Re: [RFC v4 PATCH 3/5] mm/rmqueue_bulk: alloc without touching individual page structure

[Aaron Lu <aaron.lu@intel.com>]

On Wed, Oct 17, 2018 at 12:20:42PM +0100, Mel Gorman wrote:
> On Wed, Oct 17, 2018 at 02:33:28PM +0800, Aaron Lu wrote:
> > Profile on Intel Skylake server shows the most time consuming part
> > under zone->lock on allocation path is accessing those to-be-returned
> > page's "struct page" on the free_list inside zone->lock. One explanation
> > is, different CPUs are releasing pages to the head of free_list and
> > those page's 'struct page' may very well be cache cold for the allocating
> > CPU when it grabs these pages from free_list' head. The purpose here
> > is to avoid touching these pages one by one inside zone->lock.
> > 
> 
> I didn't read this one in depth because it's somewhat ortogonal to the
> lazy buddy merging which I think would benefit from being finalised and
> ensuring that there are no reductions in high-order allocation success
> rates.  Pages being allocated on one CPU and freed on another is not that
> unusual -- ping-pong workloads or things like netperf used to exhibit
> this sort of pattern.
> 
> However, this part stuck out
> 
> > +static inline void zone_wait_cluster_alloc(struct zone *zone)
> > +{
> > +	while (atomic_read(&zone->cluster.in_progress))
> > +		cpu_relax();
> > +}
> > +
> 
> RT has had problems with cpu_relax in the past but more importantly, as
> this delay for parallel compactions and allocations of contig ranges,
> we could be stuck here for very long periods of time with interrupts

The longest possible time is one CPU accessing pcp->batch number cold
cachelines. Reason:
When zone_wait_cluster_alloc() is called, we already held zone lock so
no more allocations are possible. Waiting in_progress to become zero
means waiting any CPU that increased in_progress to finish processing
their allocated pages. Since they will at most allocate pcp->batch pages
and worse case are all these page structres are cache cold, so the
longest wait time is one CPU accessing pcp->batch number cold cache lines.

I have no idea if this time is too long though.

> disabled. It gets even worse if it's from an interrupt context such as
> jumbo frame allocation or a high-order slab allocation that is atomic.

My understanding is atomic allocation won't trigger compaction, no?

> These potentially large periods of time with interrupts disabled is very
> hazardous.

I see and agree, thanks for pointing this out.
Hopefully, the above mentioned worst case time won't be regarded as
unbound or too long.

> It may be necessary to consider instead minimising the number
> of struct page update when merging to PCP and then either increasing the
> size of the PCP or allowing it to exceed pcp->high for short periods of
> time to batch the struct page updates.

I don't quite follow this part. It doesn't seem possible we can exceed
pcp->high in allocation path, or are you talking about free path?

And thanks a lot for the review!