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

perf bench: Add atomic operations benchmark

2026-10-21T22:25:29+00:00

Add a new 'atomic' collection to perf bench for benchmarking
atomic read-modify-write operations with multi-threaded
contention testing. It provides two subcommands:

  - atomic cas: contended compare-and-swap operations
    (__atomic_compare_exchange_n)
  - atomic inc: contended fetch-add operations (__atomic_fetch_add)

Both support a configurable thread count and runtime to measure
atomic contention effects.

Why this benchmark is needed:
- Atomic read-modify-write operations are fundamental to lock-free
  algorithms and data structures. Understanding their performance
  characteristics under contention is critical for designing
  high-performance concurrent applications.
- The benchmark helps identify atomic operation latency and
  scalability issues across different thread counts, revealing
  contention patterns that are not visible in single-threaded tests.
- Useful for evaluating atomic implementation quality on different
  architectures and for regression testing after changes to atomic
  primitives or memory ordering.

Measurement methodology:
- By default two threads contend for the counter, enough to expose
  the contention effect; use -t to scale the count.
- Each thread starts a private timer (clock_gettime CLOCK_MONOTONIC)
  after synchronizing on a pthread_barrier, ensuring all threads
  begin simultaneously.
- Each thread performs a hot loop of the selected atomic operation
  (compare-and-swap or fetch-add) on a shared unsigned long counter
  for a fixed runtime. It counts its own completed operations in a
  local variable, and the loop runs in batches so the termination
  flag is only polled between batches.
- The main thread sleeps for the runtime and then sets the flag; the
  throughput of a repeat is the total operation count divided by the
  measured wall time (the max per-thread runtime).
- The shared counter is cache-line aligned (64 bytes) to isolate
  contention to the target cache line and avoid false sharing.
- The per-thread operation counts of the last repeat show whether
  any thread failed to make progress under contention.
- The first repeat is excluded from statistics as a warmup phase
  to avoid cache-cold effects.

Example usage:
  $ perf bench atomic all
  # Running atomic/cas benchmark...
    Threads: 2, runtime: 2 sec, repeats: 10 (warmup: 1)
    Avg throughput: 68,599,954 ops/sec (+- 0.13%)
    Per-thread ops (last repeat):
      fastest:  74,842,112 ops
      slowest:  63,176,704 ops
      avg:      69,009,408 ops

  # Running atomic/inc benchmark...
    Threads: 2, runtime: 2 sec, repeats: 10 (warmup: 1)
    Avg throughput: 128,428,772 ops/sec (+- 0.54%)
    Per-thread ops (last repeat):
      fastest:  146,230,272 ops
      slowest:  112,825,344 ops
      avg:      129,527,808 ops

Output fields explained:
- Threads: number of contending threads
- runtime: seconds each thread runs per repeat
- repeats: number of test runs (first is warmup)
- Avg throughput: mean aggregate ops/sec across repeats
- +- %: relative standard error of the mean throughput across repeats
- Per-thread ops: fastest/slowest/avg operations per thread in the
  last repeat (shows scheduling imbalance and lack of progress)

Reviewed-by: Namhyung Kim 
Signed-off-by: Changbin Du 
Assisted-by: opencode:DeepSeek-V4-Pro
Signed-off-by: Arnaldo Carvalho de Melo

perf bench numa: Add NULL check after calloc()

2026-09-25T15:02:31+00:00

Two calloc() calls in the numa benchmark lack NULL return checks,
leading to potential NULL pointer dereferences on allocation failure:

1. worker_process(): the allocated `pthreads` array is used in the
   following for-loop via pthreads[t] without checking for NULL.

2. __bench_numa(): the allocated `pids` array is used in the following
   for-loop via pids[i] without checking for NULL.

Add BUG_ON() checks after each calloc(), consistent with the existing
NULL check style used elsewhere in the same file (e.g., node_present
and nodes allocations).

Signed-off-by: longlong yan 
Signed-off-by: Arnaldo Carvalho de Melo

perf bench messaging: Fix the formatting issue when printing.

2026-09-17T15:27:34+00:00

The unsigned int type should use the %u format specifier when
printing with printf or related functions.

