Message ID | 1385387840-17307-5-git-send-email-pl@kamp.de |
---|---|
State | New |
Headers | show |
Il 25/11/2013 14:57, Peter Lieven ha scritto: > + /* Guess the internal cluster (page) size of the iscsi target by the means > + * of opt_unmap_gran. Transfer the unmap granularity only if it has a > + * reasonable size for bdi->cluster_size */ > + if (iscsilun->bl.opt_unmap_gran * iscsilun->block_size >= 64 * 1024 && > + iscsilun->bl.opt_unmap_gran * iscsilun->block_size <= 16 * 1024 * 1024) { > + bdi->cluster_size = iscsilun->bl.opt_unmap_gran * iscsilun->block_size; > + } I think you are mixing many different concepts: * The optimal unmap granularity is good as a suggestion for the cluster size of higher-level formats. * The optimal transfer granularity (block limits page, bytes 6-7, min_io_size in Linux) could be used to adjust the length of transfers in "qemu-img convert". I have not really thought much about *how* to do it. * The optimal transfer (block limits page, bytes 12-15, opt_io_size in Linux) should not be used in "qemu-img convert", I think, unless you can actually report performance improvements. This is because in "qemu-img convert" we need to write the data anyway to the target. We cannot schedule other commands between two transfers. So I don't think any delays incurred by a very large write should matter. The maximum transfer length (block limits page, bytes 8-11) should be handled instead in the iSCSI driver, but I do not see the need to do this unless we have reports of something not working. Paolo
Il 25/11/2013 14:57, Peter Lieven ha scritto: > this patch aims to set bdi->cluster_size to the internal page size > of the iscsi target so that enabled callers can align requests > properly. > > Signed-off-by: Peter Lieven <pl@kamp.de> > --- > block/iscsi.c | 7 +++++++ > 1 file changed, 7 insertions(+) > > diff --git a/block/iscsi.c b/block/iscsi.c > index 93fee6d..75d6b87 100644 > --- a/block/iscsi.c > +++ b/block/iscsi.c > @@ -1580,6 +1580,13 @@ static int iscsi_get_info(BlockDriverState *bs, BlockDriverInfo *bdi) > IscsiLun *iscsilun = bs->opaque; > bdi->unallocated_blocks_are_zero = !!iscsilun->lbprz; > bdi->can_write_zeroes_with_unmap = iscsilun->lbprz && iscsilun->lbp.lbpws; > + /* Guess the internal cluster (page) size of the iscsi target by the means > + * of opt_unmap_gran. Transfer the unmap granularity only if it has a > + * reasonable size for bdi->cluster_size */ > + if (iscsilun->bl.opt_unmap_gran * iscsilun->block_size >= 64 * 1024 && > + iscsilun->bl.opt_unmap_gran * iscsilun->block_size <= 16 * 1024 * 1024) { > + bdi->cluster_size = iscsilun->bl.opt_unmap_gran * iscsilun->block_size; > + } > return 0; > } > > After looking at patch 5, I think this one is correct. Reviewed-by: Paolo Bonzini <pbonzini@redhat.com> but I think you should make higher layers (qcow2) use the lower BDS's granularity, at least as a default. Paolo
diff --git a/block/iscsi.c b/block/iscsi.c index 93fee6d..75d6b87 100644 --- a/block/iscsi.c +++ b/block/iscsi.c @@ -1580,6 +1580,13 @@ static int iscsi_get_info(BlockDriverState *bs, BlockDriverInfo *bdi) IscsiLun *iscsilun = bs->opaque; bdi->unallocated_blocks_are_zero = !!iscsilun->lbprz; bdi->can_write_zeroes_with_unmap = iscsilun->lbprz && iscsilun->lbp.lbpws; + /* Guess the internal cluster (page) size of the iscsi target by the means + * of opt_unmap_gran. Transfer the unmap granularity only if it has a + * reasonable size for bdi->cluster_size */ + if (iscsilun->bl.opt_unmap_gran * iscsilun->block_size >= 64 * 1024 && + iscsilun->bl.opt_unmap_gran * iscsilun->block_size <= 16 * 1024 * 1024) { + bdi->cluster_size = iscsilun->bl.opt_unmap_gran * iscsilun->block_size; + } return 0; }
this patch aims to set bdi->cluster_size to the internal page size of the iscsi target so that enabled callers can align requests properly. Signed-off-by: Peter Lieven <pl@kamp.de> --- block/iscsi.c | 7 +++++++ 1 file changed, 7 insertions(+)