[PATCHSET v2 0/3] io_uring: support for linked SQEs

[PATCHSET v2 0/3] io_uring: support for linked SQEs

[Jens Axboe]

Here's v2 of the linked SQE patchset. For a full description of this
feature, please refer to the v1 posting:

https://lore.kernel.org/linux-block/20190517214131.5925-1-axboe@kernel.dk/

Changes since v1 are just fixes, and nothing major at that. Some of the
v1 error handling wasn't quite correct, this should be. The liburing
repo contains the link-cp copy implementation using linked SQEs, and
also has a test case that exercises a few of the link combinations.

Patches are against my for-linus branch, but should apply to Linus
master just fine as well.

 block/blk-core.c              |  74 +++++++++-
 block/blk-sysfs.c             |  47 ++----
 block/blk.h                   |   1 +
 drivers/block/loop.c          |  18 +--
 fs/aio.c                      |   9 +-
 fs/block_dev.c                |  25 ++--
 fs/io_uring.c                 | 268 ++++++++++++++++++++++++++--------
 fs/splice.c                   |   8 +-
 include/linux/uio.h           |   4 +-
 include/uapi/linux/io_uring.h |   1 +
 lib/iov_iter.c                |  15 +-
 net/compat.c                  |   3 +-
 net/socket.c                  |   3 +-
 13 files changed, 329 insertions(+), 147 deletions(-)

-- 
Jens Axboe

[PATCH 1/3] uio: make import_iovec()/compat_import_iovec() return bytes on success

[Jens Axboe]

Currently these functions return < 0 on error, and 0 for success.
Change that so that we return < 0 on error, but number of bytes
for success.

Some callers already treat the return value that way, others need a
slight tweak.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 fs/aio.c            |  9 +++++----
 fs/io_uring.c       | 16 ++++++++--------
 fs/splice.c         |  8 ++++----
 include/linux/uio.h |  4 ++--
 lib/iov_iter.c      | 15 ++++++++-------
 net/compat.c        |  3 ++-
 net/socket.c        |  3 ++-
 7 files changed, 31 insertions(+), 27 deletions(-)

diff --git a/fs/aio.c b/fs/aio.c
index 3490d1fa0e16..41824c710b36 100644
--- a/fs/aio.c
+++ b/fs/aio.c
@@ -1479,8 +1479,9 @@ static int aio_prep_rw(struct kiocb *req, const struct iocb *iocb)
 	return 0;
 }
 
-static int aio_setup_rw(int rw, const struct iocb *iocb, struct iovec **iovec,
-		bool vectored, bool compat, struct iov_iter *iter)
+static ssize_t aio_setup_rw(int rw, const struct iocb *iocb,
+		struct iovec **iovec, bool vectored, bool compat,
+		struct iov_iter *iter)
 {
 	void __user *buf = (void __user *)(uintptr_t)iocb->aio_buf;
 	size_t len = iocb->aio_nbytes;
@@ -1537,7 +1538,7 @@ static int aio_read(struct kiocb *req, const struct iocb *iocb,
 		return -EINVAL;
 
 	ret = aio_setup_rw(READ, iocb, &iovec, vectored, compat, &iter);
-	if (ret)
+	if (ret < 0)
 		return ret;
 	ret = rw_verify_area(READ, file, &req->ki_pos, iov_iter_count(&iter));
 	if (!ret)
@@ -1565,7 +1566,7 @@ static int aio_write(struct kiocb *req, const struct iocb *iocb,
 		return -EINVAL;
 
 	ret = aio_setup_rw(WRITE, iocb, &iovec, vectored, compat, &iter);
-	if (ret)
+	if (ret < 0)
 		return ret;
 	ret = rw_verify_area(WRITE, file, &req->ki_pos, iov_iter_count(&iter));
 	if (!ret) {
diff --git a/fs/io_uring.c b/fs/io_uring.c
index 0fbb486a320e..23e08c10f486 100644
--- a/fs/io_uring.c
+++ b/fs/io_uring.c
@@ -1003,9 +1003,9 @@ static int io_import_fixed(struct io_ring_ctx *ctx, int rw,
 	return 0;
 }
 
