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

Re: [Nbd] [Qemu-devel] [PATCH v5 13/14] nbd: Implement NBD_CMD_WRITE_ZEROES on server

On 07/19/2016 09:34 PM, Fam Zheng wrote:
> On Tue, 07/19 17:45, Paolo Bonzini wrote:
>> On 19/07/2016 17:28, Eric Blake wrote:
>>>> If I'm reading the NBD proto.md correctly, this is not enough if
>>>> NBD_CMD_FLAG_NO_HOLE is specified. We probably need to use a zeroed buffer with
>>>> blk_pwrite, or pass a new flag (BDRV_RED_NO_HOLE) to blk_pwrite_zeroes to
>>>> enforce the bdrv_driver_pwritev() branch in bdrv_co_do_pwrite_zeroes().
>> I agree with Eric's interpretation.  It's a bit weird to have the
>> direction inverted, but I'm not sure I see the ambiguity.  Can you explain?
> Write zeroes _means_ "punch hole" on a raw file.
> In block/raw-posix.c:handle_aiocb_write_zeroes():
>>     if (s->has_discard && s->has_fallocate) {
>>         int ret = do_fallocate(s->fd,
>>                                FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
>>                                aiocb->aio_offset, aiocb->aio_nbytes);
>>         if (ret == 0) {
>>             ret = do_fallocate(s->fd, 0, aiocb->aio_offset, aiocb->aio_nbytes);

That is just implementation: punch a hole, BUT THEN reallocate it back,
so that in the end, the file is still not sparse in that region.  Or am
I reading it wrong?

But the implementation under the hood is not visible to the guest - as
long as the end result is that a guest requesting NO_HOLE ends up with a
non-sparse file, and the data reads back as all 0, the client doesn't
care whether the zeros were written byte-by-byte or sped up by punching
a hole then reallocating.

> And unmap is translated to "punch hole", too.
> In block/raw-posix.c:handle_aiocb_discard():
>>         ret = do_fallocate(s->fd, FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
>>                            aiocb->aio_offset, aiocb->aio_nbytes);
>> #endif

No, this call is different - it punches a hole, then stops.  There is no
followup do_fallocate(,0,,) to reallocate, so the file remains sparse.

> So I agree that NBD_CMD_FLAG_NO_HOLE is a poorly named flag, because there is
> always going to be a hole event if it's set.

If we are punching holes even when BDRV_REQ_MAY_UNMAP is not set, that
seems like we have a bug in qemu (unless we are immediately then
reallocating so that there is no resulting hole).

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

Attachment: signature.asc
Description: OpenPGP digital signature

Reply to: