From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.4.0 (2014-02-07) on aws-us-west-2-korg-lkml-1.web.codeaurora.org Received: from vger.kernel.org (vger.kernel.org [23.128.96.18]) by smtp.lore.kernel.org (Postfix) with ESMTP id 29A86C38A2D for ; Mon, 24 Oct 2022 17:39:34 +0000 (UTC) Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S230471AbiJXRjc (ORCPT ); Mon, 24 Oct 2022 13:39:32 -0400 Received: from lindbergh.monkeyblade.net ([23.128.96.19]:58142 "EHLO lindbergh.monkeyblade.net" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S232904AbiJXRiz (ORCPT ); Mon, 24 Oct 2022 13:38:55 -0400 Received: from mail-yb1-xb31.google.com (mail-yb1-xb31.google.com [IPv6:2607:f8b0:4864:20::b31]) by lindbergh.monkeyblade.net (Postfix) with ESMTPS id 48947D40; Mon, 24 Oct 2022 09:14:29 -0700 (PDT) Received: by mail-yb1-xb31.google.com with SMTP id o70so11551513yba.7; Mon, 24 Oct 2022 09:14:28 -0700 (PDT) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=gmail.com; s=20210112; h=content-transfer-encoding:cc:to:subject:message-id:date:from :in-reply-to:references:mime-version:from:to:cc:subject:date :message-id:reply-to; bh=ATFwkykOyAaci5wDcDcUKYVL1U6HaRl0qiEFMft9KL8=; b=aWFz44F/QL8du5av45lr19sZAQqs7pr6MnQ5zcqA1mvNJ5T9cmr03R4rH1OK4rGXqv FNEfnwTrtP7DUXQyzNHrUyq0xOX3V9RlRtsk3OxDzUCwWmwYz+SqmUTAYUfpY/FZqCN2 HBMiXYgysl5JeL1/sc0ViSnp331OyZtzCRGuyAD1h0W3Mir2cIY10ZO3/URSz2536dkz js9ADn9ensvqbtXyby2HpvNsOg9r52Spc+URFueRCak+Z4ba2DaXxWbe+5hV0iEWzrY9 a8giUZEFb8R+m2TvFGVmPZYmYSQiUDPQmXC1r7KnbRaf6KHqxU1gd5NhMnm9ye8eeQGm EA6w== X-Google-DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=1e100.net; s=20210112; h=content-transfer-encoding:cc:to:subject:message-id:date:from :in-reply-to:references:mime-version:x-gm-message-state:from:to:cc :subject:date:message-id:reply-to; bh=ATFwkykOyAaci5wDcDcUKYVL1U6HaRl0qiEFMft9KL8=; b=tc8J35dY784XKGLqvQk+4jgcI74cS/NCQ1UUS7ag1Ab0jOYe8QxOTweMm9NFsQ2Hfh qC2H2dd4aojzcP7dbGQt+/bRDrgA+RfW85M4ZdiiPswCKTyPTfmEVMCTH8JQG0qyNrZ7 4KqkMHTyYPKiRjoiCcx0OfsC1FUuosTFknlyVm7w/4KvesUeB/cxrQ+8zjDQJzlEFQ8a qGRQBxw1J/dwGauIUyb5qv3XFCoQmYimVH9ylbzxClNtB1dvujf1ly1ku2wQfI6btGA5 W6+Wb6aXJGs43HS2xdPKE1SGD3nRRZDKIx18wwFzH1JUQAvl5ygSC6W9fzKkaNnnx0QJ 9jPQ== X-Gm-Message-State: ACrzQf2T1UksWgZCzh090uqFe10XRDqlUVJujjmvhdJMzZ2DDTePZQgu AEwj/tJ0c3co0qVq9i0LX5aNkypeucDFlP8j1FU= X-Google-Smtp-Source: AMsMyM50bchyetcRXhNbsQcJcX0g/9HlmAxvXz+5fzmEhJaaVce0rvzBnz0E3qMTMl/j7jcBecEjSTDdXmow7LMNHpM= X-Received: by 2002:a25:b098:0:b0:6ca:4484:486c with SMTP id f24-20020a25b098000000b006ca4484486cmr19631908ybj.209.1666627918532; Mon, 24 Oct 2022 09:11:58 -0700 (PDT) MIME-Version: 1.0 References: <50827796-af93-4af5-4121-dc13c31a67fc@linux.alibaba.com> In-Reply-To: From: Stefan Hajnoczi Date: Mon, 24 Oct 2022 12:11:46 -0400 Message-ID: Subject: Re: ublk-qcow2: ublk-qcow2 is available To: Ming Lei Cc: Yongji Xie , "Michael S. Tsirkin" , Ziyang Zhang , Stefan Hajnoczi , io-uring@vger.kernel.org, linux-block@vger.kernel.org, linux-kernel , "Denis V. Lunev" , Xiaoguang Wang Content-Type: text/plain; charset="UTF-8" Content-Transfer-Encoding: quoted-printable Precedence: bulk List-ID: X-Mailing-List: io-uring@vger.kernel.org On Wed, 19 Oct 2022 at 05:09, Ming Lei wrote: > > On Tue, Oct 18, 2022 at 10:54:45AM -0400, Stefan Hajnoczi wrote: > > On Tue, 18 Oct 2022 at 09:17, Yongji Xie wrot= e: > > > > > > On Tue, Oct 18, 2022 at 2:59 PM Ming Lei wrot= e: > > > > > > > > On Mon, Oct 17, 2022 at 07:11:59PM +0800, Yongji Xie wrote: > > > > > On Fri, Oct 14, 2022 at 8:57 PM Ming Lei = wrote: > > > > > > > > > > > > On Thu, Oct 13, 2022 at 02:48:04PM +0800, Yongji Xie wrote: > > > > > > > On Wed, Oct 12, 2022 at 10:22 PM Stefan Hajnoczi wrote: > > > > > > > > > > > > > > > > On Sat, 8 Oct 2022 at 04:43, Ziyang Zhang wrote: > > > > > > > > > > > > > > > > > > On 2022/10/5 12:18, Ming Lei wrote: > > > > > > > > > > On Tue, Oct 04, 2022 at 09:53:32AM -0400, Stefan Hajnoc= zi wrote: > > > > > > > > > >> On Tue, 4 Oct 2022 at 05:44, Ming Lei wrote: > > > > > > > > > >>> > > > > > > > > > >>> On Mon, Oct 03, 2022 at 03:53:41PM -0400, Stefan Hajn= oczi wrote: > > > > > > > > > >>>> On Fri, Sep 30, 2022 at 05:24:11PM +0800, Ming Lei w= rote: > > > > > > > > > >>>>> ublk-qcow2 is available now. > > > > > > > > > >>>> > > > > > > > > > >>>> Cool, thanks for sharing! > > > > > > > > > >>>> > > > > > > > > > >>>>> > > > > > > > > > >>>>> So far it provides basic read/write function, and c= ompression and snapshot > > > > > > > > > >>>>> aren't supported yet. The target/backend implementa= tion is completely > > > > > > > > > >>>>> based on io_uring, and share the same io_uring with= ublk IO command > > > > > > > > > >>>>> handler, just like what ublk-loop does. > > > > > > > > > >>>>> > > > > > > > > > >>>>> Follows the main motivations of ublk-qcow2: > > > > > > > > > >>>>> > > > > > > > > > >>>>> - building one complicated target from scratch help= s libublksrv APIs/functions > > > > > > > > > >>>>> become mature/stable more quickly, since qcow2 is= complicated and needs more > > > > > > > > > >>>>> requirement from libublksrv compared with other s= imple ones(loop, null) > > > > > > > > > >>>>> > > > > > > > > > >>>>> - there are several attempts of implementing qcow2 = driver in kernel, such as > > > > > > > > > >>>>> ``qloop`` [2], ``dm-qcow2`` [3] and ``in kernel q= cow2(ro)`` [4], so ublk-qcow2 > > > > > > > > > >>>>> might useful be for covering requirement in this = field > > > > > > > > > >>>>> > > > > > > > > > >>>>> - performance comparison with qemu-nbd, and it was = my 1st thought to evaluate > > > > > > > > > >>>>> performance of ublk/io_uring backend by writing o= ne ublk-qcow2 since ublksrv > > > > > > > > > >>>>> is started > > > > > > > > > >>>>> > > > > > > > > > >>>>> - help to abstract common building block or design = pattern for writing new ublk > > > > > > > > > >>>>> target/backend > > > > > > > > > >>>>> > > > > > > > > > >>>>> So far it basically passes xfstest(XFS) test by usi= ng ublk-qcow2 block > > > > > > > > > >>>>> device as TEST_DEV, and kernel building workload is= verified too. Also > > > > > > > > > >>>>> soft update approach is applied in meta flushing, a= nd meta data > > > > > > > > > >>>>> integrity is guaranteed, 'make test T=3Dqcow2/040' = covers this kind of > > > > > > > > > >>>>> test, and only cluster leak is reported during this= test. > > > > > > > > > >>>>> > > > > > > > > > >>>>> The performance data looks much better compared wit= h qemu-nbd, see > > > > > > > > > >>>>> details in commit log[1], README[5] and STATUS[6]. = And the test covers both > > > > > > > > > >>>>> empty image and pre-allocated image, for example of= pre-allocated qcow2 > > > > > > > > > >>>>> image(8GB): > > > > > > > > > >>>>> > > > > > > > > > >>>>> - qemu-nbd (make test T=3Dqcow2/002) > > > > > > > > > >>>> > > > > > > > > > >>>> Single queue? > > > > > > > > > >>> > > > > > > > > > >>> Yeah. > > > > > > > > > >>> > > > > > > > > > >>>> > > > > > > > > > >>>>> randwrite(4k): jobs 