]> git.proxmox.com Git - mirror_ubuntu-hirsute-kernel.git/commitdiff
perf/core: Fix race in the perf_mmap_close() function
authorJiri Olsa <jolsa@redhat.com>
Wed, 16 Sep 2020 11:53:11 +0000 (13:53 +0200)
committerIngo Molnar <mingo@kernel.org>
Mon, 12 Oct 2020 11:24:26 +0000 (13:24 +0200)
There's a possible race in perf_mmap_close() when checking ring buffer's
mmap_count refcount value. The problem is that the mmap_count check is
not atomic because we call atomic_dec() and atomic_read() separately.

  perf_mmap_close:
  ...
   atomic_dec(&rb->mmap_count);
   ...
   if (atomic_read(&rb->mmap_count))
      goto out_put;

   <ring buffer detach>
   free_uid

out_put:
  ring_buffer_put(rb); /* could be last */

The race can happen when we have two (or more) events sharing same ring
buffer and they go through atomic_dec() and then they both see 0 as refcount
value later in atomic_read(). Then both will go on and execute code which
is meant to be run just once.

The code that detaches ring buffer is probably fine to be executed more
than once, but the problem is in calling free_uid(), which will later on
demonstrate in related crashes and refcount warnings, like:

  refcount_t: addition on 0; use-after-free.
  ...
  RIP: 0010:refcount_warn_saturate+0x6d/0xf
  ...
  Call Trace:
  prepare_creds+0x190/0x1e0
  copy_creds+0x35/0x172
  copy_process+0x471/0x1a80
  _do_fork+0x83/0x3a0
  __do_sys_wait4+0x83/0x90
  __do_sys_clone+0x85/0xa0
  do_syscall_64+0x5b/0x1e0
  entry_SYSCALL_64_after_hwframe+0x44/0xa9

Using atomic decrease and check instead of separated calls.

Tested-by: Michael Petlan <mpetlan@redhat.com>
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Acked-by: Namhyung Kim <namhyung@kernel.org>
Acked-by: Wade Mealing <wmealing@redhat.com>
Fixes: 9bb5d40cd93c ("perf: Fix mmap() accounting hole");
Link: https://lore.kernel.org/r/20200916115311.GE2301783@krava
kernel/events/core.c

index 45edb85344a1b274f8767cf54bd8822920025578..fb662eb4fb694f869a107de45ebcb2f79f41ef6e 100644 (file)
@@ -5895,11 +5895,11 @@ static void perf_pmu_output_stop(struct perf_event *event);
 static void perf_mmap_close(struct vm_area_struct *vma)
 {
        struct perf_event *event = vma->vm_file->private_data;
-
        struct perf_buffer *rb = ring_buffer_get(event);
        struct user_struct *mmap_user = rb->mmap_user;
        int mmap_locked = rb->mmap_locked;
        unsigned long size = perf_data_size(rb);
+       bool detach_rest = false;
 
        if (event->pmu->event_unmapped)
                event->pmu->event_unmapped(event, vma->vm_mm);
@@ -5930,7 +5930,8 @@ static void perf_mmap_close(struct vm_area_struct *vma)
                mutex_unlock(&event->mmap_mutex);
        }
 
-       atomic_dec(&rb->mmap_count);
+       if (atomic_dec_and_test(&rb->mmap_count))
+               detach_rest = true;
 
        if (!atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex))
                goto out_put;
@@ -5939,7 +5940,7 @@ static void perf_mmap_close(struct vm_area_struct *vma)
        mutex_unlock(&event->mmap_mutex);
 
        /* If there's still other mmap()s of this buffer, we're done. */
-       if (atomic_read(&rb->mmap_count))
+       if (!detach_rest)
                goto out_put;
 
        /*