Inline sqe_submit

PATCH In review Pavel Begunkov 2 participants 6 messages Nov 6, 2019

Subsystems (3)

VFS Block layer io_uring

lore

Share & tools

lore

Pavel Begunkov

author

2019-11-06 22:41 UTC #0 lore raw

v3 cover Inline sqe_submit

The idea is to not pass struct sqe_submit as a separate entity, but always use req->submit instead, so there will be less stuff to care about.

Also, I've got steady +1% throughput improvement for nop tests. Though, it's highly system-dependent, and I wouldn't count on it.

v2: fix use-after-free catched by Jens

v3: -EAGAIN, in case submission failed

Pavel Begunkov (3):
io_uring: allocate io_kiocb upfront
io_uring: Use submit info inlined into req
io_uring: use inlined struct sqe_submit

 fs/io_uring.c | 134 +++++++++++++++++++++++++-------------------------
 1 file changed, 67 insertions(+), 67 deletions(-)

Signature · 1 line

2.23.0

Pavel Begunkov

author

2019-11-06 22:41 UTC #1 lore raw

PATCH v3 1/3 io_uring: allocate io_kiocb upfront

Let io_submit_sqes() to allocate io_kiocb before fetching an sqe.

Signed-off-by Pavel Begunkov <[email protected]>

---
 fs/io_uring.c | 27 ++++++++++++++-------------
 1 file changed, 14 insertions(+), 13 deletions(-)

fs/io_uring.c +14 ~~-13~~

index 6524898831e0..ceb616dbe710 100644
@@ -2551,30 +2551,23 @@ static int io_queue_link_head(struct io_ring_ctx *ctx, struct io_kiocb *req,
 
 #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)
+static void io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
+			  struct sqe_submit *s, struct io_submit_state *state,
+			  struct io_kiocb **link)
 {
 	struct io_uring_sqe *sqe_copy;
-	struct io_kiocb *req;
 	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;
+		goto err_req;
 	}
 
 	ret = io_req_set_file(ctx, s, state, req);
 	if (unlikely(ret)) {
 err_req:
 		io_free_req(req, NULL);
-err:
 		io_cqring_add_event(ctx, s->sqe->user_data, ret);
 		return;
 	}

@@ -2710,9 +2703,15 @@ static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
 
 	for (i = 0; i < nr; i++) {
 		struct sqe_submit s;
+		struct io_kiocb *req;
 
-		if (!io_get_sqring(ctx, &s))
+		req = io_get_req(ctx, statep);
+		if (unlikely(!req))
 			break;
+		if (!io_get_sqring(ctx, &s)) {
+			__io_free_req(req);
+			break;
+		}
 
 		if (io_sqe_needs_user(s.sqe) && !*mm) {
 			mm_fault = mm_fault || !mmget_not_zero(ctx->sqo_mm);

@@ -2740,7 +2739,7 @@ static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
 		s.in_async = async;
 		s.needs_fixed_file = async;
 		trace_io_uring_submit_sqe(ctx, s.sqe->user_data, true, async);
-		io_submit_sqe(ctx, &s, statep, &link);
+		io_submit_sqe(ctx, req, &s, statep, &link);
 		submitted++;
 
 		/*

@@ -4009,6 +4008,8 @@ SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
 		cur_mm = ctx->sqo_mm;
 		submitted = io_submit_sqes(ctx, to_submit, f.file, fd,
 					   &cur_mm, false);
+		if (!submitted)
+			submitted = -EAGAIN;
 		mutex_unlock(&ctx->uring_lock);
 	}
 	if (flags & IORING_ENTER_GETEVENTS) {

Signature · 1 line

2.23.0

Pavel Begunkov

author

2019-11-06 22:47 UTC #2 lore raw

Re: [PATCH v3 1/3] io_uring: allocate io_kiocb upfront

Quoted 73 lines · Pavel Begunkov

Let io_submit_sqes() to allocate io_kiocb before fetching an sqe.

