[PATCH 2/4] io_uring: patch up IOPOLL overflow_flush sync

[Pavel Begunkov <asml.silence@xxxxxxxxx>]

IOPOLL skips completion locking but keeps it under uring_lock, thus
io_cqring_overflow_flush() and so io_cqring_events() need extra care.
Add extra conditional locking around them.

Signed-off-by: Pavel Begunkov <asml.silence@xxxxxxxxx>
---
 fs/io_uring.c | 23 +++++++++++++++++------
 1 file changed, 17 insertions(+), 6 deletions(-)

diff --git a/fs/io_uring.c b/fs/io_uring.c
index 00dd85acd039..a4deef746bc3 100644
--- a/fs/io_uring.c
+++ b/fs/io_uring.c
@@ -2312,7 +2312,8 @@ static void io_double_put_req(struct io_kiocb *req)
 		io_free_req(req);
 }
 
-static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
+static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush,
+				 bool iopoll_lock)
 {
 	if (test_bit(0, &ctx->cq_check_overflow)) {
 		/*
@@ -2323,7 +2324,15 @@ static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
 		if (noflush)
 			return -1U;
 
+		/*
+		 * iopoll doesn't care about ctx->completion_lock but uses
+		 * ctx->uring_lock
+		 */
+		if (iopoll_lock)
+			mutex_lock(&ctx->uring_lock);
 		io_cqring_overflow_flush(ctx, false, NULL, NULL);
+		if (iopoll_lock)
+			mutex_unlock(&ctx->uring_lock);
 	}
 
 	/* See comment at the top of this file */
@@ -2550,7 +2559,7 @@ static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
 		 * If we do, we can potentially be spinning for commands that
 		 * already triggered a CQE (eg in error).
 		 */
-		if (io_cqring_events(ctx, false))
+		if (io_cqring_events(ctx, false, false))
 			break;
 
 		/*
@@ -7097,7 +7106,7 @@ static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
 	 * started waiting. For timeouts, we always want to return to userspace,
 	 * regardless of event count.
 	 */
-	return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
+	return io_cqring_events(ctx, noflush, false) >= iowq->to_wait ||
 			atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
 }
 
@@ -7142,13 +7151,14 @@ static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
 		.ctx		= ctx,
 		.to_wait	= min_events,
 	};
+	bool iopoll = ctx->flags & IORING_SETUP_IOPOLL;
 	struct io_rings *rings = ctx->rings;
 	struct timespec64 ts;
 	signed long timeout = 0;
 	int ret = 0;
 
 	do {
-		if (io_cqring_events(ctx, false) >= min_events)
+		if (io_cqring_events(ctx, false, iopoll) >= min_events)
 			return 0;
 		if (!io_run_task_work())
 			break;
@@ -7184,7 +7194,8 @@ static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
 			continue;
 		else if (ret < 0)
 			break;
-		if (io_should_wake(&iowq, false))
+		/* iopoll ignores completion_lock, so not safe to flush */
+		if (io_should_wake(&iowq, iopoll))
 			break;
 		if (uts) {
 			timeout = schedule_timeout(timeout);
@@ -8623,7 +8634,7 @@ static __poll_t io_uring_poll(struct file *file, poll_table *wait)
 	smp_rmb();
 	if (!io_sqring_full(ctx))
 		mask |= EPOLLOUT | EPOLLWRNORM;
-	if (io_cqring_events(ctx, false))
+	if (io_cqring_events(ctx, false, ctx->flags & IORING_SETUP_IOPOLL))
 		mask |= EPOLLIN | EPOLLRDNORM;
 
 	return mask;
-- 
2.24.0