Message-ID: <8fc4e1f156b075ec3f4c65acc1e10439f767bf81.camel@nvidia.com>
Date: Mon, 18 Dec 2023 13:58:22 +0000
From: Dragos Tatulea <dtatulea@...dia.com>
To: "eperezma@...hat.com" <eperezma@...hat.com>
CC: "xuanzhuo@...ux.alibaba.com" <xuanzhuo@...ux.alibaba.com>, Parav Pandit
	<parav@...dia.com>, Gal Pressman <gal@...dia.com>,
	"virtualization@...ts.linux-foundation.org"
	<virtualization@...ts.linux-foundation.org>, "linux-kernel@...r.kernel.org"
	<linux-kernel@...r.kernel.org>, "si-wei.liu@...cle.com"
	<si-wei.liu@...cle.com>, "kvm@...r.kernel.org" <kvm@...r.kernel.org>,
	"mst@...hat.com" <mst@...hat.com>, Saeed Mahameed <saeedm@...dia.com>,
	"jasowang@...hat.com" <jasowang@...hat.com>, "leon@...nel.org"
	<leon@...nel.org>
Subject: Re: [PATCH vhost v2 4/8] vdpa/mlx5: Mark vq addrs for modification in
 hw vq

On Mon, 2023-12-18 at 13:06 +0100, Eugenio Perez Martin wrote:
> On Mon, Dec 18, 2023 at 11:52 AM Dragos Tatulea <dtatulea@...dia.com> wrote:
> > 
> > On Mon, 2023-12-18 at 11:16 +0100, Eugenio Perez Martin wrote:
> > > On Sat, Dec 16, 2023 at 12:03 PM Dragos Tatulea <dtatulea@...dia.com> wrote:
> > > > 
> > > > On Fri, 2023-12-15 at 18:56 +0100, Eugenio Perez Martin wrote:
> > > > > On Fri, Dec 15, 2023 at 3:13 PM Dragos Tatulea <dtatulea@...dia.com> wrote:
> > > > > > 
> > > > > > On Fri, 2023-12-15 at 12:35 +0000, Dragos Tatulea wrote:
> > > > > > > On Thu, 2023-12-14 at 19:30 +0100, Eugenio Perez Martin wrote:
> > > > > > > > On Thu, Dec 14, 2023 at 4:51 PM Dragos Tatulea <dtatulea@...dia.com> wrote:
> > > > > > > > > 
> > > > > > > > > On Thu, 2023-12-14 at 08:45 -0500, Michael S. Tsirkin wrote:
> > > > > > > > > > On Thu, Dec 14, 2023 at 01:39:55PM +0000, Dragos Tatulea wrote:
> > > > > > > > > > > On Tue, 2023-12-12 at 15:44 -0800, Si-Wei Liu wrote:
> > > > > > > > > > > > 
> > > > > > > > > > > > On 12/12/2023 11:21 AM, Eugenio Perez Martin wrote:
> > > > > > > > > > > > > On Tue, Dec 5, 2023 at 11:46 AM Dragos Tatulea <dtatulea@...dia.com> wrote:
> > > > > > > > > > > > > > Addresses get set by .set_vq_address. hw vq addresses will be updated on
> > > > > > > > > > > > > > next modify_virtqueue.
> > > > > > > > > > > > > > 
> > > > > > > > > > > > > > Signed-off-by: Dragos Tatulea <dtatulea@...dia.com>
> > > > > > > > > > > > > > Reviewed-by: Gal Pressman <gal@...dia.com>
> > > > > > > > > > > > > > Acked-by: Eugenio Pérez <eperezma@...hat.com>
> > > > > > > > > > > > > I'm kind of ok with this patch and the next one about state, but I
> > > > > > > > > > > > > didn't ack them in the previous series.
> > > > > > > > > > > > > 
> > > > > > > > > > > > > My main concern is that it is not valid to change the vq address after
> > > > > > > > > > > > > DRIVER_OK in VirtIO, which vDPA follows. Only memory maps are ok to
> > > > > > > > > > > > > change at this moment. I'm not sure about vq state in vDPA, but vhost
> > > > > > > > > > > > > forbids changing it with an active backend.
> > > > > > > > > > > > > 
> > > > > > > > > > > > > Suspend is not defined in VirtIO at this moment though, so maybe it is
> > > > > > > > > > > > > ok to decide that all of these parameters may change during suspend.
> > > > > > > > > > > > > Maybe the best thing is to protect this with a vDPA feature flag.