-static int io_import_iovec(struct io_ring_ctx *ctx, int rw,
-			   const struct sqe_submit *s, struct iovec **iovec,
-			   struct iov_iter *iter)
+static ssize_t io_import_iovec(struct io_ring_ctx *ctx, int rw,
+			       const struct sqe_submit *s, struct iovec **iovec,
+			       struct iov_iter *iter)
 {
 	const struct io_uring_sqe *sqe = s->sqe;
 	void __user *buf = u64_to_user_ptr(READ_ONCE(sqe->addr));
@@ -1023,7 +1023,7 @@ static int io_import_iovec(struct io_ring_ctx *ctx, int rw,
 	opcode = READ_ONCE(sqe->opcode);
 	if (opcode == IORING_OP_READ_FIXED ||
 	    opcode == IORING_OP_WRITE_FIXED) {
-		int ret = io_import_fixed(ctx, rw, sqe, iter);
+		ssize_t ret = io_import_fixed(ctx, rw, sqe, iter);
 		*iovec = NULL;
 		return ret;
 	}
@@ -1089,7 +1089,7 @@ static int io_read(struct io_kiocb *req, const struct sqe_submit *s,
 	struct iov_iter iter;
 	struct file *file;
 	size_t iov_count;
-	int ret;
+	ssize_t ret;
 
 	ret = io_prep_rw(req, s, force_nonblock);
 	if (ret)
@@ -1102,7 +1102,7 @@ static int io_read(struct io_kiocb *req, const struct sqe_submit *s,
 		return -EINVAL;
 
 	ret = io_import_iovec(req->ctx, READ, s, &iovec, &iter);
-	if (ret)
+	if (ret < 0)
 		return ret;
 
 	iov_count = iov_iter_count(&iter);
@@ -1136,7 +1136,7 @@ static int io_write(struct io_kiocb *req, const struct sqe_submit *s,
 	struct iov_iter iter;
 	struct file *file;
 	size_t iov_count;
-	int ret;
+	ssize_t ret;
 
 	ret = io_prep_rw(req, s, force_nonblock);
 	if (ret)
@@ -1149,7 +1149,7 @@ static int io_write(struct io_kiocb *req, const struct sqe_submit *s,
 		return -EINVAL;
 
 	ret = io_import_iovec(req->ctx, WRITE, s, &iovec, &iter);
-	if (ret)
+	if (ret < 0)
 		return ret;
 
 	iov_count = iov_iter_count(&iter);
diff --git a/fs/splice.c b/fs/splice.c
index 14cb602d9a2f..98412721f056 100644
--- a/fs/splice.c
+++ b/fs/splice.c
@@ -1356,7 +1356,7 @@ SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, uiov,
 	struct iovec iovstack[UIO_FASTIOV];
 	struct iovec *iov = iovstack;
 	struct iov_iter iter;
-	long error;
+	ssize_t error;
 	struct fd f;
 	int type;
 
@@ -1367,7 +1367,7 @@ SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, uiov,
 
 	error = import_iovec(type, uiov, nr_segs,
 			     ARRAY_SIZE(iovstack), &iov, &iter);
-	if (!error) {
+	if (error >= 0) {
 		error = do_vmsplice(f.file, &iter, flags);
 		kfree(iov);
 	}
@@ -1382,7 +1382,7 @@ COMPAT_SYSCALL_DEFINE4(vmsplice, int, fd, const struct compat_iovec __user *, io
 	struct iovec iovstack[UIO_FASTIOV];
 	struct iovec *iov = iovstack;
 	struct iov_iter iter;
-	long error;
+	ssize_t error;
 	struct fd f;
 	int type;
 
