Re: [PATCH net-next-2.6] net: Introduce u64_stats_sync infrastructure

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

Le mardi 15 juin 2010 à 20:25 +1000, Nick Piggin a écrit :
> On Tue, Jun 15, 2010 at 12:14:16PM +0200, Eric Dumazet wrote:
> > Here is the followup patch to abstract things a bit, before upcoming
> > conversions.
> > 
> > Thanks !
> > 
> > [PATCH net-next-2.6] net: Introduce u64_stats_sync infrastructure
> > 
> > To properly implement 64bits network statistics on 32bit or 64bit hosts,
> > we provide one new type and four methods, to ease conversions.
> > 
> > Stats producer should use following template granted it already got an
> > exclusive access to counters (a previous lock is taken, or per cpu
> > data [used in a non preemptable context])
> > 
> > Let me repeat : stats producers must be serialized by other means before
> > using this template. Preemption must be disabled too.
> > 
> > u64_stats_update_begin(&stats->syncp);
> > stats->bytes += len;
> > stats->packets++;
> > u64_stats_update_end(&stats->syncp);
> > 
> > While a consumer should use following template to get consistent
> > snapshot :
> > 
> > u64 tbytes, tpackets;
> > unsigned int start;
> > 
> > do {
> > 	start = u64_stats_fetch_begin(&stats->syncp);
> > 	tbytes = stats->bytes;
> > 	tpackets = stats->packets;
> > } while (u64_stats_fetch_retry(&stats->lock, syncp));
> > 
> > This patch uses this infrastructure in net loopback driver, instead of
> > specific one added in commit 6b10de38f0ef (loopback: Implement 64bit
> > stats on 32bit arches)
> > 
> > Suggested by David Miller
> 
> Cool, I don't mind this, but perhaps could you add some comments
> because it _will_ either be misused or copied and misused elsewhere :)
> 
> Callers must:
> - write side must ensure mutual exclusion (even if it was previously OK
>   to have lost updates on the writer side, the seqlock will explodde if
>   it is taken concurrently for write)
> - write side must not sleep
> - readside and writeside must have local-CPU exclusion from one another;
>   preempt, irq, bh as appropriate
> - will only protect 64-bit sizes from tearing -- eg updating 2 different
>   stats under the same write side will not ensure they are both seen in
>   the same read side

Hmm, I am not sure I got this one, could you please give me a buggy
example ?

> 
> But I do like the minimal design.

Thanks !

I'll submit a v2 patch after my lunch to add all your comments, because
all clarifications are indeed very very welcomed !