[Top][All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Qemu-devel] [PATCH 09/13] qed: Convert to bdrv_co_pwrite_zeroes()

From: Eric Blake
Subject: Re: [Qemu-devel] [PATCH 09/13] qed: Convert to bdrv_co_pwrite_zeroes()
Date: Wed, 25 May 2016 08:28:41 -0600
User-agent: Mozilla/5.0 (X11; Linux x86_64; rv:38.0) Gecko/20100101 Thunderbird/38.3.0

On 05/25/2016 08:07 AM, Kevin Wolf wrote:
> Am 25.05.2016 um 00:25 hat Eric Blake geschrieben:
>> Another step on our continuing quest to switch to byte-based
>> interfaces.
>> Kill an abuse of the comma operator while at it (fortunately,
>> the semantics were still right).
>> Signed-off-by: Eric Blake <address@hidden>
>> ---
>>  block/qed.c | 25 +++++++++++++------------
>>  1 file changed, 13 insertions(+), 12 deletions(-)

>> @@ -1443,10 +1443,10 @@ static int coroutine_fn 
>> bdrv_qed_co_write_zeroes(BlockDriverState *bs,
>>      /* Refuse if there are untouched backing file sectors */

The comment wasn't very helpful, so I may rewort it, too
(s/untouched/unaligned/, or something like that)

>>      if (bs->backing) {
>> -        if (qed_offset_into_cluster(s, sector_num * BDRV_SECTOR_SIZE) != 0) 
>> {
>> +        if (qed_offset_into_cluster(s, offset) != 0) {
>>              return -ENOTSUP;
>>          }
>> -        if (qed_offset_into_cluster(s, nb_sectors * BDRV_SECTOR_SIZE) != 0) 
>> {
>> +        if (qed_offset_into_cluster(s, count) != 0) {
>>              return -ENOTSUP;
>>          }
>>      }
> Unaligned requests are only emulated if there is no backing file...
>> @@ -1454,12 +1454,13 @@ static int coroutine_fn 
>> bdrv_qed_co_write_zeroes(BlockDriverState *bs,
>>      /* Zero writes start without an I/O buffer.  If a buffer becomes 
>> necessary
>>       * then it will be allocated during request processing.
>>       */
>> -    iov.iov_base = NULL,
>> -    iov.iov_len  = nb_sectors * BDRV_SECTOR_SIZE,
>> +    iov.iov_base = NULL;
>> +    iov.iov_len = count;
>>      qemu_iovec_init_external(&qiov, &iov, 1);
>> -    blockacb = qed_aio_setup(bs, sector_num, &qiov, nb_sectors,
>> -                             qed_co_write_zeroes_cb, &cb,
>> +    blockacb = qed_aio_setup(bs, offset >> BDRV_SECTOR_BITS, &qiov,
>> +                             count >> BDRV_SECTOR_BITS,
> ...so offset and count can still be unaligned here and we end up zeroing
> out the wrong part of the sector. I guess we need to return -ENOTSUP for
> all sub-sector requests, even without a backing file.

Hmm. Wouldn't it be nicer if we could guarantee that blk_pwrite_zeroes()
will never call bdrv_co_pwrite_zeroes() with less than
request_alignment?  That is, if the block layer takes care of
read-modify-write for any unaligned byte offset less than
request_alignment, then the driver only has to worry about sector
alignment.  Except qed.c doesn't seem to set request_alignment, but is
just relying on io.c currently setting it to MAX(BDRV_SECTOR_SIZE,
bs->bl.request_alignment) everywhere. (And the fact that
request_alignment is a sibling rather than a member to BlockLimits bl is

Maybe we want three limits in BlockLimits, rather than two: the current
max_pwrite_zeroes does a good job at saying how small blk_pwrite_zeroes
must fragment large requests, and pwrite_zeroes_alignment does a good
job at saying how large a request must be to potentially punch a hole,
but at least in the case of qcow2, where we want to optimize a partial
write to potentially zeroing an entire cluster, we still want to limit
things to sector boundaries when checking for whether the rest of the
cluster already reads as zeroes, whether or not we also want to support
request_alignment of 1 instead of 512.

There are other drivers that I touched in this series that were relying
on the fact that the block layer currently guarantees sector alignment,
and maybe they should be setting request_alignment, or maybe we want to
add yet another BlockLimit member.  So even if we want normal read/write
to allow request_alignment of 1 in the case where we don't need the
block layer to do a read-modify-write, I'm still wondering whether we
want the write_zeroes engine to have a different minimum alignment and
ALWAYS hand off to normal read-modify-write for anything smaller.

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

Attachment: signature.asc
Description: OpenPGP digital signature

reply via email to

[Prev in Thread] Current Thread [Next in Thread]