Re: [RFC PATCH] mm: Replace ->launder_folio() with flush and wait

[Miklos Szeredi]

On Tue, 19 Mar 2024 at 17:13, Miklos Szeredi <miklos@xxxxxxxxxx> wrote:
>
> On Tue, 19 Mar 2024 at 15:15, David Howells <dhowells@xxxxxxxxxx> wrote:
>
> > What particular usage case of invalidate_inode_pages2() are you thinking of?
>
> FUSE_NOTIFY_INVAL_INODE will trigger invalidate_inode_pages2_range()
> to clean up the cache.
>
> The server is free to discard writes resulting from this invalidation
> and delay reads in the region until the invalidation finishes.  This
> would no longer work with your change, since the mapping could
> silently be reinstated between the writeback and the removal from the
> cache due to the page being unlocked/relocked.

This would also matter if a distributed filesystem wanted to implement
coherence even if there are mmaps.   I.e. a client could get exclusive
access to a region by issuing FUSE_NOTIFY_INVAL_INODE on all other
clients and blocking reads.  With your change this would fail.

Again, this is purely theoretical, and without a way to differentiate
between the read-only and write cases it has limited usefulness.
Adding leases to fuse (which I plan to do) would make this much more
useful.

Thanks,
Miklos