Message-ID: <CAJaqyWfB-nxpdtv+HRPB=7oc-MmquZ=7-mDqdVi42Mwf28CX7A@mail.gmail.com>
Date: Wed, 13 Aug 2025 12:11:52 +0200
From: Eugenio Perez Martin <eperezma@...hat.com>
To: Jason Wang <jasowang@...hat.com>
Cc: "Michael S . Tsirkin" <mst@...hat.com>, Cindy Lu <lulu@...hat.com>, 
	Yongji Xie <xieyongji@...edance.com>, Stefano Garzarella <sgarzare@...hat.com>, 
	virtualization@...ts.linux.dev, Laurent Vivier <lvivier@...hat.com>, 
	linux-kernel@...r.kernel.org, Xuan Zhuo <xuanzhuo@...ux.alibaba.com>, 
	Maxime Coquelin <mcoqueli@...hat.com>
Subject: Re: [RFC v2 2/7] vduse: add vq group support

On Tue, Aug 12, 2025 at 5:01 AM Jason Wang <jasowang@...hat.com> wrote:
>
> On Mon, Aug 11, 2025 at 5:52 PM Eugenio Perez Martin
> <eperezma@...hat.com> wrote:
> >
> > On Mon, Aug 11, 2025 at 5:10 AM Jason Wang <jasowang@...hat.com> wrote:
> > >
> > > On Thu, Aug 7, 2025 at 7:58 PM Eugenio Pérez <eperezma@...hat.com> wrote:
> > > >
> > > > This allows sepparate the different virtqueues in groups that shares the
> > > > same address space.  Asking the VDUSE device for the groups of the vq at
> > > > the beginning as they're needed for the DMA API.
> > > >
> > > > Allocating 3 vq groups as net is the device that need the most groups:
> > > > * Dataplane (guest passthrough)
> > > > * CVQ
> > > > * Shadowed vrings.
> > > >
> > > > Future versions of the series can include dynamic allocation of the
> > > > groups array so VDUSE can declare more groups.
> > > >
> > > > Signed-off-by: Eugenio Pérez <eperezma@...hat.com>
> > > > ---
> > > > v2:
> > > > * Cache group information in kernel, as we need to provide the vq map
> > > >   tokens properly.
> > > > * Add descs vq group to optimize SVQ forwarding and support indirect
> > > >   descriptors out of the box.
> > > > ---
> > > >  drivers/vdpa/vdpa_user/vduse_dev.c | 71 +++++++++++++++++++++++++++++-
> > > >  include/uapi/linux/vduse.h         | 19 +++++++-
> > > >  2 files changed, 88 insertions(+), 2 deletions(-)
> > > >
> > > > diff --git a/drivers/vdpa/vdpa_user/vduse_dev.c b/drivers/vdpa/vdpa_user/vduse_dev.c
> > > > index d858c4389cc1..d1f6d00a9c71 100644
> > > > --- a/drivers/vdpa/vdpa_user/vduse_dev.c
> > > > +++ b/drivers/vdpa/vdpa_user/vduse_dev.c
> > > > @@ -46,6 +46,11 @@
> > > >  #define VDUSE_IOVA_SIZE (VDUSE_MAX_BOUNCE_SIZE + 128 * 1024 * 1024)
> > > >  #define VDUSE_MSG_DEFAULT_TIMEOUT 30
> > > >
> > > > +/*
> > > > + * Let's make it 3 for simplicity.
> > > > + */
> > > > +#define VDUSE_MAX_VQ_GROUPS 3
> > >
> > > I think we can release this to something like 64. Otherwise we might
> > > bump the version again just to increase the limitation? Or having a
> > > sysfs entry like bounce_size?
> > >
> >
> > I think it should not be linked to the version, but it is true there
> > is no way for VDUSE devices to check the maximum VQ groups / ASID that
> > the kernel supports.
> >
> > To handle as bounce_size seems the best option, good point. I'll send
> > a new version with that!
> >
> > > > +
> > > >  #define IRQ_UNBOUND -1
> > > >
> > > >  struct vduse_virtqueue {
> > > > @@ -58,6 +63,8 @@ struct vduse_virtqueue {
> > > >         struct vdpa_vq_state state;
> > > >         bool ready;
> > > >         bool kicked;
> > > > +       u32 vq_group;
> > > > +       u32 vq_desc_group;
> > > >         spinlock_t kick_lock;
> > > >         spinlock_t irq_lock;
> > > >         struct eventfd_ctx *kickfd;
> > > > @@ -114,6 +121,7 @@ struct vduse_dev {
> > > >         u8 status;
> > > >         u32 vq_num;
> > > >         u32 vq_align;
> > > > +       u32 ngroups;
> > > >         struct vduse_umem *umem;
> > > >         struct mutex mem_lock;
> > > >         unsigned int bounce_size;
> > > > @@ -592,6 +600,20 @@ static int vduse_vdpa_set_vq_state(struct vdpa_device *vdpa, u16 idx,
> > > >         return 0;
> > > >  }
> > > >
> > > > +static u32 vduse_get_vq_group(struct vdpa_device *vdpa, u16 idx)
> > > > +{
> > > > +       struct vduse_dev *dev = vdpa_to_vduse(vdpa);
> > > > +
> > > > +       return dev->vqs[idx]->vq_group;
> > > > +}
> > > > +
> > > > +static u32 vduse_get_vq_desc_group(struct vdpa_device *vdpa, u16 idx)
> > > > +{
> > > > +       struct vduse_dev *dev = vdpa_to_vduse(vdpa);
> > > > +
> > > > +       return dev->vqs[idx]->vq_desc_group;
> > > > +}
> > > > +
> > > >  static int vduse_vdpa_get_vq_state(struct vdpa_device *vdpa, u16 idx,
> > > >                                 struct vdpa_vq_state *state)
> > > >  {
> > > > @@ -678,13 +700,48 @@ static u8 vduse_vdpa_get_status(struct vdpa_device *vdpa)
> > > >         return dev->status;
> > > >  }
> > > >
> > > > +static int vduse_fill_vq_groups(struct vduse_dev *dev)
> > > > +{
> > > > +       if (dev->api_version < VDUSE_API_VERSION_1)
> > > > +               return 0;
> > > > +
> > > > +       for (int i = 0; i < dev->vdev->vdpa.nvqs; ++i) {
> > > > +               struct vduse_dev_msg msg = { 0 };
> > > > +               int ret;
> > > > +
> > > > +               msg.req.type = VDUSE_GET_VQ_GROUP;
> > > > +               msg.req.vq_group.index = i;
> > > > +               ret = vduse_dev_msg_sync(dev, &msg);
> > >
> > > I fail to understand why the default group mapping is not done during
> > > device creation.
> > >
> >
> > Because it changes depending on the features.
> >
> > If a new device has 5 virtqueues and the device wants to isolate the
> > CVQ, the CVQ position depends on the features that the guest's acks:
> > * If MQ is acked the isolated vq is #5
> > * If MQ is not acked the isolated vq is #3.
>
> I see we are still damaged by the design of the cvq index. But this is
> just a static branch not a dynamic one. If we can find ways to make it
> static it would be better.
>
> >
> > > > +               if (ret)
> > > > +                       return ret;
> > > > +
> > > > +               dev->vqs[i]->vq_group = msg.resp.vq_group.num;
> > > > +
> > > > +               msg.req.type = VDUSE_GET_VRING_DESC_GROUP;
> > > > +               ret = vduse_dev_msg_sync(dev, &msg);
> > > > +               if (ret)
> > > > +                       return ret;
> > > > +
> > > > +               dev->vqs[i]->vq_desc_group = msg.resp.vq_group.num;
> > > > +       }
> > > > +
> > > > +       return 0;
> > > > +}
> > > > +
> > > >  static void vduse_vdpa_set_status(struct vdpa_device *vdpa, u8 status)
> > > >  {
> > > >         struct vduse_dev *dev = vdpa_to_vduse(vdpa);
> > > > +       u8 previous_status = dev->status;
> > > >
> > > >         if (vduse_dev_set_status(dev, status))
> > > >                 return;
> > > >
> > > > +       if ((dev->status ^ previous_status) &
> > > > +            BIT_ULL(VIRTIO_CONFIG_S_FEATURES_OK) &&
> > > > +           status & (1ULL << VIRTIO_CONFIG_S_FEATURES_OK))
> > > > +               if (vduse_fill_vq_groups(dev))
> > >
> > > Can we merge the two messages into a single one? Or can we use a
> > > shared memory for storing such mapping?
> > >
> > > For example, if we have 256 queues it would be very slow.
> > >
> >
> > To merge it in the same message is good to me, sure.
>
> We can start from this if we can't find a way to provision vq to group
> mapping during device creation.
>

