mm, slab: faster active and free stats
authorGreg Thelen <gthelen@google.com>
Tue, 13 Dec 2016 00:41:41 +0000 (16:41 -0800)
committerLinus Torvalds <torvalds@linux-foundation.org>
Tue, 13 Dec 2016 02:55:06 +0000 (18:55 -0800)
Reading /proc/slabinfo or monitoring slabtop(1) can become very
expensive if there are many slab caches and if there are very lengthy
per-node partial and/or free lists.

Commit 07a63c41fa1f ("mm/slab: improve performance of gathering slabinfo
stats") addressed the per-node full lists which showed a significant
improvement when no objects were freed.  This patch has the same
motivation and optimizes the remainder of the usecases where there are
very lengthy partial and free lists.

This patch maintains per-node active_slabs (full and partial) and
free_slabs rather than iterating the lists at runtime when reading
/proc/slabinfo.

When allocating 100GB of slab from a test cache where every slab page is
on the partial list, reading /proc/slabinfo (includes all other slab
caches on the system) takes ~247ms on average with 48 samples.

As a result of this patch, the same read takes ~0.856ms on average.

[rientjes@google.com: changelog]
Link: http://lkml.kernel.org/r/alpine.DEB.2.10.1611081505240.13403@chino.kir.corp.google.com
Signed-off-by: Greg Thelen <gthelen@google.com>
Signed-off-by: David Rientjes <rientjes@google.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/slab.c
mm/slab.h

index 7ea765cd7e936aadee2fb847a2615ec370c4f893..e06da6ceaf7348934d3ddd7fdebbbec93d0f60c5 100644 (file)
--- a/mm/slab.c
+++ b/mm/slab.c
@@ -227,13 +227,14 @@ static void kmem_cache_node_init(struct kmem_cache_node *parent)
        INIT_LIST_HEAD(&parent->slabs_full);
        INIT_LIST_HEAD(&parent->slabs_partial);
        INIT_LIST_HEAD(&parent->slabs_free);
+       parent->active_slabs = 0;
+       parent->free_slabs = 0;
        parent->shared = NULL;
        parent->alien = NULL;
        parent->colour_next = 0;
        spin_lock_init(&parent->list_lock);
        parent->free_objects = 0;
        parent->free_touched = 0;
-       parent->num_slabs = 0;
 }
 
 #define MAKE_LIST(cachep, listp, slab, nodeid)                         \
@@ -1366,7 +1367,6 @@ slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
 {
 #if DEBUG
        struct kmem_cache_node *n;
-       struct page *page;
        unsigned long flags;
        int node;
        static DEFINE_RATELIMIT_STATE(slab_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
@@ -1381,32 +1381,20 @@ slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
                cachep->name, cachep->size, cachep->gfporder);
 
        for_each_kmem_cache_node(cachep, node, n) {
-               unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
-               unsigned long active_slabs = 0, num_slabs = 0;
-               unsigned long num_slabs_partial = 0, num_slabs_free = 0;
-               unsigned long num_slabs_full;
+               unsigned long active_objs = 0, free_objs = 0;
+               unsigned long active_slabs, num_slabs;
 
                spin_lock_irqsave(&n->list_lock, flags);
-               num_slabs = n->num_slabs;
-               list_for_each_entry(page, &n->slabs_partial, lru) {
-                       active_objs += page->active;
-                       num_slabs_partial++;
-               }
-               list_for_each_entry(page, &n->slabs_free, lru)
-                       num_slabs_free++;
+               active_slabs = n->active_slabs;
+               num_slabs = active_slabs + n->free_slabs;
 
-               free_objects += n->free_objects;
+               active_objs += (num_slabs * cachep->num) - n->free_objects;
+               free_objs += n->free_objects;
                spin_unlock_irqrestore(&n->list_lock, flags);
 
