Re: KASAN poisoning for skb linear data

[Dmitry Vyukov]

On Thu, Mar 8, 2018 at 1:20 AM, Dmitry Vyukov <dvyukov@xxxxxxxxxx> wrote:
> On Mon, Jan 15, 2018 at 3:15 PM, Dmitry Vyukov <dvyukov@xxxxxxxxxx> wrote:
>> Hi,
>>
>> As far as I understand pskb_may_pull() plays important role in packet
>> parsing for all protocols. And we did custom fragmentation of packets
>> emitted via tun (IFF_NAPI_FRAGS). However, it seems that it does not
>> give any results (bugs found), and I think the reason for this is that
>> linear data is rounded up and is usually quite large. So if a parsing
>> function does pskb_may_pull(1), or does not do it at all, it can
>> usually access more and it will go unnoticed. KASAN has an ability to
>> do custom poisoning: it can poison/unpoison any memory range, and then
>> detect any reads/writes to that range. What do you think about adding
>> custom KASAN poisoning to pskb_may_pull() and switching it to
>> non-eager mode (pull only what was requested) under KASAN? Do you
>> think it has potential for finding important bugs? What amount of work
>> is this?
>
> Filed https://bugzilla.kernel.org/show_bug.cgi?id=199055 for this so
> it's not get lost.

Bringing this up after we discussed this with Dave on plumbers.

There are 2 strategies for making KASAN aware of exact skb linear
buffer semantics.
1. Just using kmalloc/free each time with precise size.
2. Using KASAN annotations:

void kasan_poison_shadow(const void *address, size_t size, u8 value);
void kasan_unpoison_shadow(const void *address, size_t size);

https://github.com/torvalds/linux/blob/master/mm/kasan/kasan.c#L57
https://github.com/torvalds/linux/blob/master/include/linux/kasan.h#L38

If we use annotations we can keep more of the existing skb logic.
But AFAIU this way we won't be able to detect all accesses after a
potential reallocation.

There are also annotations for explicit checks:
https://github.com/torvalds/linux/blob/master/include/linux/kasan-checks.h#L6
But not sure we need them here (maybe more appropriate for places
where KASAN does not see memory accesses e.g. a driver handing off a
packet to DMA).


I don't think it makes sense to make any more complex than necessary
in the name of performance, at least initially. This will be enabled
only under #ifdef KASAN.

If somebody gives us any prototype, we can assess (1) if it works and
(2) if it catches any new bugs.

Thanks