Skip to contentSkip to conversation

liburing: support for mix sized sqe's

PATCH Superseded Keith Busch 3 participants 12 messages Oct 13, 2025 – Oct 21, 2025 io_uring

v5 of 6 Superseded Latest is v6 →

lore
Share & tools
lore
#0 lore raw

v5 cover liburing: support for mix sized sqe's

From: Keith Busch <[email protected]>

Changes from v3 and v4:

- Removed the old artifacts leaked from previous versions

- A prep patch introducing a helper for uring_cmd

Keith Busch (4):
liburing: provide uring_cmd prep function
Add support IORING_SETUP_SQE_MIXED
Add nop testing for IORING_SETUP_SQE_MIXED
Add mixed sqe test for uring commands

 src/include/liburing.h          |  82 ++++++++++++++++++-
 src/include/liburing/io_uring.h |   8 ++
 test/Makefile                   |   3 +
 test/io_uring_passthrough.c     |  14 +---
 test/sqe-mixed-bad-wrap.c       |  87 ++++++++++++++++++++
 test/sqe-mixed-nop.c            |  82 +++++++++++++++++++
 test/sqe-mixed-uring_cmd.c      | 140 ++++++++++++++++++++++++++++++++
 7 files changed, 402 insertions(+), 14 deletions(-)
 create mode 100644 test/sqe-mixed-bad-wrap.c
 create mode 100644 test/sqe-mixed-nop.c
 create mode 100644 test/sqe-mixed-uring_cmd.c
Signature · 1 line
2.47.3
#1 lore raw

PATCH v5 1/1 io_uring: add support for IORING_SETUP_SQE_MIXED

From: Keith Busch <[email protected]>

Normal rings support 64b SQEs for posting submissions, while certain features require the ring to be configured with IORING_SETUP_SQE128, as they need to convey more information per submission. This, in turn, makes ALL the SQEs be 128b in size. This is somewhat wasteful and inefficient, particularly when only certain SQEs need to be of the bigger variant.

This adds support for setting up a ring with mixed SQE sizes, using IORING_SETUP_SQE_MIXED. When setup in this mode, SQEs posted to the ring may be either 64b or 128b in size. If a SQE is 128b in size, then opcode will be set to a variante to indicate that this is the case. Any other non-128b opcode will assume the SQ's default size.

SQEs on these types of mixed rings may also utilize NOP with skip success set. This can happen if the ring is one (small) SQE entry away from wrapping, and an attempt is made to get a 128b SQE. As SQEs must be contiguous in the SQ ring, a 128b SQE cannot wrap the ring. For this case, a single NOP SQE should be inserted with the SKIP_SUCCESS flag set. The kernel will process this as a normal NOP and without posting a CQE.

---
 include/uapi/linux/io_uring.h |  8 ++++++++
 io_uring/fdinfo.c             | 34 +++++++++++++++++++++++++++-------
 io_uring/io_uring.c           | 35 +++++++++++++++++++++++++++++++----
 io_uring/io_uring.h           | 14 ++------------
 io_uring/opdef.c              | 26 ++++++++++++++++++++++++++
 io_uring/opdef.h              |  2 ++
 io_uring/register.c           |  2 +-
 io_uring/uring_cmd.c          | 17 +++++++++++++++--
 8 files changed, 112 insertions(+), 26 deletions(-)
include/uapi/linux/io_uring.h +8 -0
index 263bed13473ef..04797a9b76bc2 100644
@@ -231,6 +231,12 @@ enum io_uring_sqe_flags_bit {
  */
 #define IORING_SETUP_CQE_MIXED		(1U << 18)
 
+/*
+ * Allow both 64b and 128b SQEs. If a 128b SQE is posted, it will have
+ * a 128b opcode.
+ */
+#define IORING_SETUP_SQE_MIXED		(1U << 19)
+
 enum io_uring_op {
 	IORING_OP_NOP,
 	IORING_OP_READV,
@@ -295,6 +301,8 @@ enum io_uring_op {
 	IORING_OP_READV_FIXED,
 	IORING_OP_WRITEV_FIXED,
 	IORING_OP_PIPE,
+	IORING_OP_NOP128,
+	IORING_OP_URING_CMD128,
 
 	/* this goes last, obviously */
 	IORING_OP_LAST,
io_uring/fdinfo.c +27 -7
index ff3364531c77b..d14d2e983b623 100644
@@ -14,6 +14,7 @@
 #include "fdinfo.h"
 #include "cancel.h"
 #include "rsrc.h"
+#include "opdef.h"
 
 #ifdef CONFIG_NET_RX_BUSY_POLL
 static __cold void common_tracking_show_fdinfo(struct io_ring_ctx *ctx,
@@ -66,7 +67,6 @@ static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
 	unsigned int cq_head = READ_ONCE(r->cq.head);
 	unsigned int cq_tail = READ_ONCE(r->cq.tail);
 	unsigned int sq_shift = 0;
-	unsigned int sq_entries;
 	int sq_pid = -1, sq_cpu = -1;
 	u64 sq_total_time = 0, sq_work_time = 0;
 	unsigned int i;
@@ -89,26 +89,45 @@ static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
 	seq_printf(m, "CqTail:\t%u\n", cq_tail);
 	seq_printf(m, "CachedCqTail:\t%u\n", data_race(ctx->cached_cq_tail));
 	seq_printf(m, "SQEs:\t%u\n", sq_tail - sq_head);
-	sq_entries = min(sq_tail - sq_head, ctx->sq_entries);
-	for (i = 0; i < sq_entries; i++) {
-		unsigned int entry = i + sq_head;
+	while (sq_head < sq_tail) {
 		struct io_uring_sqe *sqe;
 		unsigned int sq_idx;
+		bool sqe128 = false;
+		u8 opcode;
 
 		if (ctx->flags & IORING_SETUP_NO_SQARRAY)
 			break;
-		sq_idx = READ_ONCE(ctx->sq_array[entry & sq_mask]);
+		sq_idx = READ_ONCE(ctx->sq_array[sq_head & sq_mask]);
 		if (sq_idx > sq_mask)
 			continue;
+
+		opcode = READ_ONCE(sqe->opcode);
 		sqe = &ctx->sq_sqes[sq_idx << sq_shift];
+		if (sq_shift)
+			sqe128 = true;
+		else if (io_issue_defs[opcode].is_128) {
+			if (!(ctx->flags & IORING_SETUP_SQE_MIXED)) {
+				seq_printf(m,
+					"%5u: invalid sqe, 128B entry on non-mixed sq\n",
+					sq_idx);
+				break;
+			}
+			if ((++sq_head & sq_mask) == 0) {
+				seq_printf(m,
+					"%5u: corrupted sqe, wrapping 128B entry\n",
+					sq_idx);
+				break;
+			}
+			sqe128 = true;
+		}
 		seq_printf(m, "%5u: opcode:%s, fd:%d, flags:%x, off:%llu, "
 			      "addr:0x%llx, rw_flags:0x%x, buf_index:%d "
 			      "user_data:%llu",
-			   sq_idx, io_uring_get_opcode(sqe->opcode), sqe->fd,
+			   sq_idx, io_uring_get_opcode(opcode), sqe->fd,
 			   sqe->flags, (unsigned long long) sqe->off,
 			   (unsigned long long) sqe->addr, sqe->rw_flags,
 			   sqe->buf_index, sqe->user_data);
-		if (sq_shift) {
+		if (sqe128) {
 			u64 *sqeb = (void *) (sqe + 1);
 			int size = sizeof(struct io_uring_sqe) / sizeof(u64);
 			int j;
@@ -120,6 +139,7 @@ static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
 			}
 		}
 		seq_printf(m, "\n");
+		sq_head++;
 	}
 	seq_printf(m, "CQEs:\t%u\n", cq_tail - cq_head);
 	while (cq_head < cq_tail) {
io_uring/io_uring.c +31 -4
index 820ef05276667..cd84eb4f2d4ca 100644
@@ -2151,7 +2151,7 @@ static __cold int io_init_fail_req(struct io_kiocb *req, int err)
 }
 
