Skip to contentSkip to conversation

rps: misc changes

PATCH Applied Eric Dumazet 3 participants 10 messages Apr 7, 2025 – Apr 8, 2025

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

cover rps: misc changes

Minor changes in rps:

skb_flow_limit() is probably unused these days,
and data-races are quite theoretical.

Eric Dumazet (4):
net: rps: change skb_flow_limit() hash function
net: rps: annotate data-races around (struct sd_flow_limit)->count
net: add data-race annotations in softnet_seq_show()
net: rps: remove kfree_rcu_mightsleep() use

 include/net/rps.h          |  5 +++--
 net/core/dev.c             | 11 +++++++----
 net/core/dev.h             |  5 +++--
 net/core/net-procfs.c      |  9 +++++----
 net/core/sysctl_net_core.c |  6 +++---
 5 files changed, 21 insertions(+), 15 deletions(-)
Signature · 1 line
2.49.0.504.g3bcea36a83-goog
#1 lore raw

PATCH 1/4 net: rps: change skb_flow_limit() hash function

As explained in commit f3483c8e1da6 ("net: rfs: hash function change"), masking low order bits of skb_get_hash(skb) has low entropy.

A NIC with 32 RX queues uses the 5 low order bits of rss key to select a queue. This means all packets landing to a given queue share the same 5 low order bits.

Switch to hash_32() to reduce hash collisions.

---
 net/core/dev.c             | 2 +-
 net/core/dev.h             | 2 +-
 net/core/sysctl_net_core.c | 2 +-
 3 files changed, 3 insertions(+), 3 deletions(-)
net/core/dev.c +1 -1
index 0608605cfc24..f674236f34be 100644
@@ -5038,7 +5038,7 @@ static bool skb_flow_limit(struct sk_buff *skb, unsigned int qlen)
 	rcu_read_lock();
 	fl = rcu_dereference(sd->flow_limit);
 	if (fl) {
-		new_flow = skb_get_hash(skb) & (fl->num_buckets - 1);
+		new_flow = hash_32(skb_get_hash(skb), fl->log_buckets);
 		old_flow = fl->history[fl->history_head];
 		fl->history[fl->history_head] = new_flow;
 
net/core/dev.h +1 -1
index 7ee203395d8e..b1cd44b5f009 100644
@@ -16,7 +16,7 @@ struct cpumask;
 #define FLOW_LIMIT_HISTORY	(1 << 7)  /* must be ^2 and !overflow buckets */
 struct sd_flow_limit {
 	u64			count;
-	unsigned int		num_buckets;
+	u8			log_buckets;
 	unsigned int		history_head;
 	u16			history[FLOW_LIMIT_HISTORY];
 	u8			buckets[];
net/core/sysctl_net_core.c +1 -1
index c7769ee0d9c5..5cfe76ede523 100644
@@ -248,7 +248,7 @@ static int flow_limit_cpu_sysctl(const struct ctl_table *table, int write,
 					ret = -ENOMEM;
 					goto write_unlock;
 				}
-				cur->num_buckets = netdev_flow_limit_table_len;
+				cur->log_buckets = ilog2(netdev_flow_limit_table_len);
 				rcu_assign_pointer(sd->flow_limit, cur);
 			}
 		}
Signature · 1 line
2.49.0.504.g3bcea36a83-goog
Reviewed-by on v1 1/4
#2 lore raw

Re: [PATCH net-next 1/4] net: rps: change skb_flow_limit() hash function

Quoted 10 lines · Eric Dumazet

As explained in commit f3483c8e1da6 ("net: rfs: hash function change"), masking low order bits of skb_get_hash(skb) has low entropy.

A NIC with 32 RX queues uses the 5 low order bits of rss key to select a queue. This means all packets landing to a given queue share the same 5 low order bits.

Switch to hash_32() to reduce hash collisions.

#3 lore raw

PATCH 2/4 net: rps: annotate data-races around (struct sd_flow_limit)->count

In reply to Eric Dumazet

softnet_seq_show() can read fl->count while another cpu updates this field from skb_flow_limit().

Make this field an 'unsigned int', as its only consumer only deals with 32 bit.

