package/kernel/mac80211/patches/300-mac80211-add-an-intermediate-software-queue-implemen.patch

   1 From: Felix Fietkau <nbd@openwrt.org>
   2 Date: Tue, 18 Nov 2014 23:58:51 +0100
   3 Subject: [PATCH] mac80211: add an intermediate software queue implementation
   4
   5 This allows drivers to request per-vif and per-sta-tid queues from which
   6 they can pull frames. This makes it easier to keep the hardware queues
   7 short, and to improve fairness between clients and vifs.
   8
   9 The task of scheduling packet transmission is left up to the driver -
  10 queueing is controlled by mac80211. Drivers can only dequeue packets by
  11 calling ieee80211_tx_dequeue. This makes it possible to add active queue
  12 management later without changing drivers using this code.
  13
  14 This can also be used as a starting point to implement A-MSDU
  15 aggregation in a way that does not add artificially induced latency.
  16
  17 Signed-off-by: Felix Fietkau <nbd@openwrt.org>
  18 ---
  19
  20 --- a/include/net/mac80211.h
  21 +++ b/include/net/mac80211.h
  22 @@ -84,6 +84,35 @@
  23   *
  24   */
  25
  26 +/**
  27 + * DOC: mac80211 software tx queueing
  28 + *
  29 + * mac80211 provides an optional intermediate queueing implementation designed
  30 + * to allow the driver to keep hardware queues short and provide some fairness
  31 + * between different stations/interfaces.
  32 + * In this model, the driver pulls data frames from the mac80211 queue instead
  33 + * of letting mac80211 push them via drv_tx().
  34 + * Other frames (e.g. control or management) are still pushed using drv_tx().
  35 + *
  36 + * Intermediate queues (struct ieee80211_txq) are kept per-sta per-tid, with a
  37 + * single per-vif queue for multicast data frames.
  38 + *
  39 + * The driver is expected to initialize its private per-queue data for stations
  40 + * and interfaces in the .add_interface and .sta_add ops.
  41 + *
  42 + * The driver can not access the queue directly. To dequeue a frame, it calls
  43 + * ieee80211_tx_dequeue(). Whenever mac80211 adds a new frame to a queue, it
  44 + * calls the .wake_tx_queue driver op.
  45 + *
  46 + * For AP powersave TIM handling, the driver only needs to indicate if it has
  47 + * buffered packets in the driver specific data structures by calling
  48 + * ieee80211_sta_set_buffered(). For frames buffered in the ieee80211_txq
  49 + * struct, mac80211 sets the appropriate TIM PVB bits and calls
  50 + * .release_buffered_frames().
  51 + * That callback is expected to release its own buffered frames and afterwards
  52 + * also frames from the ieee80211_txq (obtained via ieee80211_tx_dequeue).
  53 + */
  54 +
  55  struct device;
  56
  57  /**
  58 @@ -1257,6 +1286,8 @@ struct ieee80211_vif {
  59         u8 cab_queue;
  60         u8 hw_queue[IEEE80211_NUM_ACS];
  61
  62 +       struct ieee80211_txq *txq;
  63 +
  64         struct ieee80211_chanctx_conf __rcu *chanctx_conf;
  65
  66         u32 driver_flags;
  67 @@ -1519,6 +1550,8 @@ struct ieee80211_sta {
  68         bool tdls_initiator;
  69         bool mfp;
  70
  71 +       struct ieee80211_txq *txq[IEEE80211_NUM_TIDS];
  72 +
  73         /* must be last */
  74         u8 drv_priv[0] __aligned(sizeof(void *));
  75  };
  76 @@ -1547,6 +1580,27 @@ struct ieee80211_tx_control {
  77  };
  78
  79  /**
  80 + * struct ieee80211_txq - Software intermediate tx queue
  81 + *
  82 + * @vif: &struct ieee80211_vif pointer from the add_interface callback.
  83 + * @sta: station table entry, %NULL for per-vif queue
  84 + * @tid: the TID for this queue (unused for per-vif queue)
  85 + * @ac: the AC for this queue
  86 + *
  87 + * The driver can obtain packets from this queue by calling
  88 + * ieee80211_tx_dequeue().
  89 + */
  90 +struct ieee80211_txq {
  91 +       struct ieee80211_vif *vif;
  92 +       struct ieee80211_sta *sta;
  93 +       u8 tid;
  94 +       u8 ac;
  95 +
  96 +       /* must be last */
  97 +       u8 drv_priv[0] __aligned(sizeof(void *));
  98 +};
  99 +
 100 +/**
 101   * enum ieee80211_hw_flags - hardware flags
 102   *
 103   * These flags are used to indicate hardware capabilities to
 104 @@ -1770,6 +1824,8 @@ enum ieee80211_hw_flags {
 105   *     within &struct ieee80211_sta.
 106   * @chanctx_data_size: size (in bytes) of the drv_priv data area
 107   *     within &struct ieee80211_chanctx_conf.
 108 + * @txq_data_size: size (in bytes) of the drv_priv data area
 109 + *     within @struct ieee80211_txq.
 110   *
 111   * @max_rates: maximum number of alternate rate retry stages the hw
 112   *     can handle.
 113 @@ -1818,6 +1874,9 @@ enum ieee80211_hw_flags {
 114   * @n_cipher_schemes: a size of an array of cipher schemes definitions.
