Skip to contentSkip to conversation

Add FMODE_NOWAIT support to pipes

PATCH Superseded Jens Axboe 4 participants 13 messages Mar 8, 2023 – Mar 14, 2023

Subsystems (3)

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore

Jens Axboe

author maintainer
#0 lore raw

cover Add FMODE_NOWAIT support to pipes

Hi,

One thing that's always been a bit slower than I'd like with io_uring is dealing with pipes. They don't support IOCB_NOWAIT, and hence we need to punt them to io-wq for handling.

This series adds support for FMODE_NOWAIT to pipes.

Patch 1 extends pipe_buf_operations->confirm() to accept a nonblock parameter, and wires up the caller, pipe_buf_confirm(), to have that argument too.

Patch 2 makes pipes deal with IOCB_NOWAIT for locking the pipe, calling pipe_buf_confirm(), and for allocating new pages on writes.

Patch 3 flicks the switch and enables FMODE_NOWAIT for pipes.

Curious on how big of a difference this makes, I wrote a small benchmark that simply opens 128 pipes and then does 256 rounds of reading and writing to them. This was run 10 times, discarding the first run as it's always a bit slower. Before the patch:

Avg: 262.52 msec
Stdev: 2.12 msec
Min: 261.07 msec
Max 267.91 msec

and after the patch:

Avg: 24.14 msec
Stdev: 9.61 msec
Min: 17.84 msec
Max: 43.75 msec

or about a 10x improvement in performance (and efficiency).

I ran the patches through the ltp pipe and splice tests, no regressions observed. Looking at io_uring traces, we can see that we no longer have any io_uring_queue_async_work() traces after the patch, where previously everything was done via io-wq.

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

PATCH 1/3 fs: add 'nonblock' parameter to pipe_buf_confirm() and fops method

In preparation for being able to do a nonblocking confirm attempt of a pipe buffer, plumb a parameter through the stack to indicate if this is a nonblocking attempt or not.

Each caller is passing down 'false' right now, but the only confirm method in the tree, page_cache_pipe_buf_confirm(), is converted to do a trylock_page() if nonblock == true.

---
 fs/fuse/dev.c             |  4 ++--
 fs/pipe.c                 |  4 ++--
 fs/splice.c               | 11 +++++++----
 include/linux/pipe_fs_i.h |  7 ++++---
 4 files changed, 15 insertions(+), 11 deletions(-)
fs/fuse/dev.c +2 -2
index eb4f88e3dc97..0bd1b0870f2d 100644
@@ -700,7 +700,7 @@ static int fuse_copy_fill(struct fuse_copy_state *cs)
 		struct pipe_buffer *buf = cs->pipebufs;
 
 		if (!cs->write) {
-			err = pipe_buf_confirm(cs->pipe, buf);
+			err = pipe_buf_confirm(cs->pipe, buf, false);
 			if (err)
 				return err;
 
@@ -800,7 +800,7 @@ static int fuse_try_move_page(struct fuse_copy_state *cs, struct page **pagep)
 
 	fuse_copy_finish(cs);
 
-	err = pipe_buf_confirm(cs->pipe, buf);
+	err = pipe_buf_confirm(cs->pipe, buf, false);
 	if (err)
 		goto out_put_old;
 
fs/pipe.c +2 -2
index 42c7ff41c2db..340f253913a2 100644
@@ -297,7 +297,7 @@ pipe_read(struct kiocb *iocb, struct iov_iter *to)
 				chars = total_len;
 			}
 
-			error = pipe_buf_confirm(pipe, buf);
+			error = pipe_buf_confirm(pipe, buf, false);
 			if (error) {
 				if (!ret)
 					ret = error;
@@ -461,7 +461,7 @@ pipe_write(struct kiocb *iocb, struct iov_iter *from)
 
 		if ((buf->flags & PIPE_BUF_FLAG_CAN_MERGE) &&
 		    offset + chars <= PAGE_SIZE) {
-			ret = pipe_buf_confirm(pipe, buf);
+			ret = pipe_buf_confirm(pipe, buf, false);
 			if (ret)
 				goto out;
 
fs/splice.c +7 -4
index 2e76dbb81a8f..5ae6b4a202df 100644
@@ -100,13 +100,16 @@ static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
  * is a page cache page, IO may be in flight.
  */
 static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
-				       struct pipe_buffer *buf)
+				       struct pipe_buffer *buf, bool nonblock)
 {
 	struct page *page = buf->page;
 	int err;
 
 	if (!PageUptodate(page)) {
-		lock_page(page);
+		if (nonblock && !trylock_page(page))
+			return -EAGAIN;
+		else
+			lock_page(page);
 
 		/*
 		 * Page got truncated/unhashed. This will cause a 0-byte
@@ -498,7 +501,7 @@ static int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_des
 		if (sd->len > sd->total_len)
 			sd->len = sd->total_len;
 
-		ret = pipe_buf_confirm(pipe, buf);
+		ret = pipe_buf_confirm(pipe, buf, false);
 		if (unlikely(ret)) {
 			if (ret == -ENODATA)
 				ret = 0;
@@ -761,7 +764,7 @@ iter_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 				continue;
 			this_len = min(this_len, left);
 
-			ret = pipe_buf_confirm(pipe, buf);
+			ret = pipe_buf_confirm(pipe, buf, false);
 			if (unlikely(ret)) {
 				if (ret == -ENODATA)
 					ret = 0;
include/linux/pipe_fs_i.h +4 -3
index d2c3f16cf6b1..d63278bb0797 100644
@@ -100,7 +100,8 @@ struct pipe_buf_operations {
 	 * hook. Returns 0 for good, or a negative error value in case of
 	 * error.  If not present all pages are considered good.
 	 */
-	int (*confirm)(struct pipe_inode_info *, struct pipe_buffer *);
+	int (*confirm)(struct pipe_inode_info *, struct pipe_buffer *,
+			bool nonblock);
 
 	/*
 	 * When the contents of this pipe buffer has been completely
@@ -209,11 +210,11 @@ static inline void pipe_buf_release(struct pipe_inode_info *pipe,
  * @buf:	the buffer to confirm
  */
 static inline int pipe_buf_confirm(struct pipe_inode_info *pipe,
-				   struct pipe_buffer *buf)
+				   struct pipe_buffer *buf, bool nonblock)
 {
 	if (!buf->ops->confirm)
 		return 0;
-	return buf->ops->confirm(pipe, buf);
+	return buf->ops->confirm(pipe, buf, nonblock);
 }
 
 /**
Signature · 1 line
2.39.2
maintainer Reviewed-by on v1 1/3
#2 lore raw

Re: [PATCH 1/3] fs: add 'nonblock' parameter to pipe_buf_confirm() and fops method

Quoted 10 lines · Jens Axboe

In preparation for being able to do a nonblocking confirm attempt of a pipe buffer, plumb a parameter through the stack to indicate if this is a nonblocking attempt or not.

Each caller is passing down 'false' right now, but the only confirm method in the tree, page_cache_pipe_buf_confirm(), is converted to do a trylock_page() if nonblock == true.

---

Looks good,

author maintainer
#3 lore raw

PATCH 2/3 pipe: enable handling of IOCB_NOWAIT

In reply to Jens Axboe

In preparation for enabling FMODE_NOWAIT for pipes, ensure that the read and write path handle it correctly. This includes the pipe locking, page allocation for writes, and confirming pipe buffers.

---
 fs/pipe.c | 34 ++++++++++++++++++++++++++--------
 1 file changed, 26 insertions(+), 8 deletions(-)
fs/pipe.c +26 -8
index 340f253913a2..10366a6cb5b6 100644
@@ -108,6 +108,16 @@ static inline void __pipe_unlock(struct pipe_inode_info *pipe)
 	mutex_unlock(&pipe->mutex);
 }
 
+static inline bool __pipe_trylock(struct pipe_inode_info *pipe, bool nonblock)
+{
+	if (!nonblock) {
+		__pipe_lock(pipe);
+		return true;
+	}
+
+	return mutex_trylock(&pipe->mutex);
+}
+
 void pipe_double_lock(struct pipe_inode_info *pipe1,
 		      struct pipe_inode_info *pipe2)
 {
@@ -234,6 +244,7 @@ pipe_read(struct kiocb *iocb, struct iov_iter *to)
 	struct file *filp = iocb->ki_filp;
 	struct pipe_inode_info *pipe = filp->private_data;
 	bool was_full, wake_next_reader = false;
+	const bool nonblock = iocb->ki_flags & IOCB_NOWAIT;
 	ssize_t ret;
 
 	/* Null read succeeds. */
@@ -241,7 +252,8 @@ pipe_read(struct kiocb *iocb, struct iov_iter *to)
 		return 0;
 
 	ret = 0;
-	__pipe_lock(pipe);
+	if (!__pipe_trylock(pipe, nonblock))
+		return -EAGAIN;
 
 	/*
 	 * We only wake up writers if the pipe was full when we started
@@ -297,7 +309,7 @@ pipe_read(struct kiocb *iocb, struct iov_iter *to)
 				chars = total_len;
 			}
 
-			error = pipe_buf_confirm(pipe, buf, false);
+			error = pipe_buf_confirm(pipe, buf, nonblock);
 			if (error) {
 				if (!ret)
 					ret = error;
@@ -342,7 +354,7 @@ pipe_read(struct kiocb *iocb, struct iov_iter *to)
 			break;
 		if (ret)
 			break;
-		if (filp->f_flags & O_NONBLOCK) {
+		if (filp->f_flags & O_NONBLOCK || nonblock) {
 			ret = -EAGAIN;
 			break;
 		}
@@ -423,12 +435,14 @@ pipe_write(struct kiocb *iocb, struct iov_iter *from)
 	ssize_t chars;
 	bool was_empty = false;
 	bool wake_next_writer = false;
+	const bool nonblock = iocb->ki_flags & IOCB_NOWAIT;
 
 	/* Null write succeeds. */
 	if (unlikely(total_len == 0))
 		return 0;
 
-	__pipe_lock(pipe);
+	if (!__pipe_trylock(pipe, nonblock))
+		return -EAGAIN;
 
 	if (!pipe->readers) {
 		send_sig(SIGPIPE, current, 0);
@@ -461,7 +475,7 @@ pipe_write(struct kiocb *iocb, struct iov_iter *from)
 
 		if ((buf->flags & PIPE_BUF_FLAG_CAN_MERGE) &&
 		    offset + chars <= PAGE_SIZE) {
-			ret = pipe_buf_confirm(pipe, buf, false);
+			ret = pipe_buf_confirm(pipe, buf, nonblock);
 			if (ret)
 				goto out;
 
@@ -493,9 +507,13 @@ pipe_write(struct kiocb *iocb, struct iov_iter *from)
 			int copied;
 
 			if (!page) {
-				page = alloc_page(GFP_HIGHUSER | __GFP_ACCOUNT);
+				gfp_t gfp = __GFP_HIGHMEM | __GFP_ACCOUNT;
+
+				if (!nonblock)
+					gfp |= GFP_USER;
+				page = alloc_page(gfp);
 				if (unlikely(!page)) {
-					ret = ret ? : -ENOMEM;
+					ret = ret ? : nonblock ? -EAGAIN : -ENOMEM;
 					break;
 				}
 				pipe->tmp_page = page;
@@ -547,7 +565,7 @@ pipe_write(struct kiocb *iocb, struct iov_iter *from)
 			continue;
 
 		/* Wait for buffer space to become available. */
-		if (filp->f_flags & O_NONBLOCK) {
+		if (filp->f_flags & O_NONBLOCK || nonblock) {
 			if (!ret)
 				ret = -EAGAIN;
 			break;
Signature · 1 line
2.39.2
#4 lore raw

Re: [PATCH 2/3] pipe: enable handling of IOCB_NOWAIT

Quoted 92 lines · Jens Axboe

In preparation for enabling FMODE_NOWAIT for pipes, ensure that the read and write path handle it correctly. This includes the pipe locking, page allocation for writes, and confirming pipe buffers.

---
 fs/pipe.c | 34 ++++++++++++++++++++++++++--------
 1 file changed, 26 insertions(+), 8 deletions(-)
fs/pipe.c +19 -5
index 340f253913a2..10366a6cb5b6 100644
@@ -108,6 +108,16 @@ static inline void __pipe_unlock(struct pipe_inode_info *pipe)
 	mutex_unlock(&pipe->mutex);
 }
 
+static inline bool __pipe_trylock(struct pipe_inode_info *pipe, bool nonblock)
+{
+	if (!nonblock) {
+		__pipe_lock(pipe);
+		return true;
+	}
+
+	return mutex_trylock(&pipe->mutex);
+}
+
 void pipe_double_lock(struct pipe_inode_info *pipe1,
 		      struct pipe_inode_info *pipe2)
 {
@@ -234,6 +244,7 @@ pipe_read(struct kiocb *iocb, struct iov_iter *to)
 	struct file *filp = iocb->ki_filp;
 	struct pipe_inode_info *pipe = filp->private_data;
 	bool was_full, wake_next_reader = false;
+	const bool nonblock = iocb->ki_flags & IOCB_NOWAIT;
 	ssize_t ret;
 
 	/* Null read succeeds. */
@@ -241,7 +252,8 @@ pipe_read(struct kiocb *iocb, struct iov_iter *to)
 		return 0;
 
 	ret = 0;
-	__pipe_lock(pipe);
+	if (!__pipe_trylock(pipe, nonblock))
+		return -EAGAIN;
 
 	/*
 	 * We only wake up writers if the pipe was full when we started
@@ -297,7 +309,7 @@ pipe_read(struct kiocb *iocb, struct iov_iter *to)
 				chars = total_len;
 			}
 
-			error = pipe_buf_confirm(pipe, buf, false);
+			error = pipe_buf_confirm(pipe, buf, nonblock);
 			if (error) {
 				if (!ret)
 					ret = error;
@@ -342,7 +354,7 @@ pipe_read(struct kiocb *iocb, struct iov_iter *to)
 			break;
 		if (ret)
 			break;
-		if (filp->f_flags & O_NONBLOCK) {
+		if (filp->f_flags & O_NONBLOCK || nonblock) {
 			ret = -EAGAIN;
 			break;
 		}
@@ -423,12 +435,14 @@ pipe_write(struct kiocb *iocb, struct iov_iter *from)
 	ssize_t chars;
 	bool was_empty = false;
 	bool wake_next_writer = false;
+	const bool nonblock = iocb->ki_flags & IOCB_NOWAIT;
 
 	/* Null write succeeds. */
 	if (unlikely(total_len == 0))
 		return 0;
 
