net: avoid indirect calls in L4 checksum calculation
authorMatteo Croce <mcroce@redhat.com>
Wed, 29 May 2019 15:13:48 +0000 (17:13 +0200)
committerDavid S. Miller <davem@davemloft.net>
Thu, 30 May 2019 21:34:12 +0000 (14:34 -0700)
Commit 283c16a2dfd3 ("indirect call wrappers: helpers to speed-up
indirect calls of builtin") introduces some macros to avoid doing
indirect calls.

Use these helpers to remove two indirect calls in the L4 checksum
calculation for devices which don't have hardware support for it.

As a test I generate packets with pktgen out to a dummy interface
with HW checksumming disabled, to have the checksum calculated in
every sent packet.
The packet rate measured with an i7-6700K CPU and a single pktgen
thread raised from 6143 to 6608 Kpps, an increase by 7.5%

Suggested-by: Davide Caratti <dcaratti@redhat.com>
Signed-off-by: Matteo Croce <mcroce@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/core/skbuff.c

index e89be628269371f8f4e9f4b1ef9df06e31d26a6a..0c2e7d4946ef2bb7354356dc9b7917945eaf1731 100644 (file)
@@ -76,6 +76,7 @@
 #include <linux/highmem.h>
 #include <linux/capability.h>
 #include <linux/user_namespace.h>
+#include <linux/indirect_call_wrapper.h>
 
 #include "datagram.h"
 
@@ -2507,7 +2508,8 @@ __wsum __skb_checksum(const struct sk_buff *skb, int offset, int len,
        if (copy > 0) {
                if (copy > len)
                        copy = len;
-               csum = ops->update(skb->data + offset, copy, csum);
+               csum = INDIRECT_CALL_1(ops->update, csum_partial_ext,
+                                      skb->data + offset, copy, csum);
                if ((len -= copy) == 0)
                        return csum;
                offset += copy;
@@ -2534,9 +2536,13 @@ __wsum __skb_checksum(const struct sk_buff *skb, int offset, int len,
                                              frag->page_offset + offset - start,
                                              copy, p, p_off, p_len, copied) {
                                vaddr = kmap_atomic(p);
-                               csum2 = ops->update(vaddr + p_off, p_len, 0);
+                               csum2 = INDIRECT_CALL_1(ops->update,
+                                                       csum_partial_ext,
+                                                       vaddr + p_off, p_len, 0);
                                kunmap_atomic(vaddr);
-                               csum = ops->combine(csum, csum2, pos, p_len);
+                               csum = INDIRECT_CALL_1(ops->combine,
+                                                      csum_block_add_ext, csum,
+                                                      csum2, pos, p_len);
                                pos += p_len;
                        }
 
@@ -2559,7 +2565,8 @@ __wsum __skb_checksum(const struct sk_buff *skb, int offset, int len,
                                copy = len;
                        csum2 = __skb_checksum(frag_iter, offset - start,
                                               copy, 0, ops);
-                       csum = ops->combine(csum, csum2, pos, copy);
+                       csum = INDIRECT_CALL_1(ops->combine, csum_block_add_ext,
+                                              csum, csum2, pos, copy);
                        if ((len -= copy) == 0)
                                return csum;
                        offset += copy;