Re: [PATCH v2 0/2] virtio-blk: add iothread-vq-mapping parameter

[Stefan Hajnoczi <stefanha@redhat.com>]

[-- Attachment #1: Type: text/plain, Size: 2926 bytes --]

On Thu, Nov 02, 2023 at 03:10:52PM +0100, Kevin Wolf wrote:
> Am 18.09.2023 um 18:16 hat Stefan Hajnoczi geschrieben:
> > virtio-blk and virtio-scsi devices need a way to specify the mapping between
> > IOThreads and virtqueues. At the moment all virtqueues are assigned to a single
> > IOThread or the main loop. This single thread can be a CPU bottleneck, so it is
> > necessary to allow finer-grained assignment to spread the load. With this
> > series applied, "pidstat -t 1" shows that guests with -smp 2 or higher are able
> > to exploit multiple IOThreads.
> > 
> > This series introduces command-line syntax for the new iothread-vq-mapping
> > property is as follows:
> > 
> >   --device '{"driver":"virtio-blk-pci","iothread-vq-mapping":[{"iothread":"iothread0","vqs":[0,1,2]},...]},...'
> > 
> > IOThreads are specified by name and virtqueues are specified by 0-based
> > index.
> > 
> > It will be common to simply assign virtqueues round-robin across a set
> > of IOThreads. A convenient syntax that does not require specifying
> > individual virtqueue indices is available:
> > 
> >   --device '{"driver":"virtio-blk-pci","iothread-vq-mapping":[{"iothread":"iothread0"},{"iothread":"iothread1"},...]},...'
> > 
> > There is no way to reassign virtqueues at runtime and I expect that to be a
> > very rare requirement.
> > 
> > Note that JSON --device syntax is required for the iothread-vq-mapping
> > parameter because it's non-scalar.
> > 
> > Based-on: 20230912231037.826804-1-stefanha@redhat.com ("[PATCH v3 0/5] block-backend: process I/O in the current AioContext")
> 
> Does this strictly depend on patch 5/5 of that series, or would it just
> be a missed opportunity for optimisation by unnecessarily running some
> requests from a different thread?

"[PATCH v3 5/5] block-coroutine-wrapper: use
qemu_get_current_aio_context()" is necessary so that
virtio_blk_sect_range_ok -> blk_get_geometry -> blk_nb_sectors ->
bdrv_refresh_total_sectors -> bdrv_poll_co can be called without holding
the AioContext lock.

That case only happens when the BlockDriverState is a file-posix host
CD-ROM or a file-win32 host_device. Most users will never hit this
problem, but it would be unsafe to proceed merging code without this
patch.

> I suspect it does depend on the other virtio-blk series, though:
> 
> [PATCH 0/4] virtio-blk: prepare for the multi-queue block layer
> https://patchew.org/QEMU/20230914140101.1065008-1-stefanha@redhat.com/
> 
> Is this right?

Yes, it depends on "[PATCH 0/4] virtio-blk: prepare for the multi-queue
block layer" so that every AioContext is able to handle Linux AIO or
io_uring I/O and to stop using the AioContext lock in the virtio-blk I/O
code path.

Stefan

> 
> Given that soft freeze is early next week, maybe we should try to merge
> just the bare minimum of strictly necessary dependencies.
> 
> Kevin
> 

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 488 bytes --]