-               num_objs = num_slabs * cachep->num;
-               active_slabs = num_slabs - num_slabs_free;
-               num_slabs_full = num_slabs -
-                       (num_slabs_partial + num_slabs_free);
-               active_objs += (num_slabs_full * cachep->num);
-
                pr_warn("  node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
-                       node, active_slabs, num_slabs, active_objs, num_objs,
-                       free_objects);
+                       node, active_slabs, num_slabs, active_objs,
+                       num_slabs * cachep->num, free_objs);
        }
 #endif
 }
@@ -2318,7 +2306,7 @@ static int drain_freelist(struct kmem_cache *cache,
 
                page = list_entry(p, struct page, lru);
                list_del(&page->lru);
-               n->num_slabs--;
+               n->free_slabs--;
                /*
                 * Safe to drop the lock. The slab is no longer linked
                 * to the cache.
@@ -2753,12 +2741,14 @@ static void cache_grow_end(struct kmem_cache *cachep, struct page *page)
        n = get_node(cachep, page_to_nid(page));
 
        spin_lock(&n->list_lock);
-       if (!page->active)
+       if (!page->active) {
                list_add_tail(&page->lru, &(n->slabs_free));
-       else
+               n->free_slabs++;
+       } else {
                fixup_slab_list(cachep, n, page, &list);
+               n->active_slabs++;
+       }
 
-       n->num_slabs++;
        STATS_INC_GROWN(cachep);
        n->free_objects += cachep->num - page->active;
        spin_unlock(&n->list_lock);
@@ -2884,7 +2874,7 @@ static inline void fixup_slab_list(struct kmem_cache *cachep,
 
 /* Try to find non-pfmemalloc slab if needed */
 static noinline struct page *get_valid_first_slab(struct kmem_cache_node *n,
-                                       struct page *page, bool pfmemalloc)
+                       struct page *page, bool *page_is_free, bool pfmemalloc)
 {
        if (!page)
                return NULL;
@@ -2903,9 +2893,11 @@ static noinline struct page *get_valid_first_slab(struct kmem_cache_node *n,
 
        /* Move pfmemalloc slab to the end of list to speed up next search */
        list_del(&page->lru);
-       if (!page->active)
+       if (*page_is_free) {
+               WARN_ON(page->active);
                list_add_tail(&page->lru, &n->slabs_free);
-       else
+               *page_is_free = false;
+       } else
                list_add_tail(&page->lru, &n->slabs_partial);
 
        list_for_each_entry(page, &n->slabs_partial, lru) {
@@ -2913,9 +2905,12 @@ static noinline struct page *get_valid_first_slab(struct kmem_cache_node *n,
                        return page;
        }
 
+       n->free_touched = 1;
        list_for_each_entry(page, &n->slabs_free, lru) {
-               if (!PageSlabPfmemalloc(page))
+               if (!PageSlabPfmemalloc(page)) {
+                       *page_is_free = true;
                        return page;
+               }
        }
 
        return NULL;
@@ -2924,17 +2919,26 @@ static noinline struct page *get_valid_first_slab(struct kmem_cache_node *n,
 static struct page *get_first_slab(struct kmem_cache_node *n, bool pfmemalloc)
 {
        struct page *page;
+       bool page_is_free = false;
 
+       assert_spin_locked(&n->list_lock);
        page = list_first_entry_or_null(&n->slabs_partial,
                        struct page, lru);
        if (!page) {
                n->free_touched = 1;
                page = list_first_entry_or_null(&n->slabs_free,
                                struct page, lru);
+               if (page)
+                       page_is_free = true;
        }
 
        if (sk_memalloc_socks())
-               return get_valid_first_slab(n, page, pfmemalloc);
+               page = get_valid_first_slab(n, page, &page_is_free, pfmemalloc);
+
+       if (page && page_is_free) {
+               n->active_slabs++;
+               n->free_slabs--;
+       }
 
        return page;
 }
@@ -3434,9 +3438,11 @@ static void free_block(struct kmem_cache *cachep, void **objpp,
                STATS_DEC_ACTIVE(cachep);
 
