On Thu 07-02-13 13:43:42, Dave Chinner wrote: > On Wed, Feb 06, 2013 at 08:25:34PM +0100, Jan Kara wrote: > > On Wed 06-02-13 10:25:12, Dave Chinner wrote: > > > On Mon, Feb 04, 2013 at 01:38:31PM +0100, Jan Kara wrote: > > > > On Thu 31-01-13 16:07:57, Andrew Morton wrote: > > > > > > c) i_mutex doesn't allow any paralellism of operations using it and some > > > > > > filesystems workaround this for specific cases (e.g. DIO reads). Using > > > > > > range locking allows for concurrent operations (e.g. writes, DIO) on > > > > > > different parts of the file. Of course, range locking itself isn't > > > > > > enough to make the parallelism possible. Filesystems still have to > > > > > > somehow deal with the concurrency when manipulating inode allocation > > > > > > data. But the range locking at least provides a common VFS mechanism for > > > > > > serialization VFS itself needs and it's upto each filesystem to > > > > > > serialize more if it needs to. > > > > > > > > > > That would be useful to end-users, but I'm having trouble predicting > > > > > *how* useful. > > > > As Zheng said, there are people interested in this for DIO. Currently > > > > filesystems each invent their own tweaks to avoid the serialization at > > > > least for the easiest cases. > > > > > > The thing is, this won't replace the locking those filesystems use > > > to parallelise DIO - it just adds another layer of locking they'll > > > need to use. The locks filesystems like XFS use to serialise IO > > > against hole punch also serialise against many more internal > > > functions and so if these range locks don't have the same capability > > > we're going to have to retain those locks even after the range locks > > > are introduced. It basically means we're going to have two layers > > > of range locks - one for IO sanity and atomicity, and then this > > > layer just for hole punch vs mmap. > > > > > > As i've said before, what we really need in XFS is IO range locks > > > because we need to be able to serialise operations against IO in > > > progress, not page cache operations in progress. > > Hum, I'm not sure I follow you here. So mapping tree lock + PageLocked + > > PageWriteback serialize all IO for part of the file underlying the page. > > I.e. at most one of truncate (punch hole), DIO, writeback, buffered write, > > buffered read, page fault can run on that part of file. > > Right, it serialises page cache operations sufficient to avoid > page cache coherence problems, but it does not serialise operations > sufficiently to provide atomicity between operations that should be > atomic w.r.t. each other. > > > So how come it > > doesn't provide enough serialization for XFS? > > > > Ah, is it the problem that if two threads do overlapping buffered writes > > to a file then we can end up with data mixed from the two writes (if we > > didn't have something like i_mutex)? > > That's one case of specific concern - the POSIX write() atomicity > guarantee So I was searching for this both yesterday and today and I didn't find anywhere any comment explaining how concurrent writes should behave. But regardless how spec defines it, we provided write vs write exclusion so far and I can imagine changing this could break some applications. So here I agree we probably have no choice... > - but it indicates the cause of many of my other concerns, > too. e.g. write vs prealloc, write vs punch, read vs truncate, write > vs truncate, buffered vs direct write, etc. > > Basically, page-cache granularity locking for buffered IO means that > it cannot be wholly serialised against any other operation in > progress. That means we can't use the range lock to provide a > barrier to guarantee that no IO is currently in progress at all, and > hence it doesn't provide the IO barrier semantics we need for > various operations within XFS. Well, I never really thought about write() call as about a single operation but rather as about a sequence of block-sized operations. But I agree this is a developer / implementation centric view and doesn't make much sence from user POV. Users could reasonably expect that if they do write to a range and punch hole to the same range then they either see the whole range empty or whole range written. Or the similar problem with a truncate you mention below. The page granularity of buffered IO locking doesn't provide that level of consistency. So I agree with you that if we want to get rid of i_mutex we'd need to lock the whole range specified by the syscall (at least for writes, for reads I'm not that convinced because that would heavily reduce the amount of parallelism we have in the code now for some workloads - there the page granularity would be better I think and consistency would be unchanged from what we have now). But if we are going to do the locking for the range specified by syscall it's going to make the write path heavier on CPU by 2% or so for small writes workload from my measurements. And I'm not sure that's acceptable :( Sigh... But I guess I'll code it without handling mmap for now and see what the situation exactly is (how much we gain from dropping i_mutex etc.). Honza -- Jan Kara <jack@xxxxxxx> SUSE Labs, CR -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>