Signed-off-by Pavel Begunkov <[email protected]>

---
 fs/io_uring.c | 27 ++++++++++++++-------------
 1 file changed, 14 insertions(+), 13 deletions(-)

fs/io_uring.c +12 ~~-13~~

index 6524898831e0..ceb616dbe710 100644
@@ -2551,30 +2551,23 @@ static int io_queue_link_head(struct io_ring_ctx *ctx, struct io_kiocb *req,
 
 #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)
+static void io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
+			  struct sqe_submit *s, struct io_submit_state *state,
+			  struct io_kiocb **link)
 {
 	struct io_uring_sqe *sqe_copy;
-	struct io_kiocb *req;
 	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;
+		goto err_req;
 	}
 
 	ret = io_req_set_file(ctx, s, state, req);
 	if (unlikely(ret)) {
 err_req:
 		io_free_req(req, NULL);
-err:
 		io_cqring_add_event(ctx, s->sqe->user_data, ret);
 		return;
 	}

@@ -2710,9 +2703,15 @@ static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
 
 	for (i = 0; i < nr; i++) {
 		struct sqe_submit s;
+		struct io_kiocb *req;
 
-		if (!io_get_sqring(ctx, &s))
+		req = io_get_req(ctx, statep);
+		if (unlikely(!req))
 			break;
+		if (!io_get_sqring(ctx, &s)) {
+			__io_free_req(req);
+			break;
+		}
 
 		if (io_sqe_needs_user(s.sqe) && !*mm) {
 			mm_fault = mm_fault || !mmget_not_zero(ctx->sqo_mm);

@@ -2740,7 +2739,7 @@ static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
 		s.in_async = async;
 		s.needs_fixed_file = async;
 		trace_io_uring_submit_sqe(ctx, s.sqe->user_data, true, async);
-		io_submit_sqe(ctx, &s, statep, &link);
+		io_submit_sqe(ctx, req, &s, statep, &link);
 		submitted++;
 
 		/*

fs/io_uring.c SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit, · line 4012

Quoted 5 lines

fs/io_uring.c +2 -0

@@ -4009,6 +4008,8 @@ SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
 		cur_mm = ctx->sqo_mm;
 		submitted = io_submit_sqes(ctx, to_submit, f.file, fd,
 					   &cur_mm, false);
+		if (!submitted)
+			submitted = -EAGAIN;

Need to say, that left io_submit_sqes() errors of more priority than ones of poll/wait.

Quoted 4 lines

fs/io_uring.c +0 -0

 		mutex_unlock(&ctx->uring_lock);
 	}
 	if (flags & IORING_ENTER_GETEVENTS) {

Signature · 1 line

Pavel Begunkov

Pavel Begunkov

author

2019-11-06 22:41 UTC #3 lore raw

PATCH v3 2/3 io_uring: Use submit info inlined into req

In reply to Pavel Begunkov

Stack allocated struct sqe_submit is passed down to the submission path along with a request (a.k.a. struct io_kiocb), and will be copied into req->submit for async requests.

As space for it is already allocated, fill req->submit in the first place instead of using on-stack one. As a result:

1. sqe->submit is the only place for sqe_submit and is always valid, so we don't need to track which one to use.
2. don't need to copy in case of async
3. allows to simplify the code by not carrying it as an argument all the way down
4. allows to reduce number of function arguments / potentially improve spilling

The downside is that stack is most probably be cached, that's not true for just allocated memory for a request. Another concern is cache pollution. Though, a request would be touched and fetched along with req->submit at some point anyway, so shouldn't be a problem.

Signed-off-by Pavel Begunkov <[email protected]>

---
 fs/io_uring.c | 32 ++++++++++++++++----------------
 1 file changed, 16 insertions(+), 16 deletions(-)

fs/io_uring.c +16 ~~-16~~

index ceb616dbe710..c0d2601dc17b 100644
@@ -2456,7 +2456,6 @@ static int __io_queue_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
 		sqe_copy = kmemdup(s->sqe, sizeof(*sqe_copy), GFP_KERNEL);
 		if (sqe_copy) {
 			s->sqe = sqe_copy;
-			memcpy(&req->submit, s, sizeof(*s));
 			if (req->work.flags & IO_WQ_WORK_NEEDS_FILES) {
 				ret = io_grab_files(ctx, req);
 				if (ret) {

@@ -2591,13 +2590,11 @@ static void io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
 		}
 
 		s->sqe = sqe_copy;
-		memcpy(&req->submit, s, sizeof(*s));
 		trace_io_uring_link(ctx, req, prev);
 		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 {