                /* fixup slab chains */
-               if (page->active == 0)
+               if (page->active == 0) {
                        list_add(&page->lru, &n->slabs_free);
-               else {
+                       n->free_slabs++;
+                       n->active_slabs--;
+               } else {
                        /* Unconditionally move a slab to the end of the
                         * partial list on free - maximum time for the
                         * other objects to be freed, too.
@@ -3450,7 +3456,7 @@ static void free_block(struct kmem_cache *cachep, void **objpp,
 
                page = list_last_entry(&n->slabs_free, struct page, lru);
                list_move(&page->lru, list);
-               n->num_slabs--;
+               n->free_slabs--;
        }
 }
 
@@ -4102,43 +4108,21 @@ out:
 #ifdef CONFIG_SLABINFO
 void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
 {
-       struct page *page;
-       unsigned long active_objs;
-       unsigned long num_objs;
-       unsigned long active_slabs = 0;
-       unsigned long num_slabs, free_objects = 0, shared_avail = 0;
-       unsigned long num_slabs_partial = 0, num_slabs_free = 0;
-       unsigned long num_slabs_full = 0;
-       const char *name;
-       char *error = NULL;
+       unsigned long active_objs, num_objs, active_slabs;
+       unsigned long num_slabs = 0, free_objs = 0, shared_avail = 0;
+       unsigned long num_slabs_free = 0;
        int node;
        struct kmem_cache_node *n;
 
-       active_objs = 0;
-       num_slabs = 0;
        for_each_kmem_cache_node(cachep, node, n) {
-
                check_irq_on();
                spin_lock_irq(&n->list_lock);
 
-               num_slabs += n->num_slabs;
+               num_slabs += n->active_slabs + n->free_slabs;
+               num_slabs_free += n->free_slabs;
 
-               list_for_each_entry(page, &n->slabs_partial, lru) {
-                       if (page->active == cachep->num && !error)
-                               error = "slabs_partial accounting error";
-                       if (!page->active && !error)
-                               error = "slabs_partial accounting error";
-                       active_objs += page->active;
-                       num_slabs_partial++;
-               }
+               free_objs += n->free_objects;
 
-               list_for_each_entry(page, &n->slabs_free, lru) {
-                       if (page->active && !error)
-                               error = "slabs_free accounting error";
-                       num_slabs_free++;
-               }
-
-               free_objects += n->free_objects;
                if (n->shared)
                        shared_avail += n->shared->avail;
 
@@ -4146,15 +4130,8 @@ void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
        }
        num_objs = num_slabs * cachep->num;
        active_slabs = num_slabs - num_slabs_free;
-       num_slabs_full = num_slabs - (num_slabs_partial + num_slabs_free);
-       active_objs += (num_slabs_full * cachep->num);
 
-       if (num_objs - active_objs != free_objects && !error)
-               error = "free_objects accounting error";
-
-       name = cachep->name;
-       if (error)
-               pr_err("slab: cache %s error: %s\n", name, error);
+       active_objs = num_objs - free_objs;
 
        sinfo->active_objs = active_objs;
        sinfo->num_objs = num_objs;
index 699b072dc46e4678d6c482255e8f7e7263693272..26123c512fee6a7330228b9a27d33b4c952fdc0e 100644 (file)
--- a/mm/slab.h
+++ b/mm/slab.h
@@ -447,7 +447,8 @@ struct kmem_cache_node {
        struct list_head slabs_partial; /* partial list first, better asm code */
        struct list_head slabs_full;
        struct list_head slabs_free;
-       unsigned long num_slabs;
+       unsigned long active_slabs;     /* length of slabs_partial+slabs_full */
+       unsigned long free_slabs;       /* length of slabs_free */
        unsigned long free_objects;
        unsigned int free_limit;
        unsigned int colour_next;       /* Per-node cache coloring */