per-zone and reclaim enhancements for memory controller: modifies vmscan.c for isolat...
authorKAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Thu, 7 Feb 2008 08:14:37 +0000 (00:14 -0800)
committerLinus Torvalds <torvalds@woody.linux-foundation.org>
Thu, 7 Feb 2008 16:42:22 +0000 (08:42 -0800)
When using memory controller, there are 2 levels of memory reclaim.
 1. zone memory reclaim because of system/zone memory shortage.
 2. memory cgroup memory reclaim because of hitting limit.

These two can be distinguished by sc->mem_cgroup parameter.
(scan_global_lru() macro)

This patch tries to make memory cgroup reclaim routine avoid affecting
system/zone memory reclaim. This patch inserts if (scan_global_lru()) and
hook to memory_cgroup reclaim support functions.

This patch can be a help for isolating system lru activity and group lru
activity and shows what additional functions are necessary.

 * mem_cgroup_calc_mapped_ratio() ... calculate mapped ratio for cgroup.
 * mem_cgroup_reclaim_imbalance() ... calculate active/inactive balance in
                                        cgroup.
 * mem_cgroup_calc_reclaim_active() ... calculate the number of active pages to
                                be scanned in this priority in mem_cgroup.

 * mem_cgroup_calc_reclaim_inactive() ... calculate the number of inactive pages
                                to be scanned in this priority in mem_cgroup.

 * mem_cgroup_all_unreclaimable() .. checks cgroup's page is all unreclaimable
                                     or not.
 * mem_cgroup_get_reclaim_priority() ...
 * mem_cgroup_note_reclaim_priority() ... record reclaim priority (temporal)
 * mem_cgroup_remember_reclaim_priority()
                             .... record reclaim priority as
                                  zone->prev_priority.
                                  This value is used for calc reclaim_mapped.

[akpm@linux-foundation.org: fix unused var warning]
Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: "Eric W. Biederman" <ebiederm@xmission.com>
Cc: Balbir Singh <balbir@linux.vnet.ibm.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Herbert Poetzl <herbert@13thfloor.at>
Cc: Kirill Korotaev <dev@sw.ru>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Paul Menage <menage@google.com>
Cc: Pavel Emelianov <xemul@openvz.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Vaidyanathan Srinivasan <svaidy@linux.vnet.ibm.com>
Cc: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/vmscan.c

index be4dfe87be03eef25206710409d9ae4e44143975..a26dabd62fed40c8ec7832dc4656feaf7d3909f9 100644 (file)
@@ -856,7 +856,8 @@ static unsigned long shrink_inactive_list(unsigned long max_scan,
                __mod_zone_page_state(zone, NR_ACTIVE, -nr_active);
                __mod_zone_page_state(zone, NR_INACTIVE,
                                                -(nr_taken - nr_active));
-               zone->pages_scanned += nr_scan;
+               if (scan_global_lru(sc))
+                       zone->pages_scanned += nr_scan;
                spin_unlock_irq(&zone->lru_lock);
 
                nr_scanned += nr_scan;
@@ -888,8 +889,9 @@ static unsigned long shrink_inactive_list(unsigned long max_scan,
                if (current_is_kswapd()) {
                        __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scan);
                        __count_vm_events(KSWAPD_STEAL, nr_freed);
-               } else
+               } else if (scan_global_lru(sc))
                        __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scan);
+
                __count_zone_vm_events(PGSTEAL, zone, nr_freed);
 
                if (nr_taken == 0)
@@ -942,6 +944,113 @@ static inline int zone_is_near_oom(struct zone *zone)
                                + zone_page_state(zone, NR_INACTIVE))*3;
 }
 
