[PATCH v3] mac80211: Optimize sta lookup for many VIFs

[PATCH v3] mac80211: Optimize sta lookup for many VIFs

[greearb]

From: Ben Greear <greearb@candelatech.com>

The sta_info hash is designed to deal with an AP
with lots of stations associated, or a station interface
connected to a single AP.

However, when you have lots of station VIFs connected
to the same AP, the sta_info hash becomes worthless
as there is a single hash bucket that contains all the
entries in a linked list.

So, have the sdata object cache one of its station
interfaces.  If we are a station VIF with a single
sta_info, then this means we can ignore the sta_info
hash entirely.

On a test case with 128 stations and 50 TCP streams,
tx performance went from around 80Mbps to 124Mbps.

Signed-off-by: Ben Greear <greearb@candelatech.com>
---

v3:  Use rcu_dereference_protected as suggested.

:100644 100644 a618bda... 9b10bbf... M	net/mac80211/cfg.c
:100644 100644 493e2e8... c6386c7... M	net/mac80211/ieee80211_i.h
:100644 100644 415f9c6... a958ebc... M	net/mac80211/sta_info.c
 net/mac80211/cfg.c         |    5 +++++
 net/mac80211/ieee80211_i.h |    6 ++++++
 net/mac80211/sta_info.c    |   22 +++++++++++++++++++++-
 3 files changed, 32 insertions(+), 1 deletions(-)

diff --git a/net/mac80211/cfg.c b/net/mac80211/cfg.c
index a618bda..9b10bbf 100644
--- a/net/mac80211/cfg.c
+++ b/net/mac80211/cfg.c
@@ -1287,6 +1287,7 @@ static int ieee80211_change_station(struct wiphy *wiphy,
 	if (params->vlan && params->vlan != sta->sdata->dev) {
 		bool prev_4addr = false;
 		bool new_4addr = false;
+		struct sta_info *some_sta;
 
 		vlansdata = IEEE80211_DEV_TO_SUB_IF(params->vlan);
 
@@ -1312,7 +1313,11 @@ static int ieee80211_change_station(struct wiphy *wiphy,
 			prev_4addr = true;
 		}
 
+		some_sta = rcu_dereference_protected(sta->sdata->some_sta, 1);
+		if (some_sta == sta)
+			rcu_assign_pointer(sta->sdata->some_sta, NULL);
 		sta->sdata = vlansdata;
+		rcu_assign_pointer(sta->sdata->some_sta, sta);
 
 		if (sta->sta_state == IEEE80211_STA_AUTHORIZED &&
 		    prev_4addr != new_4addr) {
diff --git a/net/mac80211/ieee80211_i.h b/net/mac80211/ieee80211_i.h
index 493e2e8..c6386c7 100644
--- a/net/mac80211/ieee80211_i.h
+++ b/net/mac80211/ieee80211_i.h
@@ -669,6 +669,12 @@ struct ieee80211_sub_if_data {
 	/* count for keys needing tailroom space allocation */
 	int crypto_tx_tailroom_needed_cnt;
 
+	/* A pointer to some station associated with this interface, or
+	 * NULL.  This allows opportunistic lookup for sta_info objects when
+	 * sdata is a station with a single sta_info.
+	 */
+	struct sta_info __rcu *some_sta;
+
 	struct net_device *dev;
 	struct ieee80211_local *local;
 
diff --git a/net/mac80211/sta_info.c b/net/mac80211/sta_info.c
index 415f9c6..a958ebc 100644
--- a/net/mac80211/sta_info.c
+++ b/net/mac80211/sta_info.c
@@ -67,7 +67,11 @@
 static int sta_info_hash_del(struct ieee80211_local *local,
 			     struct sta_info *sta)
 {
-	struct sta_info *s;
+	struct sta_info *s, *some_sta;
+
+	some_sta = rcu_dereference_protected(sta->sdata->some_sta, 1);
+	if (some_sta == sta)
+		rcu_assign_pointer(sta->sdata->some_sta, NULL);
 
 	s = rcu_dereference_protected(local->sta_hash[STA_HASH(sta->sta.addr)],
 				      lockdep_is_held(&local->sta_mtx));
@@ -195,6 +199,20 @@ struct sta_info *sta_info_get(struct ieee80211_sub_if_data *sdata,
 	struct ieee80211_local *local = sdata->local;
 	struct sta_info *sta;
 
