* Re: [PATCH v2] perf/bpf: Fix data races in BPF perf event handling
2026-08-13 6:05 [PATCH v2] perf/bpf: Fix data races in BPF perf event handling Deepanshu Kartikey
@ 2026-08-13 7:11 ` bot+bpf-ci
2026-08-13 11:05 ` Peter Zijlstra
1 sibling, 0 replies; 3+ messages in thread
From: bot+bpf-ci @ 2026-08-13 7:11 UTC (permalink / raw)
To: kartikey406, peterz, mingo, acme, namhyung, mark.rutland,
alexander.shishkin, jolsa, olsajiri, irogers, adrian.hunter,
james.clark, song, kpsingh, mattbobrowski, ast, daniel, andrii,
eddyz87, memxor, martin.lau, yonghong.song, emil, rostedt,
mhiramat, mathieu.desnoyers
Cc: linux-perf-users, linux-kernel, bpf, linux-trace-kernel,
kartikey406, syzbot+651d2774bd1d8395595f, ast, andrii, daniel,
martin.lau, eddyz87, yonghong.song, clm, ihor.solodrai
[-- Attachment #1: Type: text/plain, Size: 13991 bytes --]
> diff --git a/include/linux/perf_event.h b/include/linux/perf_event.h
> index 48d851fbd8ea5..5c6dabb6dccb1 100644
> --- a/include/linux/perf_event.h
> +++ b/include/linux/perf_event.h
> @@ -2136,4 +2136,6 @@ static inline void perf_lopwr_cb(bool mode)
> }
> #endif
>
> +extern struct mutex bpf_event_mutex;
> +
> #endif /* _LINUX_PERF_EVENT_H */
bpf_event_mutex is defined in kernel/trace/bpf_trace.c, which is only
built when CONFIG_BPF_EVENTS=y:
kernel/trace/Makefile: obj-$(CONFIG_BPF_EVENTS) += bpf_trace.o
CONFIG_BPF_EVENTS depends on BPF_SYSCALL and also requires either
KPROBE_EVENTS or UPROBE_EVENTS (kernel/trace/Kconfig:853-855). However,
the new reference in perf_event_alloc() is guarded by the weaker condition
'#if defined(CONFIG_BPF_SYSCALL) && defined(CONFIG_EVENT_TRACING)'.
CONFIG_EVENT_TRACING is selected by CONFIG_TRACING and has no dependency
on KPROBE_EVENTS or UPROBE_EVENTS. This means a config with
CONFIG_PERF_EVENTS=y, CONFIG_BPF_SYSCALL=y, CONFIG_TRACING=y,
CONFIG_KPROBES=n, and CONFIG_UPROBE_EVENTS=n will have CONFIG_BPF_EVENTS=n
and produce an undefined reference to bpf_event_mutex at link time.
Should the declaration be moved to include/linux/trace_events.h with
appropriate CONFIG_BPF_EVENTS guards, similar to how
perf_event_attach_bpf_prog() and perf_event_detach_bpf_prog() are
declared there?
> diff --git a/kernel/events/core.c b/kernel/events/core.c
> index ba5bd6a78fe7b..1416066ef460e 100644
> --- a/kernel/events/core.c
> +++ b/kernel/events/core.c
> @@ -10652,20 +10652,19 @@ static inline bool sample_is_allowed(struct perf_event *event, struct pt_regs *r
> #ifdef CONFIG_BPF_SYSCALL
> static int bpf_overflow_handler(struct perf_event *event,
> struct perf_sample_data *data,
> - struct pt_regs *regs)
> + struct pt_regs *regs,
> + struct bpf_prog *prog)
> {
> struct bpf_perf_event_data_kern ctx = {
> .data = data,
> .event = event,
> };
> - struct bpf_prog *prog;
> int ret = 0;
>
> ctx.regs = perf_arch_bpf_user_pt_regs(regs);
> if (unlikely(__this_cpu_inc_return(bpf_prog_active) != 1))
> goto out;
> rcu_read_lock();
> - prog = READ_ONCE(event->prog);
> if (prog) {
> perf_prepare_sample(data, event, regs);
> ret = bpf_prog_run(prog, &ctx);
The function signature was updated to add the fourth parameter, but the
!CONFIG_BPF_SYSCALL stub at the end of this file was not updated:
#else
static inline int bpf_overflow_handler(struct perf_event *event,
struct perf_sample_data *data,
struct pt_regs *regs)
{
return 1;
}
The call site at __perf_event_overflow() passes four arguments and is not
inside any CONFIG_BPF_SYSCALL guard. Since kernel/events/core.c is built
unconditionally for CONFIG_PERF_EVENTS=y, any CONFIG_PERF_EVENTS=y,
CONFIG_BPF_SYSCALL=n build will fail with "error: too many arguments to
function bpf_overflow_handler".
