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 /* needs rcu_read_lock */
 108 static struct ieee80211_sta *__carl9170_get_tx_sta(struct ar9170 *ar,
 109                                                    struct sk_buff *skb)
 110 {
 111         struct _carl9170_tx_superframe *super = (void *) skb->data;
 112         struct ieee80211_hdr *hdr = (void *) super->frame_data;
 113         struct ieee80211_vif *vif;
 114         unsigned int vif_id;
 115
 116         vif_id = (super->s.misc & CARL9170_TX_SUPER_MISC_VIF_ID) >>
 117                  CARL9170_TX_SUPER_MISC_VIF_ID_S;
 118
 119         if (WARN_ON_ONCE(vif_id >= AR9170_MAX_VIRTUAL_MAC))
 120                 return NULL;
 121
 122         vif = rcu_dereference(ar->vif_priv[vif_id].vif);
 123         if (unlikely(!vif))
 124                 return NULL;
 125
 126         /*
 127          * Normally we should use wrappers like ieee80211_get_DA to get
 128          * the correct peer ieee80211_sta.
 129          *
 130          * But there is a problem with indirect traffic (broadcasts, or
 131          * data which is designated for other stations) in station mode.
 132          * The frame will be directed to the AP for distribution and not
 133          * to the actual destination.
 134          */
 135
 136         return ieee80211_find_sta(vif, hdr->addr1);
 137 }
 138
 139 static void carl9170_tx_ps_unblock(struct ar9170 *ar, struct sk_buff *skb)
 140 {
 141         struct ieee80211_sta *sta;
 142         struct carl9170_sta_info *sta_info;
 143
 144         rcu_read_lock();
 145         sta = __carl9170_get_tx_sta(ar, skb);
 146         if (unlikely(!sta))
 147                 goto out_rcu;
 148
 149         sta_info = (struct carl9170_sta_info *) sta->drv_priv;
 150         if (atomic_dec_return(&sta_info->pending_frames) == 0)
 151                 ieee80211_sta_block_awake(ar->hw, sta, false);
 152
 153 out_rcu:
 154         rcu_read_unlock();
 155 }
 156
 157 static void carl9170_tx_accounting_free(struct ar9170 *ar, struct sk_buff *skb)
 158 {
 159         int queue;
 160
 161         queue = skb_get_queue_mapping(skb);
 162
 163         spin_lock_bh(&ar->tx_stats_lock);
 164
 165         ar->tx_stats[queue].len--;
 166
 167         if (!is_mem_full(ar)) {
 168                 unsigned int i;
 169                 for (i = 0; i < ar->hw->queues; i++) {
 170                         if (ar->tx_stats[i].len >= CARL9170_NUM_TX_LIMIT_SOFT)
 171                                 continue;
 172
 173                         if (ieee80211_queue_stopped(ar->hw, i)) {
 174                                 unsigned long tmp;
 175
 176                                 tmp = jiffies - ar->queue_stop_timeout[i];
 177                                 if (tmp > ar->max_queue_stop_timeout[i])
 178                                         ar->max_queue_stop_timeout[i] = tmp;
 179                         }
 180
 181                         ieee80211_wake_queue(ar->hw, i);
 182                 }
 183         }
 184
 185         spin_unlock_bh(&ar->tx_stats_lock);
 186
 187         if (atomic_dec_and_test(&ar->tx_total_queued))
 188                 complete(&ar->tx_flush);
 189 }
 190
 191 static int carl9170_alloc_dev_space(struct ar9170 *ar, struct sk_buff *skb)
 192 {
 193         struct _carl9170_tx_superframe *super = (void *) skb->data;
 194         unsigned int chunks;
 195         int cookie = -1;
 196
 197         atomic_inc(&ar->mem_allocs);
 198
 199         chunks = DIV_ROUND_UP(skb->len, ar->fw.mem_block_size);
 200         if (unlikely(atomic_sub_return(chunks, &ar->mem_free_blocks) < 0)) {
 201                 atomic_add(chunks, &ar->mem_free_blocks);
 202                 return -ENOSPC;
 203         }
 204
 205         spin_lock_bh(&ar->mem_lock);
 206         cookie = bitmap_find_free_region(ar->mem_bitmap, ar->fw.mem_blocks, 0);
 207         spin_unlock_bh(&ar->mem_lock);
 208
 209         if (unlikely(cookie < 0)) {
 210                 atomic_add(chunks, &ar->mem_free_blocks);
 211                 return -ENOSPC;
 212         }
 213
 214         super = (void *) skb->data;
 215
 216         /*
 217          * Cookie #0 serves two special purposes:
 218          *  1. The firmware might use it generate BlockACK frames
 219          *     in responds of an incoming BlockAckReqs.
 220          *
 221          *  2. Prevent double-free bugs.
 222          */
 223         super->s.cookie = (u8) cookie + 1;
 224         return 0;
 225 }
 226
 227 static void carl9170_release_dev_space(struct ar9170 *ar, struct sk_buff *skb)
 228 {
 229         struct _carl9170_tx_superframe *super = (void *) skb->data;
 230         int cookie;
 231
 232         /* make a local copy of the cookie */
 233         cookie = super->s.cookie;
 234         /* invalidate cookie */
 235         super->s.cookie = 0;
 236
 237         /*
 238          * Do a out-of-bounds check on the cookie:
 239          *
 240          *  * cookie "0" is reserved and won't be assigned to any
 241          *    out-going frame. Internally however, it is used to
 242          *    mark no longer/un-accounted frames and serves as a
 243          *    cheap way of preventing frames from being freed
 244          *    twice by _accident_. NB: There is a tiny race...
 245          *
 246          *  * obviously, cookie number is limited by the amount
 247          *    of available memory blocks, so the number can
 248          *    never execeed the mem_blocks count.
 249          */
 250         if (unlikely(WARN_ON_ONCE(cookie == 0) ||
 251             WARN_ON_ONCE(cookie > ar->fw.mem_blocks)))
 252                 return;
 253
 254         atomic_add(DIV_ROUND_UP(skb->len, ar->fw.mem_block_size),
 255                    &ar->mem_free_blocks);
 256
 257         spin_lock_bh(&ar->mem_lock);
 258         bitmap_release_region(ar->mem_bitmap, cookie - 1, 0);
 259         spin_unlock_bh(&ar->mem_lock);
 260 }
 261
 262 /* Called from any context */
 263 static void carl9170_tx_release(struct kref *ref)
 264 {
 265         struct ar9170 *ar;
 266         struct carl9170_tx_info *arinfo;
 267         struct ieee80211_tx_info *txinfo;
 268         struct sk_buff *skb;
 269
 270         arinfo = container_of(ref, struct carl9170_tx_info, ref);
 271         txinfo = container_of((void *) arinfo, struct ieee80211_tx_info,
 272                               rate_driver_data);
 273         skb = container_of((void *) txinfo, struct sk_buff, cb);
 274
 275         ar = arinfo->ar;
 276         if (WARN_ON_ONCE(!ar))
 277                 return;
 278
 279         BUILD_BUG_ON(
 280             offsetof(struct ieee80211_tx_info, status.ampdu_ack_len) != 23);
 281
 282         memset(&txinfo->status.ampdu_ack_len, 0,
 283                sizeof(struct ieee80211_tx_info) -
 284                offsetof(struct ieee80211_tx_info, status.ampdu_ack_len));
 285
 286         if (atomic_read(&ar->tx_total_queued))
 287                 ar->tx_schedule = true;
 288
 289         if (txinfo->flags & IEEE80211_TX_CTL_AMPDU) {
 290                 if (!atomic_read(&ar->tx_ampdu_upload))
 291                         ar->tx_ampdu_schedule = true;
 292
 293                 if (txinfo->flags & IEEE80211_TX_STAT_AMPDU) {
 294                         struct _carl9170_tx_superframe *super;
 295
 296                         super = (void *)skb->data;
 297                         txinfo->status.ampdu_len = super->s.rix;
 298                         txinfo->status.ampdu_ack_len = super->s.cnt;
 299                 } else if ((txinfo->flags & IEEE80211_TX_STAT_ACK) &&
 300                            !(txinfo->flags & IEEE80211_TX_CTL_REQ_TX_STATUS)) {
 301                         /*
 302                          * drop redundant tx_status reports:
 303                          *
 304                          * 1. ampdu_ack_len of the final tx_status does
 305                          *    include the feedback of this particular frame.
