Message ID | 1433742974-20128-4-git-send-email-famz@redhat.com |
---|---|
State | New |
Headers | show |
Am 08.06.2015 um 07:56 hat Fam Zheng geschrieben: > If guest discards a source cluster, mirroring with bdrv_aio_readv is overkill. > Some protocols do zero upon discard, where it's best to use > bdrv_aio_write_zeroes, otherwise, bdrv_aio_discard will be enough. > > Signed-off-by: Fam Zheng <famz@redhat.com> > --- > block/mirror.c | 20 ++++++++++++++++++-- > 1 file changed, 18 insertions(+), 2 deletions(-) > > diff --git a/block/mirror.c b/block/mirror.c > index d2515c7..3c38695 100644 > --- a/block/mirror.c > +++ b/block/mirror.c > @@ -164,6 +164,8 @@ static uint64_t coroutine_fn mirror_iteration(MirrorBlockJob *s) > int64_t end, sector_num, next_chunk, next_sector, hbitmap_next_sector; > uint64_t delay_ns = 0; > MirrorOp *op; > + int pnum; > + int64_t ret; > > s->sector_num = hbitmap_iter_next(&s->hbi); > if (s->sector_num < 0) { > @@ -290,8 +292,22 @@ static uint64_t coroutine_fn mirror_iteration(MirrorBlockJob *s) > s->in_flight++; > s->sectors_in_flight += nb_sectors; > trace_mirror_one_iteration(s, sector_num, nb_sectors); > - bdrv_aio_readv(source, sector_num, &op->qiov, nb_sectors, > - mirror_read_complete, op); > + > + ret = bdrv_get_block_status_above(source, NULL, sector_num, > + nb_sectors, &pnum); > + if (ret < 0 || pnum < nb_sectors || Earlier today I told Richard Jones that qemu-img commit should really be using zero cluster support in the backing file since 2.4 because I remembered this commit. Turns out it doesn't actually use it but writes explicit zeros instead. The reason is the condition 'pnum < nb_sectors' here, which makes mirror fall back to explicit writes if bdrv_get_block_status_above() doesn't return enough sectors (enough being relatively large here, I think in qemu-img commit it's always the full 10 MB buffer). In other words, we are ignoring any zero areas smaller than 10 MB! (What made this worse is that qcow2 had a bug that reports only a single zero cluster at a time, so it would never report more than 10 MB, even if the image was completely zeroed. I've sent a fix for that one.) In order to fix this, we'll probably need to move the call to bdrv_get_block_status_above() before actually allocating memory and all that for the full nb_chunks. We should detect zeros on the usual block job granularity (64k by default, I think). > + (ret & BDRV_BLOCK_DATA && !(ret & BDRV_BLOCK_ZERO))) { > + bdrv_aio_readv(source, sector_num, &op->qiov, nb_sectors, > + mirror_read_complete, op); > + } else if (ret & BDRV_BLOCK_ZERO) { > + bdrv_aio_write_zeroes(s->target, sector_num, op->nb_sectors, > + s->unmap ? BDRV_REQ_MAY_UNMAP : 0, > + mirror_write_complete, op); > + } else { > + assert(!(ret & BDRV_BLOCK_DATA)); > + bdrv_aio_discard(s->target, sector_num, op->nb_sectors, > + mirror_write_complete, op); > + } > return delay_ns; > } Paolo also noticed that there's no reason at all to allocate buffers and a qiov for the write_zeroes and discard cases. Kevin
On Wed, 11/04 19:35, Kevin Wolf wrote: > Am 08.06.2015 um 07:56 hat Fam Zheng geschrieben: > > If guest discards a source cluster, mirroring with bdrv_aio_readv is overkill. > > Some protocols do zero upon discard, where it's best to use > > bdrv_aio_write_zeroes, otherwise, bdrv_aio_discard will be enough. > > > > Signed-off-by: Fam Zheng <famz@redhat.com> > > --- > > block/mirror.c | 20 ++++++++++++++++++-- > > 1 file changed, 18 insertions(+), 2 deletions(-) > > > > diff --git a/block/mirror.c b/block/mirror.c > > index d2515c7..3c38695 100644 > > --- a/block/mirror.c > > +++ b/block/mirror.c > > @@ -164,6 +164,8 @@ static uint64_t coroutine_fn mirror_iteration(MirrorBlockJob *s) > > int64_t end, sector_num, next_chunk, next_sector, hbitmap_next_sector; > > uint64_t delay_ns = 0; > > MirrorOp *op; > > + int pnum; > > + int64_t ret; > > > > s->sector_num = hbitmap_iter_next(&s->hbi); > > if (s->sector_num < 0) { > > @@ -290,8 +292,22 @@ static uint64_t coroutine_fn mirror_iteration(MirrorBlockJob *s) > > s->in_flight++; > > s->sectors_in_flight += nb_sectors; > > trace_mirror_one_iteration(s, sector_num, nb_sectors); > > - bdrv_aio_readv(source, sector_num, &op->qiov, nb_sectors, > > - mirror_read_complete, op); > > + > > + ret = bdrv_get_block_status_above(source, NULL, sector_num, > > + nb_sectors, &pnum); > > + if (ret < 0 || pnum < nb_sectors || > > Earlier today I told Richard Jones that qemu-img commit should really > be using zero cluster support in the backing file since 2.4 because I > remembered this commit. Turns out it doesn't actually use it but writes > explicit zeros instead. > > The reason is the condition 'pnum < nb_sectors' here, which makes mirror > fall back to explicit writes if bdrv_get_block_status_above() doesn't > return enough sectors (enough being relatively large here, I think in > qemu-img commit it's always the full 10 MB buffer). > > In other words, we are ignoring any zero areas smaller than 10 MB! > > (What made this worse is that qcow2 had a bug that reports only a single > zero cluster at a time, so it would never report more than 10 MB, even > if the image was completely zeroed. I've sent a fix for that one.) > > In order to fix this, we'll probably need to move the call to > bdrv_get_block_status_above() before actually allocating memory and > all that for the full nb_chunks. We should detect zeros on the usual > block job granularity (64k by default, I think). > > > + (ret & BDRV_BLOCK_DATA && !(ret & BDRV_BLOCK_ZERO))) { > > + bdrv_aio_readv(source, sector_num, &op->qiov, nb_sectors, > > + mirror_read_complete, op); > > + } else if (ret & BDRV_BLOCK_ZERO) { > > + bdrv_aio_write_zeroes(s->target, sector_num, op->nb_sectors, > > + s->unmap ? BDRV_REQ_MAY_UNMAP : 0, > > + mirror_write_complete, op); > > + } else { > > + assert(!(ret & BDRV_BLOCK_DATA)); > > + bdrv_aio_discard(s->target, sector_num, op->nb_sectors, > > + mirror_write_complete, op); > > + } > > return delay_ns; > > } > > Paolo also noticed that there's no reason at all to allocate buffers > and a qiov for the write_zeroes and discard cases. I'll write a patch to address these. Thanks! Fam
Am 05.11.2015 um 06:42 hat Fam Zheng geschrieben: > On Wed, 11/04 19:35, Kevin Wolf wrote: > > Am 08.06.2015 um 07:56 hat Fam Zheng geschrieben: > > > If guest discards a source cluster, mirroring with bdrv_aio_readv is overkill. > > > Some protocols do zero upon discard, where it's best to use > > > bdrv_aio_write_zeroes, otherwise, bdrv_aio_discard will be enough. > > > > > > Signed-off-by: Fam Zheng <famz@redhat.com> > > > --- > > > block/mirror.c | 20 ++++++++++++++++++-- > > > 1 file changed, 18 insertions(+), 2 deletions(-) > > > > > > diff --git a/block/mirror.c b/block/mirror.c > > > index d2515c7..3c38695 100644 > > > --- a/block/mirror.c > > > +++ b/block/mirror.c > > > @@ -164,6 +164,8 @@ static uint64_t coroutine_fn mirror_iteration(MirrorBlockJob *s) > > > int64_t end, sector_num, next_chunk, next_sector, hbitmap_next_sector; > > > uint64_t delay_ns = 0; > > > MirrorOp *op; > > > + int pnum; > > > + int64_t ret; > > > > > > s->sector_num = hbitmap_iter_next(&s->hbi); > > > if (s->sector_num < 0) { > > > @@ -290,8 +292,22 @@ static uint64_t coroutine_fn mirror_iteration(MirrorBlockJob *s) > > > s->in_flight++; > > > s->sectors_in_flight += nb_sectors; > > > trace_mirror_one_iteration(s, sector_num, nb_sectors); > > > - bdrv_aio_readv(source, sector_num, &op->qiov, nb_sectors, > > > - mirror_read_complete, op); > > > + > > > + ret = bdrv_get_block_status_above(source, NULL, sector_num, > > > + nb_sectors, &pnum); > > > + if (ret < 0 || pnum < nb_sectors || > > > > Earlier today I told Richard Jones that qemu-img commit should really > > be using zero cluster support in the backing file since 2.4 because I > > remembered this commit. Turns out it doesn't actually use it but writes > > explicit zeros instead. > > > > The reason is the condition 'pnum < nb_sectors' here, which makes mirror > > fall back to explicit writes if bdrv_get_block_status_above() doesn't > > return enough sectors (enough being relatively large here, I think in > > qemu-img commit it's always the full 10 MB buffer). > > > > In other words, we are ignoring any zero areas smaller than 10 MB! > > > > (What made this worse is that qcow2 had a bug that reports only a single > > zero cluster at a time, so it would never report more than 10 MB, even > > if the image was completely zeroed. I've sent a fix for that one.) > > > > In order to fix this, we'll probably need to move the call to > > bdrv_get_block_status_above() before actually allocating memory and > > all that for the full nb_chunks. We should detect zeros on the usual > > block job granularity (64k by default, I think). > > > > > + (ret & BDRV_BLOCK_DATA && !(ret & BDRV_BLOCK_ZERO))) { > > > + bdrv_aio_readv(source, sector_num, &op->qiov, nb_sectors, > > > + mirror_read_complete, op); > > > + } else if (ret & BDRV_BLOCK_ZERO) { > > > + bdrv_aio_write_zeroes(s->target, sector_num, op->nb_sectors, > > > + s->unmap ? BDRV_REQ_MAY_UNMAP : 0, > > > + mirror_write_complete, op); > > > + } else { > > > + assert(!(ret & BDRV_BLOCK_DATA)); > > > + bdrv_aio_discard(s->target, sector_num, op->nb_sectors, > > > + mirror_write_complete, op); > > > + } > > > return delay_ns; > > > } > > > > Paolo also noticed that there's no reason at all to allocate buffers > > and a qiov for the write_zeroes and discard cases. > > I'll write a patch to address these. Thanks! Thanks, Fam! Kevin
diff --git a/block/mirror.c b/block/mirror.c index d2515c7..3c38695 100644 --- a/block/mirror.c +++ b/block/mirror.c @@ -164,6 +164,8 @@ static uint64_t coroutine_fn mirror_iteration(MirrorBlockJob *s) int64_t end, sector_num, next_chunk, next_sector, hbitmap_next_sector; uint64_t delay_ns = 0; MirrorOp *op; + int pnum; + int64_t ret; s->sector_num = hbitmap_iter_next(&s->hbi); if (s->sector_num < 0) { @@ -290,8 +292,22 @@ static uint64_t coroutine_fn mirror_iteration(MirrorBlockJob *s) s->in_flight++; s->sectors_in_flight += nb_sectors; trace_mirror_one_iteration(s, sector_num, nb_sectors); - bdrv_aio_readv(source, sector_num, &op->qiov, nb_sectors, - mirror_read_complete, op); + + ret = bdrv_get_block_status_above(source, NULL, sector_num, + nb_sectors, &pnum); + if (ret < 0 || pnum < nb_sectors || + (ret & BDRV_BLOCK_DATA && !(ret & BDRV_BLOCK_ZERO))) { + bdrv_aio_readv(source, sector_num, &op->qiov, nb_sectors, + mirror_read_complete, op); + } else if (ret & BDRV_BLOCK_ZERO) { + bdrv_aio_write_zeroes(s->target, sector_num, op->nb_sectors, + s->unmap ? BDRV_REQ_MAY_UNMAP : 0, + mirror_write_complete, op); + } else { + assert(!(ret & BDRV_BLOCK_DATA)); + bdrv_aio_discard(s->target, sector_num, op->nb_sectors, + mirror_write_complete, op); + } return delay_ns; }
If guest discards a source cluster, mirroring with bdrv_aio_readv is overkill. Some protocols do zero upon discard, where it's best to use bdrv_aio_write_zeroes, otherwise, bdrv_aio_discard will be enough. Signed-off-by: Fam Zheng <famz@redhat.com> --- block/mirror.c | 20 ++++++++++++++++++-- 1 file changed, 18 insertions(+), 2 deletions(-)