| lists.openwall.net | lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC | |
|
Open Source and information security mailing list archives
| ||
|
Message-ID: <DM8PR12MB5400DBB3EBA91C250E703E1EAB899@DM8PR12MB5400.namprd12.prod.outlook.com>
Date: Wed, 13 Jul 2022 05:18:09 +0000
From: Eli Cohen <elic@...dia.com>
To: Jason Wang <jasowang@...hat.com>,
Eugenio Perez Martin <eperezma@...hat.com>
CC: mst <mst@...hat.com>,
virtualization <virtualization@...ts.linux-foundation.org>,
linux-kernel <linux-kernel@...r.kernel.org>,
Si-Wei Liu <si-wei.liu@...cle.com>,
Parav Pandit <parav@...dia.com>
Subject: RE: [PATCH RFC 1/3] vdpa/mlx5: Implement susupend virtqueue callback
> From: Jason Wang <jasowang@...hat.com>
> Sent: Wednesday, July 13, 2022 6:29 AM
> To: Eugenio Perez Martin <eperezma@...hat.com>
> Cc: Eli Cohen <elic@...dia.com>; mst <mst@...hat.com>; virtualization <virtualization@...ts.linux-foundation.org>; linux-kernel
> <linux-kernel@...r.kernel.org>; Si-Wei Liu <si-wei.liu@...cle.com>; Parav Pandit <parav@...dia.com>
> Subject: Re: [PATCH RFC 1/3] vdpa/mlx5: Implement susupend virtqueue callback
>
> On Tue, Jul 12, 2022 at 5:16 PM Eugenio Perez Martin
> <eperezma@...hat.com> wrote:
> >
> > On Tue, Jul 12, 2022 at 10:14 AM Jason Wang <jasowang@...hat.com> wrote:
> > >
> > > On Mon, Jul 11, 2022 at 2:14 PM Eli Cohen <elic@...dia.com> wrote:
> > > >
> > > > > From: Jason Wang <jasowang@...hat.com>
> > > > > Sent: Tuesday, June 21, 2022 6:05 AM
> > > > > To: Eugenio Perez Martin <eperezma@...hat.com>
> > > > > Cc: Eli Cohen <elic@...dia.com>; mst <mst@...hat.com>; virtualization <virtualization@...ts.linux-foundation.org>; linux-
> kernel
> > > > > <linux-kernel@...r.kernel.org>; Si-Wei Liu <si-wei.liu@...cle.com>; Parav Pandit <parav@...dia.com>
> > > > > Subject: Re: [PATCH RFC 1/3] vdpa/mlx5: Implement susupend virtqueue callback
> > > > >
> > > > > On Mon, Jun 20, 2022 at 5:59 PM Eugenio Perez Martin
> > > > > <eperezma@...hat.com> wrote:
> > > > > >
> > > > > > On Mon, Jun 20, 2022 at 10:56 AM Jason Wang <jasowang@...hat.com> wrote:
> > > > > > >
> > > > > > > On Thu, Jun 16, 2022 at 9:27 PM Eli Cohen <elic@...dia.com> wrote:
> > > > > > > >
> > > > > > > > Implement the suspend callback allowing to suspend the virtqueues so
> > > > > > > > they stop processing descriptors. This is required to allow the shadow
> > > > > > > > virtqueue to kick in.