 306                          *
 307                          * 2. tx_status_irqsafe only queues up to 128
 308                          *    tx feedback reports and discards the rest.
 309                          *
 310                          * 3. minstrel_ht is picky, it only accepts
 311                          *    reports of frames with the TX_STATUS_AMPDU flag.
 312                          *
 313                          * 4. mac80211 is not particularly interested in
 314                          *    feedback either [CTL_REQ_TX_STATUS not set]
 315                          */
 316
 317                         ieee80211_free_txskb(ar->hw, skb);
 318                         return;
 319                 } else {
 320                         /*
 321                          * Either the frame transmission has failed or
 322                          * mac80211 requested tx status.
 323                          */
 324                 }
 325         }
 326
 327         skb_pull(skb, sizeof(struct _carl9170_tx_superframe));
 328         ieee80211_tx_status_irqsafe(ar->hw, skb);
 329 }
 330
 331 void carl9170_tx_get_skb(struct sk_buff *skb)
 332 {
 333         struct carl9170_tx_info *arinfo = (void *)
 334                 (IEEE80211_SKB_CB(skb))->rate_driver_data;
 335         kref_get(&arinfo->ref);
 336 }
 337
 338 int carl9170_tx_put_skb(struct sk_buff *skb)
 339 {
 340         struct carl9170_tx_info *arinfo = (void *)
 341                 (IEEE80211_SKB_CB(skb))->rate_driver_data;
 342
 343         return kref_put(&arinfo->ref, carl9170_tx_release);
 344 }
 345
 346 /* Caller must hold the tid_info->lock & rcu_read_lock */
 347 static void carl9170_tx_shift_bm(struct ar9170 *ar,
 348         struct carl9170_sta_tid *tid_info, u16 seq)
 349 {
 350         u16 off;
 351
 352         off = SEQ_DIFF(seq, tid_info->bsn);
 353
 354         if (WARN_ON_ONCE(off >= CARL9170_BAW_BITS))
 355                 return;
 356
 357         /*
 358          * Sanity check. For each MPDU we set the bit in bitmap and
 359          * clear it once we received the tx_status.
 360          * But if the bit is already cleared then we've been bitten
 361          * by a bug.
 362          */
 363         WARN_ON_ONCE(!test_and_clear_bit(off, tid_info->bitmap));
 364
 365         off = SEQ_DIFF(tid_info->snx, tid_info->bsn);
 366         if (WARN_ON_ONCE(off >= CARL9170_BAW_BITS))
 367                 return;
 368
 369         if (!bitmap_empty(tid_info->bitmap, off))
 370                 off = find_first_bit(tid_info->bitmap, off);
 371
 372         tid_info->bsn += off;
 373         tid_info->bsn &= 0x0fff;
 374
 375         bitmap_shift_right(tid_info->bitmap, tid_info->bitmap,
 376                            off, CARL9170_BAW_BITS);
 377 }
 378
 379 static void carl9170_tx_status_process_ampdu(struct ar9170 *ar,
 380         struct sk_buff *skb, struct ieee80211_tx_info *txinfo)
 381 {
 382         struct _carl9170_tx_superframe *super = (void *) skb->data;
 383         struct ieee80211_hdr *hdr = (void *) super->frame_data;
 384         struct ieee80211_sta *sta;
 385         struct carl9170_sta_info *sta_info;
 386         struct carl9170_sta_tid *tid_info;
 387         u8 tid;
 388
 389         if (!(txinfo->flags & IEEE80211_TX_CTL_AMPDU) ||
 390             txinfo->flags & IEEE80211_TX_CTL_INJECTED ||
 391            (!(super->f.mac_control & cpu_to_le16(AR9170_TX_MAC_AGGR))))
 392                 return;
 393
 394         rcu_read_lock();
 395         sta = __carl9170_get_tx_sta(ar, skb);
 396         if (unlikely(!sta))
 397                 goto out_rcu;
 398
 399         tid = get_tid_h(hdr);
 400
 401         sta_info = (void *) sta->drv_priv;
 402         tid_info = rcu_dereference(sta_info->agg[tid]);
 403         if (!tid_info)
 404                 goto out_rcu;
 405
 406         spin_lock_bh(&tid_info->lock);
 407         if (likely(tid_info->state >= CARL9170_TID_STATE_IDLE))
 408                 carl9170_tx_shift_bm(ar, tid_info, get_seq_h(hdr));
 409
 410         if (sta_info->stats[tid].clear) {
 411                 sta_info->stats[tid].clear = false;
 412                 sta_info->stats[tid].req = false;
 413                 sta_info->stats[tid].ampdu_len = 0;
 414                 sta_info->stats[tid].ampdu_ack_len = 0;
 415         }
 416
 417         sta_info->stats[tid].ampdu_len++;
 418         if (txinfo->status.rates[0].count == 1)
 419                 sta_info->stats[tid].ampdu_ack_len++;
 420
 421         if (!(txinfo->flags & IEEE80211_TX_STAT_ACK))
 422                 sta_info->stats[tid].req = true;
 423
 424         if (super->f.mac_control & cpu_to_le16(AR9170_TX_MAC_IMM_BA)) {
 425                 super->s.rix = sta_info->stats[tid].ampdu_len;
 426                 super->s.cnt = sta_info->stats[tid].ampdu_ack_len;
 427                 txinfo->flags |= IEEE80211_TX_STAT_AMPDU;
 428                 if (sta_info->stats[tid].req)
 429                         txinfo->flags |= IEEE80211_TX_STAT_AMPDU_NO_BACK;
 430
 431                 sta_info->stats[tid].clear = true;
 432         }
 433         spin_unlock_bh(&tid_info->lock);
 434
 435 out_rcu:
 436         rcu_read_unlock();
 437 }
 438
 439 void carl9170_tx_status(struct ar9170 *ar, struct sk_buff *skb,
 440                         const bool success)
 441 {
 442         struct ieee80211_tx_info *txinfo;
 443
 444         carl9170_tx_accounting_free(ar, skb);
 445
 446         txinfo = IEEE80211_SKB_CB(skb);
 447
 448         if (success)
 449                 txinfo->flags |= IEEE80211_TX_STAT_ACK;
 450         else
 451                 ar->tx_ack_failures++;
 452
 453         if (txinfo->flags & IEEE80211_TX_CTL_AMPDU)
 454                 carl9170_tx_status_process_ampdu(ar, skb, txinfo);
 455
 456         carl9170_tx_ps_unblock(ar, skb);
 457         carl9170_tx_put_skb(skb);
 458 }
 459
 460 /* This function may be called form any context */
 461 void carl9170_tx_callback(struct ar9170 *ar, struct sk_buff *skb)
 462 {
