Le jeudi 17 mars 2011 à 10:18 -0500, Christoph Lameter a écrit :

On Thu, 17 Mar 2011, David Miller wrote:

quoted

I had been meaning to bring this up from another perspective.

In networking, we often only ever access objects in base or
BH context.  Therefore in BH context cases we can do just
normal counter bumps without any of the special atomic or
IRQ disabling code at all.

We have the __ functions for that purpose. __this_cpu_inc f.e. falls back
to a simply ++ operation if the arch cannot provide something better.
irqsafe_xx are only used if the context does not provide any protection
and if there is the potential of the counter being incremented from an
interrupt context.

What David and I have in mind is to use one array per mib instead of
two. This is an old idea.

https://patchwork.kernel.org/patch/15883/

When we know we run from BH context, we can use __this_cpu_inc(), but if
we dont know or run from user/process context, we would need irqsafe_inc
variant.

For x86 this maps to same single instruction, but for other arches, this
might be too expensive.

BTW, I think following patch is possible to save some text and useless
tests (on 64bit platform at least)

size vmlinux.old vmlinux

Thanks

[PATCH] snmp: SNMP_UPD_PO_STATS_BH() always called from softirq

We dont need to test if we run from softirq context.

Signed-off-by: Eric Dumazet <redacted>
Cc: Christoph Lameter <redacted>
---
 include/net/snmp.h |    4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/include/net/snmp.h b/include/net/snmp.h
index 762e2ab..be2424d 100644
--- a/include/net/snmp.h
+++ b/include/net/snmp.h

@@ -149,8 +149,8 @@ struct linux_xfrm_mib {
 	} while (0)
 #define SNMP_UPD_PO_STATS_BH(mib, basefield, addend)	\
 	do { \
-		__typeof__(*mib[0]) *ptr = \
-			__this_cpu_ptr((mib)[!in_softirq()]); \
+		__typeof__(*mib[0]) *ptr = __this_cpu_ptr((mib)[0]);	\
+									\
 		ptr->mibs[basefield##PKTS]++; \
 		ptr->mibs[basefield##OCTETS] += addend;\
 	} while (0)