Re: [PATCH] perf record: Skip affinity for per-thread mmap buffers

[Namhyung Kim <namhyung@kernel.org>]

On Fri, Nov 06, 2026 at 07:32:29AM +0000, Paul Moses wrote:
> 
> 
> 
> 
> On Friday, November 6th, 2026 at 1:18 AM, Namhyung Kim <namhyung@kernel.org> wrote:
> 
> > Hello,
> > 
> > On Fri, Oct 30, 2026 at 06:59:28AM +0000, Paul wrote:
> > > From: Paul Moses <p@1g4.org>
> > >
> > > Per-thread mmap buffers use CPU -1 to mean any CPU. Combining --per-thread
> > > with --affinity=cpu passes this value to __set_bit() when constructing the
> > > reader affinity mask, causing an out-of-bounds bitmap access and SIGSEGV.
> > >
> > > Program received signal SIGSEGV, Segmentation fault.
> > > 0x0000555555f3488e in mmap.mmap ()
> > > #0  0x0000555555f3488e in mmap.mmap ()
> > > #1  0x0000555555e2f7bf in mmap_per_evsel (evlist=evlist@entry=0x7fffedcc1150, ops=ops@entry=0x7fffffffaf50, idx=idx@entry=0, mp=mp@entry=0x7fffffffaf70, cpu_idx=cpu_idx@entry=0, thread=thread@entry=0, _output=0x7fffffffaf08, _output_overwrite=0x7fffffffaf04, nr_mmaps=0x7fffffffaf0c) at evlist.c:598
> > > #2  0x0000555555e2ef96 in mmap_per_thread (evlist=0x7fffedcc1150, ops=0x7fffffffaf50, mp=0x7fffffffaf70) at evlist.c:650
> > > #3  perf_evlist__mmap_ops (evlist=0x7fffedcc1150, ops=0x7fffffffaf50, mp=0x7fffffffaf70) at evlist.c:748
> > > #4  0x0000555555f2150d in evlist.mmap_ex ()
> > > #5  0x0000555555db0b0a in record.open ()
> > > #6  0x0000555555dadea7 in __cmd_record ()
> > > #7  0x0000555555daca89 in cmd_record ()
> > > #8  0x0000555555e27d53 in run_builtin ()
> > > #9  0x0000555555e277cf in main ()
> > 
> > Please indent terminal outpus otherwise git will eat up lines start with
> > the hash (#).
> > 
> > > rip            0x555555f3488e      0x555555f3488e <mmap.mmap+382>
> > > rax            0x7fffef9e9bd0      140737213537232
> > > rbx            0x7fffedd249a0      140737183369632
> > > rcx            0x3ffffffffffffff   288230376151711743
> > > rdx            0x8000000000000000  -9223372036854775808
> > > rsi            0x0                 0
> > > rdi            0x7fffef9e9bd0      140737213537232
> > >    0x555555f34876 <mmap__mmap+358>:	movswq 0x10(%rbx),%rcx
> > >    0x555555f3487b <mmap__mmap+363>:	mov    0xb8(%rbx),%rax
> > >    0x555555f34882 <mmap__mmap+370>:	mov    $0x1,%edx
> > >    0x555555f34887 <mmap__mmap+375>:	shl    %cl,%rdx
> > >    0x555555f3488a <mmap__mmap+378>:	shr    $0x6,%rcx
> > > => 0x555555f3488e <mmap__mmap+382>:	or     %rdx,(%rax,%rcx,8)
> > >    0x555555f34892 <mmap__mmap+386>:	lea    0x17711b(%rip),%r14        # 0x5555560ab9b4 <verbose>
> > >    0x555555f34899 <mmap__mmap+393>:	cmpl   $0x2,(%r14)
> > >    0x555555f3489d <mmap__mmap+397>:	jne    0x555555f348f7 <mmap__mmap+487>
> > >    0x555555f3489f <mmap__mmap+399>:	mov    0xb8(%rbx),%rdi
> > >    0x555555f348a6 <mmap__mmap+406>:	mov    0xc0(%rbx),%esi
> > >    0x555555f348ac <mmap__mmap+412>:	lea    -0x430(%rbp),%r12
> > >
> > > Only construct and apply reader affinity masks for buffers with a
> > > nonnegative CPU. Skip the reader-side bitmap operations as well, since
> > > per-thread buffers no longer allocate an affinity mask.
> > 
> > Maybe we can just reject those options when they are used together.
> > 
> > Thanks,
> > Namhyung
> > 
> > >
> > > Fixes: 8384a2600c7d ("perf record: Adapt affinity to machines with #CPUs > 1K")
> > > Signed-off-by: Paul Moses <p@1g4.org>
> > > ---
> > >  tools/perf/builtin-record.c | 2 +-
> > >  tools/perf/util/mmap.c      | 4 ++--
> > >  2 files changed, 3 insertions(+), 3 deletions(-)
> > >
> > > diff --git a/tools/perf/builtin-record.c b/tools/perf/builtin-record.c
> > > index a57987851cf0..d7c8f490cb62 100644
> > > --- a/tools/perf/builtin-record.c
> > > +++ b/tools/perf/builtin-record.c
> > > @@ -1565,7 +1565,7 @@ static struct perf_event_header finished_init_event = {
> > >
> > >  static void record__adjust_affinity(struct record *rec, struct mmap *map)
> > >  {
> > > -	if (rec->opts.affinity != PERF_AFFINITY_SYS &&
> > > +	if (map->core.cpu.cpu >= 0 && rec->opts.affinity != PERF_AFFINITY_SYS &&
> > >  	    !bitmap_equal(thread->mask->affinity.bits, map->affinity_mask.bits,
> > >  			  thread->mask->affinity.nbits)) {
> > >  		bitmap_zero(thread->mask->affinity.bits, thread->mask->affinity.nbits);
> > > diff --git a/tools/perf/util/mmap.c b/tools/perf/util/mmap.c
> > > index 358e70c4f3ed..be3dffdaede8 100644
> > > --- a/tools/perf/util/mmap.c
> > > +++ b/tools/perf/util/mmap.c
> > > @@ -292,8 +292,8 @@ int mmap__mmap(struct mmap *map, struct mmap_params *mp, int fd, struct perf_cpu
> > >  		return -1;
> > >  	}
> > >
> > > -	if (mp->affinity != PERF_AFFINITY_SYS &&
> > > -		perf_mmap__setup_affinity_mask(map, mp)) {
> > > +	if (cpu.cpu >= 0 && mp->affinity != PERF_AFFINITY_SYS &&
> > > +	    perf_mmap__setup_affinity_mask(map, mp)) {
> > >  		pr_debug2("failed to alloc mmap affinity mask, error %d\n",
> > >  			  errno);
> > >  		return -1;
> > > --
> > > 2.56.GIT
> > >
> > >
> > 
> 
> 
> Hi Namhyung,
> 
> Can we please get this one settled first. There have been issues
> with important patches being overlooked for months.
> 
> https://lore.kernel.org/linux-perf-users/20261020171946.3593127-1-p@1g4.org/

Sorry about the delay.  But please understand that there can be some
delays during a merge window.  I'll take a look at the link but you
can ping me if it takes too long.

Arnaldo, I think it's simple enough to go to v7.4.

Thanks,
Namhyung