drivers/net/wireless/ath/carl9170/tx.c

   1 /*
   2  * Atheros CARL9170 driver
   3  *
   4  * 802.11 xmit & status routines
   5  *
   6  * Copyright 2008, Johannes Berg <johannes@sipsolutions.net>
   7  * Copyright 2009, 2010, Christian Lamparter <chunkeey@googlemail.com>
   8  *
   9  * This program is free software; you can redistribute it and/or modify
  10  * it under the terms of the GNU General Public License as published by
  11  * the Free Software Foundation; either version 2 of the License, or
  12  * (at your option) any later version.
  13  *
  14  * This program is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  * GNU General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU General Public License
  20  * along with this program; see the file COPYING.  If not, see
  21  * http://www.gnu.org/licenses/.
  22  *
  23  * This file incorporates work covered by the following copyright and
  24  * permission notice:
  25  *    Copyright (c) 2007-2008 Atheros Communications, Inc.
  26  *
  27  *    Permission to use, copy, modify, and/or distribute this software for any
  28  *    purpose with or without fee is hereby granted, provided that the above
  29  *    copyright notice and this permission notice appear in all copies.
  30  *
  31  *    THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
  32  *    WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
  33  *    MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
  34  *    ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
  35  *    WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
  36  *    ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
  37  *    OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
  38  */
  39
  40 #include <linux/init.h>
  41 #include <linux/slab.h>
  42 #include <linux/module.h>
  43 #include <linux/etherdevice.h>
  44 #include <net/mac80211.h>
  45 #include "carl9170.h"
  46 #include "hw.h"
  47 #include "cmd.h"
  48
  49 static inline unsigned int __carl9170_get_queue(struct ar9170 *ar,
  50                                                 unsigned int queue)
  51 {
  52         if (unlikely(modparam_noht)) {
  53                 return queue;
  54         } else {
  55                 /*
  56                  * This is just another workaround, until
  57                  * someone figures out how to get QoS and
  58                  * AMPDU to play nicely together.
  59                  */
  60
  61                 return 2;               /* AC_BE */
  62         }
  63 }
  64
  65 static inline unsigned int carl9170_get_queue(struct ar9170 *ar,
  66                                               struct sk_buff *skb)
  67 {
  68         return __carl9170_get_queue(ar, skb_get_queue_mapping(skb));
  69 }
  70
  71 static bool is_mem_full(struct ar9170 *ar)
  72 {
  73         return (DIV_ROUND_UP(IEEE80211_MAX_FRAME_LEN, ar->fw.mem_block_size) >
  74                 atomic_read(&ar->mem_free_blocks));
  75 }
  76
  77 static void carl9170_tx_accounting(struct ar9170 *ar, struct sk_buff *skb)
  78 {
  79         int queue, i;
  80         bool mem_full;
  81
  82         atomic_inc(&ar->tx_total_queued);
  83
  84         queue = skb_get_queue_mapping(skb);
  85         spin_lock_bh(&ar->tx_stats_lock);
  86
  87         /*
  88          * The driver has to accept the frame, regardless if the queue is
  89          * full to the brim, or not. We have to do the queuing internally,
  90          * since mac80211 assumes that a driver which can operate with
  91          * aggregated frames does not reject frames for this reason.
  92          */
  93         ar->tx_stats[queue].len++;
  94         ar->tx_stats[queue].count++;
  95
  96         mem_full = is_mem_full(ar);
  97         for (i = 0; i < ar->hw->queues; i++) {
  98                 if (mem_full || ar->tx_stats[i].len >= ar->tx_stats[i].limit) {
  99                         ieee80211_stop_queue(ar->hw, i);
 100                         ar->queue_stop_timeout[i] = jiffies;
 101                 }
 102         }
 103
 104         spin_unlock_bh(&ar->tx_stats_lock);
 105 }
 106
 107 static void carl9170_tx_accounting_free(struct ar9170 *ar, struct sk_buff *skb)
 108 {
 109         struct ieee80211_tx_info *txinfo;
 110         int queue;
 111
 112         txinfo = IEEE80211_SKB_CB(skb);
 113         queue = skb_get_queue_mapping(skb);
 114
 115         spin_lock_bh(&ar->tx_stats_lock);
 116
 117         ar->tx_stats[queue].len--;
 118
 119         if (!is_mem_full(ar)) {
 120                 unsigned int i;
 121                 for (i = 0; i < ar->hw->queues; i++) {
 122                         if (ar->tx_stats[i].len >= CARL9170_NUM_TX_LIMIT_SOFT)
 123                                 continue;
 124
 125                         if (ieee80211_queue_stopped(ar->hw, i)) {
 126                                 unsigned long tmp;
 127
 128                                 tmp = jiffies - ar->queue_stop_timeout[i];
 129                                 if (tmp > ar->max_queue_stop_timeout[i])
 130                                         ar->max_queue_stop_timeout[i] = tmp;
 131                         }
 132
 133                         ieee80211_wake_queue(ar->hw, i);
 134                 }
 135         }
 136
 137         spin_unlock_bh(&ar->tx_stats_lock);
 138         if (atomic_dec_and_test(&ar->tx_total_queued))
 139                 complete(&ar->tx_flush);
 140 }
 141
 142 static int carl9170_alloc_dev_space(struct ar9170 *ar, struct sk_buff *skb)
 143 {
 144         struct _carl9170_tx_superframe *super = (void *) skb->data;
 145         unsigned int chunks;
 146         int cookie = -1;
 147
 148         atomic_inc(&ar->mem_allocs);
 149
 150         chunks = DIV_ROUND_UP(skb->len, ar->fw.mem_block_size);
 151         if (unlikely(atomic_sub_return(chunks, &ar->mem_free_blocks) < 0)) {
 152                 atomic_add(chunks, &ar->mem_free_blocks);
 153                 return -ENOSPC;
 154         }
 155
 156         spin_lock_bh(&ar->mem_lock);
 157         cookie = bitmap_find_free_region(ar->mem_bitmap, ar->fw.mem_blocks, 0);
 158         spin_unlock_bh(&ar->mem_lock);
 159
 160         if (unlikely(cookie < 0)) {
 161                 atomic_add(chunks, &ar->mem_free_blocks);
 162                 return -ENOSPC;
 163         }
 164
 165         super = (void *) skb->data;
 166
 167         /*
 168          * Cookie #0 serves two special purposes:
 169          *  1. The firmware might use it generate BlockACK frames
 170          *     in responds of an incoming BlockAckReqs.
 171          *
 172          *  2. Prevent double-free bugs.