@@ -1393,7 +1393,7 @@ COMPAT_SYSCALL_DEFINE4(vmsplice, int, fd, const struct compat_iovec __user *, io
 
 	error = compat_import_iovec(type, iov32, nr_segs,
 			     ARRAY_SIZE(iovstack), &iov, &iter);
-	if (!error) {
+	if (error >= 0) {
 		error = do_vmsplice(f.file, &iter, flags);
 		kfree(iov);
 	}
diff --git a/include/linux/uio.h b/include/linux/uio.h
index 2d0131ad4604..a61ceb6575ab 100644
--- a/include/linux/uio.h
+++ b/include/linux/uio.h
@@ -279,13 +279,13 @@ bool csum_and_copy_from_iter_full(void *addr, size_t bytes, __wsum *csum, struct
 size_t hash_and_copy_to_iter(const void *addr, size_t bytes, void *hashp,
 		struct iov_iter *i);
 
-int import_iovec(int type, const struct iovec __user * uvector,
+ssize_t import_iovec(int type, const struct iovec __user * uvector,
 		 unsigned nr_segs, unsigned fast_segs,
 		 struct iovec **iov, struct iov_iter *i);
 
 #ifdef CONFIG_COMPAT
 struct compat_iovec;
-int compat_import_iovec(int type, const struct compat_iovec __user * uvector,
+ssize_t compat_import_iovec(int type, const struct compat_iovec __user * uvector,
 		 unsigned nr_segs, unsigned fast_segs,
 		 struct iovec **iov, struct iov_iter *i);
 #endif
diff --git a/lib/iov_iter.c b/lib/iov_iter.c
index f99c41d4eb54..f1e0569b4539 100644
--- a/lib/iov_iter.c
+++ b/lib/iov_iter.c
@@ -1634,9 +1634,9 @@ EXPORT_SYMBOL(dup_iter);
  * on-stack array was used or not (and regardless of whether this function
  * returns an error or not).
  *
- * Return: 0 on success or negative error code on error.
+ * Return: Negative error code on error, bytes imported on success
  */
-int import_iovec(int type, const struct iovec __user * uvector,
+ssize_t import_iovec(int type, const struct iovec __user * uvector,
 		 unsigned nr_segs, unsigned fast_segs,
 		 struct iovec **iov, struct iov_iter *i)
 {
@@ -1652,16 +1652,17 @@ int import_iovec(int type, const struct iovec __user * uvector,
 	}
 	iov_iter_init(i, type, p, nr_segs, n);
 	*iov = p == *iov ? NULL : p;
-	return 0;
+	return n;
 }
 EXPORT_SYMBOL(import_iovec);
 
 #ifdef CONFIG_COMPAT
 #include <linux/compat.h>
 
-int compat_import_iovec(int type, const struct compat_iovec __user * uvector,
-		 unsigned nr_segs, unsigned fast_segs,
-		 struct iovec **iov, struct iov_iter *i)
+ssize_t compat_import_iovec(int type,
+		const struct compat_iovec __user * uvector,
+		unsigned nr_segs, unsigned fast_segs,
+		struct iovec **iov, struct iov_iter *i)
 {
 	ssize_t n;
 	struct iovec *p;
@@ -1675,7 +1676,7 @@ int compat_import_iovec(int type, const struct compat_iovec __user * uvector,
 	}
 	iov_iter_init(i, type, p, nr_segs, n);
 	*iov = p == *iov ? NULL : p;
-	return 0;
+	return n;
 }
 #endif
 
diff --git a/net/compat.c b/net/compat.c
index 3f9ce609397f..0f7ded26059e 100644
--- a/net/compat.c
+++ b/net/compat.c
@@ -80,9 +80,10 @@ int get_compat_msghdr(struct msghdr *kmsg,
 
 	kmsg->msg_iocb = NULL;
 
-	return compat_import_iovec(save_addr ? READ : WRITE,
+	err = compat_import_iovec(save_addr ? READ : WRITE,
 				   compat_ptr(msg.msg_iov), msg.msg_iovlen,
 				   UIO_FASTIOV, iov, &kmsg->msg_iter);
+	return err < 0 ? err : 0;
 }
 
 /* Bleech... */
diff --git a/net/socket.c b/net/socket.c
index 72372dc5dd70..bffec466b4f1 100644
--- a/net/socket.c
+++ b/net/socket.c
@@ -2208,9 +2208,10 @@ static int copy_msghdr_from_user(struct msghdr *kmsg,
 
 	kmsg->msg_iocb = NULL;
 
