Re: [PATCH bpf-next v2 4/5] bpf: lru: Fix unintended eviction when updating lru hash maps

diff --git a/kernel/bpf/hashtab.c b/kernel/bpf/hashtab.c index d029690246f8..8665eb6b8a7d 100644 --- a/kernel/bpf/hashtab.c +++ b/kernel/bpf/hashtab.c

[ ... ]

+static int htab_lru_map_update_elem_in_place(struct bpf_htab *htab, void *key, void *value,

• 			     u64 map_flags, struct bucket *b,
  

• 			     struct hlist_nulls_head *head, u32 hash,
  

• 			     bool percpu, bool onallcpus)
  

+{

• struct htab_elem *l_new, *l_old, *l_free;
• struct bpf_map *map = &htab->map;
• u32 key_size = map->key_size;
• struct bpf_lru_node *node;
• unsigned long flags;
• void *l_val;
• int ret;
• node = bpf_lru_pop_extra(&htab->lru);
• if (!node)

• return -ENOENT;
  

• l_new = container_of(node, struct htab_elem, lru_node);
• l_new->hash = hash;
• memcpy(l_new->key, key, key_size);
• if (!percpu) {

• l_val = htab_elem_value(l_new, map->key_size);
  

• copy_map_value(map, l_val, value);
  

• bpf_obj_free_fields(map->record, l_val);
  

• }

Should bpf_obj_free_fields() be called here on l_new? The value was just copied into l_new and l_new will be inserted into the hash table when l_old is found. Calling bpf_obj_free_fields() on l_new's value frees the special fields (kptr, bpf_list_head, etc.) that were just copied, then l_new is inserted with freed fields. This appears to be a use-after-free for maps with special fields.

Looking at htab_lru_map_update_elem(), the pattern is to copy the value and NOT call bpf_obj_free_fields() on the new element. The fields are only freed when an element is removed from the hash table.

• ret = htab_lock_bucket(b, &flags);
• if (ret)

• goto err_lock_bucket;
  

• 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_new->lru_node);
  

• if (percpu) {
  

• 	/* per-cpu hash map can update value in-place.
  

• 	 * Keep the same logic in __htab_lru_percpu_map_update_elem().
  

• 	 */
  

• 	pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
  

• 			value, onallcpus);
  

• 	l_free = l_new;
  

• } else {
  

• 	hlist_nulls_add_head_rcu(&l_new->hash_node, head);
  

• 	hlist_nulls_del_rcu(&l_old->hash_node);
  

• 	l_free = l_old;
  

• }
  

• } else {

• ret = -ENOENT;
  

• }

+err:

• htab_unlock_bucket(b, flags);

+err_lock_bucket:

• bpf_lru_push_free(&htab->lru, ret ? node : &l_free->lru_node);

^^^^^^^^^^^^^^^^^^

When ret == 0 and !percpu, l_free points to l_old which is being removed from the hash table. The code calls bpf_lru_push_free() directly instead of htab_lru_push_free(). Looking at htab_lru_push_free():

static void htab_lru_push_free(struct bpf_htab *htab, struct htab_elem *elem) { check_and_free_fields(htab, elem); bpf_map_dec_elem_count(&htab->map); bpf_lru_push_free(&htab->lru, &elem->lru_node); }

It calls check_and_free_fields() to free special fields in the element before pushing to the free list. The code here bypasses that, which would leak special fields (kptr, etc.) in l_old when it has them.

Should this use htab_lru_push_free() instead of bpf_lru_push_free() for the non-error path where l_old is being freed?

• return ret;

+}

--- AI reviewed your patch. Please fix the bug or email reply why it's not a bug. See: https://github.com/kernel-patches/vmtest/blob/master/ci/claude/README.md

CI run summary: https://github.com/kernel-patches/bpf/actions/runs/20720201621