When offline a node in try_offline_node, pgdat is not released. So that
pgdat could be reused in hotadd_new_pgdat. While we re-allocate
pgdat->per_cpu_nodestats if this pgdat is reused.
This patch prevents the memory leak by just allocate per_cpu_nodestats
when it is a new pgdat.
NOTE: This is not tested since I didn't manage to create a case to
offline a whole node. If my analysis is not correct, please let me know.
Signed-off-by: Wei Yang <[email protected]>
---
mm/memory_hotplug.c | 10 +++++++++-
1 file changed, 9 insertions(+), 1 deletion(-)
diff --git a/mm/memory_hotplug.c b/mm/memory_hotplug.c
index c73f09913165..efaf9e6f580a 100644
--- a/mm/memory_hotplug.c
+++ b/mm/memory_hotplug.c
@@ -933,8 +933,11 @@ static pg_data_t __ref *hotadd_new_pgdat(int nid, u64 start)
if (!pgdat)
return NULL;
+ pgdat->per_cpu_nodestats =
+ alloc_percpu(struct per_cpu_nodestat);
arch_refresh_nodedata(nid, pgdat);
} else {
+ int cpu;
/*
* Reset the nr_zones, order and classzone_idx before reuse.
* Note that kswapd will init kswapd_classzone_idx properly
@@ -943,6 +946,12 @@ static pg_data_t __ref *hotadd_new_pgdat(int nid, u64 start)
pgdat->nr_zones = 0;
pgdat->kswapd_order = 0;
pgdat->kswapd_classzone_idx = 0;
+ for_each_online_cpu(cpu) {
+ struct per_cpu_nodestat *p;
+
+ p = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu);
+ memset(p, 0, sizeof(*p));
+ }
}
/* we can use NODE_DATA(nid) from here */
@@ -952,7 +961,6 @@ static pg_data_t __ref *hotadd_new_pgdat(int nid, u64 start)
/* init node's zones as empty zones, we don't have any present pages.*/
free_area_init_core_hotplug(nid);
- pgdat->per_cpu_nodestats = alloc_percpu(struct per_cpu_nodestat);
/*
* The node we allocated has no zone fallback lists. For avoiding
--
2.17.1
On Tue 13-08-19 10:06:08, Wei Yang wrote:
> When offline a node in try_offline_node, pgdat is not released. So that
> pgdat could be reused in hotadd_new_pgdat. While we re-allocate
> pgdat->per_cpu_nodestats if this pgdat is reused.
>
> This patch prevents the memory leak by just allocate per_cpu_nodestats
> when it is a new pgdat.
Yes this makes sense! I was slightly confused why we haven't initialized
the allocated pcp area because __alloc_percpu does GFP_KERNEL without
__GFP_ZERO but then I've just found out that the zeroying is done
regardless. A bit unexpected...
> NOTE: This is not tested since I didn't manage to create a case to
> offline a whole node. If my analysis is not correct, please let me know.
>
> Signed-off-by: Wei Yang <[email protected]>
Acked-by: Michal Hocko <[email protected]>
Thanks!
> ---
> mm/memory_hotplug.c | 10 +++++++++-
> 1 file changed, 9 insertions(+), 1 deletion(-)
>
> diff --git a/mm/memory_hotplug.c b/mm/memory_hotplug.c
> index c73f09913165..efaf9e6f580a 100644
> --- a/mm/memory_hotplug.c
> +++ b/mm/memory_hotplug.c
> @@ -933,8 +933,11 @@ static pg_data_t __ref *hotadd_new_pgdat(int nid, u64 start)
> if (!pgdat)
> return NULL;
>
> + pgdat->per_cpu_nodestats =
> + alloc_percpu(struct per_cpu_nodestat);
> arch_refresh_nodedata(nid, pgdat);
> } else {
> + int cpu;
> /*
> * Reset the nr_zones, order and classzone_idx before reuse.
