linux-stable.git/tools/perf/tests, branch linux-rolling-stable

perf tools: Fix uninitialized pathname on uncompressed fallback in filename__decompress()

2026-07-24T14:20:11+00:00

[ Upstream commit 51cdb188edeaf389e4377859b9c483c19ce5a259 ]

filename__decompress() has an early return path for files that are not
actually compressed.  This path returns the fd from open() directly but
never writes to the pathname output parameter, leaving the caller with
an uninitialized buffer despite a successful return.

Callers like dso__decompress_kmodule_path() pass pathname to
decompress_kmodule() which uses it to set the decompressed file path.
If pathname is uninitialized, subsequent operations on the path produce
undefined behavior.

Fix by setting pathname to an empty string on the uncompressed path.
Callers already check for an empty pathname to distinguish temporary
decompressed files (which need unlink) from the original file.

Reported-by: sashiko-bot 
Fixes: 7ac22b088afe26a4 ("perf tools: Add filename__decompress function")
Cc: Jiri Olsa 
Assisted-by: Claude:claude-opus-4.6
Signed-off-by: Arnaldo Carvalho de Melo 
Signed-off-by: Sasha Levin

perf event: Fix size of synthesized sample with branch stacks

2026-07-24T14:20:02+00:00

[ Upstream commit 059e9100d82aae2254f1b06835a55755936b1417 ]

Synthesizing branch stacks for Intel-PT highlighted an issue where
PERF_SAMPLE_BRANCH_HW_INDEX was assumed to always be set in the
perf_event_attr branch_sample_type. This caused an incorrect size
calculation.

Fix the writing of the nr and hw_idx values during sample event
synthesis by passing the branch_sample_type into the sample size
and synthesis functions. Also update hardware tracers (Intel PT,
ARM SPE, CS-ETM) to retrieve and pass their branch_sample_type
dynamically to prevent payload misalignment.

Fixes: d3f85437ad6a5511 ("perf evsel: Support PERF_SAMPLE_BRANCH_HW_INDEX")
Assisted-by: Gemini:gemini-3.1-pro-preview
Signed-off-by: Ian Rogers 
Acked-by: Namhyung Kim 
Cc: Adrian Hunter 
Cc: Dapeng Mi 
Cc: Ingo Molnar 
Cc: James Clark 
Cc: Kan Liang 
Cc: Leo Yan 
Cc: Peter Zijlstra 
Cc: Ravi Bangoria 
Cc: Thomas Falcon 
Signed-off-by: Arnaldo Carvalho de Melo 
Signed-off-by: Sasha Levin

perf pmu: Skip test on Arm64 when #slots is zero

2026-07-24T14:20:00+00:00

[ Upstream commit 2e2ba7d1ea554ee6e9e751a53eebf3e9270b0670 ]

Some Arm64 PMUs expose 'caps/slots' as 0 when the slot count is not
implemented, tool_pmu__read_event() currently returns false for this,
so metrics that reference #slots are reported as syntax error.

Since the commit 3a61fd866ef9 ("perf expr: Return -EINVAL for syntax
error in expr__find_ids()"), these syntax errors are populated as
failures and make the PMU metric test fail:

    9.3: Parsing of PMU event table metrics:
    --- start ---
    ...

    Found metric 'backend_bound'
    metric expr 100 * (stall_slot_backend / (#slots * cpu_cycles)) for backend_bound
    parsing metric: 100 * (stall_slot_backend / (#slots * cpu_cycles))
    Failure to read '#slots'
    literal: #slots = nan
    syntax error
    Fail to parse metric or group `backend_bound'

    ...
    ---- end(-1) ----
    9.3: Parsing of PMU event table metrics    : FAILED!

This commit introduces a new function is_expected_broken_metric() to
identify broken metrics, and treats metrics containing "#slots" as
expected broken when #slots == 0 on Arm64 platforms.

Fixes: 3a61fd866ef9aaa1 ("perf expr: Return -EINVAL for syntax error in expr__find_ids()")
Reviewed-by: Ian Rogers 
Reviewed-by: James Clark 
Signed-off-by: Leo Yan 
Cc: Adrian Hunter 
Cc: Alexander Shishkin 
Cc: Jiri Olsa 
Cc: Mark Rutland 
Cc: Namhyung Kim 
Signed-off-by: Arnaldo Carvalho de Melo 
Signed-off-by: Sasha Levin

perf test: Fixes for check branch stack sampling

2026-04-14T06:37:59+00:00

When filtering branch stack samples on user events they sample in user
land but may have come from the kernel. Aarch64 avoids leaking the
kernel address for kaslr reasons but other platforms, for now,
don't. Be more permissive in allowing kernel addresses in the source
of user branch stacks.

When filtering branch stack samples on kernel events they sample in
kernel land but may have come from user land. Avoid the target being a
user address but allow the source to be in user land. Aarch64 may not
leak the user land addresses (making them 0) but other platforms
do. As the kernel address sampling implies privelege, just allow this.

Increase the duration of the system call sampling test to make the
likelihood of sampling a system call higher (increased from 1000 to
8000 loops - a number found through experimentation on an Intel
Tigerlake laptop), also make the period of the event a prime number.

Put unneeded perf record output into a temporary file so that the test
output isn't cluttered. More clearly state which test is running and
the pass, fail or skipped result of the test.

These changes make the test on an Intel tigerlake laptop reliably pass
rather than reliably fail.

Signed-off-by: Ian Rogers 
Reviewed-by: James Clark 
Signed-off-by: Namhyung Kim

perf test: Fix inet_pton probe failure and unroll call graph

2026-04-14T06:28:14+00:00

When adding a probe for libc's inet_pton, perf probe may create multiple
probe points (e.g., due to inlining or multiple symbol resolutions),
resulting in multiple identical event names being output (e.g.,
`probe_libc:inet_pton_1`).

