]> git.ipfire.org Git - thirdparty/kernel/stable.git/commitdiff
mm: memcontrol: don't batch updates of local VM stats and events
authorJohannes Weiner <hannes@cmpxchg.org>
Thu, 13 Jun 2019 22:55:46 +0000 (15:55 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Fri, 14 Jun 2019 03:34:56 +0000 (17:34 -1000)
The kernel test robot noticed a 26% will-it-scale pagefault regression
from commit 42a300353577 ("mm: memcontrol: fix recursive statistics
correctness & scalabilty").  This appears to be caused by bouncing the
additional cachelines from the new hierarchical statistics counters.

We can fix this by getting rid of the batched local counters instead.

Originally, there were *only* group-local counters, and they were fully
maintained per cpu.  A reader of a stats file high up in the cgroup tree
would have to walk the entire subtree and collect each level's per-cpu
counters to get the recursive view.  This was prohibitively expensive,
and so we switched to per-cpu batched updates of the local counters
during a983b5ebee57 ("mm: memcontrol: fix excessive complexity in
memory.stat reporting"), reducing the complexity from nr_subgroups *
nr_cpus to nr_subgroups.

With growing machines and cgroup trees, the tree walk itself became too
expensive for monitoring top-level groups, and this is when the culprit
patch added hierarchy counters on each cgroup level.  When the per-cpu
batch size would be reached, both the local and the hierarchy counters
would get batch-updated from the per-cpu delta simultaneously.

This makes local and hierarchical counter reads blazingly fast, but it
unfortunately makes the write-side too cache line intense.

Since local counter reads were never a problem - we only centralized
them to accelerate the hierarchy walk - and use of the local counters
are becoming rarer due to replacement with hierarchical views (ongoing
rework in the page reclaim and workingset code), we can make those local
counters unbatched per-cpu counters again.

The scheme will then be as such:

   when a memcg statistic changes, the writer will:
   - update the local counter (per-cpu)
   - update the batch counter (per-cpu). If the batch is full:
   - spill the batch into the group's atomic_t
   - spill the batch into all ancestors' atomic_ts
   - empty out the batch counter (per-cpu)

   when a local memcg counter is read, the reader will:
   - collect the local counter from all cpus

   when a hiearchy memcg counter is read, the reader will:
   - read the atomic_t

We might be able to simplify this further and make the recursive
counters unbatched per-cpu counters as well (batch upward propagation,
but leave per-cpu collection to the readers), but that will require a
more in-depth analysis and testing of all the callsites.  Deal with the
immediate regression for now.

Link: http://lkml.kernel.org/r/20190521151647.GB2870@cmpxchg.org
Fixes: 42a300353577 ("mm: memcontrol: fix recursive statistics correctness & scalabilty")
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reported-by: kernel test robot <rong.a.chen@intel.com>
Tested-by: kernel test robot <rong.a.chen@intel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Roman Gushchin <guro@fb.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
include/linux/memcontrol.h
mm/memcontrol.c

index edf9e8f32d7025fd53412f414f4d136fb6867d1f..1dcb763bb610a85295893645fe285f2659532f1a 100644 (file)
@@ -117,9 +117,12 @@ struct memcg_shrinker_map {
 struct mem_cgroup_per_node {
        struct lruvec           lruvec;
 
+       /* Legacy local VM stats */
+       struct lruvec_stat __percpu *lruvec_stat_local;
+
+       /* Subtree VM stats (batched updates) */
        struct lruvec_stat __percpu *lruvec_stat_cpu;
        atomic_long_t           lruvec_stat[NR_VM_NODE_STAT_ITEMS];
-       atomic_long_t           lruvec_stat_local[NR_VM_NODE_STAT_ITEMS];
 
        unsigned long           lru_zone_size[MAX_NR_ZONES][NR_LRU_LISTS];
 
@@ -265,17 +268,18 @@ struct mem_cgroup {
        atomic_t                moving_account;
        struct task_struct      *move_lock_task;
 
-       /* memory.stat */
+       /* Legacy local VM stats and events */
+       struct memcg_vmstats_percpu __percpu *vmstats_local;
+
+       /* Subtree VM stats and events (batched updates) */
        struct memcg_vmstats_percpu __percpu *vmstats_percpu;
 
        MEMCG_PADDING(_pad2_);
 
        atomic_long_t           vmstats[MEMCG_NR_STAT];
-       atomic_long_t           vmstats_local[MEMCG_NR_STAT];
-
        atomic_long_t           vmevents[NR_VM_EVENT_ITEMS];
-       atomic_long_t           vmevents_local[NR_VM_EVENT_ITEMS];
 
+       /* memory.events */
        atomic_long_t           memory_events[MEMCG_NR_MEMORY_EVENTS];
 
        unsigned long           socket_pressure;
@@ -567,7 +571,11 @@ static inline unsigned long memcg_page_state(struct mem_cgroup *memcg, int idx)
 static inline unsigned long memcg_page_state_local(struct mem_cgroup *memcg,
                                                   int idx)
 {
-       long x = atomic_long_read(&memcg->vmstats_local[idx]);
+       long x = 0;
+       int cpu;
+
+       for_each_possible_cpu(cpu)
+               x += per_cpu(memcg->vmstats_local->stat[idx], cpu);
 #ifdef CONFIG_SMP
        if (x < 0)
                x = 0;
@@ -641,13 +649,15 @@ static inline unsigned long lruvec_page_state_local(struct lruvec *lruvec,
                                                    enum node_stat_item idx)
 {
        struct mem_cgroup_per_node *pn;
-       long x;
+       long x = 0;
+       int cpu;
 
        if (mem_cgroup_disabled())
                return node_page_state(lruvec_pgdat(lruvec), idx);
 
        pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
-       x = atomic_long_read(&pn->lruvec_stat_local[idx]);
+       for_each_possible_cpu(cpu)
+               x += per_cpu(pn->lruvec_stat_local->count[idx], cpu);
 #ifdef CONFIG_SMP
        if (x < 0)
                x = 0;
index ca0bc6e6be1316602def0ac42bc4ffddef14cf08..ba9138a4a1de37f744eaf488fc0930b82f60b3e1 100644 (file)
@@ -691,11 +691,12 @@ void __mod_memcg_state(struct mem_cgroup *memcg, int idx, int val)
        if (mem_cgroup_disabled())
                return;
 
+       __this_cpu_add(memcg->vmstats_local->stat[idx], val);
+
        x = val + __this_cpu_read(memcg->vmstats_percpu->stat[idx]);
        if (unlikely(abs(x) > MEMCG_CHARGE_BATCH)) {
                struct mem_cgroup *mi;
 
-               atomic_long_add(x, &memcg->vmstats_local[idx]);
                for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
                        atomic_long_add(x, &mi->vmstats[idx]);
                x = 0;
@@ -745,11 +746,12 @@ void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
        __mod_memcg_state(memcg, idx, val);
 
        /* Update lruvec */
+       __this_cpu_add(pn->lruvec_stat_local->count[idx], val);
+
        x = val + __this_cpu_read(pn->lruvec_stat_cpu->count[idx]);
        if (unlikely(abs(x) > MEMCG_CHARGE_BATCH)) {
                struct mem_cgroup_per_node *pi;
 
-               atomic_long_add(x, &pn->lruvec_stat_local[idx]);
                for (pi = pn; pi; pi = parent_nodeinfo(pi, pgdat->node_id))
                        atomic_long_add(x, &pi->lruvec_stat[idx]);
                x = 0;
@@ -771,11 +773,12 @@ void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
        if (mem_cgroup_disabled())
                return;
 
+       __this_cpu_add(memcg->vmstats_local->events[idx], count);
+
        x = count + __this_cpu_read(memcg->vmstats_percpu->events[idx]);
        if (unlikely(x > MEMCG_CHARGE_BATCH)) {
                struct mem_cgroup *mi;
 
-               atomic_long_add(x, &memcg->vmevents_local[idx]);
                for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
                        atomic_long_add(x, &mi->vmevents[idx]);
                x = 0;
@@ -790,7 +793,12 @@ static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
 
 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
 {
-       return atomic_long_read(&memcg->vmevents_local[event]);
+       long x = 0;
+       int cpu;
+
+       for_each_possible_cpu(cpu)
+               x += per_cpu(memcg->vmstats_local->events[event], cpu);
+       return x;
 }
 
 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
@@ -2191,11 +2199,9 @@ static int memcg_hotplug_cpu_dead(unsigned int cpu)
                        long x;
 
                        x = this_cpu_xchg(memcg->vmstats_percpu->stat[i], 0);
-                       if (x) {
-                               atomic_long_add(x, &memcg->vmstats_local[i]);
+                       if (x)
                                for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
                                        atomic_long_add(x, &memcg->vmstats[i]);
-                       }
 
                        if (i >= NR_VM_NODE_STAT_ITEMS)
                                continue;
@@ -2205,12 +2211,10 @@ static int memcg_hotplug_cpu_dead(unsigned int cpu)
 
                                pn = mem_cgroup_nodeinfo(memcg, nid);
                                x = this_cpu_xchg(pn->lruvec_stat_cpu->count[i], 0);
-                               if (x) {
-                                       atomic_long_add(x, &pn->lruvec_stat_local[i]);
+                               if (x)
                                        do {
                                                atomic_long_add(x, &pn->lruvec_stat[i]);
                                        } while ((pn = parent_nodeinfo(pn, nid)));
-                               }
                        }
                }
 
@@ -2218,11 +2222,9 @@ static int memcg_hotplug_cpu_dead(unsigned int cpu)
                        long x;
 
                        x = this_cpu_xchg(memcg->vmstats_percpu->events[i], 0);
-                       if (x) {
-                               atomic_long_add(x, &memcg->vmevents_local[i]);
+                       if (x)
                                for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
                                        atomic_long_add(x, &memcg->vmevents[i]);
-                       }
                }
        }
 
@@ -4483,8 +4485,15 @@ static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
        if (!pn)
                return 1;
 
+       pn->lruvec_stat_local = alloc_percpu(struct lruvec_stat);
+       if (!pn->lruvec_stat_local) {
+               kfree(pn);
+               return 1;
+       }
+
        pn->lruvec_stat_cpu = alloc_percpu(struct lruvec_stat);
        if (!pn->lruvec_stat_cpu) {
+               free_percpu(pn->lruvec_stat_local);
                kfree(pn);
                return 1;
        }
@@ -4506,6 +4515,7 @@ static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
                return;
 
        free_percpu(pn->lruvec_stat_cpu);
+       free_percpu(pn->lruvec_stat_local);
        kfree(pn);
 }
 
@@ -4516,6 +4526,7 @@ static void __mem_cgroup_free(struct mem_cgroup *memcg)
        for_each_node(node)
                free_mem_cgroup_per_node_info(memcg, node);
        free_percpu(memcg->vmstats_percpu);
+       free_percpu(memcg->vmstats_local);
        kfree(memcg);
 }
 
@@ -4544,6 +4555,10 @@ static struct mem_cgroup *mem_cgroup_alloc(void)
        if (memcg->id.id < 0)
                goto fail;
 
+       memcg->vmstats_local = alloc_percpu(struct memcg_vmstats_percpu);
+       if (!memcg->vmstats_local)
+               goto fail;
+
        memcg->vmstats_percpu = alloc_percpu(struct memcg_vmstats_percpu);
        if (!memcg->vmstats_percpu)
                goto fail;