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.ack_signal) != 20);
 281
 282         memset(&txinfo->status.ack_signal, 0,
 283                sizeof(struct ieee80211_tx_info) -
 284                offsetof(struct ieee80211_tx_info, status.ack_signal));
 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 static void carl9170_tx_bar_status(struct ar9170 *ar, struct sk_buff *skb,
 440         struct ieee80211_tx_info *tx_info)
 441 {
 442         struct _carl9170_tx_superframe *super = (void *) skb->data;
 443         struct ieee80211_bar *bar = (void *) super->frame_data;
 444
 445         /*
 446          * Unlike all other frames, the status report for BARs does
 447          * not directly come from the hardware as it is incapable of
 448          * matching a BA to a previously send BAR.
 449          * Instead the RX-path will scan for incoming BAs and set the
 450          * IEEE80211_TX_STAT_ACK if it sees one that was likely
 451          * caused by a BAR from us.
 452          */
 453
 454         if (unlikely(ieee80211_is_back_req(bar->frame_control)) &&
 455            !(tx_info->flags & IEEE80211_TX_STAT_ACK)) {
 456                 struct carl9170_bar_list_entry *entry;
 457                 int queue = skb_get_queue_mapping(skb);
 458
 459                 rcu_read_lock();
 460                 list_for_each_entry_rcu(entry, &ar->bar_list[queue], list) {
 461                         if (entry->skb == skb) {
 462                                 spin_lock_bh(&ar->bar_list_lock[queue]);
 463                                 list_del_rcu(&entry->list);
 464                                 spin_unlock_bh(&ar->bar_list_lock[queue]);
 465                                 kfree_rcu(entry, head);
 466                                 goto out;
 467                         }
 468                 }
 469
 470                 WARN(1, "bar not found in %d - ra:%pM ta:%pM c:%x ssn:%x\n",
 471                        queue, bar->ra, bar->ta, bar->control,
 472                         bar->start_seq_num);
 473 out:
 474                 rcu_read_unlock();
 475         }
 476 }
 477
 478 void carl9170_tx_status(struct ar9170 *ar, struct sk_buff *skb,
 479                         const bool success)
 480 {
 481         struct ieee80211_tx_info *txinfo;
 482
 483         carl9170_tx_accounting_free(ar, skb);
 484
 485         txinfo = IEEE80211_SKB_CB(skb);
 486
 487         carl9170_tx_bar_status(ar, skb, txinfo);
 488
 489         if (success)
 490                 txinfo->flags |= IEEE80211_TX_STAT_ACK;
 491         else
 492                 ar->tx_ack_failures++;
 493
 494         if (txinfo->flags & IEEE80211_TX_CTL_AMPDU)
 495                 carl9170_tx_status_process_ampdu(ar, skb, txinfo);
 496
 497         carl9170_tx_ps_unblock(ar, skb);
 498         carl9170_tx_put_skb(skb);
 499 }
 500
 501 /* This function may be called form any context */
 502 void carl9170_tx_callback(struct ar9170 *ar, struct sk_buff *skb)
 503 {
 504         struct ieee80211_tx_info *txinfo = IEEE80211_SKB_CB(skb);
 505
 506         atomic_dec(&ar->tx_total_pending);
 507
 508         if (txinfo->flags & IEEE80211_TX_CTL_AMPDU)
 509                 atomic_dec(&ar->tx_ampdu_upload);
 510
 511         if (carl9170_tx_put_skb(skb))
 512                 tasklet_hi_schedule(&ar->usb_tasklet);
 513 }
 514
 515 static struct sk_buff *carl9170_get_queued_skb(struct ar9170 *ar, u8 cookie,
 516                                                struct sk_buff_head *queue)
 517 {
 518         struct sk_buff *skb;
 519
 520         spin_lock_bh(&queue->lock);
 521         skb_queue_walk(queue, skb) {
 522                 struct _carl9170_tx_superframe *txc = (void *) skb->data;
 523
 524                 if (txc->s.cookie != cookie)
 525                         continue;
 526
 527                 __skb_unlink(skb, queue);
 528                 spin_unlock_bh(&queue->lock);
 529
 530                 carl9170_release_dev_space(ar, skb);
 531                 return skb;
 532         }
 533         spin_unlock_bh(&queue->lock);
 534
 535         return NULL;
 536 }
 537
 538 static void carl9170_tx_fill_rateinfo(struct ar9170 *ar, unsigned int rix,
 539         unsigned int tries, struct ieee80211_tx_info *txinfo)
 540 {
 541         unsigned int i;
 542
 543         for (i = 0; i < IEEE80211_TX_MAX_RATES; i++) {
 544                 if (txinfo->status.rates[i].idx < 0)
 545                         break;
 546
 547                 if (i == rix) {
 548                         txinfo->status.rates[i].count = tries;