 463         struct ieee80211_tx_info *txinfo = IEEE80211_SKB_CB(skb);
 464
 465         atomic_dec(&ar->tx_total_pending);
 466
 467         if (txinfo->flags & IEEE80211_TX_CTL_AMPDU)
 468                 atomic_dec(&ar->tx_ampdu_upload);
 469
 470         if (carl9170_tx_put_skb(skb))
 471                 tasklet_hi_schedule(&ar->usb_tasklet);
 472 }
 473
 474 static struct sk_buff *carl9170_get_queued_skb(struct ar9170 *ar, u8 cookie,
 475                                                struct sk_buff_head *queue)
 476 {
 477         struct sk_buff *skb;
 478
 479         spin_lock_bh(&queue->lock);
 480         skb_queue_walk(queue, skb) {
 481                 struct _carl9170_tx_superframe *txc = (void *) skb->data;
 482
 483                 if (txc->s.cookie != cookie)
 484                         continue;
 485
 486                 __skb_unlink(skb, queue);
 487                 spin_unlock_bh(&queue->lock);
 488
 489                 carl9170_release_dev_space(ar, skb);
 490                 return skb;
 491         }
 492         spin_unlock_bh(&queue->lock);
 493
 494         return NULL;
 495 }
 496
 497 static void carl9170_tx_fill_rateinfo(struct ar9170 *ar, unsigned int rix,
 498         unsigned int tries, struct ieee80211_tx_info *txinfo)
 499 {
 500         unsigned int i;
 501
 502         for (i = 0; i < IEEE80211_TX_MAX_RATES; i++) {
 503                 if (txinfo->status.rates[i].idx < 0)
 504                         break;
 505
 506                 if (i == rix) {
 507                         txinfo->status.rates[i].count = tries;
 508                         i++;
 509                         break;
 510                 }
 511         }
 512
 513         for (; i < IEEE80211_TX_MAX_RATES; i++) {
 514                 txinfo->status.rates[i].idx = -1;
 515                 txinfo->status.rates[i].count = 0;
 516         }
 517 }
 518
 519 static void carl9170_check_queue_stop_timeout(struct ar9170 *ar)
 520 {
 521         int i;
 522         struct sk_buff *skb;
 523         struct ieee80211_tx_info *txinfo;
 524         struct carl9170_tx_info *arinfo;
 525         bool restart = false;
 526
 527         for (i = 0; i < ar->hw->queues; i++) {
 528                 spin_lock_bh(&ar->tx_status[i].lock);
 529
 530                 skb = skb_peek(&ar->tx_status[i]);
 531
 532                 if (!skb)
 533                         goto next;
 534
 535                 txinfo = IEEE80211_SKB_CB(skb);
 536                 arinfo = (void *) txinfo->rate_driver_data;
 537
 538                 if (time_is_before_jiffies(arinfo->timeout +
 539                     msecs_to_jiffies(CARL9170_QUEUE_STUCK_TIMEOUT)) == true)
 540                         restart = true;
 541
 542 next:
 543                 spin_unlock_bh(&ar->tx_status[i].lock);
 544         }
 545
 546         if (restart) {
 547                 /*
 548                  * At least one queue has been stuck for long enough.
 549                  * Give the device a kick and hope it gets back to
 550                  * work.
 551                  *
 552                  * possible reasons may include:
 553                  *  - frames got lost/corrupted (bad connection to the device)
 554                  *  - stalled rx processing/usb controller hiccups
 555                  *  - firmware errors/bugs
 556                  *  - every bug you can think of.
 557                  *  - all bugs you can't...
 558                  *  - ...
 559                  */
 560                 carl9170_restart(ar, CARL9170_RR_STUCK_TX);
 561         }
 562 }
 563
 564 static void carl9170_tx_ampdu_timeout(struct ar9170 *ar)
 565 {
 566         struct carl9170_sta_tid *iter;
 567         struct sk_buff *skb;
 568         struct ieee80211_tx_info *txinfo;
 569         struct carl9170_tx_info *arinfo;
 570         struct ieee80211_sta *sta;
 571
 572         rcu_read_lock();
 573         list_for_each_entry_rcu(iter, &ar->tx_ampdu_list, list) {
 574                 if (iter->state < CARL9170_TID_STATE_IDLE)
 575                         continue;
 576
 577                 spin_lock_bh(&iter->lock);
 578                 skb = skb_peek(&iter->queue);
 579                 if (!skb)
 580                         goto unlock;
 581
 582                 txinfo = IEEE80211_SKB_CB(skb);
 583                 arinfo = (void *)txinfo->rate_driver_data;
 584                 if (time_is_after_jiffies(arinfo->timeout +
 585                     msecs_to_jiffies(CARL9170_QUEUE_TIMEOUT)))
 586                         goto unlock;
 587
 588                 sta = __carl9170_get_tx_sta(ar, skb);
 589                 if (WARN_ON(!sta))
 590                         goto unlock;
 591
 592                 ieee80211_stop_tx_ba_session(sta, iter->tid);
 593 unlock:
 594                 spin_unlock_bh(&iter->lock);
 595
 596         }
 597         rcu_read_unlock();
 598 }
 599
 600 void carl9170_tx_janitor(struct work_struct *work)
 601 {
 602         struct ar9170 *ar = container_of(work, struct ar9170,
 603                                          tx_janitor.work);
 604         if (!IS_STARTED(ar))
 605                 return;
 606
 607         ar->tx_janitor_last_run = jiffies;
 608
 609         carl9170_check_queue_stop_timeout(ar);
 610         carl9170_tx_ampdu_timeout(ar);
 611
 612         if (!atomic_read(&ar->tx_total_queued))
 613                 return;
 614
 615         ieee80211_queue_delayed_work(ar->hw, &ar->tx_janitor,
 616                 msecs_to_jiffies(CARL9170_TX_TIMEOUT));
 617 }
 618
 619 static void __carl9170_tx_process_status(struct ar9170 *ar,
 620         const uint8_t cookie, const uint8_t info)
 621 {
 622         struct sk_buff *skb;
 623         struct ieee80211_tx_info *txinfo;
 624         unsigned int r, t, q;
 625         bool success = true;
 626
 627         q = ar9170_qmap[info & CARL9170_TX_STATUS_QUEUE];
 628
 629         skb = carl9170_get_queued_skb(ar, cookie, &ar->tx_status[q]);
 630         if (!skb) {
 631                 /*
 632                  * We have lost the race to another thread.
