LKML Archive on lore.kernel.org
help / color / mirror / Atom feed
From: Yongji Xie <xieyongji@bytedance.com>
To: "Michael S. Tsirkin" <mst@redhat.com>
Cc: "Jason Wang" <jasowang@redhat.com>,
	"Stefan Hajnoczi" <stefanha@redhat.com>,
	"Stefano Garzarella" <sgarzare@redhat.com>,
	"Parav Pandit" <parav@nvidia.com>,
	"Christoph Hellwig" <hch@infradead.org>,
	"Christian Brauner" <christian.brauner@canonical.com>,
	"Randy Dunlap" <rdunlap@infradead.org>,
	"Matthew Wilcox" <willy@infradead.org>,
	"Al Viro" <viro@zeniv.linux.org.uk>,
	"Jens Axboe" <axboe@kernel.dk>,
	bcrl@kvack.org, "Jonathan Corbet" <corbet@lwn.net>,
	"Mika Penttilä" <mika.penttila@nextfour.com>,
	"Dan Carpenter" <dan.carpenter@oracle.com>,
	joro@8bytes.org, "Greg KH" <gregkh@linuxfoundation.org>,
	"He Zhe" <zhe.he@windriver.com>,
	"Liu Xiaodong" <xiaodong.liu@intel.com>,
	"Joe Perches" <joe@perches.com>,
	"Robin Murphy" <robin.murphy@arm.com>,
	"Will Deacon" <will@kernel.org>,
	"John Garry" <john.garry@huawei.com>,
	songmuchun@bytedance.com,
	virtualization <virtualization@lists.linux-foundation.org>,
	netdev@vger.kernel.org, kvm <kvm@vger.kernel.org>,
	linux-fsdevel@vger.kernel.org, iommu@lists.linux-foundation.org,
	linux-kernel <linux-kernel@vger.kernel.org>
Subject: Re: [PATCH v13 05/13] vdpa: Add reset callback in vdpa_config_ops
Date: Mon, 6 Sep 2021 20:13:04 +0800	[thread overview]
Message-ID: <CACycT3uChWmr50FGutFz0hK+2jSjx0KJsJ5w-NVBq-E_nmJm9w@mail.gmail.com> (raw)
In-Reply-To: <20210906053210-mutt-send-email-mst@kernel.org>

On Mon, Sep 6, 2021 at 6:43 PM Michael S. Tsirkin <mst@redhat.com> wrote:
>
> On Mon, Sep 06, 2021 at 04:45:55PM +0800, Yongji Xie wrote:
> > On Mon, Sep 6, 2021 at 4:01 PM Michael S. Tsirkin <mst@redhat.com> wrote:
> > >
> > > On Mon, Sep 06, 2021 at 03:06:44PM +0800, Yongji Xie wrote:
> > > > On Mon, Sep 6, 2021 at 2:37 PM Michael S. Tsirkin <mst@redhat.com> wrote:
> > > > >
> > > > > On Mon, Sep 06, 2021 at 02:09:25PM +0800, Yongji Xie wrote:
> > > > > > On Mon, Sep 6, 2021 at 1:56 PM Michael S. Tsirkin <mst@redhat.com> wrote:
> > > > > > >
> > > > > > > On Tue, Aug 31, 2021 at 06:36:26PM +0800, Xie Yongji wrote:
> > > > > > > > This adds a new callback to support device specific reset
> > > > > > > > behavior. The vdpa bus driver will call the reset function
> > > > > > > > instead of setting status to zero during resetting.
> > > > > > > >
> > > > > > > > Signed-off-by: Xie Yongji <xieyongji@bytedance.com>
> > > > > > >
> > > > > > >
> > > > > > > This does gloss over a significant change though:
> > > > > > >
> > > > > > >
> > > > > > > > ---
> > > > > > > > @@ -348,12 +352,12 @@ static inline struct device *vdpa_get_dma_dev(struct vdpa_device *vdev)
> > > > > > > >       return vdev->dma_dev;
> > > > > > > >  }
> > > > > > > >
> > > > > > > > -static inline void vdpa_reset(struct vdpa_device *vdev)
> > > > > > > > +static inline int vdpa_reset(struct vdpa_device *vdev)
> > > > > > > >  {
> > > > > > > >       const struct vdpa_config_ops *ops = vdev->config;
> > > > > > > >
> > > > > > > >       vdev->features_valid = false;
> > > > > > > > -     ops->set_status(vdev, 0);
> > > > > > > > +     return ops->reset(vdev);
> > > > > > > >  }
> > > > > > > >
> > > > > > > >  static inline int vdpa_set_features(struct vdpa_device *vdev, u64 features)
> > > > > > >
> > > > > > >
> > > > > > > Unfortunately this breaks virtio_vdpa:
> > > > > > >
> > > > > > >
> > > > > > > static void virtio_vdpa_reset(struct virtio_device *vdev)
> > > > > > > {
> > > > > > >         struct vdpa_device *vdpa = vd_get_vdpa(vdev);
> > > > > > >
> > > > > > >         vdpa_reset(vdpa);
> > > > > > > }
> > > > > > >
> > > > > > >
> > > > > > > and there's no easy way to fix this, kernel can't recover
> > > > > > > from a reset failure e.g. during driver unbind.
> > > > > > >
> > > > > >
> > > > > > Yes, but it should be safe with the protection of software IOTLB even
> > > > > > if the reset() fails during driver unbind.
> > > > > >
> > > > > > Thanks,
> > > > > > Yongji
> > > > >
> > > > > Hmm. I don't see it.
> > > > > What exactly will happen? What prevents device from poking at
> > > > > memory after reset? Note that dma unmap in e.g. del_vqs happens
> > > > > too late.
> > > >
> > > > But I didn't see any problems with touching the memory for virtqueues.
> > >
> > > Drivers make the assumption that after reset returns no new
> > > buffers will be consumed. For example a bunch of drivers
> > > call virtqueue_detach_unused_buf.
> >
> > I'm not sure if I get your point. But it looks like
> > virtqueue_detach_unused_buf() will check the driver's metadata first
> > rather than read the memory from virtqueue.
> >
> > > I can't say whether block makes this assumption anywhere.
> > > Needs careful auditing.
> > >
> > > > The memory should not be freed after dma unmap?
> > >
> > > But unmap does not happen until after the reset.
> > >
> >
> > I mean the memory is totally allocated and controlled by the VDUSE
> > driver. The VDUSE driver will not return them to the buddy system
> > unless userspace unmap it.
>
> Right. But what stops VDUSE from poking at memory after
> reset failed?
>

