<SNIP>
Would a qcow2 image with preallocation=metadata be possible on an iSCSI data store? ayal?
nope. metadata preallocation means that each logical block has a corresponding physical block. Ayal, by saying "logical block" and physical block here, what do they stand for in linux systems? I guess, physical block is "the scsi lun disk", logical block is "lvm disk"? right?
No, guest writing to block X, qcow maps X to Y on underlying device (e.g. LV) X is logical in example above. Y is 'physical' *Warning*, following explanation is a bit convoluted ;) Metadata preallocation means that all qcow clusters are already preset with every X mapped to a Y. Now on block storage, if guest writes to an X where X is mapped to Y which is beyond device size (because it's thinly provisioned), we would need to extend device to at least Y if not beyond. Worst case is if the guest I/O is to a block which is mapped to offset = size of virtual disk, which would force us to preallocate the entire disk at this point for a single block.
With files this is fine as you can seek wherever you want and the file will remain sparse. With block devices this makes little sense as the second the guest accesses a block which is mapped to an unallocated physical block we'd have to allocate all the area up to that point. (btw, qemu-img will fail if you try to create such an image on a block device) _______________________________________________ Users mailing list Users@ovirt.org http://lists.ovirt.org/mailman/listinfo/users
-- Shu Ming <shuming@linux.vnet.ibm.com> IBM China Systems and Technology Laboratory