---
 net/core/dev.c        | 3 ++-
 net/core/dev.h        | 2 +-
 net/core/net-procfs.c | 3 ++-
 3 files changed, 5 insertions(+), 3 deletions(-)
net/core/dev.c +2 -1
index f674236f34be..969883173182 100644
@@ -5049,7 +5049,8 @@ static bool skb_flow_limit(struct sk_buff *skb, unsigned int qlen)
 			fl->buckets[old_flow]--;
 
 		if (++fl->buckets[new_flow] > (FLOW_LIMIT_HISTORY >> 1)) {
-			fl->count++;
+			/* Pairs with READ_ONCE() in softnet_seq_show() */
+			WRITE_ONCE(fl->count, fl->count + 1);
 			rcu_read_unlock();
 			return true;
 		}
net/core/dev.h +1 -1
index b1cd44b5f009..e855e1cb43fd 100644
@@ -15,7 +15,7 @@ struct cpumask;
 /* Random bits of netdevice that don't need to be exposed */
 #define FLOW_LIMIT_HISTORY	(1 << 7)  /* must be ^2 and !overflow buckets */
 struct sd_flow_limit {
-	u64			count;
+	unsigned int		count;
 	u8			log_buckets;
 	unsigned int		history_head;
 	u16			history[FLOW_LIMIT_HISTORY];
net/core/net-procfs.c +2 -1
index 3e92bf0f9060..69782d62fbe1 100644
@@ -132,8 +132,9 @@ static int softnet_seq_show(struct seq_file *seq, void *v)
 
 	rcu_read_lock();
 	fl = rcu_dereference(sd->flow_limit);
+	/* Pairs with WRITE_ONCE() in skb_flow_limit() */
 	if (fl)
-		flow_limit_count = fl->count;
+		flow_limit_count = READ_ONCE(fl->count);
 	rcu_read_unlock();
 #endif
 
Signature · 1 line
2.49.0.504.g3bcea36a83-goog
#5 lore raw

PATCH 3/4 net: add data-race annotations in softnet_seq_show()

In reply to Eric Dumazet

softnet_seq_show() reads several fields that might be updated concurrently. Add READ_ONCE() and WRITE_ONCE() annotations.

---
 net/core/dev.c        | 6 ++++--
 net/core/net-procfs.c | 6 +++---
 2 files changed, 7 insertions(+), 5 deletions(-)
net/core/dev.c +4 -2
index 969883173182..4ccc6dc5303e 100644
@@ -4953,7 +4953,8 @@ static void rps_trigger_softirq(void *data)
 	struct softnet_data *sd = data;
 
 	____napi_schedule(sd, &sd->backlog);
-	sd->received_rps++;
+	/* Pairs with READ_ONCE() in softnet_seq_show() */
+	WRITE_ONCE(sd->received_rps, sd->received_rps + 1);
 }
 
 #endif /* CONFIG_RPS */
@@ -7523,7 +7524,8 @@ static __latent_entropy void net_rx_action(void)
 		 */
 		if (unlikely(budget <= 0 ||
 			     time_after_eq(jiffies, time_limit))) {
-			sd->time_squeeze++;
+			/* Pairs with READ_ONCE() in softnet_seq_show() */
+			WRITE_ONCE(sd->time_squeeze, sd->time_squeeze + 1);
 			break;
 		}
 	}
net/core/net-procfs.c +3 -3
index 69782d62fbe1..4f0f0709a1cb 100644
@@ -145,11 +145,11 @@ static int softnet_seq_show(struct seq_file *seq, void *v)
 	seq_printf(seq,
 		   "%08x %08x %08x %08x %08x %08x %08x %08x %08x %08x %08x %08x %08x "
 		   "%08x %08x\n",
-		   sd->processed, atomic_read(&sd->dropped),
-		   sd->time_squeeze, 0,
+		   READ_ONCE(sd->processed), atomic_read(&sd->dropped),
+		   READ_ONCE(sd->time_squeeze), 0,
 		   0, 0, 0, 0, /* was fastroute */
 		   0,	/* was cpu_collision */
-		   sd->received_rps, flow_limit_count,
+		   READ_ONCE(sd->received_rps), flow_limit_count,
 		   input_qlen + process_qlen, (int)seq->index,
 		   input_qlen, process_qlen);
 	return 0;
Signature · 1 line
2.49.0.504.g3bcea36a83-goog
#7 lore raw

PATCH 4/4 net: rps: remove kfree_rcu_mightsleep() use

In reply to Eric Dumazet

Add an rcu_head to sd_flow_limit and rps_sock_flow_table structs to use the more conventional and predictable k[v]free_rcu().