 173          */
 174         super->s.cookie = (u8) cookie + 1;
 175         return 0;
 176 }
 177
 178 static void carl9170_release_dev_space(struct ar9170 *ar, struct sk_buff *skb)
 179 {
 180         struct _carl9170_tx_superframe *super = (void *) skb->data;
 181         int cookie;
 182
 183         /* make a local copy of the cookie */
 184         cookie = super->s.cookie;
 185         /* invalidate cookie */
 186         super->s.cookie = 0;
 187
 188         /*
 189          * Do a out-of-bounds check on the cookie:
 190          *
 191          *  * cookie "0" is reserved and won't be assigned to any
 192          *    out-going frame. Internally however, it is used to
 193          *    mark no longer/un-accounted frames and serves as a
 194          *    cheap way of preventing frames from being freed
 195          *    twice by _accident_. NB: There is a tiny race...
 196          *
 197          *  * obviously, cookie number is limited by the amount
 198          *    of available memory blocks, so the number can
 199          *    never execeed the mem_blocks count.
 200          */
 201         if (unlikely(WARN_ON_ONCE(cookie == 0) ||
 202             WARN_ON_ONCE(cookie > ar->fw.mem_blocks)))
 203                 return;
 204
 205         atomic_add(DIV_ROUND_UP(skb->len, ar->fw.mem_block_size),
 206                    &ar->mem_free_blocks);
 207
 208         spin_lock_bh(&ar->mem_lock);
 209         bitmap_release_region(ar->mem_bitmap, cookie - 1, 0);
 210         spin_unlock_bh(&ar->mem_lock);
 211 }
 212
 213 /* Called from any context */
 214 static void carl9170_tx_release(struct kref *ref)
 215 {
 216         struct ar9170 *ar;
 217         struct carl9170_tx_info *arinfo;
 218         struct ieee80211_tx_info *txinfo;
 219         struct sk_buff *skb;
 220
 221         arinfo = container_of(ref, struct carl9170_tx_info, ref);
 222         txinfo = container_of((void *) arinfo, struct ieee80211_tx_info,
 223                               rate_driver_data);
 224         skb = container_of((void *) txinfo, struct sk_buff, cb);
 225
 226         ar = arinfo->ar;
 227         if (WARN_ON_ONCE(!ar))
 228                 return;
 229
 230         BUILD_BUG_ON(
 231             offsetof(struct ieee80211_tx_info, status.ampdu_ack_len) != 23);
 232
 233         memset(&txinfo->status.ampdu_ack_len, 0,
 234                sizeof(struct ieee80211_tx_info) -
 235                offsetof(struct ieee80211_tx_info, status.ampdu_ack_len));
 236
 237         if (atomic_read(&ar->tx_total_queued))
 238                 ar->tx_schedule = true;
 239
 240         if (txinfo->flags & IEEE80211_TX_CTL_AMPDU) {
 241                 if (!atomic_read(&ar->tx_ampdu_upload))
 242                         ar->tx_ampdu_schedule = true;
 243
 244                 if (txinfo->flags & IEEE80211_TX_STAT_AMPDU) {
 245                         txinfo->status.ampdu_len = txinfo->pad[0];
 246                         txinfo->status.ampdu_ack_len = txinfo->pad[1];
 247                         txinfo->pad[0] = txinfo->pad[1] = 0;
 248                 } else if (txinfo->flags & IEEE80211_TX_STAT_ACK) {
 249                         /*
 250                          * drop redundant tx_status reports:
 251                          *
 252                          * 1. ampdu_ack_len of the final tx_status does
 253                          *    include the feedback of this particular frame.
 254                          *
 255                          * 2. tx_status_irqsafe only queues up to 128
 256                          *    tx feedback reports and discards the rest.
 257                          *
 258                          * 3. minstrel_ht is picky, it only accepts
 259                          *    reports of frames with the TX_STATUS_AMPDU flag.
 260                          */
 261
 262                         dev_kfree_skb_any(skb);
 263                         return;
 264                 } else {
 265                         /*
 266                          * Frame has failed, but we want to keep it in
 267                          * case it was lost due to a power-state
 268                          * transition.
 269                          */
 270                 }
 271         }
 272
 273         skb_pull(skb, sizeof(struct _carl9170_tx_superframe));
 274         ieee80211_tx_status_irqsafe(ar->hw, skb);
 275 }
 276
 277 void carl9170_tx_get_skb(struct sk_buff *skb)
 278 {
 279         struct carl9170_tx_info *arinfo = (void *)
 280                 (IEEE80211_SKB_CB(skb))->rate_driver_data;
 281         kref_get(&arinfo->ref);
 282 }
 283
 284 int carl9170_tx_put_skb(struct sk_buff *skb)
 285 {
 286         struct carl9170_tx_info *arinfo = (void *)
 287                 (IEEE80211_SKB_CB(skb))->rate_driver_data;
 288
 289         return kref_put(&arinfo->ref, carl9170_tx_release);
 290 }
 291
 292 /* Caller must hold the tid_info->lock & rcu_read_lock */
 293 static void carl9170_tx_shift_bm(struct ar9170 *ar,
 294         struct carl9170_sta_tid *tid_info, u16 seq)
 295 {
 296         u16 off;
 297
 298         off = SEQ_DIFF(seq, tid_info->bsn);
 299
 300         if (WARN_ON_ONCE(off >= CARL9170_BAW_BITS))
 301                 return;
 302
 303         /*
 304          * Sanity check. For each MPDU we set the bit in bitmap and
 305          * clear it once we received the tx_status.
 306          * But if the bit is already cleared then we've been bitten
 307          * by a bug.
 308          */
 309         WARN_ON_ONCE(!test_and_clear_bit(off, tid_info->bitmap));
 310
 311         off = SEQ_DIFF(tid_info->snx, tid_info->bsn);
 312         if (WARN_ON_ONCE(off >= CARL9170_BAW_BITS))
 313                 return;
 314
 315         if (!bitmap_empty(tid_info->bitmap, off))
 316                 off = find_first_bit(tid_info->bitmap, off);
 317
 318         tid_info->bsn += off;
 319         tid_info->bsn &= 0x0fff;
 320
 321         bitmap_shift_right(tid_info->bitmap, tid_info->bitmap,
 322                            off, CARL9170_BAW_BITS);
 323 }
 324
 325 static void carl9170_tx_status_process_ampdu(struct ar9170 *ar,
 326         struct sk_buff *skb, struct ieee80211_tx_info *txinfo)
 327 {
 328         struct _carl9170_tx_superframe *super = (void *) skb->data;
 329         struct ieee80211_hdr *hdr = (void *) super->frame_data;
 330         struct ieee80211_tx_info *tx_info;
 331         struct carl9170_tx_info *ar_info;
 332         struct carl9170_sta_info *sta_info;
 333         struct ieee80211_sta *sta;
 334         struct carl9170_sta_tid *tid_info;
 335         struct ieee80211_vif *vif;
 336         unsigned int vif_id;
 337         u8 tid;
 338
 339         if (!(txinfo->flags & IEEE80211_TX_CTL_AMPDU) ||
 340             txinfo->flags & IEEE80211_TX_CTL_INJECTED)
 341                 return;
 342
 343         tx_info = IEEE80211_SKB_CB(skb);
 344         ar_info = (void *) tx_info->rate_driver_data;
 345
 346         vif_id = (super->s.misc & CARL9170_TX_SUPER_MISC_VIF_ID) >>
 347                  CARL9170_TX_SUPER_MISC_VIF_ID_S;
 348
 349         if (WARN_ON_ONCE(vif_id >= AR9170_MAX_VIRTUAL_MAC))
 350                 return;
 351
 352         rcu_read_lock();
 353         vif = rcu_dereference(ar->vif_priv[vif_id].vif);
 354         if (unlikely(!vif))
 355                 goto out_rcu;
 356
 357         /*
 358          * Normally we should use wrappers like ieee80211_get_DA to get
 359          * the correct peer ieee80211_sta.
