fio: fix aio trim completion latencies

PATCH In review [email protected] 2 participants 2 messages Jul 8, 2019 – Jul 9, 2019

lore

Share & tools

lore

[email protected]

author

2019-07-08 19:23 UTC #0 lore raw

PATCH 1/1 fio: fix aio trim completion latencies

From: Vincent Fu <[email protected]>

The io_uring, libaio, and posixaio ioengines actually carry out synchronous trim operations, but latency timestamps are recorded as if the trims were issued asynchronously. This patch fixes how timestamps are recorded for trim operations issued by these ioengines.

Fixes https://github.com/axboe/fio/issues/764

---
 engines/io_uring.c | 1 +
 engines/libaio.c   | 1 +
 engines/posixaio.c | 1 +
 ioengines.c        | 8 ++++++--
 ioengines.h        | 2 ++
 5 files changed, 11 insertions(+), 2 deletions(-)

engines/io_uring.c +1 -0

index a5e77d8f..9bcfec17 100644
@@ -533,6 +533,7 @@ static int fio_ioring_io_u_init(struct thread_data *td, struct io_u *io_u)
 static struct ioengine_ops ioengine = {
 	.name			= "io_uring",
 	.version		= FIO_IOOPS_VERSION,
+	.flags			= FIO_ASYNCIO_SYNC_TRIM,
 	.init			= fio_ioring_init,
 	.post_init		= fio_ioring_post_init,
 	.io_u_init		= fio_ioring_io_u_init,

engines/libaio.c +1 -0

index 8844ac8b..cc6ca66b 100644
@@ -429,6 +429,7 @@ static int fio_libaio_init(struct thread_data *td)
 static struct ioengine_ops ioengine = {
 	.name			= "libaio",
 	.version		= FIO_IOOPS_VERSION,
+	.flags			= FIO_ASYNCIO_SYNC_TRIM,
 	.init			= fio_libaio_init,
 	.post_init		= fio_libaio_post_init,
 	.prep			= fio_libaio_prep,

engines/posixaio.c +1 -0

index 4ac01957..82c6aa65 100644
@@ -243,6 +243,7 @@ static int fio_posixaio_init(struct thread_data *td)
 static struct ioengine_ops ioengine = {
 	.name		= "posixaio",
 	.version	= FIO_IOOPS_VERSION,
+	.flags		= FIO_ASYNCIO_SYNC_TRIM,
 	.init		= fio_posixaio_init,
 	.prep		= fio_posixaio_prep,
 	.queue		= fio_posixaio_queue,

ioengines.c +6 -2

index 7e5a50cc..aa4ccd27 100644
@@ -308,7 +308,9 @@ enum fio_q_status td_io_queue(struct thread_data *td, struct io_u *io_u)
 	io_u->error = 0;
 	io_u->resid = 0;
 
-	if (td_ioengine_flagged(td, FIO_SYNCIO)) {
+	if (td_ioengine_flagged(td, FIO_SYNCIO) ||
+		(td_ioengine_flagged(td, FIO_ASYNCIO_SYNC_TRIM) && 
+		io_u->ddir == DDIR_TRIM)) {
 		if (fio_fill_issue_time(td))
 			fio_gettime(&io_u->issue_time, NULL);

@@ -389,7 +391,9 @@ enum fio_q_status td_io_queue(struct thread_data *td, struct io_u *io_u)
 			td_io_commit(td);
 	}
 
-	if (!td_ioengine_flagged(td, FIO_SYNCIO)) {
+	if (!td_ioengine_flagged(td, FIO_SYNCIO) &&
+		(!td_ioengine_flagged(td, FIO_ASYNCIO_SYNC_TRIM) ||
+		 io_u->ddir != DDIR_TRIM)) {
 		if (fio_fill_issue_time(td))
 			fio_gettime(&io_u->issue_time, NULL);

ioengines.h +2 -0

index b9cd33d5..01a9b586 100644
@@ -63,6 +63,8 @@ enum fio_ioengine_flags {
 	FIO_FAKEIO	= 1 << 11,	/* engine pretends to do IO */
 	FIO_NOSTATS	= 1 << 12,	/* don't do IO stats */
 	FIO_NOFILEHASH	= 1 << 13,	/* doesn't hash the files for lookup later. */
+	FIO_ASYNCIO_SYNC_TRIM
+			= 1 << 14	/* io engine has async ->queue except for trim */
 };
 
 /*

Signature · 1 line

2.17.1

Jens Axboe

2019-07-09 14:37 UTC #1 lore raw

Quoted 6 lines · On 7/8/19 1:23 PM, [email protected] wrote:

From: Vincent Fu <[email protected]>
The io_uring, libaio, and posixaio ioengines actually carry out synchronous trim operations, but latency timestamps are recorded as if the trims were issued asynchronously. This patch fixes how timestamps are recorded for trim operations issued by these ioengines.

Thanks for fixing this Vincent, applied.

Signature · 1 line

Jens Axboe