[v4,7/8] block: core copy-on-read logic

Signed-off-by: Stefan Hajnoczi <stefanha@linux.vnet.ibm.com>
---
 block.c      |   72 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
 trace-events |    1 +
 2 files changed, 73 insertions(+), 0 deletions(-)

On Wed, Nov 23, 2011 at 11:47:57AM +0000, Stefan Hajnoczi wrote:
> Signed-off-by: Stefan Hajnoczi <stefanha@linux.vnet.ibm.com>
> ---
>  block.c      |   72 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
>  trace-events |    1 +
>  2 files changed, 73 insertions(+), 0 deletions(-)
> 
> diff --git a/block.c b/block.c
> index c30c8f2..a598a19 100644
> --- a/block.c
> +++ b/block.c
> @@ -1469,6 +1469,61 @@ int bdrv_pwrite_sync(BlockDriverState *bs, int64_t offset,
>      return 0;
>  }
>  
> +static int coroutine_fn bdrv_co_copy_on_readv(BlockDriverState *bs,
> +        int64_t sector_num, int nb_sectors, QEMUIOVector *qiov)
> +{
> +    /* Perform I/O through a temporary buffer so that users who scribble over
> +     * their read buffer while the operation is in progress do not end up
> +     * modifying the image file.  This is critical for zero-copy guest I/O
> +     * where anything might happen inside guest memory.
> +     */
> +    void *bounce_buffer;
> +
> +    struct iovec iov;
> +    QEMUIOVector bounce_qiov;
> +    int64_t cluster_sector_num;
> +    int cluster_nb_sectors;
> +    size_t skip_bytes;
> +    int ret;
> +
> +    /* Cover entire cluster so no additional backing file I/O is required when
> +     * allocating cluster in the image file.
> +     */
> +    round_to_clusters(bs, sector_num, nb_sectors,
> +                      &cluster_sector_num, &cluster_nb_sectors);
> +
> +    trace_bdrv_co_copy_on_readv(bs, sector_num, nb_sectors,
> +                                cluster_sector_num, cluster_nb_sectors);
> +
> +    iov.iov_len = cluster_nb_sectors * BDRV_SECTOR_SIZE;
> +    iov.iov_base = bounce_buffer = qemu_blockalign(bs, iov.iov_len);
> +    qemu_iovec_init_external(&bounce_qiov, &iov, 1);
> +
> +    ret = bs->drv->bdrv_co_readv(bs, cluster_sector_num, cluster_nb_sectors,
> +                                 &bounce_qiov);
> +    if (ret < 0) {
> +        goto err;
> +    }
> +
> +    ret = bs->drv->bdrv_co_writev(bs, cluster_sector_num, cluster_nb_sectors,
> +                                  &bounce_qiov);
> +    if (ret < 0) {
> +        /* It might be okay to ignore write errors for guest requests.  If this
> +         * is a deliberate copy-on-read then we don't want to ignore the error.
> +         * Simply report it in all cases.
> +         */
> +        goto err;
> +    }
> +
> +    skip_bytes = (sector_num - cluster_sector_num) * BDRV_SECTOR_SIZE;
> +    qemu_iovec_from_buffer(qiov, bounce_buffer + skip_bytes,
> +                           nb_sectors * BDRV_SECTOR_SIZE);
> +
> +err:
> +    qemu_vfree(bounce_buffer);
> +    return ret;
> +}
> +
>  /*
>   * Handle a read request in coroutine context
>   */
> @@ -1496,7 +1551,24 @@ static int coroutine_fn bdrv_co_do_readv(BlockDriverState *bs,
>      }
>  
>      tracked_request_begin(&req, bs, sector_num, nb_sectors, false);
> +
> +    if (bs->copy_on_read) {
> +        int pnum;
> +
> +        ret = bdrv_co_is_allocated(bs, sector_num, nb_sectors, &pnum);
> +        if (ret < 0) {
> +            goto out;
> +        }

Stefan,

It is not clear where support for shared backing files would fit.
Let us consider the following block copy example:

1) Original chain:
[ BASE ] -> [ IMAGE-1 ] -> [ IMAGE-2 ] -> [ IMAGE-3 ]

2) Final chain:
[ BASE ] -> [ IMAGE-3 ]

I was talking to Kevin and we don't have code/monitor command to the
switch from 1) to 2). But that is a separate issue.

Question is, how do you plan to stream the contents of IMAGE-1 and
IMAGE-2 (but not BASE), into IMAGE-3 ? That is an important use case.

Also, do you have status on the external COW file work, for raw images?

