Re: [PATCH net-next 4/5] ipv4: route: dissect flow in input path if fib rules need it

On Mon, Feb 26, 2018 at 1:10 AM, Paolo Abeni [off-list ref] wrote:

On Sat, 2018-02-24 at 21:44 -0800, Roopa Prabhu wrote:

quoted

From: Roopa Prabhu <redacted>

Dissect flow in fwd path if fib rules require it. Controlled by
a flag to avoid penatly for the common case. Flag is set when fib
rules with sport, dport and proto match that require flow dissect
are installed. Also passes the dissected hash keys to the multipath
hash function when applicable to avoid dissecting the flow again.
icmp packets will continue to use inner header for hash
calculations (Thanks to Nikolay Aleksandrov for some review here).

Signed-off-by: Roopa Prabhu <redacted>
---
 include/net/ip_fib.h     |  2 +-
 include/net/netns/ipv4.h |  1 +
 net/ipv4/fib_rules.c     |  6 ++++++
 net/ipv4/fib_semantics.c |  2 +-
 net/ipv4/route.c         | 52 +++++++++++++++++++++++++++++++++++-------------
 5 files changed, 47 insertions(+), 16 deletions(-)

diff --git a/include/net/ip_fib.h b/include/net/ip_fib.h
index f805243..5ada772 100644
--- a/include/net/ip_fib.h
+++ b/include/net/ip_fib.h

@@ -371,7 +371,7 @@ int fib_sync_up(struct net_device *dev, unsigned int nh_flags);

 #ifdef CONFIG_IP_ROUTE_MULTIPATH
 int fib_multipath_hash(const struct fib_info *fi, const struct flowi4 *fl4,
-                    const struct sk_buff *skb);
+                    const struct sk_buff *skb, struct flow_keys *flkeys);
 #endif
 void fib_select_multipath(struct fib_result *res, int hash);
 void fib_select_path(struct net *net, struct fib_result *res,

diff --git a/include/net/netns/ipv4.h b/include/net/netns/ipv4.h
index 44668c2..87b8fdc 100644
--- a/include/net/netns/ipv4.h
+++ b/include/net/netns/ipv4.h

@@ -52,6 +52,7 @@ struct netns_ipv4 {
 #ifdef CONFIG_IP_MULTIPLE_TABLES
      struct fib_rules_ops    *rules_ops;
      bool                    fib_has_custom_rules;
+     bool                    fib_rules_require_fldissect;
      struct fib_table __rcu  *fib_main;
      struct fib_table __rcu  *fib_default;
 #endif

diff --git a/net/ipv4/fib_rules.c b/net/ipv4/fib_rules.c
index 9d55c90..83aa786 100644
--- a/net/ipv4/fib_rules.c
+++ b/net/ipv4/fib_rules.c

@@ -253,6 +253,11 @@ static int fib4_rule_configure(struct fib_rule *rule, struct sk_buff *skb,
      }
 #endif

+     if (rule->ip_proto ||
+         fib_rule_port_range_valid(&rule->sport_range) ||
+         fib_rule_port_range_valid(&rule->dport_range))
+             net->ipv4.fib_rules_require_fldissect = true;
+
      rule4->src_len = frh->src_len;
      rule4->srcmask = inet_make_mask(rule4->src_len);
      rule4->dst_len = frh->dst_len;

What about using 'fib_rules_require_fldissect' to conditionally avoid
all the tests introduced by patch 2/5 ? Perhaps even using a static key
for that?

yep, ack

It would be great if the kernel would be able to clear this flag when
no more needed. I know it's not the current behaviour for other similar
flags, but I hope we can improve ;)

yep agreed. I did lean on the existing fib rule flags behavior.
Nikolay suggested I can move this to a counter (which i had been
debating as well).
that gives the ability to clear on deletes.

Both points above also apply to the ipv6 code path.

ack.