Re: af_packet: use after free in prb_retire_rx_blk_timer_expired

[Ding Tianhong <dingtianhong@huawei.com>]

On 2017/7/23 3:02, Cong Wang wrote:
> Hello,
> 
> On Sat, Jul 22, 2017 at 2:55 AM, liujian (CE) <liujian56@huawei.com> wrote:
>> I also hit this issue with trinity test:
>>
>> The call trace:
>>   [exception RIP: prb_retire_rx_blk_timer_expired+70]
>>     RIP: ffffffff81633be6  RSP: ffff8801bec03dc0  RFLAGS: 00010246
>>     RAX: 0000000000000000  RBX: ffff8801b49d0948  RCX: 0000000000000000
>>     RDX: ffff8801b31057a0  RSI: a56b6b6b6b6b6b6b  RDI: ffff8801b49d09ec
>>     RBP: ffff8801bec03dd8   R8: 0000000000000001   R9: ffffffff83e1bf80
>>     R10: 0000000000000002  R11: 0000000000000005  R12: ffff8801b49d09ec
>>     R13: 0000000000000100  R14: ffffffff81633ba0  R15: ffff8801b49d0948
>>     ORIG_RAX: ffffffffffffffff  CS: 0010  SS: 0018
>>  #7 [ffff8801bec03de0] call_timer_fn at ffffffff8108cb76
>>  #8 [ffff8801bec03e18] run_timer_softirq at ffffffff8108f87c
>>  #9 [ffff8801bec03e90] __do_softirq at ffffffff8108629f
>> #10 [ffff8801bec03f00] call_softirq at ffffffff8166a01c
>> #11 [ffff8801bec03f18] do_softirq at ffffffff810172ad
>> #12 [ffff8801bec03f30] irq_exit at ffffffff81086655
>> #13 [ffff8801bec03f48] msa_irq_exit at ffffffff810b1ab3
>> #14 [ffff8801bec03f88] smp_apic_timer_interrupt at ffffffff8166aeae
>> #15 [ffff8801bec03fb0] apic_timer_interrupt at ffffffff816692dd
>> --- <IRQ stack> ---
>>
>> And from vmcore, I can see the pointer GET_CURR_PBLOCK_DESC_FROM_CORE(pkc); is a56b6b6b6b6b6b6b
>>
> 
> Does the following quick fix help?
> 
> 
> diff --git a/net/packet/af_packet.c b/net/packet/af_packet.c
> index 008bb34ee324..09ec1640e5f7 100644
> --- a/net/packet/af_packet.c
> +++ b/net/packet/af_packet.c
> @@ -4264,6 +4264,7 @@ static int packet_set_ring(struct sock *sk,
> union tpacket_req_u *req_u,
>                         /* Block transmit is not supported yet */
>                         if (!tx_ring) {
>                                 init_prb_bdqc(po, rb, pg_vec, req_u);
> +                               pg_vec = NULL;
>                         } else {
>                                 struct tpacket_req3 *req3 = &req_u->req3;
> 

Hi, Cong:

Thanks for your quirk solution, but I still has some doubts about it,
it looks like fix the problem in the packet_setsockopt->packet_set_ring processing,
but when in packet_release processing, it may could not release the
real pg_vec for the TPACKET_V3 ring, and then cause the mem leak,
maybe I miss something here, nice to hear from your feedback. :)

what about fix it this way:
--- a/net/packet/af_packet.c
+++ b/net/packet/af_packet.c
@@ -4335,9 +4335,13 @@ static int packet_set_ring(struct sock *sk, union tpacket_req_u *req_u,
                /* Because we don't support block-based V3 on tx-ring */
                if (!tx_ring)
                        prb_shutdown_retire_blk_timer(po, rb_queue);
+
+               if (pg_vec)
+                       free_pg_vec(pg_vec, order, req->tp_block_nr);
+
        }

-       if (pg_vec)
+       if (pg_vec && (po->tp_version < TPACKET_V3))
                free_pg_vec(pg_vec, order, req->tp_block_nr);
 out:
        release_sock(sk);


Regards
Ding

> .
>