 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
-		       const struct io_uring_sqe *sqe)
+		       const struct io_uring_sqe *sqe, unsigned int *left)
 	__must_hold(&ctx->uring_lock)
 {
 	const struct io_issue_def *def;
@@ -2177,6 +2177,22 @@ static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
 	opcode = array_index_nospec(opcode, IORING_OP_LAST);
 
 	def = &io_issue_defs[opcode];
+	if (def->is_128 && !(ctx->flags & IORING_SETUP_SQE128)) {
+		/*
+		 * A 128b op on a non-128b SQ requires mixed SQE support as
+		 * well as 2 contiguous entries.
+		 */
+		if (!(ctx->flags & IORING_SETUP_SQE_MIXED) || *left < 2 ||
+		    !(ctx->cached_sq_head & (ctx->sq_entries - 1)))
+			return io_init_fail_req(req, -EINVAL);
+		/*
+		 * A 128b operation on a mixed SQ uses two entries, so we have
+		 * to increment the head and decrement what's left.
+		 */
+		ctx->cached_sq_head++;
+		(*left)--;
+	}
+
 	if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) {
 		/* enforce forwards compatibility on users */
 		if (sqe_flags & ~SQE_VALID_FLAGS)
@@ -2286,13 +2302,13 @@ static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe,
 }
 
 static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
-			 const struct io_uring_sqe *sqe)
+			 const struct io_uring_sqe *sqe, unsigned int *left)
 	__must_hold(&ctx->uring_lock)
 {
 	struct io_submit_link *link = &ctx->submit_state.link;
 	int ret;
 
-	ret = io_init_req(ctx, req, sqe);
+	ret = io_init_req(ctx, req, sqe, left);
 	if (unlikely(ret))
 		return io_submit_fail_init(sqe, req, ret);
 
@@ -2444,7 +2460,7 @@ int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
 		 * Continue submitting even for sqe failure if the
 		 * ring was setup with IORING_SETUP_SUBMIT_ALL
 		 */
-		if (unlikely(io_submit_sqe(ctx, req, sqe)) &&
+		if (unlikely(io_submit_sqe(ctx, req, sqe, &left)) &&
 		    !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) {
 			left--;
 			break;
@@ -2789,6 +2805,10 @@ unsigned long rings_size(unsigned int flags, unsigned int sq_entries,
 		if (cq_entries < 2)
 			return SIZE_MAX;
 	}
+	if (flags & IORING_SETUP_SQE_MIXED) {
+		if (sq_entries < 2)
+			return SIZE_MAX;
+	}
 
 #ifdef CONFIG_SMP
 	off = ALIGN(off, SMP_CACHE_BYTES);
@@ -3715,6 +3735,13 @@ static int io_uring_sanitise_params(struct io_uring_params *p)
 	if ((flags & (IORING_SETUP_CQE32|IORING_SETUP_CQE_MIXED)) ==
 	    (IORING_SETUP_CQE32|IORING_SETUP_CQE_MIXED))
 		return -EINVAL;
+	/*
+	 * Nonsensical to ask for SQE128 and mixed SQE support, it's not
+	 * supported to post 64b SQEs on a ring setup with SQE128.
+	 */
+	if ((flags & (IORING_SETUP_SQE128|IORING_SETUP_SQE_MIXED)) ==
+	    (IORING_SETUP_SQE128|IORING_SETUP_SQE_MIXED))
+		return -EINVAL;
 
 	return 0;
 }
io_uring/io_uring.h +2 -12
index 46d9141d772a7..85ed8eb7df80c 100644
@@ -54,7 +54,8 @@
 			IORING_SETUP_REGISTERED_FD_ONLY |\
 			IORING_SETUP_NO_SQARRAY |\
 			IORING_SETUP_HYBRID_IOPOLL |\
-			IORING_SETUP_CQE_MIXED)
+			IORING_SETUP_CQE_MIXED |\
+			IORING_SETUP_SQE_MIXED)
 
 #define IORING_ENTER_FLAGS (IORING_ENTER_GETEVENTS |\
 			IORING_ENTER_SQ_WAKEUP |\
@@ -578,17 +579,6 @@ static inline void io_req_queue_tw_complete(struct io_kiocb *req, s32 res)
 	io_req_task_work_add(req);
 }
 
-/*
- * IORING_SETUP_SQE128 contexts allocate twice the normal SQE size for each
- * slot.
- */
-static inline size_t uring_sqe_size(struct io_ring_ctx *ctx)
-{
-	if (ctx->flags & IORING_SETUP_SQE128)
-		return 2 * sizeof(struct io_uring_sqe);
-	return sizeof(struct io_uring_sqe);
-}
-
 static inline bool io_file_can_poll(struct io_kiocb *req)
 {
 	if (req->flags & REQ_F_CAN_POLL)
io_uring/opdef.c +26 -0
index 932319633eac2..df52d760240e4 100644
@@ -575,6 +575,24 @@ const struct io_issue_def io_issue_defs[] = {
 		.prep			= io_pipe_prep,
 		.issue			= io_pipe,
 	},
+	[IORING_OP_NOP128] = {
+		.audit_skip		= 1,
+		.iopoll			= 1,
+		.is_128			= 1,
+		.prep			= io_nop_prep,
+		.issue			= io_nop,
+	},
+	[IORING_OP_URING_CMD128] = {
+		.buffer_select		= 1,
+		.needs_file		= 1,
+		.plug			= 1,
+		.iopoll			= 1,
+		.iopoll_queue		= 1,
+		.is_128			= 1,
+		.async_size		= sizeof(struct io_async_cmd),
+		.prep			= io_uring_cmd_prep,
+		.issue			= io_uring_cmd,
+	},
 };
 
 const struct io_cold_def io_cold_defs[] = {
@@ -825,6 +843,14 @@ const struct io_cold_def io_cold_defs[] = {
 	[IORING_OP_PIPE] = {
 		.name			= "PIPE",
 	},
+	[IORING_OP_NOP128] = {
+		.name			= "NOP128",
+	},
+	[IORING_OP_URING_CMD128] = {
+		.name			= "URING_CMD128",
+		.sqe_copy		= io_uring_cmd_sqe_copy,
+		.cleanup		= io_uring_cmd_cleanup,
+	},
 };
 
 const char *io_uring_get_opcode(u8 opcode)
io_uring/opdef.h +2 -0
index c2f0907ed78cc..aa37846880ffd 100644
@@ -27,6 +27,8 @@ struct io_issue_def {
 	unsigned		iopoll_queue : 1;
 	/* vectored opcode, set if 1) vectored, and 2) handler needs to know */
 	unsigned		vectored : 1;
+	/* set to 1 if this opcode uses 128b sqes in a mixed sq */
+	unsigned		is_128 : 1;
 
