perf/core: Fix mlock accounting in perf_mmap()

commit 0034615 upstream.

Decreasing sysctl_perf_event_mlock between two consecutive perf_mmap()s of
a perf ring buffer may lead to an integer underflow in locked memory
accounting. This may lead to the undesired behaviors, such as failures in
BPF map creation.

Address this by adjusting the accounting logic to take into account the
possibility that the amount of already locked memory may exceed the
current limit.

Fixes: c4b7547 ("perf/core: Make the mlock accounting simple again")
Suggested-by: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Signed-off-by: Song Liu <songliubraving@fb.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Cc: <stable@vger.kernel.org>
Acked-by: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Link: https://lkml.kernel.org/r/20200123181146.2238074-1-songliubraving@fb.com
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

kernel/events/core.c

@@ -5709,7 +5709,15 @@ static int perf_mmap(struct file *file, struct vm_area_struct *vma)
 	 */
 	user_lock_limit *= num_online_cpus();
 
-	user_locked = atomic_long_read(&user->locked_vm) + user_extra;
+	user_locked = atomic_long_read(&user->locked_vm);
+
+	/*
+	 * sysctl_perf_event_mlock may have changed, so that
+	 *     user->locked_vm > user_lock_limit
+	 */
+	if (user_locked > user_lock_limit)
+		user_locked = user_lock_limit;
+	user_locked += user_extra;
 
 	if (user_locked > user_lock_limit)
 		extra = user_locked - user_lock_limit;