On 3/8/22 6:13 PM, Ming Lei wrote: > On Tue, Mar 08, 2022 at 06:02:50PM -0700, Jens Axboe wrote: >> On 3/7/22 11:53 AM, Mike Snitzer wrote: >>> From: Ming Lei <ming.lei@xxxxxxxxxx> >>> >>> Support bio(REQ_POLLED) polling in the following approach: >>> >>> 1) only support io polling on normal READ/WRITE, and other abnormal IOs >>> still fallback to IRQ mode, so the target io is exactly inside the dm >>> io. >>> >>> 2) hold one refcnt on io->io_count after submitting this dm bio with >>> REQ_POLLED >>> >>> 3) support dm native bio splitting, any dm io instance associated with >>> current bio will be added into one list which head is bio->bi_private >>> which will be recovered before ending this bio >>> >>> 4) implement .poll_bio() callback, call bio_poll() on the single target >>> bio inside the dm io which is retrieved via bio->bi_bio_drv_data; call >>> dm_io_dec_pending() after the target io is done in .poll_bio() >>> >>> 5) enable QUEUE_FLAG_POLL if all underlying queues enable QUEUE_FLAG_POLL, >>> which is based on Jeffle's previous patch. >> >> It's not the prettiest thing in the world with the overlay on bi_private, >> but at least it's nicely documented now. >> >> I would encourage you to actually test this on fast storage, should make >> a nice difference. I can run this on a gen2 optane, it's 10x the IOPS >> of what it was tested on and should help better highlight where it >> makes a difference. >> >> If either of you would like that, then send me a fool proof recipe for >> what should be setup so I have a poll capable dm device. > > Follows steps for setup dm stripe over two nvmes, then run io_uring on > the dm stripe dev. Thanks! Much easier when I don't have to figure it out... Setup: CPU: 12900K Drives: 2x P5800X gen2 optane (~5M IOPS each at 512b) Baseline kernel: sudo taskset -c 10 t/io_uring -d128 -b512 -s31 -c16 -p1 -F1 -B1 -n1 -R1 -X1 /dev/dm-0 Added file /dev/dm-0 (submitter 0) polled=1, fixedbufs=1/0, register_files=1, buffered=0, QD=128 Engine=io_uring, sq_ring=128, cq_ring=128 submitter=0, tid=1004 IOPS=2794K, BW=1364MiB/s, IOS/call=31/30, inflight=(124) IOPS=2793K, BW=1363MiB/s, IOS/call=31/31, inflight=(62) IOPS=2789K, BW=1362MiB/s, IOS/call=31/30, inflight=(124) IOPS=2779K, BW=1357MiB/s, IOS/call=31/31, inflight=(124) IOPS=2780K, BW=1357MiB/s, IOS/call=31/31, inflight=(62) IOPS=2779K, BW=1357MiB/s, IOS/call=31/31, inflight=(62) ^CExiting on signal Maximum IOPS=2794K generating about 500K ints/sec, and using 4k blocks: sudo taskset -c 10 t/io_uring -d128 -b4096 -s31 -c16 -p1 -F1 -B1 -n1 -R1 -X1 /dev/dm-0 Added file /dev/dm-0 (submitter 0) polled=1, fixedbufs=1/0, register_files=1, buffered=0, QD=128 Engine=io_uring, sq_ring=128, cq_ring=128 submitter=0, tid=967 IOPS=1683K, BW=6575MiB/s, IOS/call=24/24, inflight=(93) IOPS=1685K, BW=6584MiB/s, IOS/call=24/24, inflight=(124) IOPS=1686K, BW=6588MiB/s, IOS/call=24/24, inflight=(124) IOPS=1684K, BW=6581MiB/s, IOS/call=24/24, inflight=(93) IOPS=1686K, BW=6589MiB/s, IOS/call=24/24, inflight=(124) IOPS=1687K, BW=6593MiB/s, IOS/call=24/24, inflight=(128) IOPS=1687K, BW=6590MiB/s, IOS/call=24/24, inflight=(93) ^CExiting