 	/* size of async data needed, if any */
 	unsigned short		async_size;
io_uring/register.c +1 -1
index 43f04c47522c0..e97d9cbba7111 100644
@@ -395,7 +395,7 @@ static void io_register_free_rings(struct io_ring_ctx *ctx,
 #define RESIZE_FLAGS	(IORING_SETUP_CQSIZE | IORING_SETUP_CLAMP)
 #define COPY_FLAGS	(IORING_SETUP_NO_SQARRAY | IORING_SETUP_SQE128 | \
 			 IORING_SETUP_CQE32 | IORING_SETUP_NO_MMAP | \
-			 IORING_SETUP_CQE_MIXED)
+			 IORING_SETUP_CQE_MIXED | IORING_SETUP_SQE_MIXED)
 
 static int io_register_resize_rings(struct io_ring_ctx *ctx, void __user *arg)
 {
io_uring/uring_cmd.c +15 -2
index d1e3ba62ee8e8..a89b29cc5d199 100644
@@ -216,6 +216,18 @@ int io_uring_cmd_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
 	return 0;
 }
 
+/*
+ * IORING_SETUP_SQE128 contexts allocate twice the normal SQE size for each
+ * slot.
+ */
+static inline size_t uring_sqe_size(struct io_kiocb *req)
+{
+	if (req->ctx->flags & IORING_SETUP_SQE128 ||
+	    req->opcode == IORING_OP_URING_CMD128)
+		return 2 * sizeof(struct io_uring_sqe);
+	return sizeof(struct io_uring_sqe);
+}
+
 void io_uring_cmd_sqe_copy(struct io_kiocb *req)
 {
 	struct io_uring_cmd *ioucmd = io_kiocb_to_cmd(req, struct io_uring_cmd);
@@ -224,7 +236,7 @@ void io_uring_cmd_sqe_copy(struct io_kiocb *req)
 	/* Should not happen, as REQ_F_SQE_COPIED covers this */
 	if (WARN_ON_ONCE(ioucmd->sqe == ac->sqes))
 		return;
-	memcpy(ac->sqes, ioucmd->sqe, uring_sqe_size(req->ctx));
+	memcpy(ac->sqes, ioucmd->sqe, uring_sqe_size(req));
 	ioucmd->sqe = ac->sqes;
 }
 
@@ -242,7 +254,8 @@ int io_uring_cmd(struct io_kiocb *req, unsigned int issue_flags)
 	if (ret)
 		return ret;
 
-	if (ctx->flags & IORING_SETUP_SQE128)
+	if (ctx->flags & IORING_SETUP_SQE128 ||
+	    req->opcode == IORING_OP_URING_CMD128)
 		issue_flags |= IO_URING_F_SQE128;
 	if (ctx->flags & (IORING_SETUP_CQE32 | IORING_SETUP_CQE_MIXED))
 		issue_flags |= IO_URING_F_CQE32;
Signature · 1 line
2.47.3
#2 lore raw

Re: [PATCHv5 1/1] io_uring: add support for IORING_SETUP_SQE_MIXED

Quoted 163 lines · Keith Busch

From: Keith Busch <[email protected]>

Normal rings support 64b SQEs for posting submissions, while certain features require the ring to be configured with IORING_SETUP_SQE128, as they need to convey more information per submission. This, in turn, makes ALL the SQEs be 128b in size. This is somewhat wasteful and inefficient, particularly when only certain SQEs need to be of the bigger variant.

This adds support for setting up a ring with mixed SQE sizes, using IORING_SETUP_SQE_MIXED. When setup in this mode, SQEs posted to the ring may be either 64b or 128b in size. If a SQE is 128b in size, then opcode will be set to a variante to indicate that this is the case. Any other non-128b opcode will assume the SQ's default size.

SQEs on these types of mixed rings may also utilize NOP with skip success set. This can happen if the ring is one (small) SQE entry away from wrapping, and an attempt is made to get a 128b SQE. As SQEs must be contiguous in the SQ ring, a 128b SQE cannot wrap the ring. For this case, a single NOP SQE should be inserted with the SKIP_SUCCESS flag set. The kernel will process this as a normal NOP and without posting a CQE.

---
 include/uapi/linux/io_uring.h |  8 ++++++++
 io_uring/fdinfo.c             | 34 +++++++++++++++++++++++++++-------
 io_uring/io_uring.c           | 35 +++++++++++++++++++++++++++++++----
 io_uring/io_uring.h           | 14 ++------------
 io_uring/opdef.c              | 26 ++++++++++++++++++++++++++
 io_uring/opdef.h              |  2 ++
 io_uring/register.c           |  2 +-
 io_uring/uring_cmd.c          | 17 +++++++++++++++--
 8 files changed, 112 insertions(+), 26 deletions(-)
include/uapi/linux/io_uring.h +8 -0
index 263bed13473ef..04797a9b76bc2 100644
@@ -231,6 +231,12 @@ enum io_uring_sqe_flags_bit {
  */
 #define IORING_SETUP_CQE_MIXED         (1U << 18)
 
+/*
+ * Allow both 64b and 128b SQEs. If a 128b SQE is posted, it will have
+ * a 128b opcode.
+ */
+#define IORING_SETUP_SQE_MIXED         (1U << 19)
+
 enum io_uring_op {
        IORING_OP_NOP,
        IORING_OP_READV,
@@ -295,6 +301,8 @@ enum io_uring_op {
        IORING_OP_READV_FIXED,
        IORING_OP_WRITEV_FIXED,
        IORING_OP_PIPE,
+       IORING_OP_NOP128,
+       IORING_OP_URING_CMD128,
 
        /* this goes last, obviously */
        IORING_OP_LAST,
io_uring/fdinfo.c +27 -7
index ff3364531c77b..d14d2e983b623 100644
@@ -14,6 +14,7 @@
 #include "fdinfo.h"
 #include "cancel.h"
 #include "rsrc.h"
+#include "opdef.h"
 
 #ifdef CONFIG_NET_RX_BUSY_POLL
 static __cold void common_tracking_show_fdinfo(struct io_ring_ctx *ctx,
@@ -66,7 +67,6 @@ static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
        unsigned int cq_head = READ_ONCE(r->cq.head);
        unsigned int cq_tail = READ_ONCE(r->cq.tail);
        unsigned int sq_shift = 0;
-       unsigned int sq_entries;
        int sq_pid = -1, sq_cpu = -1;
        u64 sq_total_time = 0, sq_work_time = 0;
        unsigned int i;
@@ -89,26 +89,45 @@ static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
        seq_printf(m, "CqTail:\t%u\n", cq_tail);
        seq_printf(m, "CachedCqTail:\t%u\n", data_race(ctx->cached_cq_tail));
        seq_printf(m, "SQEs:\t%u\n", sq_tail - sq_head);
-       sq_entries = min(sq_tail - sq_head, ctx->sq_entries);
-       for (i = 0; i < sq_entries; i++) {
-               unsigned int entry = i + sq_head;
+       while (sq_head < sq_tail) {
                struct io_uring_sqe *sqe;
                unsigned int sq_idx;
+               bool sqe128 = false;
+               u8 opcode;
 
