Skip to contentSkip to conversation

Add support for ring mapped provided buffers

RFC PATCH Superseded Jens Axboe 2 participants 17 messages Apr 29, 2022 – May 1, 2022

Subsystems (2)

v2 of 6 Superseded Latest is v6 →

lore
Share & tools
lore

Jens Axboe

author maintainer
#0 lore raw

v2 cover Add support for ring mapped provided buffers

Hi,

This series builds to adding support for a different way of doing provided buffers. The interesting bits here are patch 11, which also has some performance numbers an an explanation of it.

Patches 1..5 are cleanups that should just applied separately, I think the clean up the existing code quite nicely.

Patch 6 is a generic optimization for the buffer list lookups.

Patch 7 adds NOP support for provided buffers, just so that we can benchmark the last change.

Patches 8..10 are prep for patch 11.

Patch 11 finally adds the feature.

This passes the full liburing suite - obviously this just means that it didn't break anything existing (that I know of), the only test case for the ring buffers is the nop peak benchmark referenced in patch 11.

v2: - Minor optimizations
- Fix 4k PAGE_SIZE assumption
- Style cleanup

Can also be found in my git repo, for-5.19/io_uring-pbuf branch:

https://git.kernel.dk/cgit/linux-block/log/?h=for-5.19/io_uring-pbuf

 fs/io_uring.c                 | 463 +++++++++++++++++++++++++---------
 include/uapi/linux/io_uring.h |  26 ++
 2 files changed, 370 insertions(+), 119 deletions(-)
Signature · 1 line
Jens Axboe
author maintainer
#1 lore raw

PATCH v2 1/11 io_uring: kill io_recv_buffer_select() wrapper

It's just a thin wrapper around io_buffer_select(), get rid of it.

---
 fs/io_uring.c | 12 ++----------
 1 file changed, 2 insertions(+), 10 deletions(-)
fs/io_uring.c +2 -10
index dfebbf3a272a..12f61ce429dc 100644
@@ -5897,14 +5897,6 @@ static int io_recvmsg_copy_hdr(struct io_kiocb *req,
 	return __io_recvmsg_copy_hdr(req, iomsg);
 }
 
-static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
-					       unsigned int issue_flags)
-{
-	struct io_sr_msg *sr = &req->sr_msg;
-
-	return io_buffer_select(req, &sr->len, sr->bgid, issue_flags);
-}
-
 static int io_recvmsg_prep_async(struct io_kiocb *req)
 {
 	int ret;
@@ -5961,7 +5953,7 @@ static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
 	}
 
 	if (req->flags & REQ_F_BUFFER_SELECT) {
-		kbuf = io_recv_buffer_select(req, issue_flags);
+		kbuf = io_buffer_select(req, &sr->len, sr->bgid, issue_flags);
 		if (IS_ERR(kbuf))
 			return PTR_ERR(kbuf);
 		kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
@@ -6022,7 +6014,7 @@ static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
 		return -ENOTSOCK;
 
 	if (req->flags & REQ_F_BUFFER_SELECT) {
-		kbuf = io_recv_buffer_select(req, issue_flags);
+		kbuf = io_buffer_select(req, &sr->len, sr->bgid, issue_flags);
 		if (IS_ERR(kbuf))
 			return PTR_ERR(kbuf);
 		buf = u64_to_user_ptr(kbuf->addr);
Signature · 1 line
2.35.1
author maintainer
#2 lore raw

PATCH v2 2/11 io_uring: make io_buffer_select() return the user address directly

In reply to Jens Axboe

There's no point in having callers provide a kbuf, we're just returning the address anyway.

---
 fs/io_uring.c | 42 ++++++++++++++++++------------------------
 1 file changed, 18 insertions(+), 24 deletions(-)
fs/io_uring.c +18 -24
index 12f61ce429dc..19dfa974ebcf 100644
@@ -3571,15 +3571,15 @@ static void io_buffer_add_list(struct io_ring_ctx *ctx,
 	list_add(&bl->list, list);
 }
 
-static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
-					  int bgid, unsigned int issue_flags)
+static void __user *io_buffer_select(struct io_kiocb *req, size_t *len,
+				     int bgid, unsigned int issue_flags)
 {
 	struct io_buffer *kbuf = req->kbuf;
 	struct io_ring_ctx *ctx = req->ctx;
 	struct io_buffer_list *bl;
 
 	if (req->flags & REQ_F_BUFFER_SELECTED)
-		return kbuf;
+		return u64_to_user_ptr(kbuf->addr);
 
 	io_ring_submit_lock(req->ctx, issue_flags);
 
@@ -3591,25 +3591,18 @@ static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
 			*len = kbuf->len;
 		req->flags |= REQ_F_BUFFER_SELECTED;
 		req->kbuf = kbuf;
-	} else {
-		kbuf = ERR_PTR(-ENOBUFS);
+		io_ring_submit_unlock(req->ctx, issue_flags);
+		return u64_to_user_ptr(kbuf->addr);
 	}
 
 	io_ring_submit_unlock(req->ctx, issue_flags);
-	return kbuf;
+	return ERR_PTR(-ENOBUFS);
 }
 
 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
 					unsigned int issue_flags)
 {
-	struct io_buffer *kbuf;
-	u16 bgid;
-
-	bgid = req->buf_index;
-	kbuf = io_buffer_select(req, len, bgid, issue_flags);
-	if (IS_ERR(kbuf))
-		return kbuf;
-	return u64_to_user_ptr(kbuf->addr);
+	return io_buffer_select(req, len, req->buf_index, issue_flags);
 }
 
 #ifdef CONFIG_COMPAT
@@ -5934,7 +5927,6 @@ static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
 	struct io_async_msghdr iomsg, *kmsg;
 	struct io_sr_msg *sr = &req->sr_msg;
 	struct socket *sock;
-	struct io_buffer *kbuf;
 	unsigned flags;
 	int ret, min_ret = 0;
 	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
@@ -5953,10 +5945,12 @@ static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
 	}
 
 	if (req->flags & REQ_F_BUFFER_SELECT) {
-		kbuf = io_buffer_select(req, &sr->len, sr->bgid, issue_flags);
-		if (IS_ERR(kbuf))
-			return PTR_ERR(kbuf);
-		kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
+		void __user *buf;
+
+		buf = io_buffer_select(req, &sr->len, sr->bgid, issue_flags);
+		if (IS_ERR(buf))
+			return PTR_ERR(buf);
+		kmsg->fast_iov[0].iov_base = buf;
 		kmsg->fast_iov[0].iov_len = req->sr_msg.len;
 		iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
 				1, req->sr_msg.len);
@@ -5999,7 +5993,6 @@ static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
 
