On Sun, Jan 28, 2024 at 05:54:34PM +0100, Christoph Hellwig wrote: > On Fri, Jan 26, 2024 at 03:49:30PM +0000, Matthew Wilcox wrote: > > This doesn't quite make sense to me. Do you mean: > > > > * If the caller modifies data in the folio, it must call folio_mark_dirty() > > * before unlocking the folio to ensure that the folio is not reclaimed. > > * There is no equivalent to write_begin/write_end for shmem. > > > > (also this should go before the Return: section; the return section > > should be the last thing in the kernel-doc) > > So the first sentence and moving the section makes total sense. > The second sentence I don't think is very useful. write_begin/write_end > are relaly just a way for generic_perform_write to do the space > reservation and extending i_size and not really methods in the classic > sense. They should go away from a_ops and certainly don't end up > being mentioned in shmem.c. > > What I have now is this: > > If the caller modifies data in the folio, it must call folio_mark_dirty() > before unlocking the folio to ensure that the folio is not reclaimed. > These is no need to reserve space before calling folio_mark_dirty(). That's totally fine with me. Could I trouble you to elaborate on what you'd like to see a filesystem like ubifs do to replace write_begin/write_end? After my recent patches, those are the only places in ubifs that have a struct page reference. I've been holding off on converting those and writepage because we have plans to eliminate them, but I'm not sure how much longer we can hold off.
