Skip to contentSkip to conversation

[PATCHv3 0/3]

PATCH Superseded Keith Busch 4 participants 12 messages Sep 24, 2025 – Sep 25, 2025 io_uring

v3 of 6 Superseded Latest is v6 →

lore
Share & tools
lore
#0 lore raw

v3 cover [PATCHv3 0/3]

From: Keith Busch <[email protected]>

Previous version:

https://lore.kernel.org/io-uring/[email protected]/

The CQ supports mixed size entries, so why not for SQ's too? There are use cases that currently allocate different queues just to keep these things separated, but we can efficiently handle both cases in a single ring.

Changes since v2:

- Define 128B opcodes to be used on mixed SQs. This is done instead of using the last SQE flags bit to generically identify a command as such. The new opcodes are valid only on a mixed SQ.

- Fixed up the accounting of sqes left to dispatch. The big sqes on a mixed sq count for two entries, so previously would have fetched too many.

- liburing won't bother submitting the nop-skip for the wrap-around condition if there are not enoungh free entries for the big-sqe.

kernel:

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           | 27 +++++++++++++++++++++++----
 io_uring/io_uring.h           |  8 +++++---
 io_uring/opdef.c              | 26 ++++++++++++++++++++++++++
 io_uring/opdef.h              |  2 ++
 io_uring/register.c           |  2 +-
 io_uring/uring_cmd.c          | 12 ++++++++++--
 io_uring/uring_cmd.h          |  1 +
 9 files changed, 103 insertions(+), 17 deletions(-)

liburing:

Keith Busch (3):
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          |  50 +++++++++++
 src/include/liburing/io_uring.h |  11 +++
 test/Makefile                   |   3 +
 test/sqe-mixed-bad-wrap.c       |  89 ++++++++++++++++++++
 test/sqe-mixed-nop.c            |  82 ++++++++++++++++++
 test/sqe-mixed-uring_cmd.c      | 142 ++++++++++++++++++++++++++++++++
 6 files changed, 377 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
Signature · 1 line
2.47.3
#1 lore raw

PATCH v3 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 post 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 posted with the SKIP_SUCCESS flag set. The kernel should simply ignore those.

---
 include/uapi/linux/io_uring.h |  8 ++++++++
 io_uring/fdinfo.c             | 34 +++++++++++++++++++++++++++-------
 io_uring/io_uring.c           | 27 +++++++++++++++++++++++----
 io_uring/io_uring.h           |  8 +++++---
 io_uring/opdef.c              | 26 ++++++++++++++++++++++++++
 io_uring/opdef.h              |  2 ++
 io_uring/register.c           |  2 +-
 io_uring/uring_cmd.c          | 12 ++++++++++--
 io_uring/uring_cmd.h          |  1 +
 9 files changed, 103 insertions(+), 17 deletions(-)
include/uapi/linux/io_uring.h +8 -0
index a0cc1cc0dd015..ce01f043fceec 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 +23 -4
index 1bfa124565f71..f9bc442bb4188 100644
@@ -2153,7 +2153,7 @@ 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;
@@ -2179,6 +2179,14 @@ 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) {
+		if (!(ctx->flags & IORING_SETUP_SQE_MIXED) || *left < 2 ||
+		    (ctx->cached_sq_head & (ctx->sq_entries - 1)) == 0)
+			return io_init_fail_req(req, -EINVAL);
+		ctx->cached_sq_head++;
+		(*left)--;
+	}
+
 	if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) {
 		/* enforce forwards compatibility on users */
 		if (sqe_flags & ~SQE_VALID_FLAGS)
@@ -2288,13 +2296,13 @@ 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);
 
@@ -2446,7 +2454,7 @@ 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;
@@ -2791,6 +2799,10 @@ 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);
@@ -3717,6 +3729,13 @@ 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 +5 -3
index af4c113106523..074908d5884e4 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 |\
@@ -582,9 +583,10 @@ static inline void io_req_queue_tw_complete(struct io_kiocb *req, s32 res)
  * 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)
+static inline size_t uring_sqe_size(struct io_kiocb *req)
 {
-	if (ctx->flags & IORING_SETUP_SQE128)
+	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);
 }
io_uring/opdef.c +26 -0
index 932319633eac2..36feebcce3827 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_cmd128_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 +10 -2
index a688c9f1a21cd..5fa3c260bc142 100644
@@ -216,6 +216,13 @@ int io_uring_cmd_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
 	return 0;
 }
 
