On Wed, 04 Oct 2017 21:13:47 +0200, Daniel Borkmann wrote:
> On 10/04/2017 05:43 AM, Alexei Starovoitov wrote:
> > On Tue, Oct 03, 2017 at 08:24:06PM -0700, Eric Dumazet wrote:
> >> On Tue, 2017-10-03 at 19:52 -0700, Alexei Starovoitov wrote:
> >>
> >>> yep. looks great.
> >>> Please test it and submit officially :)
> >>> The commit aafe6ae9cee3 ("bpf: dynamically allocate digest scratch
> >>> buffer")
> >>> fixed the other case where we were relying on the above mutex.
> >>> The only other spot to be adjusted is to add spin_lock/mutex or DO_ONCE()
> >>> to
> >>> bpf_get_skb_set_tunnel_proto() to protect md_dst init.
> >>> imo that would be it.
> >>> Daniel, anything else comes to mind?
>
> Yes, this should be all. DO_ONCE() for the tunnel proto seems a
> good choice.
Hm. I actually did:
if (!dst) {
tmp = alloc();
if (!tmp)
return;
if (cmpxchg(&dst, NULL, tmp))
free(tmp);
}
I don't like how DO_ONCE() doesn't handle errors from the init
function :(
> >> 16 MB of log (unswappable kernel memory) per active checker.
> >>
> >> We might offer a way to oom hosts.
> >
> > right. good point!
> > we need to switch to continuous copy_to_user() after a page or so.
> > Can even do it after every vscnprintf()
> > but page at a time is probably faster.
>
> Also worst case upper limits on verification side for holding state
> aside from the log would need to be checked in terms of how much mem
> we end up holding that is not accounted against any process (and not
> really "rate-limited" anymore once we drop the mutex).
