Re: [Qemu-devel] [PATCH 2/5] qcow2: Implement .bdrv_co_preadv()

[Eric Blake]

On 06/03/2016 11:21 AM, Kevin Wolf wrote:
> Reading from qcow2 images is now byte granularity.
> 
> Most of the affected code in qcow2 actually gets simpler with this
> change. The only exception is encryption, which is fixed on 512 bytes
> blocks; in order to keep this working, bs->request_alignment is set for
> encrypted images.
> 
> Signed-off-by: Kevin Wolf <address@hidden>
> ---
>  block/qcow2-cluster.c |  18 ++++-----
>  block/qcow2.c         | 108 
> +++++++++++++++++++++++++++-----------------------
>  block/qcow2.h         |   2 +-
>  3 files changed, 67 insertions(+), 61 deletions(-)
> 

> @@ -467,16 +468,16 @@ out:
>   * For a given offset of the disk image, find the cluster offset in
>   * qcow2 file. The offset is stored in *cluster_offset.
>   *
> - * on entry, *num is the number of contiguous sectors we'd like to
> + * on entry, *bytes is the number of contiguous bytes we'd like to

maybe s/number/maximum number/

>   * access following offset.
>   *
> - * on exit, *num is the number of contiguous sectors we can read.
> + * on exit, *bytes is the number of contiguous bytes we can read.

maybe s/we can read/with the same cluster type/

>   *
>   * Returns the cluster type (QCOW2_CLUSTER_*) on success, -errno in error
>   * cases.
>   */
>  int qcow2_get_cluster_offset(BlockDriverState *bs, uint64_t offset,
> -    int *num, uint64_t *cluster_offset)
> +                             unsigned int *bytes, uint64_t *cluster_offset)
>  {
>      BDRVQcow2State *s = bs->opaque;
>      unsigned int l2_index;
> @@ -485,12 +486,9 @@ int qcow2_get_cluster_offset(BlockDriverState *bs, 
> uint64_t offset,
>      unsigned int offset_in_cluster, nb_clusters;
>      uint64_t bytes_available, bytes_needed;
>      int ret;
> -    unsigned int bytes;
> -
> -    bytes = *num * BDRV_SECTOR_SIZE;

One potential overflow gone...

>  
>      offset_in_cluster = offset_into_cluster(s, offset);
> -    bytes_needed = bytes + offset_in_cluster;
> +    bytes_needed = *bytes + offset_in_cluster;

...but not the other.  Looks like your callers limit their input 'bytes'
to at most INT_MAX, and therefore it happens to not overflow unsigned
int in practice, but you may want an assertion?

> +++ b/block/qcow2.c
> @@ -975,6 +975,9 @@ static int qcow2_open(BlockDriverState *bs, QDict 
> *options, int flags,
>          }
>  
>          bs->encrypted = 1;
> +
> +        /* Encryption works on a sector granularity */
> +        bs->request_alignment = BDRV_SECTOR_SIZE;

Trivial conflict with my patch 5/5 that moves request_alignment into
BlockLimits (if we even want that, since I still have to find why my
patch makes qemu-iotests 77 hang)

>  
> -static coroutine_fn int qcow2_co_readv(BlockDriverState *bs, int64_t 
> sector_num,
> -                          int remaining_sectors, QEMUIOVector *qiov)
> +static coroutine_fn int qcow2_co_preadv(BlockDriverState *bs, uint64_t 
> offset,
> +                                        uint64_t bytes, QEMUIOVector *qiov,
> +                                        int flags)

Wait a minute.  .bdrv_co_preadv() takes uint64_t bytes, while
bdrv_co_preadv() takes only unsigned int bytes?  Eww.  We've got some
more scrubbing work to do.  At least it is going to get easier to
universally turn on full 64-bit byte interfaces everywhere, especially
once my patches for auto-fragmenting at max_transfer_length land (which
in turn won't be posted before your conversion of bdrv_aligned_preadv()
to a byte interface).  So no impact to this patch.

>  {
>      BDRVQcow2State *s = bs->opaque;
> -    int index_in_cluster, n1;
> +    int offset_in_cluster, n1;
>      int ret;
> -    int cur_nr_sectors; /* number of sectors in current iteration */
> +    unsigned int cur_bytes; /* number of sectors in current iteration */

comment is stale now

>      uint64_t cluster_offset = 0;
>      uint64_t bytes_done = 0;
>      QEMUIOVector hd_qiov;
> @@ -1389,26 +1402,24 @@ static coroutine_fn int 
> qcow2_co_readv(BlockDriverState *bs, int64_t sector_num,
>  
>      qemu_co_mutex_lock(&s->lock);
>  
> -    while (remaining_sectors != 0) {
> +    while (bytes != 0) {
>  
>          /* prepare next request */
> -        cur_nr_sectors = remaining_sectors;
> +        cur_bytes = MIN(bytes, INT_MAX);
>          if (s->cipher) {
> -            cur_nr_sectors = MIN(cur_nr_sectors,
> -                QCOW_MAX_CRYPT_CLUSTERS * s->cluster_sectors);
> +            cur_bytes = MIN(cur_bytes,
> +                            QCOW_MAX_CRYPT_CLUSTERS * s->cluster_size);

Again, my work on auto-fragmenting at the block layer should make it so
that we can eventually further simplify this part to just assert that
bytes doesn't exceed max_transfer_length, rather than having to fragment
it at INT_MAX ourselves.

Couple of tweaks to fix as pointed out above, but mostly looks sane.

-- 
Eric Blake   eblake redhat com    +1-919-301-3266
Libvirt virtualization library http://libvirt.org

signature.asc
Description: OpenPGP digital signature