Messages in this thread | | | Date | Wed, 4 Jan 2012 19:00:33 +0100 | From | Jan Kara <> | Subject | Re: linux-next: build failure after merge of the vfs tree |
| |
On Wed 04-01-12 13:50:20, Dave Chinner wrote: > On Wed, Jan 04, 2012 at 02:17:54AM +0000, Al Viro wrote: > > I'm still not > > sure about ->statfs(), BTW - any input on that would be welcome. Can > > it end up blocked on a frozen fs until said fs is thawed? > > I don't see why this should ever happen - ->statfs has to work on > read-only filesystems so shoul dnot be modifying state, and hence > should never need to care about the frozen state of the superblock. Well, I'm also not aware of a filesystem where ->statfs would wait on frozen filesystem. Just note that e.g. for stat(2) frozen filesystem and RO filesystem *are* different because of atime updates. So stat(2) can block on frozen fs because of atime update while on RO filesystem it is just fine.
> So from a ->statfs POV, a frozen filesystem should look just like a > read-only filesystem. If frozen filesystems are holding locks that > ->statfs can block on until the filesystem us thawed, then I'd > consider that a bug in the filesystem freeze implementation.... In an ideal world yes. Practically, current freeze code has races (vfs_check_frozen() is a totally racy check) which can leave processes waiting for frozen fs with filesystem locks held. I believe we need something like mnt_want_write()/mnt_drop_write() for freezing code in ->page_mkwrite() and ->write_begin/->write_end. I'm now looking into how to do that in the best way.
> > to convert ustat(2) to "wait for thaw" semantics (should be interruptible, > > BTW) or document that ->statfs() is not allowed to wait for thawing. > > It's far too subtle to leave undocumented... > > The latter, IMO. Agreed.
Honza -- Jan Kara <jack@suse.cz> SUSE Labs, CR
| |