On Wed, Oct 02, 2024 at 12:41:29AM -0700, Christoph Hellwig wrote: > > > This looks pretty ugly. What speaks against a version of > > > filemap_add_folio that doesn't charge the memcg? > > What I'd propose is something like the patch below, plus proper > documentation. I like this much better as well. > Note that this now does the uncharge on the unlocked folio in the > error case. From a quick look that should be fine, but someone who > actually knows the code needs to confirm that. That's fine. For the same reason the non-atomic __folio_clear_locked() is fine in that case. The folio just has to be exclusive.
