On Fri, May 13, 2022 at 05:19:18PM +0200, Nicolas Saenz Julienne wrote: > On Fri, 2022-05-13 at 16:04 +0100, Mel Gorman wrote: > > On Thu, May 12, 2022 at 12:37:43PM -0700, Andrew Morton wrote: > > > On Thu, 12 May 2022 09:50:43 +0100 Mel Gorman <mgorman@xxxxxxxxxxxxxxxxxxx> wrote: > > > > > > > From: Nicolas Saenz Julienne <nsaenzju@xxxxxxxxxx> > > > > > > > > Some setups, notably NOHZ_FULL CPUs, are too busy to handle the per-cpu > > > > drain work queued by __drain_all_pages(). So introduce a new mechanism to > > > > remotely drain the per-cpu lists. It is made possible by remotely locking > > > > 'struct per_cpu_pages' new per-cpu spinlocks. A benefit of this new scheme > > > > is that drain operations are now migration safe. > > > > > > > > There was no observed performance degradation vs. the previous scheme. > > > > Both netperf and hackbench were run in parallel to triggering the > > > > __drain_all_pages(NULL, true) code path around ~100 times per second. > > > > The new scheme performs a bit better (~5%), although the important point > > > > here is there are no performance regressions vs. the previous mechanism. > > > > Per-cpu lists draining happens only in slow paths. > > > > > > > > Minchan Kim tested this independently and reported; > > > > > > > > My workload is not NOHZ CPUs but run apps under heavy memory > > > > pressure so they goes to direct reclaim and be stuck on > > > > drain_all_pages until work on workqueue run. > > > > > > > > unit: nanosecond > > > > max(dur) avg(dur) count(dur) > > > > 166713013 487511.77786438033 1283 > > > > > > > > From traces, system encountered the drain_all_pages 1283 times and > > > > worst case was 166ms and avg was 487us. > > > > > > > > The other problem was alloc_contig_range in CMA. The PCP draining > > > > takes several hundred millisecond sometimes though there is no > > > > memory pressure or a few of pages to be migrated out but CPU were > > > > fully booked. > > > > > > > > Your patch perfectly removed those wasted time. > > > > > > I'm not getting a sense here of the overall effect upon userspace > > > performance. As Thomas said last year in > > > https://lkml.kernel.org/r/87v92sgt3n.ffs@tglx > > > > > > : The changelogs and the cover letter have a distinct void vs. that which > > > : means this is just another example of 'scratch my itch' changes w/o > > > : proper justification. > > > > > > Is there more to all of this than itchiness and if so, well, you know > > > the rest ;) > > > > > > > I think Minchan's example is clear-cut. The draining operation can take > > an arbitrary amount of time waiting for the workqueue to run on each CPU > > and can cause severe delays under reclaim or CMA and the patch fixes > > it. Maybe most users won't even notice but I bet phone users do if a > > camera app takes too long to open. > > > > The first paragraphs was written by Nicolas and I did not want to modify > > it heavily and still put his Signed-off-by on it. Maybe it could have > > been clearer though because "too busy" is vague when the actual intent > > is to avoid interfering with RT tasks. Does this sound better to you? > > > > Some setups, notably NOHZ_FULL CPUs, may be running realtime or > > latency-sensitive applications that cannot tolerate interference > > due to per-cpu drain work queued by __drain_all_pages(). Introduce > > a new mechanism to remotely drain the per-cpu lists. It is made > > possible by remotely locking 'struct per_cpu_pages' new per-cpu > > spinlocks. This has two advantages, the time to drain is more > > predictable and other unrelated tasks are not interrupted. > > > > You raise a very valid point with Thomas' mail and it is a concern that > > the local_lock is no longer strictly local. We still need preemption to > > be disabled between the percpu lookup and the lock acquisition but that > > can be done with get_cpu_var() to make the scope clear. > > This isn't going to work in RT :( > > get_cpu_var() disables preemption hampering RT spinlock use. There is more to > it in Documentation/locking/locktypes.rst. > Bah, you're right. A helper that called preempt_disable() on !RT and migrate_disable() on RT would work although similar to local_lock with a different name. I'll look on Monday to see how the code could be restructured to always have the get_cpu_var() call immediately before the lock acquisition. Once that is done, I'll look what sort of helper that "disables preempt/migration, lookup pcp structure, acquire lock, enable preempt/migration". It's effectively the magic trick that local_lock uses to always lock the right pcpu lock but we want the spinlock semantics for remote drain. -- Mel Gorman SUSE Labs