 115   * @cipher_schemes: a pointer to an array of cipher scheme definitions
 116   *     supported by HW.
 117 + *
 118 + * @txq_ac_max_pending: maximum number of frames per AC pending in all txq
 119 + *     entries for a vif.
 120   */
 121  struct ieee80211_hw {
 122         struct ieee80211_conf conf;
 123 @@ -1830,6 +1889,7 @@ struct ieee80211_hw {
 124         int vif_data_size;
 125         int sta_data_size;
 126         int chanctx_data_size;
 127 +       int txq_data_size;
 128         u16 queues;
 129         u16 max_listen_interval;
 130         s8 max_signal;
 131 @@ -1846,6 +1906,7 @@ struct ieee80211_hw {
 132         u8 uapsd_max_sp_len;
 133         u8 n_cipher_schemes;
 134         const struct ieee80211_cipher_scheme *cipher_schemes;
 135 +       int txq_ac_max_pending;
 136  };
 137
 138  /**
 139 @@ -3007,6 +3068,8 @@ enum ieee80211_reconfig_type {
 140   *     response template is provided, together with the location of the
 141   *     switch-timing IE within the template. The skb can only be used within
 142   *     the function call.
 143 + *
 144 + * @wake_tx_queue: Called when new packets have been added to the queue.
 145   */
 146  struct ieee80211_ops {
 147         void (*tx)(struct ieee80211_hw *hw,
 148 @@ -3238,6 +3301,9 @@ struct ieee80211_ops {
 149         void (*tdls_recv_channel_switch)(struct ieee80211_hw *hw,
 150                                          struct ieee80211_vif *vif,
 151                                          struct ieee80211_tdls_ch_sw_params *params);
 152 +
 153 +       void (*wake_tx_queue)(struct ieee80211_hw *hw,
 154 +                             struct ieee80211_txq *txq);
 155  };
 156
 157  /**
 158 @@ -5249,4 +5315,17 @@ void ieee80211_unreserve_tid(struct ieee
 159   */
 160  size_t ieee80211_ie_split(const u8 *ies, size_t ielen,
 161                           const u8 *ids, int n_ids, size_t offset);
 162 +
 163 +/**
 164 + * ieee80211_tx_dequeue - dequeue a packet from a software tx queue
 165 + *
 166 + * @hw: pointer as obtained from ieee80211_alloc_hw()
 167 + * @txq: pointer obtained from .add_tx_queue() call
 168 + *
 169 + * Returns the skb if successful, %NULL if no frame was available.
 170 + */
 171 +struct sk_buff *ieee80211_tx_dequeue(struct ieee80211_hw *hw,
 172 +                                    struct ieee80211_txq *txq);
 173 +
 174 +
 175  #endif /* MAC80211_H */
 176 --- a/net/mac80211/driver-ops.h
 177 +++ b/net/mac80211/driver-ops.h
 178 @@ -1367,4 +1367,16 @@ drv_tdls_recv_channel_switch(struct ieee
 179         trace_drv_return_void(local);
 180  }
 181
 182 +static inline void drv_wake_tx_queue(struct ieee80211_local *local,
 183 +                                    struct txq_info *txq)
 184 +{
 185 +       struct ieee80211_sub_if_data *sdata = vif_to_sdata(txq->txq.vif);
 186 +
 187 +       if (!check_sdata_in_driver(sdata))
 188 +               return;
 189 +
 190 +       trace_drv_wake_tx_queue(local, sdata, txq->txq.sta, txq->txq.tid);
 191 +       local->ops->wake_tx_queue(&local->hw, &txq->txq);
 192 +}
 193 +
 194  #endif /* __MAC80211_DRIVER_OPS */
 195 --- a/net/mac80211/ieee80211_i.h
 196 +++ b/net/mac80211/ieee80211_i.h
 197 @@ -809,6 +809,19 @@ struct mac80211_qos_map {
 198         struct rcu_head rcu_head;
 199  };
 200
 201 +enum txq_info_flags {
 202 +       IEEE80211_TXQ_STOP,
 203 +       IEEE80211_TXQ_AMPDU,
 204 +};
 205 +
 206 +struct txq_info {
 207 +       struct sk_buff_head queue;
 208 +       unsigned long flags;
 209 +
 210 +       /* keep last! */
 211 +       struct ieee80211_txq txq;
 212 +};
 213 +
 214  struct ieee80211_sub_if_data {
 215         struct list_head list;
 216
 217 @@ -853,6 +866,7 @@ struct ieee80211_sub_if_data {
 218         bool control_port_no_encrypt;
 219         int encrypt_headroom;
 220
 221 +       atomic_t txqs_len[IEEE80211_NUM_ACS];
 222         struct ieee80211_tx_queue_params tx_conf[IEEE80211_NUM_ACS];
 223         struct mac80211_qos_map __rcu *qos_map;
 224
 225 @@ -1453,6 +1467,10 @@ static inline struct ieee80211_local *hw
 226         return container_of(hw, struct ieee80211_local, hw);
 227  }
 228
 229 +static inline struct txq_info *to_txq_info(struct ieee80211_txq *txq)
 230 +{
 231 +       return container_of(txq, struct txq_info, txq);
 232 +}
 233
 234  static inline int ieee80211_bssid_match(const u8 *raddr, const u8 *addr)
 235  {