 360          *
 361          * But there is a problem with indirect traffic (broadcasts, or
 362          * data which is designated for other stations) in station mode.
 363          * The frame will be directed to the AP for distribution and not
 364          * to the actual destination.
 365          */
 366         sta = ieee80211_find_sta(vif, hdr->addr1);
 367         if (unlikely(!sta))
 368                 goto out_rcu;
 369
 370         tid = get_tid_h(hdr);
 371
 372         sta_info = (void *) sta->drv_priv;
 373         tid_info = rcu_dereference(sta_info->agg[tid]);
 374         if (!tid_info)
 375                 goto out_rcu;
 376
 377         spin_lock_bh(&tid_info->lock);
 378         if (likely(tid_info->state >= CARL9170_TID_STATE_IDLE))
 379                 carl9170_tx_shift_bm(ar, tid_info, get_seq_h(hdr));
 380
 381         if (sta_info->stats[tid].clear) {
 382                 sta_info->stats[tid].clear = false;
 383                 sta_info->stats[tid].ampdu_len = 0;
 384                 sta_info->stats[tid].ampdu_ack_len = 0;
 385         }
 386
 387         sta_info->stats[tid].ampdu_len++;
 388         if (txinfo->status.rates[0].count == 1)
 389                 sta_info->stats[tid].ampdu_ack_len++;
 390
 391         if (super->f.mac_control & cpu_to_le16(AR9170_TX_MAC_IMM_BA)) {
 392                 txinfo->pad[0] = sta_info->stats[tid].ampdu_len;
 393                 txinfo->pad[1] = sta_info->stats[tid].ampdu_ack_len;
 394                 txinfo->flags |= IEEE80211_TX_STAT_AMPDU;
 395                 sta_info->stats[tid].clear = true;
 396         }
 397         spin_unlock_bh(&tid_info->lock);
 398
 399 out_rcu:
 400         rcu_read_unlock();
 401 }
 402
 403 void carl9170_tx_status(struct ar9170 *ar, struct sk_buff *skb,
 404                         const bool success)
 405 {
 406         struct ieee80211_tx_info *txinfo;
 407
 408         carl9170_tx_accounting_free(ar, skb);
 409
 410         txinfo = IEEE80211_SKB_CB(skb);
 411
 412         if (success)
 413                 txinfo->flags |= IEEE80211_TX_STAT_ACK;
 414         else
 415                 ar->tx_ack_failures++;
 416
 417         if (txinfo->flags & IEEE80211_TX_CTL_AMPDU)
 418                 carl9170_tx_status_process_ampdu(ar, skb, txinfo);
 419
 420         carl9170_tx_put_skb(skb);
 421 }
 422
 423 /* This function may be called form any context */
 424 void carl9170_tx_callback(struct ar9170 *ar, struct sk_buff *skb)
 425 {
 426         struct ieee80211_tx_info *txinfo = IEEE80211_SKB_CB(skb);
 427
 428         atomic_dec(&ar->tx_total_pending);
 429
 430         if (txinfo->flags & IEEE80211_TX_CTL_AMPDU)
 431                 atomic_dec(&ar->tx_ampdu_upload);
 432
 433         if (carl9170_tx_put_skb(skb))
 434                 tasklet_hi_schedule(&ar->usb_tasklet);
 435 }
 436
 437 static struct sk_buff *carl9170_get_queued_skb(struct ar9170 *ar, u8 cookie,
 438                                                struct sk_buff_head *queue)
 439 {
 440         struct sk_buff *skb;
 441
 442         spin_lock_bh(&queue->lock);
 443         skb_queue_walk(queue, skb) {
 444                 struct _carl9170_tx_superframe *txc = (void *) skb->data;
 445
 446                 if (txc->s.cookie != cookie)
 447                         continue;
 448
 449                 __skb_unlink(skb, queue);
 450                 spin_unlock_bh(&queue->lock);
 451
 452                 carl9170_release_dev_space(ar, skb);
 453                 return skb;
 454         }
 455         spin_unlock_bh(&queue->lock);
 456
 457         return NULL;
 458 }
 459
 460 static void carl9170_tx_fill_rateinfo(struct ar9170 *ar, unsigned int rix,
 461         unsigned int tries, struct ieee80211_tx_info *txinfo)
 462 {
 463         unsigned int i;
 464
 465         for (i = 0; i < IEEE80211_TX_MAX_RATES; i++) {
 466                 if (txinfo->status.rates[i].idx < 0)
 467                         break;
 468
 469                 if (i == rix) {
 470                         txinfo->status.rates[i].count = tries;
 471                         i++;
 472                         break;
 473                 }
 474         }
 475
 476         for (; i < IEEE80211_TX_MAX_RATES; i++) {
 477                 txinfo->status.rates[i].idx = -1;
 478                 txinfo->status.rates[i].count = 0;
 479         }
 480 }
 481
 482 static void carl9170_check_queue_stop_timeout(struct ar9170 *ar)
 483 {
 484         int i;
 485         struct sk_buff *skb;
 486         struct ieee80211_tx_info *txinfo;
 487         struct carl9170_tx_info *arinfo;
 488         bool restart = false;
 489
 490         for (i = 0; i < ar->hw->queues; i++) {
 491                 spin_lock_bh(&ar->tx_status[i].lock);
 492
 493                 skb = skb_peek(&ar->tx_status[i]);
 494
 495                 if (!skb)
 496                         goto next;
 497
 498                 txinfo = IEEE80211_SKB_CB(skb);
 499                 arinfo = (void *) txinfo->rate_driver_data;
 500
 501                 if (time_is_before_jiffies(arinfo->timeout +
 502                     msecs_to_jiffies(CARL9170_QUEUE_STUCK_TIMEOUT)) == true)
 503                         restart = true;
 504
 505 next:
 506                 spin_unlock_bh(&ar->tx_status[i].lock);
 507         }
 508
 509         if (restart) {
 510                 /*
 511                  * At least one queue has been stuck for long enough.
 512                  * Give the device a kick and hope it gets back to
 513                  * work.
 514                  *
 515                  * possible reasons may include:
 516                  *  - frames got lost/corrupted (bad connection to the device)
 517                  *  - stalled rx processing/usb controller hiccups
 518                  *  - firmware errors/bugs
 519                  *  - every bug you can think of.
 520                  *  - all bugs you can't...
