bpf: minor cleanups for dev_map
authorDaniel Borkmann <daniel@iogearbox.net>
Tue, 22 Aug 2017 23:47:54 +0000 (01:47 +0200)
committerDavid S. Miller <davem@davemloft.net>
Wed, 23 Aug 2017 04:26:29 +0000 (21:26 -0700)
Some minor code cleanups, while going over it I also noticed that
we're accounting the bitmap only for one CPU currently, so fix that
up as well.

Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
kernel/bpf/devmap.c

index fa08181d1c3d3971e490246bc16d18c9ba740f37..bfecabfd4974d73a5e42ed3ddcded3f23cd56a9b 100644 (file)
  * calls will fail at this point.
  */
 #include <linux/bpf.h>
-#include <linux/jhash.h>
 #include <linux/filter.h>
-#include <linux/rculist_nulls.h>
-#include "percpu_freelist.h"
-#include "bpf_lru_list.h"
-#include "map_in_map.h"
 
 struct bpf_dtab_netdev {
        struct net_device *dev;
-       int key;
-       struct rcu_head rcu;
        struct bpf_dtab *dtab;
+       unsigned int bit;
+       struct rcu_head rcu;
 };
 
 struct bpf_dtab {
        struct bpf_map map;
        struct bpf_dtab_netdev **netdev_map;
-       unsigned long int __percpu *flush_needed;
+       unsigned long __percpu *flush_needed;
        struct list_head list;
 };
 
 static DEFINE_SPINLOCK(dev_map_lock);
 static LIST_HEAD(dev_map_list);
 
+static u64 dev_map_bitmap_size(const union bpf_attr *attr)
+{
+       return BITS_TO_LONGS(attr->max_entries) * sizeof(unsigned long);
+}
+
 static struct bpf_map *dev_map_alloc(union bpf_attr *attr)
 {
        struct bpf_dtab *dtab;
@@ -95,11 +95,9 @@ static struct bpf_map *dev_map_alloc(union bpf_attr *attr)
        dtab->map.map_flags = attr->map_flags;
        dtab->map.numa_node = bpf_map_attr_numa_node(attr);
 
-       err = -ENOMEM;
-
        /* make sure page count doesn't overflow */
        cost = (u64) dtab->map.max_entries * sizeof(struct bpf_dtab_netdev *);
-       cost += BITS_TO_LONGS(attr->max_entries) * sizeof(unsigned long);
+       cost += dev_map_bitmap_size(attr) * num_possible_cpus();
        if (cost >= U32_MAX - PAGE_SIZE)
                goto free_dtab;
 
@@ -110,12 +108,9 @@ static struct bpf_map *dev_map_alloc(union bpf_attr *attr)
        if (err)
                goto free_dtab;
 
-       err = -ENOMEM;
        /* A per cpu bitfield with a bit per possible net device */
-       dtab->flush_needed = __alloc_percpu(
-                               BITS_TO_LONGS(attr->max_entries) *
-                               sizeof(unsigned long),
-                               __alignof__(unsigned long));
+       dtab->flush_needed = __alloc_percpu(dev_map_bitmap_size(attr),
+                                           __alignof__(unsigned long));
        if (!dtab->flush_needed)
                goto free_dtab;
 
@@ -128,12 +123,12 @@ static struct bpf_map *dev_map_alloc(union bpf_attr *attr)
        spin_lock(&dev_map_lock);
        list_add_tail_rcu(&dtab->list, &dev_map_list);
        spin_unlock(&dev_map_lock);
-       return &dtab->map;
 
+       return &dtab->map;
 free_dtab:
        free_percpu(dtab->flush_needed);
        kfree(dtab);
-       return ERR_PTR(err);
+       return ERR_PTR(-ENOMEM);
 }
 
 static void dev_map_free(struct bpf_map *map)
@@ -178,9 +173,6 @@ static void dev_map_free(struct bpf_map *map)
                kfree(dev);
        }
 
-       /* At this point bpf program is detached and all pending operations
-        * _must_ be complete
-        */
        free_percpu(dtab->flush_needed);
        bpf_map_area_free(dtab->netdev_map);
        kfree(dtab);
@@ -190,7 +182,7 @@ static int dev_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
 {
        struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
        u32 index = key ? *(u32 *)key : U32_MAX;
-       u32 *next = (u32 *)next_key;
+       u32 *next = next_key;
 
        if (index >= dtab->map.max_entries) {
                *next = 0;
@@ -199,29 +191,16 @@ static int dev_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
 
        if (index == dtab->map.max_entries - 1)
                return -ENOENT;
-
        *next = index + 1;
        return 0;
 }
 
-void __dev_map_insert_ctx(struct bpf_map *map, u32 key)
+void __dev_map_insert_ctx(struct bpf_map *map, u32 bit)
 {
        struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
        unsigned long *bitmap = this_cpu_ptr(dtab->flush_needed);
 
