From: Yu Kuai <yuku...@huawei.com> So that if caller didn't use plug, for example, __blkdev_direct_IO_simple() and __blkdev_direct_IO_async(), block layer can still benefit from caching nsec time in the plug.
Signed-off-by: Yu Kuai <yuku...@huawei.com> Link: https://lore.kernel.org/r/20240509123825.3225207-1-yuku...@huaweicloud.com Signed-off-by: Jens Axboe <ax...@kernel.dk> +++ block: fix lost bio for plug enabled bio based device With the following two conditions, bio will be lost: 1) blk plug is not enabled, for example, __blkdev_direct_IO_simple() and __blkdev_direct_IO_async(); 2) bio plug is enabled, for example write IO for raid1/raid10 while bitmap is enabled; Root cause is that blk_finish_plug() will add the bio to curent->bio_list, while such bio will not be handled: __submit_bio_noacct current->bio_list = bio_list_on_stack; blk_start_plug do { dm_submit_bio md_handle_request raid10_write_request -> generate new bio for underlying disks raid1_add_bio_to_plug -> bio is added to plug } while ((bio = bio_list_pop(&bio_list_on_stack[0]))) -> previous bio are all handled blk_finish_plug raid10_unplug raid1_submit_write submit_bio_noacct if (current->bio_list) bio_list_add(¤t->bio_list[0], bio) -> add new bio current->bio_list = NULL -> new bio is lost Fix the problem by moving the plug into the while loop, so that current->bio_list will still be handled after blk_finish_plug(). By the way, enable plug for raid1/raid10 in this case will also prevent delay IO handling into daemon thread, which should also improve IO performance. Fixes: 060406c61c7c ("block: add plug while submitting IO") Reported-by: Changhui Zhong <czh...@redhat.com> Closes: https://lore.kernel.org/all/cagvvp+xsmzy2g9yueatfmt6qv1m--ydocq0g7z7ovmctbbx...@mail.gmail.com/ Signed-off-by: Yu Kuai <yuku...@huawei.com> Tested-by: Changhui Zhong <czh...@redhat.com> Link: https://lore.kernel.org/r/20240521200308.983986-1-yuku...@huaweicloud.com Signed-off-by: Jens Axboe <ax...@kernel.dk> We see decrease of performance between vz9.40.9 and vz9.78.5 kernels, due to patch [1] being ported via RHEL. The patch [1] basically removes plug when submitting bio on this path: +-> blkdev_write_iter +-> blk_start_plug # removed in [1] +-> __generic_file_write_iter +-> generic_file_direct_write +-> blkdev_direct_IO +-> __blkdev_direct_IO_async +-> submit_bio +-> submit_bio_noacct +-> submit_bio_noacct_nocheck +-> __submit_bio +-> blk_start_plug # added back in [2] and [3] +-> blk_mq_submit_bio And there already are two mainstream patches which bring this plug back as block layer can benefit from it, so let's port them, after that performance degradation disappears. As fixup patch [3] basically reverts [2], let's merge them together to simplify further rebases of those patches. Fixes: 712c7364655f6 ("block: don't plug in blkdev_write_iter") [1] (cherry picked from commit 060406c61c7cb4bbd82a02d179decca9c9bb3443) [2] (cherry picked from commit 9a42891c35d50a8472b42c61256867b4dfcc1941) [3] https://virtuozzo.atlassian.net/browse/VSTOR-94335 Signed-off-by: Pavel Tikhomirov <ptikhomi...@virtuozzo.com> --- block/blk-core.c | 7 +++++++ 1 file changed, 7 insertions(+) diff --git a/block/blk-core.c b/block/blk-core.c index 6abad29dd501e..12a0e48d01125 100644 --- a/block/blk-core.c +++ b/block/blk-core.c @@ -592,9 +592,14 @@ static inline blk_status_t blk_check_zone_append(struct request_queue *q, static void __submit_bio(struct bio *bio) { + /* If plug is not used, add new plug here to cache nsecs time. */ + struct blk_plug plug; + if (unlikely(!blk_crypto_bio_prep(&bio))) return; + blk_start_plug(&plug); + if (!bio->bi_bdev->bd_has_submit_bio) { blk_mq_submit_bio(bio); } else if (likely(bio_queue_enter(bio) == 0)) { @@ -603,6 +608,8 @@ static void __submit_bio(struct bio *bio) disk->fops->submit_bio(bio); blk_queue_exit(disk->queue); } + + blk_finish_plug(&plug); } /* -- 2.47.1 _______________________________________________ Devel mailing list Devel@openvz.org https://lists.openvz.org/mailman/listinfo/devel
