Re: [PATCH bpf-next] bpf: add support for other map types to bpf_map_lookup_and_delete_elem

[Andrii Nakryiko <andrii.nakryiko@gmail.com>]

On Thu, Sep 17, 2020 at 7:16 AM Luka Oreskovic
<luka.oreskovic@sartura.hr> wrote:
>
> Since this function already exists, it made sense to implement it for
> map types other than stack and queue. This patch adds the necessary parts
> from bpf_map_lookup_elem and bpf_map_delete_elem so it works as expected
> for all map types.
>
> Signed-off-by: Luka Oreskovic <luka.oreskovic@sartura.hr>
> CC: Juraj Vijtiuk <juraj.vijtiuk@sartura.hr>
> CC: Luka Perkov <luka.perkov@sartura.hr>
> ---
>  kernel/bpf/syscall.c | 30 ++++++++++++++++++++++++++++--
>  1 file changed, 28 insertions(+), 2 deletions(-)
>
> diff --git a/kernel/bpf/syscall.c b/kernel/bpf/syscall.c
> index 2ce32cad5c8e..955de6ca8c45 100644
> --- a/kernel/bpf/syscall.c
> +++ b/kernel/bpf/syscall.c
> @@ -1475,6 +1475,9 @@ static int map_lookup_and_delete_elem(union bpf_attr *attr)
>         if (CHECK_ATTR(BPF_MAP_LOOKUP_AND_DELETE_ELEM))
>                 return -EINVAL;
>
> +       if (attr->flags & ~BPF_F_LOCK)

If you want to use attr->flags, you need to update
BPF_MAP_LOOKUP_AND_DELETE_ELEM_LAST_FIELD few lines above. And every
new feature needs to come with selftests, so please check
tools/testing/selftests/bpf and latest patch sets adding new selftests
to see how it's done.

> +               return -EINVAL;
> +
>         f = fdget(ufd);
>         map = __bpf_map_get(f);
>         if (IS_ERR(map))
> @@ -1485,13 +1488,19 @@ static int map_lookup_and_delete_elem(union bpf_attr *attr)
>                 goto err_put;
>         }
>
> +       if ((attr->flags & BPF_F_LOCK) &&
> +           !map_value_has_spin_lock(map)) {
> +               err = -EINVAL;
> +               goto err_put;
> +       }
> +
>         key = __bpf_copy_key(ukey, map->key_size);
>         if (IS_ERR(key)) {
>                 err = PTR_ERR(key);
>                 goto err_put;
>         }
>
> -       value_size = map->value_size;
> +       value_size = bpf_map_value_size(map);
>
>         err = -ENOMEM;
>         value = kmalloc(value_size, GFP_USER | __GFP_NOWARN);
> @@ -1502,7 +1511,24 @@ static int map_lookup_and_delete_elem(union bpf_attr *attr)
>             map->map_type == BPF_MAP_TYPE_STACK) {
>                 err = map->ops->map_pop_elem(map, value);
>         } else {
> -               err = -ENOTSUPP;
> +               err = bpf_map_copy_value(map, key, value, attr->flags);
> +               if (err)
> +                       goto free_value;
> +
> +               if (bpf_map_is_dev_bound(map)) {
> +                       err = bpf_map_offload_delete_elem(map, key);
> +               } else if (IS_FD_PROG_ARRAY(map) ||
> +                          map->map_type == BPF_MAP_TYPE_STRUCT_OPS) {
> +                       /* These maps require sleepable context */
> +                       err = map->ops->map_delete_elem(map, key);
> +               } else {
> +                       bpf_disable_instrumentation();
> +                       rcu_read_lock();
> +                       err = map->ops->map_delete_elem(map, key);
> +                       rcu_read_unlock();
> +                       bpf_enable_instrumentation();
> +                       maybe_wait_bpf_programs(map);
> +               }

The whole point of this operation is to do lookup and deletion of
elements atomically. You can't do it with a separate lookup, followed
by a separate delete operation. Those two have to be implemented by
each type of map specifically. E.g., for hashmap, you'd have a
separate function implementation that takes a bucket lock, copies
data, and deletes entry, while still holding the lock. Of course
internally you'd want to reuse as much code as possible, but it will
be a separate bpf_map_ops operation.

>         }
>
>         if (err)
> --
> 2.26.2
>