 549                         i++;
 550                         break;
 551                 }
 552         }
 553
 554         for (; i < IEEE80211_TX_MAX_RATES; i++) {
 555                 txinfo->status.rates[i].idx = -1;
 556                 txinfo->status.rates[i].count = 0;
 557         }
 558 }
 559
 560 static void carl9170_check_queue_stop_timeout(struct ar9170 *ar)
 561 {
 562         int i;
 563         struct sk_buff *skb;
 564         struct ieee80211_tx_info *txinfo;
 565         struct carl9170_tx_info *arinfo;
 566         bool restart = false;
 567
 568         for (i = 0; i < ar->hw->queues; i++) {
 569                 spin_lock_bh(&ar->tx_status[i].lock);
 570
 571                 skb = skb_peek(&ar->tx_status[i]);
 572
 573                 if (!skb)
 574                         goto next;
 575
 576                 txinfo = IEEE80211_SKB_CB(skb);
 577                 arinfo = (void *) txinfo->rate_driver_data;
 578
 579                 if (time_is_before_jiffies(arinfo->timeout +
 580                     msecs_to_jiffies(CARL9170_QUEUE_STUCK_TIMEOUT)) == true)
 581                         restart = true;
 582
 583 next:
 584                 spin_unlock_bh(&ar->tx_status[i].lock);
 585         }
 586
 587         if (restart) {
 588                 /*
 589                  * At least one queue has been stuck for long enough.
 590                  * Give the device a kick and hope it gets back to
 591                  * work.
 592                  *
 593                  * possible reasons may include:
 594                  *  - frames got lost/corrupted (bad connection to the device)
 595                  *  - stalled rx processing/usb controller hiccups
 596                  *  - firmware errors/bugs
 597                  *  - every bug you can think of.
 598                  *  - all bugs you can't...
 599                  *  - ...
 600                  */
 601                 carl9170_restart(ar, CARL9170_RR_STUCK_TX);
 602         }
 603 }
 604
 605 static void carl9170_tx_ampdu_timeout(struct ar9170 *ar)
 606 {
 607         struct carl9170_sta_tid *iter;
 608         struct sk_buff *skb;
 609         struct ieee80211_tx_info *txinfo;
 610         struct carl9170_tx_info *arinfo;
 611         struct ieee80211_sta *sta;
 612
 613         rcu_read_lock();
 614         list_for_each_entry_rcu(iter, &ar->tx_ampdu_list, list) {
 615                 if (iter->state < CARL9170_TID_STATE_IDLE)
 616                         continue;
 617
 618                 spin_lock_bh(&iter->lock);
 619                 skb = skb_peek(&iter->queue);
 620                 if (!skb)
 621                         goto unlock;
 622
 623                 txinfo = IEEE80211_SKB_CB(skb);
 624                 arinfo = (void *)txinfo->rate_driver_data;
 625                 if (time_is_after_jiffies(arinfo->timeout +
 626                     msecs_to_jiffies(CARL9170_QUEUE_TIMEOUT)))
 627                         goto unlock;
 628
 629                 sta = __carl9170_get_tx_sta(ar, skb);
 630                 if (WARN_ON(!sta))
 631                         goto unlock;
 632
 633                 ieee80211_stop_tx_ba_session(sta, iter->tid);
 634 unlock:
 635                 spin_unlock_bh(&iter->lock);
 636
 637         }
 638         rcu_read_unlock();
 639 }
 640
 641 void carl9170_tx_janitor(struct work_struct *work)
 642 {
 643         struct ar9170 *ar = container_of(work, struct ar9170,
 644                                          tx_janitor.work);
 645         if (!IS_STARTED(ar))
 646                 return;
 647
 648         ar->tx_janitor_last_run = jiffies;
 649
 650         carl9170_check_queue_stop_timeout(ar);
 651         carl9170_tx_ampdu_timeout(ar);
 652
 653         if (!atomic_read(&ar->tx_total_queued))
 654                 return;
 655
 656         ieee80211_queue_delayed_work(ar->hw, &ar->tx_janitor,
 657                 msecs_to_jiffies(CARL9170_TX_TIMEOUT));
 658 }
 659
 660 static void __carl9170_tx_process_status(struct ar9170 *ar,
 661         const uint8_t cookie, const uint8_t info)
 662 {
 663         struct sk_buff *skb;
 664         struct ieee80211_tx_info *txinfo;
 665         unsigned int r, t, q;
 666         bool success = true;
 667
 668         q = ar9170_qmap[info & CARL9170_TX_STATUS_QUEUE];
 669
 670         skb = carl9170_get_queued_skb(ar, cookie, &ar->tx_status[q]);
 671         if (!skb) {
 672                 /*
 673                  * We have lost the race to another thread.
