io_uring/rw: ensure poll based multishot read retries appropriately

io_read_mshot() always relies on poll triggering retries, and this works
fine as long as we do a retry per size of the buffer being read. The
buffer size is given by the size of the buffer(s) in the given buffer
group ID.

But if we're reading less than what is available, then we don't always
get to read everything that is available. For example, if the buffers
available are 32 bytes and we have 64 bytes to read, then we'll
correctly read the first 32 bytes and then wait for another poll trigger
before we attempt the next read. This next poll trigger may never
happen, in which case we just sit forever and never make progress, or it
may trigger at some point in the future, and now we're just delivering
the available data much later than we should have.

io_read_mshot() could do retries itself, but that is wasteful as we'll
be going through all of __io_read() again, and most likely in vain.
Rather than do that, bump our poll reference count and have
io_poll_check_events() do one more loop and check with vfs_poll() if we
have more data to read. If we do, io_read_mshot() will get invoked again
directly and we'll read the next chunk.

io_poll_multishot_retry() must only get called from inside
io_poll_issue(), which is our multishot retry handler, as we know we
already "own" the request at this point.

Cc: stable@vger.kernel.org
Link: https://github.com/axboe/liburing/issues/1041
Fixes: fc68fcda0491 ("io_uring/rw: add support for IORING_OP_READ_MULTISHOT")
Signed-off-by: Jens Axboe <axboe@kernel.dk>

2 files changed, 18 insertions, 1 deletions

diff --git a/io_uring/poll.h b/io_uring/poll.h
index ff4d5d753387e8..1dacae9e816c92 100644
--- a/io_uring/poll.h
+++ b/io_uring/poll.h
@@ -24,6 +24,15 @@ struct async_poll {
 	struct io_poll		*double_poll;
 };
 
+/*
+ * Must only be called inside issue_flags & IO_URING_F_MULTISHOT, or
+ * potentially other cases where we already "own" this poll request.
+ */
+static inline void io_poll_multishot_retry(struct io_kiocb *req)
+{
+	atomic_inc(&req->poll_refs);
+}
+
 int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe);
 int io_poll_add(struct io_kiocb *req, unsigned int issue_flags);
 
diff --git a/io_uring/rw.c b/io_uring/rw.c
index 118cc9f1cf1602..d5e79d9bdc717b 100644
--- a/io_uring/rw.c
+++ b/io_uring/rw.c
@@ -18,6 +18,7 @@
 #include "opdef.h"
 #include "kbuf.h"
 #include "rsrc.h"
+#include "poll.h"
 #include "rw.h"
 
 struct io_rw {
@@ -962,8 +963,15 @@ int io_read_mshot(struct io_kiocb *req, unsigned int issue_flags)
 		if (io_fill_cqe_req_aux(req,
 					issue_flags & IO_URING_F_COMPLETE_DEFER,
 					ret, cflags | IORING_CQE_F_MORE)) {
-			if (issue_flags & IO_URING_F_MULTISHOT)
+			if (issue_flags & IO_URING_F_MULTISHOT) {
+				/*
+				 * Force retry, as we might have more data to
+				 * be read and otherwise it won't get retried
+				 * until (if ever) another poll is triggered.
+				 */
+				io_poll_multishot_retry(req);
 				return IOU_ISSUE_SKIP_COMPLETE;
+			}
 			return -EAGAIN;
 		}
 	}