                if (ctx->flags & IORING_SETUP_NO_SQARRAY)
                        break;
-               sq_idx = READ_ONCE(ctx->sq_array[entry & sq_mask]);
+               sq_idx = READ_ONCE(ctx->sq_array[sq_head & sq_mask]);
                if (sq_idx > sq_mask)
                        continue;
+
+               opcode = READ_ONCE(sqe->opcode);
                sqe = &ctx->sq_sqes[sq_idx << sq_shift];
+               if (sq_shift)
+                       sqe128 = true;
+               else if (io_issue_defs[opcode].is_128) {
+                       if (!(ctx->flags & IORING_SETUP_SQE_MIXED)) {
+                               seq_printf(m,
+                                       "%5u: invalid sqe, 128B entry on non-mixed sq\n",
+                                       sq_idx);
+                               break;
+                       }
+                       if ((++sq_head & sq_mask) == 0) {
+                               seq_printf(m,
+                                       "%5u: corrupted sqe, wrapping 128B entry\n",
+                                       sq_idx);
+                               break;
+                       }
+                       sqe128 = true;
+               }
                seq_printf(m, "%5u: opcode:%s, fd:%d, flags:%x, off:%llu, "
                              "addr:0x%llx, rw_flags:0x%x, buf_index:%d "
                              "user_data:%llu",
-                          sq_idx, io_uring_get_opcode(sqe->opcode), sqe->fd,
+                          sq_idx, io_uring_get_opcode(opcode), sqe->fd,
                           sqe->flags, (unsigned long long) sqe->off,
                           (unsigned long long) sqe->addr, sqe->rw_flags,
                           sqe->buf_index, sqe->user_data);
-               if (sq_shift) {
+               if (sqe128) {
                        u64 *sqeb = (void *) (sqe + 1);
                        int size = sizeof(struct io_uring_sqe) / sizeof(u64);
                        int j;
@@ -120,6 +139,7 @@ static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
                        }
                }
                seq_printf(m, "\n");
+               sq_head++;
        }
        seq_printf(m, "CQEs:\t%u\n", cq_tail - cq_head);
        while (cq_head < cq_tail) {
io_uring/io_uring.c +5 -1
index 820ef05276667..cd84eb4f2d4ca 100644
@@ -2151,7 +2151,7 @@ static __cold int io_init_fail_req(struct io_kiocb *req, int err)
 }
 
 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
-                      const struct io_uring_sqe *sqe)
+                      const struct io_uring_sqe *sqe, unsigned int *left)
        __must_hold(&ctx->uring_lock)
 {
        const struct io_issue_def *def;
@@ -2177,6 +2177,22 @@ static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
        opcode = array_index_nospec(opcode, IORING_OP_LAST);
 
        def = &io_issue_defs[opcode];
+       if (def->is_128 && !(ctx->flags & IORING_SETUP_SQE128)) {
+               /*
+                * A 128b op on a non-128b SQ requires mixed SQE support as
+                * well as 2 contiguous entries.

io_uring/io_uring.c static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req, · line 2193

Quoted 10 lines
io_uring/io_uring.c +10 -0
+                */
+               if (!(ctx->flags & IORING_SETUP_SQE_MIXED) || *left < 2 ||
+                   !(ctx->cached_sq_head & (ctx->sq_entries - 1)))
+                       return io_init_fail_req(req, -EINVAL);
+               /*
+                * A 128b operation on a mixed SQ uses two entries, so we have
+                * to increment the head and decrement what's left.
+                */
+               ctx->cached_sq_head++;
+               (*left)--;

Hmm, io_submit_sqes() calls io_get_task_refs() at the start to decrement cached_refs by the number of SQEs (counting 128-byte SQEs twice) but io_put_task() only increments it once for each completed request (counting 128-byte SQEs once). Does that mean there's a refcount leak? Perhaps io_submit_sqes() or this block needs to increment cached_refs to account for each 128-byte SQE?

Otherwise, this looks good to me.

Best,
Caleb

Quoted 202 lines
io_uring/io_uring.c +16 -3
+       }
+
        if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) {
                /* enforce forwards compatibility on users */
                if (sqe_flags & ~SQE_VALID_FLAGS)
@@ -2286,13 +2302,13 @@ static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe,
 }
 
 static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
-                        const struct io_uring_sqe *sqe)
+                        const struct io_uring_sqe *sqe, unsigned int *left)
        __must_hold(&ctx->uring_lock)
 {
        struct io_submit_link *link = &ctx->submit_state.link;
        int ret;
 
-       ret = io_init_req(ctx, req, sqe);
+       ret = io_init_req(ctx, req, sqe, left);
        if (unlikely(ret))
                return io_submit_fail_init(sqe, req, ret);
 
@@ -2444,7 +2460,7 @@ int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
                 * Continue submitting even for sqe failure if the
                 * ring was setup with IORING_SETUP_SUBMIT_ALL
                 */
-               if (unlikely(io_submit_sqe(ctx, req, sqe)) &&
+               if (unlikely(io_submit_sqe(ctx, req, sqe, &left)) &&
                    !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) {
                        left--;
                        break;
@@ -2789,6 +2805,10 @@ unsigned long rings_size(unsigned int flags, unsigned int sq_entries,
                if (cq_entries < 2)
                        return SIZE_MAX;
        }
+       if (flags & IORING_SETUP_SQE_MIXED) {
+               if (sq_entries < 2)
+                       return SIZE_MAX;
+       }
 
 #ifdef CONFIG_SMP
        off = ALIGN(off, SMP_CACHE_BYTES);
@@ -3715,6 +3735,13 @@ static int io_uring_sanitise_params(struct io_uring_params *p)
        if ((flags & (IORING_SETUP_CQE32|IORING_SETUP_CQE_MIXED)) ==
            (IORING_SETUP_CQE32|IORING_SETUP_CQE_MIXED))
                return -EINVAL;
+       /*
+        * Nonsensical to ask for SQE128 and mixed SQE support, it's not
+        * supported to post 64b SQEs on a ring setup with SQE128.
+        */
+       if ((flags & (IORING_SETUP_SQE128|IORING_SETUP_SQE_MIXED)) ==
+           (IORING_SETUP_SQE128|IORING_SETUP_SQE_MIXED))
+               return -EINVAL;
 
        return 0;
 }
io_uring/io_uring.h +2 -12
index 46d9141d772a7..85ed8eb7df80c 100644
@@ -54,7 +54,8 @@
                        IORING_SETUP_REGISTERED_FD_ONLY |\
                        IORING_SETUP_NO_SQARRAY |\
                        IORING_SETUP_HYBRID_IOPOLL |\
-                       IORING_SETUP_CQE_MIXED)
+                       IORING_SETUP_CQE_MIXED |\
+                       IORING_SETUP_SQE_MIXED)
 
 #define IORING_ENTER_FLAGS (IORING_ENTER_GETEVENTS |\
                        IORING_ENTER_SQ_WAKEUP |\
@@ -578,17 +579,6 @@ static inline void io_req_queue_tw_complete(struct io_kiocb *req, s32 res)
        io_req_task_work_add(req);
 }
 
-/*
- * IORING_SETUP_SQE128 contexts allocate twice the normal SQE size for each
- * slot.
- */
-static inline size_t uring_sqe_size(struct io_ring_ctx *ctx)
-{
-       if (ctx->flags & IORING_SETUP_SQE128)
-               return 2 * sizeof(struct io_uring_sqe);
-       return sizeof(struct io_uring_sqe);
-}
-
 static inline bool io_file_can_poll(struct io_kiocb *req)
 {
        if (req->flags & REQ_F_CAN_POLL)
io_uring/opdef.c +26 -0
index 932319633eac2..df52d760240e4 100644
@@ -575,6 +575,24 @@ const struct io_issue_def io_issue_defs[] = {
                .prep                   = io_pipe_prep,
                .issue                  = io_pipe,
        },
+       [IORING_OP_NOP128] = {
+               .audit_skip             = 1,
+               .iopoll                 = 1,
+               .is_128                 = 1,
+               .prep                   = io_nop_prep,
+               .issue                  = io_nop,
+       },
+       [IORING_OP_URING_CMD128] = {
+               .buffer_select          = 1,
+               .needs_file             = 1,
+               .plug                   = 1,
+               .iopoll                 = 1,
+               .iopoll_queue           = 1,
+               .is_128                 = 1,
+               .async_size             = sizeof(struct io_async_cmd),
+               .prep                   = io_uring_cmd_prep,
+               .issue                  = io_uring_cmd,
+       },
 };
 
 const struct io_cold_def io_cold_defs[] = {
@@ -825,6 +843,14 @@ const struct io_cold_def io_cold_defs[] = {
        [IORING_OP_PIPE] = {
                .name                   = "PIPE",
        },
+       [IORING_OP_NOP128] = {
+               .name                   = "NOP128",
+       },
+       [IORING_OP_URING_CMD128] = {
+               .name                   = "URING_CMD128",
+               .sqe_copy               = io_uring_cmd_sqe_copy,
+               .cleanup                = io_uring_cmd_cleanup,
+       },
 };
 
 const char *io_uring_get_opcode(u8 opcode)
io_uring/opdef.h +2 -0
index c2f0907ed78cc..aa37846880ffd 100644
@@ -27,6 +27,8 @@ struct io_issue_def {
        unsigned                iopoll_queue : 1;
        /* vectored opcode, set if 1) vectored, and 2) handler needs to know */
        unsigned                vectored : 1;
+       /* set to 1 if this opcode uses 128b sqes in a mixed sq */
+       unsigned                is_128 : 1;
 