 236 @@ -1905,6 +1923,9 @@ static inline bool ieee80211_can_run_wor
 237         return true;
 238  }
 239
 240 +void ieee80211_init_tx_queue(struct ieee80211_sub_if_data *sdata,
 241 +                            struct sta_info *sta,
 242 +                            struct txq_info *txq, int tid);
 243  void ieee80211_send_auth(struct ieee80211_sub_if_data *sdata,
 244                          u16 transaction, u16 auth_alg, u16 status,
 245                          const u8 *extra, size_t extra_len, const u8 *bssid,
 246 --- a/net/mac80211/iface.c
 247 +++ b/net/mac80211/iface.c
 248 @@ -969,6 +969,13 @@ static void ieee80211_do_stop(struct iee
 249         }
 250         spin_unlock_irqrestore(&local->queue_stop_reason_lock, flags);
 251
 252 +       if (sdata->vif.txq) {
 253 +               struct txq_info *txqi = to_txq_info(sdata->vif.txq);
 254 +
 255 +               ieee80211_purge_tx_queue(&local->hw, &txqi->queue);
 256 +               atomic_set(&sdata->txqs_len[txqi->txq.ac], 0);
 257 +       }
 258 +
 259         if (local->open_count == 0)
 260                 ieee80211_clear_tx_pending(local);
 261
 262 @@ -1674,6 +1681,7 @@ int ieee80211_if_add(struct ieee80211_lo
 263  {
 264         struct net_device *ndev = NULL;
 265         struct ieee80211_sub_if_data *sdata = NULL;
 266 +       struct txq_info *txqi;
 267         int ret, i;
 268         int txqs = 1;
 269
 270 @@ -1693,10 +1701,18 @@ int ieee80211_if_add(struct ieee80211_lo
 271                 ieee80211_assign_perm_addr(local, wdev->address, type);
 272                 memcpy(sdata->vif.addr, wdev->address, ETH_ALEN);
 273         } else {
 274 +               int size = ALIGN(sizeof(*sdata) + local->hw.vif_data_size,
 275 +                                sizeof(void *));
 276 +               int txq_size = 0;
 277 +
 278 +               if (local->ops->wake_tx_queue)
 279 +                       txq_size += sizeof(struct txq_info) +
 280 +                                   local->hw.txq_data_size;
 281 +
 282                 if (local->hw.queues >= IEEE80211_NUM_ACS)
 283                         txqs = IEEE80211_NUM_ACS;
 284
 285 -               ndev = alloc_netdev_mqs(sizeof(*sdata) + local->hw.vif_data_size,
 286 +               ndev = alloc_netdev_mqs(size + txq_size,
 287                                         name, NET_NAME_UNKNOWN,
 288                                         ieee80211_if_setup, txqs, 1);
 289                 if (!ndev)
 290 @@ -1731,6 +1747,9 @@ int ieee80211_if_add(struct ieee80211_lo
 291                 memcpy(sdata->vif.addr, ndev->dev_addr, ETH_ALEN);
 292                 memcpy(sdata->name, ndev->name, IFNAMSIZ);
 293
 294 +               txqi = netdev_priv(ndev) + size;
 295 +               ieee80211_init_tx_queue(sdata, NULL, txqi, 0);
 296 +
 297                 sdata->dev = ndev;
 298         }
 299
 300 @@ -1773,6 +1792,15 @@ int ieee80211_if_add(struct ieee80211_lo
 301         ieee80211_setup_sdata(sdata, type);
 302
 303         if (ndev) {
 304 +               struct txq_info *txqi = NULL;
 305 +
 306 +               if (local->ops->wake_tx_queue) {
 307 +                       txqi = kzalloc(sizeof(*txqi) +
 308 +                                     local->hw.txq_data_size, GFP_KERNEL);
 309 +                       if (txqi)
 310 +                               ieee80211_init_tx_queue(sdata, NULL, txqi, 0);
 311 +               }
 312 +
 313                 if (params) {
 314                         ndev->ieee80211_ptr->use_4addr = params->use_4addr;
 315                         if (type == NL80211_IFTYPE_STATION)
 316 @@ -1785,6 +1813,7 @@ int ieee80211_if_add(struct ieee80211_lo
 317
 318                 ret = register_netdevice(ndev);
 319                 if (ret) {
 320 +                       kfree(txqi);
 321                         free_netdev(ndev);
 322                         return ret;
 323                 }
 324 --- a/net/mac80211/main.c
 325 +++ b/net/mac80211/main.c
 326 @@ -1019,6 +1019,9 @@ int ieee80211_register_hw(struct ieee802
 327
 328         local->dynamic_ps_forced_timeout = -1;
 329
 330 +       if (!local->hw.txq_ac_max_pending)
 331 +               local->hw.txq_ac_max_pending = 64;
 332 +
 333         result = ieee80211_wep_init(local);
 334         if (result < 0)
 335                 wiphy_debug(local->hw.wiphy, "Failed to initialize wep: %d\n",
 336 --- a/net/mac80211/sta_info.c
 337 +++ b/net/mac80211/sta_info.c
