On Tue, Oct 25, 2022 at 8:02 PM Stefan Hajnoczi <stefanha@xxxxxxxxx> wrote:
>
> On Tue, 25 Oct 2022 at 04:17, Yongji Xie <xieyongji@xxxxxxxxxxxxx> wrote:
> >
> > On Fri, Oct 21, 2022 at 2:30 PM Jason Wang <jasowang@xxxxxxxxxx> wrote:
> > >
> > >
> > > 在 2022/10/21 13:33, Yongji Xie 写道:
> > > > On Tue, Oct 18, 2022 at 10:54 PM Stefan Hajnoczi <stefanha@xxxxxxxxx> wrote:
> > > >> On Tue, 18 Oct 2022 at 09:17, Yongji Xie <xieyongji@xxxxxxxxxxxxx> wrote:
> > > >>> On Tue, Oct 18, 2022 at 2:59 PM Ming Lei <tom.leiming@xxxxxxxxx> wrote:
> > > >>>> On Mon, Oct 17, 2022 at 07:11:59PM +0800, Yongji Xie wrote:
> > > >>>>> On Fri, Oct 14, 2022 at 8:57 PM Ming Lei <tom.leiming@xxxxxxxxx> 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 <stefanha@xxxxxxxxx> wrote:
> > > >>>>>>>> On Sat, 8 Oct 2022 at 04:43, Ziyang Zhang <ZiyangZhang@xxxxxxxxxxxxxxxxx> wrote:
> > > >>>>>>>>> On 2022/10/5 12:18, Ming Lei wrote:
> > > >>>>>>>>>> On Tue, Oct 04, 2022 at 09:53:32AM -0400, Stefan Hajnoczi wrote:
> > > >>>>>>>>>>> On Tue, 4 Oct 2022 at 05:44, Ming Lei <tom.leiming@xxxxxxxxx> wrote:
> > > >>>>>>>>>>>> On Mon, Oct 03, 2022 at 03:53:41PM -0400, Stefan Hajnoczi wrote:
> > > >>>>>>>>>>>>> On Fri, Sep 30, 2022 at 05:24:11PM +0800, Ming Lei wrote:
> > > >>>>>>>>>>>>>> ublk-qcow2 is available now.
> > > >>>>>>>>>>>>> Cool, thanks for sharing!
> > > >>>>>>>>>>>>>
> > > >>>>>>>>>>>>>> So far it provides basic read/write function, and compression and snapshot
> > > >>>>>>>>>>>>>> aren't supported yet. The target/backend implementation 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 helps libublksrv APIs/functions
> > > >>>>>>>>>>>>>> become mature/stable more quickly, since qcow2 is complicated and needs more
> > > >>>>>>>>>>>>>> requirement from libublksrv compared with other simple ones(loop, null)
> > > >>>>>>>>>>>>>>
> > > >>>>>>>>>>>>>> - there are several attempts of implementing qcow2 driver in kernel, such as
> > > >>>>>>>>>>>>>> ``qloop`` [2], ``dm-qcow2`` [3] and ``in kernel qcow2(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 one 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 using ublk-qcow2 block
> > > >>>>>>>>>>>>>> device as TEST_DEV, and kernel building workload is verified too. Also
> > > >>>>>>>>>>>>>> soft update approach is applied in meta flushing, and meta data
> > > >>>>>>>>>>>>>> integrity is guaranteed, 'make test T=qcow2/040' covers this kind of
> > > >>>>>>>>>>>>>> test, and only cluster leak is reported during this test.
> > > >>>>>>>>>>>>>>
> > > >>>>>>>>>>>>>> The performance data looks much better compared with 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=qcow2/002)
> > > >>>>>>>>>>>>> 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 as 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 all the following
> > > >>>>>>>>>>>> options:
> > > >>>>>>>>>>>>
> > > >>>>>>>>>>>> --- vDPA drivers
> > > >>>>>>>>>>>> <M> vDPA device simulator core
> > > >>>>>>>>>>>> <M> vDPA simulator for networking device
> > > >>>>>>>>>>>> <M> vDPA simulator for block device
> > > >>>>>>>>>>>> <M> VDUSE (vDPA Device in Userspace) support
> > > >>>>>>>>>>>> <M> Intel IFC VF vDPA driver
> > > >>>>>>>>>>>> <M> Virtio PCI bridge vDPA driver
> > > >>>>>>>>>>>> <M> 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 VDUSE.
