Skip to contentSkip to conversation

block: disable iopoll for split bio

PATCH Superseded Jeffle Xu 3 participants 9 messages Nov 25, 2020 Block layer

v6 of 9 Superseded Latest is v9 →

lore
Share & tools
lore

Jeffle Xu

author
#0 lore raw

PATCH v6 1/1 block: disable iopoll for split bio

iopoll is initially for small size, latency sensitive IO. It doesn't work well for big IO, especially when it needs to be split to multiple bios. In this case, the returned cookie of __submit_bio_noacct_mq() is indeed the cookie of the last split bio. The completion of *this* last split bio done by iopoll doesn't mean the whole original bio has completed. Callers of iopoll still need to wait for completion of other split bios.

Besides bio splitting may cause more trouble for iopoll which isn't supposed to be used in case of big IO.

iopoll for split bio may cause potential race if CPU migration happens during bio submission. Since the returned cookie is that of the last split bio, polling on the corresponding hardware queue doesn't help complete other split bios, if these split bios are enqueued into different hardware queues. Since interrupts are disabled for polling queues, the completion of these other split bios depends on timeout mechanism, thus causing a potential hang.

iopoll for split bio may also cause hang for sync polling. Currently both the blkdev and iomap-based fs (ext4/xfs, etc) support sync polling in direct IO routine. These routines will submit bio without REQ_NOWAIT flag set, and then start sync polling in current process context. The process may hang in blk_mq_get_tag() if the submitted bio has to be split into multiple bios and can rapidly exhaust the queue depth. The process are waiting for the completion of the previously allocated requests, which should be reaped by the following polling, and thus causing a deadlock.

To avoid these subtle trouble described above, just disable iopoll for split bio.

---
 block/bio.c               |  2 ++
 block/blk-merge.c         | 12 ++++++++++++
 block/blk-mq.c            |  3 +++
 include/linux/blk_types.h |  1 +
 4 files changed, 18 insertions(+)
block/bio.c +2 -0
index fa01bef35bb1..7f7ddc22a30d 100644
@@ -684,6 +684,8 @@ void __bio_clone_fast(struct bio *bio, struct bio *bio_src)
 	bio_set_flag(bio, BIO_CLONED);
 	if (bio_flagged(bio_src, BIO_THROTTLED))
 		bio_set_flag(bio, BIO_THROTTLED);
+	if (bio_flagged(bio_src, BIO_SPLIT))
+		bio_set_flag(bio, BIO_SPLIT);
 	bio->bi_opf = bio_src->bi_opf;
 	bio->bi_ioprio = bio_src->bi_ioprio;
 	bio->bi_write_hint = bio_src->bi_write_hint;
block/blk-merge.c +12 -0
index bcf5e4580603..a2890cebf99f 100644
@@ -279,6 +279,18 @@ static struct bio *blk_bio_segment_split(struct request_queue *q,
 	return NULL;
 split:
 	*segs = nsegs;
+
+	/*
+	 * Bio splitting may cause subtle trouble such as hang when doing sync
+	 * iopoll in direct IO routine. Given performance gain of iopoll for
+	 * big IO can be trival, disable iopoll when split needed. We need
+	 * BIO_SPLIT to identify bios need this workaround. Since currently
+	 * only normal IO under mq routine may suffer this issue, BIO_SPLIT is
+	 * only marked here.
+	 */
+	bio->bi_opf &= ~REQ_HIPRI;
+	bio_set_flag(bio, BIO_SPLIT);
+
 	return bio_split(bio, sectors, GFP_NOIO, bs);
 }
 
block/blk-mq.c +3 -0
index 55bcee5dc032..ce1f3628e4c2 100644
@@ -2265,6 +2265,9 @@ blk_qc_t blk_mq_submit_bio(struct bio *bio)
 		blk_mq_sched_insert_request(rq, false, true, true);
 	}
 
+	if (bio_flagged(bio, BIO_SPLIT))
+		return BLK_QC_T_NONE;
+
 	return cookie;
 queue_exit:
 	blk_queue_exit(q);
include/linux/blk_types.h +1 -0
index d9b69bbde5cc..528054c5e0bd 100644
@@ -284,6 +284,7 @@ enum {
 				 * of this bio. */
 	BIO_CGROUP_ACCT,	/* has been accounted to a cgroup */
 	BIO_TRACKED,		/* set if bio goes through the rq_qos path */
+	BIO_SPLIT,		/* bio got split because of queue limits*/
 	BIO_FLAG_LAST
 };
 
Signature · 1 line
2.27.0
#1 lore raw
Quoted 83 lines · Jeffle Xu

iopoll is initially for small size, latency sensitive IO. It doesn't work well for big IO, especially when it needs to be split to multiple bios. In this case, the returned cookie of __submit_bio_noacct_mq() is indeed the cookie of the last split bio. The completion of *this* last split bio done by iopoll doesn't mean the whole original bio has completed. Callers of iopoll still need to wait for completion of other split bios.