 633                  */
 634
 635                 return ;
 636         }
 637
 638         txinfo = IEEE80211_SKB_CB(skb);
 639
 640         if (!(info & CARL9170_TX_STATUS_SUCCESS))
 641                 success = false;
 642
 643         r = (info & CARL9170_TX_STATUS_RIX) >> CARL9170_TX_STATUS_RIX_S;
 644         t = (info & CARL9170_TX_STATUS_TRIES) >> CARL9170_TX_STATUS_TRIES_S;
 645
 646         carl9170_tx_fill_rateinfo(ar, r, t, txinfo);
 647         carl9170_tx_status(ar, skb, success);
 648 }
 649
 650 void carl9170_tx_process_status(struct ar9170 *ar,
 651                                 const struct carl9170_rsp *cmd)
 652 {
 653         unsigned int i;
 654
 655         for (i = 0;  i < cmd->hdr.ext; i++) {
 656                 if (WARN_ON(i > ((cmd->hdr.len / 2) + 1))) {
 657                         print_hex_dump_bytes("UU:", DUMP_PREFIX_NONE,
 658                                              (void *) cmd, cmd->hdr.len + 4);
 659                         break;
 660                 }
 661
 662                 __carl9170_tx_process_status(ar, cmd->_tx_status[i].cookie,
 663                                              cmd->_tx_status[i].info);
 664         }
 665 }
 666
 667 static void carl9170_tx_rate_tpc_chains(struct ar9170 *ar,
 668         struct ieee80211_tx_info *info, struct ieee80211_tx_rate *txrate,
 669         unsigned int *phyrate, unsigned int *tpc, unsigned int *chains)
 670 {
 671         struct ieee80211_rate *rate = NULL;
 672         u8 *txpower;
 673         unsigned int idx;
 674
 675         idx = txrate->idx;
 676         *tpc = 0;
 677         *phyrate = 0;
 678
 679         if (txrate->flags & IEEE80211_TX_RC_MCS) {
 680                 if (txrate->flags & IEEE80211_TX_RC_40_MHZ_WIDTH) {
 681                         /* +1 dBm for HT40 */
 682                         *tpc += 2;
 683
 684                         if (info->band == IEEE80211_BAND_2GHZ)
 685                                 txpower = ar->power_2G_ht40;
 686                         else
 687                                 txpower = ar->power_5G_ht40;
 688                 } else {
 689                         if (info->band == IEEE80211_BAND_2GHZ)
 690                                 txpower = ar->power_2G_ht20;
 691                         else
 692                                 txpower = ar->power_5G_ht20;
 693                 }
 694
 695                 *phyrate = txrate->idx;
 696                 *tpc += txpower[idx & 7];
 697         } else {
 698                 if (info->band == IEEE80211_BAND_2GHZ) {
 699                         if (idx < 4)
 700                                 txpower = ar->power_2G_cck;
 701                         else
 702                                 txpower = ar->power_2G_ofdm;
 703                 } else {
 704                         txpower = ar->power_5G_leg;
 705                         idx += 4;
 706                 }
 707
 708                 rate = &__carl9170_ratetable[idx];
 709                 *tpc += txpower[(rate->hw_value & 0x30) >> 4];
 710                 *phyrate = rate->hw_value & 0xf;
 711         }
 712
 713         if (ar->eeprom.tx_mask == 1) {
 714                 *chains = AR9170_TX_PHY_TXCHAIN_1;
 715         } else {
 716                 if (!(txrate->flags & IEEE80211_TX_RC_MCS) &&
 717                     rate && rate->bitrate >= 360)
 718                         *chains = AR9170_TX_PHY_TXCHAIN_1;
 719                 else
 720                         *chains = AR9170_TX_PHY_TXCHAIN_2;
 721         }
 722
 723         *tpc = min_t(unsigned int, *tpc, ar->hw->conf.power_level * 2);
 724 }
 725
 726 static __le32 carl9170_tx_physet(struct ar9170 *ar,
 727         struct ieee80211_tx_info *info, struct ieee80211_tx_rate *txrate)
 728 {
 729         unsigned int power = 0, chains = 0, phyrate = 0;
 730         __le32 tmp;
 731
 732         tmp = cpu_to_le32(0);
 733
 734         if (txrate->flags & IEEE80211_TX_RC_40_MHZ_WIDTH)
 735                 tmp |= cpu_to_le32(AR9170_TX_PHY_BW_40MHZ <<
 736                         AR9170_TX_PHY_BW_S);
 737         /* this works because 40 MHz is 2 and dup is 3 */
 738         if (txrate->flags & IEEE80211_TX_RC_DUP_DATA)
 739                 tmp |= cpu_to_le32(AR9170_TX_PHY_BW_40MHZ_DUP <<
 740                         AR9170_TX_PHY_BW_S);
 741
 742         if (txrate->flags & IEEE80211_TX_RC_SHORT_GI)
 743                 tmp |= cpu_to_le32(AR9170_TX_PHY_SHORT_GI);
 744
 745         if (txrate->flags & IEEE80211_TX_RC_MCS) {
 746                 SET_VAL(AR9170_TX_PHY_MCS, phyrate, txrate->idx);
 747
 748                 /* heavy clip control */
 749                 tmp |= cpu_to_le32((txrate->idx & 0x7) <<
 750                         AR9170_TX_PHY_TX_HEAVY_CLIP_S);
 751
 752                 tmp |= cpu_to_le32(AR9170_TX_PHY_MOD_HT);
 753
 754                 /*
 755                  * green field preamble does not work.
 756                  *
 757                  * if (txrate->flags & IEEE80211_TX_RC_GREEN_FIELD)
 758                  * tmp |= cpu_to_le32(AR9170_TX_PHY_GREENFIELD);
 759                  */
 760         } else {
 761                 if (info->band == IEEE80211_BAND_2GHZ) {
 762                         if (txrate->idx <= AR9170_TX_PHY_RATE_CCK_11M)
 763                                 tmp |= cpu_to_le32(AR9170_TX_PHY_MOD_CCK);
 764                         else
 765                                 tmp |= cpu_to_le32(AR9170_TX_PHY_MOD_OFDM);
 766                 } else {
 767                         tmp |= cpu_to_le32(AR9170_TX_PHY_MOD_OFDM);
 768                 }
 769
 770                 /*
 771                  * short preamble seems to be broken too.
 772                  *
 773                  * if (txrate->flags & IEEE80211_TX_RC_USE_SHORT_PREAMBLE)
 774                  *      tmp |= cpu_to_le32(AR9170_TX_PHY_SHORT_PREAMBLE);
 775                  */
 776         }
 777         carl9170_tx_rate_tpc_chains(ar, info, txrate,
 778                                     &phyrate, &power, &chains);
 779
 780         tmp |= cpu_to_le32(SET_CONSTVAL(AR9170_TX_PHY_MCS, phyrate));
 781         tmp |= cpu_to_le32(SET_CONSTVAL(AR9170_TX_PHY_TX_PWR, power));
 782         tmp |= cpu_to_le32(SET_CONSTVAL(AR9170_TX_PHY_TXCHAIN, chains));
 783         return tmp;
 784 }
 785
 786 static bool carl9170_tx_rts_check(struct ar9170 *ar,
 787                                   struct ieee80211_tx_rate *rate,
 788                                   bool ampdu, bool multi)
 789 {
 790         switch (ar->erp_mode) {
 791         case CARL9170_ERP_AUTO:
 792                 if (ampdu)
 793                         break;
 794
 795         case CARL9170_ERP_MAC80211:
 796                 if (!(rate->flags & IEEE80211_TX_RC_USE_RTS_CTS))
 797                         break;
 798
 799         case CARL9170_ERP_RTS:
 800                 if (likely(!multi))
 801                         return true;
 802
 803         default:
 804                 break;
 805         }
 806
 807         return false;
 808 }
 809
 810 static bool carl9170_tx_cts_check(struct ar9170 *ar,
 811                                   struct ieee80211_tx_rate *rate)
 812 {
 813         switch (ar->erp_mode) {
 814         case CARL9170_ERP_AUTO:
 815         case CARL9170_ERP_MAC80211:
 816                 if (!(rate->flags & IEEE80211_TX_RC_USE_CTS_PROTECT))
 817                         break;
 818
 819         case CARL9170_ERP_CTS:
 820                 return true;
 821
 822         default:
 823                 break;
 824         }
 825
 826         return false;
 827 }
 828
 829 static int carl9170_tx_prepare(struct ar9170 *ar, struct sk_buff *skb)
 830 {
 831         struct ieee80211_hdr *hdr;
 832         struct _carl9170_tx_superframe *txc;
 833         struct carl9170_vif_info *cvif;
 834         struct ieee80211_tx_info *info;
 835         struct ieee80211_tx_rate *txrate;
 836         struct ieee80211_sta *sta;
 837         struct carl9170_tx_info *arinfo;
 838         unsigned int hw_queue;
 839         int i;
 840         __le16 mac_tmp;
 841         u16 len;
 842         bool ampdu, no_ack;
 843
 844         BUILD_BUG_ON(sizeof(*arinfo) > sizeof(info->rate_driver_data));
 845         BUILD_BUG_ON(sizeof(struct _carl9170_tx_superdesc) !=
 846                      CARL9170_TX_SUPERDESC_LEN);
 847
 848         BUILD_BUG_ON(sizeof(struct _ar9170_tx_hwdesc) !=
 849                      AR9170_TX_HWDESC_LEN);
 850
 851         BUILD_BUG_ON(IEEE80211_TX_MAX_RATES < CARL9170_TX_MAX_RATES);
 852
 853         BUILD_BUG_ON(AR9170_MAX_VIRTUAL_MAC >
 854                 ((CARL9170_TX_SUPER_MISC_VIF_ID >>
 855                  CARL9170_TX_SUPER_MISC_VIF_ID_S) + 1));
 856
 857         hw_queue = ar9170_qmap[carl9170_get_queue(ar, skb)];
 858
 859         hdr = (void *)skb->data;
 860         info = IEEE80211_SKB_CB(skb);
 861         len = skb->len;
 862
 863         /*
 864          * Note: If the frame was sent through a monitor interface,
 865          * the ieee80211_vif pointer can be NULL.