> > > > > > > > > > > > I think protect with vDPA feature flag could work, while on the other
> > > > > > > > > > > > hand vDPA means vendor specific optimization is possible around suspend
> > > > > > > > > > > > and resume (in case it helps performance), which doesn't have to be
> > > > > > > > > > > > backed by virtio spec. Same applies to vhost user backend features,
> > > > > > > > > > > > variations there were not backed by spec either. Of course, we should
> > > > > > > > > > > > try best to make the default behavior backward compatible with
> > > > > > > > > > > > virtio-based backend, but that circles back to no suspend definition in
> > > > > > > > > > > > the current virtio spec, for which I hope we don't cease development on
> > > > > > > > > > > > vDPA indefinitely. After all, the virtio based vdap backend can well
> > > > > > > > > > > > define its own feature flag to describe (minor difference in) the
> > > > > > > > > > > > suspend behavior based on the later spec once it is formed in future.
> > > > > > > > > > > > 
> > > > > > > > > > > So what is the way forward here? From what I understand the options are:
> > > > > > > > > > > 
> > > > > > > > > > > 1) Add a vdpa feature flag for changing device properties while suspended.
> > > > > > > > > > > 
> > > > > > > > > > > 2) Drop these 2 patches from the series for now. Not sure if this makes sense as
> > > > > > > > > > > this. But then Si-Wei's qemu device suspend/resume poc [0] that exercises this
> > > > > > > > > > > code won't work anymore. This means the series would be less well tested.
> > > > > > > > > > > 
> > > > > > > > > > > Are there other possible options? What do you think?
> > > > > > > > > > > 
> > > > > > > > > > > [0] https://github.com/siwliu-kernel/qemu/tree/svq-resume-wip
> > > > > > > > > > 
> > > > > > > > > > I am fine with either of these.
> > > > > > > > > > 
> > > > > > > > > How about allowing the change only under the following conditions:
> > > > > > > > >   vhost_vdpa_can_suspend && vhost_vdpa_can_resume &&
> > > > > > > > > VHOST_BACKEND_F_ENABLE_AFTER_DRIVER_OK is set
> > > > > > > > > 
> > > > > > > > > ?
> > > > > > > > 
> > > > > > > > I think the best option by far is 1, as there is no hint in the
> > > > > > > > combination of these 3 indicating that you can change device
> > > > > > > > properties in the suspended state.
> > > > > > > > 
> > > > > > > Sure. Will respin a v3 without these two patches.
> > > > > > > 
> > > > > > > Another series can implement option 2 and add these 2 patches on top.
> > > > > > Hmm...I misunderstood your statement and sent a erroneus v3. You said that
> > > > > > having a feature flag is the best option.
> > > > > > 
> > > > > > Will add a feature flag in v4: is this similar to the
> > > > > > VHOST_BACKEND_F_ENABLE_AFTER_DRIVER_OK flag?
> > > > > > 
> > > > > 
> > > > > Right, it should be easy to return it from .get_backend_features op if
> > > > > the FW returns that capability, isn't it?
> > > > > 
> > > > Yes, that's easy. But I wonder if we need one feature bit for each type of
> > > > change:
> > > > - VHOST_BACKEND_F_CHANGEABLE_VQ_ADDR_IN_SUSPEND
> > > > - VHOST_BACKEND_F_CHANGEABLE_VQ_STATE_IN_SUSPEND
> > > > 
> > > 
> > > I'd say yes. Although we could configure SVQ initial state in userland
> > > as different than 0 for this first step, it would be needed in the
> > > long term.
> > > 
> > > > Or would a big one VHOST_BACKEND_F_CAN_RECONFIG_VQ_IN_SUSPEND suffice?
> > > > 
> > > 
> > > I'd say "reconfig vq" is not valid as mlx driver doesn't allow
> > > changing queue sizes, for example, isn't it?
> > > 
> > Modifying the queue size for a vq is indeed not supported by the mlx device.
> > 
> > > To define that it is
> > > valid to change "all parameters" seems very confident.
> > > 
> > Ack
> > 
> > > > To me having individual feature bits makes sense. But it could also takes too
> > > > many bits if more changes are required.
> > > > 
> > > 
> > > Yes, that's a good point. Maybe it is valid to define a subset of
> > > features that can be changed., but I think it is way clearer to just
> > > check for individual feature bits.
> > > 
> > I will prepare extra patches with the 2 feature bits approach.
> > 
> > Is it necessary to add checks in the vdpa core that block changing these
> > properties if the state is driver ok and the device doesn't support the feature?
> > 
> 
> Yes, I think it is better to protect for changes in vdpa core.
> 
Hmmm... there is no suspended state available. I would only add checks for the
DRIVER_OK state of the device because adding a is_suspended state/op seems out
of scope for this series. Any thoughts?

Powered by blists - more mailing lists