Besides bio splitting may cause more trouble for iopoll which isn't supposed to be used in case of big IO.

iopoll for split bio may cause potential race if CPU migration happens during bio submission. Since the returned cookie is that of the last split bio, polling on the corresponding hardware queue doesn't help complete other split bios, if these split bios are enqueued into different hardware queues. Since interrupts are disabled for polling queues, the completion of these other split bios depends on timeout mechanism, thus causing a potential hang.

iopoll for split bio may also cause hang for sync polling. Currently both the blkdev and iomap-based fs (ext4/xfs, etc) support sync polling in direct IO routine. These routines will submit bio without REQ_NOWAIT flag set, and then start sync polling in current process context. The process may hang in blk_mq_get_tag() if the submitted bio has to be split into multiple bios and can rapidly exhaust the queue depth. The process are waiting for the completion of the previously allocated requests, which should be reaped by the following polling, and thus causing a deadlock.

To avoid these subtle trouble described above, just disable iopoll for split bio.

---
 block/bio.c               |  2 ++
 block/blk-merge.c         | 12 ++++++++++++
 block/blk-mq.c            |  3 +++
 include/linux/blk_types.h |  1 +
 4 files changed, 18 insertions(+)
block/bio.c +2 -0
index fa01bef35bb1..7f7ddc22a30d 100644
@@ -684,6 +684,8 @@ void __bio_clone_fast(struct bio *bio, struct bio *bio_src)
 	bio_set_flag(bio, BIO_CLONED);
 	if (bio_flagged(bio_src, BIO_THROTTLED))
 		bio_set_flag(bio, BIO_THROTTLED);
+	if (bio_flagged(bio_src, BIO_SPLIT))
+		bio_set_flag(bio, BIO_SPLIT);
 	bio->bi_opf = bio_src->bi_opf;
 	bio->bi_ioprio = bio_src->bi_ioprio;
 	bio->bi_write_hint = bio_src->bi_write_hint;
block/blk-merge.c +12 -0
index bcf5e4580603..a2890cebf99f 100644
@@ -279,6 +279,18 @@ static struct bio *blk_bio_segment_split(struct request_queue *q,
 	return NULL;
 split:
 	*segs = nsegs;
+
+	/*
+	 * Bio splitting may cause subtle trouble such as hang when doing sync
+	 * iopoll in direct IO routine. Given performance gain of iopoll for
+	 * big IO can be trival, disable iopoll when split needed. We need
+	 * BIO_SPLIT to identify bios need this workaround. Since currently
+	 * only normal IO under mq routine may suffer this issue, BIO_SPLIT is
+	 * only marked here.
+	 */
+	bio->bi_opf &= ~REQ_HIPRI;
+	bio_set_flag(bio, BIO_SPLIT);
+
 	return bio_split(bio, sectors, GFP_NOIO, bs);
 }
 
block/blk-mq.c +0 -0
index 55bcee5dc032..ce1f3628e4c2 100644

block/blk-mq.c blk_qc_t blk_mq_submit_bio(struct bio *bio) · line 2270

Quoted 6 lines
block/blk-mq.c +3 -0
@@ -2265,6 +2265,9 @@ blk_qc_t blk_mq_submit_bio(struct bio *bio)
 		blk_mq_sched_insert_request(rq, false, true, true);
 	}
 
+	if (bio_flagged(bio, BIO_SPLIT))
+		return BLK_QC_T_NONE;
+

Not sure the new bio flag is really required for this case, just wondering why not take the following simple way? BTW we are really going to run out of bio flag.

block/blk-mq.c +5 -0
index 55bcee5dc032..1139b1efd712 100644
@@ -2157,6 +2157,7 @@ blk_qc_t blk_mq_submit_bio(struct bio *bio)
 	unsigned int nr_segs;
 	blk_qc_t cookie;
 	blk_status_t ret;
+	struct bio *orig_bio = bio;
 
 	blk_queue_bounce(q, &bio);
 	__blk_queue_split(&bio, &nr_segs);
@@ -2265,6 +2266,10 @@ blk_qc_t blk_mq_submit_bio(struct bio *bio)
 		blk_mq_sched_insert_request(rq, false, true, true);
 	}
 
+	/* don't poll splitted bio */
+	if (orig_bio != bio)
+		return BLK_QC_T_NONE;
+
 	return cookie;
 queue_exit:
 	blk_queue_exit(q);