 866          */
 867         if (likely(info->control.vif))
 868                 cvif = (void *) info->control.vif->drv_priv;
 869         else
 870                 cvif = NULL;
 871
 872         sta = info->control.sta;
 873
 874         txc = (void *)skb_push(skb, sizeof(*txc));
 875         memset(txc, 0, sizeof(*txc));
 876
 877         SET_VAL(CARL9170_TX_SUPER_MISC_QUEUE, txc->s.misc, hw_queue);
 878
 879         if (likely(cvif))
 880                 SET_VAL(CARL9170_TX_SUPER_MISC_VIF_ID, txc->s.misc, cvif->id);
 881
 882         if (unlikely(info->flags & IEEE80211_TX_CTL_SEND_AFTER_DTIM))
 883                 txc->s.misc |= CARL9170_TX_SUPER_MISC_CAB;
 884
 885         if (unlikely(info->flags & IEEE80211_TX_CTL_ASSIGN_SEQ))
 886                 txc->s.misc |= CARL9170_TX_SUPER_MISC_ASSIGN_SEQ;
 887
 888         if (unlikely(ieee80211_is_probe_resp(hdr->frame_control)))
 889                 txc->s.misc |= CARL9170_TX_SUPER_MISC_FILL_IN_TSF;
 890
 891         mac_tmp = cpu_to_le16(AR9170_TX_MAC_HW_DURATION |
 892                               AR9170_TX_MAC_BACKOFF);
 893         mac_tmp |= cpu_to_le16((hw_queue << AR9170_TX_MAC_QOS_S) &
 894                                AR9170_TX_MAC_QOS);
 895
 896         no_ack = !!(info->flags & IEEE80211_TX_CTL_NO_ACK);
 897         if (unlikely(no_ack))
 898                 mac_tmp |= cpu_to_le16(AR9170_TX_MAC_NO_ACK);
 899
 900         if (info->control.hw_key) {
 901                 len += info->control.hw_key->icv_len;
 902
 903                 switch (info->control.hw_key->cipher) {
 904                 case WLAN_CIPHER_SUITE_WEP40:
 905                 case WLAN_CIPHER_SUITE_WEP104:
 906                 case WLAN_CIPHER_SUITE_TKIP:
 907                         mac_tmp |= cpu_to_le16(AR9170_TX_MAC_ENCR_RC4);
 908                         break;
 909                 case WLAN_CIPHER_SUITE_CCMP:
 910                         mac_tmp |= cpu_to_le16(AR9170_TX_MAC_ENCR_AES);
 911                         break;
 912                 default:
 913                         WARN_ON(1);
 914                         goto err_out;
 915                 }
 916         }
 917
 918         ampdu = !!(info->flags & IEEE80211_TX_CTL_AMPDU);
 919         if (ampdu) {
 920                 unsigned int density, factor;
 921
 922                 if (unlikely(!sta || !cvif))
 923                         goto err_out;
 924
 925                 factor = min_t(unsigned int, 1u, sta->ht_cap.ampdu_factor);
 926                 density = sta->ht_cap.ampdu_density;
 927
 928                 if (density) {
 929                         /*
 930                          * Watch out!
 931                          *
 932                          * Otus uses slightly different density values than
 933                          * those from the 802.11n spec.
 934                          */
 935
 936                         density = max_t(unsigned int, density + 1, 7u);
 937                 }
 938
 939                 SET_VAL(CARL9170_TX_SUPER_AMPDU_DENSITY,
 940                         txc->s.ampdu_settings, density);
 941
 942                 SET_VAL(CARL9170_TX_SUPER_AMPDU_FACTOR,
 943                         txc->s.ampdu_settings, factor);
 944
 945                 for (i = 0; i < CARL9170_TX_MAX_RATES; i++) {
 946                         txrate = &info->control.rates[i];
 947                         if (txrate->idx >= 0) {
 948                                 txc->s.ri[i] =
 949                                         CARL9170_TX_SUPER_RI_AMPDU;
 950
 951                                 if (WARN_ON(!(txrate->flags &
 952                                               IEEE80211_TX_RC_MCS))) {
 953                                         /*
 954                                          * Not sure if it's even possible
 955                                          * to aggregate non-ht rates with
 956                                          * this HW.
 957                                          */
 958                                         goto err_out;
 959                                 }
 960                                 continue;
 961                         }
 962
 963                         txrate->idx = 0;
 964                         txrate->count = ar->hw->max_rate_tries;
 965                 }
 966
 967                 mac_tmp |= cpu_to_le16(AR9170_TX_MAC_AGGR);
 968         }
 969
 970         /*
 971          * NOTE: For the first rate, the ERP & AMPDU flags are directly
 972          * taken from mac_control. For all fallback rate, the firmware
 973          * updates the mac_control flags from the rate info field.
 974          */
 975         for (i = 1; i < CARL9170_TX_MAX_RATES; i++) {
 976                 txrate = &info->control.rates[i];
 977                 if (txrate->idx < 0)
 978                         break;
 979
 980                 SET_VAL(CARL9170_TX_SUPER_RI_TRIES, txc->s.ri[i],
 981                         txrate->count);
 982
 983                 if (carl9170_tx_rts_check(ar, txrate, ampdu, no_ack))
 984                         txc->s.ri[i] |= (AR9170_TX_MAC_PROT_RTS <<
 985                                 CARL9170_TX_SUPER_RI_ERP_PROT_S);
 986                 else if (carl9170_tx_cts_check(ar, txrate))
 987                         txc->s.ri[i] |= (AR9170_TX_MAC_PROT_CTS <<
 988                                 CARL9170_TX_SUPER_RI_ERP_PROT_S);
 989
 990                 txc->s.rr[i - 1] = carl9170_tx_physet(ar, info, txrate);
 991         }
 992
 993         txrate = &info->control.rates[0];
 994         SET_VAL(CARL9170_TX_SUPER_RI_TRIES, txc->s.ri[0], txrate->count);
 995
 996         if (carl9170_tx_rts_check(ar, txrate, ampdu, no_ack))
 997                 mac_tmp |= cpu_to_le16(AR9170_TX_MAC_PROT_RTS);
 998         else if (carl9170_tx_cts_check(ar, txrate))
 999                 mac_tmp |= cpu_to_le16(AR9170_TX_MAC_PROT_CTS);
1000
1001         txc->s.len = cpu_to_le16(skb->len);
1002         txc->f.length = cpu_to_le16(len + FCS_LEN);
1003         txc->f.mac_control = mac_tmp;
1004         txc->f.phy_control = carl9170_tx_physet(ar, info, txrate);
1005
1006         arinfo = (void *)info->rate_driver_data;
1007         arinfo->timeout = jiffies;
1008         arinfo->ar = ar;
1009         kref_init(&arinfo->ref);
1010         return 0;
1011
1012 err_out:
1013         skb_pull(skb, sizeof(*txc));
1014         return -EINVAL;
1015 }
1016
1017 static void carl9170_set_immba(struct ar9170 *ar, struct sk_buff *skb)
1018 {
1019         struct _carl9170_tx_superframe *super;
1020
1021         super = (void *) skb->data;
1022         super->f.mac_control |= cpu_to_le16(AR9170_TX_MAC_IMM_BA);
1023 }
1024
1025 static void carl9170_set_ampdu_params(struct ar9170 *ar, struct sk_buff *skb)
1026 {
1027         struct _carl9170_tx_superframe *super;
1028         int tmp;
1029
1030         super = (void *) skb->data;
1031
1032         tmp = (super->s.ampdu_settings & CARL9170_TX_SUPER_AMPDU_DENSITY) <<
1033                 CARL9170_TX_SUPER_AMPDU_DENSITY_S;
1034
1035         /*
1036          * If you haven't noticed carl9170_tx_prepare has already filled
1037          * in all ampdu spacing & factor parameters.