1, iops 24605 > > > > > > > > > >>>>> randread(4k): jobs 1, iops 30938 > > > > > > > > > >>>>> randrw(4k): jobs 1, iops read 13981 write 14001 > > > > > > > > > >>>>> rw(512k): jobs 1, iops read 724 write 728 > > > > > > > > > >>>> > > > > > > > > > >>>> Please try qemu-storage-daemon's VDUSE export type a= s well. The > > > > > > > > > >>>> command-line should be similar to this: > > > > > > > > > >>>> > > > > > > > > > >>>> # modprobe virtio_vdpa # attaches vDPA devices to = host kernel > > > > > > > > > >>> > > > > > > > > > >>> Not found virtio_vdpa module even though I enabled al= l the following > > > > > > > > > >>> options: > > > > > > > > > >>> > > > > > > > > > >>> --- vDPA drivers > > > > > > > > > >>> vDPA device simulator core > > > > > > > > > >>> vDPA simulator for networking devic= e > > > > > > > > > >>> vDPA simulator for block device > > > > > > > > > >>> VDUSE (vDPA Device in Userspace) supp= ort > > > > > > > > > >>> Intel IFC VF vDPA driver > > > > > > > > > >>> Virtio PCI bridge vDPA driver > > > > > > > > > >>> vDPA driver for Alibaba ENI > > > > > > > > > >>> > > > > > > > > > >>> BTW, my test environment is VM and the shared data is= done in VM too, and > > > > > > > > > >>> can virtio_vdpa be used inside VM? > > > > > > > > > >> > > > > > > > > > >> I hope Xie Yongji can help explain how to benchmark VD= USE. > > > > > > > > > >> > > > > > > > > > >> virtio_vdpa is available inside guests too. Please che= ck that > > > > > > > > > >> VIRTIO_VDPA ("vDPA driver for virtio devices") is enab= led in "Virtio > > > > > > > > > >> drivers" menu. > > > > > > > > > >> > > > > > > > > > >>> > > > > > > > > > >>>> # modprobe vduse > > > > > > > > > >>>> # qemu-storage-daemon \ > > > > > > > > > >>>> --blockdev file,filename=3Dtest.qcow2,cache.di= rect=3Dof|off,aio=3Dnative,node-name=3Dfile \ > > > > > > > > > >>>> --blockdev qcow2,file=3Dfile,node-name=3Dqcow2= \ > > > > > > > > > >>>> --object iothread,id=3Diothread0 \ > > > > > > > > > >>>> --export vduse-blk,id=3Dvduse0,name=3Dvduse0,n= um-queues=3D$(nproc),node-name=3Dqcow2,writable=3Don,iothread=3Diothread0 > > > > > > > > > >>>> # vdpa dev add name vduse0 mgmtdev vduse > > > > > > > > > >>>> > > > > > > > > > >>>> A virtio-blk device should appear and xfstests can b= e run on it > > > > > > > > > >>>> (typically /dev/vda unless you already have other vi= rtio-blk devices). > > > > > > > > > >>>> > > > > > > > > > >>>> Afterwards you can destroy the device using: > > > > > > > > > >>>> > > > > > > > > > >>>> # vdpa dev del vduse0 > > > > > > > > > >>>> > > > > > > > > > >>>>> > > > > > > > > > >>>>> - ublk-qcow2 (make test T=3Dqcow2/022) > > > > > > > > > >>>> > > > > > > > > > >>>> There are a lot of other factors not directly relate= d to NBD vs ublk. In > > > > > > > > > >>>> order to get an apples-to-apples comparison with qem= u-* a ublk export > > > > > > > > > >>>> type is needed in qemu-storage-daemon. That way only= the difference is > > > > > > > > > >>>> the ublk interface and the rest of the code path is = identical, making it > > > > > > > > > >>>> possible to compare NBD, VDUSE, ublk, etc more preci= sely. > > > > > > > > > >>> > > > > > > > > > >>> Maybe not true. > > > > > > > > > >>> > > > > > > > > > >>> ublk-qcow2 uses io_uring to handle all backend IO(inc= lude meta IO) completely, > > > > > > > > > >>> and so far single io_uring/pthread is for handling al= l qcow2 IOs and IO > > > > > > > > > >>> command. > > > > > > > > > >> > > > > > > > > > >> qemu-nbd doesn't use io_uring to handle the backend IO= , so we don't > > > > > > > > > > > > > > > > > > > > I tried to use it via --aio=3Dio_uring for setting up q= emu-nbd, but not succeed. > > > > > > > > > > > > > > > > > > > >> know whether the benchmark demonstrates that ublk is f= aster than NBD, > > > > > > > > > >> that the ublk-qcow2 implementation is faster than qemu= -nbd's qcow2, > > > > > > > > > >> whether there are miscellaneous implementation differe= nces between > > > > > > > > > >> ublk-qcow2 and qemu-nbd (like using the same io_uring = context for both > > > > > > > > > >> ublk and backend IO), or something else. > > > > > > > > > > > > > > > > > > > > The theory shouldn't be too complicated: > > > > > > > > > > > > > > > > > > > > 1) io uring passthough(pt) communication is fast than s= ocket, and io command > > > > > > > > > > is carried over io_uring pt commands, and should be fas= t than virio > > > > > > > > > > communication too. > > > > > > > > > > > > > > > > > > > > 2) io uring io handling is fast than libaio which is ta= ken in the > > > > > > > > > > test on qemu-nbd, and all qcow2 backend io(include meta= io) is handled > > > > > > > > > > by io_uring. > > > > > > > > > > > > > > > > > > > > https://github.com/ming1/ubdsrv/blob/master/tests/commo= n/qcow2_common > > > > > > > > > > > > > > > > > > > > 3) ublk uses one single io_uring to handle all io comma= nds and qcow2 > > > > > > > > > > backend IOs, so batching handling is common, and it is = easy to see > > > > > > > > > > dozens of IOs/io commands handled in single syscall, or= even more. > > > > > > > > > > > > > > > > > > > >> > > > > > > > > > >> I'm suggesting measuring changes to just 1 variable at= a time. > > > > > > > > > >> Otherwise it's hard to reach a conclusion about the ro= ot cause of the > > > > > > > > > >> performance difference. Let's learn why ublk-qcow2 per= forms well. > > > > > > > > > > > > > > > > > > > > Turns out the latest Fedora 37-beta doesn't support vdp= a yet, so I built > > > > > > > > > > qemu from the latest github tree, and finally it starts= to work. And test kernel > > > > > > > > > > is v6.0 release. > > > > > > > > > > > > > > > > > > > > Follows the test result, and all three devices are setu= p as single > > > > > > > > > > queue, and all tests are run in single job, still done = in one VM, and > > > > > > > > > > the test images are stored on XFS/virito-scsi backed SS= D. > > > > > > > > > > > > > > > > > > > > The 1st group tests all three block device which is bac= ked by empty > > > > > > > > > > qcow2 image. > > > > > > > > > > > > > > > > > > > > The 2nd group tests all the three block devices backed = by pre-allocated > > > > > > > > > > qcow2 image. > > > > > > > > > > > > > > > > > > > > Except for big sequential IO(512K), there is still not = small gap between > > > > > > > > > > vdpa-virtio-blk and ublk. > > > > > > > > > > > > > > > > > > > > 1. run fio on block device over empty qcow2 image > > > > > > > > > > 1) qemu-nbd > > > > > > > > > > running qcow2/001 > > > > > > > > > > run perf test on empty qcow2 image via nbd > > > > > > > > > > fio (nbd(/mnt/data/ublk_null_8G_nYbgF.qcow2), lib= aio, bs 4k, dio, hw queues:1)... > > > > > > > > > > randwrite: jobs 1, iops 8549 > > > > > > > > > > randread: jobs 1, iops 34829 > > > > > > > > > > randrw: jobs 1, iops read 11363 write 11333 > > > > > > > > > > rw(512k): jobs 1, iops read 590 write 597 > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > 2) ublk-qcow2 > > > > > > > > > > running qcow2/021 > > > > > > > > > > run perf test on empty qcow2 image via ublk > > > > > > > > > > fio (ublk/qcow2( -f /mnt/data/ublk_null_8G_s761j.