On Wed 27-01-21 15:05:10, Sascha Hauer wrote: > On Tue, Jan 26, 2021 at 05:18:29PM +0100, Jan Kara wrote: > > On Tue 26-01-21 13:34:06, Christoph Hellwig wrote: > > > On Tue, Jan 26, 2021 at 02:17:52PM +0100, Jan Kara wrote: > > > > Well, I don't think that "wait until unfrozen" is that strange e.g. for > > > > Q_SETQUOTA - it behaves like setxattr() or any other filesystem > > > > modification operation. And IMO it is desirable that filesystem freezing is > > > > transparent for operations like these. For stuff like Q_QUOTAON, I agree > > > > that returning EBUSY makes sense but then I'm not convinced it's really > > > > simpler or more useful behavior... > > > > > > If we want it to behave like other syscalls we'll just need to throw in > > > a mnt_want_write/mnt_drop_write pair. Than it behaves exactly like other > > > syscalls. > > > > Right, we could do that. I'd just note that the "wait until unfrozen" and > > holding of sb->s_umount semaphore is equivalent to > > mnt_want_write/mnt_drop_write pair. But I agree > > mnt_want_write/mnt_drop_write is easier to understand and there's no reason > > not to use it. So I'm for that simplification in the new syscall. > > Due to the user_path_at() to the mountpoint the fs won't go away, so I > guess for non-exclusive, non-write quota command I don't need any > additional locking. They still need s_umount in read mode - to serialize against quotaon / quotaoff (happening either through quotactl or filesystem remount). > For non-exclusive, write commands I'll need a > mnt_want_write/mnt_drop_write pair. What about the exclusive, write > commands (Q_QUOTAON/Q_QUOTAOFF)? And these still need s_umount in write mode for similar reasons. So the only thing you really save is the "wait-for-unfreeze" code. Honza -- Jan Kara <jack@xxxxxxxx> SUSE Labs, CR