> * Note that kswapd will init kswapd_classzone_idx properly
> @@ -943,6 +946,12 @@ static pg_data_t __ref *hotadd_new_pgdat(int nid, u64 start)
> pgdat->nr_zones = 0;
> pgdat->kswapd_order = 0;
> pgdat->kswapd_classzone_idx = 0;
> + for_each_online_cpu(cpu) {
> + struct per_cpu_nodestat *p;
> +
> + p = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu);
> + memset(p, 0, sizeof(*p));
> + }
> }
>
> /* we can use NODE_DATA(nid) from here */
> @@ -952,7 +961,6 @@ static pg_data_t __ref *hotadd_new_pgdat(int nid, u64 start)
>
> /* init node's zones as empty zones, we don't have any present pages.*/
> free_area_init_core_hotplug(nid);
> - pgdat->per_cpu_nodestats = alloc_percpu(struct per_cpu_nodestat);
>
> /*
> * The node we allocated has no zone fallback lists. For avoiding
> --
> 2.17.1
>
--
Michal Hocko
SUSE Labs
On Tue, Aug 13, 2019 at 09:57:07AM +0200, Michal Hocko wrote:
>On Tue 13-08-19 10:06:08, Wei Yang wrote:
>> When offline a node in try_offline_node, pgdat is not released. So that
>> pgdat could be reused in hotadd_new_pgdat. While we re-allocate
>> pgdat->per_cpu_nodestats if this pgdat is reused.
>>
>> This patch prevents the memory leak by just allocate per_cpu_nodestats
>> when it is a new pgdat.
>
>Yes this makes sense! I was slightly confused why we haven't initialized
>the allocated pcp area because __alloc_percpu does GFP_KERNEL without
>__GFP_ZERO but then I've just found out that the zeroying is done
>regardless. A bit unexpected...
>
>> NOTE: This is not tested since I didn't manage to create a case to
>> offline a whole node. If my analysis is not correct, please let me know.
>>
>> Signed-off-by: Wei Yang <[email protected]>
>
>Acked-by: Michal Hocko <[email protected]>
>
>Thanks!
>
Thanks :-)
>> ---
>> mm/memory_hotplug.c | 10 +++++++++-
>> 1 file changed, 9 insertions(+), 1 deletion(-)
>>
>> diff --git a/mm/memory_hotplug.c b/mm/memory_hotplug.c
>> index c73f09913165..efaf9e6f580a 100644
>> --- a/mm/memory_hotplug.c
>> +++ b/mm/memory_hotplug.c
>> @@ -933,8 +933,11 @@ static pg_data_t __ref *hotadd_new_pgdat(int nid, u64 start)
>> if (!pgdat)
>> return NULL;
>>
>> + pgdat->per_cpu_nodestats =
>> + alloc_percpu(struct per_cpu_nodestat);
>> arch_refresh_nodedata(nid, pgdat);
>> } else {
>> + int cpu;
>> /*
>> * Reset the nr_zones, order and classzone_idx before reuse.
>> * Note that kswapd will init kswapd_classzone_idx properly
>> @@ -943,6 +946,12 @@ static pg_data_t __ref *hotadd_new_pgdat(int nid, u64 start)
>> pgdat->nr_zones = 0;
>> pgdat->kswapd_order = 0;
>> pgdat->kswapd_classzone_idx = 0;
>> + for_each_online_cpu(cpu) {
>> + struct per_cpu_nodestat *p;
>> +
>> + p = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu);
>> + memset(p, 0, sizeof(*p));
>> + }
>> }
>>
>> /* we can use NODE_DATA(nid) from here */
>> @@ -952,7 +961,6 @@ static pg_data_t __ref *hotadd_new_pgdat(int nid, u64 start)
>>
>> /* init node's zones as empty zones, we don't have any present pages.*/
>> free_area_init_core_hotplug(nid);
>> - pgdat->per_cpu_nodestats = alloc_percpu(struct per_cpu_nodestat);
>>
>> /*
>> * The node we allocated has no zone fallback lists. For avoiding
>> --
>> 2.17.1
>>
>
>--
>Michal Hocko
>SUSE Labs
--
Wei Yang
Help you, Help me