Thanks,
Ming

#2 lore raw
Quoted 17 lines · Ming Lei

Not sure the new bio flag is really required for this case, just wondering why not take the following simple way? BTW we are really going to run out of bio flag.

block/blk-mq.c +1 -0
index 55bcee5dc032..1139b1efd712 100644
@@ -2157,6 +2157,7 @@ blk_qc_t blk_mq_submit_bio(struct bio *bio)
 	unsigned int nr_segs;
 	blk_qc_t cookie;
 	blk_status_t ret;
+	struct bio *orig_bio = bio;
 
 	blk_queue_bounce(q, &bio);
 	__blk_queue_split(&bio, &nr_segs);

block/blk-mq.c blk_qc_t blk_mq_submit_bio(struct bio *bio) · line 2275

Quoted 10 lines
block/blk-mq.c +4 -0
@@ -2265,6 +2266,10 @@ blk_qc_t blk_mq_submit_bio(struct bio *bio)
 		blk_mq_sched_insert_request(rq, false, true, true);
 	}
 
+	/* don't poll splitted bio */
+	if (orig_bio != bio)
+		return BLK_QC_T_NONE;
+
 	return cookie;
 queue_exit:
 	blk_queue_exit(q);

Yeah, this looks simpler and more obvious, and it also catches the bounce buffering case (not that polling + bounce buffering should ever happen).

author
#3 lore raw
In reply to Ming Lei
Quoted 87 lines · Ming Lei
Quoted 86 lines · Jeffle Xu

iopoll is initially for small size, latency sensitive IO. It doesn't work well for big IO, especially when it needs to be split to multiple bios. In this case, the returned cookie of __submit_bio_noacct_mq() is indeed the cookie of the last split bio. The completion of *this* last split bio done by iopoll doesn't mean the whole original bio has completed. Callers of iopoll still need to wait for completion of other split bios.

Besides bio splitting may cause more trouble for iopoll which isn't supposed to be used in case of big IO.

iopoll for split bio may cause potential race if CPU migration happens during bio submission. Since the returned cookie is that of the last split bio, polling on the corresponding hardware queue doesn't help complete other split bios, if these split bios are enqueued into different hardware queues. Since interrupts are disabled for polling queues, the completion of these other split bios depends on timeout mechanism, thus causing a potential hang.

iopoll for split bio may also cause hang for sync polling. Currently both the blkdev and iomap-based fs (ext4/xfs, etc) support sync polling in direct IO routine. These routines will submit bio without REQ_NOWAIT flag set, and then start sync polling in current process context. The process may hang in blk_mq_get_tag() if the submitted bio has to be split into multiple bios and can rapidly exhaust the queue depth. The process are waiting for the completion of the previously allocated requests, which should be reaped by the following polling, and thus causing a deadlock.

To avoid these subtle trouble described above, just disable iopoll for split bio.

---
 block/bio.c               |  2 ++
 block/blk-merge.c         | 12 ++++++++++++
 block/blk-mq.c            |  3 +++
 include/linux/blk_types.h |  1 +
 4 files changed, 18 insertions(+)
block/bio.c +2 -0
index fa01bef35bb1..7f7ddc22a30d 100644
@@ -684,6 +684,8 @@ void __bio_clone_fast(struct bio *bio, struct bio *bio_src)
 	bio_set_flag(bio, BIO_CLONED);
 	if (bio_flagged(bio_src, BIO_THROTTLED))
 		bio_set_flag(bio, BIO_THROTTLED);
+	if (bio_flagged(bio_src, BIO_SPLIT))
+		bio_set_flag(bio, BIO_SPLIT);
 	bio->bi_opf = bio_src->bi_opf;
 	bio->bi_ioprio = bio_src->bi_ioprio;
 	bio->bi_write_hint = bio_src->bi_write_hint;
block/blk-merge.c +12 -0
index bcf5e4580603..a2890cebf99f 100644
@@ -279,6 +279,18 @@ static struct bio *blk_bio_segment_split(struct request_queue *q,
 	return NULL;
 split:
 	*segs = nsegs;
+
+	/*
+	 * Bio splitting may cause subtle trouble such as hang when doing sync
+	 * iopoll in direct IO routine. Given performance gain of iopoll for
+	 * big IO can be trival, disable iopoll when split needed. We need
+	 * BIO_SPLIT to identify bios need this workaround. Since currently
+	 * only normal IO under mq routine may suffer this issue, BIO_SPLIT is
+	 * only marked here.
+	 */
+	bio->bi_opf &= ~REQ_HIPRI;
+	bio_set_flag(bio, BIO_SPLIT);
+
 	return bio_split(bio, sectors, GFP_NOIO, bs);
 }
 
block/blk-mq.c +0 -0
index 55bcee5dc032..ce1f3628e4c2 100644
@@ -2265,6 +2265,9 @@ blk_qc_t blk_mq_submit_bio(struct bio *bio)
 		blk_mq_sched_insert_request(rq, false, true, true);
 	}
 