        /* size of async data needed, if any */
        unsigned short          async_size;
io_uring/register.c +1 -1
index 43f04c47522c0..e97d9cbba7111 100644
@@ -395,7 +395,7 @@ static void io_register_free_rings(struct io_ring_ctx *ctx,
 #define RESIZE_FLAGS   (IORING_SETUP_CQSIZE | IORING_SETUP_CLAMP)
 #define COPY_FLAGS     (IORING_SETUP_NO_SQARRAY | IORING_SETUP_SQE128 | \
                         IORING_SETUP_CQE32 | IORING_SETUP_NO_MMAP | \
-                        IORING_SETUP_CQE_MIXED)
+                        IORING_SETUP_CQE_MIXED | IORING_SETUP_SQE_MIXED)
 
 static int io_register_resize_rings(struct io_ring_ctx *ctx, void __user *arg)
 {
io_uring/uring_cmd.c +15 -3
index d1e3ba62ee8e8..a89b29cc5d199 100644
@@ -216,6 +216,18 @@ int io_uring_cmd_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
        return 0;
 }
 
+/*
+ * IORING_SETUP_SQE128 contexts allocate twice the normal SQE size for each
+ * slot.
+ */
+static inline size_t uring_sqe_size(struct io_kiocb *req)
+{
+       if (req->ctx->flags & IORING_SETUP_SQE128 ||
+           req->opcode == IORING_OP_URING_CMD128)
+               return 2 * sizeof(struct io_uring_sqe);
+       return sizeof(struct io_uring_sqe);
+}
+
 void io_uring_cmd_sqe_copy(struct io_kiocb *req)
 {
        struct io_uring_cmd *ioucmd = io_kiocb_to_cmd(req, struct io_uring_cmd);
@@ -224,7 +236,7 @@ void io_uring_cmd_sqe_copy(struct io_kiocb *req)
        /* Should not happen, as REQ_F_SQE_COPIED covers this */
        if (WARN_ON_ONCE(ioucmd->sqe == ac->sqes))
                return;
-       memcpy(ac->sqes, ioucmd->sqe, uring_sqe_size(req->ctx));
+       memcpy(ac->sqes, ioucmd->sqe, uring_sqe_size(req));
        ioucmd->sqe = ac->sqes;
 }
 
@@ -242,7 +254,8 @@ int io_uring_cmd(struct io_kiocb *req, unsigned int issue_flags)
        if (ret)
                return ret;
 
-       if (ctx->flags & IORING_SETUP_SQE128)
+       if (ctx->flags & IORING_SETUP_SQE128 ||
+           req->opcode == IORING_OP_URING_CMD128)
                issue_flags |= IO_URING_F_SQE128;
        if (ctx->flags & (IORING_SETUP_CQE32 | IORING_SETUP_CQE_MIXED))
                issue_flags |= IO_URING_F_CQE32;
--

2.47.3

#3 lore raw

Re: [PATCHv5 1/1] io_uring: add support for IORING_SETUP_SQE_MIXED

Quoted 15 lines · Caleb Sander Mateos

io_uring/io_uring.c static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req, · line 2193

Quoted 13 lines · Keith Busch
io_uring/io_uring.c +13 -0
+               /*
+                * A 128b op on a non-128b SQ requires mixed SQE support as
+                * well as 2 contiguous entries.
+                */
+               if (!(ctx->flags & IORING_SETUP_SQE_MIXED) || *left < 2 ||
+                   !(ctx->cached_sq_head & (ctx->sq_entries - 1)))
+                       return io_init_fail_req(req, -EINVAL);
+               /*
+                * A 128b operation on a mixed SQ uses two entries, so we have
+                * to increment the head and decrement what's left.
+                */
+               ctx->cached_sq_head++;
+               (*left)--;
Quoted 6 lines

Hmm, io_submit_sqes() calls io_get_task_refs() at the start to decrement cached_refs by the number of SQEs (counting 128-byte SQEs twice) but io_put_task() only increments it once for each completed request (counting 128-byte SQEs once). Does that mean there's a refcount leak? Perhaps io_submit_sqes() or this block needs to increment cached_refs to account for each 128-byte SQE?

It looks like you're right. I think the increment option is the easiest way to deal with it, just adding this line into the above:

+ current->io_uring->cached_refs++;

I'm going to take a moment to figure out a good way to test this because I don't think I'm hitting any problem with the admittedly simple tests I've offered to liburing, so I may be missing something.

#4 lore raw

Re: [PATCHv5 1/1] io_uring: add support for IORING_SETUP_SQE_MIXED

Quoted 28 lines · Keith Busch
Quoted 15 lines · Caleb Sander Mateos

io_uring/io_uring.c static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req, · line 2193

Quoted 13 lines · Keith Busch
io_uring/io_uring.c +13 -0
+               /*
+                * A 128b op on a non-128b SQ requires mixed SQE support as
+                * well as 2 contiguous entries.
+                */
+               if (!(ctx->flags & IORING_SETUP_SQE_MIXED) || *left < 2 ||
+                   !(ctx->cached_sq_head & (ctx->sq_entries - 1)))
+                       return io_init_fail_req(req, -EINVAL);
+               /*
+                * A 128b operation on a mixed SQ uses two entries, so we have
+                * to increment the head and decrement what's left.
+                */
+               ctx->cached_sq_head++;
+               (*left)--;
Quoted 6 lines

Hmm, io_submit_sqes() calls io_get_task_refs() at the start to decrement cached_refs by the number of SQEs (counting 128-byte SQEs twice) but io_put_task() only increments it once for each completed request (counting 128-byte SQEs once). Does that mean there's a refcount leak? Perhaps io_submit_sqes() or this block needs to increment cached_refs to account for each 128-byte SQE?

It looks like you're right. I think the increment option is the easiest way to deal with it, just adding this line into the above:

(quoted diff) +1 -0
+ 	current->io_uring->cached_refs++;
Quoted 3 lines

I'm going to take a moment to figure out a good way to test this because I don't think I'm hitting any problem with the admittedly simple tests I've offered to liburing, so I may be missing something.

So the tests were in fact causing missing putting some usage references, but I'm not sure how to check for such leakage. Everything ends up clearing up once the ring closes, and there's no other visibility into the refcount from user space. I had to add some trace_printks just to verify it. The increment above gets everything back to normal, at least, so will send a new version with that in.

#5 lore raw

PATCH v5 1/4 liburing: provide uring_cmd prep function

In reply to Keith Busch

From: Keith Busch <[email protected]>

The rw prep doesn't clear __pad1, which is a reserved field for uring_cmd. If a prior submission in that entry did use that field, the uring_cmd will fail the kernel's checks.

