On 11/18/2016 04:24 AM, Olaf Hering wrote:
> The guest sends discard requests as u64 sector/count pairs, but the
> block layer operates internally with s64/s32 pairs. The conversion
> leads to IO errors in the guest, the discard request is not processed.
> 
>   domU.cfg:
>   'vdev=xvda, format=qcow2, backendtype=qdisk, target=/x.qcow2'
>   domU:
>   mkfs.ext4 -F /dev/xvda
>   Discarding device blocks: failed - Input/output error
> 
> Fix this by splitting the request into chunks of BDRV_REQUEST_MAX_SECTORS.
> Add input range checking to avoid overflow.
> 
> Signed-off-by: Olaf Hering <olaf@aepfle.de>
> ---
>  hw/block/xen_disk.c | 45 +++++++++++++++++++++++++++++++++++++++------
>  1 file changed, 39 insertions(+), 6 deletions(-)
> 
> diff --git a/hw/block/xen_disk.c b/hw/block/xen_disk.c
> index 3a7dc19..c3f572f 100644
> --- a/hw/block/xen_disk.c
> +++ b/hw/block/xen_disk.c
> @@ -660,6 +660,41 @@ static void qemu_aio_complete(void *opaque, int ret)
>      qemu_bh_schedule(ioreq->blkdev->bh);
>  }
>  
> +static bool blk_split_discard(struct ioreq *ioreq, blkif_sector_t sector_number,
> +                              uint64_t nr_sectors)
> +{
> +    struct XenBlkDev *blkdev = ioreq->blkdev;
> +    int64_t byte_offset;
> +    int byte_chunk;
> +    uint64_t sec_start = sector_number;
> +    uint64_t sec_count = nr_sectors;
> +    uint64_t byte_remaining;
> +    uint64_t limit = BDRV_REQUEST_MAX_SECTORS << BDRV_SECTOR_BITS;

[For reference, this limit is the same as rounding INT32_MAX down to the
nearest 512-byte limit, or 0x7ffffe00]

> +
> +    /* Wrap around? */
> +    if ((sec_start + sec_count) < sec_count) {
> +        return false;
> +    }
> +    /* Overflowing byte limit? */
> +    if ((sec_start + sec_count) > ((INT64_MAX + INT_MAX) >> BDRV_SECTOR_BITS)) {

This is undefined.  INT64_MAX + anything non-negative overflows int64,
and even if you treat overflow as defined by twos-complement
representation (which creates a negative number), shifting a negative
number is also undefined.

If you are trying to detect guests that make a request that would cover
more than INT64_MAX bytes, you can simplify.  Besides, for as much
storage as there is out there, I seriously doubt ANYONE will ever have
2^63 bytes addressable through a single device.  Why not just write it as:

if ((INT64_MAX >> BDRV_SECTOR_BITS) - sec_count < sec_start) {

> +        return false;
> +    }
> +
> +    byte_offset = sec_start << BDRV_SECTOR_BITS;
> +    byte_remaining = sec_count << BDRV_SECTOR_BITS;
> +
> +    do {
> +        byte_chunk = byte_remaining > limit ? limit : byte_remaining;
> +        ioreq->aio_inflight++;
> +        blk_aio_pdiscard(blkdev->blk, byte_offset, byte_chunk,
> +                         qemu_aio_complete, ioreq);
> +        byte_remaining -= byte_chunk;
> +        byte_offset += byte_chunk;
> +    } while (byte_remaining > 0);

This part looks reasonable.

> +
> +    return true;
> +}
> +
>  static int ioreq_runio_qemu_aio(struct ioreq *ioreq)
>  {
>      struct XenBlkDev *blkdev = ioreq->blkdev;
> @@ -708,12 +743,10 @@ static int ioreq_runio_qemu_aio(struct ioreq *ioreq)
>          break;
>      case BLKIF_OP_DISCARD:
>      {
> -        struct blkif_request_discard *discard_req = (void *)&ioreq->req;

The old code had it...

> -        ioreq->aio_inflight++;
> -        blk_aio_pdiscard(blkdev->blk,
> -                         discard_req->sector_number << BDRV_SECTOR_BITS,
> -                         discard_req->nr_sectors << BDRV_SECTOR_BITS,
> -                         qemu_aio_complete, ioreq);
> +        struct blkif_request_discard *req = (void *)&ioreq->req;

...but C doesn't require a cast to void*. As long as you are touching
this, you could remove the cast (unless I'm missing something, and the
cast is also there to cast away const).

> +        if (!blk_split_discard(ioreq, req->sector_number, req->nr_sectors)) {
> +            goto err;
> +        }
>          break;
>      }
>      default:
> 
> 

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