Quoted 8 lines

block/blk-mq.c blk_qc_t blk_mq_submit_bio(struct bio *bio) · line 2272

Quoted 3 lines
block/blk-mq.c +3 -0
+	if (bio_flagged(bio, BIO_SPLIT))
+		return BLK_QC_T_NONE;
+

Not sure the new bio flag is really required for this case, just wondering why not take the following simple way? BTW we are really going to run out of bio flag.

Please consider the following case:

One big bio got split into two split bios. At the first call of blk_mq_submit_bio(), the input @bio (actually the original big bio) indeed gets split. The split bio gets enqueued to hw queue and the returned cookie is BLK_QC_T_NONE, while the remained bio gets buffered in bio_list. So far so good.

Then when calling blk_mq_submit_bio() the second time, the input @bio is indeed the remained bio. At this time, it will not get split and you will get a *valid* cookie. And since the cookie of last split bio will actually overrides the previous cookie, you will get a *valid* cookie as a result.

Quoted 27 lines
block/blk-mq.c +5 -0
index 55bcee5dc032..1139b1efd712 100644
@@ -2157,6 +2157,7 @@ blk_qc_t blk_mq_submit_bio(struct bio *bio)
 	unsigned int nr_segs;
 	blk_qc_t cookie;
 	blk_status_t ret;
+	struct bio *orig_bio = bio;
 
 	blk_queue_bounce(q, &bio);
 	__blk_queue_split(&bio, &nr_segs);
@@ -2265,6 +2266,10 @@ blk_qc_t blk_mq_submit_bio(struct bio *bio)
 		blk_mq_sched_insert_request(rq, false, true, true);
 	}
 
+	/* don't poll splitted bio */
+	if (orig_bio != bio)
+		return BLK_QC_T_NONE;
+
 	return cookie;
 queue_exit:
 	blk_queue_exit(q);

Thanks,
Ming

Signature · 2 lines
Thanks,
Jeffle
#4 lore raw
Quoted 99 lines · JeffleXu
Quoted 87 lines · Ming Lei
Quoted 86 lines · Jeffle Xu

iopoll is initially for small size, latency sensitive IO. It doesn't work well for big IO, especially when it needs to be split to multiple bios. In this case, the returned cookie of __submit_bio_noacct_mq() is indeed the cookie of the last split bio. The completion of *this* last split bio done by iopoll doesn't mean the whole original bio has completed. Callers of iopoll still need to wait for completion of other split bios.

Besides bio splitting may cause more trouble for iopoll which isn't supposed to be used in case of big IO.

iopoll for split bio may cause potential race if CPU migration happens during bio submission. Since the returned cookie is that of the last split bio, polling on the corresponding hardware queue doesn't help complete other split bios, if these split bios are enqueued into different hardware queues. Since interrupts are disabled for polling queues, the completion of these other split bios depends on timeout mechanism, thus causing a potential hang.

iopoll for split bio may also cause hang for sync polling. Currently both the blkdev and iomap-based fs (ext4/xfs, etc) support sync polling in direct IO routine. These routines will submit bio without REQ_NOWAIT flag set, and then start sync polling in current process context. The process may hang in blk_mq_get_tag() if the submitted bio has to be split into multiple bios and can rapidly exhaust the queue depth. The process are waiting for the completion of the previously allocated requests, which should be reaped by the following polling, and thus causing a deadlock.

To avoid these subtle trouble described above, just disable iopoll for split bio.

