Re: Null dereference in bdrv_unregister_buf() probably memory-backend-file related?

[Stefan Hajnoczi <stefanha@gmail.com>]

On Mon, 21 Nov 2022 at 12:29, Jonathan Cameron
<Jonathan.Cameron@huawei.com> wrote:
>
> On Mon, 21 Nov 2022 11:47:48 -0500
> Stefan Hajnoczi <stefanha@gmail.com> wrote:
>
> > On Mon, 21 Nov 2022 at 11:22, Jonathan Cameron via
> > <qemu-devel@nongnu.org> wrote:
> > >
> > > First CC list is a guess as I haven't managed to root cause where things are
> > > going wrong yet.
> > >
> > > Originally hit this whilst rebasing some CXL patches on v7.2.0-rc1.
> > > CXL makes extensive use of memory-backends and most my tests happen
> > > to use memory-backend-file
> > >
> > > Issue seen on arm64 and x86 though helpfully on x86 the crash appears in an entirely
> > > unrelated location (though the 'fix' works).
> > >
> > > Fairly minimal test command line.
> > >
> > > qemu-system-aarch64 \
> > >     -M virt  \
> > >     -drive if=none,file=full.qcow2,format=qcow2,id=hd \
> > >     -device virtio-blk,drive=hd \
> > >     -object memory-backend-file,id=cxl-mem1,mem-path=/tmp/cxltest.raw,size=256M,align=256M \
> > >
> > > Powerdown the machine or ctrl-c during boot gives a segfault.
> > > On arm64 it was in a stable location that made at least some sense in that
> > > bs in the below snippet is NULL.
> > >
> > > I added the follow work around and the segfault goes away...
> > >
> > >  [PATCH] temp
> > >
> > > ---
> > >  block/io.c | 3 +++
> > >  1 file changed, 3 insertions(+)
> > >
> > > diff --git a/block/io.c b/block/io.c
> > > index b9424024f9..750e1366aa 100644
> > > --- a/block/io.c
> > > +++ b/block/io.c
> > > @@ -3324,6 +3324,9 @@ void bdrv_unregister_buf(BlockDriverState *bs, void *host, size_t size)
> > >  {
> > >      BdrvChild *child;
> > >
> > > +    if (!bs) {
> > > +        return;
> > > +    }
> > >      GLOBAL_STATE_CODE();
> > >      if (bs->drv && bs->drv->bdrv_unregister_buf) {
> > >          bs->drv->bdrv_unregister_buf(bs, host, size);
> >
> > bdrv_*() APIs generally don't accept NULL bs arguments.
> >
> > I think blk_unregister_buf() needs to handle the blk_bs() NULL return
> > value. Can you confirm that the parent function is
> > blk_unregister_buf()?
> >
> > This bug may have been introduced by commit baf422684d73 ("virtio-blk:
> > use BDRV_REQ_REGISTERED_BUF optimization hint").
> Got it in one. I just bisected to exactly that patch
>
> + using the below change indeed works just as well as the above.
> Now I'd send this as a patch, but I don't yet sufficiently understand what that change you
> referenced did to break things Seems it registered a notifier that is getting
> called for all ram blocks, not just the one virtio-blk ones?
>
> Perhaps better if you send a fix with an explanation :)

Sure, no problem. I have reproduced the bug and will send a patch.

Thanks,
Stefan