Thanks

On Fri, Dec 2, 2011 at 5:13 PM, Marcelo Tosatti <mtosatti@redhat.com> wrote:
> On Wed, Nov 23, 2011 at 11:47:57AM +0000, Stefan Hajnoczi wrote:
>> Signed-off-by: Stefan Hajnoczi <stefanha@linux.vnet.ibm.com>
>> ---
>>  block.c      |   72 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
>>  trace-events |    1 +
>>  2 files changed, 73 insertions(+), 0 deletions(-)
>>
>> diff --git a/block.c b/block.c
>> index c30c8f2..a598a19 100644
>> --- a/block.c
>> +++ b/block.c
>> @@ -1469,6 +1469,61 @@ int bdrv_pwrite_sync(BlockDriverState *bs, int64_t offset,
>>      return 0;
>>  }
>>
>> +static int coroutine_fn bdrv_co_copy_on_readv(BlockDriverState *bs,
>> +        int64_t sector_num, int nb_sectors, QEMUIOVector *qiov)
>> +{
>> +    /* Perform I/O through a temporary buffer so that users who scribble over
>> +     * their read buffer while the operation is in progress do not end up
>> +     * modifying the image file.  This is critical for zero-copy guest I/O
>> +     * where anything might happen inside guest memory.
>> +     */
>> +    void *bounce_buffer;
>> +
>> +    struct iovec iov;
>> +    QEMUIOVector bounce_qiov;
>> +    int64_t cluster_sector_num;
>> +    int cluster_nb_sectors;
>> +    size_t skip_bytes;
>> +    int ret;
>> +
>> +    /* Cover entire cluster so no additional backing file I/O is required when
>> +     * allocating cluster in the image file.
>> +     */
>> +    round_to_clusters(bs, sector_num, nb_sectors,
>> +                      &cluster_sector_num, &cluster_nb_sectors);
>> +
>> +    trace_bdrv_co_copy_on_readv(bs, sector_num, nb_sectors,
>> +                                cluster_sector_num, cluster_nb_sectors);
>> +
>> +    iov.iov_len = cluster_nb_sectors * BDRV_SECTOR_SIZE;
>> +    iov.iov_base = bounce_buffer = qemu_blockalign(bs, iov.iov_len);
>> +    qemu_iovec_init_external(&bounce_qiov, &iov, 1);
>> +
>> +    ret = bs->drv->bdrv_co_readv(bs, cluster_sector_num, cluster_nb_sectors,
>> +                                 &bounce_qiov);
>> +    if (ret < 0) {
>> +        goto err;
>> +    }
>> +
>> +    ret = bs->drv->bdrv_co_writev(bs, cluster_sector_num, cluster_nb_sectors,
>> +                                  &bounce_qiov);
>> +    if (ret < 0) {
>> +        /* It might be okay to ignore write errors for guest requests.  If this
>> +         * is a deliberate copy-on-read then we don't want to ignore the error.
>> +         * Simply report it in all cases.
>> +         */
>> +        goto err;
>> +    }
>> +
>> +    skip_bytes = (sector_num - cluster_sector_num) * BDRV_SECTOR_SIZE;
>> +    qemu_iovec_from_buffer(qiov, bounce_buffer + skip_bytes,
>> +                           nb_sectors * BDRV_SECTOR_SIZE);
>> +
>> +err:
>> +    qemu_vfree(bounce_buffer);
>> +    return ret;
>> +}
>> +
>>  /*
>>   * Handle a read request in coroutine context
>>   */
>> @@ -1496,7 +1551,24 @@ static int coroutine_fn bdrv_co_do_readv(BlockDriverState *bs,
>>      }
>>
>>      tracked_request_begin(&req, bs, sector_num, nb_sectors, false);
>> +
>> +    if (bs->copy_on_read) {
>> +        int pnum;
>> +
>> +        ret = bdrv_co_is_allocated(bs, sector_num, nb_sectors, &pnum);
>> +        if (ret < 0) {
>> +            goto out;
>> +        }
>
> Stefan,
>
> It is not clear where support for shared backing files would fit.
> Let us consider the following block copy example:
>
> 1) Original chain:
> [ BASE ] -> [ IMAGE-1 ] -> [ IMAGE-2 ] -> [ IMAGE-3 ]
>
> 2) Final chain:
> [ BASE ] -> [ IMAGE-3 ]
>
> I was talking to Kevin and we don't have code/monitor command to the
> switch from 1) to 2). But that is a separate issue.
>
> Question is, how do you plan to stream the contents of IMAGE-1 and
> IMAGE-2 (but not BASE), into IMAGE-3 ? That is an important use case.

We need to introduce a "deep" bdrv_is_allocated() which takes a "base"
argument at which to terminate the search.  Here's what it will do:

[ BASE ] -> [ IMAGE-1 ] -> [ IMAGE-2 ] -> [ IMAGE-3 ]

def bdrv_co_is_allocated(IMAGE-3, sector_num=0, base=BASE):
    if IMAGE-3 sector_num=0 is allocated:
        return True  # already allocated
    if IMAGE-2 sector_num=0 is allocated:
        return False # copy it!
    if IMAGE-1 sector_num=0 is allocated:
        return False # copy it!
    return True      # reached base image so we don't care

This is the hardcoded version of the function, in reality it will be a
loop but it shows the idea.

> Also, do you have status on the external COW file work, for raw images?

Sure, please help review Dong Xu Wang's add-cow patches on the list,
he is looking for feedback.

Stefan