---
 block/bio.c               |  2 ++
 block/blk-merge.c         | 12 ++++++++++++
 block/blk-mq.c            |  3 +++
 include/linux/blk_types.h |  1 +
 4 files changed, 18 insertions(+)
block/bio.c +2 -0
index fa01bef35bb1..7f7ddc22a30d 100644
@@ -684,6 +684,8 @@ void __bio_clone_fast(struct bio *bio, struct bio *bio_src)
 	bio_set_flag(bio, BIO_CLONED);
 	if (bio_flagged(bio_src, BIO_THROTTLED))
 		bio_set_flag(bio, BIO_THROTTLED);
+	if (bio_flagged(bio_src, BIO_SPLIT))
+		bio_set_flag(bio, BIO_SPLIT);
 	bio->bi_opf = bio_src->bi_opf;
 	bio->bi_ioprio = bio_src->bi_ioprio;
 	bio->bi_write_hint = bio_src->bi_write_hint;
block/blk-merge.c +12 -0
index bcf5e4580603..a2890cebf99f 100644
@@ -279,6 +279,18 @@ static struct bio *blk_bio_segment_split(struct request_queue *q,
 	return NULL;
 split:
 	*segs = nsegs;
+
+	/*
+	 * Bio splitting may cause subtle trouble such as hang when doing sync
+	 * iopoll in direct IO routine. Given performance gain of iopoll for
+	 * big IO can be trival, disable iopoll when split needed. We need
+	 * BIO_SPLIT to identify bios need this workaround. Since currently
+	 * only normal IO under mq routine may suffer this issue, BIO_SPLIT is
+	 * only marked here.
+	 */
+	bio->bi_opf &= ~REQ_HIPRI;
+	bio_set_flag(bio, BIO_SPLIT);
+
 	return bio_split(bio, sectors, GFP_NOIO, bs);
 }
 
block/blk-mq.c +0 -0
index 55bcee5dc032..ce1f3628e4c2 100644
@@ -2265,6 +2265,9 @@ blk_qc_t blk_mq_submit_bio(struct bio *bio)
 		blk_mq_sched_insert_request(rq, false, true, true);
 	}
 
Quoted 8 lines

block/blk-mq.c blk_qc_t blk_mq_submit_bio(struct bio *bio) · line 2272

Quoted 3 lines
block/blk-mq.c +3 -0
+	if (bio_flagged(bio, BIO_SPLIT))
+		return BLK_QC_T_NONE;
+

Not sure the new bio flag is really required for this case, just wondering why not take the following simple way? BTW we are really going to run out of bio flag.

Quoted 7 lines

Please consider the following case:

One big bio got split into two split bios. At the first call of blk_mq_submit_bio(), the input @bio (actually the original big bio) indeed gets split. The split bio gets enqueued to hw queue and the returned cookie is BLK_QC_T_NONE, while the remained bio gets buffered in bio_list. So far so good.

When this bio gets splitted, REQ_HIPRI is cleared for this bio, and all splitted bios won't set this flag too.

Quoted 6 lines

Then when calling blk_mq_submit_bio() the second time, the input @bio is indeed the remained bio. At this time, it will not get split and you will get a *valid* cookie. And since the cookie of last split bio will actually overrides the previous cookie, you will get a *valid* cookie as a result.

Then valid cookie can be returned only for bio with REQ_HIPRI.

Thanks,
Ming

author
#5 lore raw
Quoted 102 lines · Ming Lei
Quoted 101 lines · JeffleXu
Quoted 87 lines · Ming Lei
Quoted 86 lines · Jeffle Xu

iopoll is initially for small size, latency sensitive IO. It doesn't work well for big IO, especially when it needs to be split to multiple bios. In this case, the returned cookie of __submit_bio_noacct_mq() is indeed the cookie of the last split bio. The completion of *this* last split bio done by iopoll doesn't mean the whole original bio has completed. Callers of iopoll still need to wait for completion of other split bios.

Besides bio splitting may cause more trouble for iopoll which isn't supposed to be used in case of big IO.

iopoll for split bio may cause potential race if CPU migration happens during bio submission. Since the returned cookie is that of the last split bio, polling on the corresponding hardware queue doesn't help complete other split bios, if these split bios are enqueued into different hardware queues. Since interrupts are disabled for polling queues, the completion of these other split bios depends on timeout mechanism, thus causing a potential hang.

iopoll for split bio may also cause hang for sync polling. Currently both the blkdev and iomap-based fs (ext4/xfs, etc) support sync polling in direct IO routine. These routines will submit bio without REQ_NOWAIT flag set, and then start sync polling in current process context. The process may hang in blk_mq_get_tag() if the submitted bio has to be split into multiple bios and can rapidly exhaust the queue depth. The process are waiting for the completion of the previously allocated requests, which should be reaped by the following polling, and thus causing a deadlock.

To avoid these subtle trouble described above, just disable iopoll for split bio.