1038          * Now it's the time to check whenever the settings have to be
1039          * updated by the firmware, or if everything is still the same.
1040          *
1041          * There's no sane way to handle different density values with
1042          * this hardware, so we may as well just do the compare in the
1043          * driver.
1044          */
1045
1046         if (tmp != ar->current_density) {
1047                 ar->current_density = tmp;
1048                 super->s.ampdu_settings |=
1049                         CARL9170_TX_SUPER_AMPDU_COMMIT_DENSITY;
1050         }
1051
1052         tmp = (super->s.ampdu_settings & CARL9170_TX_SUPER_AMPDU_FACTOR) <<
1053                 CARL9170_TX_SUPER_AMPDU_FACTOR_S;
1054
1055         if (tmp != ar->current_factor) {
1056                 ar->current_factor = tmp;
1057                 super->s.ampdu_settings |=
1058                         CARL9170_TX_SUPER_AMPDU_COMMIT_FACTOR;
1059         }
1060 }
1061
1062 static bool carl9170_tx_rate_check(struct ar9170 *ar, struct sk_buff *_dest,
1063                                    struct sk_buff *_src)
1064 {
1065         struct _carl9170_tx_superframe *dest, *src;
1066
1067         dest = (void *) _dest->data;
1068         src = (void *) _src->data;
1069
1070         /*
1071          * The mac80211 rate control algorithm expects that all MPDUs in
1072          * an AMPDU share the same tx vectors.
1073          * This is not really obvious right now, because the hardware
1074          * does the AMPDU setup according to its own rulebook.
1075          * Our nicely assembled, strictly monotonic increasing mpdu
1076          * chains will be broken up, mashed back together...
1077          */
1078
1079         return (dest->f.phy_control == src->f.phy_control);
1080 }
1081
1082 static void carl9170_tx_ampdu(struct ar9170 *ar)
1083 {
1084         struct sk_buff_head agg;
1085         struct carl9170_sta_tid *tid_info;
1086         struct sk_buff *skb, *first;
1087         unsigned int i = 0, done_ampdus = 0;
1088         u16 seq, queue, tmpssn;
1089
1090         atomic_inc(&ar->tx_ampdu_scheduler);
1091         ar->tx_ampdu_schedule = false;
1092
1093         if (atomic_read(&ar->tx_ampdu_upload))
1094                 return;
1095
1096         if (!ar->tx_ampdu_list_len)
1097                 return;
1098
1099         __skb_queue_head_init(&agg);
1100
1101         rcu_read_lock();
1102         tid_info = rcu_dereference(ar->tx_ampdu_iter);
1103         if (WARN_ON_ONCE(!tid_info)) {
1104                 rcu_read_unlock();
1105                 return;
1106         }
1107
1108 retry:
1109         list_for_each_entry_continue_rcu(tid_info, &ar->tx_ampdu_list, list) {
1110                 i++;
1111
1112                 if (tid_info->state < CARL9170_TID_STATE_PROGRESS)
1113                         continue;
1114
1115                 queue = TID_TO_WME_AC(tid_info->tid);
1116
1117                 spin_lock_bh(&tid_info->lock);
1118                 if (tid_info->state != CARL9170_TID_STATE_XMIT)
1119                         goto processed;
1120
1121                 tid_info->counter++;
1122                 first = skb_peek(&tid_info->queue);
1123                 tmpssn = carl9170_get_seq(first);
1124                 seq = tid_info->snx;
1125
1126                 if (unlikely(tmpssn != seq)) {
1127                         tid_info->state = CARL9170_TID_STATE_IDLE;
1128
1129                         goto processed;
1130                 }
1131
1132                 while ((skb = skb_peek(&tid_info->queue))) {
1133                         /* strict 0, 1, ..., n - 1, n frame sequence order */
1134                         if (unlikely(carl9170_get_seq(skb) != seq))
1135                                 break;
1136
1137                         /* don't upload more than AMPDU FACTOR allows. */
1138                         if (unlikely(SEQ_DIFF(tid_info->snx, tid_info->bsn) >=
1139                             (tid_info->max - 1)))
1140                                 break;
1141
1142                         if (!carl9170_tx_rate_check(ar, skb, first))
1143                                 break;
1144
1145                         atomic_inc(&ar->tx_ampdu_upload);
1146                         tid_info->snx = seq = SEQ_NEXT(seq);
1147                         __skb_unlink(skb, &tid_info->queue);
1148
1149                         __skb_queue_tail(&agg, skb);
1150
1151                         if (skb_queue_len(&agg) >= CARL9170_NUM_TX_AGG_MAX)
1152                                 break;
1153                 }
1154
1155                 if (skb_queue_empty(&tid_info->queue) ||
1156                     carl9170_get_seq(skb_peek(&tid_info->queue)) !=
1157                     tid_info->snx) {
1158                         /*
1159                          * stop TID, if A-MPDU frames are still missing,
1160                          * or whenever the queue is empty.