 521                  *  - ...
 522                  */
 523                 carl9170_restart(ar, CARL9170_RR_STUCK_TX);
 524         }
 525 }
 526
 527 void carl9170_tx_janitor(struct work_struct *work)
 528 {
 529         struct ar9170 *ar = container_of(work, struct ar9170,
 530                                          tx_janitor.work);
 531         if (!IS_STARTED(ar))
 532                 return;
 533
 534         ar->tx_janitor_last_run = jiffies;
 535
 536         carl9170_check_queue_stop_timeout(ar);
 537
 538         if (!atomic_read(&ar->tx_total_queued))
 539                 return;
 540
 541         ieee80211_queue_delayed_work(ar->hw, &ar->tx_janitor,
 542                 msecs_to_jiffies(CARL9170_TX_TIMEOUT));
 543 }
 544
 545 static void __carl9170_tx_process_status(struct ar9170 *ar,
 546         const uint8_t cookie, const uint8_t info)
 547 {
 548         struct sk_buff *skb;
 549         struct ieee80211_tx_info *txinfo;
 550         struct carl9170_tx_info *arinfo;
 551         unsigned int r, t, q;
 552         bool success = true;
 553
 554         q = ar9170_qmap[info & CARL9170_TX_STATUS_QUEUE];
 555
 556         skb = carl9170_get_queued_skb(ar, cookie, &ar->tx_status[q]);
 557         if (!skb) {
 558                 /*
 559                  * We have lost the race to another thread.
 560                  */
 561
 562                 return ;
 563         }
 564
 565         txinfo = IEEE80211_SKB_CB(skb);
 566         arinfo = (void *) txinfo->rate_driver_data;
 567
 568         if (!(info & CARL9170_TX_STATUS_SUCCESS))
 569                 success = false;
 570
 571         r = (info & CARL9170_TX_STATUS_RIX) >> CARL9170_TX_STATUS_RIX_S;
 572         t = (info & CARL9170_TX_STATUS_TRIES) >> CARL9170_TX_STATUS_TRIES_S;
 573
 574         carl9170_tx_fill_rateinfo(ar, r, t, txinfo);
 575         carl9170_tx_status(ar, skb, success);
 576 }
 577
 578 void carl9170_tx_process_status(struct ar9170 *ar,
 579                                 const struct carl9170_rsp *cmd)
 580 {
 581         unsigned int i;
 582
 583         for (i = 0;  i < cmd->hdr.ext; i++) {
 584                 if (WARN_ON(i > ((cmd->hdr.len / 2) + 1))) {
 585                         print_hex_dump_bytes("UU:", DUMP_PREFIX_NONE,
 586                                              (void *) cmd, cmd->hdr.len + 4);
 587                         break;
 588                 }
 589
 590                 __carl9170_tx_process_status(ar, cmd->_tx_status[i].cookie,
 591                                              cmd->_tx_status[i].info);
 592         }
 593 }
 594
 595 static __le32 carl9170_tx_physet(struct ar9170 *ar,
 596         struct ieee80211_tx_info *info, struct ieee80211_tx_rate *txrate)
 597 {
 598         struct ieee80211_rate *rate = NULL;
 599         u32 power, chains;
 600         __le32 tmp;
 601
 602         tmp = cpu_to_le32(0);
 603
 604         if (txrate->flags & IEEE80211_TX_RC_40_MHZ_WIDTH)
 605                 tmp |= cpu_to_le32(AR9170_TX_PHY_BW_40MHZ <<
 606                         AR9170_TX_PHY_BW_S);
 607         /* this works because 40 MHz is 2 and dup is 3 */
 608         if (txrate->flags & IEEE80211_TX_RC_DUP_DATA)
 609                 tmp |= cpu_to_le32(AR9170_TX_PHY_BW_40MHZ_DUP <<
 610                         AR9170_TX_PHY_BW_S);
 611
 612         if (txrate->flags & IEEE80211_TX_RC_SHORT_GI)
 613                 tmp |= cpu_to_le32(AR9170_TX_PHY_SHORT_GI);
 614
 615         if (txrate->flags & IEEE80211_TX_RC_MCS) {
 616                 u32 r = txrate->idx;
 617                 u8 *txpower;
 618
 619                 /* heavy clip control */
 620                 tmp |= cpu_to_le32((r & 0x7) <<
 621                         AR9170_TX_PHY_TX_HEAVY_CLIP_S);
 622
 623                 if (txrate->flags & IEEE80211_TX_RC_40_MHZ_WIDTH) {
 624                         if (info->band == IEEE80211_BAND_5GHZ)
 625                                 txpower = ar->power_5G_ht40;
 626                         else
 627                                 txpower = ar->power_2G_ht40;
 628                 } else {
 629                         if (info->band == IEEE80211_BAND_5GHZ)
 630                                 txpower = ar->power_5G_ht20;
 631                         else
 632                                 txpower = ar->power_2G_ht20;
 633                 }
 634
 635                 power = txpower[r & 7];
 636
 637                 /* +1 dBm for HT40 */
 638                 if (txrate->flags & IEEE80211_TX_RC_40_MHZ_WIDTH)
 639                         power += 2;
 640
 641                 r <<= AR9170_TX_PHY_MCS_S;
 642                 BUG_ON(r & ~AR9170_TX_PHY_MCS);
 643
 644                 tmp |= cpu_to_le32(r & AR9170_TX_PHY_MCS);
 645                 tmp |= cpu_to_le32(AR9170_TX_PHY_MOD_HT);
 646
 647                 /*
 648                  * green field preamble does not work.
 649                  *
 650                  * if (txrate->flags & IEEE80211_TX_RC_GREEN_FIELD)
 651                  * tmp |= cpu_to_le32(AR9170_TX_PHY_GREENFIELD);
 652                  */
 653         } else {
 654                 u8 *txpower;
 655                 u32 mod;
 656                 u32 phyrate;
 657                 u8 idx = txrate->idx;
 658
 659                 if (info->band != IEEE80211_BAND_2GHZ) {
 660                         idx += 4;
 661                         txpower = ar->power_5G_leg;
 662                         mod = AR9170_TX_PHY_MOD_OFDM;
 663                 } else {
 664                         if (idx < 4) {
 665                                 txpower = ar->power_2G_cck;
 666                                 mod = AR9170_TX_PHY_MOD_CCK;
 667                         } else {
 668                                 mod = AR9170_TX_PHY_MOD_OFDM;
 669                                 txpower = ar->power_2G_ofdm;
 670                         }
 671                 }
 672
 673                 rate = &__carl9170_ratetable[idx];
 674
 675                 phyrate = rate->hw_value & 0xF;
 676                 power = txpower[(rate->hw_value & 0x30) >> 4];
 677                 phyrate <<= AR9170_TX_PHY_MCS_S;
 678
 679                 tmp |= cpu_to_le32(mod);
 680                 tmp |= cpu_to_le32(phyrate);
 681
 682                 /*
 683                  * short preamble seems to be broken too.