+int io_uring_cmd128_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
+{
+	if (!(req->ctx->flags & IORING_SETUP_SQE_MIXED))
+		return -EINVAL;
+	return io_uring_cmd_prep(req, 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 +231,7 @@ 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 +249,8 @@ 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;
io_uring/uring_cmd.h +1 -0
index 041aef8a8aa3f..0d6068fba7d0d 100644
@@ -10,6 +10,7 @@ struct io_async_cmd {
 
 int io_uring_cmd(struct io_kiocb *req, unsigned int issue_flags);
 int io_uring_cmd_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe);
+int io_uring_cmd128_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe);
 void io_uring_cmd_sqe_copy(struct io_kiocb *req);
 void io_uring_cmd_cleanup(struct io_kiocb *req);
 
Signature · 1 line
2.47.3
maintainer
#2 lore raw

Re: [PATCHv3 1/1] io_uring: add support for IORING_SETUP_SQE_MIXED

Quoted 3 lines · Keith Busch

contiguous in the SQ ring, a 128b SQE cannot wrap the ring. For this case, a single NOP SQE should be posted with the SKIP_SUCCESS flag set. The kernel should simply ignore those.

I think this mirrors the CQE side too much - the kernel doesn't ignore then, they get processed just like any other NOP that has SKIP_SUCCESS set. They don't post a CQE, but that's not because they are ignored, that's just the nature of a successful NOP w/SKIP_SUCCESS set.

io_uring/io_uring.c static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req, · line 2188

Quoted 11 lines
io_uring/io_uring.c +7 -0
@@ -2179,6 +2179,14 @@ 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) {
+		if (!(ctx->flags & IORING_SETUP_SQE_MIXED) || *left < 2 ||
+		    (ctx->cached_sq_head & (ctx->sq_entries - 1)) == 0)
+			return io_init_fail_req(req, -EINVAL);
+		ctx->cached_sq_head++;
+		(*left)--;
+	}

This could do with a comment!

io_uring/io_uring.h static inline void io_req_queue_tw_complete(struct io_kiocb *req, s32 res) · line 591

Quoted 12 lines
io_uring/io_uring.h +3 -2
@@ -582,9 +583,10 @@ static inline void io_req_queue_tw_complete(struct io_kiocb *req, s32 res)
  * 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)
+static inline size_t uring_sqe_size(struct io_kiocb *req)
 {
-	if (ctx->flags & IORING_SETUP_SQE128)
+	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);

This one really confused me, but then I grep'ed, and it's uring_cmd specific. Should probably move this one to uring_cmd.c rather than have it elsewhere.

io_uring/uring_cmd.c int io_uring_cmd_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) · line 224

Quoted 6 lines
io_uring/uring_cmd.c +6 -0
+int io_uring_cmd128_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
+{
+	if (!(req->ctx->flags & IORING_SETUP_SQE_MIXED))
+		return -EINVAL;
+	return io_uring_cmd_prep(req, sqe);
+}

Why isn't this just allowed for SQE128 as well? There should be no reason to disallow explicitly 128b sqe commands in SQE128 mode, they should work for any mode that supports 128b SQEs which is either SQE_MIXED or SQE128?

Signature · 1 line
Jens Axboe
#3 lore raw

Re: [PATCHv3 1/1] io_uring: add support for IORING_SETUP_SQE_MIXED

Quoted 42 lines · Jens Axboe
Quoted 3 lines · Keith Busch

contiguous in the SQ ring, a 128b SQE cannot wrap the ring. For this case, a single NOP SQE should be posted with the SKIP_SUCCESS flag set. The kernel should simply ignore those.

I think this mirrors the CQE side too much - the kernel doesn't ignore then, they get processed just like any other NOP that has SKIP_SUCCESS set. They don't post a CQE, but that's not because they are ignored, that's just the nature of a successful NOP w/SKIP_SUCCESS set.

io_uring/io_uring.c static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req, · line 2188

Quoted 11 lines
io_uring/io_uring.c +7 -0
@@ -2179,6 +2179,14 @@ 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) {
+             if (!(ctx->flags & IORING_SETUP_SQE_MIXED) || *left < 2 ||
+                 (ctx->cached_sq_head & (ctx->sq_entries - 1)) == 0)
+                     return io_init_fail_req(req, -EINVAL);
+             ctx->cached_sq_head++;
+             (*left)--;
+     }

This could do with a comment!

io_uring/io_uring.h static inline void io_req_queue_tw_complete(struct io_kiocb *req, s32 res) · line 591

Quoted 12 lines
io_uring/io_uring.h +3 -2
@@ -582,9 +583,10 @@ static inline void io_req_queue_tw_complete(struct io_kiocb *req, s32 res)
  * 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)
+static inline size_t uring_sqe_size(struct io_kiocb *req)
 {
-     if (ctx->flags & IORING_SETUP_SQE128)
+     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);

This one really confused me, but then I grep'ed, and it's uring_cmd specific. Should probably move this one to uring_cmd.c rather than have it elsewhere.

