Linux Ethernet Bridging - Mar 2017 - [Bridge] [PATCH 07/17] net: convert sock.sk_refcnt from atomic_t to refcount_t

On Mon, 2017-03-20 at 07:51 -0700, Eric Dumazet wrote:
> atomic_cmpxchg() on PowerPC is horribly more expensive because of the
> added two SYNC instructions.
Although I just saw that refcount was using atomic_cmpxchg_relaxed()

Time to find some documentation (probably missing) or get some specs for
this thing.

On Mon, 2017-03-20 at 07:59 -0700, Eric Dumazet wrote:
> On Mon, 2017-03-20 at 07:51 -0700, Eric Dumazet wrote:
> 
> > atomic_cmpxchg() on PowerPC is horribly more expensive because of the
> > added two SYNC instructions.
> 
> Although I just saw that refcount was using atomic_cmpxchg_relaxed()
> 
> Time to find some documentation (probably missing) or get some specs for
> this thing.
Interesting.

UDP ipv4 xmit path gets a ~25 % improvement on PPC with this patch.

( 20 concurrent netperf -t UDP_STREAM  : 2.45 Mpps -> 3.07 Mpps )

diff --git a/net/ipv4/route.c b/net/ipv4/route.c
index
8471dd116771462d149e1da2807e446b69b74bcc..9f14aebf0ae1f5f366cfff0fbf58c48603916bc7
100644
--- a/net/ipv4/route.c
+++ b/net/ipv4/route.c
@@ -497,14 +497,14 @@ u32 ip_idents_reserve(u32 hash, int segs)
 	u32 now = (u32)jiffies;
 	u32 new, delta = 0;
 
-	if (old != now && cmpxchg(p_tstamp, old, now) == old)
+	if (old != now && cmpxchg_relaxed(p_tstamp, old, now) == old)
 		delta = prandom_u32_max(now - old);
 
 	/* Do not use atomic_add_return() as it makes UBSAN unhappy */
 	do {
 		old = (u32)atomic_read(p_id);
 		new = old + delta + segs;
-	} while (atomic_cmpxchg(p_id, old, new) != old);
+	} while (atomic_cmpxchg_relaxed(p_id, old, new) != old);
 
 	return new - segs;
 }