On Wed, Jun 06, 2012 at 01:15:30AM +0200, Jan Kara wrote:
> On Tue 05-06-12 15:51:50, Dave Chinner wrote:
> > On Thu, May 24, 2012 at 02:35:38PM +0200, Jan Kara wrote:
> > > > To me the issue at hand is that we have no method of serialising
> > > > multi-page operations on the mapping tree between the filesystem and
> > > > the VM, and that seems to be the fundamental problem we face in this
> > > > whole area of mmap/buffered/direct IO/truncate/holepunch coherency.
> > > > Hence it might be better to try to work out how to fix this entire
> > > > class of problems rather than just adding a complex kuldge that just
> > > > papers over the current "hot" symptom....
> > > Yes, looking at the above table, the amount of different synchronization
> > > mechanisms is really striking. So probably we should look at some
> > > possibility of unifying at least some cases.
> > It seems to me that we need some thing in between the fine grained
> > page lock and the entire-file IO exclusion lock. We need to maintain
> > fine grained locking for mmap scalability, but we also need to be
> > able to atomically lock ranges of pages.
> Yes, we also need to keep things fine grained to keep scalability of
> direct IO and buffered reads...
> > I guess if we were to nest a fine grained multi-state lock
> > inside both the IO exclusion lock and the mmap_sem, we might be able
> > to kill all problems in one go.
> > Exclusive access on a range needs to be granted to:
> > - direct IO
> > - truncate
> > - hole punch
> > so they can be serialised against mmap based page faults, writeback
> > and concurrent buffered IO. Serialisation against themselves is an
> > IO/fs exclusion problem.
> > Shared access for traversal or modification needs to be granted to:
> > - buffered IO
> > - mmap page faults
> > - writeback
> > Each of these cases can rely on the existing page locks or IO
> > exclusion locks to provide safety for concurrent access to the same
> > ranges. This means that once we have access granted to a range we
> > can check truncate races once and ignore the problem until we drop
> > the access. And the case of taking a page fault within a buffered
> > IO won't deadlock because both take a shared lock....
> You cannot just use a lock (not even a shared one) both above and under
> mmap_sem. That is deadlockable in presence of other requests for exclusive
Well, that's assuming that exclusive lock requests form a barrier to
new shared requests. Remember that I'm talking about a range lock
here, which we can make up whatever semantics we'd need, including
having "shared lock if already locked shared" nested locking
semantics which avoids this page-fault-in-buffered-IO-copy-in/out
It also allows writeback to work the same way it does write now when
we take a page fult on a page that is under writeback
> Luckily, with buffered writes the situation isn't that bad. You
> need mmap_sem only before each page is processed (in
> iov_iter_fault_in_readable()). Later on in the loop we use
> iov_iter_copy_from_user_atomic() which doesn't need mmap_sem. So we can
> just get our shared lock after iov_iter_fault_in_readable() (or simply
> leave it for ->write_begin() if we want to give control over the locking to
That would probably work as well, but it much more likely that
people would get it wrong as opposed to special casing the nested
lock semantic in the page fault code...
> > We'd need some kind of efficient shared/exclusive range lock for
> > this sort of exclusion, and it's entirely possible that it would
> > have too much overhead to be acceptible in the page fault path. It's
> > the best I can think of right now.....
> > As it is, a range lock of this kind would be very handy for other
> > things, too (like the IO exclusion locks so we can do concurrent
> > buffered writes in XFS ;).
> Yes, that's what I thought as well. In particular it should be pretty
> efficient in locking a single page range because that's going to be
> majority of calls. I'll try to write something and see how fast it can