package/kernel/mac80211/patches/subsys/330-mac80211-add-AQL-support-for-broadcast-packets.patch

   1 From: Felix Fietkau <nbd@nbd.name>
   2 Date: Fri, 9 Feb 2024 19:43:40 +0100
   3 Subject: [PATCH] mac80211: add AQL support for broadcast packets
   4
   5 Excessive broadcast traffic with little competing unicast traffic can easily
   6 flood hardware queues, leading to throughput issues. Additionally, filling
   7 the hardware queues with too many packets breaks FQ for broadcast data.
   8 Fix this by enabling AQL for broadcast packets.
   9
  10 Signed-off-by: Felix Fietkau <nbd@nbd.name>
  11 ---
  12
  13 --- a/include/net/cfg80211.h
  14 +++ b/include/net/cfg80211.h
  15 @@ -3324,6 +3324,7 @@ enum wiphy_params_flags {
  16  /* The per TXQ device queue limit in airtime */
  17  #define IEEE80211_DEFAULT_AQL_TXQ_LIMIT_L      5000
  18  #define IEEE80211_DEFAULT_AQL_TXQ_LIMIT_H      12000
  19 +#define IEEE80211_DEFAULT_AQL_TXQ_LIMIT_BC     50000
  20
  21  /* The per interface airtime threshold to switch to lower queue limit */
  22  #define IEEE80211_AQL_THRESHOLD                        24000
  23 --- a/net/mac80211/debugfs.c
  24 +++ b/net/mac80211/debugfs.c
  25 @@ -215,11 +215,13 @@ static ssize_t aql_pending_read(struct f
  26                         "VI     %u us\n"
  27                         "BE     %u us\n"
  28                         "BK     %u us\n"
  29 +                       "BC/MC  %u us\n"
  30                         "total  %u us\n",
  31                         atomic_read(&local->aql_ac_pending_airtime[IEEE80211_AC_VO]),
  32                         atomic_read(&local->aql_ac_pending_airtime[IEEE80211_AC_VI]),
  33                         atomic_read(&local->aql_ac_pending_airtime[IEEE80211_AC_BE]),
  34                         atomic_read(&local->aql_ac_pending_airtime[IEEE80211_AC_BK]),
  35 +                       atomic_read(&local->aql_bc_pending_airtime),
  36                         atomic_read(&local->aql_total_pending_airtime));
  37         return simple_read_from_buffer(user_buf, count, ppos,
  38                                        buf, len);
  39 @@ -245,7 +247,8 @@ static ssize_t aql_txq_limit_read(struct
  40                         "VO     %u              %u\n"
  41                         "VI     %u              %u\n"
  42                         "BE     %u              %u\n"
  43 -                       "BK     %u              %u\n",
  44 +                       "BK     %u              %u\n"
  45 +                       "BC/MC  %u\n",
  46                         local->aql_txq_limit_low[IEEE80211_AC_VO],
  47                         local->aql_txq_limit_high[IEEE80211_AC_VO],
  48                         local->aql_txq_limit_low[IEEE80211_AC_VI],
  49 @@ -253,7 +256,8 @@ static ssize_t aql_txq_limit_read(struct
  50                         local->aql_txq_limit_low[IEEE80211_AC_BE],
  51                         local->aql_txq_limit_high[IEEE80211_AC_BE],
  52                         local->aql_txq_limit_low[IEEE80211_AC_BK],
  53 -                       local->aql_txq_limit_high[IEEE80211_AC_BK]);
  54 +                       local->aql_txq_limit_high[IEEE80211_AC_BK],
  55 +                       local->aql_txq_limit_bc);
  56         return simple_read_from_buffer(user_buf, count, ppos,
  57                                        buf, len);
  58  }
  59 @@ -279,6 +283,11 @@ static ssize_t aql_txq_limit_write(struc
  60         else
  61                 buf[count] = '\0';
  62
  63 +       if (sscanf(buf, "mcast %u", &q_limit_low) == 1) {
  64 +               local->aql_txq_limit_bc = q_limit_low;
  65 +               return count;
  66 +       }
  67 +
  68         if (sscanf(buf, "%u %u %u", &ac, &q_limit_low, &q_limit_high) != 3)
  69                 return -EINVAL;
  70
  71 --- a/net/mac80211/ieee80211_i.h
  72 +++ b/net/mac80211/ieee80211_i.h
  73 @@ -1328,10 +1328,12 @@ struct ieee80211_local {
  74         spinlock_t handle_wake_tx_queue_lock;
  75
  76         u16 airtime_flags;
  77 +       u32 aql_txq_limit_bc;
  78         u32 aql_txq_limit_low[IEEE80211_NUM_ACS];
  79         u32 aql_txq_limit_high[IEEE80211_NUM_ACS];
  80         u32 aql_threshold;
  81         atomic_t aql_total_pending_airtime;
  82 +       atomic_t aql_bc_pending_airtime;
  83         atomic_t aql_ac_pending_airtime[IEEE80211_NUM_ACS];
  84
  85         const struct ieee80211_ops *ops;
  86 --- a/net/mac80211/main.c
  87 +++ b/net/mac80211/main.c
  88 @@ -788,6 +788,7 @@ struct ieee80211_hw *ieee80211_alloc_hw_
  89         spin_lock_init(&local->rx_path_lock);
  90         spin_lock_init(&local->queue_stop_reason_lock);
  91
  92 +       local->aql_txq_limit_bc = IEEE80211_DEFAULT_AQL_TXQ_LIMIT_BC;
  93         for (i = 0; i < IEEE80211_NUM_ACS; i++) {
  94                 INIT_LIST_HEAD(&local->active_txqs[i]);
  95                 spin_lock_init(&local->active_txq_lock[i]);
  96 --- a/net/mac80211/sta_info.c
  97 +++ b/net/mac80211/sta_info.c
  98 @@ -2341,28 +2341,27 @@ void ieee80211_sta_update_pending_airtim
  99                                           struct sta_info *sta, u8 ac,
 100                                           u16 tx_airtime, bool tx_completed)
 101  {
 102 +       atomic_t *counter;
 103         int tx_pending;
 104
 105         if (!wiphy_ext_feature_isset(local->hw.wiphy, NL80211_EXT_FEATURE_AQL))
 106                 return;
 107
 108 -       if (!tx_completed) {
 109 -               if (sta)
 110 -                       atomic_add(tx_airtime,
 111 -                                  &sta->airtime[ac].aql_tx_pending);
 112 +       if (sta)
 113 +               counter = &sta->airtime[ac].aql_tx_pending;
 114 +       else
 115 +               counter = &local->aql_bc_pending_airtime;
 116
 117 +       if (!tx_completed) {
 118 +               atomic_add(tx_airtime, counter);
 119                 atomic_add(tx_airtime, &local->aql_total_pending_airtime);
 120                 atomic_add(tx_airtime, &local->aql_ac_pending_airtime[ac]);
 121                 return;
 122         }
 123
 124 -       if (sta) {
 125 -               tx_pending = atomic_sub_return(tx_airtime,
 126 -                                              &sta->airtime[ac].aql_tx_pending);
 127 -               if (tx_pending < 0)
 128 -                       atomic_cmpxchg(&sta->airtime[ac].aql_tx_pending,
 129 -                                      tx_pending, 0);
 130 -       }
 131 +       tx_pending = atomic_sub_return(tx_airtime, counter);
 132 +       if (tx_pending < 0)
 133 +               atomic_cmpxchg(counter, tx_pending, 0);
 134
 135         atomic_sub(tx_airtime, &local->aql_total_pending_airtime);
 136         tx_pending = atomic_sub_return(tx_airtime,
 137 --- a/net/mac80211/tx.c
 138 +++ b/net/mac80211/tx.c
 139 @@ -3958,9 +3958,8 @@ begin:
 140  encap_out:
 141         IEEE80211_SKB_CB(skb)->control.vif = vif;
 142
 143 -       if (tx.sta &&