 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
 {
-	struct io_buffer *kbuf;
 	struct io_sr_msg *sr = &req->sr_msg;
 	struct msghdr msg;
 	void __user *buf = sr->buf;
@@ -6014,10 +6007,11 @@ static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
 		return -ENOTSOCK;
 
 	if (req->flags & REQ_F_BUFFER_SELECT) {
-		kbuf = io_buffer_select(req, &sr->len, sr->bgid, issue_flags);
-		if (IS_ERR(kbuf))
-			return PTR_ERR(kbuf);
-		buf = u64_to_user_ptr(kbuf->addr);
+		void __user *buf;
+
+		buf = io_buffer_select(req, &sr->len, sr->bgid, issue_flags);
+		if (IS_ERR(buf))
+			return PTR_ERR(buf);
 	}
 
 	ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
Signature · 1 line
2.35.1
author maintainer
#3 lore raw

PATCH v2 3/11 io_uring: kill io_rw_buffer_select() wrapper

In reply to Jens Axboe

After the recent changes, this is direct call to io_buffer_select() anyway. With this change, there are no wrappers left for provided buffer selection.

---
 fs/io_uring.c | 15 +++++----------
 1 file changed, 5 insertions(+), 10 deletions(-)
fs/io_uring.c +5 -10
index 19dfa974ebcf..cdb23f9861c5 100644
@@ -3599,12 +3599,6 @@ static void __user *io_buffer_select(struct io_kiocb *req, size_t *len,
 	return ERR_PTR(-ENOBUFS);
 }
 
-static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
-					unsigned int issue_flags)
-{
-	return io_buffer_select(req, len, req->buf_index, issue_flags);
-}
-
 #ifdef CONFIG_COMPAT
 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
 				unsigned int issue_flags)
@@ -3612,7 +3606,7 @@ static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
 	struct compat_iovec __user *uiov;
 	compat_ssize_t clen;
 	void __user *buf;
-	ssize_t len;
+	size_t len;
 
 	uiov = u64_to_user_ptr(req->rw.addr);
 	if (!access_ok(uiov, sizeof(*uiov)))
@@ -3623,7 +3617,7 @@ static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
 		return -EINVAL;
 
 	len = clen;
-	buf = io_rw_buffer_select(req, &len, issue_flags);
+	buf = io_buffer_select(req, &len, req->buf_index, issue_flags);
 	if (IS_ERR(buf))
 		return PTR_ERR(buf);
 	iov[0].iov_base = buf;
@@ -3645,7 +3639,7 @@ static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
 	len = iov[0].iov_len;
 	if (len < 0)
 		return -EINVAL;
-	buf = io_rw_buffer_select(req, &len, issue_flags);
+	buf = io_buffer_select(req, &len, req->buf_index, issue_flags);
 	if (IS_ERR(buf))
 		return PTR_ERR(buf);
 	iov[0].iov_base = buf;
@@ -3701,7 +3695,8 @@ static struct iovec *__io_import_iovec(int rw, struct io_kiocb *req,
 
 	if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
 		if (req->flags & REQ_F_BUFFER_SELECT) {
-			buf = io_rw_buffer_select(req, &sqe_len, issue_flags);
+			buf = io_buffer_select(req, &sqe_len, req->buf_index,
+						issue_flags);
 			if (IS_ERR(buf))
 				return ERR_CAST(buf);
 			req->rw.len = sqe_len;
Signature · 1 line
2.35.1
author maintainer
#4 lore raw

PATCH v2 4/11 io_uring: ignore ->buf_index if REQ_F_BUFFER_SELECT isn't set

In reply to Jens Axboe

There's no point in validity checking buf_index if the request doesn't have REQ_F_BUFFER_SELECT set, as we will never use it for that case.

---
 fs/io_uring.c | 4 ----
 1 file changed, 4 deletions(-)
fs/io_uring.c +0 -4
index cdb23f9861c5..a570f47e3f76 100644
@@ -3686,10 +3686,6 @@ static struct iovec *__io_import_iovec(int rw, struct io_kiocb *req,
 		return NULL;
 	}
 
-	/* buffer index only valid with fixed read/write, or buffer select  */
-	if (unlikely(req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT)))
-		return ERR_PTR(-EINVAL);
-
 	buf = u64_to_user_ptr(req->rw.addr);
 	sqe_len = req->rw.len;
 
Signature · 1 line
2.35.1
author maintainer
#5 lore raw

PATCH v2 5/11 io_uring: always use req->buf_index for the provided buffer group

In reply to Jens Axboe

The read/write opcodes use it already, but the recv/recvmsg do not. If we switch them over and read and validate this at init time while we're checking if the opcode supports it anyway, then we can do it in one spot and we don't have to pass in a separate group ID for io_buffer_select().

---
 fs/io_uring.c | 26 +++++++++++++-------------
 1 file changed, 13 insertions(+), 13 deletions(-)
fs/io_uring.c +13 -13
index a570f47e3f76..41205548180d 100644
@@ -644,7 +644,6 @@ struct io_sr_msg {
 		void __user			*buf;
 	};
 	int				msg_flags;
-	int				bgid;
 	size_t				len;
 	size_t				done_io;
 };
@@ -3412,6 +3411,7 @@ static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
 	req->rw.addr = READ_ONCE(sqe->addr);
 	req->rw.len = READ_ONCE(sqe->len);
 	req->rw.flags = READ_ONCE(sqe->rw_flags);
+	/* used for fixed read/write too - just read unconditionally */
 	req->buf_index = READ_ONCE(sqe->buf_index);
 	return 0;
 }
@@ -3572,7 +3572,7 @@ static void io_buffer_add_list(struct io_ring_ctx *ctx,
 }
 
 static void __user *io_buffer_select(struct io_kiocb *req, size_t *len,
-				     int bgid, unsigned int issue_flags)
+				     unsigned int issue_flags)
 {
 	struct io_buffer *kbuf = req->kbuf;
 	struct io_ring_ctx *ctx = req->ctx;
@@ -3583,7 +3583,7 @@ static void __user *io_buffer_select(struct io_kiocb *req, size_t *len,
 
 	io_ring_submit_lock(req->ctx, issue_flags);
 