Should the stub also take the fourth 'struct bpf_prog *prog' parameter?
> @@ -10708,7 +10707,7 @@ static inline int perf_event_set_bpf_handler(struct perf_event *event,
> return -EPROTO;
> }
>
> - event->prog = prog;
> + WRITE_ONCE(event->prog, prog);
> event->bpf_cookie = bpf_cookie;
> return 0;
> }
> @@ -10720,7 +10719,7 @@ static inline void perf_event_free_bpf_handler(struct perf_event *event)
> if (!prog)
> return;
>
> - event->prog = NULL;
> + WRITE_ONCE(event->prog, NULL);
> bpf_prog_put(prog);
> }
> #else
> @@ -10753,6 +10752,7 @@ static int __perf_event_overflow(struct perf_event *event,
> {
> int events = atomic_read(&event->event_limit);
> int ret = 0;
> + struct bpf_prog *prog;
>
> /*
> * Non-sampling counters might still use the PMI to fold short
> @@ -10766,8 +10766,9 @@ static int __perf_event_overflow(struct perf_event *event,
> if (event->attr.aux_pause)
> perf_event_aux_pause(event->aux_event, true);
>
> - if (event->prog && event->prog->type == BPF_PROG_TYPE_PERF_EVENT &&
> - !bpf_overflow_handler(event, data, regs))
> + prog = READ_ONCE(event->prog);
> + if (prog && prog->type == BPF_PROG_TYPE_PERF_EVENT &&
> + !bpf_overflow_handler(event, data, regs, prog))
> goto out;
>
> /*
[ ... ]
> @@ -13433,12 +13434,15 @@ perf_event_alloc(struct perf_event_attr *attr, int cpu,
> overflow_handler = parent_event->overflow_handler;
> context = parent_event->overflow_handler_context;
> #if defined(CONFIG_BPF_SYSCALL) && defined(CONFIG_EVENT_TRACING)
> - if (parent_event->prog) {
> - struct bpf_prog *prog = parent_event->prog;
> + struct bpf_prog *prog;
>
> + mutex_lock(&bpf_event_mutex);
> + prog = parent_event->prog;
> + if (prog) {
> bpf_prog_inc(prog);
> event->prog = prog;
> }
> + mutex_unlock(&bpf_event_mutex);
> #endif
> }
>
bpf_event_mutex is taken to serialize against concurrent detach during
fork, but does this actually protect against all writers?