-	__pipe_lock(pipe);
+	if (!__pipe_trylock(pipe, nonblock))
+		return -EAGAIN;
 
 	if (!pipe->readers) {
 		send_sig(SIGPIPE, current, 0);
@@ -461,7 +475,7 @@ pipe_write(struct kiocb *iocb, struct iov_iter *from)
 
 		if ((buf->flags & PIPE_BUF_FLAG_CAN_MERGE) &&
 		    offset + chars <= PAGE_SIZE) {
-			ret = pipe_buf_confirm(pipe, buf, false);
+			ret = pipe_buf_confirm(pipe, buf, nonblock);
 			if (ret)
 				goto out;
 

fs/pipe.c pipe_write(struct kiocb *iocb, struct iov_iter *from) · line 513

Quoted 8 lines
fs/pipe.c +4 -1
@@ -493,9 +507,13 @@ pipe_write(struct kiocb *iocb, struct iov_iter *from)
 			int copied;
 
 			if (!page) {
-				page = alloc_page(GFP_HIGHUSER | __GFP_ACCOUNT);
+				gfp_t gfp = __GFP_HIGHMEM | __GFP_ACCOUNT;
+
+				if (!nonblock)
+					gfp |= GFP_USER;

Just for my education: Does this encode the assumpation that the non-blocking code can only be reached from io_uring and thus GFP_USER can be dropped for that case? IOW, if there's other code that could in the future reach the non blocking condition would this still be correct?

fs/pipe.c pipe_write(struct kiocb *iocb, struct iov_iter *from) · line 516

Quoted 4 lines
fs/pipe.c +2 -1
+				page = alloc_page(gfp);
 				if (unlikely(!page)) {
-					ret = ret ? : -ENOMEM;
+					ret = ret ? : nonblock ? -EAGAIN : -ENOMEM;

Hm, could we try and avoid the nested "?:?:" please. Imho, that's easy to misparse. Idk, doesn't need to be exactly that code but sm like:

if (!nonblock) {

   					gfp |= GFP_USER;
					ret = -EAGAIN;
				} else {
					ret = -ENOMEM;
				}

page = alloc_page(gfp); if (unlikely(!page))

					break;
				else
					ret = 0;

pipe->tmp_page = page;

or sm else.

author maintainer
#5 lore raw

Re: [PATCH 2/3] pipe: enable handling of IOCB_NOWAIT

Quoted 4 lines · Christian Brauner
Quoted 3 lines · Jens Axboe
fs/pipe.c +0 -0
@@ -493,9 +507,13 @@ pipe_write(struct kiocb *iocb, struct iov_iter *from)
 			int copied;
 
Quoted 11 lines

fs/pipe.c pipe_write(struct kiocb *iocb, struct iov_iter *from) · line 513

Quoted 6 lines
fs/pipe.c +4 -1
 			if (!page) {
-				page = alloc_page(GFP_HIGHUSER | __GFP_ACCOUNT);
+				gfp_t gfp = __GFP_HIGHMEM | __GFP_ACCOUNT;
+
+				if (!nonblock)
+					gfp |= GFP_USER;

Just for my education: Does this encode the assumpation that the non-blocking code can only be reached from io_uring and thus GFP_USER can be dropped for that case? IOW, if there's other code that could in the future reach the non blocking condition would this still be correct?

You can already reach that if you do preadv2(..., RWF_NOWAIT). There should be no assumptions here on the user of it, semantics should be the same. The gfp mask is just split so we avoid __GFP_WAIT for the nonblocking case.

Quoted 23 lines

fs/pipe.c pipe_write(struct kiocb *iocb, struct iov_iter *from) · line 516

Quoted 5 lines
fs/pipe.c +2 -1
+				page = alloc_page(gfp);
 				if (unlikely(!page)) {
-					ret = ret ? : -ENOMEM;
+					ret = ret ? : nonblock ? -EAGAIN : -ENOMEM;

Hm, could we try and avoid the nested "?:?:" please. Imho, that's easy to misparse. Idk, doesn't need to be exactly that code but sm like:

   				if (!nonblock) {
   					gfp |= GFP_USER;
					ret = -EAGAIN;
				} else {
					ret = -ENOMEM;
				}

   				page = alloc_page(gfp);
   				if (unlikely(!page))
					break;
				else
					ret = 0;
   				pipe->tmp_page = page;
Quoted 1 line

or sm else.