= qcow2), libaio, bs 4k, dio, hw queues:1, uring_comp: 0, get_data: 0). > > > > > > > > > > randwrite: jobs 1, iops 16086 > > > > > > > > > > randread: jobs 1, iops 172720 > > > > > > > > > > randrw: jobs 1, iops read 35760 write 35702 > > > > > > > > > > rw(512k): jobs 1, iops read 1140 write 1149 > > > > > > > > > > > > > > > > > > > > 3) vdpa-virtio-blk > > > > > > > > > > running debug/test_dev > > > > > > > > > > run io test on specified device > > > > > > > > > > fio (vdpa(/dev/vdc), libaio, bs 4k, dio, hw queue= s:1)... > > > > > > > > > > randwrite: jobs 1, iops 8626 > > > > > > > > > > randread: jobs 1, iops 126118 > > > > > > > > > > randrw: jobs 1, iops read 17698 write 17665 > > > > > > > > > > rw(512k): jobs 1, iops read 1023 write 1031 > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > 2. run fio on block device over pre-allocated qcow2 ima= ge > > > > > > > > > > 1) qemu-nbd > > > > > > > > > > running qcow2/002 > > > > > > > > > > run perf test on pre-allocated qcow2 image via nbd > > > > > > > > > > fio (nbd(/mnt/data/ublk_data_8G_sc0SB.qcow2), lib= aio, bs 4k, dio, hw queues:1)... > > > > > > > > > > randwrite: jobs 1, iops 21439 > > > > > > > > > > randread: jobs 1, iops 30336 > > > > > > > > > > randrw: jobs 1, iops read 11476 write 11449 > > > > > > > > > > rw(512k): jobs 1, iops read 718 write 722 > > > > > > > > > > > > > > > > > > > > 2) ublk-qcow2 > > > > > > > > > > running qcow2/022 > > > > > > > > > > run perf test on pre-allocated qcow2 image via ublk > > > > > > > > > > fio (ublk/qcow2( -f /mnt/data/ublk_data_8G_yZiaJ.= qcow2), libaio, bs 4k, dio, hw queues:1, uring_comp: 0, get_data: 0). > > > > > > > > > > randwrite: jobs 1, iops 98757 > > > > > > > > > > randread: jobs 1, iops 110246 > > > > > > > > > > randrw: jobs 1, iops read 47229 write 47161 > > > > > > > > > > rw(512k): jobs 1, iops read 1416 write 1427 > > > > > > > > > > > > > > > > > > > > 3) vdpa-virtio-blk > > > > > > > > > > running debug/test_dev > > > > > > > > > > run io test on specified device > > > > > > > > > > fio (vdpa(/dev/vdc), libaio, bs 4k, dio, hw queue= s:1)... > > > > > > > > > > randwrite: jobs 1, iops 47317 > > > > > > > > > > randread: jobs 1, iops 74092 > > > > > > > > > > randrw: jobs 1, iops read 27196 write 27234 > > > > > > > > > > rw(512k): jobs 1, iops read 1447 write 1458 > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > Hi All, > > > > > > > > > > > > > > > > > > We are interested in VDUSE vs UBLK, too. And I have teste= d them with nullblk backend. > > > > > > > > > Let me share some results here. > > > > > > > > > > > > > > > > > > I setup UBLK with: > > > > > > > > > ublk add -t loop -f /dev/nullb0 -d QUEUE_DEPTH -q NR_QU= EUE > > > > > > > > > > > > > > > > > > I setup VDUSE with: > > > > > > > > > qemu-storage-daemon \ > > > > > > > > > --chardev socket,id=3Dcharmonitor,path=3D/tmp/qmp.