flow: structurize flow cache
authorTimo Teräs <timo.teras@iki.fi>
Wed, 31 Mar 2010 00:17:06 +0000 (00:17 +0000)
committerDavid S. Miller <davem@davemloft.net>
Fri, 2 Apr 2010 02:41:36 +0000 (19:41 -0700)
Group all per-cpu data to one structure instead of having many
globals. Also prepare the internals so that we can have multiple
instances of the flow cache if needed.

Only the kmem_cache is left as a global as all flow caches share
the same element size, and benefit from using a common cache.

Signed-off-by: Timo Teras <timo.teras@iki.fi>
Acked-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/core/flow.c

index 96015871eceadeacc65d8b572fdcf0d757ecccca..1d27ca6b421d9d25a39272cc4ee3cece8c2866d0 100644 (file)
@@ -35,104 +35,105 @@ struct flow_cache_entry {
        atomic_t                *object_ref;
 };
 
-atomic_t flow_cache_genid = ATOMIC_INIT(0);
-
-static u32 flow_hash_shift;
-#define flow_hash_size (1 << flow_hash_shift)
-static DEFINE_PER_CPU(struct flow_cache_entry **, flow_tables) = { NULL };
-
-#define flow_table(cpu) (per_cpu(flow_tables, cpu))
-
-static struct kmem_cache *flow_cachep __read_mostly;
-
-static int flow_lwm, flow_hwm;
-
-struct flow_percpu_info {
-       int hash_rnd_recalc;
-       u32 hash_rnd;
-       int count;
+struct flow_cache_percpu {
+       struct flow_cache_entry **      hash_table;
+       int                             hash_count;
+       u32                             hash_rnd;
+       int                             hash_rnd_recalc;
+       struct tasklet_struct           flush_tasklet;
 };
-static DEFINE_PER_CPU(struct flow_percpu_info, flow_hash_info) = { 0 };
-
-#define flow_hash_rnd_recalc(cpu) \
-       (per_cpu(flow_hash_info, cpu).hash_rnd_recalc)
-#define flow_hash_rnd(cpu) \
-       (per_cpu(flow_hash_info, cpu).hash_rnd)
-#define flow_count(cpu) \
-       (per_cpu(flow_hash_info, cpu).count)
-
-static struct timer_list flow_hash_rnd_timer;
-
-#define FLOW_HASH_RND_PERIOD   (10 * 60 * HZ)
 
 struct flow_flush_info {
-       atomic_t cpuleft;
-       struct completion completion;
+       struct flow_cache *             cache;
+       atomic_t                        cpuleft;
+       struct completion               completion;
 };
-static DEFINE_PER_CPU(struct tasklet_struct, flow_flush_tasklets) = { NULL };
 
-#define flow_flush_tasklet(cpu) (&per_cpu(flow_flush_tasklets, cpu))
+struct flow_cache {
+       u32                             hash_shift;
+       unsigned long                   order;
+       struct flow_cache_percpu *      percpu;
+       struct notifier_block           hotcpu_notifier;
+       int                             low_watermark;
+       int                             high_watermark;
+       struct timer_list               rnd_timer;
+};
+
+atomic_t flow_cache_genid = ATOMIC_INIT(0);
+static struct flow_cache flow_cache_global;
+static struct kmem_cache *flow_cachep;
+
+#define flow_cache_hash_size(cache)    (1 << (cache)->hash_shift)
+#define FLOW_HASH_RND_PERIOD           (10 * 60 * HZ)
 
 static void flow_cache_new_hashrnd(unsigned long arg)
 {
+       struct flow_cache *fc = (void *) arg;
        int i;
 
        for_each_possible_cpu(i)
-               flow_hash_rnd_recalc(i) = 1;
+               per_cpu_ptr(fc->percpu, i)->hash_rnd_recalc = 1;
 
-       flow_hash_rnd_timer.expires = jiffies + FLOW_HASH_RND_PERIOD;
-       add_timer(&flow_hash_rnd_timer);
+       fc->rnd_timer.expires = jiffies + FLOW_HASH_RND_PERIOD;
+       add_timer(&fc->rnd_timer);
 }
 
-static void flow_entry_kill(int cpu, struct flow_cache_entry *fle)
+static void flow_entry_kill(struct flow_cache *fc,
+                           struct flow_cache_percpu *fcp,
+                           struct flow_cache_entry *fle)
 {
        if (fle->object)
                atomic_dec(fle->object_ref);
        kmem_cache_free(flow_cachep, fle);
-       flow_count(cpu)--;
+       fcp->hash_count--;
 }
 
-static void __flow_cache_shrink(int cpu, int shrink_to)
+static void __flow_cache_shrink(struct flow_cache *fc,
+                               struct flow_cache_percpu *fcp,
+                               int shrink_to)
 {
        struct flow_cache_entry *fle, **flp;
        int i;
 
-       for (i = 0; i < flow_hash_size; i++) {
+       for (i = 0; i < flow_cache_hash_size(fc); i++) {
                int k = 0;
 