@@ -2702,18 +2699,18 @@ static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
 	}
 
 	for (i = 0; i < nr; i++) {
-		struct sqe_submit s;
 		struct io_kiocb *req;
+		unsigned int sqe_flags;
 
 		req = io_get_req(ctx, statep);
 		if (unlikely(!req))
 			break;
-		if (!io_get_sqring(ctx, &s)) {
+		if (!io_get_sqring(ctx, &req->submit)) {
 			__io_free_req(req);
 			break;
 		}
 
-		if (io_sqe_needs_user(s.sqe) && !*mm) {
+		if (io_sqe_needs_user(req->submit.sqe) && !*mm) {
 			mm_fault = mm_fault || !mmget_not_zero(ctx->sqo_mm);
 			if (!mm_fault) {
 				use_mm(ctx->sqo_mm);

@@ -2721,7 +2718,9 @@ static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
 			}
 		}
 
-		if (link && (s.sqe->flags & IOSQE_IO_DRAIN)) {
+		sqe_flags = req->submit.sqe->flags;
+
+		if (link && (sqe_flags & IOSQE_IO_DRAIN)) {
 			if (!shadow_req) {
 				shadow_req = io_get_req(ctx, NULL);
 				if (unlikely(!shadow_req))

@@ -2729,24 +2728,25 @@ static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
 				shadow_req->flags |= (REQ_F_IO_DRAIN | REQ_F_SHADOW_DRAIN);
 				refcount_dec(&shadow_req->refs);
 			}
-			shadow_req->sequence = s.sequence;
+			shadow_req->sequence = req->submit.sequence;
 		}
 
 out:
-		s.ring_file = ring_file;
-		s.ring_fd = ring_fd;
-		s.has_user = *mm != NULL;
-		s.in_async = async;
-		s.needs_fixed_file = async;
-		trace_io_uring_submit_sqe(ctx, s.sqe->user_data, true, async);
-		io_submit_sqe(ctx, req, &s, statep, &link);
+		req->submit.ring_file = ring_file;
+		req->submit.ring_fd = ring_fd;
+		req->submit.has_user = *mm != NULL;
+		req->submit.in_async = async;
+		req->submit.needs_fixed_file = async;
+		trace_io_uring_submit_sqe(ctx, req->submit.sqe->user_data,
+					  true, async);
+		io_submit_sqe(ctx, req, &req->submit, statep, &link);
 		submitted++;
 
 		/*
 		 * If previous wasn't linked and we have a linked command,
 		 * that's the end of the chain. Submit the previous link.
 		 */
-		if (!(s.sqe->flags & IOSQE_IO_LINK) && link) {
+		if (!(sqe_flags & IOSQE_IO_LINK) && link) {
 			io_queue_link_head(ctx, link, &link->submit, shadow_req);
 			link = NULL;
 			shadow_req = NULL;

Signature · 1 line

2.23.0

Pavel Begunkov

author

2019-11-06 22:41 UTC #4 lore raw

PATCH v3 3/3 io_uring: use inlined struct sqe_submit

In reply to Pavel Begunkov

req->submit is always up-to-date, use it directly

Signed-off-by Pavel Begunkov <[email protected]>

---
 fs/io_uring.c | 87 +++++++++++++++++++++++++--------------------------
 1 file changed, 43 insertions(+), 44 deletions(-)

fs/io_uring.c +43 ~~-44~~

index c0d2601dc17b..2c56c3b9c828 100644
@@ -1155,10 +1155,9 @@ static bool io_file_supports_async(struct file *file)
 	return false;
 }
 
-static int io_prep_rw(struct io_kiocb *req, const struct sqe_submit *s,
-		      bool force_nonblock)
+static int io_prep_rw(struct io_kiocb *req, bool force_nonblock)
 {
-	const struct io_uring_sqe *sqe = s->sqe;
+	const struct io_uring_sqe *sqe = req->submit.sqe;
 	struct io_ring_ctx *ctx = req->ctx;
 	struct kiocb *kiocb = &req->rw;
 	unsigned ioprio;

@@ -1406,8 +1405,8 @@ static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
 	return ret;
 }
 
-static int io_read(struct io_kiocb *req, const struct sqe_submit *s,
-		   struct io_kiocb **nxt, bool force_nonblock)
+static int io_read(struct io_kiocb *req, struct io_kiocb **nxt,
+		   bool force_nonblock)
 {
 	struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
 	struct kiocb *kiocb = &req->rw;

