On Tue, Jan 14, 2025 at 09:55:21PM -0800, Christoph Hellwig wrote:
> On Mon, Jan 13, 2025 at 08:40:51AM -0500, Brian Foster wrote:
> > Sorry if this is out of left field as I haven't followed the discussion
> > closely, but I presumed one of the reasons Darrick and Christoph raised
> > the idea of using the folio batch thing I'm playing around with on zero
> > range for buffered writes would be to acquire and lock all targeted
> > folios up front. If so, would that help with what you're trying to
> > achieve here? (If not, nothing to see here, move along.. ;).
>
> I mostly thought about acquiring, as locking doesn't really have much
> batching effects. That being said, no that you got the idea in my mind
> here's my early morning brainfart on it:
>
> Let's ignore DIRECT I/O for the first step. In that case lookup /
> allocation and locking all folios for write before copying data will
> remove the need for i_rwsem in the read and write path. In a way that
> sounds perfect, and given that btrfs already does that (although in a
> very convoluted way) we know it's possible.
Yes, this seems like a sane, general approach to allowing concurrent
buffered writes (and reads).
> But direct I/O throws a big monkey wrench here as already mentioned by
> others. Now one interesting thing some file systems have done is
> to serialize buffered against direct I/O, either by waiting for one
> to finish, or by simply forcing buffered I/O when direct I/O would
> conflict.
Right. We really don't want to downgrade to buffered IO if we can
help it, though.
> It's easy to detect outstanding direct I/O using i_dio_count
> so buffered I/O could wait for that, and downgrading to buffered I/O
> (potentially using the new uncached mode from Jens) if there are any
> pages on the mapping after the invalidation also sounds pretty doable.
It's much harder to sanely serialise DIO against buffered writes
this way, because i_dio_count only forms a submission barrier in
conjunction with the i_rwsem being held exclusively. e.g. ongoing
DIO would result in the buffered write being indefinitely delayed.
I think the model and method that bcachefs uses is probably the best
way to move forward - the "two-state exclusive shared" lock which it
uses to do buffered vs direct exclusion is a simple, easy way to
handle this problem. The same-state shared locking fast path is a
single atomic cmpxchg operation, so it has neglible extra overhead
compared to using a rwsem in the shared DIO fast path.
The lock also has non-owner semantics, so DIO can take it during
submission and then drop it during IO completion. This solves the
problem we currently use the i_rwsem and
inode_dio_{start,end/wait}() to solve (i.e. create a DIO submission
barrier and waiting for all existing DIO to drain).
IOWs, a two-state shared lock provides the mechanism to allow DIO
to be done without holding the i_rwsem at all, as well as being able
to elide two atomic operations per DIO to track in-flight DIOs.
We'd get this whilst maintaining buffered/DIO coherency without
adding any new overhead to the DIO path, and allow concurrent
buffered reads and writes that have their atomicity defined by the
batched folio locking strategy that Brian is working on...
This only leaves DIO coherency issues with mmap() based IO as an
issue, but that's a problem for a different day...
> I don't really have time to turn this hand waving into, but maybe we
> should think if it's worthwhile or if I'm missing something important.
If people are OK with XFS moving to exclusive buffered or DIO
submission model, then I can find some time to work on the
converting the IO path locking to use a two-state shared lock in
preparation for the batched folio stuff that will allow concurrent
buffered writes...
-Dave.
--
Dave Chinner
david@xxxxxxxxxxxxx
