On Thu, Nov 12, 2020 at 11:35 PM Daniel Borkmann <daniel@xxxxxxxxxxxxx> wrote:
>
> On 11/12/20 9:03 PM, KP Singh wrote:
> > From: KP Singh <kpsingh@xxxxxxxxxx>
> >
> > Update the set of sleepable hooks with the ones that do not trigger
> > a warning with might_fault() when exercised with the correct kernel
> > config options enabled, i.e.
[...]
>
> I think this is very useful info. I was wondering whether it would make sense
> to annotate these more closely to the code so there's less chance this info
> becomes stale? Maybe something like below, not sure ... issue is if you would
> just place a cant_sleep() in there it might be wrong since this should just
> document that it can be invoked from non-sleepable context but it might not
> have to.
Indeed, this is why I did not make an explicit cant_sleep() call for these hooks
in __bpf_prog_enter (with a change in the signature to pass struct *prog).
> diff --git a/security/security.c b/security/security.c
> index a28045dc9e7f..7899bf32cdaa 100644
> --- a/security/security.c
> +++ b/security/security.c
> @@ -94,6 +94,11 @@ static __initdata bool debug;
> pr_info(__VA_ARGS__); \
> } while (0)
>
> +/*
> + * Placeholder for now to document that hook implementation cannot sleep
> + * since it could potentially be called from non-sleepable context, too.
> + */
> +#define hook_cant_sleep() do { } while (0)
Good idea!
At the very least, we can update the comments in lsm_hooks.h
which already mention some of the LSM hooks as being called from
non-sleepable contexts.
I will remove this comment, send a separate patch to security folks
and respin these patches.
-KP
> +
> static bool __init is_enabled(struct lsm_info *lsm)
> {
> if (!lsm->enabled)
> @@ -2522,6 +2527,7 @@ void security_bpf_map_free(struct bpf_map *map)
> }
> void security_bpf_prog_free(struct bpf_prog_aux *aux)
> {
> + hook_cant_sleep();
> call_void_hook(bpf_prog_free_security, aux);
> }
> #endif /* CONFIG_BPF_SYSCALL */
