On Wed, Oct 23, 2019 at 05:28:17PM -0400, Michael S. Tsirkin wrote:
> On Fri, Oct 18, 2019 at 02:55:47PM +0300, Denis Plotnikov wrote:
> > From: "Denis V. Lunev" <den@openvz.org>
> > 
> > Linux guests submit IO requests no longer than PAGE_SIZE * max_seg
> > field reported by SCSI controler. Thus typical sequential read with
> > 1 MB size results in the following pattern of the IO from the guest:
> >   8,16   1    15754     2.766095122  2071  D   R 2095104 + 1008 [dd]
> >   8,16   1    15755     2.766108785  2071  D   R 2096112 + 1008 [dd]
> >   8,16   1    15756     2.766113486  2071  D   R 2097120 + 32 [dd]
> >   8,16   1    15757     2.767668961     0  C   R 2095104 + 1008 [0]
> >   8,16   1    15758     2.768534315     0  C   R 2096112 + 1008 [0]
> >   8,16   1    15759     2.768539782     0  C   R 2097120 + 32 [0]
> > The IO was generated by
> >   dd if=/dev/sda of=/dev/null bs=1024 iflag=direct
> > 
> > This effectively means that on rotational disks we will observe 3 IOPS
> > for each 2 MBs processed. This definitely negatively affects both
> > guest and host IO performance.
> > 
> > The cure is relatively simple - we should report lengthy scatter-gather
> > ability of the SCSI controller. Fortunately the situation here is very
> > good. VirtIO transport layer can accomodate 1024 items in one request
> > while we are using only 128. This situation is present since almost
> > very beginning. 2 items are dedicated for request metadata thus we
> > should publish VIRTQUEUE_MAX_SIZE - 2 as max_seg.
> > 
> > The following pattern is observed after the patch:
> >   8,16   1     9921     2.662721340  2063  D   R 2095104 + 1024 [dd]
> >   8,16   1     9922     2.662737585  2063  D   R 2096128 + 1024 [dd]
> >   8,16   1     9923     2.665188167     0  C   R 2095104 + 1024 [0]
> >   8,16   1     9924     2.665198777     0  C   R 2096128 + 1024 [0]
> > which is much better.
> > 
> > The dark side of this patch is that we are tweaking guest visible
> > parameter, though this should be relatively safe as above transport
> > layer support is present in QEMU/host Linux for a very long time.
> > The patch adds configurable property for VirtIO SCSI with a new default
> > and hardcode option for VirtBlock which does not provide good
> > configurable framework.
> > 
> > Unfortunately the commit can not be applied as is. For the real cure we
> > need guest to be fixed to accomodate that queue length, which is done
> > only in the latest 4.14 kernel. Thus we are going to expose the property
> > and tweak it on machine type level.
> > 
> > The problem with the old kernels is that they have
> > max_segments <= virtqueue_size restriction which cause the guest
> > crashing in the case of violation.
> 
> This isn't just in the guests: virtio spec also seems to imply this,
> or at least be vague on this point.
> 
> So I think it'll need a feature bit.
> Doing that in a safe way will also allow being compatible with old guests.

The spec is quite explicit about this:

  2.6.5 The Virtqueue Descriptor Table

  The number of descriptors in the table is defined by the queue size for this virtqueue: this is the maximum possible descriptor chain length.

and:

  2.6.5.3.1 Driver Requirements: Indirect Descriptors

  A driver MUST NOT create a descriptor chain longer than the Queue Size of the device.

If some drivers or devices allow longer descriptor chains today that's
an implementation quirk but a new feature bit is definitely required to
officially allow longer descriptor chains.

Stefan