On 16-09-12 07:20 PM, Eric Dumazet wrote:
On Mon, 2016-09-12 at 16:10 -0700, Eric Dumazet wrote:diff --git a/net/core/gen_stats.c b/net/core/gen_stats.c index 508e051304fb62627e61b5065b2325edd1b84f2e..dc9dd8ae7d5405f76c775278dac7689655b21041 100644 --- a/net/core/gen_stats.c +++ b/net/core/gen_stats.c @@ -142,10 +142,14 @@ __gnet_stats_copy_basic(const seqcount_t *running, return; } do { - if (running) + if (running) { + local_bh_disable(); seq = read_seqcount_begin(running); + } bstats->bytes = b->bytes; bstats->packets = b->packets; + if (running) + local_bh_enable(); } while (running && read_seqcount_retry(running, seq)); }Ah well, forget this patch, re-enabling bh right before read_seqcount_retry() is not going to help.
I have to say I have seen some odd issues once in a while reading generic action stats. I had a program that opened a netlink socket into the kernel. Every X seconds it does a dump of all the actions to read the stats. There is a very reproducible behavior that the stats are not in sync with the kernel. Given generic stats is lockless I thought maybe rcu or per-cpu stats was the issue. I havent had time to look closely. The solution is instead of keeping the socket open all the time; I open, read stats, close (repeat every x seconds). If there is something you want me to try - I could do sometimes this week. Your patch above may be useful! cheers, jamal
