Commit 8fce4d8e authored by Christoph Lameter's avatar Christoph Lameter Committed by Linus Torvalds
Browse files

[PATCH] slab: Node rotor for freeing alien caches and remote per cpu pages.

The cache reaper currently tries to free all alien caches and all remote
per cpu pages in each pass of cache_reap.  For a machines with large number
of nodes (such as Altix) this may lead to sporadic delays of around ~10ms.
Interrupts are disabled while reclaiming creating unacceptable delays.

This patch changes that behavior by adding a per cpu reap_node variable.
Instead of attempting to free all caches, we free only one alien cache and
the per cpu pages from one remote node.  That reduces the time spend in
cache_reap.  However, doing so will lengthen the time it takes to
completely drain all remote per cpu pagesets and all alien caches.  The
time needed will grow with the number of nodes in the system.  All caches
are drained when they overflow their respective capacity.  So the drawback
here is only that a bit of memory may be wasted for awhile longer.


1. Rename drain_remote_pages to drain_node_pages to allow the specification
   of the node to drain of pcp pages.

2. Add additional functions init_reap_node, next_reap_node for NUMA
   that manage a per cpu reap_node counter.

3. Add a reap_alien function that reaps only from the current reap_node.

For us this seems to be a critical issue.  Holdoffs of an average of ~7ms
cause some HPC benchmarks to slow down significantly.  F.e.  NAS parallel
slows down dramatically.  NAS parallel has a 12-16 seconds runtime w/o rotor
compared to 5.8 secs with the rotor patches.  It gets down to 5.05 secs with
the additional interrupt holdoff reductions.

Signed-off-by: default avatarChristoph Lameter <>
Signed-off-by: default avatarAndrew Morton <>
Signed-off-by: default avatarLinus Torvalds <>
parent 7b61fcda
......@@ -157,9 +157,9 @@ extern void FASTCALL(free_cold_page(struct page *page));
void page_alloc_init(void);
void drain_remote_pages(void);
void drain_node_pages(int node);
static inline void drain_remote_pages(void) { };
static inline void drain_node_pages(int node) { };
#endif /* __LINUX_GFP_H */
......@@ -590,21 +590,20 @@ static int rmqueue_bulk(struct zone *zone, unsigned int order,
/* Called from the slab reaper to drain remote pagesets */
void drain_remote_pages(void)
* Called from the slab reaper to drain pagesets on a particular node that
* belong to the currently executing processor.
void drain_node_pages(int nodeid)
struct zone *zone;
int i;
int i, z;
unsigned long flags;
for_each_zone(zone) {
for (z = 0; z < MAX_NR_ZONES; z++) {
struct zone *zone = NODE_DATA(nodeid)->node_zones + z;
struct per_cpu_pageset *pset;
/* Do not drain local pagesets */
if (zone->zone_pgdat->node_id == numa_node_id())
pset = zone_pcp(zone, smp_processor_id());
for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
struct per_cpu_pages *pcp;
......@@ -789,6 +789,47 @@ static void __slab_error(const char *function, struct kmem_cache *cachep, char *
* Special reaping functions for NUMA systems called from cache_reap().
* These take care of doing round robin flushing of alien caches (containing
* objects freed on different nodes from which they were allocated) and the
* flushing of remote pcps by calling drain_node_pages.
static DEFINE_PER_CPU(unsigned long, reap_node);
static void init_reap_node(int cpu)
int node;
node = next_node(cpu_to_node(cpu), node_online_map);
if (node == MAX_NUMNODES)
node = 0;
__get_cpu_var(reap_node) = node;
static void next_reap_node(void)
int node = __get_cpu_var(reap_node);
* Also drain per cpu pages on remote zones
if (node != numa_node_id())
node = next_node(node, node_online_map);
if (unlikely(node >= MAX_NUMNODES))
node = first_node(node_online_map);
__get_cpu_var(reap_node) = node;
#define init_reap_node(cpu) do { } while (0)
#define next_reap_node(void) do { } while (0)
* Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
* via the workqueue/eventd.
......@@ -806,6 +847,7 @@ static void __devinit start_cpu_timer(int cpu)
* at that time.
if (keventd_up() && reap_work->func == NULL) {
INIT_WORK(reap_work, cache_reap, NULL);
schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu);
......@@ -884,6 +926,23 @@ static void __drain_alien_cache(struct kmem_cache *cachep,
* Called from cache_reap() to regularly drain alien caches round robin.
static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
int node = __get_cpu_var(reap_node);
if (l3->alien) {
struct array_cache *ac = l3->alien[node];
if (ac && ac->avail) {
__drain_alien_cache(cachep, ac, node);
static void drain_alien_cache(struct kmem_cache *cachep, struct array_cache **alien)
int i = 0;
......@@ -902,6 +961,7 @@ static void drain_alien_cache(struct kmem_cache *cachep, struct array_cache **al
#define drain_alien_cache(cachep, alien) do { } while (0)
#define reap_alien(cachep, l3) do { } while (0)
static inline struct array_cache **alloc_alien_cache(int node, int limit)
......@@ -3497,8 +3557,7 @@ static void cache_reap(void *unused)
l3 = searchp->nodelists[numa_node_id()];
if (l3->alien)
drain_alien_cache(searchp, l3->alien);
reap_alien(searchp, l3);
drain_array_locked(searchp, cpu_cache_get(searchp), 0,
......@@ -3548,7 +3607,7 @@ static void cache_reap(void *unused)
/* Setup the next iteration */
schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC);
Supports Markdown
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment