linux-toradex.git/tools/perf/tests, branch master

perf tests mmap-basic: Fix user rdpmc detection logic

2026-10-27T20:40:24+00:00

RISC-V and Arm control userspace counter access through
/proc/sys/kernel/perf_user_access. Add that as a fallback to
set_user_read() so the test can exercise both enabled and disabled
states on those platforms.

Document disabled, enabled and legacy states in user_read_state and fail
on values outside the known 0-2 range. Preserve and restore any value in
this range, rather than just 0 or 1.

Skip when the control files exist but can't be written to
(USER_READ_ERR_PERM). We can't rely on always being able to write to
them when an exclude_kernel=0 event can be opened because opening events
might succeed for non-root users when perf_event_paranoid=-1.

This test isn't run on unsupported platforms since the previous commit,
so we can simplify the following things:

 * Test that cap_user_rdpmc is always equal to the requested state
   rather than checking it only when disabled. As long as we stop
   setting the cap in the attr unconditionally on Arm, this is ok.

 * Remove the USER_READ_UNKNOWN/rdpmc_supported fallback in the checks.
   Now the expected state is always the one requested. USER_READ_UNKNOWN
   only controls whether to skip restoration of the state on unsupported
   platforms, so call it USER_READ_ERR_NO_FILE.

Signed-off-by: James Clark 
Cc: Adrian Hunter 
Cc: Albert Ou 
Cc: Alexander Shishkin 
Cc: Alexandre Ghiti 
Cc: Ian Rogers 
Cc: Ingo Molnar 
Cc: Jiri Olsa 
Cc: Leo Yan 
Cc: Mark Rutland 
Cc: Michael Petlan 
Cc: Namhyung Kim 
Cc: Palmer Dabbelt 
Cc: Paul Walmsley 
Cc: Peter Zijlstra 
Assisted-by: Codex:GPT-6.1-Sol
[Re-write: [Re-write to fix bugs in set_user_read() and simplify tests]
Signed-off-by: Qiao Zhao 
Signed-off-by: Arnaldo Carvalho de Melo

perf tests mmap-basic: Only test userspace reading on supported platforms

2026-10-27T20:03:20+00:00

mmap_user_read_instr() and mmap_user_read_instr_disabled() are the same
test for platforms that libperf doesn't support userspace counter
reading for. Only run the 'enabled' test for platforms that do support
it in order to simplify it in the next commit and not run the same test
twice.

The test can then assert stronger guarantees and have fewer edge cases
to handle as we know it only runs in places where 'rdmpc_supported' will
be true.

"unsupported" hasn't been a relevant skip reason since
commit 588d22b40480 ("perf test: Expand user space event reading (rdpmc)
tests") added the UNKNOWN fallback for unsupported platforms. Remove it
and leave only "permissions".

Signed-off-by: James Clark 
Cc: Adrian Hunter 
Cc: Albert Ou 
Cc: Alexander Shishkin 
Cc: Alexandre Ghiti 
Cc: Ian Rogers 
Cc: Ingo Molnar 
Cc: Jiri Olsa 
Cc: Leo Yan 
Cc: Mark Rutland 
Cc: Michael Petlan 
Cc: Namhyung Kim 
Cc: Palmer Dabbelt 
Cc: Paul Walmsley 
Cc: Peter Zijlstra 
Cc: Qiao Zhao 
Signed-off-by: Arnaldo Carvalho de Melo

perf debug: Avoid calling eprintf() in pr_debugN() when verbose < n

2026-10-27T19:59:18+00:00

pr_debugN() unconditionally evaluated its arguments and called eprintf()
-> veprintf() before checking if verbose >= n. In hot paths such as
dso__load_sym_internal(), two pr_debug4() calls per symbol accounted for
~15% of total CPU time in 'perf top' even at the default verbose == 0.

Guard eprintf() in pr_debugN() with unlikely(verbose >= (n)) and define
pr_debug() in terms of pr_debugN(1, ...).

In test__parsing_callback() (tools/perf/tests/pmu-events.c), move the
test_generic_metric(mexp, 0) call out of the pr_debug() argument list so
the metric evaluation continues to run during 'perf test' when verbose
== 0.

Signed-off-by: Ian Rogers 
Cc: Adrian Hunter 
Cc: Alessio Podda 
Cc: Alireza Haghdoost 
Cc: Ingo Molnar 
Cc: James Clark 
Cc: Jiri Olsa 
Cc: Michael Liang 
Cc: Namhyung Kim 
Cc: Peter Zijlstra 
Cc: Shimin Guo 
Cc: Tanushree Shah 
Signed-off-by: Arnaldo Carvalho de Melo

perf test: Add test for events re-armed with PERF_EVENT_IOC_REFRESH

2026-10-22T20:17:20+00:00

Open a sampling event with sample_period 1, re-arm it with
PERF_EVENT_IOC_REFRESH from its overflow signal handler, as
self-profilers such as async-profiler do, and check that each of 1000
events delivers a signal. One case uses a software event (page faults),
the other the syscalls:sys_enter_getppid tracepoint. The signal is a
real-time one, so overflows cannot be coalesced, and the page fault
buffer is marked MADV_NOHUGEPAGE so that each touched page faults once.

The tracepoint case fails since commit 18dbcbfabfffc4a5 ("perf: Fix the
POLL_HUP delivery breakage") only the first overflow is delivered. The
re-arming ioctl succeeds, but the event count stays at 1. It passes on
the parent commit. The software case passes on both.

The test was written with an AI coding assistant (Claude) from a
standalone reproducer; I reviewed it and ran it in QEMU on current
mainline, on that commit and on its parent.

Reviewed-by: Ian Rogers 
Signed-off-by: Francesco Nigro 
Signed-off-by: Arnaldo Carvalho de Melo

perf test: Avoid stack trace noise in verbose=1 failure snippets

2026-10-19T12:25:42+00:00

When a test fails with a signal or times out, child_test_sig_handler()
appends '---- unexpected signal () ----' followed by a multi-line
stack trace ('    #0 ...' through '_start').

In verbose=1 mode, print_test_failure_snippet() selects up to
failure_snippet_lines (10) lines:
- Pass 1 selects lines matching failure keywords plus their immediate
  next line ('    #0 ... in child_test_sig_handler').
- Pass 2 fills the remaining quota by scanning backwards from the end
  of the output buffer, which consumes the remaining lines on outermost
  stack frames ('handle_internal_command', 'run_argv', 'main',
  '__libc_start_main', '_start') and hides the trailing test output
  showing where the test had run to.

Improve print_test_failure_snippet() by:
- Skipping blank lines and stack trace lines ('    #...') when
  selecting context in Pass 1 and when scanning backwards from the end
  in Pass 2, with a Pass 3 fallback from the top of the buffer if quota
  still remains.
- Prioritizing the non-blank line immediately preceding
  '---- unexpected signal' so there is no '...' gap before the signal
  message.
- Only printing '...' between selected lines when at least one non-blank
  line was omitted between them.

Assisted-by: Antigravity:gemini-3.1-pro
Signed-off-by: Ian Rogers 
Acked-by: Namhyung Kim 
Signed-off-by: Arnaldo Carvalho de Melo

perf test: Take turns using the perf mlock budget in the mmaping tests

2026-10-17T13:30:20+00:00

The mmap size used when none is asked for is derived from
kernel.perf_event_mlock_kb itself (see perf_event_mlock_kb_in_pages()),
that after rounding down to a power of two number of pages ends up being
about all of a single CPU's share of that per user budget, so with 'perf
test' starting all the selected tests at once, any further mapping done
while another test still has its ring buffers mapped fails with:

  Permission error mapping pages.
  Consider increasing /proc/sys/kernel/perf_event_mlock_kb,
  or try again with a smaller value of -m/--mmap_pages.

that then shows up as a failure or, worse, as an "unsupp" skip in tests
like the attr ones, i.e. as a false pass.

Instead of making those tests exclusive, which would serialize them
against every other test, keep them running in parallel with the tests
that don't map ring buffers and have them take turns using the budget,
via a robust, process-shared mutex (see mutex_init_pshared_robust()) in
a MAP_SHARED|MAP_ANONYMOUS area that is inherited by the forked test
processes: a test killed by a signal doesn't leave the budget locked,
the kernel marks the mutex owner-died and the next locker takes it over;
a signal that is caught rather than kills ends up the same way, with the
longjmp out of child_test_sig_handler() leaving it held until the child
exits and the kernel hands it over.

Perf's struct mutex is used, so the lock discipline is checked in clang
builds via -Wthread-safety; as the thread safety analysis can't verify
conditionally held locks, the tests that map ring buffers run through a
helper taking the mutex unconditionally around the test function call.

Robust mutexes aren't always available, at build time or, as glibc and
musl can do when get_robust_list(2)/set_robust_list(2) aren't usable, at
run time: in that case the budget is not shared and the tests run as
before, failing or skipping on starvation instead of hanging on a lock
left behind by a test killed by a signal.

Taking turns is unnecessary when nothing can exhaust the budget: once it is
spent, whatever maps on top of it is charged against RLIMIT_MEMLOCK and the
mapping is refused with EPERM when that runs out too, but only for the
processes the kernel is paranoid about and that have no CAP_IPC_LOCK to
lift the check, see perf_mmap_calc_limits().  Root, then, has no turns to
take: it never has a mapping refused, so the turnstile only buys it the
wall clock cost measured below, for a budget nothing is contending for.

The area is shared by the tests of one 'perf test' invocation only: a
second invocation running at the same time maps its own, so the two
don't take turns with each other.

The sysctl is left alone.  Growing it, to leave room for whatever else
the user has running, would need the original value shared between the
invocations, so that the last one out puts it back, and it only helps a
user who can write the sysctl but lacks CAP_IPC_LOCK, the one capability
that lifts the limit.

Raising the soft RLIMIT_MEMLOCK to the hard one when a mapping is refused
is left for a follow-up: it would buy back the wall clock above without
taking turns, where the hard limit allows the raise.

None of this is done when just one of the selected tests maps ring
buffers, as there is nothing to share the budget with in that case.

Taking turns isn't free: a test that hangs while holding the budget now
holds with it every other test mapping ring buffers, where before it
would fail, or skip, by itself.  And there is a cost even when nothing
goes wrong, this being the whole suite on a Ryzen 9 9950X3D, kernel 7.2.7,
32 CPUs, with the default 516 KiB budget, before and after taking turns:

  budget not shared:   42 s wall clock,  72 tests Ok, 28 failed
  tests take turns:   133 s wall clock, 101 tests Ok, 10 failed

Both are non-root runs of the whole series applied, the shell test markings
included, and none of the ten failures is about the budget: eight of them
pass as root, where it can't be exhausted, and the two that fail there too,
the addr2line inline unwinding one, with no branch stack sampling to ask
this machine for, and the BPF metadata collection one, whose output isn't
what the test expects, are not budget related either.

Reviewed-by: Ian Rogers 
Tested-by: Ian Rogers 
Assisted-by: LLM
Signed-off-by: Arnaldo Carvalho de Melo

perf tests: Mark the shell tests not mapping perf ring buffers

2026-10-17T13:30:20+00:00

Shell tests run perf, that maps ring buffers, so they take a turn using
the per user budget in kernel.perf_event_mlock_kb by default; the ones
that don't map them, like the ones running just 'perf stat', say
" (no_mlock)" in their description line and thus keep running in parallel
with the ones taking turns.

The 'perf sched stats' tests are among them, they synthesize the samples
from /proc/schedstat without mapping a ring buffer, and the tests already
marked " (exclusive)" don't need it, exclusive implies every other
requisite and keeps them running alone.

Reviewed-by: Ian Rogers 
Tested-by: Ian Rogers 
Assisted-by: LLM
Signed-off-by: Arnaldo Carvalho de Melo

perf tests: Mark the C tests mapping perf ring buffers

2026-10-17T13:30:19+00:00

The C tests that run 'perf record' in the test process itself map ring
buffers and so need room in the per user budget set in
kernel.perf_event_mlock_kb: mark them with the new 'needs_mlock' suite
flag, read by the turnstile added later in the series, so that they
take turns using that budget while the tests that don't map them keep
running in parallel.

Reviewed-by: Ian Rogers 
Tested-by: Ian Rogers 
Assisted-by: LLM
Signed-off-by: Arnaldo Carvalho de Melo

perf tests: Introduce a 'needs_mlock' suite flag

2026-10-17T13:30:19+00:00

Each perf instance maps one ring buffer per CPU, all charged to the same
per user budget set in kernel.perf_event_mlock_kb, so suites that map
them need to be told apart from the ones that don't, to later have them
take turns using that budget while the other tests keep running in
parallel.

Add the flag to 'struct test_suite' and a DEFINE_SUITE_NEEDS_MLOCK()
variant for the suites using the DEFINE_SUITE() form.

Shell tests run perf, that maps ring buffers, so they are taken to need
room in that budget by default; the requisites, a comma separated list in
the parens ending their description line, name what a test needs and take
a "no_" prefix to waive it, e.g.:

  # perf stat tests (no_mlock)

"mlock" is room in that budget, so stating it only spells the default
out, "no_mlock" waives it, and "exclusive" is the catch-all that keeps a
test from running alongside any other, so it runs alone, in sequence, and
implies every other requisite, which is why it can't be waived either.
More requisites are to come, each stating what a test needs, till the
catch-all isn't needed anymore.

'perf test requisites' lists what each test needs, reusing the listing
routine and taking the same test name fragments and numbers as 'list'.

Reviewed-by: Ian Rogers 
Tested-by: Ian Rogers 
Assisted-by: LLM
Signed-off-by: Arnaldo Carvalho de Melo

perf test attr: State, spelled out in a line, why a test is skipped

2026-10-17T13:30:19+00:00

Per skip the default output shows a self contained one liner:

  unsupp  'test-record-branch-filter-any' exit 255, expected 1, no attrs recorded to verify
  skipped [kernel newer than 6.12] 'test-record-group-sampling'
  skipped [arch = !aarch64, machine = aarch64] 'test-record-graph-default'

The extra notes explaining the reason, with a copy pasteable command to
observe the failure, come in the very same stream when the verbosity is
bumped, at -vvv, nothing is written to a separate place for later
consultation, it is just the logging ladder:

  detail: '/root/bin/perf record -o /dev/null --no-bpf-event -j any kill'
  exited with 255, the expectations want the workload's 1 exit, so there
  are no attrs recorded to verify; this is normally a capability problem,
  run '/root/bin/perf record -o /dev/null --no-bpf-event -j any kill' to
  see why

Note that 'perf test -v/-vv/-vvv attribute' appends a single -v to the
shell test entry (shell_test__run()), so to see the notes one runs the
python module directly, f.i.:

  python3 tests/shell/lib/attr.py -vvv -d tests/shell/attr -p ./perf

The skip functions now return the reason fragment when they want to skip
and None when the test can go on.

The command suggested to observe the failure writes to /dev/null, so it
doesn't leave a perf.data in the directory it is run from.

Reviewed-by: Ian Rogers 
Tested-by: Ian Rogers 
Assisted-by: LLM
Signed-off-by: Arnaldo Carvalho de Melo