[PATCHSET 0/2] io_uring -EAGAIN IOPOLL retries

[PATCHSET 0/2] io_uring -EAGAIN IOPOLL retries

[Jens Axboe]

First patch is a fix, second one is just a cleanup that gets rid of a
useless task_work bounce for retries.

-- 
Jens Axboe

[PATCH 1/2] io_uring: fix IOPOLL -EAGAIN retries

[Jens Axboe]

This normally isn't hit, as polling is mostly done on NVMe with deep
queue depths. But if we do run into request starvation, we need to
ensure that retries are properly serialized.

Reported-by: Andres Freund <andres@anarazel.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 fs/io_uring.c | 10 +++++++---
 1 file changed, 7 insertions(+), 3 deletions(-)

diff --git a/fs/io_uring.c b/fs/io_uring.c
index 6df08287c59e..d27fe2b742d8 100644
--- a/fs/io_uring.c
+++ b/fs/io_uring.c
@@ -1150,7 +1150,7 @@ static void io_prep_async_work(struct io_kiocb *req)
 	io_req_init_async(req);
 
 	if (req->flags & REQ_F_ISREG) {
-		if (def->hash_reg_file)
+		if (def->hash_reg_file || (req->ctx->flags & IORING_SETUP_IOPOLL))
 			io_wq_hash_work(&req->work, file_inode(req->file));
 	} else {
 		if (def->unbound_nonreg_file)
@@ -3157,7 +3157,8 @@ static int io_read(struct io_kiocb *req, bool force_nonblock,
 		ret = 0;
 		goto out_free;
 	} else if (ret == -EAGAIN) {
-		if (!force_nonblock)
+		/* IOPOLL retry should happen for io-wq threads */
+		if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
 			goto done;
 		/* some cases will consume bytes even on error returns */
 		iov_iter_revert(iter, iov_count - iov_iter_count(iter));
@@ -3301,11 +3302,14 @@ static int io_write(struct io_kiocb *req, bool force_nonblock,
 	if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
 		ret2 = -EAGAIN;
 	if (!force_nonblock || ret2 != -EAGAIN) {
+		/* IOPOLL retry should happen for io-wq threads */
+		if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
+			goto copy_iov;
 		kiocb_done(kiocb, ret2, cs);
 	} else {
+copy_iov:
 		/* some cases will consume bytes even on error returns */
 		iov_iter_revert(iter, iov_count - iov_iter_count(iter));
-copy_iov:
 		ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
 		if (!ret)
 			return -EAGAIN;
-- 
2.28.0

[PATCH 2/2] io_uring: don't bounce block based -EAGAIN retry off task_work

[Jens Axboe]

These events happen inline from submission, so there's no need to
bounce them through the original task. Just set them up for retry
and issue retry directly instead of going over task_work.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 fs/io_uring.c | 26 ++++++--------------------
 1 file changed, 6 insertions(+), 20 deletions(-)

diff --git a/fs/io_uring.c b/fs/io_uring.c
index d27fe2b742d8..81143458f3b6 100644
--- a/fs/io_uring.c
+++ b/fs/io_uring.c
@@ -2295,22 +2295,6 @@ static bool io_resubmit_prep(struct io_kiocb *req, int error)
 	io_req_complete(req, ret);
 	return false;
 }
-
-static void io_rw_resubmit(struct callback_head *cb)
-{
-	struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
-	struct io_ring_ctx *ctx = req->ctx;
-	int err;
-
-	err = io_sq_thread_acquire_mm(ctx, req);
-
-	if (io_resubmit_prep(req, err)) {
-		refcount_inc(&req->refs);
-		io_queue_async_work(req);
-	}
-
-	percpu_ref_put(&ctx->refs);
-}
 #endif
 
 static bool io_rw_reissue(struct io_kiocb *req, long res)
@@ -2321,12 +2305,14 @@ static bool io_rw_reissue(struct io_kiocb *req, long res)
 	if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
 		return false;
 
-	init_task_work(&req->task_work, io_rw_resubmit);
-	percpu_ref_get(&req->ctx->refs);
+	ret = io_sq_thread_acquire_mm(req->ctx, req);
 
-	ret = io_req_task_work_add(req, &req->task_work, true);
-	if (!ret)
+	if (io_resubmit_prep(req, ret)) {
+		refcount_inc(&req->refs);
+		io_queue_async_work(req);
 		return true;
+	}
+
 #endif
 	return false;
 }
-- 
2.28.0