> > > >>>>>>>>>>>
> > > >>>>>>>>>>> virtio_vdpa is available inside guests too. Please check that
> > > >>>>>>>>>>> VIRTIO_VDPA ("vDPA driver for virtio devices") is enabled in "Virtio
> > > >>>>>>>>>>> drivers" menu.
> > > >>>>>>>>>>>
> > > >>>>>>>>>>>>> # modprobe vduse
> > > >>>>>>>>>>>>> # qemu-storage-daemon \
> > > >>>>>>>>>>>>> --blockdev file,filename=test.qcow2,cache.direct=of|off,aio=native,node-name=file \
> > > >>>>>>>>>>>>> --blockdev qcow2,file=file,node-name=qcow2 \
> > > >>>>>>>>>>>>> --object iothread,id=iothread0 \
> > > >>>>>>>>>>>>> --export vduse-blk,id=vduse0,name=vduse0,num-queues=$(nproc),node-name=qcow2,writable=on,iothread=iothread0
> > > >>>>>>>>>>>>> # vdpa dev add name vduse0 mgmtdev vduse
> > > >>>>>>>>>>>>>
> > > >>>>>>>>>>>>> A virtio-blk device should appear and xfstests can be run on it
> > > >>>>>>>>>>>>> (typically /dev/vda unless you already have other virtio-blk devices).
> > > >>>>>>>>>>>>>
> > > >>>>>>>>>>>>> Afterwards you can destroy the device using:
> > > >>>>>>>>>>>>>
> > > >>>>>>>>>>>>> # vdpa dev del vduse0
> > > >>>>>>>>>>>>>
> > > >>>>>>>>>>>>>> - ublk-qcow2 (make test T=qcow2/022)
> > > >>>>>>>>>>>>> There are a lot of other factors not directly related to NBD vs ublk. In
> > > >>>>>>>>>>>>> order to get an apples-to-apples comparison with qemu-* 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 precisely.
> > > >>>>>>>>>>>> Maybe not true.
> > > >>>>>>>>>>>>
> > > >>>>>>>>>>>> ublk-qcow2 uses io_uring to handle all backend IO(include meta IO) completely,
> > > >>>>>>>>>>>> and so far single io_uring/pthread is for handling all 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=io_uring for setting up qemu-nbd, but not succeed.
> > > >>>>>>>>>>
> > > >>>>>>>>>>> know whether the benchmark demonstrates that ublk is faster than NBD,
> > > >>>>>>>>>>> that the ublk-qcow2 implementation is faster than qemu-nbd's qcow2,
> > > >>>>>>>>>>> whether there are miscellaneous implementation differences 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 socket, and io command
> > > >>>>>>>>>> is carried over io_uring pt commands, and should be fast than virio
> > > >>>>>>>>>> communication too.
> > > >>>>>>>>>>
> > > >>>>>>>>>> 2) io uring io handling is fast than libaio which is taken 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/common/qcow2_common
> > > >>>>>>>>>>
> > > >>>>>>>>>> 3) ublk uses one single io_uring to handle all io commands 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 root cause of the
> > > >>>>>>>>>>> performance difference. Let's learn why ublk-qcow2 performs well.
> > > >>>>>>>>>> Turns out the latest Fedora 37-beta doesn't support vdpa 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 setup 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 SSD.
