If there is no watch_queue, holding the pipe mutex is enough to
prevent concurrent writes, and we can avoid the spinlock.
O_NOTIFICATION_QUEUE is an exotic and rarely used feature, and of all
the pipes that exist at any given time, only very few actually have a
watch_queue, therefore it appears worthwile to optimize the common
case.
This patch does not optimize pipe_resize_ring() where the spinlocks
could be avoided as well; that does not seem like a worthwile
optimization because this function is not called often.
Related commits:
- commit 8df441294dd3 ("pipe: Check for ring full inside of the
spinlock in pipe_write()")
- commit b667b8673443 ("pipe: Advance tail pointer inside of wait
spinlock in pipe_read()")
- commit 189b0ddc2451 ("pipe: Fix missing lock in pipe_resize_ring()")
Signed-off-by: Max Kellermann <max.kellermann@xxxxxxxxx>
---
fs/pipe.c | 32 ++++++++++++++++++++++++++------
1 file changed, 26 insertions(+), 6 deletions(-)
diff --git a/fs/pipe.c b/fs/pipe.c
index 939def02c18c..da557eff9560 100644
--- a/fs/pipe.c
+++ b/fs/pipe.c
@@ -322,14 +322,34 @@ pipe_read(struct kiocb *iocb, struct iov_iter *to)
if (!buf->len) {
pipe_buf_release(pipe, buf);
- spin_lock_irq(&pipe->rd_wait.lock);
+
+ if (pipe_has_watch_queue(pipe)) {
+ /* if the pipe has a
+ * watch_queue, we need
+ * additional protection by
+ * the spinlock because
+ * notifications get posted
+ * with only this spinlock, no
+ * mutex
+ */
+
+ spin_lock_irq(&pipe->rd_wait.lock);
#ifdef CONFIG_WATCH_QUEUE
- if (buf->flags & PIPE_BUF_FLAG_LOSS)
- pipe->note_loss = true;
+ if (buf->flags & PIPE_BUF_FLAG_LOSS)
+ pipe->note_loss = true;
#endif
- tail++;
- pipe->tail = tail;
- spin_unlock_irq(&pipe->rd_wait.lock);
+ tail++;
+ pipe->tail = tail;
+ spin_unlock_irq(&pipe->rd_wait.lock);
+ } else {
+ /* without a watch_queue, we
+ * can simply increment the
+ * tail without the spinlock -
+ * the mutex is enough
+ */
+
+ pipe->tail = ++tail;
+ }
}
total_len -= chars;
if (!total_len)
--
2.39.2