-	bl = io_buffer_get_list(ctx, bgid);
+	bl = io_buffer_get_list(ctx, req->buf_index);
 	if (bl && !list_empty(&bl->buf_list)) {
 		kbuf = list_first_entry(&bl->buf_list, struct io_buffer, list);
 		list_del(&kbuf->list);
@@ -3617,7 +3617,7 @@ static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
 		return -EINVAL;
 
 	len = clen;
-	buf = io_buffer_select(req, &len, req->buf_index, issue_flags);
+	buf = io_buffer_select(req, &len, issue_flags);
 	if (IS_ERR(buf))
 		return PTR_ERR(buf);
 	iov[0].iov_base = buf;
@@ -3639,7 +3639,7 @@ static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
 	len = iov[0].iov_len;
 	if (len < 0)
 		return -EINVAL;
-	buf = io_buffer_select(req, &len, req->buf_index, issue_flags);
+	buf = io_buffer_select(req, &len, issue_flags);
 	if (IS_ERR(buf))
 		return PTR_ERR(buf);
 	iov[0].iov_base = buf;
@@ -3691,8 +3691,7 @@ static struct iovec *__io_import_iovec(int rw, struct io_kiocb *req,
 
 	if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
 		if (req->flags & REQ_F_BUFFER_SELECT) {
-			buf = io_buffer_select(req, &sqe_len, req->buf_index,
-						issue_flags);
+			buf = io_buffer_select(req, &sqe_len, issue_flags);
 			if (IS_ERR(buf))
 				return ERR_CAST(buf);
 			req->rw.len = sqe_len;
@@ -5900,7 +5899,6 @@ static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
 
 	sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
 	sr->len = READ_ONCE(sqe->len);
-	sr->bgid = READ_ONCE(sqe->buf_group);
 	sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
 	if (sr->msg_flags & MSG_DONTWAIT)
 		req->flags |= REQ_F_NOWAIT;
@@ -5938,7 +5936,7 @@ static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
 	if (req->flags & REQ_F_BUFFER_SELECT) {
 		void __user *buf;
 
-		buf = io_buffer_select(req, &sr->len, sr->bgid, issue_flags);
+		buf = io_buffer_select(req, &sr->len, issue_flags);
 		if (IS_ERR(buf))
 			return PTR_ERR(buf);
 		kmsg->fast_iov[0].iov_base = buf;
@@ -6000,7 +5998,7 @@ static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
 	if (req->flags & REQ_F_BUFFER_SELECT) {
 		void __user *buf;
 
-		buf = io_buffer_select(req, &sr->len, sr->bgid, issue_flags);
+		buf = io_buffer_select(req, &sr->len, issue_flags);
 		if (IS_ERR(buf))
 			return PTR_ERR(buf);
 	}
@@ -8273,9 +8271,11 @@ static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
 		/* enforce forwards compatibility on users */
 		if (sqe_flags & ~SQE_VALID_FLAGS)
 			return -EINVAL;
-		if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
-		    !io_op_defs[opcode].buffer_select)
-			return -EOPNOTSUPP;
+		if (sqe_flags & IOSQE_BUFFER_SELECT) {
+			if (!io_op_defs[opcode].buffer_select)
+				return -EOPNOTSUPP;
+			req->buf_index = READ_ONCE(sqe->buf_group);
+		}
 		if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS)
 			ctx->drain_disabled = true;
 		if (sqe_flags & IOSQE_IO_DRAIN) {
Signature · 1 line
2.35.1
author maintainer
#6 lore raw

PATCH v2 6/11 io_uring: cache last io_buffer_list lookup

In reply to Jens Axboe

Most use cases have 1 or few buffer groups, and even if they have multiple, there's often some locality in looking them up. Add a basic one-hit cache to avoid hashing the group ID and starting the list iteration.

---
 fs/io_uring.c | 20 +++++++++++++++-----
 1 file changed, 15 insertions(+), 5 deletions(-)
fs/io_uring.c +15 -5
index 41205548180d..4bd2f4d868c2 100644
@@ -411,14 +411,16 @@ struct io_ring_ctx {
 		struct io_mapped_ubuf	**user_bufs;
 
 		struct io_submit_state	submit_state;
+		struct list_head	*io_buffers;
+		struct io_buffer_list	*io_bl_last;
+		unsigned int		io_bl_bgid;
+		u32			pers_next;
+		struct list_head	io_buffers_cache;
 		struct list_head	timeout_list;
 		struct list_head	ltimeout_list;
 		struct list_head	cq_overflow_list;
-		struct list_head	*io_buffers;
-		struct list_head	io_buffers_cache;
 		struct list_head	apoll_cache;
 		struct xarray		personalities;
-		u32			pers_next;
 		unsigned		sq_thread_idle;
 	} ____cacheline_aligned_in_smp;
 
@@ -1616,10 +1618,17 @@ static struct io_buffer_list *io_buffer_get_list(struct io_ring_ctx *ctx,
 	struct list_head *hash_list;
 	struct io_buffer_list *bl;
 
+	if (bgid == ctx->io_bl_bgid)
+		return ctx->io_bl_last;
+
 	hash_list = &ctx->io_buffers[hash_32(bgid, IO_BUFFERS_HASH_BITS)];
-	list_for_each_entry(bl, hash_list, list)
-		if (bl->bgid == bgid || bgid == -1U)
+	list_for_each_entry(bl, hash_list, list) {
+		if (bl->bgid == bgid || bgid == -1U) {
+			ctx->io_bl_bgid = bgid;
+			ctx->io_bl_last = bl;
 			return bl;
+		}
+	}
 
 	return NULL;
 }
@@ -1760,6 +1769,7 @@ static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
 		goto err;
 	for (i = 0; i < (1U << IO_BUFFERS_HASH_BITS); i++)
 		INIT_LIST_HEAD(&ctx->io_buffers[i]);
+	ctx->io_bl_bgid = -1U;
 
 	if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
 			    PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
Signature · 1 line
2.35.1
author maintainer
#7 lore raw

PATCH v2 7/11 io_uring: add buffer selection support to IORING_OP_NOP

In reply to Jens Axboe

Obviously not really useful since it's not transferring data, but it is helpful in benchmarking overhead of provided buffers.

---
 fs/io_uring.c | 15 +++++++++++++--
 1 file changed, 13 insertions(+), 2 deletions(-)
fs/io_uring.c +13 -2
index 4bd2f4d868c2..d4004c3a88a1 100644
@@ -1052,7 +1052,9 @@ struct io_op_def {
 };
 
 static const struct io_op_def io_op_defs[] = {
-	[IORING_OP_NOP] = {},
+	[IORING_OP_NOP] = {
+		.buffer_select		= 1,
+	},
 	[IORING_OP_READV] = {
 		.needs_file		= 1,
 		.unbound_nonreg_file	= 1,
@@ -4905,11 +4907,20 @@ static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
 {
 	struct io_ring_ctx *ctx = req->ctx;
+	void __user *buf;
 
 	if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
 		return -EINVAL;
 
-	__io_req_complete(req, issue_flags, 0, 0);
+	if (req->flags & REQ_F_BUFFER_SELECT) {
+		size_t len = 1;
+
+		buf = io_buffer_select(req, &len, issue_flags);
+		if (IS_ERR(buf))
+			return PTR_ERR(buf);
+	}
+
+	__io_req_complete(req, issue_flags, 0, io_put_kbuf(req, issue_flags));
 	return 0;
 }
 
Signature · 1 line
2.35.1
author maintainer
#8 lore raw

PATCH v2 8/11 io_uring: add io_pin_pages() helper

In reply to Jens Axboe

Abstract this out from io_sqe_buffer_register() so we can use it elsewhere too without duplicating this code.

