Linux kernel mirror (for testing) git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
kernel os linux

block: enable per-cpu bio cache by default

Since after commit 12e4e8c7ab59 ("io_uring/rw: enable bio caches for
IRQ rw"), bio_put is safe for task and irq context, bio_alloc_bioset is
safe for task context and no one calls in irq context, so we can enable
per cpu bio cache by default.

Benchmarked with t/io_uring and ext4+nvme:
taskset -c 6 /root/fio/t/io_uring -p0 -d128 -b4096 -s1 -c1 -F1 -B1 -R1
-X1 -n1 -P1 /mnt/testfile
base IOPS is 562K, patch IOPS is 574K. The CPU usage of bio_alloc_bioset
decrease from 1.42% to 1.22%.

The worst case is allocate bio in CPU A but free in CPU B, still use
t/io_uring and ext4+nvme:
base IOPS is 648K, patch IOPS is 647K.

Also use fio test ext4/xfs with libaio/sync/io_uring on null_blk and
nvme, no obvious performance regression.

Signed-off-by: Fengnan Chang <changfengnan@bytedance.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

authored by

Fengnan Chang and committed by
Jens Axboe
48f22f80 05ce4c58

+12 -19
+12 -14
block/bio.c
··· 517 517 if (WARN_ON_ONCE(!mempool_initialized(&bs->bvec_pool) && nr_vecs > 0)) 518 518 return NULL; 519 519 520 - if (opf & REQ_ALLOC_CACHE) { 521 - if (bs->cache && nr_vecs <= BIO_INLINE_VECS) { 522 - bio = bio_alloc_percpu_cache(bdev, nr_vecs, opf, 523 - gfp_mask, bs); 524 - if (bio) 525 - return bio; 526 - /* 527 - * No cached bio available, bio returned below marked with 528 - * REQ_ALLOC_CACHE to particpate in per-cpu alloc cache. 529 - */ 530 - } else { 531 - opf &= ~REQ_ALLOC_CACHE; 532 - } 533 - } 520 + if (bs->cache && nr_vecs <= BIO_INLINE_VECS) { 521 + opf |= REQ_ALLOC_CACHE; 522 + bio = bio_alloc_percpu_cache(bdev, nr_vecs, opf, 523 + gfp_mask, bs); 524 + if (bio) 525 + return bio; 526 + /* 527 + * No cached bio available, bio returned below marked with 528 + * REQ_ALLOC_CACHE to participate in per-cpu alloc cache. 529 + */ 530 + } else 531 + opf &= ~REQ_ALLOC_CACHE; 534 532 535 533 /* 536 534 * submit_bio_noacct() converts recursion to iteration; this means if
-4
block/fops.c
··· 184 184 loff_t pos = iocb->ki_pos; 185 185 int ret = 0; 186 186 187 - if (iocb->ki_flags & IOCB_ALLOC_CACHE) 188 - opf |= REQ_ALLOC_CACHE; 189 187 bio = bio_alloc_bioset(bdev, nr_pages, opf, GFP_KERNEL, 190 188 &blkdev_dio_pool); 191 189 dio = container_of(bio, struct blkdev_dio, bio); ··· 331 333 loff_t pos = iocb->ki_pos; 332 334 int ret = 0; 333 335 334 - if (iocb->ki_flags & IOCB_ALLOC_CACHE) 335 - opf |= REQ_ALLOC_CACHE; 336 336 bio = bio_alloc_bioset(bdev, nr_pages, opf, GFP_KERNEL, 337 337 &blkdev_dio_pool); 338 338 dio = container_of(bio, struct blkdev_dio, bio);
-1
io_uring/rw.c
··· 855 855 ret = kiocb_set_rw_flags(kiocb, rw->flags, rw_type); 856 856 if (unlikely(ret)) 857 857 return ret; 858 - kiocb->ki_flags |= IOCB_ALLOC_CACHE; 859 858 860 859 /* 861 860 * If the file is marked O_NONBLOCK, still allow retry for it if it