watch_queue: fix pipe accounting mismatch

Currently, watch_queue_set_size() modifies the pipe buffers charged to
user->pipe_bufs without updating the pipe->nr_accounted on the pipe
itself, due to the if (!pipe_has_watch_queue()) test in
pipe_resize_ring(). This means that when the pipe is ultimately freed,
we decrement user->pipe_bufs by something other than what than we had
charged to it, potentially leading to an underflow. This in turn can
cause subsequent too_many_pipe_buffers_soft() tests to fail with -EPERM.

To remedy this, explicitly account for the pipe usage in
watch_queue_set_size() to match the number set via account_pipe_buffers()

(It's unclear why watch_queue_set_size() does not update nr_accounted;
it may be due to intentional overprovisioning in watch_queue_set_size()?)

Fixes: e95aada4cb93d ("pipe: wakeup wr_wait after setting max_usage")
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Link: https://lore.kernel.org/r/206682a8-0604-49e5-8224-fdbe0c12b460@redhat.com
Signed-off-by: Christian Brauner <brauner@kernel.org>

diff --git a/kernel/watch_queue.c b/kernel/watch_queue.c
index 5267adeaa403456ba80473e6a9584fdddfe1b939..41e4e8070923a9dc3e4eb7e6acdbd91e5e0c1ea9 100644 (file)
--- a/kernel/watch_queue.c
+++ b/kernel/watch_queue.c
@@ -269,6 +269,15 @@ long watch_queue_set_size(struct pipe_inode_info *pipe, unsigned int nr_notes)
        if (ret < 0)
                goto error;
 
+       /*
+        * pipe_resize_ring() does not update nr_accounted for watch_queue
+        * pipes, because the above vastly overprovisions. Set nr_accounted on
+        * and max_usage this pipe to the number that was actually charged to
+        * the user above via account_pipe_buffers.
+        */
+       pipe->max_usage = nr_pages;
+       pipe->nr_accounted = nr_pages;
+
        ret = -ENOMEM;
        pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
        if (!pages)