1161                          */
1162
1163                         tid_info->state = CARL9170_TID_STATE_IDLE;
1164                 }
1165                 done_ampdus++;
1166
1167 processed:
1168                 spin_unlock_bh(&tid_info->lock);
1169
1170                 if (skb_queue_empty(&agg))
1171                         continue;
1172
1173                 /* apply ampdu spacing & factor settings */
1174                 carl9170_set_ampdu_params(ar, skb_peek(&agg));
1175
1176                 /* set aggregation push bit */
1177                 carl9170_set_immba(ar, skb_peek_tail(&agg));
1178
1179                 spin_lock_bh(&ar->tx_pending[queue].lock);
1180                 skb_queue_splice_tail_init(&agg, &ar->tx_pending[queue]);
1181                 spin_unlock_bh(&ar->tx_pending[queue].lock);
1182                 ar->tx_schedule = true;
1183         }
1184         if ((done_ampdus++ == 0) && (i++ == 0))
1185                 goto retry;
1186
1187         rcu_assign_pointer(ar->tx_ampdu_iter, tid_info);
1188         rcu_read_unlock();
1189 }
1190
1191 static struct sk_buff *carl9170_tx_pick_skb(struct ar9170 *ar,
1192                                             struct sk_buff_head *queue)
1193 {
1194         struct sk_buff *skb;
1195         struct ieee80211_tx_info *info;
1196         struct carl9170_tx_info *arinfo;
1197
1198         BUILD_BUG_ON(sizeof(*arinfo) > sizeof(info->rate_driver_data));
1199
1200         spin_lock_bh(&queue->lock);
1201         skb = skb_peek(queue);
1202         if (unlikely(!skb))
1203                 goto err_unlock;
1204
1205         if (carl9170_alloc_dev_space(ar, skb))
1206                 goto err_unlock;
1207
1208         __skb_unlink(skb, queue);
1209         spin_unlock_bh(&queue->lock);
1210
1211         info = IEEE80211_SKB_CB(skb);
1212         arinfo = (void *) info->rate_driver_data;
1213
1214         arinfo->timeout = jiffies;
1215         return skb;
1216
1217 err_unlock:
1218         spin_unlock_bh(&queue->lock);
1219         return NULL;
1220 }
1221
1222 void carl9170_tx_drop(struct ar9170 *ar, struct sk_buff *skb)
1223 {
1224         struct _carl9170_tx_superframe *super;
1225         uint8_t q = 0;
1226
1227         ar->tx_dropped++;
1228
1229         super = (void *)skb->data;
1230         SET_VAL(CARL9170_TX_SUPER_MISC_QUEUE, q,
1231                 ar9170_qmap[carl9170_get_queue(ar, skb)]);
1232         __carl9170_tx_process_status(ar, super->s.cookie, q);
1233 }
1234
1235 static bool carl9170_tx_ps_drop(struct ar9170 *ar, struct sk_buff *skb)
1236 {
1237         struct ieee80211_sta *sta;
1238         struct carl9170_sta_info *sta_info;
1239         struct ieee80211_tx_info *tx_info;
1240
1241         rcu_read_lock();
1242         sta = __carl9170_get_tx_sta(ar, skb);
1243         if (!sta)
1244                 goto out_rcu;
1245
1246         sta_info = (void *) sta->drv_priv;
1247         tx_info = IEEE80211_SKB_CB(skb);
1248
1249         if (unlikely(sta_info->sleeping) &&
1250             !(tx_info->flags & (IEEE80211_TX_CTL_NO_PS_BUFFER |
1251                                 IEEE80211_TX_CTL_CLEAR_PS_FILT))) {
1252                 rcu_read_unlock();
1253
1254                 if (tx_info->flags & IEEE80211_TX_CTL_AMPDU)
1255                         atomic_dec(&ar->tx_ampdu_upload);
1256
1257                 tx_info->flags |= IEEE80211_TX_STAT_TX_FILTERED;
1258                 carl9170_release_dev_space(ar, skb);
1259                 carl9170_tx_status(ar, skb, false);
1260                 return true;
1261         }
1262
1263 out_rcu:
1264         rcu_read_unlock();
1265         return false;
1266 }
1267
1268 static void carl9170_tx(struct ar9170 *ar)
1269 {
1270         struct sk_buff *skb;
1271         unsigned int i, q;
1272         bool schedule_garbagecollector = false;
1273
1274         ar->tx_schedule = false;
1275
1276         if (unlikely(!IS_STARTED(ar)))
1277                 return;
1278
1279         carl9170_usb_handle_tx_err(ar);
1280
1281         for (i = 0; i < ar->hw->queues; i++) {
1282                 while (!skb_queue_empty(&ar->tx_pending[i])) {
1283                         skb = carl9170_tx_pick_skb(ar, &ar->tx_pending[i]);
1284                         if (unlikely(!skb))
1285                                 break;
1286
1287                         if (unlikely(carl9170_tx_ps_drop(ar, skb)))
1288                                 continue;
1289
1290                         atomic_inc(&ar->tx_total_pending);
1291
1292                         q = __carl9170_get_queue(ar, i);
1293                         /*
1294                          * NB: tx_status[i] vs. tx_status[q],
1295                          * TODO: Move into pick_skb or alloc_dev_space.
1296                          */
1297                         skb_queue_tail(&ar->tx_status[q], skb);
1298
1299                         /*
1300                          * increase ref count to "2".
1301                          * Ref counting is the easiest way to solve the
1302                          * race between the urb's completion routine:
1303                          *      carl9170_tx_callback
1304                          * and wlan tx status functions:
1305                          *      carl9170_tx_status/janitor.
1306                          */
1307                         carl9170_tx_get_skb(skb);
1308
1309                         carl9170_usb_tx(ar, skb);
1310                         schedule_garbagecollector = true;
1311                 }
1312         }
1313
1314         if (!schedule_garbagecollector)
1315                 return;
1316
1317         ieee80211_queue_delayed_work(ar->hw, &ar->tx_janitor,
1318                 msecs_to_jiffies(CARL9170_TX_TIMEOUT));
1319 }
1320
1321 static bool carl9170_tx_ampdu_queue(struct ar9170 *ar,
1322         struct ieee80211_sta *sta, struct sk_buff *skb)
1323 {
1324         struct _carl9170_tx_superframe *super = (void *) skb->data;
1325         struct carl9170_sta_info *sta_info;
1326         struct carl9170_sta_tid *agg;
1327         struct sk_buff *iter;
1328         u16 tid, seq, qseq, off;
1329         bool run = false;
1330
1331         tid = carl9170_get_tid(skb);
1332         seq = carl9170_get_seq(skb);
1333         sta_info = (void *) sta->drv_priv;
1334
1335         rcu_read_lock();
1336         agg = rcu_dereference(sta_info->agg[tid]);
1337
1338         if (!agg)
1339                 goto err_unlock_rcu;
1340
1341         spin_lock_bh(&agg->lock);
1342         if (unlikely(agg->state < CARL9170_TID_STATE_IDLE))
1343                 goto err_unlock;
1344
1345         /* check if sequence is within the BA window */
1346         if (unlikely(!BAW_WITHIN(agg->bsn, CARL9170_BAW_BITS, seq)))
1347                 goto err_unlock;
1348
1349         if (WARN_ON_ONCE(!BAW_WITHIN(agg->snx, CARL9170_BAW_BITS, seq)))
1350                 goto err_unlock;
1351
1352         off = SEQ_DIFF(seq, agg->bsn);
1353         if (WARN_ON_ONCE(test_and_set_bit(off, agg->bitmap)))
1354                 goto err_unlock;
1355
1356         if (likely(BAW_WITHIN(agg->hsn, CARL9170_BAW_BITS, seq))) {
1357                 __skb_queue_tail(&agg->queue, skb);
1358                 agg->hsn = seq;
1359                 goto queued;
1360         }
1361
1362         skb_queue_reverse_walk(&agg->queue, iter) {
1363                 qseq = carl9170_get_seq(iter);
1364
1365                 if (BAW_WITHIN(qseq, CARL9170_BAW_BITS, seq)) {
1366                         __skb_queue_after(&agg->queue, iter, skb);
1367                         goto queued;
1368                 }
1369         }
1370
1371         __skb_queue_head(&agg->queue, skb);
1372 queued:
1373
1374         if (unlikely(agg->state != CARL9170_TID_STATE_XMIT)) {
1375                 if (agg->snx == carl9170_get_seq(skb_peek(&agg->queue))) {
1376                         agg->state = CARL9170_TID_STATE_XMIT;
1377                         run = true;
1378                 }
1379         }
1380
1381         spin_unlock_bh(&agg->lock);
1382         rcu_read_unlock();
1383
1384         return run;
1385
1386 err_unlock:
1387         spin_unlock_bh(&agg->lock);
1388
1389 err_unlock_rcu:
1390         rcu_read_unlock();
1391         super->f.mac_control &= ~cpu_to_le16(AR9170_TX_MAC_AGGR);
1392         carl9170_tx_status(ar, skb, false);
1393         ar->tx_dropped++;
1394         return false;
1395 }
1396
1397 void carl9170_op_tx(struct ieee80211_hw *hw, struct sk_buff *skb)
1398 {
1399         struct ar9170 *ar = hw->priv;
1400         struct ieee80211_tx_info *info;
1401         struct ieee80211_sta *sta;
1402         bool run;
1403
1404         if (unlikely(!IS_STARTED(ar)))
1405                 goto err_free;
1406
1407         info = IEEE80211_SKB_CB(skb);
1408         sta = info->control.sta;
1409
1410         if (unlikely(carl9170_tx_prepare(ar, skb)))
1411                 goto err_free;
1412
1413         carl9170_tx_accounting(ar, skb);
1414         /*
1415          * from now on, one has to use carl9170_tx_status to free
1416          * all ressouces which are associated with the frame.