> > > > > > > >
> > > > > > > > Signed-off-by: Eli Cohen <elic@...dia.com>
> > > > > > > > ---
> > > > > > > > drivers/vdpa/mlx5/net/mlx5_vnet.c | 68 +++++++++++++++++++++++++++++-
> > > > > > > > include/linux/mlx5/mlx5_ifc_vdpa.h | 8 ++++
> > > > > > > > 2 files changed, 75 insertions(+), 1 deletion(-)
> > > > > > > >
> > > > > > > > diff --git a/drivers/vdpa/mlx5/net/mlx5_vnet.c b/drivers/vdpa/mlx5/net/mlx5_vnet.c
> > > > > > > > index fb0b23e71383..ea4bc8a0cd25 100644
> > > > > > > > --- a/drivers/vdpa/mlx5/net/mlx5_vnet.c
> > > > > > > > +++ b/drivers/vdpa/mlx5/net/mlx5_vnet.c
> > > > > > > > @@ -895,6 +895,7 @@ static int create_virtqueue(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_virtque
> > > > > > > > if (err)
> > > > > > > > goto err_cmd;
> > > > > > > >
> > > > > > > > + mvq->fw_state = MLX5_VIRTIO_NET_Q_OBJECT_STATE_INIT;
> > > > > > > > kfree(in);
> > > > > > > > mvq->virtq_id = MLX5_GET(general_obj_out_cmd_hdr, out, obj_id);
> > > > > > > >
> > > > > > > > @@ -922,6 +923,7 @@ static void destroy_virtqueue(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_virtq
> > > > > > > > mlx5_vdpa_warn(&ndev->mvdev, "destroy virtqueue 0x%x\n", mvq->virtq_id);
> > > > > > > > return;
> > > > > > > > }
> > > > > > > > + mvq->fw_state = MLX5_VIRTIO_NET_Q_OBJECT_NONE;
> > > > > > > > umems_destroy(ndev, mvq);
> > > > > > > > }
> > > > > > > >
> > > > > > > > @@ -1121,6 +1123,20 @@ static int query_virtqueue(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_virtqueu
> > > > > > > > return err;
> > > > > > > > }
> > > > > > > >
> > > > > > > > +static bool is_valid_state_change(int oldstate, int newstate)
> > > > > > > > +{
> > > > > > > > + switch (oldstate) {
> > > > > > > > + case MLX5_VIRTIO_NET_Q_OBJECT_STATE_INIT:
> > > > > > > > + return newstate == MLX5_VIRTIO_NET_Q_OBJECT_STATE_RDY;
> > > > > > > > + case MLX5_VIRTIO_NET_Q_OBJECT_STATE_RDY:
> > > > > > > > + return newstate == MLX5_VIRTIO_NET_Q_OBJECT_STATE_SUSPEND;
> > > > > > > > + case MLX5_VIRTIO_NET_Q_OBJECT_STATE_SUSPEND:
> > > > > > > > + case MLX5_VIRTIO_NET_Q_OBJECT_STATE_ERR:
> > > > > > > > + default:
> > > > > > > > + return false;
> > > > > > > > + }
> > > > > > > > +}
> > > > > > > > +
> > > > > > > > static int modify_virtqueue(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_virtqueue *mvq, int state)
> > > > > > > > {
> > > > > > > > int inlen = MLX5_ST_SZ_BYTES(modify_virtio_net_q_in);
> > > > > > > > @@ -1130,6 +1146,12 @@ static int modify_virtqueue(struct mlx5_vdpa_net *ndev, struct mlx5_vdpa_virtque
> > > > > > > > void *in;
> > > > > > > > int err;
> > > > > > > >
> > > > > > > > + if (mvq->fw_state == MLX5_VIRTIO_NET_Q_OBJECT_NONE)
> > > > > > > > + return 0;
> > > > > > > > +
> > > > > > > > + if (!is_valid_state_change(mvq->fw_state, state))
> > > > > > > > + return -EINVAL;
> > > > > > > > +
> > > > > > > > in = kzalloc(inlen, GFP_KERNEL);
> > > > > > > > if (!in)
> > > > > > > > return -ENOMEM;
> > > > > > > > @@ -1991,6 +2013,7 @@ static void mlx5_vdpa_set_vq_ready(struct vdpa_device *vdev, u16 idx, bool ready
> > > > > > > > struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
> > > > > > > > struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
> > > > > > > > struct mlx5_vdpa_virtqueue *mvq;
> > > > > > > > + int err;
> > > > > > > >
> > > > > > > > if (!mvdev->actual_features)
> > > > > > > > return;
> > > > > > > > @@ -2004,8 +2027,16 @@ static void mlx5_vdpa_set_vq_ready(struct vdpa_device *vdev, u16 idx, bool ready
> > > > > > > > }
> > > > > > > >
> > > > > > > > mvq = &ndev->vqs[idx];
> > > > > > > > - if (!ready)
> > > > > > > > + if (!ready) {
> > > > > > > > suspend_vq(ndev, mvq);
> > > > > > > > + } else {
> > > > > > > > + err = modify_virtqueue(ndev, mvq, MLX5_VIRTIO_NET_Q_OBJECT_STATE_RDY);
> > > > > > > > + if (err) {
> > > > > > > > + mlx5_vdpa_warn(mvdev, "modify VQ %d to ready failed (%d)\n", idx, err);
> > > > > > > > + ready = false;
> > > > > > > > + }
> > > > > > > > + }
> > > > > > > > +
> > > > > > > >
> > > > > > > > mvq->ready = ready;
> > > > > > > > }
> > > > > > > > @@ -2732,6 +2763,39 @@ static int mlx5_vdpa_get_vendor_vq_stats(struct vdpa_device *vdev, u16 idx,
> > > > > > > > return err;
> > > > > > > > }
> > > > > > > >
> > > > > > > > +static void mlx5_vdpa_cvq_suspend(struct mlx5_vdpa_dev *mvdev, bool suspend)
> > > > > > > > +{
> > > > > > > > + struct mlx5_control_vq *cvq;
> > > > > > > > +
> > > > > > > > + if (!(mvdev->actual_features & BIT_ULL(VIRTIO_NET_F_CTRL_VQ)))
> > > > > > > > + return;
> > > > > > > > +
> > > > > > > > + cvq = &mvdev->cvq;
> > > > > > > > + cvq->ready = !suspend;
> > > > > > > > +}
> > > > > > >
> > > > > > > It looks to me we need to synchronize this with reslock. And this
> > > > > > > probably deserve a dedicated fix.