Quoted 11 lines

io_uring/uring_cmd.c int io_uring_cmd_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) · line 224

Quoted 6 lines
io_uring/uring_cmd.c +6 -0
+int io_uring_cmd128_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
+{
+     if (!(req->ctx->flags & IORING_SETUP_SQE_MIXED))
+             return -EINVAL;
+     return io_uring_cmd_prep(req, sqe);
+}

Why isn't this just allowed for SQE128 as well? There should be no reason to disallow explicitly 128b sqe commands in SQE128 mode, they should work for any mode that supports 128b SQEs which is either SQE_MIXED or SQE128?

Not to mention, the check in io_init_req() should already have rejected a 128-byte operation on a non-IORING_SETUP_SQE_MIXED io_ring_ctx.

Best,
Caleb

maintainer
#4 lore raw

Re: [PATCHv3 1/1] io_uring: add support for IORING_SETUP_SQE_MIXED

Quoted 12 lines · Caleb Sander Mateos
Quoted 11 lines

io_uring/uring_cmd.c int io_uring_cmd_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) · line 224

Quoted 6 lines
io_uring/uring_cmd.c +6 -0
+int io_uring_cmd128_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
+{
+     if (!(req->ctx->flags & IORING_SETUP_SQE_MIXED))
+             return -EINVAL;
+     return io_uring_cmd_prep(req, sqe);
+}

Why isn't this just allowed for SQE128 as well? There should be no reason to disallow explicitly 128b sqe commands in SQE128 mode, they should work for any mode that supports 128b SQEs which is either SQE_MIXED or SQE128?

Quoted 3 lines

Not to mention, the check in io_init_req() should already have rejected a 128-byte operation on a non-IORING_SETUP_SQE_MIXED io_ring_ctx.

Yes good point, it should all be caught there, no need for any per-op checking (or the separate prep handler).

Signature · 1 line
Jens Axboe
#5 lore raw

PATCH v3 1/3 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 should simply ignore as it's just a pad filler that is posted when required.

---
 src/include/liburing.h          | 50 +++++++++++++++++++++++++++++++++
 src/include/liburing/io_uring.h | 11 ++++++++
 2 files changed, 61 insertions(+)
src/include/liburing.h +50 -0
index 052d6b56..66f1b990 100644
@@ -575,6 +575,7 @@ IOURINGINLINE void io_uring_initialize_sqe(struct io_uring_sqe *sqe)
 	sqe->buf_index = 0;
 	sqe->personality = 0;
 	sqe->file_index = 0;
+	sqe->addr2 = 0;
 	sqe->addr3 = 0;
 	sqe->__pad2[0] = 0;
 }
@@ -799,6 +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,
 					 struct __kernel_timespec *ts,
 					 unsigned count, unsigned flags)
@@ -1882,6 +1889,49 @@ 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);
+		if (!sqe)
+			return NULL;
+
+		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 +11 -0
index 31396057..1e0b6398 100644
@@ -126,6 +126,7 @@ enum io_uring_sqe_flags_bit {
 	IOSQE_ASYNC_BIT,
 	IOSQE_BUFFER_SELECT_BIT,
 	IOSQE_CQE_SKIP_SUCCESS_BIT,
+	IOSQE_SQE_128B_BIT,
 };
 
 /*
@@ -145,6 +146,8 @@ enum io_uring_sqe_flags_bit {
 #define IOSQE_BUFFER_SELECT	(1U << IOSQE_BUFFER_SELECT_BIT)
 /* don't post CQE if request succeeded */
 #define IOSQE_CQE_SKIP_SUCCESS	(1U << IOSQE_CQE_SKIP_SUCCESS_BIT)
+/* this is a 128b/big-sqe posting */
+#define IOSQE_SQE_128B          (1U << IOSQE_SQE_128B_BIT)
 
 /*
  * io_uring_setup() flags
@@ -211,6 +214,12 @@ 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 +284,8 @@ 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
#6 lore raw

Re: [PATCHv3 1/3] Add support IORING_SETUP_SQE_MIXED

Quoted 23 lines · 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 should simply ignore as it's just a pad filler that is posted when required.