No intended functional changes in this patch.

---
 fs/io_uring.c | 77 +++++++++++++++++++++++++++++++++------------------
 1 file changed, 50 insertions(+), 27 deletions(-)
fs/io_uring.c +50 -27
index d4004c3a88a1..67465cf3700a 100644
@@ -10160,30 +10160,18 @@ static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
 	return ret;
 }
 
-static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
-				  struct io_mapped_ubuf **pimu,
-				  struct page **last_hpage)
+static struct page **io_pin_pages(unsigned long ubuf, unsigned long len,
+				  int *npages)
 {
-	struct io_mapped_ubuf *imu = NULL;
+	unsigned long start, end, nr_pages;
 	struct vm_area_struct **vmas = NULL;
 	struct page **pages = NULL;
-	unsigned long off, start, end, ubuf;
-	size_t size;
-	int ret, pret, nr_pages, i;
-
-	if (!iov->iov_base) {
-		*pimu = ctx->dummy_ubuf;
-		return 0;
-	}
+	int i, pret, ret = -ENOMEM;
 
-	ubuf = (unsigned long) iov->iov_base;
-	end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
+	end = (ubuf + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
 	start = ubuf >> PAGE_SHIFT;
 	nr_pages = end - start;
 
-	*pimu = NULL;
-	ret = -ENOMEM;
-
 	pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
 	if (!pages)
 		goto done;
@@ -10193,10 +10181,6 @@ static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
 	if (!vmas)
 		goto done;
 
-	imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
-	if (!imu)
-		goto done;
-
 	ret = 0;
 	mmap_read_lock(current->mm);
 	pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
@@ -10214,6 +10198,7 @@ static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
 				break;
 			}
 		}
+		*npages = nr_pages;
 	} else {
 		ret = pret < 0 ? pret : -EFAULT;
 	}
@@ -10227,14 +10212,53 @@ static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
 			unpin_user_pages(pages, pret);
 		goto done;
 	}
+	ret = 0;
+done:
+	kvfree(vmas);
+	if (ret < 0) {
+		kvfree(pages);
+		pages = ERR_PTR(ret);
+	}
+	return pages;
+}
 
-	ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
+static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
+				  struct io_mapped_ubuf **pimu,
+				  struct page **last_hpage)
+{
+	struct io_mapped_ubuf *imu = NULL;
+	struct page **pages = NULL;
+	unsigned long off;
+	size_t size;
+	int ret, nr_pages, i;
+
+	if (!iov->iov_base) {
+		*pimu = ctx->dummy_ubuf;
+		return 0;
+	}
+
+	*pimu = NULL;
+	ret = -ENOMEM;
+
+	pages = io_pin_pages((unsigned long) iov->iov_base, iov->iov_len,
+				&nr_pages);
+	if (IS_ERR(pages)) {
+		ret = PTR_ERR(pages);
+		pages = NULL;
+		goto done;
+	}
+
+	imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
+	if (!imu)
+		goto done;
+
+	ret = io_buffer_account_pin(ctx, pages, nr_pages, imu, last_hpage);
 	if (ret) {
-		unpin_user_pages(pages, pret);
+		unpin_user_pages(pages, nr_pages);
 		goto done;
 	}
 
-	off = ubuf & ~PAGE_MASK;
+	off = (unsigned long) iov->iov_base & ~PAGE_MASK;
 	size = iov->iov_len;
 	for (i = 0; i < nr_pages; i++) {
 		size_t vec_len;
@@ -10247,8 +10271,8 @@ static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
 		size -= vec_len;
 	}
 	/* store original address for later verification */
-	imu->ubuf = ubuf;
-	imu->ubuf_end = ubuf + iov->iov_len;
+	imu->ubuf = (unsigned long) iov->iov_base;
+	imu->ubuf_end = imu->ubuf + iov->iov_len;
 	imu->nr_bvecs = nr_pages;
 	*pimu = imu;
 	ret = 0;
@@ -10256,7 +10280,6 @@ static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
 	if (ret)
 		kvfree(imu);
 	kvfree(pages);
-	kvfree(vmas);
 	return ret;
 }
 
Signature · 1 line
2.35.1
author maintainer
#9 lore raw

PATCH v2 9/11 io_uring: abstract out provided buffer list selection

In reply to Jens Axboe

In preparation for providing another way to select a buffer, move the existing logic into a helper.

---
 fs/io_uring.c | 33 +++++++++++++++++++++------------
 1 file changed, 21 insertions(+), 12 deletions(-)
fs/io_uring.c +21 -12
index 67465cf3700a..9e1dd33980d8 100644
@@ -3583,32 +3583,41 @@ static void io_buffer_add_list(struct io_ring_ctx *ctx,
 	list_add(&bl->list, list);
 }
 
+static void __user *io_provided_buffer_select(struct io_kiocb *req, size_t *len,
+					      struct io_buffer_list *bl,
+					      unsigned int issue_flags)
+{
+	struct io_buffer *kbuf;
+
+	kbuf = list_first_entry(&bl->buf_list, struct io_buffer, list);
+	list_del(&kbuf->list);
+	if (*len > kbuf->len)
+		*len = kbuf->len;
+	req->flags |= REQ_F_BUFFER_SELECTED;
+	req->kbuf = kbuf;
+	io_ring_submit_unlock(req->ctx, issue_flags);
+	return u64_to_user_ptr(kbuf->addr);
+}
+
 static void __user *io_buffer_select(struct io_kiocb *req, size_t *len,
 				     unsigned int issue_flags)
 {
-	struct io_buffer *kbuf = req->kbuf;
 	struct io_ring_ctx *ctx = req->ctx;
 	struct io_buffer_list *bl;
 
 	if (req->flags & REQ_F_BUFFER_SELECTED)
-		return u64_to_user_ptr(kbuf->addr);
+		return u64_to_user_ptr(req->kbuf->addr);
 
 	io_ring_submit_lock(req->ctx, issue_flags);
 
 	bl = io_buffer_get_list(ctx, req->buf_index);
-	if (bl && !list_empty(&bl->buf_list)) {
-		kbuf = list_first_entry(&bl->buf_list, struct io_buffer, list);
-		list_del(&kbuf->list);
-		if (*len > kbuf->len)
-			*len = kbuf->len;
-		req->flags |= REQ_F_BUFFER_SELECTED;
-		req->kbuf = kbuf;
+	if (unlikely(!bl)) {
 		io_ring_submit_unlock(req->ctx, issue_flags);
-		return u64_to_user_ptr(kbuf->addr);
+		return ERR_PTR(-ENOBUFS);
 	}
 
-	io_ring_submit_unlock(req->ctx, issue_flags);
-	return ERR_PTR(-ENOBUFS);
+	/* selection helpers drop the submit lock again, if needed */
+	return io_provided_buffer_select(req, len, bl, issue_flags);
 }
 
 #ifdef CONFIG_COMPAT
Signature · 1 line
2.35.1
author maintainer
#10 lore raw

PATCH v2 10/11 io_uring: relocate io_buffer_get_list()

In reply to Jens Axboe

In preparation for calling it from __io_put_kbuf(), move it up a bit to avoid a forward declaration.

