Skip to contentSkip to conversation

Provide more efficient buffer registration

PATCH In review Jens Axboe 2 participants 7 messages Sep 12, 2024 – Sep 20, 2024 io_uring

lore
Share & tools
lore

Jens Axboe

author maintainer
#0 lore raw

v3 cover Provide more efficient buffer registration

Hi,

Pretty much what the subject line says, it's about 25k to 40k times faster to provide a way to duplicate an existing rings buffer registration than it is manually map/pin/register the buffers again with a new ring.

Patch 1 is just a prep patch, patch 2 adds refs to struct io_mapped_ubuf, patch 3 abstracts out a helper, and patch 4 finally adds the register opcode to allow a ring to duplicate the registered mappings from one ring to another.

This came about from discussing overhead from the varnish cache project for cases with more dynamic ring/thread creation.

Also see the buf-copy liburing branch for support and test code:

https://git.kernel.dk/cgit/liburing/log/?h=buf-copy

 include/uapi/linux/io_uring.h | 13 +++++
 io_uring/register.c           | 60 ++++++++++++++--------
 io_uring/register.h           |  1 +
 io_uring/rsrc.c               | 96 ++++++++++++++++++++++++++++++++++-
 io_uring/rsrc.h               |  2 +
 5 files changed, 150 insertions(+), 22 deletions(-)

Since v2:
- Ensure that it works for registered rings (both src/dst)
- Little cleanups

Signature · 1 line
Jens Axboe
author maintainer
#1 lore raw

PATCH v3 1/4 io_uring/rsrc: clear 'slot' entry upfront

No functional changes in this patch, but clearing the slot pointer earlier will be required by a later change.

---
 io_uring/rsrc.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
io_uring/rsrc.c +1 -1
index 7d639a996f28..d42114845fac 100644
@@ -114,6 +114,7 @@ static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slo
 	struct io_mapped_ubuf *imu = *slot;
 	unsigned int i;
 
+	*slot = NULL;
 	if (imu != &dummy_ubuf) {
 		for (i = 0; i < imu->nr_bvecs; i++)
 			unpin_user_page(imu->bvec[i].bv_page);
@@ -121,7 +122,6 @@ static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slo
 			io_unaccount_mem(ctx, imu->acct_pages);
 		kvfree(imu);
 	}
-	*slot = NULL;
 }
 
 static void io_rsrc_put_work(struct io_rsrc_node *node)
Signature · 1 line
2.45.2
author maintainer
#2 lore raw

PATCH v3 2/4 io_uring/rsrc: add reference count to struct io_mapped_ubuf

In reply to Jens Axboe

Currently there's a single ring owner of a mapped buffer, and hence the reference count will always be 1 when it's torn down and freed. However, in preparation for being able to link io_mapped_ubuf to different spots, add a reference count to manage the lifetime of it.

---
 io_uring/rsrc.c | 3 +++
 io_uring/rsrc.h | 1 +
 2 files changed, 4 insertions(+)
io_uring/rsrc.c +3 -0
index d42114845fac..28f98de3c304 100644
@@ -116,6 +116,8 @@ static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slo
 
 	*slot = NULL;
 	if (imu != &dummy_ubuf) {
+		if (!refcount_dec_and_test(&imu->refs))
+			return;
 		for (i = 0; i < imu->nr_bvecs; i++)
 			unpin_user_page(imu->bvec[i].bv_page);
 		if (imu->acct_pages)
@@ -990,6 +992,7 @@ static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
 		imu->folio_shift = data.folio_shift;
 		imu->folio_mask = ~((1UL << data.folio_shift) - 1);
 	}
+	refcount_set(&imu->refs, 1);
 	off = (unsigned long) iov->iov_base & ~imu->folio_mask;
 	*pimu = imu;
 	ret = 0;
io_uring/rsrc.h +1 -0
index 3d0dda3556e6..98a253172c27 100644
@@ -47,6 +47,7 @@ struct io_mapped_ubuf {
 	unsigned int    folio_shift;
 	unsigned long	acct_pages;
 	unsigned long   folio_mask;
+	refcount_t	refs;
 	struct bio_vec	bvec[] __counted_by(nr_bvecs);
 };
 
Signature · 1 line
2.45.2
author maintainer
#3 lore raw

PATCH v3 3/4 io_uring/register: provide helper to get io_ring_ctx from 'fd'

In reply to Jens Axboe

Can be done in one of two ways:

1) Regular file descriptor, just fget()
2) Registered ring, index our own table for that

In preparation for adding another register use of needing to get a ctx from a file descriptor, abstract out this helper and use it in the main register syscall as well.

