Re: [PATCH] mm/memory_hotplug: Make unpopulated zones PCP structures unreachable during hot remove

[Michal Hocko]

On Fri 09-04-21 14:42:21, Mel Gorman wrote:
> On Fri, Apr 09, 2021 at 02:48:12PM +0200, Michal Hocko wrote:
> > On Fri 09-04-21 14:42:58, Michal Hocko wrote:
> > > On Fri 09-04-21 13:09:57, Mel Gorman wrote:
> > > > zone_pcp_reset allegedly protects against a race with drain_pages
> > > > using local_irq_save but this is bogus. local_irq_save only operates
> > > > on the local CPU. If memory hotplug is running on CPU A and drain_pages
> > > > is running on CPU B, disabling IRQs on CPU A does not affect CPU B and
> > > > offers no protection.
> > > 
> > > Yes, the synchronization aspect is bogus indeed.
> > > 
> > > > This patch reorders memory hotremove such that the PCP structures
> > > > relevant to the zone are no longer reachable by the time the structures
> > > > are freed.  With this reordering, no protection is required to prevent
> > > > a use-after-free and the IRQs can be left enabled. zone_pcp_reset is
> > > > renamed to zone_pcp_destroy to make it clear that the per-cpu structures
> > > > are deleted when the function returns.
> > > 
> > > Wouldn't it be much easier to simply not destroy/reset pcp of an empty
> > > zone at all? The whole point of this exercise seems to be described in
> > > 340175b7d14d5. setup_zone_pageset can check for an already allocated pcp
> > > and simply reinitialize it. 
> > 
> > I meant this
> > 
> 
> It might be simplier but if the intention is to free as much memory
> as possible during hot-remove, it seems wasteful to leave the per-cpu
> structures behind if we do not have to.

We do leave the whole pgdat behind. I do not think pagesets really do
matter.

> If a problem with my patch can
> be spotted then I'm happy to go with an alternative fix but there are
> two minor issues with your proposed fix.

I will not insist but this code has proven to bitrot and I just find it
much simpler to drop it altogether rather than conserve it in some form.
Not something I would insist though.

> > diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> > index e6a602e82860..b0fdda77e570 100644
> > --- a/mm/page_alloc.c
> > +++ b/mm/page_alloc.c
> > @@ -6496,7 +6496,13 @@ void __meminit setup_zone_pageset(struct zone *zone)
> >  	struct per_cpu_pageset *p;
> >  	int cpu;
> >  
> > -	zone->pageset = alloc_percpu(struct per_cpu_pageset);
> > +	/*
> > +	 * zone could have gone completely offline during memory hotplug
> > +	 * when the pgdat is left behind for simplicity. On a next onlining
> > +	 * we do not need to reallocate pcp state.
> > +	 */
> > +	if (!zone->pageset)
> > +		zone->pageset = alloc_percpu(struct per_cpu_pageset);
> 
> Should be "if (zone->pageset != &boot_pageset)" ?

Memory hotplug really wants the NULL
check. it doesn't use boot_pageset (if we drop rest to boot_pageset).
But you are right that the boot time initialization first sets
boot_pageset (zone_pcp_init) and initializes real pagesets later
(setup_per_cpu_pageset). But this can be handled at the memory hotplug
layer I believe

diff --git a/mm/memory_hotplug.c b/mm/memory_hotplug.c
index 754026a9164d..1cadfec323fc 100644
--- a/mm/memory_hotplug.c
+++ b/mm/memory_hotplug.c
@@ -883,7 +883,8 @@ int __ref online_pages(unsigned long pfn, unsigned long nr_pages,
 	 */
 	if (!populated_zone(zone)) {
 		need_zonelists_rebuild = 1;
-		setup_zone_pageset(zone);
+		if (!zone->pageset)
+			setup_zone_pageset(zone);
 	}

> >  	for_each_possible_cpu(cpu) {
> >  		p = per_cpu_ptr(zone->pageset, cpu);
> >  		pageset_init(p);
> > @@ -8803,25 +8809,6 @@ void zone_pcp_enable(struct zone *zone)
> >  	mutex_unlock(&pcp_batch_high_lock);
> >  }
> >  
> > -void zone_pcp_reset(struct zone *zone)
> > -{
> > -	unsigned long flags;
> > -	int cpu;
> > -	struct per_cpu_pageset *pset;
> > -
> > -	/* avoid races with drain_pages()  */
> > -	local_irq_save(flags);
> > -	if (zone->pageset != &boot_pageset) {
> > -		for_each_online_cpu(cpu) {
> > -			pset = per_cpu_ptr(zone->pageset, cpu);
> > -			drain_zonestat(zone, pset);
> > -		}
> > -		free_percpu(zone->pageset);
> > -		zone->pageset = &boot_pageset;
> > -	}
> > -	local_irq_restore(flags);
> > -}
> > -
> 
> zone_pcp_reset still needs to exist to drain the remaining vmstats or
> it'll break 5a883813845a ("memory-hotplug: fix zone stat
> mismatch").

Are you sure we are reseting vmstats in the hotremove. I do not see
anything like that. Maybe this was needed at the time. I will double
check.
-- 
Michal Hocko
SUSE Labs