On 09.08.19 14:47, Vladimir Sementsov-Ogievskiy wrote: > 09.08.2019 15:25, Max Reitz wrote: >> On 09.08.19 09:50, Vladimir Sementsov-Ogievskiy wrote: >>> 07.08.2019 21:01, Max Reitz wrote: >>>> On 07.08.19 10:07, Vladimir Sementsov-Ogievskiy wrote: >>>>> Limit block_status querying to request bounds on write notifier to >>>>> avoid extra seeking. >>>> >>>> I don’t understand this reasoning. Checking whether something is >>>> allocated for qcow2 should just mean an L2 cache lookup. Which we have >>>> to do anyway when we try to copy data off the source. >>> >>> But for raw it's seeking. >> >> (1) That’s a bug in block_status then, isn’t it? >> >> file-posix cannot determine the allocation status, or rather, everything >> is allocated. bdrv_co_block_status() should probably pass @want_zero on >> to the driver’s implementation, and file-posix should just >> unconditionally return DATA if it’s false. >> >> (2) Why would you even use sync=top for raw nodes? >> > > As I described in parallel letters, raw was bad example. NBD is good.
Does NBD support backing files? > Anyway, now I'm refactoring cluster skipping more deeply for v2. > > About top-mode: finally block-status should be used to improve other > modes too. In virtuozzo we skip unallocated for full mode too, for example. But this patch here is about sync=top. Skipping is an optimization, the block_status querying here happens because copying anything that isn’t allocated in the top layer would be wrong. Max > Unfortunately, backup is most long-term thing to upstream for me..
signature.asc
Description: OpenPGP digital signature
