Thanks Vlastimil
On Wed, Oct 13, 2021 at 05:39:36PM +0200, Vlastimil Babka wrote:
> > +/*
> > + * Account for pages written if tasks are throttled waiting on dirty
> > + * pages to clean. If enough pages have been cleaned since throttling
> > + * started then wakeup the throttled tasks.
> > + */
> > +void __acct_reclaim_writeback(pg_data_t *pgdat, struct page *page,
> > + int nr_throttled)
> > +{
> > + unsigned long nr_written;
> > +
> > + __inc_node_page_state(page, NR_THROTTLED_WRITTEN);
>
> Is this intentionally using the __ version that normally expects irqs to be
> disabled (AFAIK they are not in this path)? I think this is rarely used cold
> path so it doesn't seem worth to trade off speed for accuracy.
>
It was intentional because IRQs can be disabled and if it's race-prone,
it's not overly problematic but you're right, better to be safe. I changed
it to the safe type as it's mostly free on x86, arm64 and s390 and for
other architectures, this is a slow path.
> > + nr_written = node_page_state(pgdat, NR_THROTTLED_WRITTEN) -
> > + READ_ONCE(pgdat->nr_reclaim_start);
>
> Even if the inc above was safe, node_page_state() will return only the
> global counter, so the value we read here will only actually increment when
> some cpu's counter overflows, so it will be "bursty". Maybe it's ok, just
> worth documenting?
>
I didn't think the penalty of doing an accurate read while writeback
throttled is worth it. I'll add a comment.
> > +
> > + if (nr_written > SWAP_CLUSTER_MAX * nr_throttled)
> > + wake_up_all(&pgdat->reclaim_wait);
>
> Hm it seems a bit weird that the more tasks are throttled, the more we wait,
> and then wake up all. Theoretically this will lead to even more
> bursty/staggering herd behavior. Could be better to wake up single task each
> SWAP_CLUSTER_MAX, and bump nr_reclaim_start? But maybe it's not a problem in
> practice due to HZ/10 timeouts being short enough?
>
Yes, the more tasks are throttled the longer tasks wait because tasks are
allocating faster than writeback can complete so I wanted to reduce the
allocation pressure. I considered waking one task at a time but there is
no prioritisation of tasks on the waitqueue and it's not clear that the
additional complexity is justified. With inaccurate counters, a light
allocator could get throttled for the full timeout unnecessarily.
Even if we were to wake one task at a time, I would prefer it was done
as a potential optimisation on top.
Diff on top based on review feedback;
diff --git a/mm/vmscan.c b/mm/vmscan.c
index bcd22e53795f..735b1f2b5d9e 100644
--- a/mm/vmscan.c
+++ b/mm/vmscan.c
@@ -1048,7 +1048,15 @@ void __acct_reclaim_writeback(pg_data_t *pgdat, struct page *page,
{
unsigned long nr_written;
- __inc_node_page_state(page, NR_THROTTLED_WRITTEN);
+ inc_node_page_state(page, NR_THROTTLED_WRITTEN);
+
+ /*
+ * This is an inaccurate read as the per-cpu deltas may not
+ * be synchronised. However, given that the system is
+ * writeback throttled, it is not worth taking the penalty
+ * of getting an accurate count. At worst, the throttle
+ * timeout guarantees forward progress.
+ */
nr_written = node_page_state(pgdat, NR_THROTTLED_WRITTEN) -
READ_ONCE(pgdat->nr_reclaim_start);