 684                  *
 685                  * if (txrate->flags & IEEE80211_TX_RC_USE_SHORT_PREAMBLE)
 686                  *      tmp |= cpu_to_le32(AR9170_TX_PHY_SHORT_PREAMBLE);
 687                  */
 688         }
 689         power <<= AR9170_TX_PHY_TX_PWR_S;
 690         power &= AR9170_TX_PHY_TX_PWR;
 691         tmp |= cpu_to_le32(power);
 692
 693         /* set TX chains */
 694         if (ar->eeprom.tx_mask == 1) {
 695                 chains = AR9170_TX_PHY_TXCHAIN_1;
 696         } else {
 697                 chains = AR9170_TX_PHY_TXCHAIN_2;
 698
 699                 /* >= 36M legacy OFDM - use only one chain */
 700                 if (rate && rate->bitrate >= 360 &&
 701                     !(txrate->flags & IEEE80211_TX_RC_MCS))
 702                         chains = AR9170_TX_PHY_TXCHAIN_1;
 703         }
 704         tmp |= cpu_to_le32(chains << AR9170_TX_PHY_TXCHAIN_S);
 705
 706         return tmp;
 707 }
 708
 709 static bool carl9170_tx_rts_check(struct ar9170 *ar,
 710                                   struct ieee80211_tx_rate *rate,
 711                                   bool ampdu, bool multi)
 712 {
 713         switch (ar->erp_mode) {
 714         case CARL9170_ERP_AUTO:
 715                 if (ampdu)
 716                         break;
 717
 718         case CARL9170_ERP_MAC80211:
 719                 if (!(rate->flags & IEEE80211_TX_RC_USE_RTS_CTS))
 720                         break;
 721
 722         case CARL9170_ERP_RTS:
 723                 if (likely(!multi))
 724                         return true;
 725
 726         default:
 727                 break;
 728         }
 729
 730         return false;
 731 }
 732
 733 static bool carl9170_tx_cts_check(struct ar9170 *ar,
 734                                   struct ieee80211_tx_rate *rate)
 735 {
 736         switch (ar->erp_mode) {
 737         case CARL9170_ERP_AUTO:
 738         case CARL9170_ERP_MAC80211:
 739                 if (!(rate->flags & IEEE80211_TX_RC_USE_CTS_PROTECT))
 740                         break;
 741
 742         case CARL9170_ERP_CTS:
 743                 return true;
 744
 745         default:
 746                 break;
 747         }
 748
 749         return false;
 750 }
 751
 752 static int carl9170_tx_prepare(struct ar9170 *ar, struct sk_buff *skb)
 753 {
 754         struct ieee80211_hdr *hdr;
 755         struct _carl9170_tx_superframe *txc;
 756         struct carl9170_vif_info *cvif;
 757         struct ieee80211_tx_info *info;
 758         struct ieee80211_tx_rate *txrate;
 759         struct ieee80211_sta *sta;
 760         struct carl9170_tx_info *arinfo;
 761         unsigned int hw_queue;
 762         int i;
 763         u16 keytype = 0;
 764         u16 len, icv = 0;
 765         bool ampdu, no_ack;
 766
 767         BUILD_BUG_ON(sizeof(*arinfo) > sizeof(info->rate_driver_data));
 768         BUILD_BUG_ON(sizeof(struct _carl9170_tx_superdesc) !=
 769                      CARL9170_TX_SUPERDESC_LEN);
 770
 771         BUILD_BUG_ON(sizeof(struct _ar9170_tx_hwdesc) !=
 772                      AR9170_TX_HWDESC_LEN);
 773
 774         BUILD_BUG_ON(IEEE80211_TX_MAX_RATES < CARL9170_TX_MAX_RATES);
 775
 776         hw_queue = ar9170_qmap[carl9170_get_queue(ar, skb)];
 777
 778         hdr = (void *)skb->data;
 779         info = IEEE80211_SKB_CB(skb);
 780         len = skb->len;
 781
 782         /*
 783          * Note: If the frame was sent through a monitor interface,
 784          * the ieee80211_vif pointer can be NULL.
 785          */
 786         if (likely(info->control.vif))
 787                 cvif = (void *) info->control.vif->drv_priv;
 788         else
 789                 cvif = NULL;
 790
 791         sta = info->control.sta;
 792
 793         txc = (void *)skb_push(skb, sizeof(*txc));
 794         memset(txc, 0, sizeof(*txc));
 795
 796         ampdu = !!(info->flags & IEEE80211_TX_CTL_AMPDU);
 797         no_ack = !!(info->flags & IEEE80211_TX_CTL_NO_ACK);
 798
 799         if (info->control.hw_key) {
 800                 icv = info->control.hw_key->icv_len;
 801
 802                 switch (info->control.hw_key->cipher) {
 803                 case WLAN_CIPHER_SUITE_WEP40:
 804                 case WLAN_CIPHER_SUITE_WEP104:
 805                 case WLAN_CIPHER_SUITE_TKIP:
 806                         keytype = AR9170_TX_MAC_ENCR_RC4;
 807                         break;
 808                 case WLAN_CIPHER_SUITE_CCMP:
 809                         keytype = AR9170_TX_MAC_ENCR_AES;
 810                         break;
 811                 default:
 812                         WARN_ON(1);
 813                         goto err_out;
 814                 }
 815         }
 816
 817         BUILD_BUG_ON(AR9170_MAX_VIRTUAL_MAC >
 818                 ((CARL9170_TX_SUPER_MISC_VIF_ID >>
 819                  CARL9170_TX_SUPER_MISC_VIF_ID_S) + 1));
 820
 821         txc->s.len = cpu_to_le16(len + sizeof(*txc));
 822         txc->f.length = cpu_to_le16(len + icv + 4);
 823         SET_VAL(CARL9170_TX_SUPER_MISC_VIF_ID, txc->s.misc,
 824                 cvif ? cvif->id : 0);
 825
 826         txc->f.mac_control = cpu_to_le16(AR9170_TX_MAC_HW_DURATION |
 827                                          AR9170_TX_MAC_BACKOFF);
 828
 829         SET_VAL(CARL9170_TX_SUPER_MISC_QUEUE, txc->s.misc, hw_queue);
 830
 831         txc->f.mac_control |= cpu_to_le16(hw_queue << AR9170_TX_MAC_QOS_S);
 832         txc->f.mac_control |= cpu_to_le16(keytype);
 833         txc->f.phy_control = cpu_to_le32(0);
 834
 835         if (no_ack)
 836                 txc->f.mac_control |= cpu_to_le16(AR9170_TX_MAC_NO_ACK);
 837
 838         if (info->flags & IEEE80211_TX_CTL_SEND_AFTER_DTIM)
 839                 txc->s.misc |= CARL9170_TX_SUPER_MISC_CAB;
 840
 841         txrate = &info->control.rates[0];
 842         if (carl9170_tx_rts_check(ar, txrate, ampdu, no_ack))
 843                 txc->f.mac_control |= cpu_to_le16(AR9170_TX_MAC_PROT_RTS);
 844         else if (carl9170_tx_cts_check(ar, txrate))
 845                 txc->f.mac_control |= cpu_to_le16(AR9170_TX_MAC_PROT_CTS);
 846
 847         SET_VAL(CARL9170_TX_SUPER_RI_TRIES, txc->s.ri[0], txrate->count);
 848         txc->f.phy_control |= carl9170_tx_physet(ar, info, txrate);
 849
 850         if (info->flags & IEEE80211_TX_CTL_AMPDU) {
 851                 for (i = 1; i < CARL9170_TX_MAX_RATES; i++) {
 852                         txrate = &info->control.rates[i];
 853                         if (txrate->idx >= 0)
 854                                 continue;
 855
 856                         txrate->idx = 0;
 857                         txrate->count = ar->hw->max_rate_tries;
 858                 }
 859         }
 860
 861         /*
 862          * NOTE: For the first rate, the ERP & AMPDU flags are directly
 863          * taken from mac_control. For all fallback rate, the firmware
 864          * updates the mac_control flags from the rate info field.