-	return import_iovec(save_addr ? READ : WRITE,
+	err = import_iovec(save_addr ? READ : WRITE,
 			    msg.msg_iov, msg.msg_iovlen,
 			    UIO_FASTIOV, iov, &kmsg->msg_iter);
+	return err < 0 ? err : 0;
 }
 
 static int ___sys_sendmsg(struct socket *sock, struct user_msghdr __user *msg,
-- 
2.17.1

[PATCH 2/3] io_uring: punt short reads to async context

[Jens Axboe]

We can encounter a short read when we're doing buffered reads and the
data is partially cached. Right now we just return the short read, but
that forces the application to read that CQE, then issue another SQE
to finish the read. That read will not be cached, and hence will result
in an async punt.

It's more efficient to do that async punt from within the kernel, as
that will the not need two round trips more to the kernel.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 fs/io_uring.c | 15 +++++++++++++--
 1 file changed, 13 insertions(+), 2 deletions(-)

diff --git a/fs/io_uring.c b/fs/io_uring.c
index 23e08c10f486..92debd8be535 100644
--- a/fs/io_uring.c
+++ b/fs/io_uring.c
@@ -1089,7 +1089,7 @@ static int io_read(struct io_kiocb *req, const struct sqe_submit *s,
 	struct iov_iter iter;
 	struct file *file;
 	size_t iov_count;
-	ssize_t ret;
+	ssize_t read_size, ret;
 
 	ret = io_prep_rw(req, s, force_nonblock);
 	if (ret)
@@ -1105,13 +1105,24 @@ static int io_read(struct io_kiocb *req, const struct sqe_submit *s,
 	if (ret < 0)
 		return ret;
 
+	read_size = ret;
 	iov_count = iov_iter_count(&iter);
 	ret = rw_verify_area(READ, file, &kiocb->ki_pos, iov_count);
 	if (!ret) {
 		ssize_t ret2;
 
-		/* Catch -EAGAIN return for forced non-blocking submission */
 		ret2 = call_read_iter(file, kiocb, &iter);
+		/*
+		 * In case of a short read, punt to async. This can happen
+		 * if we have data partially cached. Alternatively we can
+		 * return the short read, in which case the application will
+		 * need to issue another SQE and wait for it. That SQE will
+		 * need async punt anyway, so it's more efficient to do it
+		 * here.
+		 */
+		if (force_nonblock && ret2 > 0 && ret2 < read_size)
+			ret2 = -EAGAIN;
+		/* Catch -EAGAIN return for forced non-blocking submission */
 		if (!force_nonblock || ret2 != -EAGAIN) {
 			io_rw_done(kiocb, ret2);
 		} else {
-- 
2.17.1

[PATCH 3/3] io_uring: add support for sqe links

[Jens Axboe]

With SQE links, we can create chains of dependent SQEs. One example
would be queueing an SQE that's a read from one file descriptor, with
the linked SQE being a write to another with the same set of buffers.

An SQE link will not stall the pipeline, it'll just ensure that
dependent SQEs aren't issued before the previous link has completed.

Any error at submission or completion time will break the chain of SQEs.
For completions, this also includes short reads or writes, as the next
SQE could depend on the previous one being fully completed.

Any SQE in a chain that gets canceled due to any of the above errors,
will get an CQE fill with -ECANCELED as the error value.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 fs/io_uring.c                 | 239 +++++++++++++++++++++++++++-------
 include/uapi/linux/io_uring.h |   1 +
 2 files changed, 192 insertions(+), 48 deletions(-)

diff --git a/fs/io_uring.c b/fs/io_uring.c
index 92debd8be535..df803a73849b 100644
--- a/fs/io_uring.c
+++ b/fs/io_uring.c
@@ -322,6 +322,7 @@ struct io_kiocb {
 
