Re: oops in tcp_v4_rcv.

[netdev added to cc list]

I think I understand now what causes the crash:
The tcp_ehash assumes that the entries are of the type 'struct inet_sock'.
But the actual entry is of the type tcp_tw_bucket. And 'sk->inet.daddr' 
is not shared between both structures.
<< net/ipv4/tcp_ipv4, line 510:
       /* Must check for a TIME_WAIT'er before going to listener hash. */
       for (sk = (head + tcp_ehash_size)->chain; sk; sk = sk->next)
              if (TCP_IPV4_MATCH(sk, acookie, saddr, daddr, ports, dif))
                   goto hit;
<<
preprocessor output:
<<
for (sk = (head + (tcp_hashinfo.__tcp_ehash_size))->chain; sk; sk = 
sk->next)
     if ((((&((struct inet_sock *)sk)->inet)->daddr == (saddr)) &&
          ((&((struct inet_sock *)sk)->inet)->rcv_saddr == (daddr)) &&
          ((*((__u32 *)&((&((struct inet_sock *)sk)->inet)->dport)))== 
(ports)) &&
          (!((sk)->bound_dev_if) || ((sk)->bound_dev_if == (dif)))))
                      goto hit;
<<


Manfred Spraul wrote:

> Hi,
>
> I'm looking at crashes that occur during network stress testing with 
> the CONFIG_DEBUG_PAGEALLOC from -mm: Pages that are not in use are 
> immediately unmapped from the linear mapping, and thus reading stale 
> pointer causes an immediate oops.
> I've now analyzed one crash:
> the oops is in __tcp_v4_lookup_established, in the 2nd look [i.e. 
> looking at TIME_WAIT sockets. Easy to identify due to the access to 
> __tcp_ehash_size].
> The entry in the hash table is an tcp_tw_bucket, and that structure is 
> only ~88 bytes long. The oops is caused by an access to objp+0x168, 
> which doesn't exist.