mirror of https://lore.kernel.org/lkml/
 help / color / mirror / Atom feed
From: Aditya Chillara <aditya.chillara@oss.qualcomm.com>
To: Daniel Wade <danjwade95@gmail.com>,
	Peter Zijlstra <peterz@infradead.org>,
	Ingo Molnar <mingo@redhat.com>,
	Arnaldo Carvalho de Melo <acme@kernel.org>,
	Namhyung Kim <namhyung@kernel.org>
Cc: Mark Rutland <mark.rutland@arm.com>,
	Alexander Shishkin <alexander.shishkin@linux.intel.com>,
	Jiri Olsa <jolsa@kernel.org>, Ian Rogers <irogers@google.com>,
	Adrian Hunter <adrian.hunter@intel.com>,
	James Clark <james.clark@linaro.org>,
	Kyle Zeng <kylebot@openai.com>,
	linux-perf-users@vger.kernel.org, linux-kernel@vger.kernel.org
Subject: Re: [PATCH v2] perf: Reject group leaders detached by CPU hot-unplug
Date: Thu, 19 Nov 2026 14:24:19 +0530	[thread overview]
Message-ID: <06758809-6748-4f7b-b7e6-c862936a961c@oss.qualcomm.com> (raw)
In-Reply-To: <20261119014730.2074-1-danjwade95@gmail.com>

On 11/19/2026 7:17 AM, Daniel Wade wrote:
> CPU hot-unplug detaches the events of a CPU context but leaves them in
> OFF or ERROR, so once the CPU is back a detached leader is still accepted
> as group_fd. perf_group_detach() does nothing for a leader that is
> already detached, so closing it frees it with the new sibling still
> linked, and closing the sibling is a use-after-free.
> 
> Detach them with DETACH_EXIT as well, so that they end up in
> PERF_EVENT_STATE_EXIT and are rejected by the check from
> commit fa091f46c383 ("perf: Reject exited events as group leaders").
> 
> Side effect: poll() on a hot-unplugged event now reports EPOLLHUP, and it
> is refused as group_fd even with PERF_FLAG_FD_NO_GROUP.
> 
> Fixes: 226424eee809 ("perf: Fix corruption of sibling list with hotplug")
> Suggested-by: Peter Zijlstra <peterz@infradead.org>
> Signed-off-by: Daniel Wade <danjwade95@gmail.com>
> Cc: stable@vger.kernel.org
> ---
> v2: use DETACH_EXIT at hot-unplug instead of a new check, as Peter
> suggested.
> v1: https://lore.kernel.org/all/20261116081248.21059-1-danjwade95@gmail.com/
> 
> Tested on mainline b3dfbde62cf1, x86-64 with KASAN and PROVE_LOCKING,
> 2 vCPUs under KVM. Without the patch a sibling is accepted on a
> hot-unplugged leader and the close gives 15 KASAN reports. With it the
> open fails with ENODEV and no report appears.
> 
>  kernel/events/core.c |    6 ++++--
>  1 file changed, 4 insertions(+), 2 deletions(-)
> 
> diff --git a/kernel/events/core.c b/kernel/events/core.c
> index f5f2261f1fe6..2de4c3cc37ac 100644
> --- a/kernel/events/core.c
> +++ b/kernel/events/core.c
> @@ -15342,8 +15342,10 @@ static void __perf_event_exit_context(void *__info)
>  
>  	raw_spin_lock(&ctx->lock);
>  	ctx_sched_out(ctx, NULL, EVENT_TIME);
> -	list_for_each_entry(event, &ctx->event_list, event_entry)
> -		__perf_remove_from_context(event, cpuctx, ctx, (void *)DETACH_GROUP);
> +	list_for_each_entry(event, &ctx->event_list, event_entry) {
> +		__perf_remove_from_context(event, cpuctx, ctx,
> +					   (void *)(DETACH_EXIT | DETACH_GROUP));
> +	}
>  	raw_spin_unlock(&ctx->lock);
>  }

The patch itself LGTM.

Reviewed-by: Aditya Chillara <aditya.chillara@oss.qualcomm.com>

PatchWise AI code review (gpt-6-astra) flagged the following though:

[medium] Adding DETACH_EXIT makes is_event_hup() true, but this path never
wakes event->waitq. __perf_remove_from_context() changes the state and
detaches the event without calling perf_event_wakeup(), unlike
perf_event_exit_event(). A thread blocked in poll/epoll on a disabled,
mmap-backed CPU event can remain asleep indefinitely after CPU unplug,
even though a fresh poll reports EPOLLHUP. Arrange a wakeup in an
appropriate context after publishing the terminal state so existing
waiters receive the hangup indication.