 338 @@ -118,6 +118,16 @@ static void __cleanup_single_sta(struct
 339                 atomic_dec(&ps->num_sta_ps);
 340         }
 341
 342 +       if (sta->sta.txq[0]) {
 343 +               for (i = 0; i < ARRAY_SIZE(sta->sta.txq); i++) {
 344 +                       struct txq_info *txqi = to_txq_info(sta->sta.txq[i]);
 345 +                       int n = skb_queue_len(&txqi->queue);
 346 +
 347 +                       ieee80211_purge_tx_queue(&local->hw, &txqi->queue);
 348 +                       atomic_sub(n, &sdata->txqs_len[txqi->txq.ac]);
 349 +               }
 350 +       }
 351 +
 352         for (ac = 0; ac < IEEE80211_NUM_ACS; ac++) {
 353                 local->total_ps_buffered -= skb_queue_len(&sta->ps_tx_buf[ac]);
 354                 ieee80211_purge_tx_queue(&local->hw, &sta->ps_tx_buf[ac]);
 355 @@ -234,6 +244,8 @@ void sta_info_free(struct ieee80211_loca
 356
 357         sta_dbg(sta->sdata, "Destroyed STA %pM\n", sta->sta.addr);
 358
 359 +       if (sta->sta.txq[0])
 360 +               kfree(to_txq_info(sta->sta.txq[0]));
 361         kfree(rcu_dereference_raw(sta->sta.rates));
 362         kfree(sta);
 363  }
 364 @@ -285,11 +297,12 @@ struct sta_info *sta_info_alloc(struct i
 365                                 const u8 *addr, gfp_t gfp)
 366  {
 367         struct ieee80211_local *local = sdata->local;
 368 +       struct ieee80211_hw *hw = &local->hw;
 369         struct sta_info *sta;
 370         struct timespec uptime;
 371         int i;
 372
 373 -       sta = kzalloc(sizeof(*sta) + local->hw.sta_data_size, gfp);
 374 +       sta = kzalloc(sizeof(*sta) + hw->sta_data_size, gfp);
 375         if (!sta)
 376                 return NULL;
 377
 378 @@ -321,11 +334,25 @@ struct sta_info *sta_info_alloc(struct i
 379         for (i = 0; i < ARRAY_SIZE(sta->chain_signal_avg); i++)
 380                 ewma_init(&sta->chain_signal_avg[i], 1024, 8);
 381
 382 -       if (sta_prepare_rate_control(local, sta, gfp)) {
 383 -               kfree(sta);
 384 -               return NULL;
 385 +       if (local->ops->wake_tx_queue) {
 386 +               void *txq_data;
 387 +               int size = sizeof(struct txq_info) +
 388 +                          ALIGN(hw->txq_data_size, sizeof(void *));
 389 +
 390 +               txq_data = kcalloc(ARRAY_SIZE(sta->sta.txq), size, gfp);
 391 +               if (!txq_data)
 392 +                       goto free;
 393 +
 394 +               for (i = 0; i < ARRAY_SIZE(sta->sta.txq); i++) {
 395 +                       struct txq_info *txq = txq_data + i * size;
 396 +
 397 +                       ieee80211_init_tx_queue(sdata, sta, txq, i);
 398 +               }
 399         }
 400
 401 +       if (sta_prepare_rate_control(local, sta, gfp))
 402 +               goto free_txq;
 403 +
 404         for (i = 0; i < IEEE80211_NUM_TIDS; i++) {
 405                 /*
 406                  * timer_to_tid must be initialized with identity mapping
 407 @@ -346,7 +373,7 @@ struct sta_info *sta_info_alloc(struct i
 408         if (sdata->vif.type == NL80211_IFTYPE_AP ||
 409             sdata->vif.type == NL80211_IFTYPE_AP_VLAN) {
 410                 struct ieee80211_supported_band *sband =
 411 -                       local->hw.wiphy->bands[ieee80211_get_sdata_band(sdata)];
 412 +                       hw->wiphy->bands[ieee80211_get_sdata_band(sdata)];
 413                 u8 smps = (sband->ht_cap.cap & IEEE80211_HT_CAP_SM_PS) >>
 414                                 IEEE80211_HT_CAP_SM_PS_SHIFT;
 415                 /*
 416 @@ -371,6 +398,13 @@ struct sta_info *sta_info_alloc(struct i
 417         sta_dbg(sdata, "Allocated STA %pM\n", sta->sta.addr);
 418
 419         return sta;
 420 +
 421 +free_txq:
 422 +       if (sta->sta.txq[0])
 423 +               kfree(to_txq_info(sta->sta.txq[0]));
 424 +free:
 425 +       kfree(sta);
 426 +       return NULL;
 427  }
 428
 429  static int sta_info_insert_check(struct sta_info *sta)
 430 @@ -640,6 +674,8 @@ static void __sta_info_recalc_tim(struct
 431
 432                 indicate_tim |=
 433                         sta->driver_buffered_tids & tids;
 434 +               indicate_tim |=
 435 +                       sta->txq_buffered_tids & tids;
 436         }
 437
 438   done:
 439 @@ -1071,7 +1107,7 @@ void ieee80211_sta_ps_deliver_wakeup(str
 440         struct ieee80211_sub_if_data *sdata = sta->sdata;