Yeah this is much better, I think I was a bit too lazy here, not really a fan of ternaries myself... I'll fix that up. Thanks!

Signature · 1 line
Jens Axboe
#6 lore raw

Re: [PATCH 2/3] pipe: enable handling of IOCB_NOWAIT

In reply to Jens Axboe

fs/pipe.c pipe_write(struct kiocb *iocb, struct iov_iter *from) · line 514

Quoted 10 lines · Jens Axboe
fs/pipe.c +5 -1
@@ -493,9 +507,13 @@ pipe_write(struct kiocb *iocb, struct iov_iter *from)
 			int copied;
 
 			if (!page) {
-				page = alloc_page(GFP_HIGHUSER | __GFP_ACCOUNT);
+				gfp_t gfp = __GFP_HIGHMEM | __GFP_ACCOUNT;
+
+				if (!nonblock)
+					gfp |= GFP_USER;
+				page = alloc_page(gfp);

Hmm, looks like you drop __GFP_HARDWALL in the nonblock case. People who use cpusets might be annoyed by that.

fs/pipe.c pipe_write(struct kiocb *iocb, struct iov_iter *from) · line 516

Quoted 3 lines
fs/pipe.c +1 -1
 				if (unlikely(!page)) {
-					ret = ret ? : -ENOMEM;
+					ret = ret ? : nonblock ? -EAGAIN : -ENOMEM;

double ternary operator? really?

					if (!ret)
						ret = nonblock ? -EAGAIN : -ENOMEM;
author maintainer
#7 lore raw

Re: [PATCH 2/3] pipe: enable handling of IOCB_NOWAIT

Quoted 4 lines · Matthew Wilcox
Quoted 3 lines · Jens Axboe
fs/pipe.c +0 -0
@@ -493,9 +507,13 @@ pipe_write(struct kiocb *iocb, struct iov_iter *from)
 			int copied;
 
Quoted 10 lines

fs/pipe.c pipe_write(struct kiocb *iocb, struct iov_iter *from) · line 514

Quoted 7 lines
fs/pipe.c +5 -1
 			if (!page) {
-				page = alloc_page(GFP_HIGHUSER | __GFP_ACCOUNT);
+				gfp_t gfp = __GFP_HIGHMEM | __GFP_ACCOUNT;
+
+				if (!nonblock)
+					gfp |= GFP_USER;
+				page = alloc_page(gfp);

Hmm, looks like you drop __GFP_HARDWALL in the nonblock case. People who use cpusets might be annoyed by that.

Ah good catch! Yes, that's an oversight, I'll rectify that in v2.

Quoted 5 lines

fs/pipe.c pipe_write(struct kiocb *iocb, struct iov_iter *from) · line 516

Quoted 3 lines
fs/pipe.c +1 -1
 				if (unlikely(!page)) {
-					ret = ret ? : -ENOMEM;
+					ret = ret ? : nonblock ? -EAGAIN : -ENOMEM;

double ternary operator? really?

yeah sorry... See reply to Christian, I'll make this cleaner.

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

PATCH 3/3 pipe: set FMODE_NOWAIT on pipes

In reply to Jens Axboe

The read/write path is now prepared to deal with IOCB_NOWAIT, hence enable support for that via setting FMODE_NOWAIT on new pipes.

---
 fs/pipe.c | 3 +++
 1 file changed, 3 insertions(+)
fs/pipe.c +3 -0
index 10366a6cb5b6..9db274f9baa5 100644
@@ -994,6 +994,9 @@ static int __do_pipe_flags(int *fd, struct file **files, int flags)
 	audit_fd_pair(fdr, fdw);
 	fd[0] = fdr;
 	fd[1] = fdw;
+	/* pipe groks IOCB_NOWAIT */
+	files[0]->f_mode |= FMODE_NOWAIT;
+	files[1]->f_mode |= FMODE_NOWAIT;
 	return 0;
 
  err_fdr:
Signature · 1 line
2.39.2
author maintainer
#10 lore raw
In reply to Jens Axboe
Quoted 10 lines · Jens Axboe

