Re: [PATCH v6 1/2] vhost-user: remove VirtQ notifier restore

[Michael S. Tsirkin]

I dropped this for now as I'm a bit lost with this patchset.
Cc Raphael maybe he'll understand it better.

On Wed, Jan 12, 2022 at 03:05:15PM +0000, Xueming(Steven) Li wrote:
> On Wed, 2021-11-03 at 16:30 -0400, Michael S. Tsirkin wrote:
> > On Wed, Nov 03, 2021 at 02:48:41PM +0000, Xueming(Steven) Li wrote:
> > > On Tue, 2021-11-02 at 02:49 -0400, Michael S. Tsirkin wrote:
> > > > On Tue, Nov 02, 2021 at 06:08:39AM +0000, Xueming(Steven) Li wrote:
> > > > > On Mon, 2021-11-01 at 17:06 -0400, Michael S. Tsirkin wrote:
> > > > > > On Mon, Nov 01, 2021 at 04:38:12PM +0800, Xueming Li wrote:
> > > > > > > When vhost-user vdpa client suspend, backend may close all 
> > > > > > > resources,
> > > > > > > VQ notifier mmap address become invalid, restore MR which contains
> > > > > > > the invalid address is wrong. vdpa client will set VQ notifier 
> > > > > > > after
> > > > > > > reconnect.
> > > > > > > 
> > > > > > > This patch removes VQ notifier restore and related flags to avoid 
> > > > > > > reusing
> > > > > > > invalid address.
> > > > > > > 
> > > > > > > Fixes: 44866521bd6e ("vhost-user: support registering external 
> > > > > > > host notifiers")
> > > > > > > Cc: qemu-stable@nongnu.org
> > > > > > > Cc: Yuwei Zhang <zhangyuwei.9149@bytedance.com>
> > > > > > > Signed-off-by: Xueming Li <xuemingl@nvidia.com>
> > > > > > > ---
> > > > > > >  hw/virtio/vhost-user.c         | 19 +------------------
> > > > > > >  include/hw/virtio/vhost-user.h |  1 -
> > > > > > >  2 files changed, 1 insertion(+), 19 deletions(-)
> > > > > > > 
> > > > > > > diff --git a/hw/virtio/vhost-user.c b/hw/virtio/vhost-user.c
> > > > > > > index bf6e50223c..c671719e9b 100644
> > > > > > > --- a/hw/virtio/vhost-user.c
> > > > > > > +++ b/hw/virtio/vhost-user.c
> > > > > > > @@ -1143,19 +1143,6 @@ static int vhost_user_set_vring_num(struct 
> > > > > > > vhost_dev *dev,
> > > > > > >      return vhost_set_vring(dev, VHOST_USER_SET_VRING_NUM, ring);
> > > > > > >  }
> > > > > > >  
> > > > > > > -static void vhost_user_host_notifier_restore(struct vhost_dev 
> > > > > > > *dev,
> > > > > > > -                                             int queue_idx)
> > > > > > > -{
> > > > > > > -    struct vhost_user *u = dev->opaque;
> > > > > > > -    VhostUserHostNotifier *n = &u->user->notifier[queue_idx];
> > > > > > > -    VirtIODevice *vdev = dev->vdev;
> > > > > > > -
> > > > > > > -    if (n->addr && !n->set) {
> > > > > > > -        virtio_queue_set_host_notifier_mr(vdev, queue_idx, 
> > > > > > > &n->mr, true);
> > > > > > > -        n->set = true;
> > > > > > > -    }
> > > > > > > -}
> > > > > > > -
> > > > > > >  static void vhost_user_host_notifier_remove(struct vhost_dev 
> > > > > > > *dev,
> > > > > > >                                              int queue_idx)
> > > > > > >  {
> > > > > > > @@ -1163,17 +1150,14 @@ static void 
> > > > > > > vhost_user_host_notifier_remove(struct vhost_dev *dev,
> > > > > > >      VhostUserHostNotifier *n = &u->user->notifier[queue_idx];
> > > > > > >      VirtIODevice *vdev = dev->vdev;
> > > > > > >  
> > > > > > > -    if (n->addr && n->set) {
> > > > > > > +    if (n->addr) {
> > > > > > >          virtio_queue_set_host_notifier_mr(vdev, queue_idx, 
> > > > > > > &n->mr, false);
> > > > > > > -        n->set = false;
> > > > > > >      }
> > > > > > >  }
> > > > > > > 
> > > > > > 
> > > > > > So on vq stop we still remove the notifier...
> > > > > >   