---
 io_uring/register.c | 54 +++++++++++++++++++++++++++------------------
 io_uring/register.h |  1 +
 2 files changed, 34 insertions(+), 21 deletions(-)
io_uring/register.c +33 -21
index 57cb85c42526..d90159478045 100644
@@ -550,21 +550,16 @@ static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
 	return ret;
 }
 
-SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
-		void __user *, arg, unsigned int, nr_args)
+/*
+ * Given an 'fd' value, return the ctx associated with if. If 'registered' is
+ * true, then the registered index is used. Otherwise, the normal fd table.
+ * Caller must call fput() on the returned file, unless it's an ERR_PTR.
+ */
+struct file *io_uring_register_get_file(int fd, bool registered)
 {
-	struct io_ring_ctx *ctx;
-	long ret = -EBADF;
 	struct file *file;
-	bool use_registered_ring;
 
-	use_registered_ring = !!(opcode & IORING_REGISTER_USE_REGISTERED_RING);
-	opcode &= ~IORING_REGISTER_USE_REGISTERED_RING;
-
-	if (opcode >= IORING_REGISTER_LAST)
-		return -EINVAL;
-
-	if (use_registered_ring) {
+	if (registered) {
 		/*
 		 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
 		 * need only dereference our task private array to find it.
@@ -572,27 +567,44 @@ SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
 		struct io_uring_task *tctx = current->io_uring;
 
 		if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX))
-			return -EINVAL;
+			return ERR_PTR(-EINVAL);
 		fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
 		file = tctx->registered_rings[fd];
-		if (unlikely(!file))
-			return -EBADF;
 	} else {
 		file = fget(fd);
-		if (unlikely(!file))
-			return -EBADF;
-		ret = -EOPNOTSUPP;
-		if (!io_is_uring_fops(file))
-			goto out_fput;
 	}
 
+	if (unlikely(!file))
+		return ERR_PTR(-EBADF);
+	if (io_is_uring_fops(file))
+		return file;
+	fput(file);
+	return ERR_PTR(-EOPNOTSUPP);
+}
+
+SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
+		void __user *, arg, unsigned int, nr_args)
+{
+	struct io_ring_ctx *ctx;
+	long ret = -EBADF;
+	struct file *file;
+	bool use_registered_ring;
+
+	use_registered_ring = !!(opcode & IORING_REGISTER_USE_REGISTERED_RING);
+	opcode &= ~IORING_REGISTER_USE_REGISTERED_RING;
+
+	if (opcode >= IORING_REGISTER_LAST)
+		return -EINVAL;
+
+	file = io_uring_register_get_file(fd, use_registered_ring);
+	if (IS_ERR(file))
+		return PTR_ERR(file);
 	ctx = file->private_data;
 
 	mutex_lock(&ctx->uring_lock);
 	ret = __io_uring_register(ctx, opcode, arg, nr_args);
 	mutex_unlock(&ctx->uring_lock);
 	trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, ret);
-out_fput:
 	if (!use_registered_ring)
 		fput(file);
 	return ret;
io_uring/register.h +1 -0
index c9da997d503c..cc69b88338fe 100644
@@ -4,5 +4,6 @@
 
 int io_eventfd_unregister(struct io_ring_ctx *ctx);
 int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id);
+struct file *io_uring_register_get_file(int fd, bool registered);
 
 #endif
Signature · 1 line
2.45.2
author maintainer
#4 lore raw

PATCH v3 4/4 io_uring: add IORING_REGISTER_COPY_BUFFERS method

In reply to Jens Axboe

Buffers can get registered with io_uring, which allows to skip the repeated pin_pages, unpin/unref pages for each O_DIRECT operation. This reduces the overhead of O_DIRECT IO.

However, registrering buffers can take some time. Normally this isn't an issue as it's done at initialization time (and hence less critical), but for cases where rings can be created and destroyed as part of an IO thread pool, registering the same buffers for multiple rings become a more time sensitive proposition. As an example, let's say an application has an IO memory pool of 500G. Initial registration takes:

Got 500 huge pages (each 1024MB)
Registered 500 pages in 409 msec

or about 0.4 seconds. If we go higher to 900 1GB huge pages being registered:

Registered 900 pages in 738 msec

which is, as expected, a fully linear scaling.

Rather than have each ring pin/map/register the same buffer pool, provide an io_uring_register(2) opcode to simply duplicate the buffers that are registered with another ring. Adding the same 900GB of registered buffers to the target ring can then be accomplished in:

Copied 900 pages in 17 usec

While timing differs a bit, this provides around a 25,000-40,000x speedup for this use case.