---
 include/net/rps.h          | 5 +++--
 net/core/dev.h             | 1 +
 net/core/sysctl_net_core.c | 4 ++--
 3 files changed, 6 insertions(+), 4 deletions(-)
include/net/rps.h +3 -2
index e358e9711f27..507f4aa5d39b 100644
@@ -57,9 +57,10 @@ struct rps_dev_flow_table {
  * meaning we use 32-6=26 bits for the hash.
  */
 struct rps_sock_flow_table {
-	u32	mask;
+	struct rcu_head	rcu;
+	u32		mask;
 
-	u32	ents[] ____cacheline_aligned_in_smp;
+	u32		ents[] ____cacheline_aligned_in_smp;
 };
 #define	RPS_SOCK_FLOW_TABLE_SIZE(_num) (offsetof(struct rps_sock_flow_table, ents[_num]))
 
net/core/dev.h +1 -0
index e855e1cb43fd..710abc05ebdb 100644
@@ -15,6 +15,7 @@ struct cpumask;
 /* Random bits of netdevice that don't need to be exposed */
 #define FLOW_LIMIT_HISTORY	(1 << 7)  /* must be ^2 and !overflow buckets */
 struct sd_flow_limit {
+	struct rcu_head		rcu;
 	unsigned int		count;
 	u8			log_buckets;
 	unsigned int		history_head;
net/core/sysctl_net_core.c +2 -2
index 5cfe76ede523..5dbb2c6f371d 100644
@@ -201,7 +201,7 @@ static int rps_sock_flow_sysctl(const struct ctl_table *table, int write,
 			if (orig_sock_table) {
 				static_branch_dec(&rps_needed);
 				static_branch_dec(&rfs_needed);
-				kvfree_rcu_mightsleep(orig_sock_table);
+				kvfree_rcu(orig_sock_table, rcu);
 			}
 		}
 	}
@@ -239,7 +239,7 @@ static int flow_limit_cpu_sysctl(const struct ctl_table *table, int write,
 				     lockdep_is_held(&flow_limit_update_mutex));
 			if (cur && !cpumask_test_cpu(i, mask)) {
 				RCU_INIT_POINTER(sd->flow_limit, NULL);
-				kfree_rcu_mightsleep(cur);
+				kfree_rcu(cur, rcu);
 			} else if (!cur && cpumask_test_cpu(i, mask)) {
 				cur = kzalloc_node(len, GFP_KERNEL,
 						   cpu_to_node(i));
Signature · 1 line
2.49.0.504.g3bcea36a83-goog
#9 lore raw
In reply to Eric Dumazet
This series was applied to netdev/net-next.git (main) by Jakub Kicinski <[email protected]>:

Hello:

This series was applied to netdev/net-next.git (main) by Jakub Kicinski <[email protected]>:

Quoted 5 lines · On Mon, 7 Apr 2025 16:35:58 +0000 you wrote:

Minor changes in rps:

skb_flow_limit() is probably unused these days,
and data-races are quite theoretical.

Quoted 7 lines

Eric Dumazet (4):
net: rps: change skb_flow_limit() hash function
net: rps: annotate data-races around (struct sd_flow_limit)->count
net: add data-race annotations in softnet_seq_show()
net: rps: remove kfree_rcu_mightsleep() use

[...]

Here is the summary with links:
- [net-next,1/4] net: rps: change skb_flow_limit() hash function

    https://git.kernel.org/netdev/net-next/c/c3025e94daa9
  - [net-next,2/4] net: rps: annotate data-races around (struct sd_flow_limit)->count
    https://git.kernel.org/netdev/net-next/c/7b6f0a852da3
  - [net-next,3/4] net: add data-race annotations in softnet_seq_show()
    https://git.kernel.org/netdev/net-next/c/22d046a778e4
  - [net-next,4/4] net: rps: remove kfree_rcu_mightsleep() use
    https://git.kernel.org/netdev/net-next/c/0a7de4a8f898

You are awesome, thank you!

Signature · 2 lines
Deet-doot-dot, I am a bot.
https://korg.docs.kernel.org/patchwork/pwbot.html