Re: [Qemu-devel] [PATCH] block: Produce zeros when protocols reading beyond end of file

[Asias He <asias@redhat.com>]

On Fri, Aug 16, 2013 at 10:41:36AM +0200, Stefan Hajnoczi wrote:
> On Mon, Aug 5, 2013 at 10:11 AM, Asias He <asias@redhat.com> wrote:
> > From: MORITA Kazutaka <morita.kazutaka@lab.ntt.co.jp>
> >
> > While Asias is debugging an issue creating qcow2 images on top of
> > non-file protocols.  It boils down to this example using NBD:
> >
> > $ qemu-io -c 'open -g nbd+unix:///?socket=/tmp/nbd.sock' -c 'read -v 0 512'
> >
> > Notice the open -g option to set bs->growable.  This means you can
> > read/write beyond end of file.  Reading beyond end of file is supposed
> > to produce zeroes.
> >
> > We rely on this behavior in qcow2_create2() during qcow2 image
> > creation.  We create a new file and then write the qcow2 header
> > structure using bdrv_pwrite().  Since QCowHeader is not a multiple of
> > sector size, block.c first uses bdrv_read() on the empty file to fetch
> > the first sector (should be all zeroes).
> >
> > Here is the output from the qemu-io NBD example above:
> >
> > $ qemu-io -c 'open -g nbd+unix:///?socket=/tmp/nbd.sock' -c 'read -v 0 512'
> > 00000000:  ab ab ab ab ab ab ab ab ab ab ab ab ab ab ab ab  ................
> > 00000010:  ab ab ab ab ab ab ab ab ab ab ab ab ab ab ab ab  ................
> > 00000020:  ab ab ab ab ab ab ab ab ab ab ab ab ab ab ab ab  ................
> > ...
> >
> > We are not zeroing the buffer!  As a result qcow2 image creation on top
> > of protocols is not guaranteed to work even when file creation is
> > supported by the protocol.
> >
> > Signed-off-by: MORITA Kazutaka <morita.kazutaka@lab.ntt.co.jp>
> > Signed-off-by: Asias He <asias@redhat.com>
> > ---
> >  block.c | 30 +++++++++++++++++++++++++++++-
> >  1 file changed, 29 insertions(+), 1 deletion(-)
> >
> > diff --git a/block.c b/block.c
> > index 01b66d8..deaf0a0 100644
> > --- a/block.c
> > +++ b/block.c
> > @@ -2544,7 +2544,35 @@ static int coroutine_fn bdrv_co_do_readv(BlockDriverState *bs,
> >          }
> >      }
> >
> > -    ret = drv->bdrv_co_readv(bs, sector_num, nb_sectors, qiov);
> > +    if (!bs->drv->protocol_name) {
> > +        ret = drv->bdrv_co_readv(bs, sector_num, nb_sectors, qiov);
> > +    } else {
> > +        /* NBD doesn't support reading beyond end of file. */
> > +        int64_t len, total_sectors, max_nb_sectors;
> > +
> > +        len = bdrv_getlength(bs);
> > +        if (len < 0) {
> > +            ret = len;
> > +            goto out;
> > +        }
> > +
> > +        total_sectors = len >> BDRV_SECTOR_BITS;
> > +        max_nb_sectors = MAX(0, total_sectors - sector_num);
> > +        if (max_nb_sectors > 0) {
> > +            ret = drv->bdrv_co_readv(bs, sector_num,
> > +                                     MIN(nb_sectors, max_nb_sectors), qiov);
> > +        } else {
> > +            ret = 0;
> > +        }
> > +
> > +        /* Reading beyond end of file is supposed to produce zeroes */
> > +        if (ret == 0 && total_sectors < sector_num + nb_sectors) {
> > +            size_t offset = MAX(0, total_sectors - sector_num);
> > +            size_t bytes = (sector_num + nb_sectors - offset) *
> > +                            BDRV_SECTOR_SIZE;
> > +            qemu_iovec_memset(qiov, offset * BDRV_SECTOR_SIZE, 0, bytes);
> > +        }
> > +    }
> 
> This patch breaks qemu-iotests ./check -qcow2 022.  This happens
> because qcow2 temporarily sets ->growable = 1 for vmstate accesses
> (which are stored beyond the end of regular image data).

I am a bit confused. This is from the other mail:

"""
> > I think it would break qcow2_load_vmstate(), which is basically a                                                                   
> > bdrv_pread() after the end of the image.                                                                                            
>                                                                                                                                       
> I see, then only protocols have to zeroing the buffer?  In case of                                                                    
> protocols, I think bdrv_getlength() returns the underlying file                                                                       
> length, so qcow2_load_vmstate() would be a bdrv_pread() within the                                                                    
> result of bdrv_getlength().                                                                                                           

Limiting it to protocols solves the problem, I think.
"""

And in v1 of this patch, Kevin wanted bs->growable check instad of the
protocol_name one.

"""
> -    ret = drv->bdrv_co_readv(bs, sector_num, nb_sectors, qiov);                                                                      
> +    if (!bs->drv->protocol_name) {                                                                                                   

I think !bs->growable is the right check.

Checking for the protocol name is always a hack and most times wrong.
"""

Switching back to the protocol_name check, ./check -qcow2 022 test passes.

> static int qcow2_load_vmstate(BlockDriverState *bs, uint8_t *buf,
>                               int64_t pos, int size)
> {
>     BDRVQcowState *s = bs->opaque;
>     int growable = bs->growable;
>     int ret;
> 
>     BLKDBG_EVENT(bs->file, BLKDBG_VMSTATE_LOAD);
>     bs->growable = 1;
>     ret = bdrv_pread(bs, qcow2_vm_state_offset(s) + pos, buf, size);
>     bs->growable = growable;
> 
>     return ret;
> }
> 
> Please *always* run qemu-iotests before submitting block patches:
> http://qemu-project.org/Documentation/QemuIoTests

OK.

> A simple but ugly way to fix this is for block.c to also have a
> ->zero_beyond_eof flag which enables the behavior you are adding.
> qcow2_load_vmstate() would disable ->zero_beyond_eof temporarily in
> addition to enabling ->growable.

I am wondering why the ->growable logic is introduced in the first
place.  Adding yet another this kind of flag looks realy ugly ;(

> It's not easy to call the internal qcow2_co_readv() from
> qcow2_load_vmstate() because the vmstate functions are byte
> granularity (like bdrv_pread()) while .bdrv_co_readv() is
> sector-granularity.
> 
> Stefan

-- 
Asias