---
 block/bio.c               |  2 ++
 block/blk-merge.c         | 12 ++++++++++++
 block/blk-mq.c            |  3 +++
 include/linux/blk_types.h |  1 +
 4 files changed, 18 insertions(+)
block/bio.c +2 -0
index fa01bef35bb1..7f7ddc22a30d 100644
@@ -684,6 +684,8 @@ void __bio_clone_fast(struct bio *bio, struct bio *bio_src)
 	bio_set_flag(bio, BIO_CLONED);
 	if (bio_flagged(bio_src, BIO_THROTTLED))
 		bio_set_flag(bio, BIO_THROTTLED);
+	if (bio_flagged(bio_src, BIO_SPLIT))
+		bio_set_flag(bio, BIO_SPLIT);
 	bio->bi_opf = bio_src->bi_opf;
 	bio->bi_ioprio = bio_src->bi_ioprio;
 	bio->bi_write_hint = bio_src->bi_write_hint;
block/blk-merge.c +12 -0
index bcf5e4580603..a2890cebf99f 100644
@@ -279,6 +279,18 @@ static struct bio *blk_bio_segment_split(struct request_queue *q,
 	return NULL;
 split:
 	*segs = nsegs;
+
+	/*
+	 * Bio splitting may cause subtle trouble such as hang when doing sync
+	 * iopoll in direct IO routine. Given performance gain of iopoll for
+	 * big IO can be trival, disable iopoll when split needed. We need
+	 * BIO_SPLIT to identify bios need this workaround. Since currently
+	 * only normal IO under mq routine may suffer this issue, BIO_SPLIT is
+	 * only marked here.
+	 */
+	bio->bi_opf &= ~REQ_HIPRI;
+	bio_set_flag(bio, BIO_SPLIT);
+
 	return bio_split(bio, sectors, GFP_NOIO, bs);
 }
 
block/blk-mq.c +0 -0
index 55bcee5dc032..ce1f3628e4c2 100644
@@ -2265,6 +2265,9 @@ blk_qc_t blk_mq_submit_bio(struct bio *bio)
 		blk_mq_sched_insert_request(rq, false, true, true);
 	}
 
Quoted 8 lines

block/blk-mq.c blk_qc_t blk_mq_submit_bio(struct bio *bio) · line 2272

Quoted 3 lines
block/blk-mq.c +3 -0
+	if (bio_flagged(bio, BIO_SPLIT))
+		return BLK_QC_T_NONE;
+

Not sure the new bio flag is really required for this case, just wondering why not take the following simple way? BTW we are really going to run out of bio flag.

Please consider the following case:

Quoted 9 lines
Quoted 5 lines

One big bio got split into two split bios. At the first call of blk_mq_submit_bio(), the input @bio (actually the original big bio) indeed gets split. The split bio gets enqueued to hw queue and the returned cookie is BLK_QC_T_NONE, while the remained bio gets buffered in bio_list. So far so good.

When this bio gets splitted, REQ_HIPRI is cleared for this bio, and all splitted bios won't set this flag too.

That's true, at least in terms of my patch.

Quoted 8 lines
Quoted 6 lines

Then when calling blk_mq_submit_bio() the second time, the input @bio is indeed the remained bio. At this time, it will not get split and you will get a *valid* cookie. And since the cookie of last split bio will actually overrides the previous cookie, you will get a *valid* cookie as a result.

Then valid cookie can be returned only for bio with REQ_HIPRI.

Sounds reasonable. The side effect is that original IO without REQ_HIPRI now also returns BLK_QC_T_NONE. But it may not be a big problem, since the returned cookie is not used at all for IO without REQ_HIPRI.

Signature · 2 lines
Thanks,
Jeffle
#6 lore raw
In reply to Jeffle Xu
Quoted 64 lines · Jeffle Xu

iopoll is initially for small size, latency sensitive IO. It doesn't work well for big IO, especially when it needs to be split to multiple bios. In this case, the returned cookie of __submit_bio_noacct_mq() is indeed the cookie of the last split bio. The completion of *this* last split bio done by iopoll doesn't mean the whole original bio has completed. Callers of iopoll still need to wait for completion of other split bios.

Besides bio splitting may cause more trouble for iopoll which isn't supposed to be used in case of big IO.

iopoll for split bio may cause potential race if CPU migration happens during bio submission. Since the returned cookie is that of the last split bio, polling on the corresponding hardware queue doesn't help complete other split bios, if these split bios are enqueued into different hardware queues. Since interrupts are disabled for polling queues, the completion of these other split bios depends on timeout mechanism, thus causing a potential hang.

iopoll for split bio may also cause hang for sync polling. Currently both the blkdev and iomap-based fs (ext4/xfs, etc) support sync polling in direct IO routine. These routines will submit bio without REQ_NOWAIT flag set, and then start sync polling in current process context. The process may hang in blk_mq_get_tag() if the submitted bio has to be split into multiple bios and can rapidly exhaust the queue depth. The process are waiting for the completion of the previously allocated requests, which should be reaped by the following polling, and thus causing a deadlock.

To avoid these subtle trouble described above, just disable iopoll for split bio.

