On Mon, Feb 22, 2021 at 02:58:04PM +0000, Chuck Lever wrote: > > There is a conflict at the end where rq_page_end gets updated. The 5.11 > > code assumes that the loop around the allocator definitely gets all > > the required pages. What tree is this patch based on and is it going in > > during this merge window? While the conflict is "trivial" to resolve, > > it would be buggy because on retry, "i" will be pointing to the wrong > > index and pages potentially leak. Rather than guessing, I'd prefer to > > base a series on code you've tested. > > I posted this patch as a proof of concept. There is a clean-up patch > that goes before it to deal properly with rq_page_end. I can post > both if you really want to apply this and play with it. > It's for the best. It doesn't belong in the series as such but it may affect what the bulk allocator usage looks like. > > > The slowpath for the bulk allocator also sucks a bit for the semantics > > required by this caller. As the bulk allocator does not walk the zonelist, > > it can return failures prematurely -- fine for an optimistic bulk allocator > > that can return a subset of pages but not for this caller which really > > wants those pages. The allocator may need NOFAIL-like semantics to walk > > the zonelist if the caller really requires success or at least walk the > > zonelist if the preferred zone is low on pages. This patch would also > > need to preserve the schedule_timeout behaviour so it does not use a lot > > of CPU time retrying allocations in the presense of memory pressure. > > Waiting half a second before trying again seems like overkill, though. > It is both overkill and time is not directly correlated with memory pressure. However, I would also suggest removing the timeout as a separate patch as it's not related to the bulk allocator in case someone does encounter a high CPU usage problem and bisects it the patch using the bulk allocator for the first time. -- Mel Gorman SUSE Labs