-               flp = &flow_table(cpu)[i];
+               flp = &fcp->hash_table[i];
                while ((fle = *flp) != NULL && k < shrink_to) {
                        k++;
                        flp = &fle->next;
                }
                while ((fle = *flp) != NULL) {
                        *flp = fle->next;
-                       flow_entry_kill(cpu, fle);
+                       flow_entry_kill(fc, fcp, fle);
                }
        }
 }
 
-static void flow_cache_shrink(int cpu)
+static void flow_cache_shrink(struct flow_cache *fc,
+                             struct flow_cache_percpu *fcp)
 {
-       int shrink_to = flow_lwm / flow_hash_size;
+       int shrink_to = fc->low_watermark / flow_cache_hash_size(fc);
 
-       __flow_cache_shrink(cpu, shrink_to);
+       __flow_cache_shrink(fc, fcp, shrink_to);
 }
 
-static void flow_new_hash_rnd(int cpu)
+static void flow_new_hash_rnd(struct flow_cache *fc,
+                             struct flow_cache_percpu *fcp)
 {
-       get_random_bytes(&flow_hash_rnd(cpu), sizeof(u32));
-       flow_hash_rnd_recalc(cpu) = 0;
-
-       __flow_cache_shrink(cpu, 0);
+       get_random_bytes(&fcp->hash_rnd, sizeof(u32));
+       fcp->hash_rnd_recalc = 0;
+       __flow_cache_shrink(fc, fcp, 0);
 }
 
-static u32 flow_hash_code(struct flowi *key, int cpu)
+static u32 flow_hash_code(struct flow_cache *fc,
+                         struct flow_cache_percpu *fcp,
+                         struct flowi *key)
 {
        u32 *k = (u32 *) key;
 
-       return (jhash2(k, (sizeof(*key) / sizeof(u32)), flow_hash_rnd(cpu)) &
-               (flow_hash_size - 1));
+       return (jhash2(k, (sizeof(*key) / sizeof(u32)), fcp->hash_rnd)
+               & (flow_cache_hash_size(fc) - 1));
 }
 
 #if (BITS_PER_LONG == 64)
@@ -168,24 +169,25 @@ static int flow_key_compare(struct flowi *key1, struct flowi *key2)
 void *flow_cache_lookup(struct net *net, struct flowi *key, u16 family, u8 dir,
                        flow_resolve_t resolver)
 {
+       struct flow_cache *fc = &flow_cache_global;
+       struct flow_cache_percpu *fcp;
        struct flow_cache_entry *fle, **head;
        unsigned int hash;
-       int cpu;
 
        local_bh_disable();
-       cpu = smp_processor_id();
+       fcp = per_cpu_ptr(fc->percpu, smp_processor_id());
 
        fle = NULL;
        /* Packet really early in init?  Making flow_cache_init a
         * pre-smp initcall would solve this.  --RR */
-       if (!flow_table(cpu))
+       if (!fcp->hash_table)
                goto nocache;
 
-       if (flow_hash_rnd_recalc(cpu))
-               flow_new_hash_rnd(cpu);
-       hash = flow_hash_code(key, cpu);
+       if (fcp->hash_rnd_recalc)
+               flow_new_hash_rnd(fc, fcp);
+       hash = flow_hash_code(fc, fcp, key);
 
-       head = &flow_table(cpu)[hash];
+       head = &fcp->hash_table[hash];
        for (fle = *head; fle; fle = fle->next) {
                if (fle->family == family &&
                    fle->dir == dir &&
@@ -204,8 +206,8 @@ void *flow_cache_lookup(struct net *net, struct flowi *key, u16 family, u8 dir,
        }
 
        if (!fle) {
-               if (flow_count(cpu) > flow_hwm)
-                       flow_cache_shrink(cpu);
+               if (fcp->hash_count > fc->high_watermark)
+                       flow_cache_shrink(fc, fcp);
 
                fle = kmem_cache_alloc(flow_cachep, GFP_ATOMIC);
                if (fle) {
@@ -215,7 +217,7 @@ void *flow_cache_lookup(struct net *net, struct flowi *key, u16 family, u8 dir,
                        fle->dir = dir;
                        memcpy(&fle->key, key, sizeof(*key));
                        fle->object = NULL;
-                       flow_count(cpu)++;
+                       fcp->hash_count++;
                }
        }
 
@@ -249,14 +251,15 @@ nocache:
 static void flow_cache_flush_tasklet(unsigned long data)
 {
        struct flow_flush_info *info = (void *)data;
+       struct flow_cache *fc = info->cache;
+       struct flow_cache_percpu *fcp;
        int i;
-       int cpu;
 
-       cpu = smp_processor_id();
-       for (i = 0; i < flow_hash_size; i++) {
+       fcp = per_cpu_ptr(fc->percpu, smp_processor_id());
+       for (i = 0; i < flow_cache_hash_size(fc); i++) {
                struct flow_cache_entry *fle;
 