---
 block/bio.c               |  2 ++
 block/blk-merge.c         | 12 ++++++++++++
 block/blk-mq.c            |  3 +++
 include/linux/blk_types.h |  1 +
 4 files changed, 18 insertions(+)
block/bio.c +2 -0
index fa01bef35bb1..7f7ddc22a30d 100644
@@ -684,6 +684,8 @@ void __bio_clone_fast(struct bio *bio, struct bio *bio_src)
 	bio_set_flag(bio, BIO_CLONED);
 	if (bio_flagged(bio_src, BIO_THROTTLED))
 		bio_set_flag(bio, BIO_THROTTLED);
+	if (bio_flagged(bio_src, BIO_SPLIT))
+		bio_set_flag(bio, BIO_SPLIT);
 	bio->bi_opf = bio_src->bi_opf;
 	bio->bi_ioprio = bio_src->bi_ioprio;
 	bio->bi_write_hint = bio_src->bi_write_hint;
block/blk-merge.c +1 -0
index bcf5e4580603..a2890cebf99f 100644
@@ -279,6 +279,18 @@ static struct bio *blk_bio_segment_split(struct request_queue *q,
 	return NULL;
 split:
 	*segs = nsegs;
+

block/blk-merge.c static struct bio *blk_bio_segment_split(struct request_queue *q, · line 292

Quoted 10 lines
block/blk-merge.c +10 -0
+	/*
+	 * Bio splitting may cause subtle trouble such as hang when doing sync
+	 * iopoll in direct IO routine. Given performance gain of iopoll for
+	 * big IO can be trival, disable iopoll when split needed. We need
+	 * BIO_SPLIT to identify bios need this workaround. Since currently
+	 * only normal IO under mq routine may suffer this issue, BIO_SPLIT is
+	 * only marked here.
+	 */
+	bio->bi_opf &= ~REQ_HIPRI;
+	bio_set_flag(bio, BIO_SPLIT);

You may need to put the above into one helper, and call the helper for other splitted cases(discard, write zero and write same) too.

thanks,
Ming

author
#7 lore raw
Quoted 70 lines · Ming Lei
Quoted 69 lines · Jeffle Xu

iopoll is initially for small size, latency sensitive IO. It doesn't work well for big IO, especially when it needs to be split to multiple bios. In this case, the returned cookie of __submit_bio_noacct_mq() is indeed the cookie of the last split bio. The completion of *this* last split bio done by iopoll doesn't mean the whole original bio has completed. Callers of iopoll still need to wait for completion of other split bios.

Besides bio splitting may cause more trouble for iopoll which isn't supposed to be used in case of big IO.

iopoll for split bio may cause potential race if CPU migration happens during bio submission. Since the returned cookie is that of the last split bio, polling on the corresponding hardware queue doesn't help complete other split bios, if these split bios are enqueued into different hardware queues. Since interrupts are disabled for polling queues, the completion of these other split bios depends on timeout mechanism, thus causing a potential hang.

iopoll for split bio may also cause hang for sync polling. Currently both the blkdev and iomap-based fs (ext4/xfs, etc) support sync polling in direct IO routine. These routines will submit bio without REQ_NOWAIT flag set, and then start sync polling in current process context. The process may hang in blk_mq_get_tag() if the submitted bio has to be split into multiple bios and can rapidly exhaust the queue depth. The process are waiting for the completion of the previously allocated requests, which should be reaped by the following polling, and thus causing a deadlock.

To avoid these subtle trouble described above, just disable iopoll for split bio.

---
 block/bio.c               |  2 ++
 block/blk-merge.c         | 12 ++++++++++++
 block/blk-mq.c            |  3 +++
 include/linux/blk_types.h |  1 +
 4 files changed, 18 insertions(+)
block/bio.c +2 -0
index fa01bef35bb1..7f7ddc22a30d 100644
@@ -684,6 +684,8 @@ void __bio_clone_fast(struct bio *bio, struct bio *bio_src)
 	bio_set_flag(bio, BIO_CLONED);
 	if (bio_flagged(bio_src, BIO_THROTTLED))
 		bio_set_flag(bio, BIO_THROTTLED);
+	if (bio_flagged(bio_src, BIO_SPLIT))
+		bio_set_flag(bio, BIO_SPLIT);
 	bio->bi_opf = bio_src->bi_opf;
 	bio->bi_ioprio = bio_src->bi_ioprio;
 	bio->bi_write_hint = bio_src->bi_write_hint;
block/blk-merge.c +6 -0
index bcf5e4580603..a2890cebf99f 100644
@@ -279,6 +279,18 @@ static struct bio *blk_bio_segment_split(struct request_queue *q,
 	return NULL;
 split:
 	*segs = nsegs;
+
+	/*
+	 * Bio splitting may cause subtle trouble such as hang when doing sync
+	 * iopoll in direct IO routine. Given performance gain of iopoll for
+	 * big IO can be trival, disable iopoll when split needed. We need
+	 * BIO_SPLIT to identify bios need this workaround. Since currently
Quoted 8 lines

block/blk-merge.c static struct bio *blk_bio_segment_split(struct request_queue *q, · line 292