event->prog has two sets of writers:
(a) tracing events: perf_event_attach_bpf_prog() and
perf_event_detach_bpf_prog() in kernel/trace/bpf_trace.c - these hold
bpf_event_mutex
(b) non-tracing events (BPF_PROG_TYPE_PERF_EVENT):
perf_event_set_bpf_handler() and perf_event_free_bpf_handler() in
kernel/events/core.c - these do not hold bpf_event_mutex
For case (b), the concurrent path remains unprotected:
bpf_perf_link_release() (kernel/bpf/syscall.c, bpf_link fd close)
-> perf_event_free_bpf_prog()
-> perf_event_free_bpf_handler()
struct bpf_prog *prog = event->prog;
WRITE_ONCE(event->prog, NULL);
bpf_prog_put(prog);
versus the fork path here:
perf_event_init_context() (holds parent_ctx->mutex only)
-> inherit_task_group -> inherit_group -> inherit_event()
-> perf_event_alloc()
mutex_lock(&bpf_event_mutex);
prog = parent_event->prog;
if (prog) bpf_prog_inc(prog);
Timeline:
CPU0: mutex_lock(&bpf_event_mutex); prog = parent_event->prog -> P
CPU0: preempted
CPU1: perf_event_free_bpf_handler(): WRITE_ONCE(event->prog, NULL);
bpf_prog_put(P); -> refcnt -> 0, free deferred by call_rcu()
CPU0: bpf_prog_inc(P) on an object already at zero and queued for
freeing
bpf_perf_link_release() takes neither ctx->mutex nor bpf_event_mutex, so
the NULL-deref/use-after-free window remains open for overflow-handler
progs.
Should perf_event_free_bpf_handler() also take bpf_event_mutex, or should
the inherit be serialized differently (e.g., under the event's ctx->mutex)?
Also, the read stayed a plain load rather than READ_ONCE(), so it still
data-races with the WRITE_ONCE(event->prog, NULL) added in this same
patch.
> diff --git a/kernel/trace/bpf_trace.c b/kernel/trace/bpf_trace.c
> index 29260951aa871..e91676ae6cfb4 100644
> --- a/kernel/trace/bpf_trace.c
> +++ b/kernel/trace/bpf_trace.c
> @@ -1944,7 +1944,7 @@ const struct bpf_verifier_ops perf_event_verifier_ops = {
> const struct bpf_prog_ops perf_event_prog_ops = {
> };
>
> -static DEFINE_MUTEX(bpf_event_mutex);
> +DEFINE_MUTEX(bpf_event_mutex);
>
> #define BPF_TRACE_MAX_PROGS 64
[ ... ]
> @@ -2018,7 +2018,7 @@ void perf_event_detach_bpf_prog(struct perf_event *event)
>
> put:
> prog = event->prog;
> - event->prog = NULL;
> + WRITE_ONCE(event->prog, NULL);
>
> unlock:
> mutex_unlock(&bpf_event_mutex);
The commit message states "Fix by using WRITE_ONCE() in all write paths
including perf_event_detach_bpf_prog()", but the matching attach function
in the same file was left as a plain store:
kernel/trace/bpf_trace.c perf_event_attach_bpf_prog():
event->prog = prog;
perf_event_attach_bpf_prog() is a writer of the same field that
__perf_event_overflow() now reads with READ_ONCE(). The attach path can
be called on already-enabled tracepoint events that are concurrently
firing (via ioctl(PERF_EVENT_IOC_SET_BPF) or BPF_LINK_CREATE), so KCSAN
can still report 'data-race in __perf_event_overflow /
perf_event_attach_bpf_prog' (marked READ_ONCE vs plain write).
Similarly, several other plain reads remain:
- perf_event_free_bpf_handler(): 'struct bpf_prog *prog = event->prog;'
- perf_event_free_bpf_prog() read before dispatching
- perf_event_set_bpf_handler() checks before writing
- perf_event_detach_bpf_prog() reads before the WRITE_ONCE
- perf_event_alloc() read in the new mutex-protected section
Should the remaining plain writes be converted to WRITE_ONCE() and the
plain reads to READ_ONCE() to complete the annotation coverage?
---
The commit message describes the changes and their intent:
> Fix multiple data races involving event->prog field:
> 1. __perf_event_overflow() reads event->prog twice without
> synchronization creating a TOCTOU race. Fix by using READ_ONCE()
> to capture prog into a local variable and pass it directly to
> bpf_overflow_handler() to avoid a second read inside that function.
The fix does eliminate the double-read and the dropped-sample bug. However,
moving the READ_ONCE() out of bpf_overflow_handler() also moves it outside
the rcu_read_lock() that previously ensured the prog couldn't be freed
before bpf_prog_run().