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

perf bench futex: Define SYS_futex on 32-bit arches with 64-bit time_t

2026-09-08T20:50:39+00:00

The kernel does not provide sys_futex() on 32-bit architectures that lack
a 32-bit time representation, such as riscv32. Consequently glibc does not
define SYS_futex there, only SYS_futex_time64, and perf bench's futex
benchmarks fail to build:

  bench/futex.h: In function 'futex_syscall':
  bench/futex.h:77:18: error: 'SYS_futex' undeclared (first use in this function)

Define SYS_futex as SYS_futex_time64 when only the latter is available.
The guard and its comment are taken verbatim from
tools/testing/selftests/futex/include/futextest.h, where they were added by
commit 04850819c65c ("selftests/futex: Define SYS_futex on 32-bit
architectures with 64-bit time_t"), keeping the two futex userspace headers
in sync.

No timespec conversion helper is needed on top of this. glibc lays out
struct timespec on 32-bit architectures with 64-bit time_t as
{ int64 tv_sec; int32 tv_nsec; 32-bit pad }, which is compatible with
struct __kernel_timespec, and every futex_wait()/futex_lock_pi()/
futex_wait_requeue_pi() call site in perf bench passes a NULL timeout
anyway, so no timespec ever crosses the syscall boundary here.

Note that an earlier and more ambitious attempt at this was
commit c1ff12dac465 ("perf bench futex: Add support for 32-bit systems
with 64-bit time_t"), reverted by
commit ba4026b09d83 ("Revert "perf bench futex: Add support for 32-bit
systems with 64-bit time_t"") because it included linux/time_types.h,
which is unavailable on older distributions. This change deliberately
avoids that dependency and adds only the one guard needed to fix
compilation.

Reviewed-by: Ian Rogers 
Signed-off-by: Nylon Chen 
Signed-off-by: Arnaldo Carvalho de Melo

Merge tag 'perf-tools-fixes-for-v7.2-1-2026-07-25' into perf-tools-next

2026-07-29T05:11:23+00:00

perf tools fixes for v7.2.

- Update header copies of kernel headers, including const.h, fs.h,
  perf_event.h, gfp_types.h, kvm.h, cpufeatures.h, rtnetlink.hp,
  msr-index.h, drm.h and socket.h.

- Add some build files related to BPF skels to .gitignore.

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

perf bench bpf: Add missing .gitignore file

2026-07-21T23:50:20+00:00

In 713eeb2279402758 ("perf build: Move BPF skeleton generation out of
Makefile.perf") the bpf_skel used with 'perf bench uprobe' was moved
from tools/perf/util/bpf_skel/ to tools/perf/bench/bpf_skel.

Copy tools/perf/util/bpf_skel/.gitignore to that new directory so that
files generated during build get ignored by git.

Reported-by: Linus Torvalds 
Cc: Adrian Hunter 
Cc: Ian Rogers 
Cc: James Clark 
Cc: Jiri Olsa 
Cc: Namhyung Kim 
Fixes: 713eeb2279402758 ("perf build: Move BPF skeleton generation out of Makefile.perf")
Signed-off-by: Arnaldo Carvalho de Melo

perf evlist: Add reference count checking

2026-06-30T17:17:11+00:00

Now the evlist is reference counted, add reference count checking so
that gets and puts are paired and easy to debug. Reference count
checking is documented here:
https://perfwiki.github.io/main/reference-count-checking/

This large patch is adding accessors to evlist functions and switching
to their use. There was some minor renaming as evlist__mmap is now an
accessor to the mmap variable, and the original evlist__mmap is
renamed to evlist__do_mmap.

Signed-off-by: Ian Rogers 
Cc: Adrian Hunter 
Cc: Alice Rogers 
Cc: Dapeng Mi 
Cc: Ingo Molnar 
Cc: James Clark 
Cc: Leo Yan 
Cc: Namhyung Kim 
Cc: Peter Zijlstra 
Cc: Thomas Richter 
Signed-off-by: Arnaldo Carvalho de Melo 
Signed-off-by: Namhyung Kim

perf evlist: Add reference count

2026-06-30T17:17:10+00:00

This a no-op for most of the perf tool. The reference count is set to
1 at allocation, the put will see the 1, decrement it and perform the
delete.

