package/kernel/mac80211/patches/subsys/354-mac80211-calculate-hash-for-fq-without-holding-fq-lo.patch

   1 From: Felix Fietkau <nbd@nbd.name>
   2 Date: Sat, 16 Mar 2019 17:57:38 +0100
   3 Subject: [PATCH] mac80211: calculate hash for fq without holding fq->lock
   4  in itxq enqueue
   5
   6 Reduces lock contention on enqueue/dequeue of iTXQ packets
   7
   8 Signed-off-by: Felix Fietkau <nbd@nbd.name>
   9 ---
  10
  11 --- a/include/net/fq_impl.h
  12 +++ b/include/net/fq_impl.h
  13 @@ -107,29 +107,31 @@ begin:
  14         return skb;
  15  }
  16
  17 -static struct fq_flow *fq_flow_classify(struct fq *fq,
  18 -                                       struct fq_tin *tin,
  19 -                                       struct sk_buff *skb,
  20 -                                       fq_flow_get_default_t get_default_func)
  21 +static u32 fq_flow_idx(struct fq *fq, struct sk_buff *skb)
  22  {
  23 -       struct fq_flow *flow;
  24 -       u32 hash;
  25 -       u32 idx;
  26 -
  27 -       lockdep_assert_held(&fq->lock);
  28 -
  29  #if LINUX_VERSION_IS_GEQ(5,3,10) || \
  30      LINUX_VERSION_IN_RANGE(4,19,83, 4,20,0) || \
  31      LINUX_VERSION_IN_RANGE(4,14,153, 4,15,0) || \
  32      LINUX_VERSION_IN_RANGE(4,9,200, 4,10,0) || \
  33      LINUX_VERSION_IN_RANGE(4,4,200, 4,5,0)
  34 -       hash = skb_get_hash_perturb(skb, &fq->perturbation);
  35 +       u32 hash = skb_get_hash_perturb(skb, &fq->perturbation);
  36  #else
  37 -       hash = skb_get_hash_perturb(skb, fq->perturbation);
  38 +       u32 hash = skb_get_hash_perturb(skb, fq->perturbation);
  39  #endif
  40 -       idx = reciprocal_scale(hash, fq->flows_cnt);
  41 -       flow = &fq->flows[idx];
  42
  43 +       return reciprocal_scale(hash, fq->flows_cnt);
  44 +}
  45 +
  46 +static struct fq_flow *fq_flow_classify(struct fq *fq,
  47 +                                       struct fq_tin *tin, u32 idx,
  48 +                                       struct sk_buff *skb,
  49 +                                       fq_flow_get_default_t get_default_func)
  50 +{
  51 +       struct fq_flow *flow;
  52 +
  53 +       lockdep_assert_held(&fq->lock);
  54 +
  55 +       flow = &fq->flows[idx];
  56         if (flow->tin && flow->tin != tin) {
  57                 flow = get_default_func(fq, tin, idx, skb);
  58                 tin->collisions++;
  59 @@ -161,7 +163,7 @@ static void fq_recalc_backlog(struct fq
  60  }
  61
  62  static void fq_tin_enqueue(struct fq *fq,
  63 -                          struct fq_tin *tin,
  64 +                          struct fq_tin *tin, u32 idx,
  65                            struct sk_buff *skb,
  66                            fq_skb_free_t free_func,
  67                            fq_flow_get_default_t get_default_func)
  68 @@ -171,7 +173,7 @@ static void fq_tin_enqueue(struct fq *fq
  69
  70         lockdep_assert_held(&fq->lock);
  71
  72 -       flow = fq_flow_classify(fq, tin, skb, get_default_func);
  73 +       flow = fq_flow_classify(fq, tin, idx, skb, get_default_func);
  74
  75         flow->tin = tin;
  76         flow->backlog += skb->len;
  77 --- a/net/mac80211/tx.c
  78 +++ b/net/mac80211/tx.c
  79 @@ -1396,11 +1396,15 @@ static void ieee80211_txq_enqueue(struct
  80  {
  81         struct fq *fq = &local->fq;
  82         struct fq_tin *tin = &txqi->tin;
  83 +       u32 flow_idx = fq_flow_idx(fq, skb);
  84
  85         ieee80211_set_skb_enqueue_time(skb);
  86 -       fq_tin_enqueue(fq, tin, skb,
  87 +
  88 +       spin_lock_bh(&fq->lock);
  89 +       fq_tin_enqueue(fq, tin, flow_idx, skb,
  90                        fq_skb_free_func,
  91                        fq_flow_get_default_func);
  92 +       spin_unlock_bh(&fq->lock);
  93  }
  94
  95  static bool fq_vlan_filter_func(struct fq *fq, struct fq_tin *tin,
  96 @@ -1570,7 +1574,6 @@ static bool ieee80211_queue_skb(struct i
  97                                 struct sta_info *sta,
  98                                 struct sk_buff *skb)
  99  {
 100 -       struct fq *fq = &local->fq;
 101         struct ieee80211_vif *vif;
 102         struct txq_info *txqi;
 103
 104 @@ -1588,9 +1591,7 @@ static bool ieee80211_queue_skb(struct i
 105         if (!txqi)
 106                 return false;
 107
 108 -       spin_lock_bh(&fq->lock);
 109         ieee80211_txq_enqueue(local, txqi, skb);
 110 -       spin_unlock_bh(&fq->lock);
 111
 112         schedule_and_wake_txq(local, txqi);
 113
 114 @@ -3223,6 +3224,7 @@ static bool ieee80211_amsdu_aggregate(st
 115         u8 max_subframes = sta->sta.max_amsdu_subframes;
 116         int max_frags = local->hw.max_tx_fragments;
 117         int max_amsdu_len = sta->sta.max_amsdu_len;
 118 +       u32 flow_idx;
 119         int orig_truesize;
 120         __be16 len;
 121         void *data;
 122 @@ -3245,6 +3247,8 @@ static bool ieee80211_amsdu_aggregate(st
 123                 max_amsdu_len = min_t(int, max_amsdu_len,
 124                                       sta->sta.max_rc_amsdu_len);
 125
 126 +       flow_idx = fq_flow_idx(fq, skb);
 127 +
 128         spin_lock_bh(&fq->lock);
 129
 130         /* TODO: Ideally aggregation should be done on dequeue to remain
 131 @@ -3252,7 +3256,8 @@ static bool ieee80211_amsdu_aggregate(st
 132          */
 133
 134         tin = &txqi->tin;
 135 -       flow = fq_flow_classify(fq, tin, skb, fq_flow_get_default_func);
 136 +       flow = fq_flow_classify(fq, tin, flow_idx, skb,
 137 +                               fq_flow_get_default_func);
 138         head = skb_peek_tail(&flow->queue);
 139         if (!head)
 140                 goto out;