> > > > > > >
> > > > > > > > +
> > > > > > > > +static int mlx5_vdpa_suspend(struct vdpa_device *vdev, bool suspend)
> > > > > > > > +{
> > > > > > > > + struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
> > > > > > > > + struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
> > > > > > > > + struct mlx5_vdpa_virtqueue *mvq;
> > > > > > > > + int i;
> > > > > > > > +
> > > > > > > > + if (!suspend) {
> > > > > > > > + mlx5_vdpa_warn(mvdev, "Resume of virtqueues is not supported\n");
> > > > > > > > + return -EOPNOTSUPP;
> > > > > > > > + }
> > > > > > > > +
> > > > > > > > + down_write(&ndev->reslock);
> > > > > > > > + for (i = 0; i < ndev->cur_num_vqs; i++) {
> > > > > > > > + mvq = &ndev->vqs[i];
> > > > > > > > + suspend_vq(ndev, mvq);
> > > > > > > > + }
> > > > > > > > + mlx5_vdpa_cvq_suspend(mvdev, suspend);
> > > > > > >
> > > > > > > Do we need to synchronize with the carrier work here? Otherwise we may
> > > > > > > get config notification after suspending.
> > > > > > >
> > > > > > > > + up_write(&ndev->reslock);
> > > > > > > > + return 0;
> > > > > > > > +}
> > > > > > > > +
> > > > > > > > static const struct vdpa_config_ops mlx5_vdpa_ops = {
> > > > > > > > .set_vq_address = mlx5_vdpa_set_vq_address,
> > > > > > > > .set_vq_num = mlx5_vdpa_set_vq_num,
> > > > > > > > @@ -2762,6 +2826,7 @@ static const struct vdpa_config_ops mlx5_vdpa_ops = {
> > > > > > > > .get_generation = mlx5_vdpa_get_generation,
> > > > > > > > .set_map = mlx5_vdpa_set_map,
> > > > > > > > .free = mlx5_vdpa_free,
> > > > > > > > + .suspend = mlx5_vdpa_suspend,
> > > > > > >
> > > > > > > I don't see the vDPA bus patch to enable this method. Or anything I missed here?
> > > > > > >
> > > > > >
> > > > > > Should we add
> > > > > > Based-on: <20220526124338.36247-1-eperezma@...hat.com>
> > > > > >
> > > > > > To this series?
> > > > >
> > > > > Probably, but that series seems to support resume while this series doesn't.
> > > > >
> > > > > Any reason for this?
> > > >
> > > > I think Eugenio agreed that resume is not really required since we're going stop using this
> > > > instance and migrate. In any case, we don't support resume for the hardware object
> > > > though it could be simulated should it be absolutely necessary.
> > >
> > > This is fine if everything is fine during the live migration. But when
> > > migration fails due to some reason, management (libvirt) may choose to
> > > restart the device in the source.
> > >
> > > This means we should either
> > >
> > > 1) support resume in the parent
> > > 2) emulate it in the qemu (with a lot of restoring of the states)
> > >
> >
> > I think it should be handled in qemu (at least the POC reset the
> > device), but I didn't exercise a lot of the failure paths there
> > because, well, it was a POC :).
>
> It looks like a must in the production environment. The failure is not
> necessarily related to shadow virtqueue itself.
>
I don't see a specific interface to resume the device after suspend.
Reset however could do the job and we already have it.
> Thanks
>
> >
> > > And it is not only used for live migration, it could be used for vmstop/start.
> > >
> >
> > I think it would be easier if we dedicate a feature flag for resuming
> > the device in the future. Qemu could take advantage of it at some
> > error paths of live migration, but less than it seems because it
> > overrides things like ring addresses. And, obviously, in the
> > vmstop/vmstart.
> >
> > Actually, net devices should be ok to restore with a full reset. The
> > problem should be filesystems etc that are not part of vdpa at the
> > moment.
> >
> > Thanks!
> >
Powered by blists - more mailing lists