Pre-patch: bpf_overflow_handler() did 'rcu_read_lock(); prog =
READ_ONCE(event->prog);' and then used that pointer. Since
perf_event_free_bpf_handler() stores NULL before bpf_prog_put(), and
freeing is RCU-deferred (call_rcu() in __bpf_prog_put_noref()), observing
a non-NULL prog inside the RCU section guaranteed the object couldn't be
freed before rcu_read_unlock().
Post-patch: the pointer is loaded in __perf_event_overflow() before
rcu_read_lock(), so in principle bpf_prog_run() could be handed an
already-freed prog if the grace period ends between the READ_ONCE() and
the rcu_read_lock().
Looking at the call paths, __perf_event_overflow() is reached with
preemption and IRQs disabled (perf_event_overflow() asserts
lockdep_assert_irqs_disabled()), which is an implicit RCU read-side
critical section covering the whole call, so the object is protected.
Should the commit message mention this trade-off, or should the code
document why the moved load is still safe despite being outside the
explicit rcu_read_lock()?
> 2. perf_event_set_bpf_handler() and perf_event_free_bpf_handler()
> perform plain writes to event->prog without WRITE_ONCE(), failing
> to pair with the READ_ONCE() in __perf_event_overflow().
Before this patch there was no READ_ONCE() in __perf_event_overflow() at
all - the only READ_ONCE(event->prog) was in bpf_overflow_handler(), and
this patch deletes it. The sentence describes the pre-existing code as
failing to pair with something the same patch introduces.
The substantive point (an unpaired marked read existed) is accurate, but
it was in bpf_overflow_handler(), not __perf_event_overflow().
Should this be clarified to say the plain writes failed to pair with the
READ_ONCE() that was in bpf_overflow_handler()?
> 3. perf_event_alloc() reads parent_event->prog locklessly during
> fork() which can race with a concurrent detach clearing and freeing
> the prog, potentially causing a NULL pointer dereference or
> use-after-free in bpf_prog_inc(). Fix by holding bpf_event_mutex
> when inheriting the BPF program.
As noted above, bpf_event_mutex is only held by the tracing event
attach/detach paths in kernel/trace/bpf_trace.c. It is not taken by
perf_event_free_bpf_handler(), which is the detach path for non-tracing
events (BPF_PROG_TYPE_PERF_EVENT).
perf_event_free_bpf_prog() dispatches based on perf_event_is_tracing()
and routes BPF_PROG_TYPE_PERF_EVENT progs to perf_event_free_bpf_handler(),
which does 'WRITE_ONCE(event->prog, NULL); bpf_prog_put(prog);' with no
lock at all.
A BPF_PROG_TYPE_PERF_EVENT prog can be attached to an inheritable sampling
event, and the inherit block in perf_event_alloc() is not restricted by
event type. For that case the NULL-deref/refcount-resurrection race
described above remains open.
Holding bpf_event_mutex does correctly close the race against
perf_event_detach_bpf_prog() (which clears event->prog under the mutex and
only calls bpf_prog_put() after unlocking).
Should the commit message clarify that the fix is partial and only covers
tracing events, or should perf_event_free_bpf_handler() also take
bpf_event_mutex?
> Make bpf_event_mutex non-static and declare it extern in perf_event.h
> so it is accessible from kernel/events/core.c.
The declaration is added to include/linux/perf_event.h outside any CONFIG
guard, but bpf_event_mutex is only defined when CONFIG_BPF_EVENTS=y
(kernel/trace/bpf_trace.o is only built for CONFIG_BPF_EVENTS).