 865          */
 866         for (i = 1; i < CARL9170_TX_MAX_RATES; i++) {
 867                 txrate = &info->control.rates[i];
 868                 if (txrate->idx < 0)
 869                         break;
 870
 871                 SET_VAL(CARL9170_TX_SUPER_RI_TRIES, txc->s.ri[i],
 872                         txrate->count);
 873
 874                 if (carl9170_tx_rts_check(ar, txrate, ampdu, no_ack))
 875                         txc->s.ri[i] |= (AR9170_TX_MAC_PROT_RTS <<
 876                                 CARL9170_TX_SUPER_RI_ERP_PROT_S);
 877                 else if (carl9170_tx_cts_check(ar, txrate))
 878                         txc->s.ri[i] |= (AR9170_TX_MAC_PROT_CTS <<
 879                                 CARL9170_TX_SUPER_RI_ERP_PROT_S);
 880
 881                 /*
 882                  * unaggregated fallback, in case aggregation
 883                  * proves to be unsuccessful and unreliable.
 884                  */
 885                 if (ampdu && i < 3)
 886                         txc->s.ri[i] |= CARL9170_TX_SUPER_RI_AMPDU;
 887
 888                 txc->s.rr[i - 1] = carl9170_tx_physet(ar, info, txrate);
 889         }
 890
 891         if (ieee80211_is_probe_resp(hdr->frame_control))
 892                 txc->s.misc |= CARL9170_TX_SUPER_MISC_FILL_IN_TSF;
 893
 894         if (ampdu) {
 895                 unsigned int density, factor;
 896
 897                 if (unlikely(!sta || !cvif))
 898                         goto err_out;
 899
 900                 density = info->control.sta->ht_cap.ampdu_density;
 901                 factor = info->control.sta->ht_cap.ampdu_factor;
 902
 903                 if (density) {
 904                         /*
 905                          * Watch out!
 906                          *
 907                          * Otus uses slightly different density values than
 908                          * those from the 802.11n spec.
 909                          */
 910
 911                         density = max_t(unsigned int, density + 1, 7u);
 912                 }
 913
 914                 factor = min_t(unsigned int, 1u, factor);
 915
 916                 SET_VAL(CARL9170_TX_SUPER_AMPDU_DENSITY,
 917                         txc->s.ampdu_settings, density);
 918
 919                 SET_VAL(CARL9170_TX_SUPER_AMPDU_FACTOR,
 920                         txc->s.ampdu_settings, factor);
 921
 922                 if (info->control.rates[0].flags & IEEE80211_TX_RC_MCS) {
 923                         txc->f.mac_control |= cpu_to_le16(AR9170_TX_MAC_AGGR);
 924                 } else {
 925                         /*
 926                          * Not sure if it's even possible to aggregate
 927                          * non-ht rates with this HW.
 928                          */
 929                         WARN_ON_ONCE(1);
 930                 }
 931         }
 932
 933         arinfo = (void *)info->rate_driver_data;
 934         arinfo->timeout = jiffies;
 935         arinfo->ar = ar;
 936         kref_init(&arinfo->ref);
 937         return 0;
 938
 939 err_out:
 940         skb_pull(skb, sizeof(*txc));
 941         return -EINVAL;
 942 }
 943
 944 static void carl9170_set_immba(struct ar9170 *ar, struct sk_buff *skb)
 945 {
 946         struct _carl9170_tx_superframe *super;
 947
 948         super = (void *) skb->data;
 949         super->f.mac_control |= cpu_to_le16(AR9170_TX_MAC_IMM_BA);
 950 }
 951
 952 static void carl9170_set_ampdu_params(struct ar9170 *ar, struct sk_buff *skb)
 953 {
 954         struct _carl9170_tx_superframe *super;
 955         int tmp;
 956
 957         super = (void *) skb->data;
 958
 959         tmp = (super->s.ampdu_settings & CARL9170_TX_SUPER_AMPDU_DENSITY) <<
 960                 CARL9170_TX_SUPER_AMPDU_DENSITY_S;
 961
 962         /*
 963          * If you haven't noticed carl9170_tx_prepare has already filled
 964          * in all ampdu spacing & factor parameters.
 965          * Now it's the time to check whenever the settings have to be
 966          * updated by the firmware, or if everything is still the same.
 967          *
 968          * There's no sane way to handle different density values with
 969          * this hardware, so we may as well just do the compare in the
 970          * driver.
 971          */
 972
 973         if (tmp != ar->current_density) {
 974                 ar->current_density = tmp;
 975                 super->s.ampdu_settings |=
 976                         CARL9170_TX_SUPER_AMPDU_COMMIT_DENSITY;
 977         }
 978
 979         tmp = (super->s.ampdu_settings & CARL9170_TX_SUPER_AMPDU_FACTOR) <<
 980                 CARL9170_TX_SUPER_AMPDU_FACTOR_S;
 981
 982         if (tmp != ar->current_factor) {
 983                 ar->current_factor = tmp;
 984                 super->s.ampdu_settings |=
 985                         CARL9170_TX_SUPER_AMPDU_COMMIT_FACTOR;
 986         }
 987 }
 988
 989 static bool carl9170_tx_rate_check(struct ar9170 *ar, struct sk_buff *_dest,
 990                                    struct sk_buff *_src)
 991 {
 992         struct _carl9170_tx_superframe *dest, *src;
 993
 994         dest = (void *) _dest->data;
 995         src = (void *) _src->data;
 996
 997         /*
 998          * The mac80211 rate control algorithm expects that all MPDUs in
 999          * an AMPDU share the same tx vectors.
1000          * This is not really obvious right now, because the hardware
1001          * does the AMPDU setup according to its own rulebook.
1002          * Our nicely assembled, strictly monotonic increasing mpdu
1003          * chains will be broken up, mashed back together...