+/*
+ * Determine we should try to reclaim mapped pages.
+ * This is called only when sc->mem_cgroup is NULL.
+ */
+static int calc_reclaim_mapped(struct scan_control *sc, struct zone *zone,
+                               int priority)
+{
+       long mapped_ratio;
+       long distress;
+       long swap_tendency;
+       long imbalance;
+       int reclaim_mapped = 0;
+       int prev_priority;
+
+       if (scan_global_lru(sc) && zone_is_near_oom(zone))
+               return 1;
+       /*
+        * `distress' is a measure of how much trouble we're having
+        * reclaiming pages.  0 -> no problems.  100 -> great trouble.
+        */
+       if (scan_global_lru(sc))
+               prev_priority = zone->prev_priority;
+       else
+               prev_priority = mem_cgroup_get_reclaim_priority(sc->mem_cgroup);
+
+       distress = 100 >> min(prev_priority, priority);
+
+       /*
+        * The point of this algorithm is to decide when to start
+        * reclaiming mapped memory instead of just pagecache.  Work out
+        * how much memory
+        * is mapped.
+        */
+       if (scan_global_lru(sc))
+               mapped_ratio = ((global_page_state(NR_FILE_MAPPED) +
+                               global_page_state(NR_ANON_PAGES)) * 100) /
+                                       vm_total_pages;
+       else
+               mapped_ratio = mem_cgroup_calc_mapped_ratio(sc->mem_cgroup);
+
+       /*
+        * Now decide how much we really want to unmap some pages.  The
+        * mapped ratio is downgraded - just because there's a lot of
+        * mapped memory doesn't necessarily mean that page reclaim
+        * isn't succeeding.
+        *
+        * The distress ratio is important - we don't want to start
+        * going oom.
+        *
+        * A 100% value of vm_swappiness overrides this algorithm
+        * altogether.
+        */
+       swap_tendency = mapped_ratio / 2 + distress + sc->swappiness;
+
+       /*
+        * If there's huge imbalance between active and inactive
+        * (think active 100 times larger than inactive) we should
+        * become more permissive, or the system will take too much
+        * cpu before it start swapping during memory pressure.
+        * Distress is about avoiding early-oom, this is about
+        * making swappiness graceful despite setting it to low
+        * values.
+        *
+        * Avoid div by zero with nr_inactive+1, and max resulting
+        * value is vm_total_pages.
+        */
+       if (scan_global_lru(sc)) {
+               imbalance  = zone_page_state(zone, NR_ACTIVE);
+               imbalance /= zone_page_state(zone, NR_INACTIVE) + 1;
+       } else
+               imbalance = mem_cgroup_reclaim_imbalance(sc->mem_cgroup);
+
+       /*
+        * Reduce the effect of imbalance if swappiness is low,
+        * this means for a swappiness very low, the imbalance
+        * must be much higher than 100 for this logic to make
+        * the difference.
+        *
+        * Max temporary value is vm_total_pages*100.
+        */
+       imbalance *= (vm_swappiness + 1);
+       imbalance /= 100;
+
+       /*
+        * If not much of the ram is mapped, makes the imbalance
+        * less relevant, it's high priority we refill the inactive
+        * list with mapped pages only in presence of high ratio of
+        * mapped pages.
+        *
+        * Max temporary value is vm_total_pages*100.
+        */
+       imbalance *= mapped_ratio;
+       imbalance /= 100;
+
+       /* apply imbalance feedback to swap_tendency */
+       swap_tendency += imbalance;
+
+       /*
+        * Now use this metric to decide whether to start moving mapped
+        * memory onto the inactive list.
+        */
+       if (swap_tendency >= 100)
+               reclaim_mapped = 1;
+
+       return reclaim_mapped;
+}
+
 /*
  * This moves pages from the active list to the inactive list.
  *
@@ -959,6 +1068,8 @@ static inline int zone_is_near_oom(struct zone *zone)
  * The downside is that we have to touch page->_count against each page.
  * But we had to alter page->flags anyway.
  */
+
+
 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
                                struct scan_control *sc, int priority)
 {
@@ -972,100 +1083,21 @@ static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
        struct pagevec pvec;
        int reclaim_mapped = 0;
 