---
 include/uapi/linux/io_uring.h | 13 +++++
 io_uring/register.c           |  6 +++
 io_uring/rsrc.c               | 91 +++++++++++++++++++++++++++++++++++
 io_uring/rsrc.h               |  1 +
 4 files changed, 111 insertions(+)
include/uapi/linux/io_uring.h +13 -0
index a275f91d2ac0..9dc5bb428c8a 100644
@@ -609,6 +609,9 @@ enum io_uring_register_op {
 
 	IORING_REGISTER_CLOCK			= 29,
 
+	/* copy registered buffers from source ring to current ring */
+	IORING_REGISTER_COPY_BUFFERS		= 30,
+
 	/* this goes last */
 	IORING_REGISTER_LAST,
 
@@ -694,6 +697,16 @@ struct io_uring_clock_register {
 	__u32	__resv[3];
 };
 
+enum {
+	IORING_REGISTER_SRC_REGISTERED = 1,
+};
+
+struct io_uring_copy_buffers {
+	__u32	src_fd;
+	__u32	flags;
+	__u32	pad[6];
+};
+
 struct io_uring_buf {
 	__u64	addr;
 	__u32	len;
io_uring/register.c +6 -0
index d90159478045..dab0f8024ddf 100644
@@ -542,6 +542,12 @@ static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
 			break;
 		ret = io_register_clock(ctx, arg);
 		break;
+	case IORING_REGISTER_COPY_BUFFERS:
+		ret = -EINVAL;
+		if (!arg || nr_args != 1)
+			break;
+		ret = io_register_copy_buffers(ctx, arg);
+		break;
 	default:
 		ret = -EINVAL;
 		break;
io_uring/rsrc.c +91 -0

100 lines not on this page · raw message

io_uring/rsrc.h +1 -0
index 98a253172c27..93546ab337a6 100644
@@ -68,6 +68,7 @@ int io_import_fixed(int ddir, struct iov_iter *iter,
 			   struct io_mapped_ubuf *imu,
 			   u64 buf_addr, size_t len);
 
+int io_register_copy_buffers(struct io_ring_ctx *ctx, void __user *arg);
 void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx);
 int io_sqe_buffers_unregister(struct io_ring_ctx *ctx);
 int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
Signature · 1 line
2.45.2
#5 lore raw

Re: [PATCH 4/4] io_uring: add IORING_REGISTER_COPY_BUFFERS method

Quoted 26 lines · Jens Axboe

Buffers can get registered with io_uring, which allows to skip the repeated pin_pages, unpin/unref pages for each O_DIRECT operation. This reduces the overhead of O_DIRECT IO.

However, registrering buffers can take some time. Normally this isn't an issue as it's done at initialization time (and hence less critical), but for cases where rings can be created and destroyed as part of an IO thread pool, registering the same buffers for multiple rings become a more time sensitive proposition. As an example, let's say an application has an IO memory pool of 500G. Initial registration takes:

Got 500 huge pages (each 1024MB)
Registered 500 pages in 409 msec

or about 0.4 seconds. If we go higher to 900 1GB huge pages being registered:

Registered 900 pages in 738 msec

which is, as expected, a fully linear scaling.

Rather than have each ring pin/map/register the same buffer pool, provide an io_uring_register(2) opcode to simply duplicate the buffers that are registered with another ring. Adding the same 900GB of registered buffers to the target ring can then be accomplished in:

Quoted 4 lines

Copied 900 pages in 17 usec

While timing differs a bit, this provides around a 25,000-40,000x speedup for this use case.

Looks good, but I couldn't get it to apply on top of your branches. I have only one comment, if you are doing a v4:

Quoted 22 lines
---
 include/uapi/linux/io_uring.h | 13 +++++
 io_uring/register.c           |  6 +++
 io_uring/rsrc.c               | 91 +++++++++++++++++++++++++++++++++++
 io_uring/rsrc.h               |  1 +
 4 files changed, 111 insertions(+)
io_uring/rsrc.c +1 -0
@@ -17,6 +17,7 @@
 #include "openclose.h"
 #include "rsrc.h"
 #include "memmap.h"
+#include "register.h"
 
 struct io_rsrc_update {
 	struct file			*file;

io_uring/rsrc.c int io_import_fixed(int ddir, struct iov_iter *iter, · line 1142

Quoted 5 lines
io_uring/rsrc.c +2 -0
@@ -1137,3 +1138,93 @@ int io_import_fixed(int ddir, struct iov_iter *iter,
 
 	return 0;
 }
+
+static int io_copy_buffers(struct io_ring_ctx *ctx, struct io_ring_ctx *src_ctx)

The error handling code in this function is a bit hairy, IMO. I think if you check nbufs unlocked and validate it later, it could be much simpler:

static int io_copy_buffers(struct io_ring_ctx *ctx, struct io_ring_ctx *src_ctx)

