Re: [PATCH v5 00/12] fold per-CPU vmstats remotely

[Michal Hocko]

On Tue 14-03-23 15:49:09, Marcelo Tosatti wrote:
> On Tue, Mar 14, 2023 at 03:31:21PM +0100, Michal Hocko wrote:
> > On Tue 14-03-23 09:59:37, Marcelo Tosatti wrote:
> > > On Tue, Mar 14, 2023 at 01:25:53PM +0100, Michal Hocko wrote:
> > > > On Mon 13-03-23 13:25:07, Marcelo Tosatti wrote:
> > > > > This patch series addresses the following two problems:
> > > > > 
> > > > >     1. A customer provided some evidence which indicates that
> > > > >        the idle tick was stopped; albeit, CPU-specific vmstat
> > > > >        counters still remained populated.
> > > > > 
> > > > >        Thus one can only assume quiet_vmstat() was not
> > > > >        invoked on return to the idle loop. If I understand
> > > > >        correctly, I suspect this divergence might erroneously
> > > > >        prevent a reclaim attempt by kswapd. If the number of
> > > > >        zone specific free pages are below their per-cpu drift
> > > > >        value then zone_page_state_snapshot() is used to
> > > > >        compute a more accurate view of the aforementioned
> > > > >        statistic.  Thus any task blocked on the NUMA node
> > > > >        specific pfmemalloc_wait queue will be unable to make
> > > > >        significant progress via direct reclaim unless it is
> > > > >        killed after being woken up by kswapd
> > > > >        (see throttle_direct_reclaim())
> > > > 
> > > > I have hard time to follow the actual problem described above. Are you
> > > > suggesting that a lack of pcp vmstat counters update has led to
> > > > reclaim issues? What is the said "evidence"? Could you share more of the
> > > > story please?
> > > 
> > > 
> > >   - The process was trapped in throttle_direct_reclaim().
> > >     The function wait_event_killable() was called to wait condition
> > >     allow_direct_reclaim(pgdat) for current node to be true.
> > >     The allow_direct_reclaim(pgdat) examined the number of free pages
> > >     on the node by zone_page_state() which just returns value in
> > >     zone->vm_stat[NR_FREE_PAGES].
> > > 
> > >   - On node #1, zone->vm_stat[NR_FREE_PAGES] was 0.
> > >     However, the freelist on this node was not empty.
> > > 
> > >   - This inconsistent of vmstat value was caused by percpu vmstat on
> > >     nohz_full cpus. Every increment/decrement of vmstat is performed
> > >     on percpu vmstat counter at first, then pooled diffs are cumulated
> > >     to the zone's vmstat counter in timely manner. However, on nohz_full
> > >     cpus (in case of this customer's system, 48 of 52 cpus) these pooled
> > >     diffs were not cumulated once the cpu had no event on it so that
> > >     the cpu started sleeping infinitely.
> > >     I checked percpu vmstat and found there were total 69 counts not
> > >     cumulated to the zone's vmstat counter yet.
> > > 
> > >   - In this situation, kswapd did not help the trapped process.
> > >     In pgdat_balanced(), zone_wakermark_ok_safe() examined the number
> > >     of free pages on the node by zone_page_state_snapshot() which
> > >     checks pending counts on percpu vmstat.
> > >     Therefore kswapd could know there were 69 free pages correctly.
> > >     Since zone->_watermark = {8, 20, 32}, kswapd did not work because
> > >     69 was greater than 32 as high watermark.
> > 
> > If the imprecision of allow_direct_reclaim is the underlying problem why
> > haven't you used zone_page_state_snapshot instead?
> 
> It might have dealt with problem #1 for this particular case. However,
> looking at the callers of zone_page_state:
> 
>    5   2227  mm/compaction.c <<compaction_suitable>>
>              zone_page_state(zone, NR_FREE_PAGES));
>    6    124  mm/highmem.c <<__nr_free_highpages>>
>              pages += zone_page_state(zone, NR_FREE_PAGES);
>    7    283  mm/page-writeback.c <<node_dirtyable_memory>>
>              nr_pages += zone_page_state(zone, NR_FREE_PAGES);
>    8    318  mm/page-writeback.c <<highmem_dirtyable_memory>>
>              nr_pages = zone_page_state(z, NR_FREE_PAGES);
>    9    321  mm/page-writeback.c <<highmem_dirtyable_memory>>
>              nr_pages += zone_page_state(z, NR_ZONE_INACTIVE_FILE);
>   10    322  mm/page-writeback.c <<highmem_dirtyable_memory>>
>              nr_pages += zone_page_state(z, NR_ZONE_ACTIVE_FILE);
>   11   3091  mm/page_alloc.c <<__rmqueue>>
>              zone_page_state(zone, NR_FREE_CMA_PAGES) >
>   12   3092  mm/page_alloc.c <<__rmqueue>>
>              zone_page_state(zone, NR_FREE_PAGES) / 2) {
> 
> The suggested patchset fixes the problem of where due to nohz_full,
> the delayed timer for vmstat_work can be armed but not executed (which means
> the per-cpu counters can be out of sync for as long as the cpu is in 
> idle while in nohz_full mode).
> 
> You probably do not want to convert all callers of zone_page_state
> into zone_page_state_snapshot (as a justification for the proposed
> patchset).

Yes, I do not really think we want or even need to convert all of them.
But it seems that your direct reclaim throttling example really requires
that. The thing with the remote flushing is that it would suffer from
a similar imprecations as the flushing could be deferred and under
certain conditions really starved. So it is definitely worth fixing the
issue you are seeing without such a complex scheme.

> > Anyway, this is kind of information that is really helpful to have in
> > the patch description.
> 
> Agree: resending a new version with updated commit.

I would really recommend trying out the simple fix and see if it changes
the behavior.

> > [...]
> > > > >     2. With a SCHED_FIFO task that busy loops on a given CPU,
> > > > >        and kworker for that CPU at SCHED_OTHER priority,
> > > > >        queuing work to sync per-vmstats will either cause that
> > > > >        work to never execute, or stalld (i.e. stall daemon)
> > > > >        boosts kworker priority which causes a latency
> > > > >        violation
> > > > 
> > > > Why is that a problem? Out-of-sync stats shouldn't cause major problems.
> > > > Or can they?
> > > 
> > > Consider SCHED_FIFO task that is polling the network queue (say
> > > testpmd).
> > > 
> > > 	do {
> > > 	 	if (net_registers->state & DATA_AVAILABLE) {
> > > 			process_data)();
> > > 		}
> > > 	 } while (!stopped);
> > > 
> > > Since this task runs at SCHED_FIFO priority, kworker won't 
> > > be scheduled to run (therefore per-CPU vmstats won't be
> > > flushed to global vmstats). 
> > 
> > Yes, that is certainly possible. But my main point is that vmstat
> > imprecision shouldn't cause functional problems. That is why we have
> > _snapshot readers to get an exact value where it matters for
> > consistency.
> 
> Understood. Perhaps allow_direct_reclaim should use
> zone_page_state_snapshot, as otherwise it is only precise
> at sysctl_stat_interval intervals?

or even much less than that. The flusher uses WQ infrastructure and even
when a WQ_MEM_RECLAIM one is used this doesn't mean that all workers
could be jammed.
 
> > > Or, if testpmd runs at SCHED_OTHER, then the work item to
> > > flush per-CPU vmstats causes
> > > 
> > > 	testpmd -> kworker
> > > 	kworker: flush per-CPU vmstats
> > > 	kworker -> testpmd
> >
> > And this might cause undesired latencies to the packets being                                                                                    
> > processed by the testpmd task.                                                                                                                   
> 
> > Right but can you have any latencies expectation in a situation like
> > that?
> 
> Not sure i understand what you mean. Example:
> 
> https://www.gabrieleara.it/assets/documents/papers/conferences/2021-ieee-nfv-sdn.pdf
> 
> In general, UDPDK exhibits a much lower
> latency than the in-kernel UDP stack used through the POSIX
> API (e.g., a 69 % reduction from 95 µs down to 29 µs), thanks
> to its ability to bypass the kernel entirely, which in turn
> outperforms the in-kernel TCP stack as available through the
> POSIX API, as expected.
> ...
> Alternatively, application processes can use UDPDK
> with the non-blocking API calls (using the O_NONBLOCK flag)
> and perform some other action while waiting for packets to
> be ready to be sent/received to/from the UDPDK Process,
> instead of performing continuous busy-loops on packet queues.
> However, in this case the cost of a single CPU fully busy due
> to the UDPDK Process itself is anyway unavoidab

If the userspace workload avoids the kernel completely then it is quite
unlikely that there is any pcp work to be flushed for in-kernel
counters.

That being said, I am nor saying remote flushing is not useful. I just
think that the issue you are reporting here could be fixed by a much
simpler fix that doesn't change the way how the flushing is performed.
Such a large rework should be justified by performance numbers. It
should be also explained how do we end up doing a lot of work on
isolated cpus or a pure user space workload.

-- 
Michal Hocko
SUSE Labs