]> git.baikalelectronics.ru Git - kernel.git/commitdiff
bpf: Add BPF_MAP_TYPE_LRU_PERCPU_HASH
authorMartin KaFai Lau <kafai@fb.com>
Fri, 11 Nov 2016 18:55:10 +0000 (10:55 -0800)
committerDavid S. Miller <davem@davemloft.net>
Tue, 15 Nov 2016 16:50:20 +0000 (11:50 -0500)
Provide a LRU version of the existing BPF_MAP_TYPE_PERCPU_HASH

Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
include/uapi/linux/bpf.h
kernel/bpf/hashtab.c
kernel/bpf/syscall.c

index ed8c6799fb1474d1b87cbb2c187091dd1b2e90f7..7d9b2832c280c82172cb428a8e6d793359891db0 100644 (file)
@@ -86,6 +86,7 @@ enum bpf_map_type {
        BPF_MAP_TYPE_STACK_TRACE,
        BPF_MAP_TYPE_CGROUP_ARRAY,
        BPF_MAP_TYPE_LRU_HASH,
+       BPF_MAP_TYPE_LRU_PERCPU_HASH,
 };
 
 enum bpf_prog_type {
@@ -108,7 +109,7 @@ enum bpf_prog_type {
 
 #define BPF_F_NO_PREALLOC      (1U << 0)
 /* Instead of having one common LRU list in the
- * BPF_MAP_TYPE_LRU_HASH map, use a percpu LRU list
+ * BPF_MAP_TYPE_LRU_[PERCPU_]HASH map, use a percpu LRU list
  * which can scale and perform better.
  * Note, the LRU nodes (including free nodes) cannot be moved
  * across different LRU lists.
index 4a9e71a7c41f4d0b283e86105ec8da39aa95b33a..34debc1a9641875382b6603820f3e85c4a18dffc 100644 (file)
@@ -62,7 +62,14 @@ static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node);
 
 static bool htab_is_lru(const struct bpf_htab *htab)
 {
-       return htab->map.map_type == BPF_MAP_TYPE_LRU_HASH;
+       return htab->map.map_type == BPF_MAP_TYPE_LRU_HASH ||
+               htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
+}
+
+static bool htab_is_percpu(const struct bpf_htab *htab)
+{
+       return htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH ||
+               htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
 }
 
 static inline void htab_elem_set_ptr(struct htab_elem *l, u32 key_size,
@@ -85,7 +92,7 @@ static void htab_free_elems(struct bpf_htab *htab)
 {
        int i;
 
-       if (htab->map.map_type != BPF_MAP_TYPE_PERCPU_HASH)
+       if (!htab_is_percpu(htab))
                goto free_elems;
 
        for (i = 0; i < htab->map.max_entries; i++) {
@@ -122,7 +129,7 @@ static int prealloc_init(struct bpf_htab *htab)
        if (!htab->elems)
                return -ENOMEM;
 
-       if (htab->map.map_type != BPF_MAP_TYPE_PERCPU_HASH)
+       if (!htab_is_percpu(htab))
                goto skip_percpu_elems;
 
        for (i = 0; i < htab->map.max_entries; i++) {
@@ -195,8 +202,10 @@ static int alloc_extra_elems(struct bpf_htab *htab)
 /* Called from syscall */
 static struct bpf_map *htab_map_alloc(union bpf_attr *attr)
 {
-       bool percpu = attr->map_type == BPF_MAP_TYPE_PERCPU_HASH;
-       bool lru = attr->map_type == BPF_MAP_TYPE_LRU_HASH;
+       bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
+                      attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
+       bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
+                   attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
        /* percpu_lru means each cpu has its own LRU list.
         * it is different from BPF_MAP_TYPE_PERCPU_HASH where
         * the map's value itself is percpu.  percpu_lru has
@@ -823,12 +832,84 @@ err:
        return ret;
 }
 
+static int __htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
+                                            void *value, u64 map_flags,
+                                            bool onallcpus)
+{
+       struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
+       struct htab_elem *l_new = NULL, *l_old;
+       struct hlist_head *head;
+       unsigned long flags;
+       struct bucket *b;
+       u32 key_size, hash;
+       int ret;
+
+       if (unlikely(map_flags > BPF_EXIST))
+               /* unknown flags */
+               return -EINVAL;
+
+       WARN_ON_ONCE(!rcu_read_lock_held());
+
+       key_size = map->key_size;
+
+       hash = htab_map_hash(key, key_size);
+
+       b = __select_bucket(htab, hash);
+       head = &b->head;
+
+       /* For LRU, we need to alloc before taking bucket's
+        * spinlock because LRU's elem alloc may need
+        * to remove older elem from htab and this removal
+        * operation will need a bucket lock.
+        */
+       if (map_flags != BPF_EXIST) {
+               l_new = prealloc_lru_pop(htab, key, hash);
+               if (!l_new)
+                       return -ENOMEM;
+       }
+
+       /* bpf_map_update_elem() can be called in_irq() */
+       raw_spin_lock_irqsave(&b->lock, flags);
+
+       l_old = lookup_elem_raw(head, hash, key, key_size);
+
+       ret = check_flags(htab, l_old, map_flags);
+       if (ret)
+               goto err;
+
+       if (l_old) {
+               bpf_lru_node_set_ref(&l_old->lru_node);
+
+               /* per-cpu hash map can update value in-place */
+               pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
+                               value, onallcpus);
+       } else {
+               pcpu_copy_value(htab, htab_elem_get_ptr(l_new, key_size),
+                               value, onallcpus);
+               hlist_add_head_rcu(&l_new->hash_node, head);
+               l_new = NULL;
+       }
+       ret = 0;
+err:
+       raw_spin_unlock_irqrestore(&b->lock, flags);
+       if (l_new)
+               bpf_lru_push_free(&htab->lru, &l_new->lru_node);
+       return ret;
+}
+
 static int htab_percpu_map_update_elem(struct bpf_map *map, void *key,
                                       void *value, u64 map_flags)
 {
        return __htab_percpu_map_update_elem(map, key, value, map_flags, false);
 }
 
+static int htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
+                                          void *value, u64 map_flags)
+{
+       return __htab_lru_percpu_map_update_elem(map, key, value, map_flags,
+                                                false);
+}
+
 /* Called from syscall or from eBPF program */
 static int htab_map_delete_elem(struct bpf_map *map, void *key)
 {
@@ -976,8 +1057,21 @@ static void *htab_percpu_map_lookup_elem(struct bpf_map *map, void *key)
                return NULL;
 }
 
+static void *htab_lru_percpu_map_lookup_elem(struct bpf_map *map, void *key)
+{
+       struct htab_elem *l = __htab_map_lookup_elem(map, key);
+
+       if (l) {
+               bpf_lru_node_set_ref(&l->lru_node);
+               return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
+       }
+
+       return NULL;
+}
+
 int bpf_percpu_hash_copy(struct bpf_map *map, void *key, void *value)
 {
+       struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
        struct htab_elem *l;
        void __percpu *pptr;
        int ret = -ENOENT;
@@ -993,6 +1087,8 @@ int bpf_percpu_hash_copy(struct bpf_map *map, void *key, void *value)
        l = __htab_map_lookup_elem(map, key);
        if (!l)
                goto out;
+       if (htab_is_lru(htab))
+               bpf_lru_node_set_ref(&l->lru_node);
        pptr = htab_elem_get_ptr(l, map->key_size);
        for_each_possible_cpu(cpu) {
                bpf_long_memcpy(value + off,
@@ -1008,10 +1104,16 @@ out:
 int bpf_percpu_hash_update(struct bpf_map *map, void *key, void *value,
                           u64 map_flags)
 {
+       struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
        int ret;
 
        rcu_read_lock();
-       ret = __htab_percpu_map_update_elem(map, key, value, map_flags, true);
+       if (htab_is_lru(htab))
+               ret = __htab_lru_percpu_map_update_elem(map, key, value,
+                                                       map_flags, true);
+       else
+               ret = __htab_percpu_map_update_elem(map, key, value, map_flags,
+                                                   true);
        rcu_read_unlock();
 
        return ret;
@@ -1031,11 +1133,26 @@ static struct bpf_map_type_list htab_percpu_type __read_mostly = {
        .type = BPF_MAP_TYPE_PERCPU_HASH,
 };
 
+static const struct bpf_map_ops htab_lru_percpu_ops = {
+       .map_alloc = htab_map_alloc,
+       .map_free = htab_map_free,
+       .map_get_next_key = htab_map_get_next_key,
+       .map_lookup_elem = htab_lru_percpu_map_lookup_elem,
+       .map_update_elem = htab_lru_percpu_map_update_elem,
+       .map_delete_elem = htab_lru_map_delete_elem,
+};
+
+static struct bpf_map_type_list htab_lru_percpu_type __read_mostly = {
+       .ops = &htab_lru_percpu_ops,
+       .type = BPF_MAP_TYPE_LRU_PERCPU_HASH,
+};
+
 static int __init register_htab_map(void)
 {
        bpf_register_map_type(&htab_type);
        bpf_register_map_type(&htab_percpu_type);
        bpf_register_map_type(&htab_lru_type);
+       bpf_register_map_type(&htab_lru_percpu_type);
        return 0;
 }
 late_initcall(register_htab_map);
index 233e3ac836a6689e7aefee822cbd38a3f4d00216..ce1b7de7d72c6926ac5d4a50f812df764d25a07c 100644 (file)
@@ -292,6 +292,7 @@ static int map_lookup_elem(union bpf_attr *attr)
                goto free_key;
 
        if (map->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
+           map->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH ||
            map->map_type == BPF_MAP_TYPE_PERCPU_ARRAY)
                value_size = round_up(map->value_size, 8) * num_possible_cpus();
        else
@@ -302,7 +303,8 @@ static int map_lookup_elem(union bpf_attr *attr)
        if (!value)
                goto free_key;
 
-       if (map->map_type == BPF_MAP_TYPE_PERCPU_HASH) {
+       if (map->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
+           map->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH) {
                err = bpf_percpu_hash_copy(map, key, value);
        } else if (map->map_type == BPF_MAP_TYPE_PERCPU_ARRAY) {
                err = bpf_percpu_array_copy(map, key, value);
@@ -366,6 +368,7 @@ static int map_update_elem(union bpf_attr *attr)
                goto free_key;
 
        if (map->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
+           map->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH ||
            map->map_type == BPF_MAP_TYPE_PERCPU_ARRAY)
                value_size = round_up(map->value_size, 8) * num_possible_cpus();
        else
@@ -385,7 +388,8 @@ static int map_update_elem(union bpf_attr *attr)
         */
        preempt_disable();
        __this_cpu_inc(bpf_prog_active);
-       if (map->map_type == BPF_MAP_TYPE_PERCPU_HASH) {
+       if (map->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
+           map->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH) {
                err = bpf_percpu_hash_update(map, key, value, attr->flags);
        } else if (map->map_type == BPF_MAP_TYPE_PERCPU_ARRAY) {
                err = bpf_percpu_array_update(map, key, value, attr->flags);