Re: [PATCH net] bpf: cpumap: use GFP_KERNEL instead of GFP_ATOMIC in __cpu_map_entry_alloc()

[Michal Hocko <mhocko@kernel.org>]

On Wed 14-02-18 18:34:51, Jesper Dangaard Brouer wrote:
> On Wed, 14 Feb 2018 16:06:40 +0100
> Michal Hocko <mhocko@kernel.org> wrote:
> 
> > On Wed 14-02-18 22:17:34, Jason Wang wrote:
> > > There're several implications after commit 0bf7800f1799 ("ptr_ring:
> > > try vmalloc() when kmalloc() fails") with the using of vmalloc() since
> > > can't allow GFP_ATOMIC but mandate GFP_KERNEL. This will lead a WARN
> > > since cpumap try to call with GFP_ATOMIC. Fortunately, entry
> > > allocation of cpumap can only be done through syscall path which means
> > > GFP_ATOMIC is not necessary, so fixing this by replacing GFP_ATOMIC
> > > with GFP_KERNEL.  
> > 
> > map_update_elem does the following. Unless I am missing something and
> > the callback doesn't call cpu_map_update_elem there then we are in a
> > non-preemptible context there and GFP_WAIT would blow up.
> > 		rcu_read_lock();
> > 		err = map->ops->map_update_elem(map, key, value, attr->flags);
> > 		rcu_read_unlock();
> 
> Nope - you did miss something ;-)
> 
> You are looking at the wrong place.  Look at /kernel/bpf/syscall.c line 697.
> 
>  vim +697 kernel/bpf/syscall.c
>  [...]
>         } else if (map->map_type == BPF_MAP_TYPE_CPUMAP) {
>                 err = map->ops->map_update_elem(map, key, value, attr->flags);
>                 goto out;
>         }
> 
> You missed that map type BPF_MAP_TYPE_CPUMAP is special cased, and
> is moved outside rcu_read_{lock,unlock} (because it need to create some
> kthreads).
> 
> Further more the BPF-verifier disallow BPF programs runtime changing
> the BPF_MAP_TYPE_CPUMAP.  Right now, we disallow almost everything from
> the bpf-side (even reading the value):
> 
>  vim +2057 kernel/bpf/verifier.c

OK, thanks for the clarification. I am not familiar with the code at all
so I was merely looking at call sites and this one just hit my eyes.

-- 
Michal Hocko
SUSE Labs