 	struct io_ring_ctx	*ctx;
 	struct list_head	list;
+	struct list_head	link_list;
 	unsigned int		flags;
 	refcount_t		refs;
 #define REQ_F_NOWAIT		1	/* must not punt to workers */
@@ -330,8 +331,10 @@ struct io_kiocb {
 #define REQ_F_SEQ_PREV		8	/* sequential with previous */
 #define REQ_F_IO_DRAIN		16	/* drain existing IO first */
 #define REQ_F_IO_DRAINED	32	/* drain done */
+#define REQ_F_LINK		64	/* linked sqes */
+#define REQ_F_FAIL_LINK		128	/* fail rest of links */
 	u64			user_data;
-	u32			error;	/* iopoll result from callback */
+	u32			result;
 	u32			sequence;
 
 	struct work_struct	work;
@@ -583,6 +586,7 @@ static struct io_kiocb *io_get_req(struct io_ring_ctx *ctx,
 	req->flags = 0;
 	/* one is dropped after submission, the other at completion */
 	refcount_set(&req->refs, 2);
+	req->result = 0;
 	return req;
 out:
 	io_ring_drop_ctx_refs(ctx, 1);
@@ -598,7 +602,7 @@ static void io_free_req_many(struct io_ring_ctx *ctx, void **reqs, int *nr)
 	}
 }
 
-static void io_free_req(struct io_kiocb *req)
+static void __io_free_req(struct io_kiocb *req)
 {
 	if (req->file && !(req->flags & REQ_F_FIXED_FILE))
 		fput(req->file);
@@ -606,6 +610,63 @@ static void io_free_req(struct io_kiocb *req)
 	kmem_cache_free(req_cachep, req);
 }
 
+static void io_req_link_next(struct io_kiocb *req)
+{
+	struct io_kiocb *nxt;
+
+	/*
+	 * The list should never be empty when we are called here. But could
+	 * potentially happen if the chain is messed up, check to be on the
+	 * safe side.
+	 */
+	nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb, list);
+	if (nxt) {
+		list_del(&nxt->list);
+		if (!list_empty(&req->link_list)) {
+			INIT_LIST_HEAD(&nxt->link_list);
+			list_splice(&req->link_list, &nxt->link_list);
+			nxt->flags |= REQ_F_LINK;
+		}
+
+		INIT_WORK(&nxt->work, io_sq_wq_submit_work);
+		queue_work(req->ctx->sqo_wq, &nxt->work);
+	}
+}
+
+/*
+ * Called if REQ_F_LINK is set, and we fail the head request
+ */
+static void io_fail_links(struct io_kiocb *req)
+{
+	struct io_kiocb *link;
+
+	while (!list_empty(&req->link_list)) {
+		link = list_first_entry(&req->link_list, struct io_kiocb, list);
+		list_del(&link->list);
+
+		io_cqring_add_event(req->ctx, link->user_data, -ECANCELED);
+		__io_free_req(link);
+	}
+}
+
+static void io_free_req(struct io_kiocb *req)
+{
+	/*
+	 * If LINK is set, we have dependent requests in this chain. If we
+	 * didn't fail this request, queue the first one up, moving any other
+	 * dependencies to the next request. In case of failure, fail the rest
+	 * of the chain.
+	 */
+	if (req->flags & REQ_F_LINK) {
+		if (req->flags & REQ_F_FAIL_LINK)
+			io_fail_links(req);
+		else
+			io_req_link_next(req);
+	}
+
+	__io_free_req(req);
+}
+
 static void io_put_req(struct io_kiocb *req)
 {
 	if (refcount_dec_and_test(&req->refs))
@@ -627,16 +688,17 @@ static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
 		req = list_first_entry(done, struct io_kiocb, list);
 		list_del(&req->list);
 
-		io_cqring_fill_event(ctx, req->user_data, req->error);
+		io_cqring_fill_event(ctx, req->user_data, req->result);
 		(*nr_events)++;
 
 		if (refcount_dec_and_test(&req->refs)) {
 			/* If we're not using fixed files, we have to pair the
 			 * completion part with the file put. Use regular
 			 * completions for those, only batch free for fixed
-			 * file.
+			 * file and non-linked commands.
 			 */
-			if (req->flags & REQ_F_FIXED_FILE) {
+			if ((req->flags & (REQ_F_FIXED_FILE|REQ_F_LINK)) ==
+			    REQ_F_FIXED_FILE) {
 				reqs[to_free++] = req;
 				if (to_free == ARRAY_SIZE(reqs))
 					io_free_req_many(ctx, reqs, &to_free);
@@ -775,6 +837,8 @@ static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
 
 	kiocb_end_write(kiocb);
 