---
 src/include/liburing.h          | 50 +++++++++++++++++++++++++++++++++
 src/include/liburing/io_uring.h | 11 ++++++++
 2 files changed, 61 insertions(+)
src/include/liburing.h +0 -0
index 052d6b56..66f1b990 100644

src/include/liburing.h IOURINGINLINE void io_uring_initialize_sqe(struct io_uring_sqe *sqe) · line 578

Quoted 4 lines
src/include/liburing.h +1 -0
@@ -575,6 +575,7 @@ IOURINGINLINE void io_uring_initialize_sqe(struct io_uring_sqe *sqe)
        sqe->buf_index = 0;
        sqe->personality = 0;
        sqe->file_index = 0;
+       sqe->addr2 = 0;

Why is this necessary for mixed SQE size support? It looks like this field is already initialized in io_uring_prep_rw() via the unioned off field. Though, to be honest, I can't say I understand why the initialization of the SQE fields is split between io_uring_initialize_sqe() and io_uring_prep_rw().

Quoted 38 lines
src/include/liburing.h +24 -0
        sqe->addr3 = 0;
        sqe->__pad2[0] = 0;
 }
@@ -799,6 +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,
                                         struct __kernel_timespec *ts,
                                         unsigned count, unsigned flags)
@@ -1882,6 +1889,49 @@ 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;
+

src/include/liburing.h IOURINGINLINE struct io_uring_sqe *_io_uring_get_sqe(struct io_uring *ring) · line 1919

Quoted 10 lines
src/include/liburing.h +10 -0
+       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);
+               if (!sqe)
+                       return NULL;

This case should be impossible since we just checked there is an empty SQ slot at the end of the ring plus two more at the beginning.

Quoted 23 lines
src/include/liburing.h +15 -0
+
+               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 +0 -0
index 31396057..1e0b6398 100644

src/include/liburing/io_uring.h enum io_uring_sqe_flags_bit { · line 129

Quoted 4 lines
src/include/liburing/io_uring.h +1 -0
@@ -126,6 +126,7 @@ enum io_uring_sqe_flags_bit {
        IOSQE_ASYNC_BIT,
        IOSQE_BUFFER_SELECT_BIT,
        IOSQE_CQE_SKIP_SUCCESS_BIT,
+       IOSQE_SQE_128B_BIT,

I thought we decided against using an SQE flag bit for this? Looks like this needs to be re-synced with the kernel uapi header.

Best,
Caleb

Quoted 37 lines
src/include/liburing/io_uring.h +10 -1
 };
 
 /*
@@ -145,6 +146,8 @@ enum io_uring_sqe_flags_bit {
 #define IOSQE_BUFFER_SELECT    (1U << IOSQE_BUFFER_SELECT_BIT)
 /* don't post CQE if request succeeded */
 #define IOSQE_CQE_SKIP_SUCCESS (1U << IOSQE_CQE_SKIP_SUCCESS_BIT)
+/* this is a 128b/big-sqe posting */
+#define IOSQE_SQE_128B          (1U << IOSQE_SQE_128B_BIT)
 
 /*
  * io_uring_setup() flags
@@ -211,6 +214,12 @@ 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 +284,8 @@ 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,
--

2.47.3

#7 lore raw

Re: [PATCHv3 1/3] Add support IORING_SETUP_SQE_MIXED

Quoted 6 lines · Caleb Sander Mateos
Quoted 6 lines

index 052d6b56..66f1b990 100644

src/include/liburing.h +0 -0
@@ -575,6 +575,7 @@ IOURINGINLINE void io_uring_initialize_sqe(struct io_uring_sqe *sqe)
        sqe->buf_index = 0;
        sqe->personality = 0;
Quoted 8 lines

src/include/liburing.h IOURINGINLINE void io_uring_initialize_sqe(struct io_uring_sqe *sqe) · line 578

Quoted 2 lines
src/include/liburing.h +1 -0
        sqe->file_index = 0;
+       sqe->addr2 = 0;

Why is this necessary for mixed SQE size support? It looks like this field is already initialized in io_uring_prep_rw() via the unioned off field. Though, to be honest, I can't say I understand why the initialization of the SQE fields is split between io_uring_initialize_sqe() and io_uring_prep_rw().

The nvme passthrough uring_cmd doesn't call io_uring_prep_rw(), so we'd just get a stale value in that field if we don't clear it. But you're right that many cases would end up setting the field twice when we don't need that.

Quoted 4 lines

src/include/liburing.h IOURINGINLINE void io_uring_initialize_sqe(struct io_uring_sqe *sqe) · line 579

Quoted 1 line
src/include/liburing.h +1 -0
+       IOSQE_SQE_128B_BIT,

I thought we decided against using an SQE flag bit for this? Looks like this needs to be re-synced with the kernel uapi header.

We did, and this is a left over artifact that is not supposed to be here. :( Nothing is depending on the bit in this series.

#8 lore raw

Re: [PATCHv3 1/3] Add support IORING_SETUP_SQE_MIXED

Quoted 17 lines · Keith Busch
Quoted 6 lines · Caleb Sander Mateos
Quoted 6 lines

index 052d6b56..66f1b990 100644

src/include/liburing.h +0 -0
@@ -575,6 +575,7 @@ IOURINGINLINE void io_uring_initialize_sqe(struct io_uring_sqe *sqe)
        sqe->buf_index = 0;
        sqe->personality = 0;
Quoted 8 lines

src/include/liburing.h IOURINGINLINE void io_uring_initialize_sqe(struct io_uring_sqe *sqe) · line 578

Quoted 2 lines
src/include/liburing.h +1 -0
        sqe->file_index = 0;
+       sqe->addr2 = 0;

Why is this necessary for mixed SQE size support? It looks like this field is already initialized in io_uring_prep_rw() via the unioned off field. Though, to be honest, I can't say I understand why the initialization of the SQE fields is split between io_uring_initialize_sqe() and io_uring_prep_rw().

Quoted 4 lines

The nvme passthrough uring_cmd doesn't call io_uring_prep_rw(), so we'd just get a stale value in that field if we don't clear it. But you're right that many cases would end up setting the field twice when we don't need that.

Sure, that's a reasonable concern. Perhaps a helper for initializing a NVMe passthru operation would make sense, though maybe it's difficult to do that without requiring the linux/nvme_ioctl.h uapi header. But regardless, it seems unrelated to the mixed SQE size work.

Quoted 8 lines
Quoted 5 lines

src/include/liburing.h IOURINGINLINE void io_uring_initialize_sqe(struct io_uring_sqe *sqe) · line 579

Quoted 2 lines
src/include/liburing.h +1 -0
+       IOSQE_SQE_128B_BIT,

I thought we decided against using an SQE flag bit for this? Looks like this needs to be re-synced with the kernel uapi header.

We did, and this is a left over artifact that is not supposed to be here. :( Nothing is depending on the bit in this series.

Yeah I figured this file just needed to be updated with the current version of the uapi header defined in your latest kernel patch.

Best,
Caleb

#9 lore raw

PATCH v3 2/3 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  | 89 ++++++++++++++++++++++++++++++++++++++
 test/sqe-mixed-nop.c       | 82 +++++++++++++++++++++++++++++++++++
 test/sqe-mixed-uring_cmd.c |  0
 4 files changed, 173 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 +89 -0

89 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
#10 lore raw

PATCH v3 3/3 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 | 142 +++++++++++++++++++++++++++++++++++++
 2 files changed, 143 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 +142 -0

142 lines not on this page · raw message

Signature · 1 line
2.47.3
#11 lore raw

Re: [PATCHv3 0/3] io_uring: mixed submission queue size support

In reply to Keith Busch

Resend for missing subject... I keep pointing to the wrong directory after merging the liburing and kernel cover letters.

Quoted 62 lines · Keith Busch

From: Keith Busch <[email protected]>

Previous version:

https://lore.kernel.org/io-uring/[email protected]/

The CQ supports mixed size entries, so why not for SQ's too? There are use cases that currently allocate different queues just to keep these things separated, but we can efficiently handle both cases in a single ring.

Changes since v2:

- Define 128B opcodes to be used on mixed SQs. This is done instead of using the last SQE flags bit to generically identify a command as such. The new opcodes are valid only on a mixed SQ.

- Fixed up the accounting of sqes left to dispatch. The big sqes on a mixed sq count for two entries, so previously would have fetched too many.

- liburing won't bother submitting the nop-skip for the wrap-around condition if there are not enoungh free entries for the big-sqe.

kernel:

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           | 27 +++++++++++++++++++++++----
 io_uring/io_uring.h           |  8 +++++---
 io_uring/opdef.c              | 26 ++++++++++++++++++++++++++
 io_uring/opdef.h              |  2 ++
 io_uring/register.c           |  2 +-
 io_uring/uring_cmd.c          | 12 ++++++++++--
 io_uring/uring_cmd.h          |  1 +
 9 files changed, 103 insertions(+), 17 deletions(-)

liburing:

Keith Busch (3):
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          |  50 +++++++++++
 src/include/liburing/io_uring.h |  11 +++
 test/Makefile                   |   3 +
 test/sqe-mixed-bad-wrap.c       |  89 ++++++++++++++++++++
 test/sqe-mixed-nop.c            |  82 ++++++++++++++++++
 test/sqe-mixed-uring_cmd.c      | 142 ++++++++++++++++++++++++++++++++
 6 files changed, 377 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
Signature · 1 line
2.47.3