netdev - Re: [PATCH bpf-next v2 1/2] xsk: store struct xdp_sock as a flexible array member of the XSKMAP

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <CAJ+HfNjDzNg9wdNkhx7BVkK5Udd3_WP0UMT8jTyssd254M6NsQ@mail.gmail.com>
Date:   Tue, 29 Oct 2019 07:20:55 +0100
From:   Björn Töpel <bjorn.topel@...il.com>
To:     Jakub Kicinski <jakub.kicinski@...ronome.com>
Cc:     Netdev <netdev@...r.kernel.org>,
        Alexei Starovoitov <ast@...nel.org>,
        Daniel Borkmann <daniel@...earbox.net>,
        Björn Töpel <bjorn.topel@...el.com>,
        bpf <bpf@...r.kernel.org>,
        Magnus Karlsson <magnus.karlsson@...il.com>,
        "Karlsson, Magnus" <magnus.karlsson@...el.com>,
        "Fijalkowski, Maciej" <maciej.fijalkowski@...el.com>,
        Toke Høiland-Jørgensen <toke@...hat.com>
Subject: Re: [PATCH bpf-next v2 1/2] xsk: store struct xdp_sock as a flexible
 array member of the XSKMAP

On Mon, 28 Oct 2019 at 23:26, Jakub Kicinski
<jakub.kicinski@...ronome.com> wrote:
>
> On Mon, 28 Oct 2019 23:11:50 +0100, Björn Töpel wrote:
> > On Mon, 28 Oct 2019 at 18:55, Jakub Kicinski
> > <jakub.kicinski@...ronome.com> wrote:
> > >
> > > On Fri, 25 Oct 2019 09:18:40 +0200, Björn Töpel wrote:
> > > > From: Björn Töpel <bjorn.topel@...el.com>
> > > >
> > > > Prior this commit, the array storing XDP socket instances were stored
> > > > in a separate allocated array of the XSKMAP. Now, we store the sockets
> > > > as a flexible array member in a similar fashion as the arraymap. Doing
> > > > so, we do less pointer chasing in the lookup.
> > > >
> > > > Signed-off-by: Björn Töpel <bjorn.topel@...el.com>
> > >
> > > Thanks for the re-spin.
> > >
> > > > diff --git a/kernel/bpf/xskmap.c b/kernel/bpf/xskmap.c
> > > > index 82a1ffe15dfa..a83e92fe2971 100644
> > > > --- a/kernel/bpf/xskmap.c
> > > > +++ b/kernel/bpf/xskmap.c
> > >
> > > > @@ -92,44 +93,35 @@ static struct bpf_map *xsk_map_alloc(union bpf_attr *attr)
> > > >           attr->map_flags & ~(BPF_F_NUMA_NODE | BPF_F_RDONLY | BPF_F_WRONLY))
> > > >               return ERR_PTR(-EINVAL);
> > > >
> > > > -     m = kzalloc(sizeof(*m), GFP_USER);
> > > > -     if (!m)
> > > > +     numa_node = bpf_map_attr_numa_node(attr);
> > > > +     size = struct_size(m, xsk_map, attr->max_entries);
> > > > +     cost = size + array_size(sizeof(*m->flush_list), num_possible_cpus());
> > >
> > > Now we didn't use array_size() previously because the sum here may
> > > overflow.
> > >
> > > We could use __ab_c_size() here, the name is probably too ugly to use
> > > directly and IDK what we'd have to name such a accumulation helper...
> > >
> > > So maybe just make cost and size a u64 and we should be in the clear.
> > >
> >
> > Hmm, but both:
> >   int bpf_map_charge_init(struct bpf_map_memory *mem, size_t size);
> >   void *bpf_map_area_alloc(size_t size, int numa_node);
> > pass size as size_t, so casting to u64 doesn't really help on 32-bit
> > systems, right?
>
> Yup :( IOW looks like the overflows will not be caught on 32bit
> machines in all existing code that does the (u64) cast. I hope
> I'm wrong there.
>
> > Wdyt about simply adding:
> >   if (cost < size)
> >     return ERR_PTR(-EINVAL)
> > after the cost calculation for explicit overflow checking?
>
> We'd need that for all users of these helpers. Could it perhaps makes
> sense to pass "alloc_size" and "extra_cost" as separate size_t to
> bpf_map_charge_init() and then we can do the overflow checking there,
> centrally?
>

The cost/size calculations seem to vary a bit from map to map, so I
don't know about the extra size_t arguments... but all of them do use
u64 for cost and explicit casting, in favor of u32 overflow checks.
Changing bpf_map_charge_init()/bpf_map_area_alloc() size to u64 would
be the smallest change, together with a 64-to-32 overflow check in
those functions.

> We can probably get rid of all the u64 casting too at that point,
> and use standard overflow helpers, yuppie :)
>

Yeah, that's the other path, but more churn (check_add_overflow() in every map).

Preferred path?

> > So, if size's struct_size overflows, the allocation will fail.
> > And we'll catch the cost overflow with the if-statement, no?
> >
> > Another option is changing the size_t in bpf_map_... to u64. Maybe
> > that's better, since arraymap and devmap uses u64 for cost/size.