bpf: fix incorrect kmalloc usage in lpm_trie MAP_GET_NEXT_KEY rcu region

In commit b471f2f1 ("bpf: implement MAP_GET_NEXT_KEY command for LPM_TRIE map"),
the implemented MAP_GET_NEXT_KEY callback function is guarded with rcu read lock.
In the function body, "kmalloc(size, GFP_USER | __GFP_NOWARN)" is used which may
sleep and violate rcu read lock region requirements. This patch fixed the issue
by using GFP_ATOMIC instead to avoid blocking kmalloc. Tested with
CONFIG_DEBUG_ATOMIC_SLEEP=y as suggested by Eric Dumazet.

Fixes: b471f2f1 ("bpf: implement MAP_GET_NEXT_KEY command for LPM_TRIE map")
Signed-off-by: Yonghong Song <yhs@fb.com>
Reported-by: syzbot <syzkaller@googlegroups.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>

kernel/bpf/lpm_trie.c

@@ -624,7 +624,7 @@ static int trie_get_next_key(struct bpf_map *map, void *_key, void *_next_key)
 	}
 
 	node_stack = kmalloc(trie->max_prefixlen * sizeof(struct lpm_trie_node *),
-			     GFP_USER | __GFP_NOWARN);
+			     GFP_ATOMIC | __GFP_NOWARN);
 	if (!node_stack)
 		return -ENOMEM;