Quoted 5 lines
block/blk-merge.c +5 -0
+	 * only normal IO under mq routine may suffer this issue, BIO_SPLIT is
+	 * only marked here.
+	 */
+	bio->bi_opf &= ~REQ_HIPRI;
+	bio_set_flag(bio, BIO_SPLIT);

You may need to put the above into one helper, and call the helper for other splitted cases(discard, write zero and write same) too.

It could be, though currently only normal IO could be marked with REQ_HIPRI.

Signature · 2 lines
Thanks,
Jeffle
#8 lore raw
Quoted 81 lines · JeffleXu
Quoted 70 lines · Ming Lei
Quoted 69 lines · Jeffle Xu

iopoll is initially for small size, latency sensitive IO. It doesn't work well for big IO, especially when it needs to be split to multiple bios. In this case, the returned cookie of __submit_bio_noacct_mq() is indeed the cookie of the last split bio. The completion of *this* last split bio done by iopoll doesn't mean the whole original bio has completed. Callers of iopoll still need to wait for completion of other split bios.

Besides bio splitting may cause more trouble for iopoll which isn't supposed to be used in case of big IO.

iopoll for split bio may cause potential race if CPU migration happens during bio submission. Since the returned cookie is that of the last split bio, polling on the corresponding hardware queue doesn't help complete other split bios, if these split bios are enqueued into different hardware queues. Since interrupts are disabled for polling queues, the completion of these other split bios depends on timeout mechanism, thus causing a potential hang.

iopoll for split bio may also cause hang for sync polling. Currently both the blkdev and iomap-based fs (ext4/xfs, etc) support sync polling in direct IO routine. These routines will submit bio without REQ_NOWAIT flag set, and then start sync polling in current process context. The process may hang in blk_mq_get_tag() if the submitted bio has to be split into multiple bios and can rapidly exhaust the queue depth. The process are waiting for the completion of the previously allocated requests, which should be reaped by the following polling, and thus causing a deadlock.

To avoid these subtle trouble described above, just disable iopoll for split bio.

---
 block/bio.c               |  2 ++
 block/blk-merge.c         | 12 ++++++++++++
 block/blk-mq.c            |  3 +++
 include/linux/blk_types.h |  1 +
 4 files changed, 18 insertions(+)
block/bio.c +2 -0
index fa01bef35bb1..7f7ddc22a30d 100644
@@ -684,6 +684,8 @@ void __bio_clone_fast(struct bio *bio, struct bio *bio_src)
 	bio_set_flag(bio, BIO_CLONED);
 	if (bio_flagged(bio_src, BIO_THROTTLED))
 		bio_set_flag(bio, BIO_THROTTLED);
+	if (bio_flagged(bio_src, BIO_SPLIT))
+		bio_set_flag(bio, BIO_SPLIT);
 	bio->bi_opf = bio_src->bi_opf;
 	bio->bi_ioprio = bio_src->bi_ioprio;
 	bio->bi_write_hint = bio_src->bi_write_hint;
block/blk-merge.c +6 -0
index bcf5e4580603..a2890cebf99f 100644
@@ -279,6 +279,18 @@ static struct bio *blk_bio_segment_split(struct request_queue *q,
 	return NULL;
 split:
 	*segs = nsegs;
+
+	/*
+	 * Bio splitting may cause subtle trouble such as hang when doing sync
+	 * iopoll in direct IO routine. Given performance gain of iopoll for
+	 * big IO can be trival, disable iopoll when split needed. We need
+	 * BIO_SPLIT to identify bios need this workaround. Since currently
Quoted 8 lines

block/blk-merge.c static struct bio *blk_bio_segment_split(struct request_queue *q, · line 292

Quoted 5 lines
block/blk-merge.c +5 -0
+	 * only normal IO under mq routine may suffer this issue, BIO_SPLIT is
+	 * only marked here.
+	 */
+	bio->bi_opf &= ~REQ_HIPRI;
+	bio_set_flag(bio, BIO_SPLIT);

You may need to put the above into one helper, and call the helper for other splitted cases(discard, write zero and write same) too.

Quoted 1 line

It could be, though currently only normal IO could be marked with REQ_HIPRI.

You are right, so far RWF_HIPRI is only applied on RW IO, then no need to do that for other non-RW split.

Thanks,
Ming