1004          */
1005
1006         return (dest->f.phy_control == src->f.phy_control);
1007 }
1008
1009 static void carl9170_tx_ampdu(struct ar9170 *ar)
1010 {
1011         struct sk_buff_head agg;
1012         struct carl9170_sta_tid *tid_info;
1013         struct sk_buff *skb, *first;
1014         unsigned int i = 0, done_ampdus = 0;
1015         u16 seq, queue, tmpssn;
1016
1017         atomic_inc(&ar->tx_ampdu_scheduler);
1018         ar->tx_ampdu_schedule = false;
1019
1020         if (atomic_read(&ar->tx_ampdu_upload))
1021                 return;
1022
1023         if (!ar->tx_ampdu_list_len)
1024                 return;
1025
1026         __skb_queue_head_init(&agg);
1027
1028         rcu_read_lock();
1029         tid_info = rcu_dereference(ar->tx_ampdu_iter);
1030         if (WARN_ON_ONCE(!tid_info)) {
1031                 rcu_read_unlock();
1032                 return;
1033         }
1034
1035 retry:
1036         list_for_each_entry_continue_rcu(tid_info, &ar->tx_ampdu_list, list) {
1037                 i++;
1038
1039                 if (tid_info->state < CARL9170_TID_STATE_PROGRESS)
1040                         continue;
1041
1042                 queue = TID_TO_WME_AC(tid_info->tid);
1043
1044                 spin_lock_bh(&tid_info->lock);
1045                 if (tid_info->state != CARL9170_TID_STATE_XMIT) {
1046                         first = skb_peek(&tid_info->queue);
1047                         if (first) {
1048                                 struct ieee80211_tx_info *txinfo;
1049                                 struct carl9170_tx_info *arinfo;
1050
1051                                 txinfo = IEEE80211_SKB_CB(first);
1052                                 arinfo = (void *) txinfo->rate_driver_data;
1053
1054                                 if (time_is_after_jiffies(arinfo->timeout +
1055                                     msecs_to_jiffies(CARL9170_QUEUE_TIMEOUT))
1056                                     == true)
1057                                         goto processed;
1058
1059                                 /*
1060                                  * We've been waiting for the frame which
1061                                  * matches "snx" (start sequence of the
1062                                  * next aggregate) for some time now.
1063                                  *
1064                                  * But it never arrived. Therefore
1065                                  * jump to the next available frame
1066                                  * and kick-start the transmission.
1067                                  *
1068                                  * Note: This might induce odd latency
1069                                  * spikes because the receiver will be
1070                                  * waiting for the lost frame too.
1071                                  */
1072                                 ar->tx_ampdu_timeout++;
1073
1074                                 tid_info->snx = carl9170_get_seq(first);
1075                                 tid_info->state = CARL9170_TID_STATE_XMIT;
1076                         } else {
1077                                 goto processed;
1078                         }
1079                 }
1080
1081                 tid_info->counter++;
1082                 first = skb_peek(&tid_info->queue);
1083                 tmpssn = carl9170_get_seq(first);
1084                 seq = tid_info->snx;
1085
1086                 if (unlikely(tmpssn != seq)) {
1087                         tid_info->state = CARL9170_TID_STATE_IDLE;
1088
1089                         goto processed;
1090                 }
1091
1092                 while ((skb = skb_peek(&tid_info->queue))) {
1093                         /* strict 0, 1, ..., n - 1, n frame sequence order */
1094                         if (unlikely(carl9170_get_seq(skb) != seq))
1095                                 break;
1096
1097                         /* don't upload more than AMPDU FACTOR allows. */
1098                         if (unlikely(SEQ_DIFF(tid_info->snx, tid_info->bsn) >=
1099                             (tid_info->max - 1)))
1100                                 break;
1101
1102                         if (!carl9170_tx_rate_check(ar, skb, first))
1103                                 break;
1104
1105                         atomic_inc(&ar->tx_ampdu_upload);
1106                         tid_info->snx = seq = SEQ_NEXT(seq);
1107                         __skb_unlink(skb, &tid_info->queue);
1108
1109                         __skb_queue_tail(&agg, skb);
1110
1111                         if (skb_queue_len(&agg) >= CARL9170_NUM_TX_AGG_MAX)
1112                                 break;
1113                 }
1114
1115                 if (skb_queue_empty(&tid_info->queue) ||
1116                     carl9170_get_seq(skb_peek(&tid_info->queue)) !=
1117                     tid_info->snx) {
1118                         /*
1119                          * stop TID, if A-MPDU frames are still missing,
1120                          * or whenever the queue is empty.
1121                          */
1122
1123                         tid_info->state = CARL9170_TID_STATE_IDLE;
1124                 }
1125                 done_ampdus++;
1126
1127 processed:
1128                 spin_unlock_bh(&tid_info->lock);
1129
1130                 if (skb_queue_empty(&agg))
1131                         continue;
1132
1133                 /* apply ampdu spacing & factor settings */
1134                 carl9170_set_ampdu_params(ar, skb_peek(&agg));
1135
1136                 /* set aggregation push bit */
1137                 carl9170_set_immba(ar, skb_peek_tail(&agg));
1138
1139                 spin_lock_bh(&ar->tx_pending[queue].lock);
1140                 skb_queue_splice_tail_init(&agg, &ar->tx_pending[queue]);
1141                 spin_unlock_bh(&ar->tx_pending[queue].lock);
1142                 ar->tx_schedule = true;
1143         }
1144         if ((done_ampdus++ == 0) && (i++ == 0))
1145                 goto retry;
1146
1147         rcu_assign_pointer(ar->tx_ampdu_iter, tid_info);
1148         rcu_read_unlock();
1149 }
1150
1151 static struct sk_buff *carl9170_tx_pick_skb(struct ar9170 *ar,
1152                                             struct sk_buff_head *queue)
1153 {
1154         struct sk_buff *skb;
1155         struct ieee80211_tx_info *info;
1156         struct carl9170_tx_info *arinfo;
1157
1158         BUILD_BUG_ON(sizeof(*arinfo) > sizeof(info->rate_driver_data));
1159
1160         spin_lock_bh(&queue->lock);
1161         skb = skb_peek(queue);
1162         if (unlikely(!skb))
1163                 goto err_unlock;
1164
1165         if (carl9170_alloc_dev_space(ar, skb))
1166                 goto err_unlock;
1167
1168         __skb_unlink(skb, queue);
1169         spin_unlock_bh(&queue->lock);
1170
1171         info = IEEE80211_SKB_CB(skb);
1172         arinfo = (void *) info->rate_driver_data;
1173
1174         arinfo->timeout = jiffies;
1175
1176         /*
1177          * increase ref count to "2".
1178          * Ref counting is the easiest way to solve the race between
1179          * the the urb's completion routine: carl9170_tx_callback and
1180          * wlan tx status functions: carl9170_tx_status/janitor.