-               fle = flow_table(cpu)[i];
+               fle = fcp->hash_table[i];
                for (; fle; fle = fle->next) {
                        unsigned genid = atomic_read(&flow_cache_genid);
 
@@ -272,7 +275,6 @@ static void flow_cache_flush_tasklet(unsigned long data)
                complete(&info->completion);
 }
 
-static void flow_cache_flush_per_cpu(void *) __attribute__((__unused__));
 static void flow_cache_flush_per_cpu(void *data)
 {
        struct flow_flush_info *info = data;
@@ -280,8 +282,7 @@ static void flow_cache_flush_per_cpu(void *data)
        struct tasklet_struct *tasklet;
 
        cpu = smp_processor_id();
-
-       tasklet = flow_flush_tasklet(cpu);
+       tasklet = &per_cpu_ptr(info->cache->percpu, cpu)->flush_tasklet;
        tasklet->data = (unsigned long)info;
        tasklet_schedule(tasklet);
 }
@@ -294,6 +295,7 @@ void flow_cache_flush(void)
        /* Don't want cpus going down or up during this. */
        get_online_cpus();
        mutex_lock(&flow_flush_sem);
+       info.cache = &flow_cache_global;
        atomic_set(&info.cpuleft, num_online_cpus());
        init_completion(&info.completion);
 
@@ -307,62 +309,75 @@ void flow_cache_flush(void)
        put_online_cpus();
 }
 
-static void __init flow_cache_cpu_prepare(int cpu)
+static void __init flow_cache_cpu_prepare(struct flow_cache *fc,
+                                         struct flow_cache_percpu *fcp)
 {
-       struct tasklet_struct *tasklet;
-       unsigned long order;
-
-       for (order = 0;
-            (PAGE_SIZE << order) <
-                    (sizeof(struct flow_cache_entry *)*flow_hash_size);
-            order++)
-               /* NOTHING */;
-
-       flow_table(cpu) = (struct flow_cache_entry **)
-               __get_free_pages(GFP_KERNEL|__GFP_ZERO, order);
-       if (!flow_table(cpu))
-               panic("NET: failed to allocate flow cache order %lu\n", order);
-
-       flow_hash_rnd_recalc(cpu) = 1;
-       flow_count(cpu) = 0;
-
-       tasklet = flow_flush_tasklet(cpu);
-       tasklet_init(tasklet, flow_cache_flush_tasklet, 0);
+       fcp->hash_table = (struct flow_cache_entry **)
+               __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);
+
+       fcp->hash_rnd_recalc = 1;
+       fcp->hash_count = 0;
+       tasklet_init(&fcp->flush_tasklet, flow_cache_flush_tasklet, 0);
 }
 
 static int 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;
+       struct flow_cache_percpu *fcp = per_cpu_ptr(fc->percpu, cpu);
+
        if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
-               __flow_cache_shrink((unsigned long)hcpu, 0);
+               __flow_cache_shrink(fc, fcp, 0);
        return NOTIFY_OK;
 }
 
-static int __init flow_cache_init(void)
+static int flow_cache_init(struct flow_cache *fc)
 {
+       unsigned long order;
        int i;
 
-       flow_cachep = kmem_cache_create("flow_cache",
-                                       sizeof(struct flow_cache_entry),
-                                       0, SLAB_PANIC,
-                                       NULL);
-       flow_hash_shift = 10;
-       flow_lwm = 2 * flow_hash_size;
-       flow_hwm = 4 * flow_hash_size;
+       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 flow_cache_entry *)*flow_cache_hash_size(fc));
+            order++)
+               /* NOTHING */;
+       fc->order = order;
+       fc->percpu = alloc_percpu(struct flow_cache_percpu);
 
-       setup_timer(&flow_hash_rnd_timer, flow_cache_new_hashrnd, 0);
-       flow_hash_rnd_timer.expires = jiffies + FLOW_HASH_RND_PERIOD;
-       add_timer(&flow_hash_rnd_timer);
+       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(i);
+               flow_cache_cpu_prepare(fc, per_cpu_ptr(fc->percpu, i));
+
+       fc->hotcpu_notifier = (struct notifier_block){
+               .notifier_call = flow_cache_cpu,
+       };
+       register_hotcpu_notifier(&fc->hotcpu_notifier);
 
-       hotcpu_notifier(flow_cache_cpu, 0);
        return 0;
 }
 
-module_init(flow_cache_init);
+static int __init flow_cache_init_global(void)
+{
+       flow_cachep = kmem_cache_create("flow_cache",
+                                       sizeof(struct flow_cache_entry),
+                                       0, SLAB_PANIC, NULL);
+
+       return flow_cache_init(&flow_cache_global);
+}
+
+module_init(flow_cache_init_global);
 
 EXPORT_SYMBOL(flow_cache_genid);
 EXPORT_SYMBOL(flow_cache_lookup);