Re: [PATCH net-next 3/6] bpf: bpf_progs stores all loaded programs

[Daniel Borkmann <daniel@iogearbox.net>]

[ -daniel@iogearbox.com (wrong address) ]

On 04/26/2017 08:24 PM, Hannes Frederic Sowa wrote:
> We later want to give users a quick dump of what is possible with procfs,
> so store a list of all currently loaded bpf programs. Later this list
> will be printed in procfs.
>
> Signed-off-by: Hannes Frederic Sowa <hannes@stressinduktion.org>
> ---
>   include/linux/filter.h |  4 ++--
>   kernel/bpf/core.c      | 51 +++++++++++++++++++++++---------------------------
>   kernel/bpf/syscall.c   |  4 ++--
>   3 files changed, 27 insertions(+), 32 deletions(-)
>
> diff --git a/include/linux/filter.h b/include/linux/filter.h
> index 9a7786db14fa53..63624c619e371b 100644
> --- a/include/linux/filter.h
> +++ b/include/linux/filter.h
> @@ -753,8 +753,8 @@ bpf_address_lookup(unsigned long addr, unsigned long *size,
>   	return ret;
>   }
>
> -void bpf_prog_kallsyms_add(struct bpf_prog *fp);
> -void bpf_prog_kallsyms_del(struct bpf_prog *fp);
> +void bpf_prog_link(struct bpf_prog *fp);
> +void bpf_prog_unlink(struct bpf_prog *fp);
>
>   #else /* CONFIG_BPF_JIT */
>
> diff --git a/kernel/bpf/core.c b/kernel/bpf/core.c
> index 043f634ff58d87..2139118258cdf8 100644
> --- a/kernel/bpf/core.c
> +++ b/kernel/bpf/core.c
> @@ -365,22 +365,6 @@ static struct latch_tree_root bpf_tree __cacheline_aligned;
>
>   int bpf_jit_kallsyms __read_mostly;
>
> -static void bpf_prog_ksym_node_add(struct bpf_prog_aux *aux)
> -{
> -	WARN_ON_ONCE(!list_empty(&aux->bpf_progs_head));
> -	list_add_tail_rcu(&aux->bpf_progs_head, &bpf_progs);
> -	latch_tree_insert(&aux->ksym_tnode, &bpf_tree, &bpf_tree_ops);
> -}
> -
> -static void bpf_prog_ksym_node_del(struct bpf_prog_aux *aux)
> -{
> -	if (list_empty(&aux->bpf_progs_head))
> -		return;
> -
> -	latch_tree_erase(&aux->ksym_tnode, &bpf_tree, &bpf_tree_ops);
> -	list_del_rcu(&aux->bpf_progs_head);
> -}
> -
>   static bool bpf_prog_kallsyms_candidate(const struct bpf_prog *fp)
>   {
>   	return fp->jited && !bpf_prog_was_classic(fp);
> @@ -392,38 +376,45 @@ static bool bpf_prog_kallsyms_verify_off(const struct bpf_prog *fp)
>   	       fp->aux->bpf_progs_head.prev == LIST_POISON2;
>   }
>
> -void bpf_prog_kallsyms_add(struct bpf_prog *fp)
> +void bpf_prog_link(struct bpf_prog *fp)
>   {
> -	if (!bpf_prog_kallsyms_candidate(fp) ||
> -	    !capable(CAP_SYS_ADMIN))
> -		return;
> +	struct bpf_prog_aux *aux = fp->aux;
>
>   	spin_lock_bh(&bpf_lock);
> -	bpf_prog_ksym_node_add(fp->aux);
> +	list_add_tail_rcu(&aux->bpf_progs_head, &bpf_progs);
> +	if (bpf_prog_kallsyms_candidate(fp))
> +		latch_tree_insert(&aux->ksym_tnode, &bpf_tree, &bpf_tree_ops);

Hmm, this has the side-effect that it will hook up all progs
to kallsyms (I left out !capable(CAP_SYS_ADMIN) intentionally).

>   	spin_unlock_bh(&bpf_lock);
>   }
>
> -void bpf_prog_kallsyms_del(struct bpf_prog *fp)
> +void bpf_prog_unlink(struct bpf_prog *fp)
>   {
> -	if (!bpf_prog_kallsyms_candidate(fp))
> -		return;
> +	struct bpf_prog_aux *aux = fp->aux;
>
>   	spin_lock_bh(&bpf_lock);
> -	bpf_prog_ksym_node_del(fp->aux);
> +	list_del_rcu(&aux->bpf_progs_head);
> +	if (bpf_prog_kallsyms_candidate(fp))
> +		latch_tree_erase(&aux->ksym_tnode, &bpf_tree, &bpf_tree_ops);
>   	spin_unlock_bh(&bpf_lock);
>   }
>
>   static struct bpf_prog *bpf_prog_kallsyms_find(unsigned long addr)
>   {
>   	struct latch_tree_node *n;
> +	struct bpf_prog *prog;
>
>   	if (!bpf_jit_kallsyms_enabled())
>   		return NULL;
>
>   	n = latch_tree_find((void *)addr, &bpf_tree, &bpf_tree_ops);
> -	return n ?
> -	       container_of(n, struct bpf_prog_aux, ksym_tnode)->prog :
> -	       NULL;
> +	if (!n)
> +		return NULL;
> +
> +	prog = container_of(n, struct bpf_prog_aux, ksym_tnode)->prog;
> +	if (!prog->priv_cap_sys_admin)

Where is this bit defined?

If we return NULL on them anyway, why adding them to the tree
in the first place, just wastes resources on the traversal?

> +		return NULL;
> +
> +	return prog;
>   }
>
>   const char *__bpf_address_lookup(unsigned long addr, unsigned long *size,
> @@ -474,6 +465,10 @@ int bpf_get_kallsym(unsigned int symnum, unsigned long *value, char *type,
>
>   	rcu_read_lock();
>   	list_for_each_entry_rcu(aux, &bpf_progs, bpf_progs_head) {
> +		if (!bpf_prog_kallsyms_candidate(aux->prog) ||
> +		    !aux->prog->priv_cap_sys_admin)

Same here.

> +			continue;
> +
>   		if (it++ != symnum)
>   			continue;
>
> diff --git a/kernel/bpf/syscall.c b/kernel/bpf/syscall.c
> index 13642c73dca0b4..d61d1bd3e6fee6 100644
> --- a/kernel/bpf/syscall.c
> +++ b/kernel/bpf/syscall.c
> @@ -664,7 +664,7 @@ void bpf_prog_put(struct bpf_prog *prog)
>   {
>   	if (atomic_dec_and_test(&prog->aux->refcnt)) {
>   		trace_bpf_prog_put_rcu(prog);
> -		bpf_prog_kallsyms_del(prog);
> +		bpf_prog_unlink(prog);
>   		call_rcu(&prog->aux->rcu, __bpf_prog_put_rcu);
>   	}
>   }
> @@ -858,7 +858,7 @@ static int bpf_prog_load(union bpf_attr *attr)
>   		/* failed to allocate fd */
>   		goto free_used_maps;
>
> -	bpf_prog_kallsyms_add(prog);
> +	bpf_prog_link(prog);
>   	trace_bpf_prog_load(prog, err);
>   	return err;
>
>