On Wed, Jan 16, 2019 at 4:54 PM Linus Torvalds
[off-list ref] wrote:

On Wed, Jan 16, 2019 at 11:45 AM Dave Chinner [off-list ref] wrote:

quoted

I'm assuming that you can invalidate the page cache reliably by a
means that does not repeated require probing to detect invalidation
has occurred. I've mentioned one method in this discussion
already...

Yes. And it was made clear to you that it was a bug in xfs dio and
what the right thing to do was.

Side note: I actually think we *do* the right thing. Even for xfs. I
couldn't find the alleged place that invalidates the page cache on dio
reads.

The *generic* dio code only does it for writes (which is correct and
fine). And maybe xfs has some extra invalidation, but I don't see it.

So I actually hope your "you can use direct-io read to do directed
invalidating of the page cache" isn't true. I admittedly did *not* try
to delve very deeply into it, but the invalidates I found looked
correct. The generic code does it for writes, and at least ext4 does
the "writeback and wait" for reads.

There *does* seem to be a 'invalidate_inode_pages2_range()' call in
iomap_dio_rw(). That has a *comment* that says it only is for writes,
but it looks to me like it would trigger for reads too.

Just a plain bug/oversight? Or me misreading things.

So yes, maybe xfs does that "invalidate on read", but it really seems
to be just a bug. If the xfs people insist on keeping the bug, fine
(looks like gfs2 and xfs are the only users), but it seems kind of
sad.

             Linus