> > > >>>>>>>>>>
> > > >>>>>>>>>> The 1st group tests all three block device which is backed 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), libaio, 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 queues: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 image
> > > >>>>>>>>>> 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), libaio, 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 queues: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 tested 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_QUEUE
> > > >>>>>>>>>
> > > >>>>>>>>> I setup VDUSE with:
> > > >>>>>>>>> qemu-storage-daemon \
> > > >>>>>>>>> --chardev socket,id=charmonitor,path=/tmp/qmp.sock,server=on,wait=off \
> > > >>>>>>>>> --monitor chardev=charmonitor \
> > > >>>>>>>>> --blockdev driver=host_device,cache.direct=on,filename=/dev/nullb0,node-name=disk0 \
> > > >>>>>>>>> --export vduse-blk,id=test,node-name=disk0,name=vduse_test,writable=on,num-queues=NR_QUEUE,queue-size=QUEUE_DEPTH
> > > >>>>>>>>>
> > > >>>>>>>>> Here QUEUE_DEPTH is 1, 32 or 128 and NR_QUEUE is 1 or 4.
> > > >>>>>>>>>
> > > >>>>>>>>> Note:
> > > >>>>>>>>> (1) VDUSE requires QUEUE_DEPTH >= 2. I cannot setup QUEUE_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 fair.
> > > >>>>>>>>> (4) I setup fio with direct=1, bs=4k.
> > > >>>>>>>>>
> > > >>>>>>>>> ------------------------------
> > > >>>>>>>>> 1 job 1 iodepth, lat（usec)
> > > >>>>>>>>> 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 vduse and ublk?
> > > >>>>>>>>
> > > >>>>>>> I think one reason for the latency gap of sync I/O is that vduse 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 will do
> > > >>>>>>> memcpy inside the critical section of virtqueue's spinlock in the
> > > >>>>>>> virtio-blk driver. That will hurt the performance heavily when
> > > >>>>>>> virtio_queue_rq() and virtblk_done() run concurrently. And it can be
> > > >>>>>>> mitigated by the advance DMA mapping feature [1] or irq binding
> > > >>>>>>> support [2].
> > > >>>>>> Hi Yongji,
> > > >>>>>>
> > > >>>>>> Yeah, that is the cost you paid for virtio. Wrt. userspace block 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 needed.
> > > >>>>>
> > > >>>>>> Even not sure if virtio
> > > >>>>>> ring is one good choice for such use case, given io_uring has been proved
> > > >>>>>> as very efficient(should be better than virtio ring, IMO).
> > > >>>>>>
> > > >>>>> Since vduse is aimed at creating a generic userspace device framework,
> > > >>>>> 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 much 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 in the
> > > >>>>> future.
> > > >>>> Per my understanding, at least there are two places where virtio ring 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 queue(CQ),
> > > >>>> so no contention exists between submission and completion; but virtio 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
> > > >> 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.
> > > >>
> > > >> 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.
> > > >>
> > > > Yes, we need to maintain a head pointer of the free descriptors in
> > > > both submission and completion path.
> > >
> > >
> > > Not necessarily after IN_ORDER?
> > >
> >
> > Sounds like a good idea.
>
> Submission and completion are still not 100% independent with IN_ORDER
> because descriptors are still in use until completion. It may not be
> necessary to keep a freelist, but you cannot actually use the
> descriptors for new submissions until existing requests complete. Is
> that correct?
Yes.
>
> Anyway, independent submission and completion rings aren't perfect
> either because independent submission introduces a new point of
> communication: the device must tell the driver when submitted
> descriptors have been processed. That means the driver must access a
> hardware register on the device or the device must DMA to RAM. So it
> involves extra bus traffic that is not necessary if descriptors are in
> use until completion. io_uring gets away with it because the
> io_uring_enter(2) syscall is synchronous and can therefore return the
> number of consumed sq elements for free.
Note that this is a syscall interface not a device/driver API, so
technically if it's useful, it could be added to virtio as well.
>
> There are ways to minimize that cost:
> 1. The driver only needs to fetch the device's sq index when it has
> run out of sq ring space.
> 2. The device can include sq index updates with completions. This is
> what NVMe does with the CQE SQ Head Pointer field, but the
> disadvantage is that the driver has no way of determining the sq index
> until a completion occurs.
Probably, but as replied in another thread, based on the numbers
measured from the networking test, I think the current virtio layout
should be sufficient for block I/O but might not fit for cases like
NFV.
Thanks
>
> Stefan
>