The new mutex_lock/unlock reference in kernel/events/core.c is guarded by
'#if defined(CONFIG_BPF_SYSCALL) && defined(CONFIG_EVENT_TRACING)', which
is strictly weaker than CONFIG_BPF_EVENTS (CONFIG_BPF_EVENTS additionally
requires KPROBE_EVENTS || UPROBE_EVENTS). A valid config with
CONFIG_BPF_SYSCALL=y, CONFIG_TRACING=y, CONFIG_KPROBES=n, and
CONFIG_UPROBE_EVENTS=n yields CONFIG_BPF_EVENTS=n and an undefined
reference at link time.
This creates both a build failure and a layering concern (exposing a
bpf_trace.c-internal mutex through the global perf_event.h). Other
bpf_trace.c functions (perf_event_attach_bpf_prog,
perf_event_detach_bpf_prog, perf_event_query_prog_array) are declared in
include/linux/trace_events.h under CONFIG_BPF_EVENTS with !BPF_EVENTS
stubs.
Should bpf_event_mutex follow the same pattern?
---
This commit fixes data races but lacks a Fixes: tag. The bugs were
introduced when BPF perf event support was first added. Suggested:
Fixes: aa6a5f3cb2b2 ("perf, bpf: add perf events core support for BPF_PROG_TYPE_PERF_EVENT programs")
---
AI reviewed your patch. Please fix the bug or email reply why it's not a bug.
See: https://github.com/kernel-patches/vmtest/blob/master/ci/claude/README.md
CI run summary: https://github.com/kernel-patches/bpf/actions/runs/31673322805
^ permalink raw reply [flat|nested] 3+ messages in thread* Re: [PATCH v2] perf/bpf: Fix data races in BPF perf event handling
2026-08-13 6:05 [PATCH v2] perf/bpf: Fix data races in BPF perf event handling Deepanshu Kartikey
2026-08-13 7:11 ` bot+bpf-ci
@ 2026-08-13 11:05 ` Peter Zijlstra
1 sibling, 0 replies; 3+ messages in thread
From: Peter Zijlstra @ 2026-08-13 11:05 UTC (permalink / raw)
To: Deepanshu Kartikey
Cc: mingo, acme, namhyung, mark.rutland, alexander.shishkin, jolsa,
olsajiri, irogers, adrian.hunter, james.clark, song, kpsingh,
mattbobrowski, ast, daniel, andrii, eddyz87, memxor, martin.lau,
yonghong.song, emil, rostedt, mhiramat, mathieu.desnoyers,
linux-perf-users, linux-kernel, bpf, linux-trace-kernel,
syzbot+651d2774bd1d8395595f
On Thu, Aug 13, 2026 at 11:35:28AM +0530, Deepanshu Kartikey wrote:
> Fix multiple data races involving event->prog field:
>
> 1. __perf_event_overflow() reads event->prog twice without
> synchronization creating a TOCTOU race. Fix by using READ_ONCE()
> to capture prog into a local variable and pass it directly to
> bpf_overflow_handler() to avoid a second read inside that function.
Well, first you have to show there is concurrency where this matters,
since I still don't believe in your next point.
Also, there's actually worse issues when you consider ->prog and
->bpf_cookie form a pair.
> 2. perf_event_set_bpf_handler() and perf_event_free_bpf_handler()
> perform plain writes to event->prog without WRITE_ONCE(), failing
> to pair with the READ_ONCE() in __perf_event_overflow(). Fix by
> using WRITE_ONCE() in all write paths including
> perf_event_detach_bpf_prog().
As I said yesterday, how can perf_event_detach_bpf_prog() run
concurrently with __perf_event_overflow()? Unless you answer that, this
patch ain't moving nowhere.
> 3. perf_event_alloc() reads parent_event->prog locklessly during
> fork() which can race with a concurrent detach clearing and freeing
> the prog, potentially causing a NULL pointer dereference or
> use-after-free in bpf_prog_inc(). Fix by holding bpf_event_mutex
> when inheriting the BPF program. Make bpf_event_mutex non-static
> and declare it extern in perf_event.h so it is accessible from
> kernel/events/core.c.
This seems like a separate issue and should thus be a separate patch.
^ permalink raw reply [flat|nested] 3+ messages in thread