Also, the nvme uring_cmd tests had a couple places setting the sqe addr and length, which are unused fields for the nvme uring_cmds, so they shouldn't have been doing that, though had been checking these, so it didn't cause any errors.

Provide a helper function specific to the uring_cmd preparation.

---
 src/include/liburing.h      | 19 +++++++++++++++----
 test/io_uring_passthrough.c | 14 ++++----------
 2 files changed, 19 insertions(+), 14 deletions(-)
src/include/liburing.h +15 -4
index c80bffd3..f7af20aa 100644
@@ -1517,6 +1517,19 @@ IOURINGINLINE void io_uring_prep_socket_direct_alloc(struct io_uring_sqe *sqe,
 	__io_uring_set_target_fixed_file(sqe, IORING_FILE_INDEX_ALLOC - 1);
 }
 
+IOURINGINLINE void io_uring_prep_uring_cmd(struct io_uring_sqe *sqe,
+					   int cmd_op,
+					   int fd)
+	LIBURING_NOEXCEPT
+{
+	sqe->opcode = (__u8) IORING_OP_URING_CMD;
+	sqe->fd = fd;
+	sqe->cmd_op = cmd_op;
+	sqe->__pad1 = 0;
+	sqe->addr = 0ul;
+	sqe->len = 0;
+}
+
 /*
  * Prepare commands for sockets
  */
@@ -1529,11 +1542,10 @@ IOURINGINLINE void io_uring_prep_cmd_sock(struct io_uring_sqe *sqe,
 					  int optlen)
 	LIBURING_NOEXCEPT
 {
-	io_uring_prep_rw(IORING_OP_URING_CMD, sqe, fd, NULL, 0, 0);
+	io_uring_prep_uring_cmd(sqe, cmd_op, fd);
 	sqe->optval = (unsigned long) (uintptr_t) optval;
 	sqe->optname = optname;
 	sqe->optlen = optlen;
-	sqe->cmd_op = cmd_op;
 	sqe->level = level;
 }
 
@@ -1607,8 +1619,7 @@ IOURINGINLINE void io_uring_prep_cmd_discard(struct io_uring_sqe *sqe,
 					     uint64_t offset, uint64_t nbytes)
 	LIBURING_NOEXCEPT
 {
-	io_uring_prep_rw(IORING_OP_URING_CMD, sqe, fd, 0, 0, 0);
-	sqe->cmd_op = BLOCK_URING_CMD_DISCARD;
+	io_uring_prep_uring_cmd(sqe, BLOCK_URING_CMD_DISCARD, fd);
 	sqe->addr = offset;
 	sqe->addr3 = nbytes;
 }
test/io_uring_passthrough.c +4 -10
index beaa81ad..26051710 100644
@@ -148,11 +148,9 @@ static int __test_io(const char *file, struct io_uring *ring, int tc, int read,
 		if (async)
 			sqe->flags |= IOSQE_ASYNC;
 		if (nonvec)
-			sqe->cmd_op = NVME_URING_CMD_IO;
+			io_uring_prep_uring_cmd(sqe, NVME_URING_CMD_IO, use_fd);
 		else
-			sqe->cmd_op = NVME_URING_CMD_IO_VEC;
-		sqe->fd = use_fd;
-		sqe->opcode = IORING_OP_URING_CMD;
+			io_uring_prep_uring_cmd(sqe, NVME_URING_CMD_IO_VEC, use_fd);
 		if (do_fixed)
 			sqe->uring_cmd_flags |= IORING_URING_CMD_FIXED;
 		sqe->user_data = ((uint64_t)offset << 32) | i;
@@ -328,9 +326,7 @@ static int test_invalid_passthru_submit(const char *file)
 	}
 
 	sqe = io_uring_get_sqe(&ring);
-	io_uring_prep_read(sqe, fd, vecs[0].iov_base, vecs[0].iov_len, 0);
-	sqe->cmd_op = NVME_URING_CMD_IO;
-	sqe->opcode = IORING_OP_URING_CMD;
+	io_uring_prep_uring_cmd(sqe, NVME_URING_CMD_IO, fd);
 	sqe->user_data = 1;
 	cmd = (struct nvme_uring_cmd *)sqe->cmd;
 	memset(cmd, 0, sizeof(struct nvme_uring_cmd));
@@ -401,10 +397,8 @@ static int test_io_uring_submit_enters(const char *file)
 		__u32 nlb;
 
 		sqe = io_uring_get_sqe(&ring);
-		io_uring_prep_readv(sqe, fd, &vecs[i], 1, offset);
+		io_uring_prep_uring_cmd(sqe, NVME_URING_CMD_IO, fd);
 		sqe->user_data = i;
-		sqe->opcode = IORING_OP_URING_CMD;
-		sqe->cmd_op = NVME_URING_CMD_IO;
 		cmd = (struct nvme_uring_cmd *)sqe->cmd;
 		memset(cmd, 0, sizeof(struct nvme_uring_cmd));
 
Signature · 1 line
2.47.3
#6 lore raw

Re: [PATCHv5 1/4] liburing: provide uring_cmd prep function

Looks good to me, just a few minor comments.

Quoted 4 lines · Keith Busch

From: Keith Busch <[email protected]>

The rw prep doesn't clear __pad1, which is a reserved field for

io_uring_prep_rw() does assign to sqe->off, which is unioned with cmd_op and __pad1. Though obviously __pad1 being set to 0 is dependent on a offset being passed as 0 to io_uring_prep_rw(). But I certainly agree a dedicated helper for IORING_OP_URING_CMD is a great improvement.

Quoted 6 lines

uring_cmd. If a prior submission in that entry did use that field, the uring_cmd will fail the kernel's checks.

Also, the nvme uring_cmd tests had a couple places setting the sqe addr and length, which are unused fields for the nvme uring_cmds, so they shouldn't have been doing that, though had been checking these, so it

"had" -> "hadn't"?

Quoted 15 lines

didn't cause any errors.

Provide a helper function specific to the uring_cmd preparation.

---
 src/include/liburing.h      | 19 +++++++++++++++----
 test/io_uring_passthrough.c | 14 ++++----------
 2 files changed, 19 insertions(+), 14 deletions(-)
src/include/liburing.h +0 -0
index c80bffd3..f7af20aa 100644

src/include/liburing.h IOURINGINLINE void io_uring_prep_socket_direct_alloc(struct io_uring_sqe *sqe, · line 1521

Quoted 5 lines
src/include/liburing.h +2 -0
@@ -1517,6 +1517,19 @@ IOURINGINLINE void io_uring_prep_socket_direct_alloc(struct io_uring_sqe *sqe,
        __io_uring_set_target_fixed_file(sqe, IORING_FILE_INDEX_ALLOC - 1);
 }
 
+IOURINGINLINE void io_uring_prep_uring_cmd(struct io_uring_sqe *sqe,
+                                          int cmd_op,

I see this is copied from io_uring_prep_cmd_sock(), but u32 is probably more accurate.

src/include/liburing.h IOURINGINLINE void io_uring_prep_socket_direct_alloc(struct io_uring_sqe *sqe, · line 1525

Quoted 4 lines
src/include/liburing.h +4 -0
+                                          int fd)
+       LIBURING_NOEXCEPT
+{
+       sqe->opcode = (__u8) IORING_OP_URING_CMD;

Casting the constant seems unnecessary. Do compilers really warn about this?

