Messages in this thread Patch in this message | | | Subject | Re: [PATCH v3 7/7] mm, page_alloc: disable pcplists during memory offline | From | David Hildenbrand <> | Date | Wed, 11 Nov 2020 18:58:35 +0100 |
| |
On 11.11.20 10:28, Vlastimil Babka wrote: > Memory offlining relies on page isolation to guarantee a forward > progress because pages cannot be reused while they are isolated. But the > page isolation itself doesn't prevent from races while freed pages are > stored on pcp lists and thus can be reused. This can be worked around by > repeated draining of pcplists, as done by commit 968318261221 > ("mm/memory_hotplug: drain per-cpu pages again during memory offline"). > > David and Michal would prefer that this race was closed in a way that callers > of page isolation who need stronger guarantees don't need to repeatedly drain. > David suggested disabling pcplists usage completely during page isolation, > instead of repeatedly draining them. > > To achieve this without adding special cases in alloc/free fastpath, we can use > the same approach as boot pagesets - when pcp->high is 0, any pcplist addition > will be immediately flushed. > > The race can thus be closed by setting pcp->high to 0 and draining pcplists > once, before calling start_isolate_page_range(). The draining will serialize > after processes that already disabled interrupts and read the old value of > pcp->high in free_unref_page_commit(), and processes that have not yet disabled > interrupts, will observe pcp->high == 0 when they are rescheduled, and skip > pcplists. This guarantees no stray pages on pcplists in zones where isolation > happens. > > This patch thus adds zone_pcp_disable() and zone_pcp_enable() functions that > page isolation users can call before start_isolate_page_range() and after > unisolating (or offlining) the isolated pages. > > Also, drain_all_pages() is optimized to only execute on cpus where pcplists are > not empty. The check can however race with a free to pcplist that has not yet > increased the pcp->count from 0 to 1. Thus make the drain optionally skip the > racy check and drain on all cpus, and use this option in zone_pcp_disable(). > > As we have to avoid external updates to high and batch while pcplists are > disabled, we take pcp_batch_high_lock in zone_pcp_disable() and release it in > zone_pcp_enable(). This also synchronizes multiple users of > zone_pcp_disable()/enable(). > > Currently the only user of this functionality is offline_pages(). > > Suggested-by: David Hildenbrand <david@redhat.com> > Suggested-by: Michal Hocko <mhocko@suse.com> > Signed-off-by: Vlastimil Babka <vbabka@suse.cz> > Reviewed-by: Oscar Salvador <osalvador@suse.de> > Acked-by: Michal Hocko <mhocko@suse.com> > --- > mm/internal.h | 2 ++ > mm/memory_hotplug.c | 28 ++++++++---------- > mm/page_alloc.c | 69 +++++++++++++++++++++++++++++++++++---------- > mm/page_isolation.c | 6 ++-- > 4 files changed, 71 insertions(+), 34 deletions(-) > > diff --git a/mm/internal.h b/mm/internal.h > index c43ccdddb0f6..2966496680bc 100644 > --- a/mm/internal.h > +++ b/mm/internal.h > @@ -201,6 +201,8 @@ extern int user_min_free_kbytes; > > extern void zone_pcp_update(struct zone *zone); > extern void zone_pcp_reset(struct zone *zone); > +extern void zone_pcp_disable(struct zone *zone); > +extern void zone_pcp_enable(struct zone *zone); > > #if defined CONFIG_COMPACTION || defined CONFIG_CMA > > diff --git a/mm/memory_hotplug.c b/mm/memory_hotplug.c > index 3c494ab0d075..e0a561c550b3 100644 > --- a/mm/memory_hotplug.c > +++ b/mm/memory_hotplug.c > @@ -1491,17 +1491,21 @@ int __ref offline_pages(unsigned long start_pfn, unsigned long nr_pages) > } > node = zone_to_nid(zone); > > + /* > + * Disable pcplists so that page isolation cannot race with freeing > + * in a way that pages from isolated pageblock are left on pcplists. > + */ > + zone_pcp_disable(zone); > + > /* set above range as isolated */ > ret = start_isolate_page_range(start_pfn, end_pfn, > MIGRATE_MOVABLE, > MEMORY_OFFLINE | REPORT_FAILURE); > if (ret) { > reason = "failure to isolate range"; > - goto failed_removal; > + goto failed_removal_pcplists_disabled; > } > > - drain_all_pages(zone); > - > arg.start_pfn = start_pfn; > arg.nr_pages = nr_pages; > node_states_check_changes_offline(nr_pages, zone, &arg); > @@ -1551,20 +1555,8 @@ int __ref offline_pages(unsigned long start_pfn, unsigned long nr_pages) > goto failed_removal_isolated; > } > > - /* > - * per-cpu pages are drained after start_isolate_page_range, but > - * if there are still pages that are not free, make sure that we > - * drain again, because when we isolated range we might have > - * raced with another thread that was adding pages to pcp list. > - * > - * Forward progress should be still guaranteed because > - * pages on the pcp list can only belong to MOVABLE_ZONE > - * because has_unmovable_pages explicitly checks for > - * PageBuddy on freed pages on other zones. > - */ > ret = test_pages_isolated(start_pfn, end_pfn, MEMORY_OFFLINE); > - if (ret) > - drain_all_pages(zone); > +
Why two empty lines before the "} while (ret);" ? (unless I'm confused while looking at this diff)
[...]
> +void __drain_all_pages(struct zone *zone, bool force_all_cpus) > { > int cpu; > > @@ -3076,7 +3069,13 @@ void drain_all_pages(struct zone *zone) > struct zone *z; > bool has_pcps = false; > > - if (zone) { > + if (force_all_cpus) { > + /* > + * The pcp.count check is racy, some callers need a > + * guarantee that no cpu is missed.
Why this comment is helpful, it doesn't tell the whole story. Who exactly/in which situations?
> + */ > + has_pcps = true; > + } else if (zone) { > pcp = per_cpu_ptr(zone->pageset, cpu); > if (pcp->pcp.count) > has_pcps = true; > @@ -3109,6 +3108,18 @@ void drain_all_pages(struct zone *zone) > mutex_unlock(&pcpu_drain_mutex); > } > > +/* > + * Spill all the per-cpu pages from all CPUs back into the buddy allocator. > + * > + * When zone parameter is non-NULL, spill just the single zone's pages. > + * > + * Note that this can be extremely slow as the draining happens in a workqueue. > + */ > +void drain_all_pages(struct zone *zone) > +{ > + __drain_all_pages(zone, false);
It's still somewhat unclear to me why we don't need "force_all_cpus" here. Can you clarify that? (e.g., add a comment somewhere?)
[...]
> +void __zone_set_pageset_high_and_batch(struct zone *zone, unsigned long high, > + unsigned long batch) > +{ > + struct per_cpu_pageset *p; > + int cpu; > + > + for_each_possible_cpu(cpu) { > + p = per_cpu_ptr(zone->pageset, cpu); > + pageset_update(&p->pcp, high, batch); > + } > +} > + > /* > * Calculate and set new high and batch values for all per-cpu pagesets of a > * zone, based on the zone's size and the percpu_pagelist_fraction sysctl. > @@ -6315,8 +6338,6 @@ static void pageset_init(struct per_cpu_pageset *p) > static void zone_set_pageset_high_and_batch(struct zone *zone) > { > unsigned long new_high, new_batch; > - struct per_cpu_pageset *p; > - int cpu; > > if (percpu_pagelist_fraction) { > new_high = zone_managed_pages(zone) / percpu_pagelist_fraction; > @@ -6336,10 +6357,7 @@ static void zone_set_pageset_high_and_batch(struct zone *zone) > zone->pageset_high = new_high; > zone->pageset_batch = new_batch; > > - for_each_possible_cpu(cpu) { > - p = per_cpu_ptr(zone->pageset, cpu); > - pageset_update(&p->pcp, new_high, new_batch); > - } > + __zone_set_pageset_high_and_batch(zone, new_high, new_batch); > }
These two hunks look like an unrelated cleanup, or am I missing something?
Thanks for looking into this!
-- Thanks,
David / dhildenb
| |