On Thu, Feb 13, 2020 at 11:08:35AM +0300, Denis Plotnikov wrote:
> On 12.02.2020 18:43, Stefan Hajnoczi wrote:
> > On Tue, Feb 11, 2020 at 05:14:14PM +0300, Denis Plotnikov wrote:
> > > The goal is to reduce the amount of requests issued by a guest on
> > > 1M reads/writes. This rises the performance up to 4% on that kind of
> > > disk access pattern.
> > > 
> > > The maximum chunk size to be used for the guest disk accessing is
> > > limited with seg_max parameter, which represents the max amount of
> > > pices in the scatter-geather list in one guest disk request.
> > > 
> > > Since seg_max is virqueue_size dependent, increasing the virtqueue
> > > size increases seg_max, which, in turn, increases the maximum size
> > > of data to be read/write from a guest disk.
> > > 
> > > More details in the original problem statment:
> > > https://lists.gnu.org/archive/html/qemu-devel/2017-12/msg03721.html
> > > 
> > > Suggested-by: Denis V. Lunev <den@openvz.org>
> > > Signed-off-by: Denis Plotnikov <dplotnikov@virtuozzo.com>
> > > ---
> > >   hw/block/virtio-blk.c | 4 ++--
> > >   hw/core/machine.c     | 2 ++
> > >   hw/scsi/virtio-scsi.c | 4 ++--
> > >   3 files changed, 6 insertions(+), 4 deletions(-)
> > > 
> > > diff --git a/hw/block/virtio-blk.c b/hw/block/virtio-blk.c
> > > index 09f46ed85f..6df3a7a6df 100644
> > > --- a/hw/block/virtio-blk.c
> > > +++ b/hw/block/virtio-blk.c
> > > @@ -914,7 +914,7 @@ static void virtio_blk_update_config(VirtIODevice *vdev, uint8_t *config)
> > >       memset(&blkcfg, 0, sizeof(blkcfg));
> > >       virtio_stq_p(vdev, &blkcfg.capacity, capacity);
> > >       virtio_stl_p(vdev, &blkcfg.seg_max,
> > > -                 s->conf.seg_max_adjust ? s->conf.queue_size - 2 : 128 - 2);
> > > +                 s->conf.seg_max_adjust ? s->conf.queue_size - 2 : 256 - 2);
> > This value must not change on older machine types.
> Yes, that's true, but ..
> > So does this patch
> > need to turn seg-max-adjust *on* in hw_compat_4_2 so that old machine
> > types get 126 instead of 254?
> If we set seg-max-adjust "on" in older machine types, the setups using them
> and having queue_sizes set , for example, 1024 will also set seg_max to 1024
> - 2 which isn't the expected behavior: older mt didn't change seg_max in
> that case and stuck with 128 - 2.
> So, should we, instead, leave the default 128 - 2, for seg_max?

Argh!  Good point :-).

How about a seg_max_default property that is initialized to 254 for
modern machines and 126 to old machines?

Stefan