 674                  */
 675
 676                 return ;
 677         }
 678
 679         txinfo = IEEE80211_SKB_CB(skb);
 680
 681         if (!(info & CARL9170_TX_STATUS_SUCCESS))
 682                 success = false;
 683
 684         r = (info & CARL9170_TX_STATUS_RIX) >> CARL9170_TX_STATUS_RIX_S;
 685         t = (info & CARL9170_TX_STATUS_TRIES) >> CARL9170_TX_STATUS_TRIES_S;
 686
 687         carl9170_tx_fill_rateinfo(ar, r, t, txinfo);
 688         carl9170_tx_status(ar, skb, success);
 689 }
 690
 691 void carl9170_tx_process_status(struct ar9170 *ar,
 692                                 const struct carl9170_rsp *cmd)
 693 {
 694         unsigned int i;
 695
 696         for (i = 0;  i < cmd->hdr.ext; i++) {
 697                 if (WARN_ON(i > ((cmd->hdr.len / 2) + 1))) {
 698                         print_hex_dump_bytes("UU:", DUMP_PREFIX_NONE,
 699                                              (void *) cmd, cmd->hdr.len + 4);
 700                         break;
 701                 }
 702
 703                 __carl9170_tx_process_status(ar, cmd->_tx_status[i].cookie,
 704                                              cmd->_tx_status[i].info);
 705         }
 706 }
 707
 708 static void carl9170_tx_rate_tpc_chains(struct ar9170 *ar,
 709         struct ieee80211_tx_info *info, struct ieee80211_tx_rate *txrate,
 710         unsigned int *phyrate, unsigned int *tpc, unsigned int *chains)
 711 {
 712         struct ieee80211_rate *rate = NULL;
 713         u8 *txpower;
 714         unsigned int idx;
 715
 716         idx = txrate->idx;
 717         *tpc = 0;
 718         *phyrate = 0;
 719
 720         if (txrate->flags & IEEE80211_TX_RC_MCS) {
 721                 if (txrate->flags & IEEE80211_TX_RC_40_MHZ_WIDTH) {
 722                         /* +1 dBm for HT40 */
 723                         *tpc += 2;
 724
 725                         if (info->band == IEEE80211_BAND_2GHZ)
 726                                 txpower = ar->power_2G_ht40;
 727                         else
 728                                 txpower = ar->power_5G_ht40;
 729                 } else {
 730                         if (info->band == IEEE80211_BAND_2GHZ)
 731                                 txpower = ar->power_2G_ht20;
 732                         else
 733                                 txpower = ar->power_5G_ht20;
 734                 }
 735
 736                 *phyrate = txrate->idx;
 737                 *tpc += txpower[idx & 7];
 738         } else {
 739                 if (info->band == IEEE80211_BAND_2GHZ) {
 740                         if (idx < 4)
 741                                 txpower = ar->power_2G_cck;
 742                         else
 743                                 txpower = ar->power_2G_ofdm;
 744                 } else {
 745                         txpower = ar->power_5G_leg;
 746                         idx += 4;
 747                 }
 748
 749                 rate = &__carl9170_ratetable[idx];
 750                 *tpc += txpower[(rate->hw_value & 0x30) >> 4];
 751                 *phyrate = rate->hw_value & 0xf;
 752         }
 753
 754         if (ar->eeprom.tx_mask == 1) {
 755                 *chains = AR9170_TX_PHY_TXCHAIN_1;
 756         } else {
 757                 if (!(txrate->flags & IEEE80211_TX_RC_MCS) &&
 758                     rate && rate->bitrate >= 360)
 759                         *chains = AR9170_TX_PHY_TXCHAIN_1;
 760                 else
 761                         *chains = AR9170_TX_PHY_TXCHAIN_2;
 762         }
 763
 764         *tpc = min_t(unsigned int, *tpc, ar->hw->conf.power_level * 2);
 765 }
 766
 767 static __le32 carl9170_tx_physet(struct ar9170 *ar,
