[PATCH 4/9] mm/page_alloc: keep the pagesets of an empty zone

[Vlastimil Babka (SUSE)]

From: Josef Bacik <josef@xxxxxxxxxxxxxx>

When the last memory of a zone goes offline, zone_pcp_reset() frees
zone->per_cpu_pageset and zone->per_cpu_zonestats. Nothing waits for those
who found the zone populated and still use the two pointers. They take
neither the memory hotplug lock nor any other lock which offline_pages()
holds, and some of them can sleep:

  CPU0 vmstat_update()                CPU1 offline_pages()
  refresh_cpu_vm_stats()
    for_each_populated_zone(zone)
      pcp = zone->per_cpu_pageset
      cond_resched()
                                      adjust_present_page_count()
                                      zone_pcp_reset(zone)
                                        free_percpu(zone->per_cpu_pageset)
      decay_pcp_high(zone, pcp)
      __this_cpu_write(pcp->expire, 0)

The same goes for drain_all_pages(NULL), which takes pcp->lock, for
zone_pcp_update(), show_free_areas() and the readers and writers of the
zonestats. If the per-cpu memory has got a new owner in the meantime, it
is corrupted.

The pagesets are only freed because setup_zone_pageset() would otherwise
leak them when the zone gets memory again. Keep them instead and let
setup_zone_pageset() and zone_pcp_init() use them again, like it is done
with pgdat->per_cpu_nodestats. Their lists are empty, and their limits
have been calculated for a zone without memory by the
init_per_zone_wmark_min() of offline_pages(), which is what
setup_zone_pageset() would come up with as well.

Fixes: 340175b7d14d ("mm/hotplug: free zone->pageset when a zone becomes empty")
Cc: <stable@xxxxxxxxxxxxxxx>
Assisted-by: LLM
Signed-off-by: Josef Bacik <josef@xxxxxxxxxxxxxx>
Signed-off-by: David Hildenbrand (Arm) <david@xxxxxxxxxx>
Reviewed-by: Johannes Weiner <hannes@xxxxxxxxxxx>
Signed-off-by: Vlastimil Babka (SUSE) <vbabka@xxxxxxxxxx>
---
 mm/page_alloc.c | 20 ++++++++++++++------
 1 file changed, 14 insertions(+), 6 deletions(-)

diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index 9a4d3352c037..13211b8caada 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -6312,6 +6312,10 @@ void __meminit setup_zone_pageset(struct zone *zone)
 {
 	int cpu;
 
+	/* The zone has been populated before, see zone_pcp_reset(). */
+	if (zone->per_cpu_pageset != &boot_pageset)
+		return;
+
 	/* Size may be 0 on !SMP && !NUMA */
 	if (sizeof(struct per_cpu_zonestat) > 0)
 		zone->per_cpu_zonestats = alloc_percpu(struct per_cpu_zonestat);
@@ -6404,6 +6408,10 @@ void __init setup_per_cpu_pageset(void)
 
 __meminit void zone_pcp_init(struct zone *zone)
 {
+	/* A node which is added again, see zone_pcp_reset(). */
+	if (zone->per_cpu_pageset && zone->per_cpu_pageset != &boot_pageset)
+		return;
+
 	/*
 	 * per cpu subsystem is not up at this point. The following code
 	 * relies on the ability of the linker to provide the
@@ -7684,6 +7692,12 @@ void zone_pcp_enable(struct zone *zone)
 	mutex_unlock(&pcp_batch_high_lock);
 }
 
+/*
+ * The zone has become empty. Its pagesets and zonestats are not freed: nothing
+ * keeps somebody who found the zone populated, such as refresh_cpu_vm_stats()
+ * or drain_all_pages(), from still using them. They are used again when the
+ * zone gets memory again.
+ */
 void zone_pcp_reset(struct zone *zone)
 {
 	int cpu;
@@ -7694,12 +7708,6 @@ void zone_pcp_reset(struct zone *zone)
 			pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
 			drain_zonestat(zone, pzstats);
 		}
-		free_percpu(zone->per_cpu_pageset);
-		zone->per_cpu_pageset = &boot_pageset;
-		if (zone->per_cpu_zonestats != &boot_zonestats) {
-			free_percpu(zone->per_cpu_zonestats);
-			zone->per_cpu_zonestats = &boot_zonestats;
-		}
 	}
 }
 

-- 
2.56.0