* [PATCH] perf: Reject group leaders detached by CPU hot-unplug
@ 2026-11-16 8:12 Daniel Wade
0 siblings, 0 replies; only message in thread
From: Daniel Wade @ 2026-11-16 8:12 UTC (permalink / raw)
To: Peter Zijlstra, Ingo Molnar, Arnaldo Carvalho de Melo, Namhyung Kim
Cc: Mark Rutland, Alexander Shishkin, Jiri Olsa, Ian Rogers,
Adrian Hunter, James Clark, Kyle Zeng, Aditya Chillara,
linux-perf-users, linux-kernel
When a CPU goes offline, __perf_event_exit_context() removes every event
of that CPU's context with DETACH_GROUP. That clears PERF_ATTACH_GROUP and
PERF_ATTACH_CONTEXT and leaves the event in PERF_EVENT_STATE_OFF, or in
ERROR, while the file descriptor stays open and event->ctx keeps pointing
at the per-CPU context; nothing re-attaches it when the CPU comes back.
Commit fa091f46c383 ("perf: Reject exited events as group leaders")
fixed the same defect for remove-on-exec, by rejecting a leader in a
state <= PERF_EVENT_STATE_EXIT. OFF and ERROR are both above EXIT, so
once the CPU is online again a leader detached by hot-unplug is still
accepted and perf_group_attach() links the new event to it.
perf_group_detach() returns early when PERF_ATTACH_GROUP is already
clear, which is right for the double detach it was written for, where no
group is left to break up. With a sibling added after the hot-unplug
detach it skips a sibling_list which is no longer empty: the leader is
freed with that sibling still on it and still pointing at it through
->group_leader. Closing the sibling, or a group read or ioctl on it, then
uses the freed event.
Commit 42c5ca1f0a28 ("perf/core: Fix group leader use-after-free after
sibling detach") fixed the mirror image, a sibling detached by hot-unplug
and the leader closed afterwards. It promotes the sibling from the paths
below that early return, so a leader detached first and given a sibling
afterwards still reaches the same free.
Reproduced on v7.4-rc2, which contains that commit, with KASAN in a
two-CPU guest, as root:
BUG: KASAN: slab-use-after-free in perf_event_update_time
Call Trace:
perf_event_update_time
__event_disable
__perf_remove_from_context
event_function
Freed by task 149 from an RCU callback queued by
perf_event_release_kernel(), via perf_release() and __fput().
Opening a CPU event needs CAP_PERFMON or perf_event_paranoid below 1, so
an ordinary user cannot reach this on a default configuration.
Reject a leader that is no longer attached to its context, rather than
promoting siblings of a detached leader in perf_group_detach(): the early
return there is right, and the anomaly is a leader on no context taking
new siblings for a group that can never be scheduled. Test
PERF_ATTACH_CONTEXT, which only the __perf_remove_from_context() path
clears, so every detached leader is covered whatever state it is left in,
and test it under ctx->mutex, where the leader is already known to share
this context and every detach path holds that mutex across the removal.
One open that works today fails as well: adding a hardware event to a
detached, pure software leader takes the move_group path, which installs
the leader again before the new event is linked to it, so that group was
valid. Events removed by hot-unplug are not scheduled again, which the
commit in the Fixes tag relies on, so it now gets ENODEV like any other
open on a detached leader.
Fixes: 226424eee809 ("perf: Fix corruption of sibling list with hotplug")
Signed-off-by: Daniel Wade <danjwade95@gmail.com>
Cc: stable@vger.kernel.org
---
Tested on mainline b3dfbde62cf1 (kernel/events identical to v7.4-rc2),
x86-64 with KASAN, PROVE_LOCKING and DEBUG_LIST, kasan_multi_shot=1, 2 vCPUs
under KVM, on a kernel carrying only this patch. Without it a software
sibling is accepted and the close produces 15 KASAN reports; with it both
the software and the hardware open fail with ENODEV and no report appears.
An unrelated test that moves an inherited software group to the hardware
context still passes.
Not tested: other architectures, hotplug or suspend on real hardware, the
perf tool's own test suite.
kernel/events/core.c | 12 ++++++++++++
1 file changed, 12 insertions(+)
diff --git a/kernel/events/core.c b/kernel/events/core.c
index f5f2261f1fe..21c529cd7d1 100644
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -14274,6 +14274,18 @@ SYSCALL_DEFINE5(perf_event_open,
goto err_locked;
}
+ /*
+ * Likewise against CPU hot-unplug, which detaches the events
+ * of a CPU context and breaks up its groups without putting
+ * them in a terminal state. perf_group_detach() does nothing
+ * for a leader that is already detached, so a sibling added
+ * to one would stay linked to it after the leader is freed.
+ */
+ if (!(group_leader->attach_state & PERF_ATTACH_CONTEXT)) {
+ err = -ENODEV;
+ goto err_locked;
+ }
+
/*
* Only a group leader can be exclusive or pinned
*/
^ permalink raw reply [flat|nested] only message in thread
only message in thread, other threads:[~2026-11-16 8:13 UTC | newest]
Thread overview: (only message) (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2026-11-16 8:12 [PATCH] perf: Reject group leaders detached by CPU hot-unplug Daniel Wade
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®