>"
0.01% CPU 6/TCG CPU <N>/TCG
0.01% kworker/53:2-mm kworker/<N>:<N>-mm
0.01% migration/24 migration/<N>
0.01% kworker/24:1-ev kworker/<N>:<N>-ev
0.01% llvmpipe-8 llvmpipe-<N>
Signed-off-by: Stephen Brennan <stephen.s.brennan@oracle.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
|
|
Commit f5803651b4a4 ("perf stat: Choose the most disaggregate command
line option") changed aggregation option handling for `perf stat` but
not `perf stat report` leading to parse_cache_level being passed a
struct in the `perf stat` case but erroneously an aggr_mode enum value
for `perf stat report`. Change the `perf stat report` aggregation
handling to use the same opt_aggr_mode as `perf stat`. Also, just pass
the boolean for consistency with other boolean argument handling.
Fixes: f5803651b4a4 ("perf stat: Choose the most disaggregate command line option")
Signed-off-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
|
|
The value is a void* and the address of an int, max_stack_depth, is
set up in the perf lock options. The parse_max_stack function treats
the int* as a long*, make this more correct by declaring the value to
be an int*.
Fixes: 0a277b622670 ("perf lock contention: Check --max-stack option")
Signed-off-by: Ian Rogers <irogers@google.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
|
|
commit e5e66adfe45a6 ("perf regs: Remove __weak attributive arch_sdt_arg_parse_op() function")
removes arch_sdt_arg_parse_op() functions and reveals missing s390 support.
The following warning is printed:
Unknown ELF machine 22, standard arguments parse will be skipped.
ELF machine 22 is the EM_S390 host. This happens with command
# ./perf record -v -- stress-ng -t 1s --matrix 0
when the event is not specified.
Add s390 specific __perf_sdt_arg_parse_op_s390() function to support
-architecture calls to arch_sdt_arg_parse_op() for s390.
The warning disappears.
Signed-off-by: Thomas Richter <tmricht@linux.ibm.com>
Reviewed-by: Ian Rogers <irogers@google.com>
Tested-by: Jan Polensky <japo@linux.ibm.com>
Cc: Dapeng Mi <dapeng1.mi@linux.intel.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
|
|
To get the various fixes for v7.0.
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
|
|
Frame pointer callchains are not supported on s390 and dwarf
callchains are only supported on software events.
Switch the default event from the hardware 'cycles' event to the
software 'cpu-clock' or 'task-clock' on s390 if callchains are
enabled. Move some of the target initialization earlier in builtin-top
and builtin-record, so it is ready for use by evlist__new_default.
If frame pointer callchains are requested on s390 show a
warning. Modify the '-g' option of `perf top` and `perf record` to
default to dwarf callchains on s390.
Signed-off-by: Ian Rogers <irogers@google.com>
Tested-by: Thomas Richter <tmricht@linux.ibm.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
|
|
record_opts__parse_callchain is shared by builtin-record and
builtin-trace, it is declared in callchain.h. Move the declaration to
callchain.c for consistency with the header. In other cases make the
option callback a small static stub that then calls into callchain.c.
Make the no argument '-g' callchain option just a short-cut for
'--call-graph fp' so that there is consistency in how the arguments
are handled. This requires the const char* string to be strdup-ed in
__parse_callchain_report_opt. For consistency also make
parse_callchain_record use strdup and remove some unnecessary
casts. Also, be more explicit about the '-g' behavior if there is a
.perfconfig file setting.
Signed-off-by: Ian Rogers <irogers@google.com>
Tested-by: Thomas Richter <tmricht@linux.ibm.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
|
|
The options are used to configure the evsel but are not themselves
configured. Make the arguments const to better capture this.
Signed-off-by: Ian Rogers <irogers@google.com>
Tested-by: Thomas Richter <tmricht@linux.ibm.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
|
|
Allow the target to be const in callers.