Quoted 38 lines
src/include/liburing.h +9 -4
+       sqe->fd = fd;
+       sqe->cmd_op = cmd_op;
+       sqe->__pad1 = 0;
+       sqe->addr = 0ul;
+       sqe->len = 0;
+}
+
 /*
  * Prepare commands for sockets
  */
@@ -1529,11 +1542,10 @@ IOURINGINLINE void io_uring_prep_cmd_sock(struct io_uring_sqe *sqe,
                                          int optlen)
        LIBURING_NOEXCEPT
 {
-       io_uring_prep_rw(IORING_OP_URING_CMD, sqe, fd, NULL, 0, 0);
+       io_uring_prep_uring_cmd(sqe, cmd_op, fd);
        sqe->optval = (unsigned long) (uintptr_t) optval;
        sqe->optname = optname;
        sqe->optlen = optlen;
-       sqe->cmd_op = cmd_op;
        sqe->level = level;
 }
 
@@ -1607,8 +1619,7 @@ IOURINGINLINE void io_uring_prep_cmd_discard(struct io_uring_sqe *sqe,
                                             uint64_t offset, uint64_t nbytes)
        LIBURING_NOEXCEPT
 {
-       io_uring_prep_rw(IORING_OP_URING_CMD, sqe, fd, 0, 0, 0);
-       sqe->cmd_op = BLOCK_URING_CMD_DISCARD;
+       io_uring_prep_uring_cmd(sqe, BLOCK_URING_CMD_DISCARD, fd);
        sqe->addr = offset;
        sqe->addr3 = nbytes;
 }
test/io_uring_passthrough.c +0 -0
index beaa81ad..26051710 100644

test/io_uring_passthrough.c static int __test_io(const char *file, struct io_uring *ring, int tc, int read, · line 151

Quoted 5 lines
test/io_uring_passthrough.c +1 -1
@@ -148,11 +148,9 @@ static int __test_io(const char *file, struct io_uring *ring, int tc, int read,
                if (async)
                        sqe->flags |= IOSQE_ASYNC;
                if (nonvec)
-                       sqe->cmd_op = NVME_URING_CMD_IO;
+                       io_uring_prep_uring_cmd(sqe, NVME_URING_CMD_IO, use_fd);

I guess this works because io_uring_prep_uring_cmd() doesn't touch sqe->buf_index or sqe->flags, but it seems like it would be less brittle to call io_uring_prep_uring_cmd() before setting any of the other sqe fields.

Quoted 34 lines
test/io_uring_passthrough.c +3 -10
                else
-                       sqe->cmd_op = NVME_URING_CMD_IO_VEC;
-               sqe->fd = use_fd;
-               sqe->opcode = IORING_OP_URING_CMD;
+                       io_uring_prep_uring_cmd(sqe, NVME_URING_CMD_IO_VEC, use_fd);
                if (do_fixed)
                        sqe->uring_cmd_flags |= IORING_URING_CMD_FIXED;
                sqe->user_data = ((uint64_t)offset << 32) | i;
@@ -328,9 +326,7 @@ static int test_invalid_passthru_submit(const char *file)
        }
 
        sqe = io_uring_get_sqe(&ring);
-       io_uring_prep_read(sqe, fd, vecs[0].iov_base, vecs[0].iov_len, 0);
-       sqe->cmd_op = NVME_URING_CMD_IO;
-       sqe->opcode = IORING_OP_URING_CMD;
+       io_uring_prep_uring_cmd(sqe, NVME_URING_CMD_IO, fd);
        sqe->user_data = 1;
        cmd = (struct nvme_uring_cmd *)sqe->cmd;
        memset(cmd, 0, sizeof(struct nvme_uring_cmd));
@@ -401,10 +397,8 @@ static int test_io_uring_submit_enters(const char *file)
                __u32 nlb;
 
                sqe = io_uring_get_sqe(&ring);
-               io_uring_prep_readv(sqe, fd, &vecs[i], 1, offset);
+               io_uring_prep_uring_cmd(sqe, NVME_URING_CMD_IO, fd);
                sqe->user_data = i;
-               sqe->opcode = IORING_OP_URING_CMD;
-               sqe->cmd_op = NVME_URING_CMD_IO;
                cmd = (struct nvme_uring_cmd *)sqe->cmd;
                memset(cmd, 0, sizeof(struct nvme_uring_cmd));
 
--

2.47.3

#7 lore raw

Re: [PATCHv5 1/4] liburing: provide uring_cmd prep function

Quoted 7 lines · Caleb Sander Mateos

src/include/liburing.h IOURINGINLINE void io_uring_prep_socket_direct_alloc(struct io_uring_sqe *sqe, · line 1525

Quoted 4 lines · Keith Busch
src/include/liburing.h +4 -0
+                                          int fd)
+       LIBURING_NOEXCEPT
+{
+       sqe->opcode = (__u8) IORING_OP_URING_CMD;

Casting the constant seems unnecessary. Do compilers really warn about this?

Oh, not necessary here, but the next patch wants the cast. This was copied from io_uring_prep_rw(), which passes the 'opcode' as an 'int' type. I don't know why that type was used, so just trying to match the local convention.

But I digress, I'll move the cast to the next patch where it is actually needed.

Quoted 6 lines
Quoted 1 line

+ io_uring_prep_uring_cmd(sqe, NVME_URING_CMD_IO, use_fd);

I guess this works because io_uring_prep_uring_cmd() doesn't touch sqe->buf_index or sqe->flags, but it seems like it would be less brittle to call io_uring_prep_uring_cmd() before setting any of the other sqe fields.

Good point, it happens to be "okay", but is safer to move the generic init before initializing command specific fields.

#8 lore raw

v5 cover io_uring: mixed submission queue entries sizes

In reply to Keith Busch

From: Keith Busch <[email protected]>

Resending to the correct mailing list.

Changes from v3:

- Allowed 128b opcodes on both big and mixed SQ's

- Added additional comments for clarity

- Commit message fixups

- Moved the uring specific entry size fucntion to the uring code.

Keith Busch (1):
io_uring: add support for IORING_SETUP_SQE_MIXED

 include/uapi/linux/io_uring.h |  8 ++++++++
 io_uring/fdinfo.c             | 34 +++++++++++++++++++++++++++-------
 io_uring/io_uring.c           | 35 +++++++++++++++++++++++++++++++----
 io_uring/io_uring.h           | 14 ++------------
 io_uring/opdef.c              | 26 ++++++++++++++++++++++++++
 io_uring/opdef.h              |  2 ++
 io_uring/register.c           |  2 +-
 io_uring/uring_cmd.c          | 17 +++++++++++++++--
 8 files changed, 112 insertions(+), 26 deletions(-)
Signature · 1 line
2.47.3
#9 lore raw

PATCH v5 2/4 Add support IORING_SETUP_SQE_MIXED

In reply to Keith Busch

From: Keith Busch <[email protected]>

This adds core support for mixed sized SQEs in the same SQ ring. Before this, SQEs were either 64b in size (the normal size), or 128b if IORING_SETUP_SQE128 was set in the ring initialization. With the mixed support, an SQE may be either 64b or 128b on the same SQ ring. If the SQE is 128b in size, then a 128b opcode will be set in the sqe op. When acquiring a large sqe at the end of the sq, the client may post a NOP SQE with IOSQE_CQE_SKIP_SUCCESS set that the kernel will process and skip posting a CQE.

