On Tue, Jan 04, 2022 at 12:22:15PM +1100, Dave Chinner wrote: > --- a/fs/iomap/buffered-io.c > +++ b/fs/iomap/buffered-io.c > @@ -1098,6 +1098,15 @@ iomap_ioend_can_merge(struct iomap_ioend *ioend, struct iomap_ioend *next) > return false; > if (ioend->io_offset + ioend->io_size != next->io_offset) > return false; > + /* > + * Do not merge physically discontiguous ioends. The filesystem > + * completion functions will have to iterate the physical > + * discontiguities even if we merge the ioends at a logical level, so > + * we don't gain anything by merging physical discontiguities here. > + */ > + if (ioend->io_inline_bio.bi_iter.bi_sector + (ioend->io_size >> 9) != This open codes bio_end_sector() > + next->io_inline_bio.bi_iter.bi_sector) But more importantly I don't think just using the inline_bio makes sense here as the ioend can have multiple bios. Fortunately we should always have the last built bio available in ->io_bio. > + return false; > return true; > } > > @@ -1241,6 +1250,13 @@ iomap_can_add_to_ioend(struct iomap_writepage_ctx *wpc, loff_t offset, > return false; > if (sector != bio_end_sector(wpc->ioend->io_bio)) > return false; > + /* > + * Limit ioend bio chain lengths to minimise IO completion latency. This > + * also prevents long tight loops ending page writeback on all the pages > + * in the ioend. > + */ > + if (wpc->ioend->io_size >= 4096 * PAGE_SIZE) > + return false; And this stops making sense with the impending additions of large folio support. I think we need to count the pages/folios instead as the operations are once per page/folio.