@@ -1416,7 +1415,7 @@ static int io_read(struct io_kiocb *req, const struct sqe_submit *s,
 	size_t iov_count;
 	ssize_t read_size, ret;
 
-	ret = io_prep_rw(req, s, force_nonblock);
+	ret = io_prep_rw(req, force_nonblock);
 	if (ret)
 		return ret;
 	file = kiocb->ki_filp;

@@ -1424,7 +1423,7 @@ static int io_read(struct io_kiocb *req, const struct sqe_submit *s,
 	if (unlikely(!(file->f_mode & FMODE_READ)))
 		return -EBADF;
 
-	ret = io_import_iovec(req->ctx, READ, s, &iovec, &iter);
+	ret = io_import_iovec(req->ctx, READ, &req->submit, &iovec, &iter);
 	if (ret < 0)
 		return ret;

@@ -1456,7 +1455,7 @@ static int io_read(struct io_kiocb *req, const struct sqe_submit *s,
 			ret2 = -EAGAIN;
 		/* Catch -EAGAIN return for forced non-blocking submission */
 		if (!force_nonblock || ret2 != -EAGAIN)
-			kiocb_done(kiocb, ret2, nxt, s->in_async);
+			kiocb_done(kiocb, ret2, nxt, req->submit.in_async);
 		else
 			ret = -EAGAIN;
 	}

@@ -1464,8 +1463,8 @@ static int io_read(struct io_kiocb *req, const struct sqe_submit *s,
 	return ret;
 }
 
-static int io_write(struct io_kiocb *req, const struct sqe_submit *s,
-		    struct io_kiocb **nxt, bool force_nonblock)
+static int io_write(struct io_kiocb *req, struct io_kiocb **nxt,
+		    bool force_nonblock)
 {
 	struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
 	struct kiocb *kiocb = &req->rw;

@@ -1474,7 +1473,7 @@ static int io_write(struct io_kiocb *req, const struct sqe_submit *s,
 	size_t iov_count;
 	ssize_t ret;
 
-	ret = io_prep_rw(req, s, force_nonblock);
+	ret = io_prep_rw(req, force_nonblock);
 	if (ret)
 		return ret;

@@ -1482,7 +1481,7 @@ static int io_write(struct io_kiocb *req, const struct sqe_submit *s,
 	if (unlikely(!(file->f_mode & FMODE_WRITE)))
 		return -EBADF;
 
-	ret = io_import_iovec(req->ctx, WRITE, s, &iovec, &iter);
+	ret = io_import_iovec(req->ctx, WRITE, &req->submit, &iovec, &iter);
 	if (ret < 0)
 		return ret;

@@ -1519,7 +1518,7 @@ static int io_write(struct io_kiocb *req, const struct sqe_submit *s,
 		else
 			ret2 = loop_rw_iter(WRITE, file, kiocb, &iter);
 		if (!force_nonblock || ret2 != -EAGAIN)
-			kiocb_done(kiocb, ret2, nxt, s->in_async);
+			kiocb_done(kiocb, ret2, nxt, req->submit.in_async);
 		else
 			ret = -EAGAIN;
 	}

@@ -2188,9 +2187,9 @@ static int io_async_cancel(struct io_kiocb *req, const struct io_uring_sqe *sqe,
 	return 0;
 }
 
-static int io_req_defer(struct io_ring_ctx *ctx, struct io_kiocb *req,
-			const struct io_uring_sqe *sqe)
+static int io_req_defer(struct io_ring_ctx *ctx, struct io_kiocb *req)
 {
+	const struct io_uring_sqe *sqe = req->submit.sqe;
 	struct io_uring_sqe *sqe_copy;
 
 	if (!io_sequence_defer(ctx, req) && list_empty(&ctx->defer_list))

@@ -2217,10 +2216,10 @@ static int io_req_defer(struct io_ring_ctx *ctx, struct io_kiocb *req,
 }
 
 static int __io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
-			   const struct sqe_submit *s, struct io_kiocb **nxt,
-			   bool force_nonblock)
+			   struct io_kiocb **nxt, bool force_nonblock)
 {
 	int ret, opcode;
+	struct sqe_submit *s = &req->submit;
 
 	req->user_data = READ_ONCE(s->sqe->user_data);