---
 fs/io_uring.c | 42 +++++++++++++++++++++---------------------
 1 file changed, 21 insertions(+), 21 deletions(-)
fs/io_uring.c +21 -21
index 9e1dd33980d8..42a952cb073b 100644
@@ -1558,6 +1558,27 @@ static inline void io_req_set_rsrc_node(struct io_kiocb *req,
 	}
 }
 
+static struct io_buffer_list *io_buffer_get_list(struct io_ring_ctx *ctx,
+						 unsigned int bgid)
+{
+	struct list_head *hash_list;
+	struct io_buffer_list *bl;
+
+	if (bgid == ctx->io_bl_bgid)
+		return ctx->io_bl_last;
+
+	hash_list = &ctx->io_buffers[hash_32(bgid, IO_BUFFERS_HASH_BITS)];
+	list_for_each_entry(bl, hash_list, list) {
+		if (bl->bgid == bgid || bgid == -1U) {
+			ctx->io_bl_bgid = bgid;
+			ctx->io_bl_last = bl;
+			return bl;
+		}
+	}
+
+	return NULL;
+}
+
 static unsigned int __io_put_kbuf(struct io_kiocb *req, struct list_head *list)
 {
 	struct io_buffer *kbuf = req->kbuf;
@@ -1614,27 +1635,6 @@ static inline unsigned int io_put_kbuf(struct io_kiocb *req,
 	return cflags;
 }
 
-static struct io_buffer_list *io_buffer_get_list(struct io_ring_ctx *ctx,
-						 unsigned int bgid)
-{
-	struct list_head *hash_list;
-	struct io_buffer_list *bl;
-
-	if (bgid == ctx->io_bl_bgid)
-		return ctx->io_bl_last;
-
-	hash_list = &ctx->io_buffers[hash_32(bgid, IO_BUFFERS_HASH_BITS)];
-	list_for_each_entry(bl, hash_list, list) {
-		if (bl->bgid == bgid || bgid == -1U) {
-			ctx->io_bl_bgid = bgid;
-			ctx->io_bl_last = bl;
-			return bl;
-		}
-	}
-
-	return NULL;
-}
-
 static void io_kbuf_recycle(struct io_kiocb *req, unsigned issue_flags)
 {
 	struct io_ring_ctx *ctx = req->ctx;
Signature · 1 line
2.35.1
author maintainer
#11 lore raw

PATCH v2 11/11 io_uring: add support for ring mapped supplied buffers

In reply to Jens Axboe

Provided buffers allow an application to supply io_uring with buffers that can then be grabbed for a read/receive request, when the data source is ready to deliver data. The existing scheme relies on using IORING_OP_PROVIDE_BUFFERS to do that, but it can be difficult to use in real world applications. It's pretty efficient if the application is able to supply back batches of provided buffers when they have been consumed and the application is ready to recycle them, but if fragmentation occurs in the buffer space, it can become difficult to supply enough buffers at the time. This hurts efficiency.

Add a register op, IORING_REGISTER_PBUF_RING, which allows an application to setup a shared queue for each buffer group of provided buffers. The application can then supply buffers simply by adding them to this ring, and the kernel can consume then just as easily. The ring shares the head with the application, the tail remains private in the kernel.

Provided buffers setup with IORING_REGISTER_PBUF_RING cannot use IORING_OP_{PROVIDE,REMOVE}_BUFFERS for adding or removing entries to the ring, they must use the mapped ring. Mapped provided buffer rings can co-exist with normal provided buffers, just not within the same group ID.

To gauge overhead of the existing scheme and evaluate the mapped ring approach, a simple NOP benchmark was written. It uses a ring of 128 entries, and submits/completes 32 at the time. 'Replenish' is how many buffers are provided back at the time after they have been consumed:

Test Replenish NOPs/sec
================================================================
No provided buffers NA ~30M
Provided buffers 32 ~16M
Provided buffers 1 ~10M
Ring buffers 32 ~27M
Ring buffers 1 ~27M

The ring mapped buffers perform almost as well as not using provided buffers at all, and they don't care if you provided 1 or more back at the same time. This means application can just replenish as they go, rather than need to batch and compact, further reducing overhead in the application. The NOP benchmark above doesn't need to do any compaction, so that overhead isn't even reflected in the above test.

---
 fs/io_uring.c                 | 225 +++++++++++++++++++++++++++++++---
 include/uapi/linux/io_uring.h |  26 ++++
 2 files changed, 236 insertions(+), 15 deletions(-)
fs/io_uring.c +210 -15
index 42a952cb073b..ab789b0d3431 100644
@@ -285,7 +285,16 @@ struct io_rsrc_data {
 struct io_buffer_list {
 	struct list_head list;
 	struct list_head buf_list;
+	struct page **buf_pages;
 	__u16 bgid;
+
+	/* below is for ring provided buffers */
+	__u16 buf_nr_pages;
+	__u16 nr_entries;
+	__u16 buf_per_page;
+	struct io_uring_buf_ring *buf_ring;
+	__u32 tail;
+	__u32 mask;
 };
 
