Le mercredi 12 janvier 2011 à 13:23 +0000, Jarek Poplawski a écrit :

On Wed, Jan 12, 2011 at 02:07:12PM +0100, Eric Dumazet wrote:
...

quoted

Hmm... 

So all net devices in gen_ndis_query_resp() should have a
ndo_get_stats() or ndo_get_stats64() method, not allowed to use
spin_lock_bh() / spin_unlock_bh()

If yes, we should add big fat comments to pn_net_stats()/eth_get_stats() so that nobody
tries to revert your patch ;)

I'm not sure I got your point: my patch could be replaced with
ndo_get_stats64() implementing irq safe locking or by changing
gen_ndis_query_resp() calling context. It's intended as a fast
(compatible) fix.

I was mentioning that we tried in past months to remove useless
ndo_get_stats() methods that were only doing :

return &net->stats;

random commit : b27d50a9ff5cf2775b7a4daf5

Another possibility would be to use u64_stats_sync.h for these txq
counters.

(no locking needed to read counters, only a seqcount fetch/retry)

As a bonus, no overhead on 64bit arches.