Signed-off-by: Ian Rogers <irogers@google.com>
Tested-by: Thomas Richter <tmricht@linux.ibm.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
|
|
Switch to using evsel__match rather than comparing perf_event_attr
values, this is robust on hybrid architectures.
Ensure evsel->pmu matches the evsel->core.attr.
Remove exclude bits that get set in other fallback attempts when
switching the event.
Log the event name with modifiers when switching the event on fallback.
Signed-off-by: Ian Rogers <irogers@google.com>
Tested-by: Thomas Richter <tmricht@linux.ibm.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
|
|
Previously, only the first DWARF location entry was collected for each
variable. This was based on the assumption that instruction tracking
could reconstruct the remaining state. However, variables may have
different locations across different address ranges, and relying solely
on instruction tracking can miss valid type information.
Change __die_collect_vars_cb() to iterate over all location entries
using dwarf_getlocations() in a loop. This ensures that variables with
multiple location ranges are properly tracked, improving type coverage.
Signed-off-by: Zecheng Li <zli94@ncsu.edu>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
|
|
When a function call occurs, caller-saved registers are typically
invalidated since the callee may clobber them. However, DWARF debug info
provides location ranges that indicate exactly where a variable is valid
in a register.
Track the DWARF location range end address in type_state_reg and use it
to determine if a caller-saved register should be preserved across a
call. If the current call address is within the DWARF-specified lifetime
of the variable, keep the register state valid instead of invalidating
it.
This improves type annotation for code where the compiler knows a
register value survives across calls (e.g., when the callee is known not
to clobber certain registers or when the value is reloaded after the
call at the same logical location).
Changes:
- Add `end` and `has_range` fields to die_var_type to capture DWARF
location range information
- Add `lifetime_active` and `lifetime_end` fields to type_state_reg
- Check location lifetime before invalidating caller-saved registers
Signed-off-by: Zecheng Li <zli94@ncsu.edu>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
|
|
Previously, the x86 call handler returned early without invalidating
caller-saved registers when the call target symbol could not be resolved
(func == NULL). This violated the ABI which requires caller-saved
registers to be considered clobbered after any call instruction.
Fix this by:
1. Always invalidating caller-saved registers for any call instruction
(except __fentry__ which preserves registers)
2. Using dl->ops.target.name as fallback when func->name is unavailable,
allowing return type lookup for more call targets
This is a conservative change that may reduce type coverage for indirect
calls (e.g., callq *(%rax)) where we cannot determine the return type
but it ensures correctness.
Signed-off-by: Zecheng Li <zli94@ncsu.edu>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
|
|
Add a helper function to consistently invalidate register state instead
of field assignments. This ensures kind, ok, and copied_from are all
properly cleared when a register becomes invalid.
The helper sets:
- kind = TSR_KIND_INVALID
- ok = false
- copied_from = -1
Replace all invalidation patterns with calls to this helper. No
functional change and this removes some incorrect annotations that were
caused by incomplete invalidation (e.g. a obsolete copied_from from an
invalidated register).
Signed-off-by: Zecheng Li <zli94@ncsu.edu>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
|
|
When a register holds a constant value (TSR_KIND_CONST) and is used with
a negative offset, treat it as a potential global variable access
instead of falling through to CFA (frame) handling.
This fixes cases like array indexing with computed offsets:
movzbl -0x7d72725a(%rax), %eax # array[%rax]
Where %rax contains a computed index and the negative offset points to a
global array. Previously this fell through to the CFA path which doesn't
handle global variables, resulting in "no type information".
The fix redirects such accesses to check_kernel which calls
get_global_var_type() to resolve the type from the global variable
cache. This is only done for kernel DSOs since the pattern relies on
kernel-specific global variable resolution. We could also treat
registers with integer types to the global variable path, but this
requires more changes.
Signed-off-by: Zecheng Li <zli94@ncsu.edu>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
|