= sock,server=3Don,wait=3Doff \ > > > > > > > > > --monitor chardev=3Dcharmonitor \ > > > > > > > > > --blockdev driver=3Dhost_device,cache.direct=3Don,= filename=3D/dev/nullb0,node-name=3Ddisk0 \ > > > > > > > > > --export vduse-blk,id=3Dtest,node-name=3Ddisk0,nam= e=3Dvduse_test,writable=3Don,num-queues=3DNR_QUEUE,queue-size=3DQUEUE_DEPTH > > > > > > > > > > > > > > > > > > Here QUEUE_DEPTH is 1, 32 or 128 and NR_QUEUE is 1 or 4. > > > > > > > > > > > > > > > > > > Note: > > > > > > > > > (1) VDUSE requires QUEUE_DEPTH >=3D 2. I cannot setup QUE= UE_DEPTH to 1. > > > > > > > > > (2) I use qemu 7.1.0-rc3. It supports vduse-blk. > > > > > > > > > (3) I do not use ublk null target so that the test is fai= r. > > > > > > > > > (4) I setup fio with direct=3D1, bs=3D4k. > > > > > > > > > > > > > > > > > > ------------------------------ > > > > > > > > > 1 job 1 iodepth, lat=EF=BC=88usec) > > > > > > > > > vduse ublk > > > > > > > > > seq-read 22.55 11.15 > > > > > > > > > rand-read 22.49 11.17 > > > > > > > > > seq-write 25.67 10.25 > > > > > > > > > rand-write 24.13 10.16 > > > > > > > > > > > > > > > > Thanks for sharing. Any idea what the bottlenecks are for v= duse and ublk? > > > > > > > > > > > > > > > > > > > > > > I think one reason for the latency gap of sync I/O is that vd= use uses > > > > > > > workqueue in the I/O completion path but ublk doesn't. > > > > > > > > > > > > > > And one bottleneck for the async I/O in vduse is that vduse w= ill do > > > > > > > memcpy inside the critical section of virtqueue's spinlock in= the > > > > > > > virtio-blk driver. That will hurt the performance heavily whe= n > > > > > > > virtio_queue_rq() and virtblk_done() run concurrently. And it= can be > > > > > > > mitigated by the advance DMA mapping feature [1] or irq bindi= ng > > > > > > > support [2]. > > > > > > > > > > > > Hi Yongji, > > > > > > > > > > > > Yeah, that is the cost you paid for virtio. Wrt. userspace bloc= k device > > > > > > or other sort of userspace devices, cmd completion is driven by > > > > > > userspace, not sure if one such 'irq' is needed. > > > > > > > > > > I'm not sure, it can be an optional feature in the future if need= ed. > > > > > > > > > > > Even not sure if virtio > > > > > > ring is one good choice for such use case, given io_uring has b= een proved > > > > > > as very efficient(should be better than virtio ring, IMO). > > > > > > > > > > > > > > > > Since vduse is aimed at creating a generic userspace device frame= work, > > > > > virtio should be the right way IMO. > > > > > > > > OK, it is the right way, but may not be the effective one. > > > > > > > > > > Maybe, but I think we can try to optimize it. > > > > > > > > And with the vdpa framework, the > > > > > userspace device can serve both virtual machines and containers. > > > > > > > > virtio is good for VM, but not sure it is good enough for other > > > > cases. > > > > > > > > > > > > > > Regarding the performance issue, actually I can't measure how muc= h of > > > > > the performance loss is due to the difference between virtio ring= and > > > > > iouring. But I think it should be very small. The main costs come= from > > > > > the two bottlenecks I mentioned before which could be mitigated i= n the > > > > > future. > > > > > > > > Per my understanding, at least there are two places where virtio ri= ng is > > > > less efficient than io_uring: > > > > > > > > > > I might have misunderstood what you mean by virtio ring before. My > > > previous understanding of the virtio ring does not include the > > > virtio-blk driver. > > > > > > > 1) io_uring uses standalone submission queue(SQ) and completion que= ue(CQ), > > > > so no contention exists between submission and completion; but virt= io queue > > > > requires per-vq lock in both submission