 144 -           wiphy_ext_feature_isset(local->hw.wiphy, NL80211_EXT_FEATURE_AQL)) {
 145 -               bool ampdu = txq->ac != IEEE80211_AC_VO;
 146 +       if (wiphy_ext_feature_isset(local->hw.wiphy, NL80211_EXT_FEATURE_AQL)) {
 147 +               bool ampdu = txq->sta && txq->ac != IEEE80211_AC_VO;
 148                 u32 airtime;
 149
 150                 airtime = ieee80211_calc_expected_tx_airtime(hw, vif, txq->sta,
 151 @@ -4026,6 +4025,7 @@ struct ieee80211_txq *ieee80211_next_txq
 152         struct ieee80211_txq *ret = NULL;
 153         struct txq_info *txqi = NULL, *head = NULL;
 154         bool found_eligible_txq = false;
 155 +       bool aql_check;
 156
 157         spin_lock_bh(&local->active_txq_lock[ac]);
 158
 159 @@ -4049,26 +4049,26 @@ struct ieee80211_txq *ieee80211_next_txq
 160         if (!head)
 161                 head = txqi;
 162
 163 +       aql_check = ieee80211_txq_airtime_check(hw, &txqi->txq);
 164 +       if (aql_check)
 165 +               found_eligible_txq = true;
 166 +
 167         if (txqi->txq.sta) {
 168                 struct sta_info *sta = container_of(txqi->txq.sta,
 169                                                     struct sta_info, sta);
 170 -               bool aql_check = ieee80211_txq_airtime_check(hw, &txqi->txq);
 171 -               s32 deficit = ieee80211_sta_deficit(sta, txqi->txq.ac);
 172 -
 173 -               if (aql_check)
 174 -                       found_eligible_txq = true;
 175 -
 176 -               if (deficit < 0)
 177 +               if (ieee80211_sta_deficit(sta, txqi->txq.ac) < 0) {
 178                         sta->airtime[txqi->txq.ac].deficit +=
 179                                 sta->airtime_weight << AIRTIME_QUANTUM_SHIFT;
 180 -
 181 -               if (deficit < 0 || !aql_check) {
 182 -                       list_move_tail(&txqi->schedule_order,
 183 -                                      &local->active_txqs[txqi->txq.ac]);
 184 -                       goto begin;
 185 +                       aql_check = false;
 186                 }
 187         }
 188
 189 +       if (!aql_check) {
 190 +               list_move_tail(&txqi->schedule_order,
 191 +                                  &local->active_txqs[txqi->txq.ac]);
 192 +               goto begin;
 193 +       }
 194 +
 195         if (txqi->schedule_round == local->schedule_round[ac])
 196                 goto out;
 197
 198 @@ -4133,7 +4133,8 @@ bool ieee80211_txq_airtime_check(struct
 199                 return true;
 200
 201         if (!txq->sta)
 202 -               return true;
 203 +               return atomic_read(&local->aql_bc_pending_airtime) <
 204 +                      local->aql_txq_limit_bc;
 205
 206         if (unlikely(txq->tid == IEEE80211_NUM_TIDS))
 207                 return true;
 208 @@ -4182,15 +4183,15 @@ bool ieee80211_txq_may_transmit(struct i
 209
 210         spin_lock_bh(&local->active_txq_lock[ac]);
 211
 212 -       if (!txqi->txq.sta)
 213 -               goto out;
 214 -
 215         if (list_empty(&txqi->schedule_order))
 216                 goto out;
 217
 218         if (!ieee80211_txq_schedule_airtime_check(local, ac))
 219                 goto out;
 220
 221 +       if (!txqi->txq.sta)
 222 +               goto out;
 223 +
 224         list_for_each_entry_safe(iter, tmp, &local->active_txqs[ac],
 225                                  schedule_order) {
 226                 if (iter == txqi)