 struct io_buffer {
@@ -815,6 +824,7 @@ enum {
 	REQ_F_NEED_CLEANUP_BIT,
 	REQ_F_POLLED_BIT,
 	REQ_F_BUFFER_SELECTED_BIT,
+	REQ_F_BUFFER_RING_BIT,
 	REQ_F_COMPLETE_INLINE_BIT,
 	REQ_F_REISSUE_BIT,
 	REQ_F_CREDS_BIT,
@@ -865,6 +875,8 @@ enum {
 	REQ_F_POLLED		= BIT(REQ_F_POLLED_BIT),
 	/* buffer already selected */
 	REQ_F_BUFFER_SELECTED	= BIT(REQ_F_BUFFER_SELECTED_BIT),
+	/* buffer selected from ring, needs commit */
+	REQ_F_BUFFER_RING	= BIT(REQ_F_BUFFER_RING_BIT),
 	/* completion is deferred through io_comp_state */
 	REQ_F_COMPLETE_INLINE	= BIT(REQ_F_COMPLETE_INLINE_BIT),
 	/* caller should reissue async */
@@ -995,8 +1007,15 @@ struct io_kiocb {
 	struct async_poll		*apoll;
 	/* opcode allocated if it needs to store data for async defer */
 	void				*async_data;
-	/* stores selected buf, valid IFF REQ_F_BUFFER_SELECTED is set */
-	struct io_buffer		*kbuf;
+	union {
+		/* stores selected buf, valid IFF REQ_F_BUFFER_SELECTED is set */
+		struct io_buffer	*kbuf;
+		/*
+		 * stores buffer ID for ring provided buffers, valid IFF
+		 * REQ_F_BUFFER_RING is set.
+		 */
+		__u32			bid;
+	};
 	/* linked requests, IFF REQ_F_HARDLINK or REQ_F_LINK are set */
 	struct io_kiocb			*link;
 	/* custom credentials, valid IFF REQ_F_CREDS is set */
@@ -1581,21 +1600,33 @@ static struct io_buffer_list *io_buffer_get_list(struct io_ring_ctx *ctx,
 
 static unsigned int __io_put_kbuf(struct io_kiocb *req, struct list_head *list)
 {
-	struct io_buffer *kbuf = req->kbuf;
 	unsigned int cflags;
 
-	cflags = IORING_CQE_F_BUFFER | (kbuf->bid << IORING_CQE_BUFFER_SHIFT);
-	req->flags &= ~REQ_F_BUFFER_SELECTED;
-	list_add(&kbuf->list, list);
-	req->kbuf = NULL;
-	return cflags;
+	if (req->flags & REQ_F_BUFFER_RING) {
+		struct io_buffer_list *bl;
+
+		/* mark as consumed. would be nice to avoid lookup... */
+		bl = io_buffer_get_list(req->ctx, req->buf_index);
+		bl->tail++;
+
+		cflags = req->bid << IORING_CQE_BUFFER_SHIFT;
+		req->flags &= ~REQ_F_BUFFER_RING;
+	} else {
+		struct io_buffer *kbuf = req->kbuf;
+
+		cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
+		list_add(&kbuf->list, list);
+		req->kbuf = NULL;
+		req->flags &= ~REQ_F_BUFFER_SELECTED;
+	}
+	return cflags | IORING_CQE_F_BUFFER;
 }
 
 static inline unsigned int io_put_kbuf_comp(struct io_kiocb *req)
 {
 	lockdep_assert_held(&req->ctx->completion_lock);
 
-	if (likely(!(req->flags & REQ_F_BUFFER_SELECTED)))
+	if (!(req->flags & (REQ_F_BUFFER_SELECTED|REQ_F_BUFFER_RING)))
 		return 0;
 	return __io_put_kbuf(req, &req->ctx->io_buffers_comp);
 }
@@ -1605,7 +1636,7 @@ static inline unsigned int io_put_kbuf(struct io_kiocb *req,
 {
 	unsigned int cflags;
 
-	if (likely(!(req->flags & REQ_F_BUFFER_SELECTED)))
+	if (!(req->flags & (REQ_F_BUFFER_SELECTED|REQ_F_BUFFER_RING)))
 		return 0;
 
 	/*
@@ -1620,7 +1651,10 @@ static inline unsigned int io_put_kbuf(struct io_kiocb *req,
 	 * We migrate buffers from the comp_list to the issue cache list
 	 * when we need one.
 	 */
-	if (issue_flags & IO_URING_F_UNLOCKED) {
+	if (req->flags & REQ_F_BUFFER_RING) {
+		/* no buffers to recycle for this case */
+		cflags = __io_put_kbuf(req, NULL);
+	} else if (issue_flags & IO_URING_F_UNLOCKED) {
 		struct io_ring_ctx *ctx = req->ctx;
 
 		spin_lock(&ctx->completion_lock);
@@ -1641,11 +1675,19 @@ static void io_kbuf_recycle(struct io_kiocb *req, unsigned issue_flags)
 	struct io_buffer_list *bl;
 	struct io_buffer *buf;
 
-	if (likely(!(req->flags & REQ_F_BUFFER_SELECTED)))
+	if (!(req->flags & (REQ_F_BUFFER_SELECTED|REQ_F_BUFFER_RING)))
 		return;
 	/* don't recycle if we already did IO to this buffer */
 	if (req->flags & REQ_F_PARTIAL_IO)
 		return;
+	/*
+	 * We don't need to recycle for REQ_F_BUFFER_RING, we can just clear
+	 * the flag and hence ensure that bl->tail doesn't get incremented.
+	 */
+	if (req->flags & REQ_F_BUFFER_RING) {
+		req->flags &= ~REQ_F_BUFFER_RING;
+		return;
+	}
 
 	io_ring_submit_lock(ctx, issue_flags);
 
@@ -3599,6 +3641,52 @@ static void __user *io_provided_buffer_select(struct io_kiocb *req, size_t *len,
 	return u64_to_user_ptr(kbuf->addr);
 }
 
+static void __user *io_ring_buffer_select(struct io_kiocb *req, size_t *len,
+					  struct io_buffer_list *bl,
+					  unsigned int issue_flags)
+{
+	struct io_uring_buf_ring *br = bl->buf_ring;
+	struct io_uring_buf *buf = &br->bufs[0];
+	__u32 tail = bl->tail;
+
+	if (unlikely(smp_load_acquire(&br->head) == tail))
+		return ERR_PTR(-ENOBUFS);
+
+	tail &= bl->mask;
+	if (tail < bl->buf_per_page) {
+		buf = &br->bufs[tail];
+	} else {
+		int index = tail - bl->buf_per_page;
+		int off = index & bl->buf_per_page;
+
+		index = (index >> (PAGE_SHIFT - 4)) + 1;
+		buf = page_address(bl->buf_pages[index]);
+		buf += off;
+	}
+	if (*len > buf->len)
+		*len = buf->len;
+	req->flags |= REQ_F_BUFFER_RING;
+	req->bid = buf->bid;
+
+	if (!(issue_flags & IO_URING_F_UNLOCKED))
+		return u64_to_user_ptr(buf->addr);
+
+	/*
+	 * If we came in unlocked, we have no choice but to
+	 * consume the buffer here. This does mean it'll be
+	 * pinned until the IO completes. But coming in
+	 * unlocked means we're in io-wq context, hence there
+	 * should be no further retry. For the locked case, the
+	 * caller must ensure to call the commit when the
+	 * transfer completes (or if we get -EAGAIN and must
+	 * poll or retry).
+	 */
+	req->flags &= ~REQ_F_BUFFER_RING;
+	bl->tail++;
+	io_ring_submit_unlock(req->ctx, issue_flags);
+	return u64_to_user_ptr(buf->addr);
+}
+
 static void __user *io_buffer_select(struct io_kiocb *req, size_t *len,
 				     unsigned int issue_flags)
 {
@@ -3617,6 +3705,9 @@ static void __user *io_buffer_select(struct io_kiocb *req, size_t *len,
 	}
 