+	if ((req->flags & REQ_F_LINK) && res != req->result)
+		req->flags |= REQ_F_FAIL_LINK;
 	io_cqring_add_event(req->ctx, req->user_data, res);
 	io_put_req(req);
 }
@@ -785,7 +849,9 @@ static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
 
 	kiocb_end_write(kiocb);
 
-	req->error = res;
+	if ((req->flags & REQ_F_LINK) && res != req->result)
+		req->flags |= REQ_F_FAIL_LINK;
+	req->result = res;
 	if (res != -EAGAIN)
 		req->flags |= REQ_F_IOPOLL_COMPLETED;
 }
@@ -928,7 +994,6 @@ static int io_prep_rw(struct io_kiocb *req, const struct sqe_submit *s,
 		    !kiocb->ki_filp->f_op->iopoll)
 			return -EOPNOTSUPP;
 
-		req->error = 0;
 		kiocb->ki_flags |= IOCB_HIPRI;
 		kiocb->ki_complete = io_complete_rw_iopoll;
 	} else {
@@ -1106,6 +1171,9 @@ static int io_read(struct io_kiocb *req, const struct sqe_submit *s,
 		return ret;
 
 	read_size = ret;
+	if (req->flags & REQ_F_LINK)
+		req->result = read_size;
+
 	iov_count = iov_iter_count(&iter);
 	ret = rw_verify_area(READ, file, &kiocb->ki_pos, iov_count);
 	if (!ret) {
@@ -1163,6 +1231,9 @@ static int io_write(struct io_kiocb *req, const struct sqe_submit *s,
 	if (ret < 0)
 		return ret;
 
+	if (req->flags & REQ_F_LINK)
+		req->result = ret;
+
 	iov_count = iov_iter_count(&iter);
 
 	ret = -EAGAIN;
@@ -1266,6 +1337,8 @@ static int io_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe,
 				end > 0 ? end : LLONG_MAX,
 				fsync_flags & IORING_FSYNC_DATASYNC);
 
+	if (ret < 0)
+		req->flags |= REQ_F_FAIL_LINK;
 	io_cqring_add_event(req->ctx, sqe->user_data, ret);
 	io_put_req(req);
 	return 0;
@@ -1310,6 +1383,8 @@ static int io_sync_file_range(struct io_kiocb *req,
 
 	ret = sync_file_range(req->rw.ki_filp, sqe_off, sqe_len, flags);
 
+	if (ret < 0)
+		req->flags |= REQ_F_FAIL_LINK;
 	io_cqring_add_event(req->ctx, sqe->user_data, ret);
 	io_put_req(req);
 	return 0;
@@ -1562,9 +1637,10 @@ static int __io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
 {
 	int ret, opcode;
 
+	req->user_data = READ_ONCE(s->sqe->user_data);
+
 	if (unlikely(s->index >= ctx->sq_entries))
 		return -EINVAL;
-	req->user_data = READ_ONCE(s->sqe->user_data);
 
 	opcode = READ_ONCE(s->sqe->opcode);
 	switch (opcode) {
@@ -1608,7 +1684,7 @@ static int __io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
 		return ret;
 
 	if (ctx->flags & IORING_SETUP_IOPOLL) {
-		if (req->error == -EAGAIN)
+		if (req->result == -EAGAIN)
 			return -EAGAIN;
 
 		/* workqueue context doesn't hold uring_lock, grab it now */
@@ -1834,31 +1910,11 @@ static int io_req_set_file(struct io_ring_ctx *ctx, const struct sqe_submit *s,
 	return 0;
 }
 
-static int io_submit_sqe(struct io_ring_ctx *ctx, struct sqe_submit *s,
-			 struct io_submit_state *state)
+static int io_queue_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
+			struct sqe_submit *s)
 {
-	struct io_kiocb *req;
 	int ret;
 
