Re: [PATCH] raw-format: drop WRITE and RESIZE child perms when possible

[Stefan Hajnoczi]

On Mon, Jul 26, 2021 at 05:42:47PM +0200, Kevin Wolf wrote:
> Am 26.07.2021 um 16:41 hat Vladimir Sementsov-Ogievskiy geschrieben:
> > 26.07.2021 15:28, Stefan Hajnoczi wrote:
> > > The following command-line fails due to a permissions conflict:
> > > 
> > >    $ qemu-storage-daemon \
> > >        --blockdev 
> > > driver=nvme,node-name=nvme0,device=0000:08:00.0,namespace=1 \
> > >        --blockdev 
> > > driver=raw,node-name=l1-1,file=nvme0,offset=0,size=1073741824 \
> > >        --blockdev 
> > > driver=raw,node-name=l1-2,file=nvme0,offset=1073741824,size=1073741824 \
> > >        --nbd-server 
> > > addr.type=unix,addr.path=/tmp/nbd.sock,max-connections=2 \
> > >        --export type=nbd,id=nbd-l1-1,node-name=l1-1,name=l1-1,writable=on 
> > > \
> > >        --export type=nbd,id=nbd-l1-2,node-name=l1-2,name=l1-2,writable=on
> > > 
> > >    qemu-storage-daemon: --export 
> > > type=nbd,id=nbd-l1-1,node-name=l1-1,name=l1-1,writable=on: Permission 
> > > conflict on node 'nvme0': permissions 'resize' are both required by node 
> > > 'l1-1' (uses node 'nvme0' as 'file' child) and unshared by node 'l1-2' 
> > > (uses node 'nvme0' as 'file' child).
> > > 
> > > The problem is that block/raw-format.c relies on bdrv_default_perms() to
> > > set permissions on the nvme node. The default permissions add RESIZE in
> > > anticipation of a format driver like qcow2 that needs to grow the image
> > > file. This fails because RESIZE is unshared, so we cannot get the RESIZE
> > > permission.
> > > 
> > > Max Reitz pointed out that block/crypto.c already handles this case by
> > > implementing a custom ->bdrv_child_perm() function that adjusts the
> > > result of bdrv_default_perms().
> > > 
> > > This patch takes the same approach in block/raw-format.c so that RESIZE
> > > is only required if it's actually necessary (e.g. the parent is qcow2).
> > > 
> > > Cc: Max Reitz <mreitz@redhat.com>
> > > Cc: Kevin Wolf <kwolf@redhat.com>
> > > Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
> > > ---
> > > This is not a bug fix, so I didn't mark it for QEMU 6.1. It's new
> > > behavior that hasn't been supported before. I want to split an NVMe
> > > drive using the raw format's offset=/size= feature.
> > > ---
> > >   block/raw-format.c | 21 ++++++++++++++++++++-
> > >   1 file changed, 20 insertions(+), 1 deletion(-)
> > > 
> > > diff --git a/block/raw-format.c b/block/raw-format.c
> > > index 7717578ed6..c26f493688 100644
> > > --- a/block/raw-format.c
> > > +++ b/block/raw-format.c
> > > @@ -580,6 +580,25 @@ static void raw_cancel_in_flight(BlockDriverState 
> > > *bs)
> > >       bdrv_cancel_in_flight(bs->file->bs);
> > >   }
> > > +static void raw_child_perm(BlockDriverState *bs, BdrvChild *c,
> > > +                           BdrvChildRole role,
> > > +                           BlockReopenQueue *reopen_queue,
> > > +                           uint64_t parent_perm, uint64_t parent_shared,
> > > +                           uint64_t *nperm, uint64_t *nshared)
> > > +{
> > > +    bdrv_default_perms(bs, c, role, reopen_queue, parent_perm,
> > > +                       parent_shared, nperm, nshared);
> > > +
> > > +    /*
> > > +     * bdrv_default_perms() may add WRITE and/or RESIZE (see comment in
> > > +     * bdrv_default_perms_for_storage() for an explanation) but we only 
> > > need
> > > +     * them if they are in parent_perm. Drop WRITE and RESIZE whenever 
> > > possible
> > > +     * to avoid permission conflicts.
> > > +     */
> > > +    *nperm &= ~(BLK_PERM_WRITE | BLK_PERM_RESIZE);
> > > +    *nperm |= parent_perm & (BLK_PERM_WRITE | BLK_PERM_RESIZE);
> > > +}
> > > +
> > >   BlockDriver bdrv_raw = {
> > >       .format_name          = "raw",
> > >       .instance_size        = sizeof(BDRVRawState),
> > > @@ -588,7 +607,7 @@ BlockDriver bdrv_raw = {
> > >       .bdrv_reopen_commit   = &raw_reopen_commit,
> > >       .bdrv_reopen_abort    = &raw_reopen_abort,
> > >       .bdrv_open            = &raw_open,
> > > -    .bdrv_child_perm      = bdrv_default_perms,
> > > +    .bdrv_child_perm      = raw_child_perm,
> > >       .bdrv_co_create_opts  = &raw_co_create_opts,
> > >       .bdrv_co_preadv       = &raw_co_preadv,
> > >       .bdrv_co_pwritev      = &raw_co_pwritev,
> > > 
> > 
> > I think it's OK:
> > 
> > Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
> > 
> > 
> > Still, did you consider an alternative of making
> > bdrv_filter_default_perm() function public and just do
> > ".bdrv_child_perm = bdrv_filter_default_perm," here?
> > 
> > raw_format is not considered to be filter, but for it's permissions I
> > think it works exactly like filter.
> 
> I had the same thought, but then commit 69dca43d6b6 explicitly made the
> opposite change. I seem to remember that Max never liked raw being
> treated like a filter much.

Additionally:

  static int raw_open(BlockDriverState *bs, QDict *options, int flags,
                      Error **errp)
  {
  ...
      /*
       * Without offset and a size limit, this driver behaves very much
       * like a filter.  With any such limit, it does not.
       */
      if (offset || has_size) {
          file_role = BDRV_CHILD_DATA | BDRV_CHILD_PRIMARY;
      } else {
          file_role = BDRV_CHILD_FILTERED | BDRV_CHILD_PRIMARY;
      }

Whether the raw child node acts as FILTERED or DATA depends on whether
offset=/size= were given.

Stefan

signature.asc
Description: PGP signature