on signal Maximum IOPS=1687K which ends up being bw limited for me, because the devices aren't linked gen4. That's about 1.4M ints/sec. With the patched kernel, same test: sudo taskset -c 10 t/io_uring -d128 -b512 -s31 -c16 -p1 -F1 -B1 -n1 -R1 -X1 /dev/dm-0 Added file /dev/dm-0 (submitter 0) polled=1, fixedbufs=1/0, register_files=1, buffered=0, QD=128 Engine=io_uring, sq_ring=128, cq_ring=128 submitter=0, tid=989 IOPS=4151K, BW=2026MiB/s, IOS/call=16/15, inflight=(128) IOPS=4159K, BW=2031MiB/s, IOS/call=15/15, inflight=(128) IOPS=4193K, BW=2047MiB/s, IOS/call=15/15, inflight=(128) IOPS=4191K, BW=2046MiB/s, IOS/call=15/15, inflight=(128) IOPS=4202K, BW=2052MiB/s, IOS/call=15/15, inflight=(128) ^CExiting on signal Maximum IOPS=4202K with basically zero interrupts, and 4k: sudo taskset -c 10 t/io_uring -d128 -b4096 -s31 -c16 -p1 -F1 -B1 -n1 -R1 -X1 /dev/dm-0 Added file /dev/dm-0 (submitter 0) polled=1, fixedbufs=1/0, register_files=1, buffered=0, QD=128 Engine=io_uring, sq_ring=128, cq_ring=128 submitter=0, tid=1015 IOPS=1706K, BW=6666MiB/s, IOS/call=15/15, inflight=(128) IOPS=1704K, BW=6658MiB/s, IOS/call=15/15, inflight=(128) IOPS=1704K, BW=6658MiB/s, IOS/call=15/15, inflight=(128) IOPS=1704K, BW=6658MiB/s, IOS/call=15/15, inflight=(128) IOPS=1704K, BW=6658MiB/s, IOS/call=15/15, inflight=(128) ^CExiting on signal Maximum IOPS=1706K again with basically zero interrupts. That's about a 50% improvement for polled IO. This is using 2 gen2 optanes, which are good for ~5M IOPS each. Using two threads on a single core, baseline kernel: sudo taskset -c 10,11 t/io_uring -d128 -b512 -s31 -c16 -p1 -F1 -B1 -n2 -R1 -X1 /dev/dm-0 Added file /dev/dm-0 (submitter 0) Added file /dev/dm-0 (submitter 1) polled=1, fixedbufs=1/0, register_files=1, buffered=0, QD=128 Engine=io_uring, sq_ring=128, cq_ring=128 submitter=0, tid=1081 submitter=1, tid=1082 IOPS=3515K, BW=1716MiB/s, IOS/call=31/30, inflight=(124 62) IOPS=3515K, BW=1716MiB/s, IOS/call=31/31, inflight=(62 124) IOPS=3517K, BW=1717MiB/s, IOS/call=30/30, inflight=(113 124) IOPS=3517K, BW=1717MiB/s, IOS/call=31/31, inflight=(62 62) ^CExiting on signal Maximum IOPS=3517K and patched: udo taskset -c 10,11 t/io_uring -d128 -b512 -s31 -c16 -p1 -F1 -B1 -n2 -R1 -X1 /dev/dm-0 Added file /dev/dm-0 (submitter 0) Added file /dev/dm-0 (submitter 1) polled=1, fixedbufs=1/0, register_files=1, buffered=0, QD=128 Engine=io_uring, sq_ring=128, cq_ring=128 submitter=0, tid=949 submitter=1, tid=950 IOPS=4988K, BW=2435MiB/s, IOS/call=15/15, inflight=(128 128) IOPS=4985K, BW=2434MiB/s, IOS/call=15/15, inflight=(128 128) IOPS=4970K, BW=2426MiB/s, IOS/call=15/15, inflight=(128 128) IOPS=4985K, BW=2434MiB/s, IOS/call=15/15, inflight=(128 128) ^CExiting on signal Maximum IOPS=4988K which is about a 42% improvement in IOPS. -- Jens Axboe