-	/* enforce forwards compatibility on users */
-	if (unlikely(s->sqe->flags & ~(IOSQE_FIXED_FILE | IOSQE_IO_DRAIN)))
-		return -EINVAL;
-
-	req = io_get_req(ctx, state);
-	if (unlikely(!req))
-		return -EAGAIN;
-
-	ret = io_req_set_file(ctx, s, state, req);
-	if (unlikely(ret))
-		goto out;
-
-	ret = io_req_defer(ctx, req, s->sqe);
-	if (ret) {
-		if (ret == -EIOCBQUEUED)
-			ret = 0;
-		return ret;
-	}
-
 	ret = __io_submit_sqe(ctx, req, s, true);
 	if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
 		struct io_uring_sqe *sqe_copy;
@@ -1881,24 +1937,91 @@ static int io_submit_sqe(struct io_ring_ctx *ctx, struct sqe_submit *s,
 
 			/*
 			 * Queued up for async execution, worker will release
-			 * submit reference when the iocb is actually
-			 * submitted.
+			 * submit reference when the iocb is actually submitted.
 			 */
 			return 0;
 		}
 	}
 
-out:
 	/* drop submission reference */
 	io_put_req(req);
 
 	/* and drop final reference, if we failed */
-	if (ret)
+	if (ret) {
+		io_cqring_add_event(ctx, req->user_data, ret);
+		if (req->flags & REQ_F_LINK)
+			req->flags |= REQ_F_FAIL_LINK;
 		io_put_req(req);
+	}
 
 	return ret;
 }
 
+#define SQE_VALID_FLAGS	(IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK)
+
+static void io_submit_sqe(struct io_ring_ctx *ctx, struct sqe_submit *s,
+			  struct io_submit_state *state, struct io_kiocb **link)
+{
+	struct io_uring_sqe *sqe_copy;
+	struct io_kiocb *req = NULL;
+	int ret;
+
+	/* enforce forwards compatibility on users */
+	if (unlikely(s->sqe->flags & ~SQE_VALID_FLAGS)) {
+		ret = -EINVAL;
+		goto err;
+	}
+
+	req = io_get_req(ctx, state);
+	if (unlikely(!req)) {
+		ret = -EAGAIN;
+		goto err;
+	}
+
+	ret = io_req_set_file(ctx, s, state, req);
+	if (unlikely(ret))
+		goto err;
+
+	ret = io_req_defer(ctx, req, s->sqe);
+	if (ret && ret != -EIOCBQUEUED)
+		goto err;
+
+	/*
+	 * If we already have a head request, queue this one for async
+	 * submittal once the head completes. If we don't have a head but
+	 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
+	 * submitted sync once the chain is complete. If none of those
+	 * conditions are true (normal request), then just queue it.
+	 */
+	if (*link) {
+		struct io_kiocb *prev = *link;
+
+		sqe_copy = kmemdup(s->sqe, sizeof(*sqe_copy), GFP_KERNEL);
+		if (!sqe_copy) {
+			ret = -EAGAIN;
+			goto err;
+		}
+
+		s->sqe = sqe_copy;
+		memcpy(&req->submit, s, sizeof(*s));
+		list_add_tail(&req->list, &prev->link_list);
+	} else if (s->sqe->flags & IOSQE_IO_LINK) {
+		req->flags |= REQ_F_LINK;
+
+		memcpy(&req->submit, s, sizeof(*s));
+		INIT_LIST_HEAD(&req->link_list);
+		*link = req;
+	} else {
+		io_queue_sqe(ctx, req, s);
+	}
+
+	return;
+err:
+	io_cqring_add_event(ctx, s->sqe->user_data, ret);
+	if (req)
+		io_free_req(req);
+}
+
 /*
  * Batched submission is done, ensure local IO is flushed out.
  */
@@ -1981,7 +2104,9 @@ static int io_submit_sqes(struct io_ring_ctx *ctx, struct sqe_submit *sqes,
 			  unsigned int nr, bool has_user, bool mm_fault)
 {
 	struct io_submit_state state, *statep = NULL;
-	int ret, i, submitted = 0;
+	struct io_kiocb *link = NULL;
+	bool prev_was_link = false;
+	int i, submitted = 0;
 