---
 src/include/liburing.h          | 71 +++++++++++++++++++++++++++++++--
 src/include/liburing/io_uring.h |  8 ++++
 2 files changed, 75 insertions(+), 4 deletions(-)
src/include/liburing.h +67 -4
index f7af20aa..d6a45cbb 100644
@@ -800,6 +800,12 @@ IOURINGINLINE void io_uring_prep_nop(struct io_uring_sqe *sqe)
 	io_uring_prep_rw(IORING_OP_NOP, sqe, -1, NULL, 0, 0);
 }
 
+IOURINGINLINE void io_uring_prep_nop128(struct io_uring_sqe *sqe)
+	LIBURING_NOEXCEPT
+{
+	io_uring_prep_rw(IORING_OP_NOP128, sqe, -1, NULL, 0, 0);
+}
+
 IOURINGINLINE void io_uring_prep_timeout(struct io_uring_sqe *sqe,
 					 const struct __kernel_timespec *ts,
 					 unsigned count, unsigned flags)
@@ -1517,12 +1523,13 @@ IOURINGINLINE void io_uring_prep_socket_direct_alloc(struct io_uring_sqe *sqe,
 	__io_uring_set_target_fixed_file(sqe, IORING_FILE_INDEX_ALLOC - 1);
 }
 
-IOURINGINLINE void io_uring_prep_uring_cmd(struct io_uring_sqe *sqe,
-					   int cmd_op,
-					   int fd)
+IOURINGINLINE void __io_uring_prep_uring_cmd(struct io_uring_sqe *sqe,
+					     int op,
+					     int cmd_op,
+					     int fd)
 	LIBURING_NOEXCEPT
 {
-	sqe->opcode = (__u8) IORING_OP_URING_CMD;
+	sqe->opcode = (__u8) op;
 	sqe->fd = fd;
 	sqe->cmd_op = cmd_op;
 	sqe->__pad1 = 0;
@@ -1530,6 +1537,22 @@ IOURINGINLINE void io_uring_prep_uring_cmd(struct io_uring_sqe *sqe,
 	sqe->len = 0;
 }
 
+IOURINGINLINE void io_uring_prep_uring_cmd(struct io_uring_sqe *sqe,
+					   int cmd_op,
+					   int fd)
+	LIBURING_NOEXCEPT
+{
+	__io_uring_prep_uring_cmd(sqe, IORING_OP_URING_CMD, cmd_op, fd);
+}
+
+IOURINGINLINE void io_uring_prep_uring_cmd128(struct io_uring_sqe *sqe,
+					      int cmd_op,
+					      int fd)
+	LIBURING_NOEXCEPT
+{
+	__io_uring_prep_uring_cmd(sqe, IORING_OP_URING_CMD128, cmd_op, fd);
+}
+
 /*
  * Prepare commands for sockets
  */
@@ -1894,6 +1917,46 @@ IOURINGINLINE struct io_uring_sqe *_io_uring_get_sqe(struct io_uring *ring)
 	return sqe;
 }
 
+/*
+ * Return a 128B sqe to fill. Applications must later call io_uring_submit()
+ * when it's ready to tell the kernel about it. The caller may call this
+ * function multiple times before calling io_uring_submit().
+ *
+ * Returns a vacant 128B sqe, or NULL if we're full. If the current tail is the
+ * last entry in the ring, this function will insert a nop + skip complete such
+ * that the 128b entry wraps back to the beginning of the queue for a
+ * contiguous big sq entry. It's up to the caller to use a 128b opcode in order
+ * for the kernel to know how to advance its sq head pointer.
+ */
+IOURINGINLINE struct io_uring_sqe *io_uring_get_sqe128_mixed(struct io_uring *ring)
+	LIBURING_NOEXCEPT
+{
+	struct io_uring_sq *sq = &ring->sq;
+	unsigned head = io_uring_load_sq_head(ring), tail = sq->sqe_tail;
+	struct io_uring_sqe *sqe;
+
+	if (!(ring->flags & IORING_SETUP_SQE_MIXED))
+		return NULL;
+
+	if (((tail + 1) & sq->ring_mask) == 0) {
+		if ((tail + 2) - head >= sq->ring_entries)
+			return NULL;
+
+		sqe = _io_uring_get_sqe(ring);
+		io_uring_prep_nop(sqe);
+		sqe->flags |= IOSQE_CQE_SKIP_SUCCESS;
+		tail = sq->sqe_tail;
+	} else if ((tail + 1) - head >= sq->ring_entries) {
+		return NULL;
+	}
+
+	sqe = &sq->sqes[tail & sq->ring_mask];
+	sq->sqe_tail = tail + 2;
+	io_uring_initialize_sqe(sqe);
+
+	return sqe;
+}
+
 /*
  * Return the appropriate mask for a buffer ring of size 'ring_entries'
  */
src/include/liburing/io_uring.h +8 -0
index 31396057..f2388645 100644
@@ -211,6 +211,12 @@ enum io_uring_sqe_flags_bit {
  */
 #define IORING_SETUP_CQE_MIXED		(1U << 18)
 
+/*
+ *  Allow both 64b and 128b SQEs. If a 128b SQE is posted, it will have
+ *  IOSQE_SQE_128B set in sqe->flags.
+ */
+#define IORING_SETUP_SQE_MIXED		(1U << 19)
+
 enum io_uring_op {
 	IORING_OP_NOP,
 	IORING_OP_READV,
@@ -275,6 +281,8 @@ enum io_uring_op {
 	IORING_OP_READV_FIXED,
 	IORING_OP_WRITEV_FIXED,
 	IORING_OP_PIPE,
+	IORING_OP_NOP128,
+	IORING_OP_URING_CMD128,
 
 	/* this goes last, obviously */
 	IORING_OP_LAST,
Signature · 1 line
2.47.3
#10 lore raw

PATCH v5 3/4 Add nop testing for IORING_SETUP_SQE_MIXED

In reply to Keith Busch

From: Keith Busch <[email protected]>

Test mixing 64 and 128 byte sqe entries on a queue.

Insert a bad 128b operation at the end of a mixed sqe to test the kernel's invalid entry detection.

---
 test/Makefile              |  2 +
 test/sqe-mixed-bad-wrap.c  | 87 ++++++++++++++++++++++++++++++++++++++
 test/sqe-mixed-nop.c       | 82 +++++++++++++++++++++++++++++++++++
 test/sqe-mixed-uring_cmd.c |  0
 4 files changed, 171 insertions(+)
 create mode 100644 test/sqe-mixed-bad-wrap.c
 create mode 100644 test/sqe-mixed-nop.c
 create mode 100644 test/sqe-mixed-uring_cmd.c
test/Makefile +2 -0
index 64d67a1e..2c250c81 100644
@@ -234,6 +234,8 @@ test_srcs := \
 	sq-poll-share.c \
 	sqpoll-sleep.c \
 	sq-space_left.c \
+	sqe-mixed-nop.c \
+	sqe-mixed-bad-wrap.c \
 	sqwait.c \
 	stdout.c \
 	submit-and-wait.c \
test/sqe-mixed-bad-wrap.c new file +87 -0

87 lines not on this page · raw message

test/sqe-mixed-nop.c new file +82 -0

82 lines not on this page · raw message

test/sqe-mixed-uring_cmd.c new file +0 -0
new file mode 100644
index 00000000..e69de29b
Signature · 1 line
2.47.3
#11 lore raw

PATCH v5 4/4 Add mixed sqe test for uring commands

In reply to Keith Busch

From: Keith Busch <[email protected]>

---
 test/Makefile              |   1 +
 test/sqe-mixed-uring_cmd.c | 140 +++++++++++++++++++++++++++++++++++++
 2 files changed, 141 insertions(+)
test/Makefile +1 -0
index 2c250c81..2b2e3967 100644
@@ -236,6 +236,7 @@ test_srcs := \
 	sq-space_left.c \
 	sqe-mixed-nop.c \
 	sqe-mixed-bad-wrap.c \
+	sqe-mixed-uring_cmd.c \
 	sqwait.c \
 	stdout.c \
 	submit-and-wait.c \
test/sqe-mixed-uring_cmd.c +140 -0

140 lines not on this page · raw message

Signature · 1 line
2.47.3