The script previously used a brittle pipeline (`tail -n +2 | head -n -5`)
and an awk script to extract the event name. When multiple probes were
added, awk would output the event name multiple times, which expanded
to multiple words in bash. This broke the subsequent `perf record` and
`perf probe -d` commands, causing the test to fail with:
`Error: another command except --add is set.`

Fix this by removing the brittle `tail/head` commands and appending
`| head -n 1` to the awk extraction. This ensures that only a single,
unique event name is captured, regardless of how many probe points
are created.

Additionally, the test artificially limited the backtrace size via
`max-stack=4` and did not specify dwarf call graphs for non-s390x
architectures. In newer libc versions where `inet_pton` is nested
deeper or compiled without frame pointers, `perf script` failed to resolve
the backtrace up to `/bin/ping`. Fix this by explicitly collecting
dwarf call-graphs for all architectures and increasing `max-stack` to 8.

Assisted-by: Gemini:gemini-3.1-pro-preview
Signed-off-by: Ian Rogers 
Tested-by: Thomas Richter 
Signed-off-by: Namhyung Kim

perf test: Make perf trace BTF general tests exclusive

2026-04-09T04:42:25+00:00

Running both tests cases 126 128 together causes the first test case
126 to fail:
 # for i in $(seq 3); do ./perf test 'perf trace BTF general tests' \
	'perf trace record and replay'; done
 126: perf trace BTF general tests    : FAILED!
 128: perf trace record and replay    : Ok
 126: perf trace BTF general tests    : FAILED!
 128: perf trace record and replay    : Ok
 126: perf trace BTF general tests    : FAILED!
 128: perf trace record and replay    : Ok
 #

Test case 126 fails because test case 128 runs concurrently as can
be observed using a ps -ef | grep perf output list on a different
window. Both do a perf trace command concurrently.
Make test case 'perf trace BTF general tests' exclusive.

Output after:
 # for i in $(seq 3); do ./perf test 'perf trace BTF general tests' \
	'perf trace record and replay'; done
 127: perf trace BTF general tests                   : Ok
 155: perf trace record and replay                   : Ok
 127: perf trace BTF general tests                   : Ok
 155: perf trace record and replay                   : Ok
 127: perf trace BTF general tests                   : Ok
 155: perf trace record and replay                   : Ok
 #

Signed-off-by: Thomas Richter 
Acked-by: Howard Chu 
Signed-off-by: Namhyung Kim

perf data: Clean up use_stdio and structures

2026-04-09T02:21:17+00:00

use_stdio was associated with struct perf_data and not perf_data_file
meaning there was implicit use of fd rather than fptr that may not be
safe. For example, in perf_data_file__write. Reorganize perf_data_file
to better abstract use_stdio, add kernel-doc and more consistently use
perf_data__ accessors so that use_stdio is better respected.

Signed-off-by: Ian Rogers 
Signed-off-by: Namhyung Kim

perf tools: Use calloc() where applicable

2026-04-09T02:21:05+00:00

Instead of using zalloc(nr_entries * sizeof_entry) that is what calloc()
does.

In some places where linux/zalloc.h isn't needed, remove it, add when
needed and was getting it indirectly.

Signed-off-by: Arnaldo Carvalho de Melo 
Signed-off-by: Namhyung Kim

perf tools: Make more global variables static

2026-04-09T02:21:04+00:00

`make check` will run sparse on the perf code base. A frequent warning
is "warning: symbol '...' was not declared. Should it be static?" Go
through and make global definitions without declarations static.

In some cases it is deliberate due to dlsym accessing the symbol, this
change doesn't clean up the missing declarations for perf test suites.

Sometimes things can opportunistically be made const.

Making somethings static exposed unused functions warnings, so
restructuring of ifdefs was necessary for that.

These changes reduce the size of the perf binary by 568 bytes.

Committer notes:

Refreshed the patch, the original one fell thru the cracks, updated the
size reduction.

Remove the trace-event-scripting.c changes, break the build, noticed
with container builds and with sashiko:

  https://sashiko.dev/#/patchset/20260401215306.2152898-1-acme%40kernel.org

Also make two variables static to address another sashiko review
comment:

  https://sashiko.dev/#/patchset/20260402001740.2220481-1-acme%40kernel.org

Signed-off-by: Ian Rogers 
Acked-by: Ankur Arora 
Acked-by: Namhyung Kim 
Cc: Adrian Hunter 
Cc: Albert Ou 
Cc: Alexander Shishkin 
Cc: Alexandre Ghiti 
Cc: Athira Rajeev 
Cc: Guo Ren 
Cc: Howard Chu 
Cc: Ingo Molnar 
Cc: James Clark 
Cc: Jiri Olsa 
Cc: Leo Yan 
Cc: Masami Hiramatsu 
Cc: Palmer Dabbelt 
Cc: Paul Walmsley 
Cc: Peter Zijlstra 
Cc: Yujie Liu 
Signed-off-by: Arnaldo Carvalho de Melo 
Signed-off-by: Namhyung Kim

perf test: Skip sched stats test for !root

2026-04-06T06:27:51+00:00

Running perf sched stats requires root and it fails to open the
schedstat file for regular users.  Let's skip the test.

  $ perf sched stats true
  Failed to open /proc/sys/kernel/sched_schedstats

Reviewed-by: Ian Rogers 
Tested-by: Swapnil Sapkal 
Signed-off-by: Namhyung Kim