On 2019/10/22 下午8:23, David Sterba wrote: > On Tue, Oct 22, 2019 at 02:30:22PM +0800, Qu Wenruo wrote: >> BTW, there is one important compatibility problem related to all the BGI >> related features. >> >> Although I'm putting the BG_TREE feature as incompatible feature, but in >> theory, it should be RO compatible. > > It could be RO compatible yes. > >> As except extent/bg tree, we *should* read the fs without any problem. >> >> But the problem is, current btrfs mount (including btrfs-check) still >> need to go through the block group item search, even for permanent RO mount. >> >> This get my rescue mount option patchset to be involved. >> If we have such skip_bg feature earlier, we can completely afford to >> make all these potential features as RO compatible. >> >> >> Now my question is, should we put this feature still as incompatible >> feature? > > In some way it would probably have to be incompat, either full or RO. As > unknown tree items are ignored, if the rest of the filesystem provides > enough information to access the data, then incompat RO sounds like the > best option. And that's probably independent of how exactly the new BGI > is done. > But the problem is, older fs can't mount it at all, no matter RO or RW.
It's now completely dependent on how older kernel handles missing block group items. If current kernel has something like skip_bg or fall back to RO + skip_bg by default, then it's really RO compatible. But we all know how picky current btrfs is about missing block group items... So, I guess we will have a RO compatible feature that can't really be mounted RO by older kernel at last? Thanks, Qu
signature.asc
Description: OpenPGP digital signature
