Re: [PATCH bpf v2 2/4] bpf: map_poke_descriptor is being called with an unstable poke_tab[]

[John Fastabend <john.fastabend@gmail.com>]

Alexei Starovoitov wrote:
> On Wed, Jun 16, 2021 at 03:55:19PM -0700, John Fastabend wrote:
> > When populating poke_tab[] of a subprog we call map_poke_track() after
> > doing bpf_jit_add_poke_descriptor(). But, bpf_jit_add_poke_descriptor()
> > may, likely will, realloc the poke_tab[] structure and free the old
> > one. So that prog->aux->poke_tab is not stable. However, the aux pointer
> > is referenced from bpf_array_aux and poke_tab[] is used to 'track'
> > prog<->map link. This way when progs are released the entry in the
> > map is dropped and vice versa when the map is released we don't drop
> > it too soon if a prog is in the process of calling it.
> > 
> > I wasn't able to trigger any errors here, for example having map_poke_run
> > run with a poke_tab[] pointer that was free'd from
> > bpf_jit_add_poke_descriptor(), but it looks possible and at very least
> > is very fragile.
> > 
> > This patch moves poke_track call out of loop that is calling add_poke
> > so that we only ever add stable aux->poke_tab pointers to the map's
> > bpf_array_aux struct. Further, we need this in the next patch to fix
> > a real bug where progs are not 'untracked'.
> > 
> > Signed-off-by: John Fastabend <john.fastabend@gmail.com>
> > ---
> >  kernel/bpf/verifier.c |    6 +++++-
> >  1 file changed, 5 insertions(+), 1 deletion(-)
> > 
> > diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c
> > index 6e2ebcb0d66f..066fac9b5460 100644
> > --- a/kernel/bpf/verifier.c
> > +++ b/kernel/bpf/verifier.c
> > @@ -12126,8 +12126,12 @@ static int jit_subprogs(struct bpf_verifier_env *env)
> >  			}
> >  
> >  			func[i]->insnsi[insn_idx - subprog_start].imm = ret + 1;
> > +		}
> >  
> > -			map_ptr = func[i]->aux->poke_tab[ret].tail_call.map;
> > +		for (j = 0; j < func[i]->aux->size_poke_tab; j++) {
> > +			int ret;
> > +
> > +			map_ptr = func[i]->aux->poke_tab[j].tail_call.map;
> 
> I don't see why it's necessary.

Agree its not necessary. Nothing else can get at poke_tab while we do
the realloc so I agree its fine as is. It still seems odd to me to do the
poke_track when we know we are about to do multiple reallocs in the
next round of the loop. Either way I'll reply on the feedback in 3/4 and
we can avoid this patch altogether I think.

> poke_tab pointer will be re-read after bpf_jit_add_poke_descriptor().
> The compiler is not allowed to cache it.
> 
> I've applied the patch 1.