-       if (sc->may_swap) {
-               long mapped_ratio;
-               long distress;
-               long swap_tendency;
-               long imbalance;
-
-               if (zone_is_near_oom(zone))
-                       goto force_reclaim_mapped;
-
-               /*
-                * `distress' is a measure of how much trouble we're having
-                * reclaiming pages.  0 -> no problems.  100 -> great trouble.
-                */
-               distress = 100 >> min(zone->prev_priority, priority);
-
-               /*
-                * The point of this algorithm is to decide when to start
-                * reclaiming mapped memory instead of just pagecache.  Work out
-                * how much memory
-                * is mapped.
-                */
-               mapped_ratio = ((global_page_state(NR_FILE_MAPPED) +
-                               global_page_state(NR_ANON_PAGES)) * 100) /
-                                       vm_total_pages;
-
-               /*
-                * Now decide how much we really want to unmap some pages.  The
-                * mapped ratio is downgraded - just because there's a lot of
-                * mapped memory doesn't necessarily mean that page reclaim
-                * isn't succeeding.
-                *
-                * The distress ratio is important - we don't want to start
-                * going oom.
-                *
-                * A 100% value of vm_swappiness overrides this algorithm
-                * altogether.
-                */
-               swap_tendency = mapped_ratio / 2 + distress + sc->swappiness;
-
-               /*
-                * If there's huge imbalance between active and inactive
-                * (think active 100 times larger than inactive) we should
-                * become more permissive, or the system will take too much
-                * cpu before it start swapping during memory pressure.
-                * Distress is about avoiding early-oom, this is about
-                * making swappiness graceful despite setting it to low
-                * values.
-                *
-                * Avoid div by zero with nr_inactive+1, and max resulting
-                * value is vm_total_pages.
-                */
-               imbalance  = zone_page_state(zone, NR_ACTIVE);
-               imbalance /= zone_page_state(zone, NR_INACTIVE) + 1;
-
-               /*
-                * Reduce the effect of imbalance if swappiness is low,
-                * this means for a swappiness very low, the imbalance
-                * must be much higher than 100 for this logic to make
-                * the difference.
-                *
-                * Max temporary value is vm_total_pages*100.
-                */
-               imbalance *= (vm_swappiness + 1);
-               imbalance /= 100;
-
-               /*
-                * If not much of the ram is mapped, makes the imbalance
-                * less relevant, it's high priority we refill the inactive
-                * list with mapped pages only in presence of high ratio of
-                * mapped pages.
-                *
-                * Max temporary value is vm_total_pages*100.
-                */
-               imbalance *= mapped_ratio;
-               imbalance /= 100;
-
-               /* apply imbalance feedback to swap_tendency */
-               swap_tendency += imbalance;
-
-               /*
-                * Now use this metric to decide whether to start moving mapped
-                * memory onto the inactive list.
-                */
-               if (swap_tendency >= 100)
-force_reclaim_mapped:
-                       reclaim_mapped = 1;
-       }
+       if (sc->may_swap)
+               reclaim_mapped = calc_reclaim_mapped(sc, zone, priority);
 
        lru_add_drain();
        spin_lock_irq(&zone->lru_lock);
        pgmoved = sc->isolate_pages(nr_pages, &l_hold, &pgscanned, sc->order,
                                        ISOLATE_ACTIVE, zone,
                                        sc->mem_cgroup, 1);
-       zone->pages_scanned += pgscanned;
+       /*
+        * zone->pages_scanned is used for detect zone's oom
+        * mem_cgroup remembers nr_scan by itself.
+        */
+       if (scan_global_lru(sc))
+               zone->pages_scanned += pgscanned;
+
        __mod_zone_page_state(zone, NR_ACTIVE, -pgmoved);
        spin_unlock_irq(&zone->lru_lock);
 
@@ -1155,25 +1187,39 @@ static unsigned long shrink_zone(int priority, struct zone *zone,
        unsigned long nr_to_scan;
        unsigned long nr_reclaimed = 0;
 
-       /*
-        * Add one to `nr_to_scan' just to make sure that the kernel will
-        * slowly sift through the active list.
-        */
-       zone->nr_scan_active +=
-               (zone_page_state(zone, NR_ACTIVE) >> priority) + 1;
-       nr_active = zone->nr_scan_active;
-       if (nr_active >= sc->swap_cluster_max)
-               zone->nr_scan_active = 0;
-       else
-               nr_active = 0;
+       if (scan_global_lru(sc)) {
+               /*
+                * Add one to nr_to_scan just to make sure that the kernel
+                * will slowly sift through the active list.
+                */
+               zone->nr_scan_active +=
+                       (zone_page_state(zone, NR_ACTIVE) >> priority) + 1;
+               nr_active = zone->nr_scan_active;
+               zone->nr_scan_inactive +=
+                       (zone_page_state(zone, NR_INACTIVE) >> priority) + 1;
+               nr_inactive = zone->nr_scan_inactive;
+               if (nr_inactive >= sc->swap_cluster_max)
+                       zone->nr_scan_inactive = 0;
+               else
+                       nr_inactive = 0;
+
+               if (nr_active >= sc->swap_cluster_max)
+                       zone->nr_scan_active = 0;
+               else
+                       nr_active = 0;
+       } else {
+               /*
+                * This reclaim occurs not because zone memory shortage but
+                * because memory controller hits its limit.
+                * Then, don't modify zone reclaim related data.
+                */
+               nr_active = mem_cgroup_calc_reclaim_active(sc->mem_cgroup,
+                                       zone, priority);
+
+               nr_inactive = mem_cgroup_calc_reclaim_inactive(sc->mem_cgroup,
+                                       zone, priority);
+       }
 
