flow: better memory management
Eric Dumazet [Fri, 10 Sep 2010 07:00:25 +0000 (07:00 +0000)]
Allocate hash tables for every online cpus, not every possible ones.

NUMA aware allocations.

Dont use a full page on arches where PAGE_SIZE > 1024*sizeof(void *)

misc:
  __percpu , __read_mostly, __cpuinit annotations
  flow_compare_t is just an "unsigned long"

Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

net/core/flow.c

index f67dcbf..b143b86 100644 (file)
@@ -53,8 +53,7 @@ struct flow_flush_info {
 
 struct flow_cache {
        u32                             hash_shift;
-       unsigned long                   order;
-       struct flow_cache_percpu        *percpu;
+       struct flow_cache_percpu __percpu *percpu;
        struct notifier_block           hotcpu_notifier;
        int                             low_watermark;
        int                             high_watermark;
@@ -64,7 +63,7 @@ struct flow_cache {
 atomic_t flow_cache_genid = ATOMIC_INIT(0);
 EXPORT_SYMBOL(flow_cache_genid);
 static struct flow_cache flow_cache_global;
-static struct kmem_cache *flow_cachep;
+static struct kmem_cache *flow_cachep __read_mostly;
 
 static DEFINE_SPINLOCK(flow_cache_gc_lock);
 static LIST_HEAD(flow_cache_gc_list);
@@ -181,11 +180,7 @@ static u32 flow_hash_code(struct flow_cache *fc,
                & (flow_cache_hash_size(fc) - 1));
 }
 
-#if (BITS_PER_LONG == 64)
-typedef u64 flow_compare_t;
-#else
-typedef u32 flow_compare_t;
-#endif
+typedef unsigned long flow_compare_t;
 
 /* I hear what you're saying, use memcmp.  But memcmp cannot make
  * important assumptions that we can here, such as alignment and
@@ -357,62 +352,73 @@ void flow_cache_flush(void)
        put_online_cpus();
 }
 
-static void __init flow_cache_cpu_prepare(struct flow_cache *fc,
-                                         struct flow_cache_percpu *fcp)
+static int __cpuinit flow_cache_cpu_prepare(struct flow_cache *fc, int cpu)
 {
-       fcp->hash_table = (struct hlist_head *)
-               __get_free_pages(GFP_KERNEL|__GFP_ZERO, fc->order);
-       if (!fcp->hash_table)
-               panic("NET: failed to allocate flow cache order %lu\n", fc->order);
+       struct flow_cache_percpu *fcp = per_cpu_ptr(fc->percpu, cpu);
+       size_t sz = sizeof(struct hlist_head) * flow_cache_hash_size(fc);
 
-       fcp->hash_rnd_recalc = 1;
-       fcp->hash_count = 0;
-       tasklet_init(&fcp->flush_tasklet, flow_cache_flush_tasklet, 0);
+       if (!fcp->hash_table) {
+               fcp->hash_table = kzalloc_node(sz, GFP_KERNEL, cpu_to_node(cpu));
+               if (!fcp->hash_table) {
+                       pr_err("NET: failed to allocate flow cache sz %zu\n", sz);
+                       return -ENOMEM;
+               }
+               fcp->hash_rnd_recalc = 1;
+               fcp->hash_count = 0;
+               tasklet_init(&fcp->flush_tasklet, flow_cache_flush_tasklet, 0);
+       }
+       return 0;
 }
 
-static int flow_cache_cpu(struct notifier_block *nfb,
+static int __cpuinit flow_cache_cpu(struct notifier_block *nfb,
                          unsigned long action,
                          void *hcpu)
 {
        struct flow_cache *fc = container_of(nfb, struct flow_cache, hotcpu_notifier);
-       int cpu = (unsigned long) hcpu;
+       int res, cpu = (unsigned long) hcpu;
        struct flow_cache_percpu *fcp = per_cpu_ptr(fc->percpu, cpu);
 
-       if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
+       switch (action) {
+       case CPU_UP_PREPARE:
+       case CPU_UP_PREPARE_FROZEN:
+               res = flow_cache_cpu_prepare(fc, cpu);
+               if (res)
+                       return notifier_from_errno(res);
+               break;
+       case CPU_DEAD:
+       case CPU_DEAD_FROZEN:
                __flow_cache_shrink(fc, fcp, 0);
+               break;
+       }
        return NOTIFY_OK;
 }
 
-static int flow_cache_init(struct flow_cache *fc)
+static int __init flow_cache_init(struct flow_cache *fc)
 {
-       unsigned long order;
        int i;
 
        fc->hash_shift = 10;
        fc->low_watermark = 2 * flow_cache_hash_size(fc);
        fc->high_watermark = 4 * flow_cache_hash_size(fc);
 
-       for (order = 0;
-            (PAGE_SIZE << order) <
-                    (sizeof(struct hlist_head)*flow_cache_hash_size(fc));
-            order++)
-               /* NOTHING */;
-       fc->order = order;
        fc->percpu = alloc_percpu(struct flow_cache_percpu);
+       if (!fc->percpu)
+               return -ENOMEM;
 
-       setup_timer(&fc->rnd_timer, flow_cache_new_hashrnd,
-                   (unsigned long) fc);
-       fc->rnd_timer.expires = jiffies + FLOW_HASH_RND_PERIOD;
-       add_timer(&fc->rnd_timer);
-
-       for_each_possible_cpu(i)
-               flow_cache_cpu_prepare(fc, per_cpu_ptr(fc->percpu, i));
-
+       for_each_online_cpu(i) {
+               if (flow_cache_cpu_prepare(fc, i))
+                       return -ENOMEM;
+       }
        fc->hotcpu_notifier = (struct notifier_block){
                .notifier_call = flow_cache_cpu,
        };
        register_hotcpu_notifier(&fc->hotcpu_notifier);
 
+       setup_timer(&fc->rnd_timer, flow_cache_new_hashrnd,
+                   (unsigned long) fc);
+       fc->rnd_timer.expires = jiffies + FLOW_HASH_RND_PERIOD;
+       add_timer(&fc->rnd_timer);
+
        return 0;
 }