{
	struct io_mapped_ubuf **user_bufs;
	struct io_rsrc_data *data;
	int i, ret, nbufs;

	/* Read nr_user_bufs unlocked.  Must be validated later */
	nbufs = READ_ONCE(src_ctx->nr_user_bufs);
	if (!nbufs)
		return -ENXIO;

	ret = io_rsrc_data_alloc(ctx, IORING_RSRC_BUFFER, NULL, nbufs, &data);
	if (ret)
		return ret;

	user_bufs = kcalloc(nbufs, sizeof(*ctx->user_bufs), GFP_KERNEL);
	if (!user_bufs) {
        	ret = -ENOMEM;
		goto out_free_data;
        }

	mutex_unlock(&ctx->uring_lock);
	mutex_lock(&src_ctx->uring_lock);
 	ret = -EBUSY;
	if (nbufs != src_ctx->nr_user_bufs) {
		mutex_unlock(&src_ctx->uring_lock);
		mutex_lock(&ctx->uring_lock);
		goto out;
	}
	for (i = 0; i < nbufs; i++) {
		struct io_mapped_ubuf *src = src_ctx->user_bufs[i];
		refcount_inc(&src->refs);
		user_bufs[i] = src;
	}

	/* Have a ref on the bufs now, drop src lock and re-grab our own lock */
	mutex_unlock(&src_ctx->uring_lock);
	mutex_lock(&ctx->uring_lock);

	if (!ctx->user_bufs)
                goto out_unmap;

	ctx->user_bufs = user_bufs;
	ctx->buf_data = data;
	ctx->nr_user_bufs = nbufs;

	return 0;

out_unmap:
/* someone raced setting up buffers, dump ours */ for (i = 0; i < nbufs; i++)
io_buffer_unmap(ctx, &user_bufs[i]);
out:

	kfree(user_bufs);
out_free_data:
	io_rsrc_data_free(data);
	return ret;
}

Thanks,

Signature · 1 line
Gabriel Krisman Bertazi
author maintainer
#6 lore raw

Re: [PATCH 4/4] io_uring: add IORING_REGISTER_COPY_BUFFERS method

Quoted 58 lines · Gabriel Krisman Bertazi
Quoted 26 lines · Jens Axboe

Buffers can get registered with io_uring, which allows to skip the repeated pin_pages, unpin/unref pages for each O_DIRECT operation. This reduces the overhead of O_DIRECT IO.

However, registrering buffers can take some time. Normally this isn't an issue as it's done at initialization time (and hence less critical), but for cases where rings can be created and destroyed as part of an IO thread pool, registering the same buffers for multiple rings become a more time sensitive proposition. As an example, let's say an application has an IO memory pool of 500G. Initial registration takes:

Got 500 huge pages (each 1024MB)
Registered 500 pages in 409 msec

or about 0.4 seconds. If we go higher to 900 1GB huge pages being registered:

Registered 900 pages in 738 msec

which is, as expected, a fully linear scaling.

Rather than have each ring pin/map/register the same buffer pool, provide an io_uring_register(2) opcode to simply duplicate the buffers that are registered with another ring. Adding the same 900GB of registered buffers to the target ring can then be accomplished in:

Quoted 4 lines

Copied 900 pages in 17 usec

While timing differs a bit, this provides around a 25,000-40,000x speedup for this use case.

Looks good, but I couldn't get it to apply on top of your branches. I have only one comment, if you are doing a v4:

Quoted 23 lines
---
 include/uapi/linux/io_uring.h | 13 +++++
 io_uring/register.c           |  6 +++
 io_uring/rsrc.c               | 91 +++++++++++++++++++++++++++++++++++
 io_uring/rsrc.h               |  1 +
 4 files changed, 111 insertions(+)
io_uring/rsrc.c +1 -0
@@ -17,6 +17,7 @@
 #include "openclose.h"
 #include "rsrc.h"
 #include "memmap.h"
+#include "register.h"
 
 struct io_rsrc_update {
 	struct file			*file;
@@ -1137,3 +1138,93 @@ int io_import_fixed(int ddir, struct iov_iter *iter,
 
Quoted 9 lines

io_uring/rsrc.c int io_import_fixed(int ddir, struct iov_iter *iter, · line 1142

Quoted 4 lines
io_uring/rsrc.c +2 -0
 	return 0;
 }
+
+static int io_copy_buffers(struct io_ring_ctx *ctx, struct io_ring_ctx *src_ctx)

The error handling code in this function is a bit hairy, IMO. I think if you check nbufs unlocked and validate it later, it could be much simpler:

Sorry missed this due to travel - this is upstream in this merge window. If you want to send a cleanup against for-6.12/io_uring, then please do!

Signature · 1 line
Jens Axboe