On Tue, May 03, 2022 at 07:35:41AM -0700, Christoph Hellwig wrote: > On Fri, Apr 29, 2022 at 06:24:48PM +0100, Matthew Wilcox (Oracle) wrote: > > Filesystems do not necessarily set PageError; instead they will leave > > PageUptodate clear on errors. We should also kmap() the page before > > accessing it in case the page is allocated from HIGHMEM. > > The PageError to PageUptodate change looks sane. But block device > page cache absolutely must not be in highmem, so I don't see a point > in the kmap conversion here. Huh, I didn't know that. I think I've seen that some filesystems also require non-highmem pages for directory or symlink inodes. Coincidentally, I was in Ira's session earlier today where he was talking about reusing/abusing kmap/kmap_local/kmap_atomic to implement protection keys. I'm not entirely sold on the concept, but something like this might also be useful for a CHERI-style architecture. I'm kind of inclined to say "We should always kmap() anything we get from the page cache to support the unmapped-by-default case" because it's not exactly expensive to call kmap() on a page which is permanently mapped. I dunno.
