On Fri, Aug 21, 2020 at 7:57 PM Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote: > > On Fri, Aug 21, 2020 at 05:31:52PM +0800, Zhaoyang Huang wrote: > > This patch has been verified on an android system and reduces 15% of > > UNITERRUPTIBLE_SLEEP_BLOCKIO which was used to be caused by wrong > > ra->ra_pages. > > Wait, what? Readahead doesn't sleep on the pages it's requesting. > Unless ... your file access pattern is random, so you end up submitting > a readahead I/O that's bigger than needed, so takes longer for the page > you actually wanted to be returned. I know we have the LOTSAMISS > logic, but that's not really enough. actually, async read even if hitting the marker will also introduce huge mount of read by wrong ra->ra_pages, which will be used as req_size for ondemand_readahead. > > OK, assuming this problem is really about sync mmap (ie executables), > this makes a bit more sense. I think the real problem is here: > > ra->start = max_t(long, 0, offset - ra->ra_pages / 2); > ra->size = ra->ra_pages; > ra->async_size = ra->ra_pages / 4; > ra_submit(ra, mapping, file); > > which actually skips all the logic we have in ondemand_readahead() > for adjusting the readahead size. Ugh, this is a mess. > > I think a quick fix to your problem will be just replacing ra->ra_pages > with bdi->ra_pages in do_sync_mmap_readahead() and leaving ra->ra_pages > alone everywhere else. > > We need a smarter readahead algorithm for mmap'ed files, and I don't have > time to work on it right now. So let's stick to the same dumb algorithm, > but make it responsive to bdi ra_pages being reset.