and completion. > > > > > > > > > > Yes, this is the bottleneck of the virtio-blk driver, even in the VM > > > case. We are also trying to optimize this lock. > > > > > > One way to mitigate it is making submission and completion happen in > > > the same core. > > > > QEMU sizes virtio-blk device num-queues to match the vCPU count. The > > num-queues is configurable via qemu-storage-daemon command line, and > single queue is usually common case, more queues often means more > resources. Sorry, I didn't make a distinction between the running VM case the and qemu-storage-daemon case where there is no VM. I described the VM case here because "submission and completion happen in the same core" as suggested there. For qemu-storage-daemon there is no automatic sizing of num-queues. It's up to the user to decide that manually. qemu-storage-daemon currently does not fully take advantage of SMP. All queues are processed by a single thread in qemu-storage-daemon. In the future it should be possible to assign queues to specific threads (there is ongoing "multi-queue QEMU block layer" work to do this). The resource requirements of virtqueues aren't very large: 1. Minimum 4 KB of memory for a packed vring. 2. 1 eventfd for submission notification. 3. 1 eventfd for completion notification. Having ~64 queues is not a big resource commitment. > > > virtio-blk driver is a blk-mq driver, so submissions and completions > > for a given virtqueue should already be processed by the same vCPU. > > > > Unless the device is misconfigured or the guest software chooses a > > custom vq:vCPU mapping, there should be no vq lock contention between > > vCPUs. > > Single queue or nr_queue less than nr_cpus can't be thought as mis-config= ured, > so every vCPU can submit request, but only one or a few vCPUs complete al= l. Yes. > > > > > I can think of a reason why submission and completion require > > coordination: descriptors are occupied until completion. The > > submission logic chooses free descriptors from the table. The > > completion logic returns free descriptors so they can be used in > > future submissions. > > Shared descriptors is one fundamental design of virtio ring, and > looks the reason why vq spin lock is needed in both sides. > > > > > Other ring designs expose the submission ring head AND tail index so > > that it's clear which submissions have been processed by the other > > side. Once processed, the descriptors are no longer occupied and can > > be reused for future submissions immediately. This means that > > submission and completion do not share state. > > > > This is for the split virtqueue layout. For the packed layout I think > > there is a similar dependency because descriptors are used for both > > submission and completion. > > > > I have CCed Michael Tsirkin in case he has any thoughts on the > > independence of submission and completion in the vring design. > > > > BTW I have written about difference in the VIRTIO, NVMe, and io_uring > > descriptor ring designs here: > > https://blog.vmsplice.net/2022/06/comparing-virtio-nvme-and-iouring-que= ue.html > > Except for ring, notification could be another difference. Yes, the io_uring_enter(2) syscall takes over the control flow of the current thread and can perform both submission and completion work. VIRTIO/vhost-user has separate submission and completion notifications, although they are typically implemented as eventfds that can be processed with io_uring too. Stefan