--

To be clear, this is not a regression: before your patch these events went
to OFF and poll() never reported a hangup at all, so a waiter slept just
the same. But the EPOLLHUP the patch advertises does not reach tasks that
are already waiting:

                          task already asleep in poll()   a fresh poll()
  before your patch       stays asleep                    no data
  your patch              stays asleep                    EPOLLHUP
  your patch + wakeup     woken, sees EPOLLHUP            EPOLLHUP

perf_event_exit_event() kicks perf_poll() with perf_event_wakeup() after
moving events to EXIT, but this path doesn't, so a task already sleeping
in poll() is not woken and only sees EPOLLHUP when its timeout runs out.

Deferring the wakeup to the event's irq_work, as merge_sched_in() does,
seems enough. Something like this, on top of your patch:

diff --git a/kernel/events/core.c b/kernel/events/core.c
index c1f3e1f013ec..885b6594139a 100644
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -15350,6 +15350,9 @@ static void __perf_event_exit_context(void *__info)
 	list_for_each_entry(event, &ctx->event_list, event_entry) {
 		__perf_remove_from_context(event, cpuctx, ctx,
 					   (void *)(DETACH_EXIT | DETACH_GROUP));
+		/* Kick perf_poll() for is_event_hup(). */
+		event->pending_wakeup = 1;
+		irq_work_queue(&event->pending_irq);
 	}
 	raw_spin_unlock(&ctx->lock);
 }

Tested in a 4-CPU x86_64 QEMU/KVM guest (virtme-ng), 7.4.0-rc3+ with
lockdep enabled, using perf record -C 1 and then taking CPU 1 offline:

  perf record -C 1 -e cpu-clock -o /tmp/hp.data &
  echo 0 > /sys/devices/system/cpu/cpu1/online
  jobs
  perf report -i /tmp/hp.data --stdio | head

A) Your patch only. perf keeps running after the unplug and needs Ctrl-C:

  # perf record -C 1 -e cpu-clock -o /tmp/hp.data &
  [1] 149
  # echo 0 > /sys/devices/system/cpu/cpu1/online
  # jobs
  [1]+  Running                 perf record -C 1 -e cpu-clock -o /tmp/hp.data &
  # fg
  perf record -C 1 -e cpu-clock -o /tmp/hp.data
  ^C[ perf record: Woken up 1 times to write data ]
  [ perf record: Captured and wrote 0.040 MB /tmp/hp.data (283 samples) ]
  # perf report -i /tmp/hp.data --stdio | head
  ...
  # Samples: 283  of event 'cpu-clock'
  # Event count (approx.): 70750000

B) Your patch plus the wakeup above. perf gets EPOLLHUP at once, flushes
   and exits by itself:

  # perf record -C 1 -e cpu-clock -o /tmp/hp.data &
  [1] 151
  # echo 0 > /sys/devices/system/cpu/cpu1/online
  [ perf record: Woken up 1 times to write data ]
  [ perf record: Captured and wrote 0.065 MB /tmp/hp.data (737 samples) ]

  [1]+  Done                    perf record -C 1 -e cpu-clock -o /tmp/hp.data
  # perf report -i /tmp/hp.data --stdio | head
  ...
  # Samples: 737  of event 'cpu-clock'
  # Event count (approx.): 184250000

Thank you,
Aditya


  reply	other threads:[~2026-11-19  8:54 UTC|newest]

Thread overview: 3+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2026-11-19  1:47 Daniel Wade
2026-11-19  8:54 ` Aditya Chillara [this message]
2026-11-19  9:19   ` Peter Zijlstra

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=06758809-6748-4f7b-b7e6-c862936a961c@oss.qualcomm.com \
    --to=aditya.chillara@oss.qualcomm.com \
    --cc=acme@kernel.org \
    --cc=adrian.hunter@intel.com \
    --cc=alexander.shishkin@linux.intel.com \
    --cc=danjwade95@gmail.com \
    --cc=irogers@google.com \
    --cc=james.clark@linaro.org \
    --cc=jolsa@kernel.org \
    --cc=kylebot@openai.com \
    --cc=linux-kernel@vger.kernel.org \
    --cc=linux-perf-users@vger.kernel.org \
    --cc=mark.rutland@arm.com \
    --cc=mingo@redhat.com \
    --cc=namhyung@kernel.org \
    --cc=peterz@infradead.org \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox

all inboxes | Powered by JetHome®