Note that I don't think it is worth implementing these in this series,
but to add them on top in another one. Because I don't think we will
find devices with a lot of virtqueues for now. But here are some ideas
to mitigate the startup time cost:

1) Support more than one virtqueue in the same vduse request / reply
Something like:
vduse_dev_response{
  u32 req_id;
  u32 result;
  union {
    ...
    struct vduse_vq_group {
      u32 num_vqs_requested;
      struct {
        u32 vq_idx;
        u32 vq_group;
      } vqs[15];
    }
    u32 padding[32];
  }
}

Choosing 15 to fill the current size of vduse_dev_response struct.

2) Pointer chasing in the struct written

Same as previous, but the vqs struct is actually a pointer in
userspace. This way it can be arbitrarily big.

vduse_dev_response{
  u32 req_id;
  u32 result;
  union {
    ...
    struct vduse_vq_group {
      u32 num_vqs_requested;
      struct {
        u32 vq_idx;
        u32 vq_group;
      } *vqs;
    }
    u32 padding[32];
  }
}

I cannot locate any use of this in write() data, but it is more or
less common in ioctl.

3) Allow VQ_GROUP_BATCH_BEGIN and _END, similar to how memory map
works in vhost_vdpa. As many vq_group response as needed in between.

+) Assume that any vq not mentioned in the reply is vq group 0.


> > To make it a
> > table in shm seems more complicated, unless we accept a void * in the
> > reply and VDUSE uses copy_from_user. If that is ok here, then sure.
> >
>
> This looks tricky indeed.
>
> Thanks
>

Powered by blists - more mailing lists