[PATCH net-next 2/2] net: initialize hashrnd in flow_dissector with net_get_random_once

Subsystems: networking [general], the rest

STALE4668d

5 messages, 3 authors, 2013-10-25 · open the first message on its own page

[PATCH net-next 2/2] net: initialize hashrnd in flow_dissector with net_get_random_once

From: Hannes Frederic Sowa <hidden>
Date: 2013-10-23 11:12:20

We also can defer the initialization of hashrnd in flow_dissector
to its first use. Since net_get_random_once is irqsave now we don't
have to audit the call paths if one of this functions get called by an
interrupt handler.

Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Signed-off-by: Hannes Frederic Sowa <redacted>
---
 net/core/flow_dissector.c | 34 +++++++++++++++++++++-------------
 1 file changed, 21 insertions(+), 13 deletions(-)
diff --git a/net/core/flow_dissector.c b/net/core/flow_dissector.c
index f8e25ac..5cac36e 100644
--- a/net/core/flow_dissector.c
+++ b/net/core/flow_dissector.c
@@ -184,6 +184,22 @@ ipv6:
 EXPORT_SYMBOL(skb_flow_dissect);
 
 static u32 hashrnd __read_mostly;
+static __always_inline void __flow_hash_secret_init(void)
+{
+	net_get_random_once(&hashrnd, sizeof(hashrnd));
+}
+
+static __always_inline u32 __flow_hash_3words(u32 a, u32 b, u32 c)
+{
+	__flow_hash_secret_init();
+	return jhash_3words(a, b, c, hashrnd);
+}
+
+static __always_inline u32 __flow_hash_1word(u32 a)
+{
+	__flow_hash_secret_init();
+	return jhash_1word(a, hashrnd);
+}
 
 /*
  * __skb_get_rxhash: calculate a flow hash based on src/dst addresses
@@ -210,9 +226,9 @@ void __skb_get_rxhash(struct sk_buff *skb)
 		swap(keys.port16[0], keys.port16[1]);
 	}
 
-	hash = jhash_3words((__force u32)keys.dst,
-			    (__force u32)keys.src,
-			    (__force u32)keys.ports, hashrnd);
+	hash = __flow_hash_3words((__force u32)keys.dst,
+				  (__force u32)keys.src,
+				  (__force u32)keys.ports);
 	if (!hash)
 		hash = 1;
 
@@ -248,7 +264,7 @@ u16 __skb_tx_hash(const struct net_device *dev, const struct sk_buff *skb,
 		hash = skb->sk->sk_hash;
 	else
 		hash = (__force u16) skb->protocol;
-	hash = jhash_1word(hash, hashrnd);
+	hash = __flow_hash_1word(hash);
 
 	return (u16) (((u64) hash * qcount) >> 32) + qoffset;
 }
@@ -340,7 +356,7 @@ static inline int get_xps_queue(struct net_device *dev, struct sk_buff *skb)
 				else
 					hash = (__force u16) skb->protocol ^
 					    skb->rxhash;
-				hash = jhash_1word(hash, hashrnd);
+				hash = __flow_hash_1word(hash);
 				queue_index = map->queues[
 				    ((u64)hash * map->len) >> 32];
 			}
@@ -395,11 +411,3 @@ struct netdev_queue *netdev_pick_tx(struct net_device *dev,
 	skb_set_queue_mapping(skb, queue_index);
 	return netdev_get_tx_queue(dev, queue_index);
 }
-
-static int __init initialize_hashrnd(void)
-{
-	get_random_bytes(&hashrnd, sizeof(hashrnd));
-	return 0;
-}
-
-late_initcall_sync(initialize_hashrnd);
-- 
1.8.3.1

Re: [PATCH net-next 2/2] net: initialize hashrnd in flow_dissector with net_get_random_once

From: Eric Dumazet <hidden>
Date: 2013-10-23 13:58:38

On Wed, 2013-10-23 at 13:12 +0200, Hannes Frederic Sowa wrote:
We also can defer the initialization of hashrnd in flow_dissector
to its first use. Since net_get_random_once is irqsave now we don't
have to audit the call paths if one of this functions get called by an
interrupt handler.

Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Signed-off-by: Hannes Frederic Sowa <redacted>
---
This really works well if CONFIG_JUMP_LABEL=y

I am afraid some arches do not really have this.

Re: [PATCH net-next 2/2] net: initialize hashrnd in flow_dissector with net_get_random_once

From: Hannes Frederic Sowa <hidden>
Date: 2013-10-23 14:13:10

On Wed, Oct 23, 2013 at 06:58:34AM -0700, Eric Dumazet wrote:
On Wed, 2013-10-23 at 13:12 +0200, Hannes Frederic Sowa wrote:
quoted
We also can defer the initialization of hashrnd in flow_dissector
to its first use. Since net_get_random_once is irqsave now we don't
have to audit the call paths if one of this functions get called by an
interrupt handler.

Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Signed-off-by: Hannes Frederic Sowa <redacted>
---
This really works well if CONFIG_JUMP_LABEL=y

I am afraid some arches do not really have this.
This is the problem with all users of net_get_random_once.

If an architecture does not have JUMP_LABEL or gcc does not provide
CC_HAVE_ASM_GOTO we have an unlikely branch in the code path and branch
to the epilogue of the function to generate the hash for the first
time. After that it's "just" a likely atomic boolean test.

I guess it is not that important to have a good secret hash key here as
this is solely used to dispatch the packets to cpus and does not store
anything in tables.

Greetings,

  Hannes

[PATCH net-next v2 2/2] net: initialize hashrnd in flow_dissector with net_get_random_once

From: Hannes Frederic Sowa <hidden>
Date: 2013-10-23 18:06:01

We also can defer the initialization of hashrnd in flow_dissector
to its first use. Since net_get_random_once is irq safe now we don't
have to audit the call paths if one of this functions get called by an
interrupt handler.

Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Signed-off-by: Hannes Frederic Sowa <redacted>
---
v2: Reword the commit message only. It looked horribly.

 net/core/flow_dissector.c | 34 +++++++++++++++++++++-------------
 1 file changed, 21 insertions(+), 13 deletions(-)
diff --git a/net/core/flow_dissector.c b/net/core/flow_dissector.c
index f8e25ac..5cac36e 100644
--- a/net/core/flow_dissector.c
+++ b/net/core/flow_dissector.c
@@ -184,6 +184,22 @@ ipv6:
 EXPORT_SYMBOL(skb_flow_dissect);
 
 static u32 hashrnd __read_mostly;
+static __always_inline void __flow_hash_secret_init(void)
+{
+	net_get_random_once(&hashrnd, sizeof(hashrnd));
+}
+
+static __always_inline u32 __flow_hash_3words(u32 a, u32 b, u32 c)
+{
+	__flow_hash_secret_init();
+	return jhash_3words(a, b, c, hashrnd);
+}
+
+static __always_inline u32 __flow_hash_1word(u32 a)
+{
+	__flow_hash_secret_init();
+	return jhash_1word(a, hashrnd);
+}
 
 /*
  * __skb_get_rxhash: calculate a flow hash based on src/dst addresses
@@ -210,9 +226,9 @@ void __skb_get_rxhash(struct sk_buff *skb)
 		swap(keys.port16[0], keys.port16[1]);
 	}
 
-	hash = jhash_3words((__force u32)keys.dst,
-			    (__force u32)keys.src,
-			    (__force u32)keys.ports, hashrnd);
+	hash = __flow_hash_3words((__force u32)keys.dst,
+				  (__force u32)keys.src,
+				  (__force u32)keys.ports);
 	if (!hash)
 		hash = 1;
 
@@ -248,7 +264,7 @@ u16 __skb_tx_hash(const struct net_device *dev, const struct sk_buff *skb,
 		hash = skb->sk->sk_hash;
 	else
 		hash = (__force u16) skb->protocol;
-	hash = jhash_1word(hash, hashrnd);
+	hash = __flow_hash_1word(hash);
 
 	return (u16) (((u64) hash * qcount) >> 32) + qoffset;
 }
@@ -340,7 +356,7 @@ static inline int get_xps_queue(struct net_device *dev, struct sk_buff *skb)
 				else
 					hash = (__force u16) skb->protocol ^
 					    skb->rxhash;
-				hash = jhash_1word(hash, hashrnd);
+				hash = __flow_hash_1word(hash);
 				queue_index = map->queues[
 				    ((u64)hash * map->len) >> 32];
 			}
@@ -395,11 +411,3 @@ struct netdev_queue *netdev_pick_tx(struct net_device *dev,
 	skb_set_queue_mapping(skb, queue_index);
 	return netdev_get_tx_queue(dev, queue_index);
 }
-
-static int __init initialize_hashrnd(void)
-{
-	get_random_bytes(&hashrnd, sizeof(hashrnd));
-	return 0;
-}
-
-late_initcall_sync(initialize_hashrnd);
-- 
1.8.3.1

Re: [PATCH net-next 2/2] net: initialize hashrnd in flow_dissector with net_get_random_once

From: David Miller <davem@davemloft.net>
Date: 2013-10-25 23:04:31

From: Hannes Frederic Sowa <redacted>
Date: Wed, 23 Oct 2013 13:12:19 +0200
We also can defer the initialization of hashrnd in flow_dissector
to its first use. Since net_get_random_once is irqsave now we don't
have to audit the call paths if one of this functions get called by an
interrupt handler.

Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Signed-off-by: Hannes Frederic Sowa <redacted>
Applied.
Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help