[v3,00/11] VDUSE: Improve performance

  Hi all,

This series introduces some ways to improve VDUSE performance.

Patch 1 ~ 6 bring current interrupt affinity spreading mechanism
to vduse device and make it possible for the virtio-blk driver
to build the blk-mq queues based on it. This would be useful to
mitigate the virtqueue lock contention in virtio-blk driver. In
our test, with those patches, we could get ~50% improvement (600k
iops -> 900k iops) when using per-cpu virtqueue.

Patch 7 adds a sysfs interface for each vduse virtqueue to change
the affinity for IRQ callback. It would be helpful for performance
tuning when the affinity mask contains more than one CPU.

Patch 8 ~ 9 associate an eventfd to the vdpa callback so that
we can signal it directly during irq injection without scheduling
an additional workqueue thread to do that.

Patch 10, 11 add a sysfs interface to support specifying bounce
buffer size in virtio-vdpa case. The high throughput workloads
can benefit from it. And we can also use it to reduce the memory
overhead for small throughput workloads.

Please review, thanks!

V2 to V3:
- Rebased to newest kernel tree
- Export group_cpus_evenly() instead of irq_create_affinity_masks() [MST]
- Remove the sysfs for workqueue control [Jason]
- Associate an eventfd to the vdpa callback [Jason]
- Signal the eventfd directly in vhost-vdpa case [Jason]
- Use round-robin to spread IRQs between CPUs in the affinity mask [Jason]
- Handle the cpu hotplug case on IRQ injection [Jason]
- Remove effective IRQ affinity and balance mechanism for IRQ allocation

V1 to V2:
- Export irq_create_affinity_masks()
- Add set/get_vq_affinity and set_irq_affinity callbacks in vDPA
  framework
- Add automatic irq callback affinity support in VDUSE driver [Jason]
- Add more backgrounds information in commit log [Jason]
- Only support changing effective affinity when the value is a subset
  of the IRQ callback affinity mask

Xie Yongji (11):
  lib/group_cpus: Export group_cpus_evenly()
  vdpa: Add set/get_vq_affinity callbacks in vdpa_config_ops
  vdpa: Add set_irq_affinity callback in vdpa_config_ops
  vduse: Refactor allocation for vduse virtqueues
  vduse: Support automatic irq callback affinity
  vduse: Support set/get_vq_affinity callbacks
  vduse: Add sysfs interface for irq callback affinity
  vdpa: Add eventfd for the vdpa callback
  vduse: Signal interrupt's eventfd directly in vhost-vdpa case
  vduse: Delay iova domain creation
  vduse: Support specifying bounce buffer size via sysfs

 drivers/vdpa/vdpa_user/vduse_dev.c | 490 +++++++++++++++++++++++++----
 drivers/vhost/vdpa.c               |   2 +
 drivers/virtio/virtio_vdpa.c       |  33 ++
 include/linux/vdpa.h               |  25 ++
 lib/group_cpus.c                   |   1 +
 5 files changed, 488 insertions(+), 63 deletions(-)
  

On Tue, Feb 28, 2023 at 05:40:59PM +0800, Xie Yongji wrote:
> Hi all,
> 
> This series introduces some ways to improve VDUSE performance.


Pls fix warnings reported by 0-day infra, dropping this for now.


> Patch 1 ~ 6 bring current interrupt affinity spreading mechanism
> to vduse device and make it possible for the virtio-blk driver
> to build the blk-mq queues based on it. This would be useful to
> mitigate the virtqueue lock contention in virtio-blk driver. In
> our test, with those patches, we could get ~50% improvement (600k
> iops -> 900k iops) when using per-cpu virtqueue.
> 
> Patch 7 adds a sysfs interface for each vduse virtqueue to change
> the affinity for IRQ callback. It would be helpful for performance
> tuning when the affinity mask contains more than one CPU.
> 
> Patch 8 ~ 9 associate an eventfd to the vdpa callback so that
> we can signal it directly during irq injection without scheduling
> an additional workqueue thread to do that.
> 
> Patch 10, 11 add a sysfs interface to support specifying bounce
> buffer size in virtio-vdpa case. The high throughput workloads
> can benefit from it. And we can also use it to reduce the memory
> overhead for small throughput workloads.
> 
> Please review, thanks!
> 
> V2 to V3:
> - Rebased to newest kernel tree
> - Export group_cpus_evenly() instead of irq_create_affinity_masks() [MST]
> - Remove the sysfs for workqueue control [Jason]
> - Associate an eventfd to the vdpa callback [Jason]
> - Signal the eventfd directly in vhost-vdpa case [Jason]
> - Use round-robin to spread IRQs between CPUs in the affinity mask [Jason]
> - Handle the cpu hotplug case on IRQ injection [Jason]
> - Remove effective IRQ affinity and balance mechanism for IRQ allocation
> 
> V1 to V2:
> - Export irq_create_affinity_masks()
> - Add set/get_vq_affinity and set_irq_affinity callbacks in vDPA
>   framework
> - Add automatic irq callback affinity support in VDUSE driver [Jason]
> - Add more backgrounds information in commit log [Jason]
> - Only support changing effective affinity when the value is a subset
>   of the IRQ callback affinity mask
> 
> Xie Yongji (11):
>   lib/group_cpus: Export group_cpus_evenly()
>   vdpa: Add set/get_vq_affinity callbacks in vdpa_config_ops
>   vdpa: Add set_irq_affinity callback in vdpa_config_ops
>   vduse: Refactor allocation for vduse virtqueues
>   vduse: Support automatic irq callback affinity
>   vduse: Support set/get_vq_affinity callbacks
>   vduse: Add sysfs interface for irq callback affinity
>   vdpa: Add eventfd for the vdpa callback
>   vduse: Signal interrupt's eventfd directly in vhost-vdpa case
>   vduse: Delay iova domain creation
>   vduse: Support specifying bounce buffer size via sysfs
> 
>  drivers/vdpa/vdpa_user/vduse_dev.c | 490 +++++++++++++++++++++++++----
>  drivers/vhost/vdpa.c               |   2 +
>  drivers/virtio/virtio_vdpa.c       |  33 ++
>  include/linux/vdpa.h               |  25 ++
>  lib/group_cpus.c                   |   1 +
>  5 files changed, 488 insertions(+), 63 deletions(-)
> 
> -- 
> 2.20.1
  

