Re: Process Hang in __read_seqcount_begin

[Eric Dumazet <eric.dumazet@gmail.com>]

On Fri, 2012-10-26 at 11:51 -0700, Peter LaDow wrote:
> (I've added netfilter and linux-rt-users to try to pull in more help).
> 
> On Fri, Oct 26, 2012 at 9:48 AM, Eric Dumazet <eric.dumazet@gmail.com> wrote:
> > Upstream kernel is fine, there is no race, as long as :
> >
> > local_bh_disable() disables BH and preemption.
> 
> Looking at the unpatched code in net/ipv4/netfilter/ip_tables.c, it
> doesn't appear that any of the code checks the return value for
> xt_write_receq_begin to determine if it is safe to write.  And neither
> does the newly patched code.  How did the mainline code prevent
> corruption of the tables it is updating?
> 

Do you know what is per cpu data in linux kernel ?

> Why isn't there something like
> 
>   while ( (addend = xt_write_recseq_begin()) == 0 );
> 
> To make sure that only one person has write access to the tables?
> Better yet, why not use a seqlock_t instead?
> 

Because its not needed. Really I dont know why you want that.

Once you are sure a thread cannot be interrupted by a softirq, and
cannot migrate to another cpu, access to percpu data doesnt need other
synchronization at all.

Following sequence is safe :

addend = (__this_cpu_read(xt_recseq.sequence) + 1) & 1;
/*
 * This is kind of a write_seqcount_begin(), but addend is 0 or 1
 * We dont check addend value to avoid a test and conditional jump,
 * since addend is most likely 1
 */
__this_cpu_add(xt_recseq.sequence, addend);

Because any other thread will use a different percpu data.