 	/* selection helpers drop the submit lock again, if needed */
+	if (bl->buf_pages)
+		return io_ring_buffer_select(req, len, bl, issue_flags);
+
 	return io_provided_buffer_select(req, len, bl, issue_flags);
 }
 
@@ -5207,6 +5298,17 @@ static int __io_remove_buffers(struct io_ring_ctx *ctx,
 	if (!nbufs)
 		return 0;
 
+	if (bl->buf_pages) {
+		int j;
+
+		if (WARN_ON_ONCE(nbufs != -1U))
+			return -EINVAL;
+		for (j = 0; j < bl->buf_nr_pages; j++)
+			unpin_user_page(bl->buf_pages[j]);
+		kvfree(bl->buf_pages);
+		bl->buf_pages = NULL;
+	}
+
 	/* the head kbuf is the list itself */
 	while (!list_empty(&bl->buf_list)) {
 		struct io_buffer *nxt;
@@ -5233,8 +5335,12 @@ static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
 
 	ret = -ENOENT;
 	bl = io_buffer_get_list(ctx, p->bgid);
-	if (bl)
-		ret = __io_remove_buffers(ctx, bl, p->nbufs);
+	if (bl) {
+		ret = -EINVAL;
+		/* can't use provide/remove buffers command on mapped buffers */
+		if (!bl->buf_pages)
+			ret = __io_remove_buffers(ctx, bl, p->nbufs);
+	}
 	if (ret < 0)
 		req_set_fail(req);
 
@@ -5359,13 +5465,18 @@ static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
 
 	bl = io_buffer_get_list(ctx, p->bgid);
 	if (unlikely(!bl)) {
-		bl = kmalloc(sizeof(*bl), GFP_KERNEL);
+		bl = kzalloc(sizeof(*bl), GFP_KERNEL);
 		if (!bl) {
 			ret = -ENOMEM;
 			goto err;
 		}
 		io_buffer_add_list(ctx, bl, p->bgid);
 	}
+	/* can't add buffers via this command for a mapped buffer ring */
+	if (bl->buf_pages) {
+		ret = -EINVAL;
+		goto err;
+	}
 
 	ret = io_add_buffers(ctx, p, bl);
 err:
@@ -12308,6 +12419,77 @@ static __cold int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
 	return ret;
 }
 
+static int io_register_pbuf_ring(struct io_ring_ctx *ctx, void __user *arg)
+{
+	struct io_uring_buf_ring *br;
+	struct io_uring_buf_reg reg;
+	struct io_buffer_list *bl;
+	struct page **pages;
+	int nr_pages;
+
+	if (copy_from_user(&reg, arg, sizeof(reg)))
+		return -EFAULT;
+
+	if (reg.resv[0] || reg.resv[1] || reg.resv[2])
+		return -EINVAL;
+	if (!reg.ring_addr)
+		return -EFAULT;
+	if (reg.ring_addr & ~PAGE_MASK)
+		return -EINVAL;
+	if (!is_power_of_2(reg.ring_entries))
+		return -EINVAL;
+
+	bl = io_buffer_get_list(ctx, reg.bgid);
+	if (bl)
+		return -EEXIST;
+	bl = kzalloc(sizeof(*bl), GFP_KERNEL);
+	if (!bl)
+		return -ENOMEM;
+
+	pages = io_pin_pages(reg.ring_addr,
+			     struct_size(br, bufs, reg.ring_entries),
+			     &nr_pages);
+	if (IS_ERR(pages)) {
+		kfree(bl);
+		return PTR_ERR(pages);
+	}
+
+	br = page_address(pages[0]);
+	br->head = 0;
+	bl->buf_pages = pages;
+	bl->buf_nr_pages = nr_pages;
+	bl->nr_entries = reg.ring_entries;
+	BUILD_BUG_ON(sizeof(struct io_uring_buf) != 16);
+	bl->buf_per_page = (PAGE_SIZE - sizeof(struct io_uring_buf)) /
+				sizeof(struct io_uring_buf);
+	bl->buf_ring = br;
+	bl->mask = reg.ring_entries - 1;
+	io_buffer_add_list(ctx, bl, reg.bgid);
+	return 0;
+}
+
+static int io_unregister_pbuf_ring(struct io_ring_ctx *ctx, void __user *arg)
+{
+	struct io_uring_buf_reg reg;
+	struct io_buffer_list *bl;
+
+	if (copy_from_user(&reg, arg, sizeof(reg)))
+		return -EFAULT;
+	if (reg.resv[0] || reg.resv[1] || reg.resv[2])
+		return -EINVAL;
+
+	bl = io_buffer_get_list(ctx, reg.bgid);
+	if (!bl)
+		return -ENOENT;
+	if (!bl->buf_pages)
+		return -EINVAL;
+
+	__io_remove_buffers(ctx, bl, -1U);
+	list_del(&bl->list);
+	kfree(bl);
+	return 0;
+}
+
 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
 			       void __user *arg, unsigned nr_args)
 	__releases(ctx->uring_lock)
@@ -12436,6 +12618,18 @@ static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
 	case IORING_UNREGISTER_RING_FDS:
 		ret = io_ringfd_unregister(ctx, arg, nr_args);
 		break;
+	case IORING_REGISTER_PBUF_RING:
+		ret = -EINVAL;
+		if (!arg || nr_args != 1)
+			break;
+		ret = io_register_pbuf_ring(ctx, arg);
+		break;
+	case IORING_UNREGISTER_PBUF_RING:
+		ret = -EINVAL;
+		if (!arg || nr_args != 1)
+			break;
+		ret = io_unregister_pbuf_ring(ctx, arg);
+		break;
 	default:
 		ret = -EINVAL;
 		break;
@@ -12521,6 +12715,7 @@ static int __init io_uring_init(void)
 
 	/* ->buf_index is u16 */
 	BUILD_BUG_ON(IORING_MAX_REG_BUFFERS >= (1u << 16));
+	BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 16);
 
 	/* should fit into one byte */
 	BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
include/uapi/linux/io_uring.h +26 -0
index 49d1f3994f8d..90b70071110a 100644
@@ -352,6 +352,10 @@ enum {
 	IORING_REGISTER_RING_FDS		= 20,
 	IORING_UNREGISTER_RING_FDS		= 21,
 
