]> git.ipfire.org Git - thirdparty/kernel/linux.git/commitdiff
perf: Fix refcount warning on event->mmap_count increment
authorWill Rosenberg <whrosenb@asu.edu>
Mon, 19 Jan 2026 18:49:56 +0000 (11:49 -0700)
committerPeter Zijlstra <peterz@infradead.org>
Wed, 21 Jan 2026 15:28:58 +0000 (16:28 +0100)
When calling refcount_inc(&event->mmap_count) inside perf_mmap_rb(), the
following warning is triggered:

        refcount_t: addition on 0; use-after-free.
        WARNING: lib/refcount.c:25

PoC:

    struct perf_event_attr attr = {0};
    int fd = syscall(__NR_perf_event_open, &attr, 0, -1, -1, 0);
    mmap(NULL, 0x3000, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
    int victim = syscall(__NR_perf_event_open, &attr, 0, -1, fd,
                         PERF_FLAG_FD_OUTPUT);
    mmap(NULL, 0x3000, PROT_READ | PROT_WRITE, MAP_SHARED, victim, 0);

This occurs when creating a group member event with the flag
PERF_FLAG_FD_OUTPUT. The group leader should be mmap-ed and then mmap-ing
the event triggers the warning.

Since the event has copied the output_event in perf_event_set_output(),
event->rb is set. As a result, perf_mmap_rb() calls
refcount_inc(&event->mmap_count) when event->mmap_count = 0.

Disallow the case when event->mmap_count = 0. This also prevents two
events from updating the same user_page.

Fixes: 448f97fba901 ("perf: Convert mmap() refcounts to refcount_t")
Suggested-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Will Rosenberg <whrosenb@asu.edu>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://patch.msgid.link/20260119184956.801238-1-whrosenb@asu.edu
kernel/events/core.c

index 5b5cb620499efccffcd1f322cacaec9cc6945643..a0fa488bce8465c1c1bfa1935d18775bbfa58c41 100644 (file)
@@ -6997,6 +6997,15 @@ static int perf_mmap_rb(struct vm_area_struct *vma, struct perf_event *event,
                if (data_page_nr(event->rb) != nr_pages)
                        return -EINVAL;
 
+               /*
+                * If this event doesn't have mmap_count, we're attempting to
+                * create an alias of another event's mmap(); this would mean
+                * both events will end up scribbling the same user_page;
+                * which makes no sense.
+                */
+               if (!refcount_read(&event->mmap_count))
+                       return -EBUSY;
+
                if (refcount_inc_not_zero(&event->rb->mmap_count)) {
                        /*
                         * Success -- managed to mmap() the same buffer