On Wed 30-06-21 18:15:09, Linus Torvalds wrote:

On Wed, Jun 30, 2021 at 10:25 AM Jan Kara [off-list ref] wrote:

quoted

  could you please pull from

No.

There is no way I'll merge something this broken.

Looking up a page in the page cache is just about the most critical
thing there is, and this introduces a completely pointless lock for
that situation.

Does it take the lock only when it creates the page? No. It takes the
lock in filemap_fault() even if it found a valid page in the page
cache.

Hum, fair point. I did filemap_fault() the way it is because I was mostly
just lifting fs-private lock into the VFS one in that code path and
ext4/xfs/f2fs and others grabbed this lock unconditionally in their fault
paths (before calling into filemap_fault()). But you are right that now
that we have the lock in VFS, we can actually do better and have a fast
path when everything is cached and uptodate where we can avoid grabbing the
lock. That being said I don't expect the optimization to matter too much
because in do_read_fault() we first call do_fault_around() which will
exactly map pages that are already in cache and uptodate so we usually get
into filemap_fault() only for pages that are not present or not uptodate.
So do you think the optimization is still worth it despite
do_fault_around()? I guess I can try to see how many times I can see a page
that would benefit from this optimization in filemap_fault() on my test
machine - there are also write faults that don't call do_fault_around() -
and if it's noticeable fraction reorganize filemap_fault() so that we don't
take the lock if the page is present and uptodate...

								Honza
-- 
Jan Kara [off-list ref]
SUSE Labs, CR