+	/* register ring based provide buffer group */
+	IORING_REGISTER_PBUF_RING		= 22,
+	IORING_UNREGISTER_PBUF_RING		= 23,
+
 	/* this goes last */
 	IORING_REGISTER_LAST
 };
@@ -423,6 +427,28 @@ struct io_uring_restriction {
 	__u32 resv2[3];
 };
 
+struct io_uring_buf {
+	__u64	addr;
+	__u32	len;
+	__u32	bid;
+};
+
+struct io_uring_buf_ring {
+	union {
+		__u32			head;
+		struct io_uring_buf	pad;
+	};
+	struct io_uring_buf		bufs[];
+};
+
+/* argument for IORING_(UN)REGISTER_PBUF_RING */
+struct io_uring_buf_reg {
+	__u64	ring_addr;
+	__u32	ring_entries;
+	__u32	bgid;
+	__u64	resv[3];
+};
+
 /*
  * io_uring_restriction->opcode values
  */
Signature · 1 line
2.35.1
#12 lore raw
In reply to Jens Axboe
Quoted 5 lines · Jens Axboe

Hi,

This series builds to adding support for a different way of doing provided buffers. The interesting bits here are patch 11, which also has some performance numbers an an explanation of it.

Jens, would be great if you can CC me for large changes, you know how it's with mailing lists nowadays...

1) reading "io_uring: abstract out provided buffer list selection"

Let's move io_ring_submit_unlock() to where the lock call is. In the end, it's only confusing and duplicates unlock in io_ring_buffer_select() and io_provided_buffer_select().

2) As it's a new API, let's do bucket selection right, I quite don't like io_buffer_get_list(). We can replace "bgid" with indexes into an array and let the userspace to handle indexing. Most likely it knows the index right away or can implement indexes lookup with as many tricks and caching it needs.

Quoted 28 lines

Patches 1..5 are cleanups that should just applied separately, I think the clean up the existing code quite nicely.

Patch 6 is a generic optimization for the buffer list lookups.

Patch 7 adds NOP support for provided buffers, just so that we can benchmark the last change.

Patches 8..10 are prep for patch 11.

Patch 11 finally adds the feature.

This passes the full liburing suite - obviously this just means that it didn't break anything existing (that I know of), the only test case for the ring buffers is the nop peak benchmark referenced in patch 11.

v2: - Minor optimizations
- Fix 4k PAGE_SIZE assumption
- Style cleanup

Can also be found in my git repo, for-5.19/io_uring-pbuf branch:

https://git.kernel.dk/cgit/linux-block/log/?h=for-5.19/io_uring-pbuf

  fs/io_uring.c                 | 463 +++++++++++++++++++++++++---------
  include/uapi/linux/io_uring.h |  26 ++
  2 files changed, 370 insertions(+), 119 deletions(-)
Signature · 1 line
Pavel Begunkov
author maintainer
#13 lore raw
Quoted 9 lines · Pavel Begunkov
Quoted 5 lines · Jens Axboe

Hi,

This series builds to adding support for a different way of doing provided buffers. The interesting bits here are patch 11, which also has some performance numbers an an explanation of it.

Jens, would be great if you can CC me for large changes, you know how it's with mailing lists nowadays...

You bet, I can just add you to anything posted. Starting to lose faith in email ever becoming reliable again...

Quoted 5 lines

1) reading "io_uring: abstract out provided buffer list selection"

Let's move io_ring_submit_unlock() to where the lock call is. In the end, it's only confusing and duplicates unlock in io_ring_buffer_select() and io_provided_buffer_select().

Sure, I can clean that up.

Quoted 5 lines

2) As it's a new API, let's do bucket selection right, I quite don't like io_buffer_get_list(). We can replace "bgid" with indexes into an array and let the userspace to handle indexing. Most likely it knows the index right away or can implement indexes lookup with as many tricks and caching it needs.

Maybe we can just use xarray here rather than a hashed list? It's really just a sparse array. The downside is that xarray locking isn't always very convenient, eg using it with your own locking...

Any other suggestions?

Signature · 1 line
Jens Axboe
#14 lore raw
Quoted 8 lines · Jens Axboe
Quoted 7 lines · Pavel Begunkov
Quoted 5 lines · Jens Axboe

Hi,

This series builds to adding support for a different way of doing provided buffers. The interesting bits here are patch 11, which also has some performance numbers an an explanation of it.

Quoted 5 lines
Quoted 2 lines

Jens, would be great if you can CC me for large changes, you know how it's with mailing lists nowadays...

You bet, I can just add you to anything posted. Starting to lose faith in email ever becoming reliable again...

thanks

Quoted 14 lines
Quoted 5 lines

1) reading "io_uring: abstract out provided buffer list selection"

Let's move io_ring_submit_unlock() to where the lock call is. In the end, it's only confusing and duplicates unlock in io_ring_buffer_select() and io_provided_buffer_select().

Sure, I can clean that up.

Quoted 5 lines

2) As it's a new API, let's do bucket selection right, I quite don't like io_buffer_get_list(). We can replace "bgid" with indexes into an array and let the userspace to handle indexing. Most likely it knows the index right away or can implement indexes lookup with as many tricks and caching it needs.

Quoted 5 lines

Maybe we can just use xarray here rather than a hashed list? It's really just a sparse array. The downside is that xarray locking isn't always very convenient, eg using it with your own locking...

Any other suggestions?

I'd suggest for mapped pbuffers to have an old plain array with sequential indexing, just how we do it for fixed buffers. Do normal and mapped pbuffers share something that would prevent it?

Signature · 1 line
Pavel Begunkov
author maintainer
#15 lore raw
Quoted 3 lines · Pavel Begunkov

I'd suggest for mapped pbuffers to have an old plain array with sequential indexing, just how we do it for fixed buffers. Do normal and mapped pbuffers share something that would prevent it?

Ah yes, we could do that. Registering it returns the group ID instead of providing it up front.

Signature · 1 line
Jens Axboe
author maintainer
#16 lore raw
Quoted 7 lines · Jens Axboe
Quoted 3 lines · Pavel Begunkov

I'd suggest for mapped pbuffers to have an old plain array with sequential indexing, just how we do it for fixed buffers. Do normal and mapped pbuffers share something that would prevent it?

Ah yes, we could do that. Registering it returns the group ID instead of providing it up front.

Actually I'd rather just have the app provide it, but recommendations can be made in terms of using mostly sequential indexes. I suspect that's what most would naturally do anyway.

I'm thinking just straight array of X entries, and then a fallback to xarray if we go beyond that to ensure we don't grow the buffer group array to crazy values.

I'll do this as a prep patch, not really related to the actual change here, but will benefit both the classic and ring buffers alike.

Signature · 1 line
Jens Axboe