On Fri, Mar 10, 2023 at 4:50 PM Michael S. Tsirkin <mst@redhat.com> wrote:
>
> On Tue, Feb 28, 2023 at 05:40:59PM +0800, Xie Yongji wrote:
> > Hi all,
> >
> > This series introduces some ways to improve VDUSE performance.
>
>
> Pls fix warnings reported by 0-day infra, dropping this for now.

Note that I plan to review this next week.

Thanks

>
>
> > Patch 1 ~ 6 bring current interrupt affinity spreading mechanism
> > to vduse device and make it possible for the virtio-blk driver
> > to build the blk-mq queues based on it. This would be useful to
> > mitigate the virtqueue lock contention in virtio-blk driver. In
> > our test, with those patches, we could get ~50% improvement (600k
> > iops -> 900k iops) when using per-cpu virtqueue.
> >
> > Patch 7 adds a sysfs interface for each vduse virtqueue to change
> > the affinity for IRQ callback. It would be helpful for performance
> > tuning when the affinity mask contains more than one CPU.
> >
> > Patch 8 ~ 9 associate an eventfd to the vdpa callback so that
> > we can signal it directly during irq injection without scheduling
> > an additional workqueue thread to do that.
> >
> > Patch 10, 11 add a sysfs interface to support specifying bounce
> > buffer size in virtio-vdpa case. The high throughput workloads
> > can benefit from it. And we can also use it to reduce the memory
> > overhead for small throughput workloads.
> >
> > Please review, thanks!
> >
> > V2 to V3:
> > - Rebased to newest kernel tree
> > - Export group_cpus_evenly() instead of irq_create_affinity_masks() [MST]
> > - Remove the sysfs for workqueue control [Jason]
> > - Associate an eventfd to the vdpa callback [Jason]
> > - Signal the eventfd directly in vhost-vdpa case [Jason]
> > - Use round-robin to spread IRQs between CPUs in the affinity mask [Jason]
> > - Handle the cpu hotplug case on IRQ injection [Jason]
> > - Remove effective IRQ affinity and balance mechanism for IRQ allocation
> >
> > V1 to V2:
> > - Export irq_create_affinity_masks()
> > - Add set/get_vq_affinity and set_irq_affinity callbacks in vDPA
> >   framework
> > - Add automatic irq callback affinity support in VDUSE driver [Jason]
> > - Add more backgrounds information in commit log [Jason]
> > - Only support changing effective affinity when the value is a subset
> >   of the IRQ callback affinity mask
> >
> > Xie Yongji (11):
> >   lib/group_cpus: Export group_cpus_evenly()
> >   vdpa: Add set/get_vq_affinity callbacks in vdpa_config_ops
> >   vdpa: Add set_irq_affinity callback in vdpa_config_ops
> >   vduse: Refactor allocation for vduse virtqueues
> >   vduse: Support automatic irq callback affinity
> >   vduse: Support set/get_vq_affinity callbacks
> >   vduse: Add sysfs interface for irq callback affinity
> >   vdpa: Add eventfd for the vdpa callback
> >   vduse: Signal interrupt's eventfd directly in vhost-vdpa case
> >   vduse: Delay iova domain creation
> >   vduse: Support specifying bounce buffer size via sysfs
> >
> >  drivers/vdpa/vdpa_user/vduse_dev.c | 490 +++++++++++++++++++++++++----
> >  drivers/vhost/vdpa.c               |   2 +
> >  drivers/virtio/virtio_vdpa.c       |  33 ++
> >  include/linux/vdpa.h               |  25 ++
> >  lib/group_cpus.c                   |   1 +
> >  5 files changed, 488 insertions(+), 63 deletions(-)
> >
> > --
> > 2.20.1
>