 441         struct ieee80211_local *local = sdata->local;
 442         struct sk_buff_head pending;
 443 -       int filtered = 0, buffered = 0, ac;
 444 +       int filtered = 0, buffered = 0, ac, i;
 445         unsigned long flags;
 446         struct ps_data *ps;
 447
 448 @@ -1090,10 +1126,22 @@ void ieee80211_sta_ps_deliver_wakeup(str
 449
 450         BUILD_BUG_ON(BITS_TO_LONGS(IEEE80211_NUM_TIDS) > 1);
 451         sta->driver_buffered_tids = 0;
 452 +       sta->txq_buffered_tids = 0;
 453
 454         if (!(local->hw.flags & IEEE80211_HW_AP_LINK_PS))
 455                 drv_sta_notify(local, sdata, STA_NOTIFY_AWAKE, &sta->sta);
 456
 457 +       if (sta->sta.txq[0]) {
 458 +               for (i = 0; i < ARRAY_SIZE(sta->sta.txq); i++) {
 459 +                       struct txq_info *txqi = to_txq_info(sta->sta.txq[i]);
 460 +
 461 +                       if (!skb_queue_len(&txqi->queue))
 462 +                               continue;
 463 +
 464 +                       drv_wake_tx_queue(local, txqi);
 465 +               }
 466 +       }
 467 +
 468         skb_queue_head_init(&pending);
 469
 470         /* sync with ieee80211_tx_h_unicast_ps_buf */
 471 @@ -1254,7 +1302,7 @@ ieee80211_sta_ps_deliver_response(struct
 472         struct ieee80211_sub_if_data *sdata = sta->sdata;
 473         struct ieee80211_local *local = sdata->local;
 474         bool more_data = false;
 475 -       int ac;
 476 +       int ac, tid;
 477         unsigned long driver_release_tids = 0;
 478         struct sk_buff_head frames;
 479
 480 @@ -1275,8 +1323,10 @@ ieee80211_sta_ps_deliver_response(struct
 481                 /* if we already have frames from software, then we can't also
 482                  * release from hardware queues
 483                  */
 484 -               if (skb_queue_empty(&frames))
 485 +               if (skb_queue_empty(&frames)) {
 486                         driver_release_tids |= sta->driver_buffered_tids & tids;
 487 +                       driver_release_tids |= sta->txq_buffered_tids & tids;
 488 +               }
 489
 490                 if (driver_release_tids) {
 491                         /* If the driver has data on more than one TID then
 492 @@ -1447,6 +1497,8 @@ ieee80211_sta_ps_deliver_response(struct
 493
 494                 sta_info_recalc_tim(sta);
 495         } else {
 496 +               unsigned long tids = sta->txq_buffered_tids & driver_release_tids;
 497 +
 498                 /*
 499                  * We need to release a frame that is buffered somewhere in the
 500                  * driver ... it'll have to handle that.
 501 @@ -1466,8 +1518,22 @@ ieee80211_sta_ps_deliver_response(struct
 502                  * that the TID(s) became empty before returning here from the
 503                  * release function.
 504                  * Either way, however, when the driver tells us that the TID(s)
 505 -                * became empty we'll do the TIM recalculation.
 506 +                * became empty or we find that a txq became empty, we'll do the
 507 +                * TIM recalculation.
 508                  */
 509 +
 510 +               if (!sta->sta.txq[0])
 511 +                       return;
 512 +
 513 +               for (tid = 0; tid < ARRAY_SIZE(sta->sta.txq); tid++) {
 514 +                       struct txq_info *txqi = to_txq_info(sta->sta.txq[tid]);
 515 +
 516 +                       if (!(tids & BIT(tid)) || skb_queue_len(&txqi->queue))
 517 +                               continue;
 518 +
 519 +                       sta_info_recalc_tim(sta);
 520 +                       break;
 521 +               }
 522         }
 523  }
 524
 525 --- a/net/mac80211/sta_info.h
 526 +++ b/net/mac80211/sta_info.h
 527 @@ -274,6 +274,7 @@ struct sta_ampdu_mlme {
 528   *     entered power saving state, these are also delivered to
 529   *     the station when it leaves powersave or polls for frames
 530   * @driver_buffered_tids: bitmap of TIDs the driver has data buffered on
 531 + * @txq_buffered_tids: bitmap of TIDs that mac80211 has txq data buffered on
 532   * @rx_packets: Number of MSDUs received from this STA
 533   * @rx_bytes: Number of bytes received from this STA
 534   * @last_rx: time (in jiffies) when last frame was received from this STA
 535 @@ -368,6 +369,7 @@ struct sta_info {
 536         struct sk_buff_head ps_tx_buf[IEEE80211_NUM_ACS];