1417          */
1418
1419         if (sta) {
1420                 struct carl9170_sta_info *stai = (void *) sta->drv_priv;
1421                 atomic_inc(&stai->pending_frames);
1422         }
1423
1424         if (info->flags & IEEE80211_TX_CTL_AMPDU) {
1425                 run = carl9170_tx_ampdu_queue(ar, sta, skb);
1426                 if (run)
1427                         carl9170_tx_ampdu(ar);
1428
1429         } else {
1430                 unsigned int queue = skb_get_queue_mapping(skb);
1431
1432                 skb_queue_tail(&ar->tx_pending[queue], skb);
1433         }
1434
1435         carl9170_tx(ar);
1436         return;
1437
1438 err_free:
1439         ar->tx_dropped++;
1440         ieee80211_free_txskb(ar->hw, skb);
1441 }
1442
1443 void carl9170_tx_scheduler(struct ar9170 *ar)
1444 {
1445
1446         if (ar->tx_ampdu_schedule)
1447                 carl9170_tx_ampdu(ar);
1448
1449         if (ar->tx_schedule)
1450                 carl9170_tx(ar);
1451 }
1452
1453 int carl9170_update_beacon(struct ar9170 *ar, const bool submit)
1454 {
1455         struct sk_buff *skb = NULL;
1456         struct carl9170_vif_info *cvif;
1457         struct ieee80211_tx_info *txinfo;
1458         struct ieee80211_tx_rate *rate;
1459         __le32 *data, *old = NULL;
1460         unsigned int plcp, power, chains;
1461         u32 word, ht1, off, addr, len;
1462         int i = 0, err = 0;
1463
1464         rcu_read_lock();
1465         cvif = rcu_dereference(ar->beacon_iter);
1466 retry:
1467         if (ar->vifs == 0 || !cvif)
1468                 goto out_unlock;
1469
1470         list_for_each_entry_continue_rcu(cvif, &ar->vif_list, list) {
1471                 if (cvif->active && cvif->enable_beacon)
1472                         goto found;
1473         }
1474
1475         if (!ar->beacon_enabled || i++)
1476                 goto out_unlock;
1477
1478         goto retry;
1479
1480 found:
1481         rcu_assign_pointer(ar->beacon_iter, cvif);
1482
1483         skb = ieee80211_beacon_get_tim(ar->hw, carl9170_get_vif(cvif),
1484                 NULL, NULL);
1485
1486         if (!skb) {
1487                 err = -ENOMEM;
1488                 goto err_free;
1489         }
1490
1491         txinfo = IEEE80211_SKB_CB(skb);
1492         spin_lock_bh(&ar->beacon_lock);
1493         data = (__le32 *)skb->data;
1494         if (cvif->beacon)
1495                 old = (__le32 *)cvif->beacon->data;
1496
1497         off = cvif->id * AR9170_MAC_BCN_LENGTH_MAX;
1498         addr = ar->fw.beacon_addr + off;
1499         len = roundup(skb->len + FCS_LEN, 4);
1500
1501         if ((off + len) > ar->fw.beacon_max_len) {
1502                 if (net_ratelimit()) {
1503                         wiphy_err(ar->hw->wiphy, "beacon does not "
1504                                   "fit into device memory!\n");
1505                 }
1506                 err = -EINVAL;
1507                 goto err_unlock;
1508         }
1509
1510         if (len > AR9170_MAC_BCN_LENGTH_MAX) {
1511                 if (net_ratelimit()) {
1512                         wiphy_err(ar->hw->wiphy, "no support for beacons "
1513                                 "bigger than %d (yours:%d).\n",
1514                                  AR9170_MAC_BCN_LENGTH_MAX, len);
1515                 }
1516
1517                 err = -EMSGSIZE;
1518                 goto err_unlock;
1519         }
1520
1521         ht1 = AR9170_MAC_BCN_HT1_TX_ANT0;
1522         rate = &txinfo->control.rates[0];
1523         carl9170_tx_rate_tpc_chains(ar, txinfo, rate, &plcp, &power, &chains);
1524         if (!(txinfo->control.rates[0].flags & IEEE80211_TX_RC_MCS)) {
1525                 if (plcp <= AR9170_TX_PHY_RATE_CCK_11M)
1526                         plcp |= ((skb->len + FCS_LEN) << (3 + 16)) + 0x0400;
1527                 else
1528                         plcp |= ((skb->len + FCS_LEN) << 16) + 0x0010;
1529         } else {
1530                 ht1 |= AR9170_MAC_BCN_HT1_HT_EN;
1531                 if (rate->flags & IEEE80211_TX_RC_SHORT_GI)
1532                         plcp |= AR9170_MAC_BCN_HT2_SGI;
1533
1534                 if (rate->flags & IEEE80211_TX_RC_40_MHZ_WIDTH) {
1535                         ht1 |= AR9170_MAC_BCN_HT1_BWC_40M_SHARED;
1536                         plcp |= AR9170_MAC_BCN_HT2_BW40;
1537                 }
1538                 if (rate->flags & IEEE80211_TX_RC_DUP_DATA) {
1539                         ht1 |= AR9170_MAC_BCN_HT1_BWC_40M_DUP;
1540                         plcp |= AR9170_MAC_BCN_HT2_BW40;
1541                 }
1542
1543                 SET_VAL(AR9170_MAC_BCN_HT2_LEN, plcp, skb->len + FCS_LEN);
1544         }
1545
1546         SET_VAL(AR9170_MAC_BCN_HT1_PWR_CTRL, ht1, 7);
1547         SET_VAL(AR9170_MAC_BCN_HT1_TPC, ht1, power);
1548         SET_VAL(AR9170_MAC_BCN_HT1_CHAIN_MASK, ht1, chains);
1549         if (chains == AR9170_TX_PHY_TXCHAIN_2)
1550                 ht1 |= AR9170_MAC_BCN_HT1_TX_ANT1;
1551
1552         carl9170_async_regwrite_begin(ar);
1553         carl9170_async_regwrite(AR9170_MAC_REG_BCN_HT1, ht1);
1554         if (!(txinfo->control.rates[0].flags & IEEE80211_TX_RC_MCS))
1555                 carl9170_async_regwrite(AR9170_MAC_REG_BCN_PLCP, plcp);
1556         else
1557                 carl9170_async_regwrite(AR9170_MAC_REG_BCN_HT2, plcp);
1558
1559         for (i = 0; i < DIV_ROUND_UP(skb->len, 4); i++) {
1560                 /*
1561                  * XXX: This accesses beyond skb data for up
1562                  *      to the last 3 bytes!!
1563                  */
1564
1565                 if (old && (data[i] == old[i]))
1566                         continue;
1567
1568                 word = le32_to_cpu(data[i]);
1569                 carl9170_async_regwrite(addr + 4 * i, word);
1570         }
1571         carl9170_async_regwrite_finish();
1572
1573         dev_kfree_skb_any(cvif->beacon);
1574         cvif->beacon = NULL;
1575
1576         err = carl9170_async_regwrite_result();
1577         if (!err)
1578                 cvif->beacon = skb;
1579         spin_unlock_bh(&ar->beacon_lock);
1580         if (err)
1581                 goto err_free;
1582
1583         if (submit) {
1584                 err = carl9170_bcn_ctrl(ar, cvif->id,
1585                                         CARL9170_BCN_CTRL_CAB_TRIGGER,
1586                                         addr, skb->len + FCS_LEN);
1587
1588                 if (err)
1589                         goto err_free;
1590         }
1591 out_unlock:
1592         rcu_read_unlock();
1593         return 0;
1594
1595 err_unlock:
1596         spin_unlock_bh(&ar->beacon_lock);
1597
1598 err_free:
1599         rcu_read_unlock();
1600         dev_kfree_skb_any(skb);
1601         return err;
1602 }