Re: [PATCH 9/9] mm, page_alloc: optionally disable pcplists during page isolation

[Michal Hocko]

On Fri 25-09-20 13:10:05, Vlastimil Babka wrote:
> On 9/25/20 12:54 PM, David Hildenbrand wrote:
> >>> --- a/mm/page_isolation.c
> >>> +++ b/mm/page_isolation.c
> >>> @@ -15,6 +15,22 @@
> >>>  #define CREATE_TRACE_POINTS
> >>>  #include <trace/events/page_isolation.h>
> >>>  
> >>> +void zone_pcplist_disable(struct zone *zone)
> >>> +{
> >>> +	down_read(&pcp_batch_high_lock);
> >>> +	if (atomic_inc_return(&zone->pcplist_disabled) == 1) {
> >>> +		zone_update_pageset_high_and_batch(zone, 0, 1);
> >>> +		__drain_all_pages(zone, true);
> >>> +	}
> >> Hm, if one CPU is still inside the if-clause, the other one would
> >> continue, however pcp wpould not be disabled and zones not drained when
> >> returning.
> 
> Ah, well spotted, thanks!
> 
> >> (while we only allow a single Offline_pages() call, it will be different
> >> when we use the function in other context - especially,
> >> alloc_contig_range() for some users)
> >>
> >> Can't we use down_write() here? So it's serialized and everybody has to
> >> properly wait. (and we would not have to rely on an atomic_t)
> > Sorry, I meant down_write only temporarily in this code path. Not
> > keeping it locked in write when returning (I remember there is a way to
> > downgrade).
> 
> Hmm that temporary write lock would still block new callers until previous
> finish with the downgraded-to-read lock.
> 
> But I guess something like this would work:
> 
> retry:
>   if (atomic_read(...) == 0) {
>     // zone_update... + drain
>     atomic_inc(...);
>   else if (atomic_inc_return == 1)
>     // atomic_cmpxchg from 0 to 1; if that fails, goto retry
> 
> Tricky, but races could only read to unnecessary duplicated updates + flushing
> but nothing worse?
> 
> Or add another spinlock to cover this part instead of the temp write lock...

Do you plan to post a new version or should I review this one?

-- 
Michal Hocko
SUSE Labs