Re: Bug in net/ipv6/ip6_fib.c:fib6_dump_table()

[Gao feng <gaofeng@cn.fujitsu.com>]

于 2012年06月22日 05:50, Alexey Kuznetsov 写道:
> On Thu, Jun 21, 2012 at 10:27:49PM +0200, Eric Dumazet wrote:
>> Looking at this code, it lacks proper synchronization
>> between tree updaters and tree walkers.
>>
>> fib6_walker_lock rwlock is not enough to prevent races.
> 
> Hmm. As author of this weird code, I must say I honestly believed it was correct.
> At least I tried. :-)
> 
> 
> What's about 2bec5a336.., it does not look reasonable.
> The idea was that when you change tree, you fixup sleeping walkers, moving
> their location in tree to correct point. So, walkers must not have any stale pointers
> at any times (except when you under table write lock) and no skips/counts are required.
> I remember how damn difficult was it to make this right (well, sorry, if it is not yet :-)),
> so that understand that if some update is forgotten or done incorrectly, it is not so easy to find,
> but it is definitely worth of efforts.


Actually, I spent two months to try to reproduce this crash four months ago,
But finally I give up, I don't think there was any stale pointers,
we already correct it when we change the tree.