> > > > > > >  static int vhost_user_set_vring_base(struct vhost_dev *dev,
> > > > > > >                                       struct vhost_vring_state 
> > > > > > > *ring)
> > > > > > >  {
> > > > > > > -    vhost_user_host_notifier_restore(dev, ring->index);
> > > > > > > -
> > > > > > >      return vhost_set_vring(dev, VHOST_USER_SET_VRING_BASE, ring);
> > > > > > >  }
> > > > > > >  
> > > > > > 
> > > > > > but on vq start we do not reinstate it? Does this not mean that
> > > > > > notifiers won't work after stop then start?
> > > > > 
> > > > > 
> > > > > Yes, backend initially work w/o host notifier, request VM driver to
> > > > > install notifier if needed - call this function through slave socket.
> > > > 
> > > > I think it's cleaner if qemu handles this itself like it did before, it
> > > > knows vm is stopped without getting called.
> > > 
> > > If vhost play as server, there are 2 scenario that remove the notifier:
> > > 1. VM suspend: backend still there, it's okay to keep mmap address.
> > > 2. vhost backend stopped or process killed: resources from backend
> > > should be released. That's why patch 2/2 munmap in notifier remove
> > > function. Then the restore function get nothing to restore, maybe I
> > > shouldn't reverse patch order.
> > 
> > I can't say I understand what you mean here. Do you plan to change
> > the patchset in some way?
> > When you do, pls include a cover letter with a changelog and
> > Cc all people you include on patches on the cover letter too. 
> 
> Here is the detail of the problem I encountered, my vhost backend is
> DPDK vdpa user space application. Notifier address is set when vdpa ask
> qemu to mmap a FD and an offset from vdpa, see function
> vhost_user_slave_handle_vring_host_notifier(). If the vdpa application
> restart of get killed for some reason,
> vhost_user_host_notifier_remove() is called and notifier MR is
> uninstalled, the notifier address that retrieved from mmap is
> referencing to an invalid FD, not released. This will cause HW
> resources on kernel side still referenced, most important, when vdpa
> connection restored, this invalid notifier will be be restored as
> notifier MR.
> 
> To resolve it, have to remove the notifer restore mechanism, vDPA
> application will issue client socket request again to install notifier
> to VM, so no concern that the notifier will be lost after resume.
> 
> Since vdpa might be killed, no chance to notify qemu to remove
> notifier. An alternative solution is to detect sock disconnection and
> unmmap notifier, but it looks more complex to me. How do you think?
> 
> 
> > 
> > > > 
> > > > > > 
> > > > > > 
> > > > > > > @@ -1538,7 +1522,6 @@ static int 
> > > > > > > vhost_user_slave_handle_vring_host_notifier(struct vhost_dev *dev,
> > > > > > >      }
> > > > > > >  
> > > > > > >      n->addr = addr;
> > > > > > > -    n->set = true;
> > > > > > >  
> > > > > > >      return 0;
> > > > > > >  }
> > > > > > > diff --git a/include/hw/virtio/vhost-user.h 
> > > > > > > b/include/hw/virtio/vhost-user.h
> > > > > > > index a9abca3288..f6012b2078 100644
> > > > > > > --- a/include/hw/virtio/vhost-user.h
> > > > > > > +++ b/include/hw/virtio/vhost-user.h
> > > > > > > @@ -14,7 +14,6 @@
> > > > > > >  typedef struct VhostUserHostNotifier {
> > > > > > >      MemoryRegion mr;
> > > > > > >      void *addr;
> > > > > > > -    bool set;
> > > > > > >  } VhostUserHostNotifier;
> > > > > > >  
> > > > > > >  typedef struct VhostUserState {
> > > > > > > -- 
> > > > > > > 2.33.0
> > > > > > 
> > > > > 
> > > > 
> > > 
> > 
>