Re: [PATCH vhost 20/23] vdpa/mlx5: Pre-create hardware VQs at vdpa .dev_add time

[Dragos Tatulea]

Hi Zhu Yanjun,

On Mon, 2024-07-08 at 18:22 +0200, Zhu Yanjun wrote:
> 在 2024/6/17 17:07, Dragos Tatulea 写道:
> > Currently, hardware VQs are created right when the vdpa device gets into
> > DRIVER_OK state. That is easier because most of the VQ state is known by
> > then.
> > 
> > This patch switches to creating all VQs and their associated resources
> > at device creation time. The motivation is to reduce the vdpa device
> > live migration downtime by moving the expensive operation of creating
> > all the hardware VQs and their associated resources out of downtime on
> > the destination VM.
> 
> Hi, Dragos Tatulea
> 
>  From the above, when a device is created, all the VQs and their 
> associated resources are also created.
> If VM live migration does not occur, how much resources are wasted?
> 
> I mean, to achieve a better downtime, how much resource are used?
> 
When you use the vdpa device there are no resources wasted. The HW VQs that were
previously created at VM boot (during DRIVER_OK state) are now created at vdpa
device add time.

The trade-off here is that if you configure different VQ sizes then you will pay
the price of re-creating the VQs.

This could be mitigated by adding a default VQ size parameter that is setable
via the vdpa tool. But this part is not implemented in this series.

Ah, one more thing to keep in mind: the MSIX interrupts will be now allocated at
vdpa device creation time instead of VM startup.

> "
> On a 64 CPU, 256 GB VM with 1 vDPA device of 16 VQps, the full VQ
> resource creation + resume time was ~370ms. Now it's down to 60 ms
> (only VQ config and resume). The measurements were done on a ConnectX6DX
> based vDPA device.
> "
>  From the above, the performance is amazing.
> If we expect to use it in the production hosts, how much resources 
> should we prepare to achieve this downtime?
> 

You do need to have the latest FW (22.41.1000) to be able to get the full
benefit of the optimization.

Thanks,
Dragos
> Zhu Yanjun
> 
> > 
> > The VQs are now created in a blank state. The VQ configuration will
> > happen later, on DRIVER_OK. Then the configuration will be applied when
> > the VQs are moved to the Ready state.
> > 
> > When .set_vq_ready() is called on a VQ before DRIVER_OK, special care is
> > needed: now that the VQ is already created a resume_vq() will be
> > triggered too early when no mr has been configured yet. Skip calling
> > resume_vq() in this case, let it be handled during DRIVER_OK.
> > 
> > For virtio-vdpa, the device configuration is done earlier during
> > .vdpa_dev_add() by vdpa_register_device(). Avoid calling
> > setup_vq_resources() a second time in that case.
> > 
> > Signed-off-by: Dragos Tatulea <dtatulea@xxxxxxxxxx>
> > Reviewed-by: Cosmin Ratiu <cratiu@xxxxxxxxxx>
> > ---
> >   drivers/vdpa/mlx5/net/mlx5_vnet.c | 37 ++++++++++++++++++++++++++++++++-----
> >   1 file changed, 32 insertions(+), 5 deletions(-)
> > 
> > diff --git a/drivers/vdpa/mlx5/net/mlx5_vnet.c b/drivers/vdpa/mlx5/net/mlx5_vnet.c
> > index 249b5afbe34a..b2836fd3d1dd 100644
> > --- a/drivers/vdpa/mlx5/net/mlx5_vnet.c
> > +++ b/drivers/vdpa/mlx5/net/mlx5_vnet.c
> > @@ -2444,7 +2444,7 @@ static void mlx5_vdpa_set_vq_ready(struct vdpa_device *vdev, u16 idx, bool ready
> >   	mvq = &ndev->vqs[idx];
> >   	if (!ready) {
> >   		suspend_vq(ndev, mvq);
> > -	} else {
> > +	} else if (mvdev->status & VIRTIO_CONFIG_S_DRIVER_OK) {
> >   		if (resume_vq(ndev, mvq))
> >   			ready = false;
> >   	}
> > @@ -3078,10 +3078,18 @@ static void mlx5_vdpa_set_status(struct vdpa_device *vdev, u8 status)
> >   				goto err_setup;
> >   			}
> >   			register_link_notifier(ndev);
> > -			err = setup_vq_resources(ndev, true);
> > -			if (err) {
> > -				mlx5_vdpa_warn(mvdev, "failed to setup driver\n");
> > -				goto err_driver;
> > +			if (ndev->setup) {
> > +				err = resume_vqs(ndev);
> > +				if (err) {
> > +					mlx5_vdpa_warn(mvdev, "failed to resume VQs\n");
> > +					goto err_driver;
> > +				}
> > +			} else {
> > +				err = setup_vq_resources(ndev, true);
> > +				if (err) {
> > +					mlx5_vdpa_warn(mvdev, "failed to setup driver\n");
> > +					goto err_driver;
> > +				}
> >   			}
> >   		} else {
> >   			mlx5_vdpa_warn(mvdev, "did not expect DRIVER_OK to be cleared\n");
> > @@ -3142,6 +3150,7 @@ static int mlx5_vdpa_compat_reset(struct vdpa_device *vdev, u32 flags)
> >   		if (mlx5_vdpa_create_dma_mr(mvdev))
> >   			mlx5_vdpa_warn(mvdev, "create MR failed\n");
> >   	}
> > +	setup_vq_resources(ndev, false);
> >   	up_write(&ndev->reslock);
> >   
> >   	return 0;
> > @@ -3836,8 +3845,21 @@ static int mlx5_vdpa_dev_add(struct vdpa_mgmt_dev *v_mdev, const char *name,
> >   		goto err_reg;
> >   
> >   	mgtdev->ndev = ndev;
> > +
> > +	/* For virtio-vdpa, the device was set up during device register. */
> > +	if (ndev->setup)
> > +		return 0;
> > +
> > +	down_write(&ndev->reslock);
> > +	err = setup_vq_resources(ndev, false);
> > +	up_write(&ndev->reslock);
> > +	if (err)
> > +		goto err_setup_vq_res;
> > +
> >   	return 0;
> >   
> > +err_setup_vq_res:
> > +	_vdpa_unregister_device(&mvdev->vdev);
> >   err_reg:
> >   	destroy_workqueue(mvdev->wq);
> >   err_res2:
> > @@ -3863,6 +3885,11 @@ static void mlx5_vdpa_dev_del(struct vdpa_mgmt_dev *v_mdev, struct vdpa_device *
> >   
> >   	unregister_link_notifier(ndev);
> >   	_vdpa_unregister_device(dev);
> > +
> > +	down_write(&ndev->reslock);
> > +	teardown_vq_resources(ndev);
> > +	up_write(&ndev->reslock);
> > +
> >   	wq = mvdev->wq;
> >   	mvdev->wq = NULL;
> >   	destroy_workqueue(wq);
> > 
>