On 09/01/19 17:51, Kevin Wolf wrote: > Am 09.01.2019 um 17:42 hat Paolo Bonzini geschrieben: >> On 09/01/19 12:23, Kevin Wolf wrote: >>> Also note that this is only metadata preallocation; full preallocation >>> will still return allocated for the protocol layer and so it will always >>> be slow. >> >> Full preallocation these days can create images with preallocated but >> known-zero blocks, I think? > > That would defeat one of the main purposes of preallocation because it > would still require COW and metadata updates on the first write.
Sorry I mean at the protocol level, like FALLOC_FL_ZERO_RANGE. It would still require metadata updates on the filesystem level, unlike "real" full preallocation, but no qcow2 metadata updates. > If there is demand, we could add something like preallocation=data where > data clusters are preallocated but COW/metadata updates still happen at > runtime, but so far nobody has asked for it. Not sure when you would use > it either. > > Kevin >