+	if (sdata->vif.type == NL80211_IFTYPE_STATION) {
+		struct sta_info *some_sta;
+
+		/* Shortcut for finding station entries for STATION VIFs */
+		some_sta = rcu_dereference(sdata->some_sta);
+		if (some_sta) {
+			if (WARN_ON(some_sta->sdata != sdata))
+				rcu_assign_pointer(sdata->some_sta, NULL);
+			else
+				if (ether_addr_equal(some_sta->sta.addr, addr))
+					return some_sta;
+		}
+	}
+
 	sta = rcu_dereference_check(local->sta_hash[STA_HASH(addr)],
 				    lockdep_is_held(&local->sta_mtx));
 	while (sta) {
@@ -278,6 +296,8 @@ static void sta_info_hash_add(struct ieee80211_local *local,
 	lockdep_assert_held(&local->sta_mtx);
 	sta->hnext = local->sta_hash[STA_HASH(sta->sta.addr)];
 	rcu_assign_pointer(local->sta_hash[STA_HASH(sta->sta.addr)], sta);
+
+	rcu_assign_pointer(sta->sdata->some_sta, sta);
 }
 
 static void sta_unblock(struct work_struct *wk)
-- 
1.7.3.4

Re: [PATCH v3] mac80211: Optimize sta lookup for many VIFs

[Johannes Berg]

On Mon, 2013-03-25 at 17:57 -0700, greearb@candelatech.com wrote:

> v3:  Use rcu_dereference_protected as suggested.

Heh.

> +		some_sta = rcu_dereference_protected(sta->sdata->some_sta, 1);

No no no, you're supposed to give it a proper argument like
"lockdep_is_held(&local->sta_mtx)"

johannes

Re: [PATCH v3] mac80211: Optimize sta lookup for many VIFs

[Ben Greear]

On 03/26/2013 12:42 PM, Johannes Berg wrote:
> On Mon, 2013-03-25 at 17:57 -0700, greearb@candelatech.com wrote:
>
>> v3:  Use rcu_dereference_protected as suggested.
>
> Heh.
>
>> +		some_sta = rcu_dereference_protected(sta->sdata->some_sta, 1);
>
> No no no, you're supposed to give it a proper argument like
> "lockdep_is_held(&local->sta_mtx)"

Before I screw this up again..do you prefer that I line up
method arguments and go over 80 character line length, or
force the argument leftward to meet the 80 char limit?

I see no good way to make it fit in 80 chars and also
line up properly.

Thanks,
Ben

>
> johannes
>


-- 
Ben Greear <greearb@candelatech.com>
Candela Technologies Inc  http://www.candelatech.com

Re: [PATCH v3] mac80211: Optimize sta lookup for many VIFs

[Johannes Berg]

On Wed, 2013-03-27 at 11:35 -0700, Ben Greear wrote:
> On 03/26/2013 12:42 PM, Johannes Berg wrote:
> > On Mon, 2013-03-25 at 17:57 -0700, greearb@candelatech.com wrote:
> >
> >> v3:  Use rcu_dereference_protected as suggested.
> >
> > Heh.
> >
> >> +		some_sta = rcu_dereference_protected(sta->sdata->some_sta, 1);
> >
> > No no no, you're supposed to give it a proper argument like
> > "lockdep_is_held(&local->sta_mtx)"
> 
> Before I screw this up again..do you prefer that I line up
> method arguments and go over 80 character line length, or
> force the argument leftward to meet the 80 char limit?
> 
> I see no good way to make it fit in 80 chars and also
> line up properly.

Huh, no idea. I guess I'd rather have it less than 80 cols.

johannes