Re: [Qemu-devel] [PATCH 7/7] block: add write support for VHDX images

[Jeff Cody <jcody@redhat.com>]

On Thu, Mar 07, 2013 at 04:59:05PM +0100, Stefan Hajnoczi wrote:
> On Wed, Mar 06, 2013 at 09:48:43AM -0500, Jeff Cody wrote:
> > @@ -958,12 +960,150 @@ exit:
> >      return ret;
> >  }
> >  
> > +/*
> > + * Allocate a new payload block at the end of the file.
> > + *
> > + * Allocation will happen at 1MB alignment inside the file
> > + *
> > + * Returns the file offset start of the new payload block
> > + */
> > +static int vhdx_allocate_block(BlockDriverState *bs, BDRVVHDXState *s,
> > +                                    uint64_t *new_offset)
> > +{
> > +    *new_offset = bdrv_getlength(bs->file);
> > +
> > +    /* per the spec, the address for a block is in units of 1MB */
> > +    if (*new_offset % (1024*1024)) {
> > +        *new_offset = ((*new_offset >> 20) + 1) << 20;  /* round up to 1MB */
> > +    }
> > +    return bdrv_truncate(bs->file, *new_offset + s->block_size);
> > +}
> 
> Is it necessary to resize the file?  Why not just write at EOF to grow
> it?
> 

After the recent bdrv_truncate() discussion, I think that may be best.

> > +/*
> > + * Update the BAT tablet entry with the new file offset, and the new entry
> > + * state */
> > +static int vhdx_update_bat_table_entry(BlockDriverState *bs, BDRVVHDXState *s,
> > +                                       vhdx_sector_info *sinfo, int state)
> > +{
> > +    uint64_t bat_tmp;
> > +    uint64_t bat_entry_offset;
> > +
> > +    /* The BAT entry is a uint64, with 44 bits for the file offset in units of
> > +     * 1MB, and 3 bits for the block state. */
> > +    s->bat[sinfo->bat_idx]  = ((sinfo->file_offset>>20) <<
> > +                               VHDX_BAT_FILE_OFF_BITS);
> > +
> > +    s->bat[sinfo->bat_idx] |= state & VHDX_BAT_STATE_BIT_MASK;
> >  
> > +    bat_tmp = cpu_to_le64(s->bat[sinfo->bat_idx]);
> > +    bat_entry_offset = s->bat_offset + sinfo->bat_idx * sizeof(vhdx_bat_entry);
> > +
> > +    return bdrv_pwrite_sync(bs->file, bat_entry_offset, &bat_tmp,
> > +                            sizeof(vhdx_bat_entry));
> > +}
> >  
> >  static coroutine_fn int vhdx_co_writev(BlockDriverState *bs, int64_t sector_num,
> >                                        int nb_sectors, QEMUIOVector *qiov)
> >  {
> > -    return -ENOTSUP;
> > +    int ret = -ENOTSUP;
> > +    BDRVVHDXState *s = bs->opaque;
> > +    vhdx_sector_info sinfo;
> > +    uint64_t bytes_done = 0;
> > +    QEMUIOVector hd_qiov;
> > +
> > +    qemu_iovec_init(&hd_qiov, qiov->niov);
> > +
> > +    qemu_co_mutex_lock(&s->lock);
> > +
> > +    /* Per the spec, on the first write of guest-visible data to the file the
> > +     * data write guid must be updated in the header */
> > +    if (s->first_visible_write) {
> > +        s->first_visible_write = false;
> > +        vhdx_update_headers(bs, s, true);
> > +    }
> > +
> > +    while (nb_sectors > 0) {
> > +        if (s->params.data_bits & VHDX_PARAMS_HAS_PARENT) {
> > +            /* not supported yet */
> > +            ret = -ENOTSUP;
> > +            goto exit;
> > +        } else {
> > +            vhdx_block_translate(s, sector_num, nb_sectors, &sinfo);
> > +
> > +            qemu_iovec_reset(&hd_qiov);
> > +            qemu_iovec_concat(&hd_qiov, qiov,  bytes_done, sinfo.bytes_avail);
> > +            /* check the payload block state */
> > +            switch (s->bat[sinfo.bat_idx] & VHDX_BAT_STATE_BIT_MASK) {
> > +            case PAYLOAD_BLOCK_ZERO:
> > +                /* in this case, we need to preserve zero writes for
> > +                 * data that is not part of this write, so we must pad
> > +                 * the rest of the buffer to zeroes */
> > +
> > +                /* if we are on a posix system with ftruncate() that extends
> > +                 * a file, then it is zero-filled for us.  On Win32, the raw
> > +                 * layer uses SetFilePointer and SetFileEnd, which does not
> > +                 * zero fill AFAIK */
> > +
> > +                /* TODO: queue another write of zero buffers if the host OS does
> > +                 * not zero-fill on file extension */
> > +
> > +                /* fall through */
> > +            case PAYLOAD_BLOCK_NOT_PRESENT: /* fall through */
> > +            case PAYLOAD_BLOCK_UNMAPPED:    /* fall through */
> > +            case PAYLOAD_BLOCK_UNDEFINED:   /* fall through */
> > +                ret = vhdx_allocate_block(bs, s, &sinfo.file_offset);
> > +                if (ret < 0) {
> > +                    goto exit;
> > +                }
> > +                /* once we support differencing files, this may also be
> > +                 * partially present */
> > +                /* update block state to the newly specified state */
> > +                ret = vhdx_update_bat_table_entry(bs, s, &sinfo,
> > +                                                  PAYLOAD_BLOCK_FULL_PRESENT);
> > +                if (ret < 0) {
> > +                    goto exit;
> > +                }
> 
> The BAT table entry must be written after data is already on disk.
> Otherwise a crash partway through would leave the BAT table pointing to
> undefined data.
> 
> Do you need to use the log to ensure this?

Yes, indeed.  The log support is what I am working on now - all
metadata and BAT writes are supposed to go through the log, for this
reason.  The next version will have log support, and so I will use
that for the BAT writes (and any other metadata writes, except the
header).