Only itself. But in normal cases, userspace would not poke at the
memory since there is no available data after reset.

And it makes me think whether it's better to disallow returning errors
from userspace for the reset message. Then the only case that leads to
reset failure is the request timeout, which will mark the device
broken now.

>
> > >
> > > > And the memory for the bounce buffer should also be safe to be
> > > > accessed by userspace in this case.
> > > >
> > > > > And what about e.g. interrupts?
> > > > > E.g. we have this:
> > > > >
> > > > >         /* Virtqueues are stopped, nothing can use vblk->vdev anymore. */
> > > > >         vblk->vdev = NULL;
> > > > >
> > > > > and this is no longer true at this point.
> > > > >
> > > >
> > > > You're right. But I didn't see where the interrupt handler will use
> > > > the vblk->vdev.
> > >
> > > static void virtblk_done(struct virtqueue *vq)
> > > {
> > >         struct virtio_blk *vblk = vq->vdev->priv;
> > >
> > > vq->vdev is the same as vblk->vdev.
> > >
> >
> > We will test the vq->ready (will be set to false in del_vqs()) before
> > injecting an interrupt in the VDUSE driver. So it should be OK?
>
> Maybe not ...  It's not designed for such asynchronous access, so e.g.
> there's no locking or memory ordering around accesses.
>

Yes, so we still need the below fix.

>
> > >
> > > > So it seems to be not too late to fix it:
> > > >
> > > > diff --git a/drivers/vdpa/vdpa_user/vduse_dev.c
> > > > b/drivers/vdpa/vdpa_user/vduse_dev.c
> > > > index 5c25ff6483ad..ea41a7389a26 100644
> > > > --- a/drivers/vdpa/vdpa_user/vduse_dev.c
> > > > +++ b/drivers/vdpa/vdpa_user/vduse_dev.c
> > > > @@ -665,13 +665,13 @@ static void vduse_vdpa_set_config(struct
> > > > vdpa_device *vdpa, unsigned int offset,
> > > >  static int vduse_vdpa_reset(struct vdpa_device *vdpa)
> > > >  {
> > > >         struct vduse_dev *dev = vdpa_to_vduse(vdpa);
> > > > +       int ret;
> > > >
> > > > -       if (vduse_dev_set_status(dev, 0))
> > > > -               return -EIO;
> > > > +       ret = vduse_dev_set_status(dev, 0);
> > > >
> > > >         vduse_dev_reset(dev);
> > > >
> > > > -       return 0;
> > > > +       return ret;
> > > >  }
> > > >
> > > >  static u32 vduse_vdpa_get_generation(struct vdpa_device *vdpa)
> > > >
> > > > Thanks,
> > > > Yongji
> > >
> > > Needs some comments to explain why it's done like this.
> > >
> >
> > This is used to make sure the userspace can't not inject the interrupt
> > any more after reset. The vduse_dev_reset() will clear the interrupt
> > callback and flush the irq kworker.
> >
> > > BTW device is generally wedged at this point right?
> > > E.g. if reset during initialization fails, userspace
> > > will still get the reset at some later point and be
> > > confused ...
> > >
> >
> > Sorry, I don't get why userspace will get the reset at some later point?
> >
> > Thanks,
> > Yongji
>
> I am generally a bit confused about how does reset work with vduse.
> We clearly want device to get back to its original state.
> How is that supposed to be achieved?
>

I think both the userspace daemon and kernel should reset their
internal states. Besides, userspace daemon needs to stop processing
virtqueue and clear the memory mapping.

