Re: [PATCH virtio 0/8] virtio_pci_modern: allow parallel admin queue commands execution

[Jiri Pirko <jiri@resnulli.us>]

Mon, Jun 24, 2024 at 01:23:01PM CEST, mst@redhat.com wrote:
>On Mon, Jun 24, 2024 at 05:53:52PM +0800, Heng Qi wrote:
>> On Mon, 24 Jun 2024 11:04:43 +0200, Jiri Pirko <jiri@resnulli.us> wrote:
>> > From: Jiri Pirko <jiri@nvidia.com>
>> > 
>> > Currently the admin queue command execution is serialized by a lock.
>> > This patchsets lifts this limitation allowing to execute admin queue
>> > commands in parallel. To do that, admin queue processing needs to be
>> > converted from polling to interrupt based completion.
>> > 
>> > Patches #1-#6 are preparations, making things a bit smoother as well.
>> > Patch #7 implements interrupt based completion for admin queue.
>> 
>> Hi, Jiri
>> 
>> Before this set, I pushed the cvq irq set [1], and the discussion focused on the
>> fact that the newly added irq vector may cause the IO queue to fall back to
>> shared interrupt mode.
>> But it is true that devices implemented according to the specification should
>> not encounter this problem. So what do you think?
>> 
>> [1] https://lore.kernel.org/all/20240619171708-mutt-send-email-mst@kernel.org/
>
>It's true - this can cause guest to run out of vectors for a variety of
>reasons.
>
>First we have guest irqs - I am guessing avq could use IRQF_SHARED ?
>I am not sure why we don't allow IRQF_SHARED for the config
>interrupt though. So I think addressing this part can be deferred.
>
>Second, we might not have enough msix vectors on the device. Here sharing
>with e.g. cvq and further with config interrupt would make sense.
>
>Jiri do you think you can help Heng Qi hammer out a solution for cvq?
>I feel this will work will then benefit in a similar way,
>and having us poll aggressively for cvq but not admin commands
>does not make much sense, right?


How about to enhance existing fallback in vp_find_vqs() to something
like this:

        /* Try MSI-X with one vector per queue. */
        err = vp_find_vqs_msix(vdev, nvqs, vqs, callbacks, names, ONE_FOR_EACH_VQ, ctx, desc);
        if (!err)
                return 0;
        /* Try MSI-X with one vector per data queue, share one for the
         * rest of the queues and config. */
        err = vp_find_vqs_msix(vdev, nvqs, vqs, callbacks, names, ONE_FOR_EACH_DATA_VQ_ONE_FOR_REST, ctx, desc);
        if (!err)
                return 0;
        /* Fallback: MSI-X with one vector for config, one shared for queues. */
        err = vp_find_vqs_msix(vdev, nvqs, vqs, callbacks, names, ONE_DATA_VQS_ONE_FOR_REST, ctx, desc);
        if (!err)
                return 0;

I can implement this for admin queue, control queue can benefit from
this eventually.



>
>> > Patch #8 finally removes the admin queue serialization lock.
>> > 
>> > Jiri Pirko (8):
>> >   virtio_pci: push out single vq find code to vp_find_one_vq_msix()
>> >   virtio_pci_modern: treat vp_dev->admin_vq.info.vq pointer as static
>> >   virtio: push out code to vp_avq_index()
>> >   virtio: create admin queues alongside other virtqueues
>> >   virtio_pci_modern: create admin queue of queried size
>> >   virtio_pci_modern: pass cmd as an identification token
>> >   virtio_pci_modern: use completion instead of busy loop to wait on
>> >     admin cmd result
>> >   virtio_pci_modern: remove admin queue serialization lock
>> > 
>> >  drivers/virtio/virtio.c            |  28 +----
>> >  drivers/virtio/virtio_pci_common.c | 109 ++++++++++++++------
>> >  drivers/virtio/virtio_pci_common.h |   9 +-
>> >  drivers/virtio/virtio_pci_modern.c | 160 ++++++++++++-----------------
>> >  include/linux/virtio.h             |   2 +
>> >  include/linux/virtio_config.h      |   2 -
>> >  6 files changed, 150 insertions(+), 160 deletions(-)
>> > 
>> > -- 
>> > 2.45.1
>> > 
>> > 
>