Re: [PATCH v2] virtio-blk: handle block_device_operations callbacks after hot unplug

From: Michael S. Tsirkin
Date: Tue Apr 28 2020 - 12:03:25 EST


On Tue, Apr 28, 2020 at 04:57:15PM +0100, Stefan Hajnoczi wrote:
> On Tue, Apr 28, 2020 at 11:25:07AM -0400, Michael S. Tsirkin wrote:
> > On Tue, Apr 28, 2020 at 03:30:09PM +0100, Stefan Hajnoczi wrote:
> > > A userspace process holding a file descriptor to a virtio_blk device can
> > > still invoke block_device_operations after hot unplug. For example, a
> > > program that has /dev/vdb open can call ioctl(HDIO_GETGEO) after hot
> > > unplug to invoke virtblk_getgeo().
> >
> >
> > which causes what? a use after free?
>
> Yes, use after free. I will include the kernel panic in the next
> revision.
>
> virtio_check_driver_offered_feature() accesses vdev->dev.driver, which
> doesn't work after virtblk_remove() has freed vdev :).
>
> > >
> > > Introduce a reference count in struct virtio_blk so that its lifetime
> > > covers both virtio_driver probe/remove and block_device_operations
> > > open/release users. This ensures that block_device_operations functions
> > > like virtblk_getgeo() can safely access struct virtio_blk.
> > >
> > > Add remove_mutex to prevent block_device_operations functions from
> > > accessing vblk->vdev during virtblk_remove() and let the safely check
> >
> > let the -> let them?
>
> Thanks, will fix.
>
> >
> > > for !vblk->vdev after virtblk_remove() returns.
> > >
> > > Switching to a reference count also solves the vd_index_ida leak where
> > > vda, vdb, vdc, etc indices were lost when the device was hot unplugged
> > > while the block device was still open.
> >
> > Can you move this statement up so we list both issues (use after free
> > and leak) upfront, then discuss the fix?
>
> Thanks, will fix.

Maybe add Fixes: if there's a specific commit that added geo support.