-       __set_bit(key, bitmap);
-}
-
-struct net_device  *__dev_map_lookup_elem(struct bpf_map *map, u32 key)
-{
-       struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
-       struct bpf_dtab_netdev *dev;
-
-       if (key >= map->max_entries)
-               return NULL;
-
-       dev = READ_ONCE(dtab->netdev_map[key]);
-       return dev ? dev->dev : NULL;
+       __set_bit(bit, bitmap);
 }
 
 /* __dev_map_flush is called from xdp_do_flush_map() which _must_ be signaled
@@ -248,7 +227,6 @@ void __dev_map_flush(struct bpf_map *map)
                        continue;
 
                netdev = dev->dev;
-
                __clear_bit(bit, bitmap);
                if (unlikely(!netdev || !netdev->netdev_ops->ndo_xdp_flush))
                        continue;
@@ -261,43 +239,49 @@ void __dev_map_flush(struct bpf_map *map)
  * update happens in parallel here a dev_put wont happen until after reading the
  * ifindex.
  */
-static void *dev_map_lookup_elem(struct bpf_map *map, void *key)
+struct net_device  *__dev_map_lookup_elem(struct bpf_map *map, u32 key)
 {
        struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
        struct bpf_dtab_netdev *dev;
-       u32 i = *(u32 *)key;
 
-       if (i >= map->max_entries)
+       if (key >= map->max_entries)
                return NULL;
 
-       dev = READ_ONCE(dtab->netdev_map[i]);
-       return dev ? &dev->dev->ifindex : NULL;
+       dev = READ_ONCE(dtab->netdev_map[key]);
+       return dev ? dev->dev : NULL;
 }
 
-static void dev_map_flush_old(struct bpf_dtab_netdev *old_dev)
+static void *dev_map_lookup_elem(struct bpf_map *map, void *key)
+{
+       struct net_device *dev = __dev_map_lookup_elem(map, *(u32 *)key);
+
+       return dev ? &dev->ifindex : NULL;
+}
+
+static void dev_map_flush_old(struct bpf_dtab_netdev *dev)
 {
-       if (old_dev->dev->netdev_ops->ndo_xdp_flush) {
-               struct net_device *fl = old_dev->dev;
+       if (dev->dev->netdev_ops->ndo_xdp_flush) {
+               struct net_device *fl = dev->dev;
                unsigned long *bitmap;
                int cpu;
 
                for_each_online_cpu(cpu) {
-                       bitmap = per_cpu_ptr(old_dev->dtab->flush_needed, cpu);
-                       __clear_bit(old_dev->key, bitmap);
+                       bitmap = per_cpu_ptr(dev->dtab->flush_needed, cpu);
+                       __clear_bit(dev->bit, bitmap);
 
-                       fl->netdev_ops->ndo_xdp_flush(old_dev->dev);
+                       fl->netdev_ops->ndo_xdp_flush(dev->dev);
                }
        }
 }
 
 static void __dev_map_entry_free(struct rcu_head *rcu)
 {
-       struct bpf_dtab_netdev *old_dev;
+       struct bpf_dtab_netdev *dev;
 
-       old_dev = container_of(rcu, struct bpf_dtab_netdev, rcu);
-       dev_map_flush_old(old_dev);
-       dev_put(old_dev->dev);
-       kfree(old_dev);
+       dev = container_of(rcu, struct bpf_dtab_netdev, rcu);
+       dev_map_flush_old(dev);
+       dev_put(dev->dev);
+       kfree(dev);
 }
 
 static int dev_map_delete_elem(struct bpf_map *map, void *key)
@@ -309,8 +293,8 @@ static int dev_map_delete_elem(struct bpf_map *map, void *key)
        if (k >= map->max_entries)
                return -EINVAL;
 
-       /* Use synchronize_rcu() here to ensure any rcu critical sections
-        * have completed, but this does not guarantee a flush has happened
+       /* Use call_rcu() here to ensure any rcu critical sections have
+        * completed, but this does not guarantee a flush has happened
         * yet. Because driver side rcu_read_lock/unlock only protects the
         * running XDP program. However, for pending flush operations the
         * dev and ctx are stored in another per cpu map. And additionally,
@@ -334,10 +318,8 @@ static int dev_map_update_elem(struct bpf_map *map, void *key, void *value,
 
        if (unlikely(map_flags > BPF_EXIST))
                return -EINVAL;
-
        if (unlikely(i >= dtab->map.max_entries))
                return -E2BIG;
-
        if (unlikely(map_flags == BPF_NOEXIST))
                return -EEXIST;
 
@@ -355,7 +337,7 @@ static int dev_map_update_elem(struct bpf_map *map, void *key, void *value,
                        return -EINVAL;
                }
 
-               dev->key = i;
+               dev->bit = i;
                dev->dtab = dtab;
        }