@@ -2232,18 +2231,18 @@ static int __io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
 	case IORING_OP_READV:
 		if (unlikely(s->sqe->buf_index))
 			return -EINVAL;
-		ret = io_read(req, s, nxt, force_nonblock);
+		ret = io_read(req, nxt, force_nonblock);
 		break;
 	case IORING_OP_WRITEV:
 		if (unlikely(s->sqe->buf_index))
 			return -EINVAL;
-		ret = io_write(req, s, nxt, force_nonblock);
+		ret = io_write(req, nxt, force_nonblock);
 		break;
 	case IORING_OP_READ_FIXED:
-		ret = io_read(req, s, nxt, force_nonblock);
+		ret = io_read(req, nxt, force_nonblock);
 		break;
 	case IORING_OP_WRITE_FIXED:
-		ret = io_write(req, s, nxt, force_nonblock);
+		ret = io_write(req, nxt, force_nonblock);
 		break;
 	case IORING_OP_FSYNC:
 		ret = io_fsync(req, s->sqe, nxt, force_nonblock);

@@ -2318,7 +2317,7 @@ static void io_wq_submit_work(struct io_wq_work **workptr)
 		s->has_user = (work->flags & IO_WQ_WORK_HAS_MM) != 0;
 		s->in_async = true;
 		do {
-			ret = __io_submit_sqe(ctx, req, s, &nxt, false);
+			ret = __io_submit_sqe(ctx, req, &nxt, false);
 			/*
 			 * We can get EAGAIN for polled IO even though we're
 			 * forcing a sync submission from here, since we can't

@@ -2372,9 +2371,10 @@ static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
 	return table->files[index & IORING_FILE_TABLE_MASK];
 }
 
-static int io_req_set_file(struct io_ring_ctx *ctx, const struct sqe_submit *s,
+static int io_req_set_file(struct io_ring_ctx *ctx,
 			   struct io_submit_state *state, struct io_kiocb *req)
 {
+	struct sqe_submit *s = &req->submit;
 	unsigned flags;
 	int fd;

@@ -2438,12 +2438,11 @@ static int io_grab_files(struct io_ring_ctx *ctx, struct io_kiocb *req)
 	return ret;
 }
 
-static int __io_queue_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
-			struct sqe_submit *s)
+static int __io_queue_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req)
 {
 	int ret;
 
-	ret = __io_submit_sqe(ctx, req, s, NULL, true);
+	ret = __io_submit_sqe(ctx, req, NULL, true);
 
 	/*
 	 * We async punt it if the file wasn't marked NOWAIT, or if the file

@@ -2451,6 +2450,7 @@ static int __io_queue_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
 	 */
 	if (ret == -EAGAIN && (!(req->flags & REQ_F_NOWAIT) ||
 	    (req->flags & REQ_F_MUST_PUNT))) {
+		struct sqe_submit *s = &req->submit;
 		struct io_uring_sqe *sqe_copy;
 
 		sqe_copy = kmemdup(s->sqe, sizeof(*sqe_copy), GFP_KERNEL);

@@ -2488,31 +2488,30 @@ static int __io_queue_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
 	return ret;
 }
 
-static int io_queue_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
-			struct sqe_submit *s)
+static int io_queue_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req)
 {
 	int ret;
 
-	ret = io_req_defer(ctx, req, s->sqe);
+	ret = io_req_defer(ctx, req);
 	if (ret) {
 		if (ret != -EIOCBQUEUED) {
+			io_cqring_add_event(ctx, req->submit.sqe->user_data, ret);
 			io_free_req(req, NULL);
-			io_cqring_add_event(ctx, s->sqe->user_data, ret);
 		}
 		return 0;
 	}
 
-	return __io_queue_sqe(ctx, req, s);
+	return __io_queue_sqe(ctx, req);
 }
 
 static int io_queue_link_head(struct io_ring_ctx *ctx, struct io_kiocb *req,
-			      struct sqe_submit *s, struct io_kiocb *shadow)
+			      struct io_kiocb *shadow)
 {
 	int ret;
 	int need_submit = false;
 
 	if (!shadow)
-		return io_queue_sqe(ctx, req, s);
+		return io_queue_sqe(ctx, req);
 
 	/*
 	 * Mark the first IO in link list as DRAIN, let all the following

@@ -2520,12 +2519,12 @@ static int io_queue_link_head(struct io_ring_ctx *ctx, struct io_kiocb *req,
 	 * list.
 	 */
 	req->flags |= REQ_F_IO_DRAIN;
-	ret = io_req_defer(ctx, req, s->sqe);
+	ret = io_req_defer(ctx, req);
 	if (ret) {
 		if (ret != -EIOCBQUEUED) {
+			io_cqring_add_event(ctx, req->submit.sqe->user_data, ret);
 			io_free_req(req, NULL);
 			__io_free_req(shadow);
-			io_cqring_add_event(ctx, s->sqe->user_data, ret);
 			return 0;
 		}
 	} else {

@@ -2543,7 +2542,7 @@ static int io_queue_link_head(struct io_ring_ctx *ctx, struct io_kiocb *req,
 	spin_unlock_irq(&ctx->completion_lock);
 
 	if (need_submit)
-		return __io_queue_sqe(ctx, req, s);
+		return __io_queue_sqe(ctx, req);
 
 	return 0;
 }

@@ -2551,10 +2550,10 @@ static int io_queue_link_head(struct io_ring_ctx *ctx, struct io_kiocb *req,
 #define SQE_VALID_FLAGS	(IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK)
 
 static void io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
-			  struct sqe_submit *s, struct io_submit_state *state,
-			  struct io_kiocb **link)
+			  struct io_submit_state *state, struct io_kiocb **link)
 {
 	struct io_uring_sqe *sqe_copy;
+	struct sqe_submit *s = &req->submit;
 	int ret;
 
 	/* enforce forwards compatibility on users */

@@ -2563,11 +2562,11 @@ static void io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
 		goto err_req;
 	}
 
-	ret = io_req_set_file(ctx, s, state, req);
+	ret = io_req_set_file(ctx, state, req);
 	if (unlikely(ret)) {
 err_req:
-		io_free_req(req, NULL);
 		io_cqring_add_event(ctx, s->sqe->user_data, ret);
+		io_free_req(req, NULL);
 		return;
 	}

@@ -2598,7 +2597,7 @@ static void io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
 		INIT_LIST_HEAD(&req->link_list);
 		*link = req;
 	} else {
-		io_queue_sqe(ctx, req, s);
+		io_queue_sqe(ctx, req);
 	}
 }

@@ -2739,7 +2738,7 @@ static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
 		req->submit.needs_fixed_file = async;
 		trace_io_uring_submit_sqe(ctx, req->submit.sqe->user_data,
 					  true, async);
-		io_submit_sqe(ctx, req, &req->submit, statep, &link);
+		io_submit_sqe(ctx, req, statep, &link);
 		submitted++;
 
 		/*

@@ -2747,14 +2746,14 @@ static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
 		 * that's the end of the chain. Submit the previous link.
 		 */
 		if (!(sqe_flags & IOSQE_IO_LINK) && link) {
-			io_queue_link_head(ctx, link, &link->submit, shadow_req);
+			io_queue_link_head(ctx, link, shadow_req);
 			link = NULL;
 			shadow_req = NULL;
 		}
 	}
 
 	if (link)
-		io_queue_link_head(ctx, link, &link->submit, shadow_req);
+		io_queue_link_head(ctx, link, shadow_req);
 	if (statep)
 		io_submit_state_end(&state);

Signature · 1 line

2.23.0

Jens Axboe

maintainer

2019-11-06 22:47 UTC #5 lore raw

In reply to Pavel Begunkov

Quoted 11 lines · Pavel Begunkov

The idea is to not pass struct sqe_submit as a separate entity, but always use req->submit instead, so there will be less stuff to care about.
Also, I've got steady +1% throughput improvement for nop tests. Though, it's highly system-dependent, and I wouldn't count on it.
v2: fix use-after-free catched by Jens
v3: -EAGAIN, in case submission failed

Quoted 7 lines

Pavel Begunkov (3):
io_uring: allocate io_kiocb upfront
io_uring: Use submit info inlined into req
io_uring: use inlined struct sqe_submit
  fs/io_uring.c | 134 +++++++++++++++++++++++++-------------------------
  1 file changed, 67 insertions(+), 67 deletions(-)

This looks good to me now, thanks! Applied.

Signature · 1 line

Jens Axboe