Thanks,
Yongji

  reply	other threads:[~2021-09-06 12:13 UTC|newest]

Thread overview: 28+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2021-08-31 10:36 [PATCH v13 00/13] Introduce VDUSE - vDPA Device in Userspace Xie Yongji
2021-08-31 10:36 ` [PATCH v13 01/13] iova: Export alloc_iova_fast() and free_iova_fast() Xie Yongji
2021-08-31 10:36 ` [PATCH v13 02/13] eventfd: Export eventfd_wake_count to modules Xie Yongji
2021-09-01  2:50   ` Jason Wang
2021-08-31 10:36 ` [PATCH v13 03/13] file: Export receive_fd() " Xie Yongji
2021-09-05 15:57   ` Michael S. Tsirkin
2021-09-05 16:44     ` Al Viro
2021-08-31 10:36 ` [PATCH v13 04/13] vdpa: Fix some coding style issues Xie Yongji
2021-08-31 10:36 ` [PATCH v13 05/13] vdpa: Add reset callback in vdpa_config_ops Xie Yongji
2021-09-06  5:55   ` Michael S. Tsirkin
2021-09-06  6:09     ` Yongji Xie
2021-09-06  6:37       ` Michael S. Tsirkin
2021-09-06  7:06         ` Yongji Xie
2021-09-06  8:00           ` Michael S. Tsirkin
2021-09-06  8:45             ` Yongji Xie
2021-09-06 10:43               ` Michael S. Tsirkin
2021-09-06 12:13                 ` Yongji Xie [this message]
2021-08-31 10:36 ` [PATCH v13 06/13] vhost-vdpa: Handle the failure of vdpa_reset() Xie Yongji
2021-08-31 10:36 ` [PATCH v13 07/13] vhost-iotlb: Add an opaque pointer for vhost IOTLB Xie Yongji
2021-08-31 10:36 ` [PATCH v13 08/13] vdpa: Add an opaque pointer for vdpa_config_ops.dma_map() Xie Yongji
2021-08-31 10:36 ` [PATCH v13 09/13] vdpa: factor out vhost_vdpa_pa_map() and vhost_vdpa_pa_unmap() Xie Yongji
2021-08-31 10:36 ` [PATCH v13 10/13] vdpa: Support transferring virtual addressing during DMA mapping Xie Yongji
2021-08-31 10:36 ` [PATCH v13 11/13] vduse: Implement an MMU-based software IOTLB Xie Yongji
2021-08-31 10:36 ` [PATCH v13 12/13] vduse: Introduce VDUSE - vDPA Device in Userspace Xie Yongji
2021-08-31 10:36 ` [PATCH v13 13/13] Documentation: Add documentation for VDUSE Xie Yongji
2021-10-11  7:32 ` [PATCH v13 00/13] Introduce VDUSE - vDPA Device in Userspace Liuxiangdong
2021-10-11  8:31   ` Yongji Xie
2021-10-11  8:32     ` Yongji Xie

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=CACycT3uChWmr50FGutFz0hK+2jSjx0KJsJ5w-NVBq-E_nmJm9w@mail.gmail.com \
    --to=xieyongji@bytedance.com \
    --cc=axboe@kernel.dk \
    --cc=bcrl@kvack.org \
    --cc=christian.brauner@canonical.com \
    --cc=corbet@lwn.net \
    --cc=dan.carpenter@oracle.com \
    --cc=gregkh@linuxfoundation.org \
    --cc=hch@infradead.org \
    --cc=iommu@lists.linux-foundation.org \
    --cc=jasowang@redhat.com \
    --cc=joe@perches.com \
    --cc=john.garry@huawei.com \
    --cc=joro@8bytes.org \
    --cc=kvm@vger.kernel.org \
    --cc=linux-fsdevel@vger.kernel.org \
    --cc=linux-kernel@vger.kernel.org \
    --cc=mika.penttila@nextfour.com \
    --cc=mst@redhat.com \
    --cc=netdev@vger.kernel.org \
    --cc=parav@nvidia.com \
    --cc=rdunlap@infradead.org \
    --cc=robin.murphy@arm.com \
    --cc=sgarzare@redhat.com \
    --cc=songmuchun@bytedance.com \
    --cc=stefanha@redhat.com \
    --cc=viro@zeniv.linux.org.uk \
    --cc=virtualization@lists.linux-foundation.org \
    --cc=will@kernel.org \
    --cc=willy@infradead.org \
    --cc=xiaodong.liu@intel.com \
    --cc=zhe.he@windriver.com \
    --subject='Re: [PATCH v13 05/13] vdpa: Add reset callback in vdpa_config_ops' \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).