> >
> > >
> > > Reported-by: Lance Digby <ldigby@xxxxxxxxxx>
> > > Signed-off-by: Stefan Hajnoczi <stefanha@xxxxxxxxxx>
> > > ---
> > > If someone has a simpler solution please let me know. I looked at
> > > various approaches including reusing device_lock(&vblk->vdev.dev) but
> > > they were more complex and extending the lifetime of virtio_device after
> > > remove() has been called seems questionable.
> > > ---
> > > drivers/block/virtio_blk.c | 85 ++++++++++++++++++++++++++++++++++----
> > > 1 file changed, 77 insertions(+), 8 deletions(-)
> > >
> > > diff --git a/drivers/block/virtio_blk.c b/drivers/block/virtio_blk.c
> > > index 93468b7c6701..3dd53b445cc1 100644
> > > --- a/drivers/block/virtio_blk.c
> > > +++ b/drivers/block/virtio_blk.c
> > > @@ -44,6 +44,13 @@ struct virtio_blk {
> > > /* Process context for config space updates */
> > > struct work_struct config_work;
> > >
> > > + /*
> > > + * Tracks references from block_device_operations open/release and
> > > + * virtio_driver probe/remove so this object can be freed once no
> > > + * longer in use.
> > > + */
> > > + refcount_t refs;
> > > +
> > > /* What host tells us, plus 2 for header & tailer. */
> > > unsigned int sg_elems;
> > >
> > > @@ -53,6 +60,9 @@ struct virtio_blk {
> > > /* num of vqs */
> > > int num_vqs;
> > > struct virtio_blk_vq *vqs;
> > > +
> > > + /* Provides mutual exclusion with virtblk_remove(). */
> >
> > This is not the best way to document access rules.
> > Which fields does this protect, exactly?
> > I think it's just vdev. Right?
> > Pls add to the comment.
>
> Yes, vblk->vdev cannot be dereferenced after virtblk_remove() is
> entered.
>
> I'll document this mutex as protecting vdev.
>
> >
> > > + struct mutex remove_mutex;
> > > };
> > >
> > > struct virtblk_req {
> > > @@ -295,10 +305,54 @@ static int virtblk_get_id(struct gendisk *disk, char *id_str)
> > > return err;
> > > }
> > >
> > > +static void virtblk_get(struct virtio_blk *vblk)
> > > +{
> > > + refcount_inc(&vblk->refs);
> > > +}
> > > +
> > > +static void virtblk_put(struct virtio_blk *vblk)
> > > +{
> > > + if (refcount_dec_and_test(&vblk->refs)) {
> > > + ida_simple_remove(&vd_index_ida, vblk->index);
> > > + mutex_destroy(&vblk->remove_mutex);
> > > + kfree(vblk);
> > > + }
> > > +}
> > > +
> > > +static int virtblk_open(struct block_device *bd, fmode_t mode)
> > > +{
> > > + struct virtio_blk *vblk = bd->bd_disk->private_data;
> > > + int ret = -ENXIO;
> >
> >
> > It's more common to do
> >
> > int ret = 0;
> >
> > and on error:
> > ret = -ENXIO;
> >
> >
> > let's do this.
>
> Will fix.
>
> > > +
> > > + mutex_lock(&vblk->remove_mutex);
> > > +
> > > + if (vblk->vdev) {
> > > + virtblk_get(vblk);
> > > + ret = 0;
> > > + }
> >
> > I prefer
> > else
> > ret = -ENXIO
> >
> > here.
>
> Got it.
>
> > > +
> > > + mutex_unlock(&vblk->remove_mutex);
> > > + return ret;
> > > +}
> > > +
> > > +static void virtblk_release(struct gendisk *disk, fmode_t mode)
> > > +{
> > > + struct virtio_blk *vblk = disk->private_data;
> > > +
> > > + virtblk_put(vblk);
> > > +}
> > > +
> > > /* We provide getgeo only to please some old bootloader/partitioning tools */
> > > static int virtblk_getgeo(struct block_device *bd, struct hd_geometry *geo)
> > > {
> > > struct virtio_blk *vblk = bd->bd_disk->private_data;
> > > + int ret = -ENXIO;
> >
> > It's more common to do
> >
> > int ret = 0;
> >
> > and on error:
> > ret = -ENXIO;
> >
> >
> > let's do this.
>
> Will fix.
>
> >
> > > +
> > > + mutex_lock(&vblk->remove_mutex);
> > > +
> > > + if (!vblk->vdev) {
> > > + goto out;
> > > + }
> >
> >
> > single lines are not supposed to use {}.
> > if you add ret = -ENXIO here then it won't be a single line anymore
> > though.
>
> Oops, a QEMU coding style habit :).
>
> > >
> > > /* see if the host passed in geometry config */
> > > if (virtio_has_feature(vblk->vdev, VIRTIO_BLK_F_GEOMETRY)) {
> > > @@ -314,11 +368,17 @@ static int virtblk_getgeo(struct block_device *bd, struct hd_geometry *geo)
> > > geo->sectors = 1 << 5;
> > > geo->cylinders = get_capacity(bd->bd_disk) >> 11;
> > > }
> > > - return 0;
> > > +
> > > + ret = 0;
> > > +out:
> > > + mutex_unlock(&vblk->remove_mutex);
> > > + return ret;
> > > }
> > >
> > > static const struct block_device_operations virtblk_fops = {
> > > .owner = THIS_MODULE,
> > > + .open = virtblk_open,
> > > + .release = virtblk_release,
> > > .getgeo = virtblk_getgeo,
> > > };
> > >
> > > @@ -655,6 +715,10 @@ static int virtblk_probe(struct virtio_device *vdev)
> > > goto out_free_index;
> > > }
> > >
> > > + /* This reference is dropped in virtblk_remove(). */
> > > + refcount_set(&vblk->refs, 1);
> > > + mutex_init(&vblk->remove_mutex);
> > > +
> > > vblk->vdev = vdev;
> > > vblk->sg_elems = sg_elems;
> > >
> > > @@ -820,8 +884,12 @@ static int virtblk_probe(struct virtio_device *vdev)
> > > static void virtblk_remove(struct virtio_device *vdev)
> > > {
> > > struct virtio_blk *vblk = vdev->priv;
> > > - int index = vblk->index;
> > > - int refc;
> > > +
> > > + /*
> > > + * Virtqueue processing is stopped safely here but mutual exclusion is
> > > + * needed for block_device_operations.
> > > + */
> > > + mutex_lock(&vblk->remove_mutex);
> > >
> > > /* Make sure no work handler is accessing the device. */
> > > flush_work(&vblk->config_work);
> > > @@ -834,15 +902,16 @@ static void virtblk_remove(struct virtio_device *vdev)
> > > /* Stop all the virtqueues. */
> > > vdev->config->reset(vdev);
> > >
> > > - refc = kref_read(&disk_to_dev(vblk->disk)->kobj.kref);
> > > + /* Virtqueue are stopped, nothing can use vblk->vdev anymore. */
> >
> > Virtqueues?
>
> Thanks, will fix.