 537         struct sk_buff_head tx_filtered[IEEE80211_NUM_ACS];
 538         unsigned long driver_buffered_tids;
 539 +       unsigned long txq_buffered_tids;
 540
 541         /* Updated from RX path only, no locking requirements */
 542         unsigned long rx_packets;
 543 --- a/net/mac80211/trace.h
 544 +++ b/net/mac80211/trace.h
 545 @@ -2312,6 +2312,34 @@ TRACE_EVENT(drv_tdls_recv_channel_switch
 546         )
 547  );
 548
 549 +TRACE_EVENT(drv_wake_tx_queue,
 550 +       TP_PROTO(struct ieee80211_local *local,
 551 +                struct ieee80211_sub_if_data *sdata,
 552 +                struct ieee80211_sta *sta,
 553 +                u8 tid),
 554 +
 555 +       TP_ARGS(local, sdata, sta, tid),
 556 +
 557 +       TP_STRUCT__entry(
 558 +               LOCAL_ENTRY
 559 +               VIF_ENTRY
 560 +               STA_ENTRY
 561 +               __field(u8, tid)
 562 +       ),
 563 +
 564 +       TP_fast_assign(
 565 +               LOCAL_ASSIGN;
 566 +               VIF_ASSIGN;
 567 +               STA_ASSIGN;
 568 +               __entry->tid = tid;
 569 +       ),
 570 +
 571 +       TP_printk(
 572 +               LOCAL_PR_FMT  VIF_PR_FMT  STA_PR_FMT " tid: 0x%x",
 573 +               LOCAL_PR_ARG, VIF_PR_ARG, STA_PR_ARG, __entry->tid
 574 +       )
 575 +);
 576 +
 577  #ifdef CPTCFG_MAC80211_MESSAGE_TRACING
 578  #undef TRACE_SYSTEM
 579  #define TRACE_SYSTEM mac80211_msg
 580 --- a/net/mac80211/tx.c
 581 +++ b/net/mac80211/tx.c
 582 @@ -776,12 +776,23 @@ ieee80211_tx_h_rate_ctrl(struct ieee8021
 583         return TX_CONTINUE;
 584  }
 585
 586 +static u16
 587 +ieee80211_tx_next_seq(struct sta_info *sta, int tid)
 588 +{
 589 +       u16 *seq = &sta->tid_seq[tid];
 590 +       u16 ret = cpu_to_le16(*seq);
 591 +
 592 +       /* Increase the sequence number. */
 593 +       *seq = (*seq + 0x10) & IEEE80211_SCTL_SEQ;
 594 +
 595 +       return ret;
 596 +}
 597 +
 598  static ieee80211_tx_result debug_noinline
 599  ieee80211_tx_h_sequence(struct ieee80211_tx_data *tx)
 600  {
 601         struct ieee80211_tx_info *info = IEEE80211_SKB_CB(tx->skb);
 602         struct ieee80211_hdr *hdr = (struct ieee80211_hdr *)tx->skb->data;
 603 -       u16 *seq;
 604         u8 *qc;
 605         int tid;
 606
 607 @@ -832,13 +843,10 @@ ieee80211_tx_h_sequence(struct ieee80211
 608
 609         qc = ieee80211_get_qos_ctl(hdr);
 610         tid = *qc & IEEE80211_QOS_CTL_TID_MASK;
 611 -       seq = &tx->sta->tid_seq[tid];
 612         tx->sta->tx_msdu[tid]++;
 613
 614 -       hdr->seq_ctrl = cpu_to_le16(*seq);
 615 -
 616 -       /* Increase the sequence number. */
 617 -       *seq = (*seq + 0x10) & IEEE80211_SCTL_SEQ;
 618 +       if (!tx->sta->sta.txq[0])
 619 +               hdr->seq_ctrl = ieee80211_tx_next_seq(tx->sta, tid);
 620
 621         return TX_CONTINUE;
 622  }
 623 @@ -1067,7 +1075,7 @@ static bool ieee80211_tx_prep_agg(struct
 624                  * nothing -- this aggregation session is being started
 625                  * but that might still fail with the driver
 626                  */
 627 -       } else {
 628 +       } else if (!tx->sta->sta.txq[tid]) {
 629                 spin_lock(&tx->sta->lock);
 630                 /*
 631                  * Need to re-check now, because we may get here
 632 @@ -1201,13 +1209,102 @@ ieee80211_tx_prepare(struct ieee80211_su
 633         return TX_CONTINUE;
 634  }
 635
 636 +static void ieee80211_drv_tx(struct ieee80211_local *local,
 637 +                            struct ieee80211_vif *vif,
 638 +                            struct ieee80211_sta *pubsta,
 639 +                            struct sk_buff *skb)
 640 +{
 641 +       struct ieee80211_hdr *hdr = (struct ieee80211_hdr *) skb->data;
 642 +       struct ieee80211_sub_if_data *sdata = vif_to_sdata(vif);
 643 +       struct ieee80211_tx_info *info = IEEE80211_SKB_CB(skb);
 644 +       struct ieee80211_tx_control control = {
 645 +               .sta = pubsta
 646 +       };
 647 +       struct ieee80211_txq *txq = NULL;
 648 +       struct txq_info *txqi;
 649 +       u8 ac;
 650 +
 651 +       if (info->control.flags & IEEE80211_TX_CTRL_PS_RESPONSE)
 652 +               goto tx_normal;
 653 +
 654 +       if (!ieee80211_is_data(hdr->frame_control))
 655 +               goto tx_normal;
 656 +
 657 +       if (pubsta) {
 658 +               u8 tid = skb->priority & IEEE80211_QOS_CTL_TID_MASK;
 659 +
 660 +               txq = pubsta->txq[tid];