Curious on how big of a difference this makes, I wrote a small benchmark that simply opens 128 pipes and then does 256 rounds of reading and writing to them. This was run 10 times, discarding the first run as it's always a bit slower. Before the patch:

Avg: 262.52 msec
Stdev: 2.12 msec
Min: 261.07 msec
Max 267.91 msec

Quoted 8 lines

and after the patch:

Avg: 24.14 msec
Stdev: 9.61 msec
Min: 17.84 msec
Max: 43.75 msec

or about a 10x improvement in performance (and efficiency).

The above test was for a pipe being empty when the read is issued, if the test is changed to have data when, then it looks even better:

Before:

Avg: 249.24 msec
Stdev: 0.20 msec
Min: 248.96 msec
Max: 249.53 msec

After:

Avg: 10.86 msec
Stdev: 0.91 msec
Min: 10.02 msec
Max: 12.67 msec

or about a 23x improvement.

Signature · 1 line
Jens Axboe
Acked-by
#11 lore raw
Quoted 30 lines · Jens Axboe
Quoted 10 lines · Jens Axboe

Curious on how big of a difference this makes, I wrote a small benchmark that simply opens 128 pipes and then does 256 rounds of reading and writing to them. This was run 10 times, discarding the first run as it's always a bit slower. Before the patch:

Avg: 262.52 msec
Stdev: 2.12 msec
Min: 261.07 msec
Max 267.91 msec

Quoted 8 lines

and after the patch:

Avg: 24.14 msec
Stdev: 9.61 msec
Min: 17.84 msec
Max: 43.75 msec

or about a 10x improvement in performance (and efficiency).

The above test was for a pipe being empty when the read is issued, if the test is changed to have data when, then it looks even better:

Before:

Avg: 249.24 msec
Stdev: 0.20 msec
Min: 248.96 msec
Max: 249.53 msec

Quoted 8 lines

After:

Avg: 10.86 msec
Stdev: 0.91 msec
Min: 10.02 msec
Max: 12.67 msec

or about a 23x improvement.

Nice!

Code looks OK, maybe consider s/nonblock/nowait/, but I'm not a pipe expert so I'll leave nitty gritty details to Al, et al.

Signature · 2 lines
Dave Chinner
[email protected]
author maintainer
#12 lore raw
Quoted 38 lines · Dave Chinner
Quoted 37 lines · Jens Axboe
Quoted 10 lines · Jens Axboe

Curious on how big of a difference this makes, I wrote a small benchmark that simply opens 128 pipes and then does 256 rounds of reading and writing to them. This was run 10 times, discarding the first run as it's always a bit slower. Before the patch:

Avg: 262.52 msec
Stdev: 2.12 msec
Min: 261.07 msec
Max 267.91 msec

Quoted 8 lines

and after the patch:

Avg: 24.14 msec
Stdev: 9.61 msec
Min: 17.84 msec
Max: 43.75 msec

or about a 10x improvement in performance (and efficiency).

The above test was for a pipe being empty when the read is issued, if the test is changed to have data when, then it looks even better:

Before:

Avg: 249.24 msec
Stdev: 0.20 msec
Min: 248.96 msec
Max: 249.53 msec

After:

Avg: 10.86 msec
Stdev: 0.91 msec
Min: 10.02 msec
Max: 12.67 msec

Quoted 6 lines
Quoted 1 line

or about a 23x improvement.

Nice!

Code looks OK, maybe consider s/nonblock/nowait/, but I'm not a pipe expert so I'll leave nitty gritty details to Al, et al.

We seem to use both somewhat interchangably throughout the kernel. Don't feel strongly about that one, so I'll let the majority speak on what they prefer.

Quoted 1 line

Thanks, added.

Signature · 1 line
Jens Axboe