 	if (nr > IO_PLUG_THRESHOLD) {
 		io_submit_state_start(&state, ctx, nr);
@@ -1989,22 +2114,30 @@ static int io_submit_sqes(struct io_ring_ctx *ctx, struct sqe_submit *sqes,
 	}
 
 	for (i = 0; i < nr; i++) {
+		/*
+		 * If previous wasn't linked and we have a linked command,
+		 * that's the end of the chain. Submit the previous link.
+		 */
+		if (!prev_was_link && link) {
+			io_queue_sqe(ctx, link, &link->submit);
+			link = NULL;
+		}
+		prev_was_link = (sqes[i].sqe->flags & IOSQE_IO_LINK) != 0;
+
 		if (unlikely(mm_fault)) {
-			ret = -EFAULT;
+			io_cqring_add_event(ctx, sqes[i].sqe->user_data,
+						-EFAULT);
 		} else {
 			sqes[i].has_user = has_user;
 			sqes[i].needs_lock = true;
 			sqes[i].needs_fixed_file = true;
-			ret = io_submit_sqe(ctx, &sqes[i], statep);
-		}
-		if (!ret) {
+			io_submit_sqe(ctx, &sqes[i], statep, &link);
 			submitted++;
-			continue;
 		}
-
-		io_cqring_add_event(ctx, sqes[i].sqe->user_data, ret);
 	}
 
+	if (link)
+		io_queue_sqe(ctx, link, &link->submit);
 	if (statep)
 		io_submit_state_end(&state);
 
@@ -2145,6 +2278,8 @@ static int io_sq_thread(void *data)
 static int io_ring_submit(struct io_ring_ctx *ctx, unsigned int to_submit)
 {
 	struct io_submit_state state, *statep = NULL;
+	struct io_kiocb *link = NULL;
+	bool prev_was_link = false;
 	int i, submit = 0;
 
 	if (to_submit > IO_PLUG_THRESHOLD) {
@@ -2154,22 +2289,30 @@ static int io_ring_submit(struct io_ring_ctx *ctx, unsigned int to_submit)
 
 	for (i = 0; i < to_submit; i++) {
 		struct sqe_submit s;
-		int ret;
 
 		if (!io_get_sqring(ctx, &s))
 			break;
 
+		/*
+		 * If previous wasn't linked and we have a linked command,
+		 * that's the end of the chain. Submit the previous link.
+		 */
+		if (!prev_was_link && link) {
+			io_queue_sqe(ctx, link, &link->submit);
+			link = NULL;
+		}
+		prev_was_link = (s.sqe->flags & IOSQE_IO_LINK) != 0;
+
 		s.has_user = true;
 		s.needs_lock = false;
 		s.needs_fixed_file = false;
 		submit++;
-
-		ret = io_submit_sqe(ctx, &s, statep);
-		if (ret)
-			io_cqring_add_event(ctx, s.sqe->user_data, ret);
+		io_submit_sqe(ctx, &s, statep, &link);
 	}
 	io_commit_sqring(ctx);
 
+	if (link)
+		io_queue_sqe(ctx, link, &link->submit);
 	if (statep)
 		io_submit_state_end(statep);
 
diff --git a/include/uapi/linux/io_uring.h b/include/uapi/linux/io_uring.h
index a0c460025036..10b7c45f6d57 100644
--- a/include/uapi/linux/io_uring.h
+++ b/include/uapi/linux/io_uring.h
@@ -40,6 +40,7 @@ struct io_uring_sqe {
  */
 #define IOSQE_FIXED_FILE	(1U << 0)	/* use fixed fileset */
 #define IOSQE_IO_DRAIN		(1U << 1)	/* issue after inflight IO */
+#define IOSQE_IO_LINK		(1U << 2)	/* links next sqe */
 
 /*
  * io_uring_setup() flags
-- 
2.17.1