-       zone->nr_scan_inactive +=
-               (zone_page_state(zone, NR_INACTIVE) >> priority) + 1;
-       nr_inactive = zone->nr_scan_inactive;
-       if (nr_inactive >= sc->swap_cluster_max)
-               zone->nr_scan_inactive = 0;
-       else
-               nr_inactive = 0;
 
        while (nr_active || nr_inactive) {
                if (nr_active) {
@@ -1218,25 +1264,39 @@ static unsigned long shrink_zones(int priority, struct zone **zones,
        unsigned long nr_reclaimed = 0;
        int i;
 
+
        sc->all_unreclaimable = 1;
        for (i = 0; zones[i] != NULL; i++) {
                struct zone *zone = zones[i];
 
                if (!populated_zone(zone))
                        continue;
+               /*
+                * Take care memory controller reclaiming has small influence
+                * to global LRU.
+                */
+               if (scan_global_lru(sc)) {
+                       if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
+                               continue;
+                       note_zone_scanning_priority(zone, priority);
 
-               if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
-                       continue;
-
-               note_zone_scanning_priority(zone, priority);
-
-               if (zone_is_all_unreclaimable(zone) && priority != DEF_PRIORITY)
-                       continue;       /* Let kswapd poll it */
-
-               sc->all_unreclaimable = 0;
+                       if (zone_is_all_unreclaimable(zone) &&
+                                               priority != DEF_PRIORITY)
+                               continue;       /* Let kswapd poll it */
+                       sc->all_unreclaimable = 0;
+               } else {
+                       /*
+                        * Ignore cpuset limitation here. We just want to reduce
+                        * # of used pages by us regardless of memory shortage.
+                        */
+                       sc->all_unreclaimable = 0;
+                       mem_cgroup_note_reclaim_priority(sc->mem_cgroup,
+                                                       priority);
+               }
 
                nr_reclaimed += shrink_zone(priority, zone, sc);
        }
+
        return nr_reclaimed;
 }
  
@@ -1264,16 +1324,21 @@ static unsigned long do_try_to_free_pages(struct zone **zones, gfp_t gfp_mask,
        unsigned long lru_pages = 0;
        int i;
 
-       count_vm_event(ALLOCSTALL);
-
-       for (i = 0; zones[i] != NULL; i++) {
-               struct zone *zone = zones[i];
+       if (scan_global_lru(sc))
+               count_vm_event(ALLOCSTALL);
+       /*
+        * mem_cgroup will not do shrink_slab.
+        */
+       if (scan_global_lru(sc)) {
+               for (i = 0; zones[i] != NULL; i++) {
+                       struct zone *zone = zones[i];
 
-               if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
-                       continue;
+                       if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
+                               continue;
 
-               lru_pages += zone_page_state(zone, NR_ACTIVE)
-                               + zone_page_state(zone, NR_INACTIVE);
+                       lru_pages += zone_page_state(zone, NR_ACTIVE)
+                                       + zone_page_state(zone, NR_INACTIVE);
+               }
        }
 
        for (priority = DEF_PRIORITY; priority >= 0; priority--) {
@@ -1330,14 +1395,19 @@ out:
         */
        if (priority < 0)
                priority = 0;
-       for (i = 0; zones[i] != NULL; i++) {
-               struct zone *zone = zones[i];
 
-               if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
-                       continue;
+       if (scan_global_lru(sc)) {
+               for (i = 0; zones[i] != NULL; i++) {
+                       struct zone *zone = zones[i];
+
+                       if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
+                               continue;
+
+                       zone->prev_priority = priority;
+               }
+       } else
+               mem_cgroup_record_reclaim_priority(sc->mem_cgroup, priority);
 
-               zone->prev_priority = priority;
-       }
        return ret;
 }