The purpose for adding the reference count is for the python code. Prior
to this change the python code would clone evlists, but this has issues
if events are opened, etc.

This change adds a reference count for the evlists and a later change
will add it to evsels. The combination is needed for the python code to
operate correctly (not hit asserts in the evsel clone), but the changes
are broken apart for the sake of smaller patches.

Assisted-by: Gemini:gemini-3.1-pro-preview
Signed-off-by: Ian Rogers 
Cc: Adrian Hunter 
Cc: Alice Rogers 
Cc: Dapeng Mi 
Cc: Ingo Molnar 
Cc: James Clark 
Cc: Leo Yan 
Cc: Namhyung Kim 
Cc: Peter Zijlstra 
Cc: Thomas Richter 
Signed-off-by: Arnaldo Carvalho de Melo 
Signed-off-by: Namhyung Kim

perf bench: Add --write-size option to sched pipe

2026-06-04T13:11:04+00:00

The default ping-pong uses sizeof(int) (4 bytes) per iteration, which
exercises only the pipe-buffer merge path and keeps allocation entirely
out of the picture. That makes the bench a useful scheduler / context-
switch latency probe but unable to surface anything from the pipe
page-allocation hot path.

Add a -s/--write-size option that sets the bytes written and read per
ping-pong iteration. The buffer is allocated for each side via struct
thread_data and replaces the on-stack int previously used. The default
remains sizeof(int) so existing invocations are unchanged.

With --write-size set above PAGE_SIZE the bench drives anon_pipe_write()
through alloc_page() (or the bulk pre-alloc, if the relevant patch is
applied), which is what we want when measuring pipe locking and page
allocation work.

The bench is a ping-pong: both sides call write() before read(), so a
single write_size payload must fit entirely in the pipe buffer or both
sides deadlock waiting for the other to drain.

Resize the pipe via F_SETPIPE_SZ to match write_size (skipped at the
sizeof(int) default), and error out cleanly when the request exceeds
/proc/sys/fs/pipe-max-size.

Committer testing:

  ⬢ [acme@toolbx perf-tools-next]$ perf bench sched pipe
  # Running 'sched/pipe' benchmark:
  # Executed 1000000 pipe operations between two processes

     Total time: 0.915 [sec]

       0.915493 usecs/op
        1092307 ops/sec
  ⬢ [acme@toolbx perf-tools-next]$ perf bench sched pipe --write-size 1024
  # Running 'sched/pipe' benchmark:
  # Executed 1000000 pipe operations between two processes

     Total time: 0.891 [sec]

       0.891915 usecs/op
        1121183 ops/sec
  ⬢ [acme@toolbx perf-tools-next]$ perf bench sched pipe --write-size 4096
  # Running 'sched/pipe' benchmark:
  # Executed 1000000 pipe operations between two processes

     Total time: 1.366 [sec]

       1.366073 usecs/op
         732025 ops/sec
  ⬢ [acme@toolbx perf-tools-next]$ strace -e fcntl perf bench sched pipe --write-size 4096
  # Running 'sched/pipe' benchmark:
  fcntl(4, F_SETPIPE_SZ, 4096)            = 4096
  fcntl(6, F_SETPIPE_SZ, 4096)            = 4096
  ^Cstrace: Process 17840 detached

  ⬢ [acme@toolbx perf-tools-next]$ strace -e fcntl perf bench sched pipe --write-size 1024
  # Running 'sched/pipe' benchmark:
  fcntl(4, F_SETPIPE_SZ, 1024)            = 4096
  fcntl(6, F_SETPIPE_SZ, 1024)            = 4096
  ^Cstrace: Process 17845 detached

  ⬢ [acme@toolbx perf-tools-next]$ strace -e fcntl perf bench sched pipe
  # Running 'sched/pipe' benchmark:
  ^Cstrace: Process 17851 detached

