Jason Wang
2022-Sep-07 02:09 UTC
[PATCH net] virtio-net: add cond_resched() to the command waiting loop
On Tue, Sep 6, 2022 at 6:56 PM Paolo Abeni <pabeni at redhat.com> wrote:> > On Mon, 2022-09-05 at 15:49 +0800, Jason Wang wrote: > > On Mon, Sep 5, 2022 at 3:15 PM Michael S. Tsirkin <mst at redhat.com> wrote: > > > > > > On Mon, Sep 05, 2022 at 12:53:41PM +0800, Jason Wang wrote: > > > > Adding cond_resched() to the command waiting loop for a better > > > > co-operation with the scheduler. This allows to give CPU a breath to > > > > run other task(workqueue) instead of busy looping when preemption is > > > > not allowed. > > > > > > > > What's more important. This is a must for some vDPA parent to work > > > > since control virtqueue is emulated via a workqueue for those parents. > > > > > > > > Fixes: bda324fd037a ("vdpasim: control virtqueue support") > > > > > > That's a weird commit to fix. so it fixes the simulator? > > > > Yes, since the simulator is using a workqueue to handle control virtueue. > > Uhmm... touching a driver for a simulator's sake looks a little weird.Simulator is not the only one that is using a workqueue (but should be the first). I can see that the mlx5 vDPA driver is using a workqueue as well (see mlx5_vdpa_kick_vq()). And in the case of VDUSE, it needs to wait for the response from the userspace, this means cond_resched() is probably a must for the case like UP.> > Additionally, if the bug is vdpasim, I think it's better to try to > solve it there, if possible. > > Looking at vdpasim_net_work() and vdpasim_blk_work() it looks like > neither needs a process context, so perhaps you could rework it to run > the work_fn() directly from vdpasim_kick_vq(), at least for the control > virtqueue?It's possible (but require some rework on the simulator core). But considering we have other similar use cases, it looks better to solve it in the virtio-net driver. Additionally, this may have better behaviour when using for the buggy hardware (e.g the control virtqueue takes too long to respond). We may consider switching to use interrupt/sleep in the future (but not suitable for -net). Thanks> > Thanks! > > Paolo >
Paolo Abeni
2022-Sep-07 07:07 UTC
[PATCH net] virtio-net: add cond_resched() to the command waiting loop
On Wed, 2022-09-07 at 10:09 +0800, Jason Wang wrote:> On Tue, Sep 6, 2022 at 6:56 PM Paolo Abeni <pabeni at redhat.com> wrote: > > > > On Mon, 2022-09-05 at 15:49 +0800, Jason Wang wrote: > > > On Mon, Sep 5, 2022 at 3:15 PM Michael S. Tsirkin <mst at redhat.com> wrote: > > > > > > > > On Mon, Sep 05, 2022 at 12:53:41PM +0800, Jason Wang wrote: > > > > > Adding cond_resched() to the command waiting loop for a better > > > > > co-operation with the scheduler. This allows to give CPU a breath to > > > > > run other task(workqueue) instead of busy looping when preemption is > > > > > not allowed. > > > > > > > > > > What's more important. This is a must for some vDPA parent to work > > > > > since control virtqueue is emulated via a workqueue for those parents. > > > > > > > > > > Fixes: bda324fd037a ("vdpasim: control virtqueue support") > > > > > > > > That's a weird commit to fix. so it fixes the simulator? > > > > > > Yes, since the simulator is using a workqueue to handle control virtueue. > > > > Uhmm... touching a driver for a simulator's sake looks a little weird. > > Simulator is not the only one that is using a workqueue (but should be > the first). > > I can see that the mlx5 vDPA driver is using a workqueue as well (see > mlx5_vdpa_kick_vq()). > > And in the case of VDUSE, it needs to wait for the response from the > userspace, this means cond_resched() is probably a must for the case > like UP. > > > > > Additionally, if the bug is vdpasim, I think it's better to try to > > solve it there, if possible. > > > > Looking at vdpasim_net_work() and vdpasim_blk_work() it looks like > > neither needs a process context, so perhaps you could rework it to run > > the work_fn() directly from vdpasim_kick_vq(), at least for the control > > virtqueue? > > It's possible (but require some rework on the simulator core). But > considering we have other similar use cases, it looks better to solve > it in the virtio-net driver.I see.> Additionally, this may have better behaviour when using for the buggy > hardware (e.g the control virtqueue takes too long to respond). We may > consider switching to use interrupt/sleep in the future (but not > suitable for -net).Agreed. Possibly a timeout could be useful, too. Cheers, Paolo