 661 +       } else if (vif) {
 662 +               txq = vif->txq;
 663 +       }
 664 +
 665 +       if (!txq)
 666 +               goto tx_normal;
 667 +
 668 +       ac = txq->ac;
 669 +       txqi = to_txq_info(txq);
 670 +       atomic_inc(&sdata->txqs_len[ac]);
 671 +       if (atomic_read(&sdata->txqs_len[ac]) >= local->hw.txq_ac_max_pending)
 672 +               netif_stop_subqueue(sdata->dev, ac);
 673 +
 674 +       skb_queue_tail(&txqi->queue, skb);
 675 +       drv_wake_tx_queue(local, txqi);
 676 +
 677 +       return;
 678 +
 679 +tx_normal:
 680 +       drv_tx(local, &control, skb);
 681 +}
 682 +
 683 +struct sk_buff *ieee80211_tx_dequeue(struct ieee80211_hw *hw,
 684 +                                    struct ieee80211_txq *txq)
 685 +{
 686 +       struct ieee80211_local *local = hw_to_local(hw);
 687 +       struct ieee80211_sub_if_data *sdata = vif_to_sdata(txq->vif);
 688 +       struct txq_info *txqi = container_of(txq, struct txq_info, txq);
 689 +       struct ieee80211_hdr *hdr;
 690 +       struct sk_buff *skb = NULL;
 691 +       u8 ac = txq->ac;
 692 +
 693 +       spin_lock_bh(&txqi->queue.lock);
 694 +
 695 +       if (test_bit(IEEE80211_TXQ_STOP, &txqi->flags))
 696 +               goto out;
 697 +
 698 +       skb = __skb_dequeue(&txqi->queue);
 699 +       if (!skb)
 700 +               goto out;
 701 +
 702 +       atomic_dec(&sdata->txqs_len[ac]);
 703 +       if (__netif_subqueue_stopped(sdata->dev, ac))
 704 +               ieee80211_propagate_queue_wake(local, sdata->vif.hw_queue[ac]);
 705 +
 706 +       hdr = (struct ieee80211_hdr *)skb->data;
 707 +       if (txq->sta && ieee80211_is_data_qos(hdr->frame_control)) {
 708 +               struct sta_info *sta = container_of(txq->sta, struct sta_info,
 709 +                                                   sta);
 710 +               struct ieee80211_tx_info *info = IEEE80211_SKB_CB(skb);
 711 +
 712 +               hdr->seq_ctrl = ieee80211_tx_next_seq(sta, txq->tid);
 713 +               if (test_bit(IEEE80211_TXQ_AMPDU, &txqi->flags))
 714 +                       info->flags |= IEEE80211_TX_CTL_AMPDU;
 715 +               else
 716 +                       info->flags &= ~IEEE80211_TX_CTL_AMPDU;
 717 +       }
 718 +
 719 +out:
 720 +       spin_unlock_bh(&txqi->queue.lock);
 721 +
 722 +       return skb;
 723 +}
 724 +EXPORT_SYMBOL(ieee80211_tx_dequeue);
 725 +
 726  static bool ieee80211_tx_frags(struct ieee80211_local *local,
 727                                struct ieee80211_vif *vif,
 728                                struct ieee80211_sta *sta,
 729                                struct sk_buff_head *skbs,
 730                                bool txpending)
 731  {
 732 -       struct ieee80211_tx_control control;
 733         struct sk_buff *skb, *tmp;
 734         unsigned long flags;
 735
 736 @@ -1265,10 +1362,9 @@ static bool ieee80211_tx_frags(struct ie
 737                 spin_unlock_irqrestore(&local->queue_stop_reason_lock, flags);
 738
 739                 info->control.vif = vif;
 740 -               control.sta = sta;
 741
 742                 __skb_unlink(skb, skbs);
 743 -               drv_tx(local, &control, skb);
 744 +               ieee80211_drv_tx(local, vif, sta, skb);
 745         }
 746
 747         return true;
 748 --- a/net/mac80211/util.c
 749 +++ b/net/mac80211/util.c
 750 @@ -308,6 +308,11 @@ void ieee80211_propagate_queue_wake(stru
 751                 for (ac = 0; ac < n_acs; ac++) {
 752                         int ac_queue = sdata->vif.hw_queue[ac];
 753
 754 +                       if (local->ops->wake_tx_queue &&
 755 +                           (atomic_read(&sdata->txqs_len[ac]) >
 756 +                            local->hw.txq_ac_max_pending))
 757 +                               continue;
 758 +
 759                         if (ac_queue == queue ||
 760                             (sdata->vif.cab_queue == queue &&
 761                              local->queue_stop_reasons[ac_queue] == 0 &&
 762 @@ -3307,3 +3312,20 @@ u8 *ieee80211_add_wmm_info_ie(u8 *buf, u
 763
 764         return buf;
 765  }
 766 +
 767 +void ieee80211_init_tx_queue(struct ieee80211_sub_if_data *sdata,
 768 +                            struct sta_info *sta,
 769 +                            struct txq_info *txqi, int tid)
 770 +{
 771 +       skb_queue_head_init(&txqi->queue);
 772 +       txqi->txq.vif = &sdata->vif;