  ⬢ [acme@toolbx perf-tools-next]$
  ⬢ [acme@toolbx perf-tools-next]$ perf bench sched pipe --write-size 1048577
  # Running 'sched/pipe' benchmark:
  --write-size 1048577 exceeds /proc/sys/fs/pipe-max-size
  ⬢ [acme@toolbx perf-tools-next]$ cat /proc/sys/fs/pipe-max-size
  1048576
  ⬢ [acme@toolbx perf-tools-next]$
  acme@number:~/git/perf-tools-next$

Signed-off-by: Breno Leitao 
Acked-by: Namhyung Kim 
Tested-by: Arnaldo Carvalho de Melo 
Cc: Adrian Hunter 
Cc: Alexander Shishkin 
Cc: Ian Rogers 
Cc: Ingo Molnar 
Cc: James Clark 
Cc: Jiri Olsa 
Cc: Mark Rutland 
Cc: Peter Zijlstra 
Signed-off-by: Arnaldo Carvalho de Melo

perf build: Move BPF skeleton generation out of Makefile.perf

2026-05-20T20:46:45+00:00

Currently, the top-level Makefile.perf defines a massive global bpf-skel
umbrella target that pre-compiles all 12+ BPF skeletons (%.skel.h) upfront
before launching sub-makes. This forces unrelated sub-makes to serialize
behind bpftool and clang BPF target evaluations, causing parallel build
bottlenecks.

Furthermore, bench_uprobe.bpf.c lived inside util/bpf_skel/, breaking
conceptual directory encapsulation since it is consumed purely by
bench/uprobe.c.

Refactor the BPF skeletons to better achieve directory isolation:
1. Move tools/perf/util/bpf_skel/bench_uprobe.bpf.c directly into
   tools/perf/bench/bpf_skel/.
2. Extract the skeleton generation infrastructure out of Makefile.perf into
   a shared inclusion file tools/perf/bpf_skel.mak.
3. Include bpf_skel.mak locally inside tools/perf/util/Build and
   tools/perf/bench/Build and bind precise local prerequisites.
4. Safely synchronize the shared bpftool bootstrap and vmlinux.h targets
   via the conditional prepare: umbrella to avoid parallel sub-make races,
   while evaluating the actual skeletons completely locally on demand. A
   later patch will move these targets into bpf_skel.mak.
5. Export CLANG from the global Makefile to ensure accurate tool
   propagation.
6. Clean up Makefile.perf by stripping the global bpf-skel umbrella target
   and its SKELETONS list.

While removing code from Makefile.perf generally helps build
performance, the impact here is minimal. The main motivation for the
change is to better encapsulate things in the build and simplify
Makefile.perf that has around 50 lines removed.

Reviewed-by: Namhyung Kim 
Assisted-by: Gemini:gemini-3.1-pro-preview
Signed-off-by: Ian Rogers 
Tested-by: James Clark 
Cc: Adrian Hunter 
Cc: Albert Ou 
Cc: Alexandre Chartre 
Cc: Alexandre Ghiti 
Cc: Andrii Nakryiko 
Cc: Ankur Arora 
Cc: Collin Funk 
Cc: Costa Shulyupin 
Cc: Daniel Borkmann 
Cc: Dapeng Mi 
Cc: David Sterba 
Cc: Dmitrii Dolgov <9erthalion6@gmail.com>
Cc: Eduard Zingerman 
Cc: Howard Chu 
Cc: Ingo Molnar 
Cc: Jiri Olsa 
Cc: Kumar Kartikeya Dwivedi 
Cc: Leo Yan 
Cc: Markus Mayer 
Cc: Martin KaFai Lau 
Cc: Nathan Chancellor 
Cc: Nick Terrell 
Cc: Palmer Dabbelt 
Cc: Paul Walmsley 
Cc: Peter Zijlstra 
Cc: Quentin Monnet 
Cc: Ricky Ringler 
Cc: Song Liu 
Cc: Swapnil Sapkal 
Cc: Thomas Falcon 
Cc: Tomas Glozar 
Cc: Yonghong Song 
Signed-off-by: Arnaldo Carvalho de Melo