1181          */
1182         carl9170_tx_get_skb(skb);
1183
1184         return skb;
1185
1186 err_unlock:
1187         spin_unlock_bh(&queue->lock);
1188         return NULL;
1189 }
1190
1191 void carl9170_tx_drop(struct ar9170 *ar, struct sk_buff *skb)
1192 {
1193         struct _carl9170_tx_superframe *super;
1194         uint8_t q = 0;
1195
1196         ar->tx_dropped++;
1197
1198         super = (void *)skb->data;
1199         SET_VAL(CARL9170_TX_SUPER_MISC_QUEUE, q,
1200                 ar9170_qmap[carl9170_get_queue(ar, skb)]);
1201         __carl9170_tx_process_status(ar, super->s.cookie, q);
1202 }
1203
1204 static void carl9170_tx(struct ar9170 *ar)
1205 {
1206         struct sk_buff *skb;
1207         unsigned int i, q;
1208         bool schedule_garbagecollector = false;
1209
1210         ar->tx_schedule = false;
1211
1212         if (unlikely(!IS_STARTED(ar)))
1213                 return;
1214
1215         carl9170_usb_handle_tx_err(ar);
1216
1217         for (i = 0; i < ar->hw->queues; i++) {
1218                 while (!skb_queue_empty(&ar->tx_pending[i])) {
1219                         skb = carl9170_tx_pick_skb(ar, &ar->tx_pending[i]);
1220                         if (unlikely(!skb))
1221                                 break;
1222
1223                         atomic_inc(&ar->tx_total_pending);
1224
1225                         q = __carl9170_get_queue(ar, i);
1226                         /*
1227                          * NB: tx_status[i] vs. tx_status[q],
1228                          * TODO: Move into pick_skb or alloc_dev_space.
1229                          */
1230                         skb_queue_tail(&ar->tx_status[q], skb);
1231
1232                         carl9170_usb_tx(ar, skb);
1233                         schedule_garbagecollector = true;
1234                 }
1235         }
1236
1237         if (!schedule_garbagecollector)
1238                 return;
1239
1240         ieee80211_queue_delayed_work(ar->hw, &ar->tx_janitor,
1241                 msecs_to_jiffies(CARL9170_TX_TIMEOUT));
1242 }
1243
1244 static bool carl9170_tx_ampdu_queue(struct ar9170 *ar,
1245         struct ieee80211_sta *sta, struct sk_buff *skb)
1246 {
1247         struct carl9170_sta_info *sta_info;
1248         struct carl9170_sta_tid *agg;
1249         struct sk_buff *iter;
1250         unsigned int max;
1251         u16 tid, seq, qseq, off;
1252         bool run = false;
1253
1254         tid = carl9170_get_tid(skb);
1255         seq = carl9170_get_seq(skb);
1256         sta_info = (void *) sta->drv_priv;
1257
1258         rcu_read_lock();
1259         agg = rcu_dereference(sta_info->agg[tid]);
1260         max = sta_info->ampdu_max_len;
1261
1262         if (!agg)
1263                 goto err_unlock_rcu;
1264
1265         spin_lock_bh(&agg->lock);
1266         if (unlikely(agg->state < CARL9170_TID_STATE_IDLE))
1267                 goto err_unlock;
1268
1269         /* check if sequence is within the BA window */
1270         if (unlikely(!BAW_WITHIN(agg->bsn, CARL9170_BAW_BITS, seq)))
1271                 goto err_unlock;
1272
1273         if (WARN_ON_ONCE(!BAW_WITHIN(agg->snx, CARL9170_BAW_BITS, seq)))
1274                 goto err_unlock;
1275
1276         off = SEQ_DIFF(seq, agg->bsn);
1277         if (WARN_ON_ONCE(test_and_set_bit(off, agg->bitmap)))
1278                 goto err_unlock;
1279
1280         if (likely(BAW_WITHIN(agg->hsn, CARL9170_BAW_BITS, seq))) {
1281                 __skb_queue_tail(&agg->queue, skb);
1282                 agg->hsn = seq;
1283                 goto queued;
1284         }
1285
1286         skb_queue_reverse_walk(&agg->queue, iter) {
1287                 qseq = carl9170_get_seq(iter);
1288
1289                 if (BAW_WITHIN(qseq, CARL9170_BAW_BITS, seq)) {
1290                         __skb_queue_after(&agg->queue, iter, skb);
1291                         goto queued;
1292                 }
1293         }
1294
1295         __skb_queue_head(&agg->queue, skb);
1296 queued:
1297
1298         if (unlikely(agg->state != CARL9170_TID_STATE_XMIT)) {
1299                 if (agg->snx == carl9170_get_seq(skb_peek(&agg->queue))) {
1300                         agg->state = CARL9170_TID_STATE_XMIT;
1301                         run = true;
1302                 }
1303         }
1304
1305         spin_unlock_bh(&agg->lock);
1306         rcu_read_unlock();
1307
1308         return run;
1309
1310 err_unlock:
1311         spin_unlock_bh(&agg->lock);
1312
1313 err_unlock_rcu:
1314         rcu_read_unlock();
1315         carl9170_tx_status(ar, skb, false);
1316         ar->tx_dropped++;
1317         return false;
1318 }
1319
1320 int carl9170_op_tx(struct ieee80211_hw *hw, struct sk_buff *skb)
1321 {
1322         struct ar9170 *ar = hw->priv;
1323         struct ieee80211_tx_info *info;
1324         struct ieee80211_sta *sta;
1325         bool run;
1326
1327         if (unlikely(!IS_STARTED(ar)))
1328                 goto err_free;
1329
1330         info = IEEE80211_SKB_CB(skb);
1331         sta = info->control.sta;
1332
1333         if (unlikely(carl9170_tx_prepare(ar, skb)))
1334                 goto err_free;
1335
1336         carl9170_tx_accounting(ar, skb);
1337         /*
1338          * from now on, one has to use carl9170_tx_status to free
1339          * all ressouces which are associated with the frame.
1340          */
1341
1342         if (info->flags & IEEE80211_TX_CTL_AMPDU) {
1343                 if (WARN_ON_ONCE(!sta))
1344                         goto err_free;
1345
1346                 run = carl9170_tx_ampdu_queue(ar, sta, skb);
1347                 if (run)
1348                         carl9170_tx_ampdu(ar);
1349
1350         } else {
1351                 unsigned int queue = skb_get_queue_mapping(skb);
1352
1353                 skb_queue_tail(&ar->tx_pending[queue], skb);
1354         }
1355
1356         carl9170_tx(ar);
1357         return NETDEV_TX_OK;
1358
1359 err_free:
1360         ar->tx_dropped++;
1361         dev_kfree_skb_any(skb);
1362         return NETDEV_TX_OK;
1363 }
1364
1365 void carl9170_tx_scheduler(struct ar9170 *ar)
1366 {
1367
1368         if (ar->tx_ampdu_schedule)
1369                 carl9170_tx_ampdu(ar);
1370
1371         if (ar->tx_schedule)
1372                 carl9170_tx(ar);
1373 }