 773 +
 774 +       if (sta) {
 775 +               txqi->txq.sta = &sta->sta;
 776 +               sta->sta.txq[tid] = &txqi->txq;
 777 +               txqi->txq.ac = ieee802_1d_to_ac[tid & 7];
 778 +       } else {
 779 +               sdata->vif.txq = &txqi->txq;
 780 +               txqi->txq.ac = IEEE80211_AC_BE;
 781 +       }
 782 +}
 783 --- a/net/mac80211/rx.c
 784 +++ b/net/mac80211/rx.c
 785 @@ -1176,6 +1176,7 @@ static void sta_ps_start(struct sta_info
 786         struct ieee80211_sub_if_data *sdata = sta->sdata;
 787         struct ieee80211_local *local = sdata->local;
 788         struct ps_data *ps;
 789 +       int tid;
 790
 791         if (sta->sdata->vif.type == NL80211_IFTYPE_AP ||
 792             sta->sdata->vif.type == NL80211_IFTYPE_AP_VLAN)
 793 @@ -1189,6 +1190,18 @@ static void sta_ps_start(struct sta_info
 794                 drv_sta_notify(local, sdata, STA_NOTIFY_SLEEP, &sta->sta);
 795         ps_dbg(sdata, "STA %pM aid %d enters power save mode\n",
 796                sta->sta.addr, sta->sta.aid);
 797 +
 798 +       if (!sta->sta.txq[0])
 799 +               return;
 800 +
 801 +       for (tid = 0; tid < ARRAY_SIZE(sta->sta.txq); tid++) {
 802 +               struct txq_info *txqi = to_txq_info(sta->sta.txq[tid]);
 803 +
 804 +               if (!skb_queue_len(&txqi->queue))
 805 +                       set_bit(tid, &sta->txq_buffered_tids);
 806 +               else
 807 +                       clear_bit(tid, &sta->txq_buffered_tids);
 808 +       }
 809  }
 810
 811  static void sta_ps_end(struct sta_info *sta)
 812 --- a/net/mac80211/agg-tx.c
 813 +++ b/net/mac80211/agg-tx.c
 814 @@ -188,6 +188,43 @@ ieee80211_wake_queue_agg(struct ieee8021
 815         __release(agg_queue);
 816  }
 817
 818 +static void
 819 +ieee80211_agg_stop_txq(struct sta_info *sta, int tid)
 820 +{
 821 +       struct ieee80211_txq *txq = sta->sta.txq[tid];
 822 +       struct txq_info *txqi;
 823 +
 824 +       if (!txq)
 825 +               return;
 826 +
 827 +       txqi = to_txq_info(txq);
 828 +
 829 +       /* Lock here to protect against further seqno updates on dequeue */
 830 +       spin_lock_bh(&txqi->queue.lock);
 831 +       set_bit(IEEE80211_TXQ_STOP, &txqi->flags);
 832 +       spin_unlock_bh(&txqi->queue.lock);
 833 +}
 834 +
 835 +static void
 836 +ieee80211_agg_start_txq(struct sta_info *sta, int tid, bool enable)
 837 +{
 838 +       struct ieee80211_txq *txq = sta->sta.txq[tid];
 839 +       struct txq_info *txqi;
 840 +
 841 +       if (!txq)
 842 +               return;
 843 +
 844 +       txqi = to_txq_info(txq);
 845 +
 846 +       if (enable)
 847 +               set_bit(IEEE80211_TXQ_AMPDU, &txqi->flags);
 848 +       else
 849 +               clear_bit(IEEE80211_TXQ_AMPDU, &txqi->flags);
 850 +
 851 +       clear_bit(IEEE80211_TXQ_STOP, &txqi->flags);
 852 +       drv_wake_tx_queue(sta->sdata->local, txqi);
 853 +}
 854 +
 855  /*
 856   * splice packets from the STA's pending to the local pending,
 857   * requires a call to ieee80211_agg_splice_finish later
 858 @@ -247,6 +284,7 @@ static void ieee80211_remove_tid_tx(stru
 859         ieee80211_assign_tid_tx(sta, tid, NULL);
 860
 861         ieee80211_agg_splice_finish(sta->sdata, tid);
 862 +       ieee80211_agg_start_txq(sta, tid, false);
 863
 864         kfree_rcu(tid_tx, rcu_head);
 865  }
 866 @@ -418,6 +456,8 @@ void ieee80211_tx_ba_session_handle_star
 867          */
 868         clear_bit(HT_AGG_STATE_WANT_START, &tid_tx->state);
 869
 870 +       ieee80211_agg_stop_txq(sta, tid);
 871 +
 872         /*
 873          * Make sure no packets are being processed. This ensures that
 874          * we have a valid starting sequence number and that in-flight
 875 @@ -440,6 +480,8 @@ void ieee80211_tx_ba_session_handle_star
 876                 ieee80211_agg_splice_finish(sdata, tid);
 877                 spin_unlock_bh(&sta->lock);
 878
 879 +               ieee80211_agg_start_txq(sta, tid, false);
 880 +
 881                 kfree_rcu(tid_tx, rcu_head);
 882                 return;
 883         }
 884 @@ -666,6 +708,8 @@ static void ieee80211_agg_tx_operational
 885         ieee80211_agg_splice_finish(sta->sdata, tid);
 886
 887         spin_unlock_bh(&sta->lock);
 888 +
 889 +       ieee80211_agg_start_txq(sta, tid, true);
 890  }
 891
 892  void ieee80211_start_tx_ba_cb(struct ieee80211_vif *vif, u8 *ra, u16 tid)