 768         struct ieee80211_tx_info *info, struct ieee80211_tx_rate *txrate)
 769 {
 770         unsigned int power = 0, chains = 0, phyrate = 0;
 771         __le32 tmp;
 772
 773         tmp = cpu_to_le32(0);
 774
 775         if (txrate->flags & IEEE80211_TX_RC_40_MHZ_WIDTH)
 776                 tmp |= cpu_to_le32(AR9170_TX_PHY_BW_40MHZ <<
 777                         AR9170_TX_PHY_BW_S);
 778         /* this works because 40 MHz is 2 and dup is 3 */
 779         if (txrate->flags & IEEE80211_TX_RC_DUP_DATA)
 780                 tmp |= cpu_to_le32(AR9170_TX_PHY_BW_40MHZ_DUP <<
 781                         AR9170_TX_PHY_BW_S);
 782
 783         if (txrate->flags & IEEE80211_TX_RC_SHORT_GI)
 784                 tmp |= cpu_to_le32(AR9170_TX_PHY_SHORT_GI);
 785
 786         if (txrate->flags & IEEE80211_TX_RC_MCS) {
 787                 SET_VAL(AR9170_TX_PHY_MCS, phyrate, txrate->idx);
 788
 789                 /* heavy clip control */
 790                 tmp |= cpu_to_le32((txrate->idx & 0x7) <<
 791                         AR9170_TX_PHY_TX_HEAVY_CLIP_S);
 792
 793                 tmp |= cpu_to_le32(AR9170_TX_PHY_MOD_HT);
 794
 795                 /*
 796                  * green field preamble does not work.
 797                  *
 798                  * if (txrate->flags & IEEE80211_TX_RC_GREEN_FIELD)
 799                  * tmp |= cpu_to_le32(AR9170_TX_PHY_GREENFIELD);
 800                  */
 801         } else {
 802                 if (info->band == IEEE80211_BAND_2GHZ) {
 803                         if (txrate->idx <= AR9170_TX_PHY_RATE_CCK_11M)
 804                                 tmp |= cpu_to_le32(AR9170_TX_PHY_MOD_CCK);
 805                         else
 806                                 tmp |= cpu_to_le32(AR9170_TX_PHY_MOD_OFDM);
 807                 } else {
 808                         tmp |= cpu_to_le32(AR9170_TX_PHY_MOD_OFDM);
 809                 }
 810
 811                 /*
 812                  * short preamble seems to be broken too.
 813                  *
 814                  * if (txrate->flags & IEEE80211_TX_RC_USE_SHORT_PREAMBLE)
 815                  *      tmp |= cpu_to_le32(AR9170_TX_PHY_SHORT_PREAMBLE);
 816                  */
 817         }
 818         carl9170_tx_rate_tpc_chains(ar, info, txrate,
 819                                     &phyrate, &power, &chains);
 820
 821         tmp |= cpu_to_le32(SET_CONSTVAL(AR9170_TX_PHY_MCS, phyrate));
 822         tmp |= cpu_to_le32(SET_CONSTVAL(AR9170_TX_PHY_TX_PWR, power));
 823         tmp |= cpu_to_le32(SET_CONSTVAL(AR9170_TX_PHY_TXCHAIN, chains));
 824         return tmp;
 825 }
 826
 827 static bool carl9170_tx_rts_check(struct ar9170 *ar,
 828                                   struct ieee80211_tx_rate *rate,
 829                                   bool ampdu, bool multi)
 830 {
 831         switch (ar->erp_mode) {
 832         case CARL9170_ERP_AUTO:
 833                 if (ampdu)
 834                         break;
 835
 836         case CARL9170_ERP_MAC80211:
 837                 if (!(rate->flags & IEEE80211_TX_RC_USE_RTS_CTS))
 838                         break;
 839
 840         case CARL9170_ERP_RTS:
 841                 if (likely(!multi))
 842                         return true;
 843
 844         default:
 845                 break;
 846         }
 847
 848         return false;
 849 }
 850
 851 static bool carl9170_tx_cts_check(struct ar9170 *ar,
 852                                   struct ieee80211_tx_rate *rate)
 853 {
 854         switch (ar->erp_mode) {
 855         case CARL9170_ERP_AUTO:
 856         case CARL9170_ERP_MAC80211:
 857                 if (!(rate->flags & IEEE80211_TX_RC_USE_CTS_PROTECT))
 858                         break;
 859
 860         case CARL9170_ERP_CTS:
 861                 return true;
 862
 863         default:
 864                 break;
 865         }
 866
 867         return false;
 868 }
 869
 870 static int carl9170_tx_prepare(struct ar9170 *ar, struct sk_buff *skb)
 871 {
 872         struct ieee80211_hdr *hdr;
 873         struct _carl9170_tx_superframe *txc;
 874         struct carl9170_vif_info *cvif;
 875         struct ieee80211_tx_info *info;
 876         struct ieee80211_tx_rate *txrate;
 877         struct ieee80211_sta *sta;
 878         struct carl9170_tx_info *arinfo;
 879         unsigned int hw_queue;
 880         int i;
 881         __le16 mac_tmp;
 882         u16 len;
 883         bool ampdu, no_ack;
 884
 885         BUILD_BUG_ON(sizeof(*arinfo) > sizeof(info->rate_driver_data));
 886         BUILD_BUG_ON(sizeof(struct _carl9170_tx_superdesc) !=
 887                      CARL9170_TX_SUPERDESC_LEN);
 888
 889         BUILD_BUG_ON(sizeof(struct _ar9170_tx_hwdesc) !=
 890                      AR9170_TX_HWDESC_LEN);
 891
 892         BUILD_BUG_ON(IEEE80211_TX_MAX_RATES < CARL9170_TX_MAX_RATES);
 893
 894         BUILD_BUG_ON(AR9170_MAX_VIRTUAL_MAC >
 895                 ((CARL9170_TX_SUPER_MISC_VIF_ID >>
 896                  CARL9170_TX_SUPER_MISC_VIF_ID_S) + 1));
 897
 898         hw_queue = ar9170_qmap[carl9170_get_queue(ar, skb)];
 899
 900         hdr = (void *)skb->data;
 901         info = IEEE80211_SKB_CB(skb);
 902         len = skb->len;
 903
 904         /*
 905          * Note: If the frame was sent through a monitor interface,
 906          * the ieee80211_vif pointer can be NULL.
 907          */
 908         if (likely(info->control.vif))
 909                 cvif = (void *) info->control.vif->drv_priv;
 910         else
 911                 cvif = NULL;
 912
 913         sta = info->control.sta;
 914
 915         txc = (void *)skb_push(skb, sizeof(*txc));
 916         memset(txc, 0, sizeof(*txc));
 917
 918         SET_VAL(CARL9170_TX_SUPER_MISC_QUEUE, txc->s.misc, hw_queue);
 919
 920         if (likely(cvif))
 921                 SET_VAL(CARL9170_TX_SUPER_MISC_VIF_ID, txc->s.misc, cvif->id);
 922
 923         if (unlikely(info->flags & IEEE80211_TX_CTL_SEND_AFTER_DTIM))
 924                 txc->s.misc |= CARL9170_TX_SUPER_MISC_CAB;
 925
 926         if (unlikely(info->flags & IEEE80211_TX_CTL_ASSIGN_SEQ))
 927                 txc->s.misc |= CARL9170_TX_SUPER_MISC_ASSIGN_SEQ;
 928
 929         if (unlikely(ieee80211_is_probe_resp(hdr->frame_control)))
 930                 txc->s.misc |= CARL9170_TX_SUPER_MISC_FILL_IN_TSF;
 931
 932         mac_tmp = cpu_to_le16(AR9170_TX_MAC_HW_DURATION |
 933                               AR9170_TX_MAC_BACKOFF);
 934         mac_tmp |= cpu_to_le16((hw_queue << AR9170_TX_MAC_QOS_S) &
 935                                AR9170_TX_MAC_QOS);
 936
 937         no_ack = !!(info->flags & IEEE80211_TX_CTL_NO_ACK);
 938         if (unlikely(no_ack))
 939                 mac_tmp |= cpu_to_le16(AR9170_TX_MAC_NO_ACK);
 940
 941         if (info->control.hw_key) {
 942                 len += info->control.hw_key->icv_len;
 943
 944                 switch (info->control.hw_key->cipher) {
 945                 case WLAN_CIPHER_SUITE_WEP40:
 946                 case WLAN_CIPHER_SUITE_WEP104:
 947                 case WLAN_CIPHER_SUITE_TKIP:
 948                         mac_tmp |= cpu_to_le16(AR9170_TX_MAC_ENCR_RC4);
 949                         break;
 950                 case WLAN_CIPHER_SUITE_CCMP:
 951                         mac_tmp |= cpu_to_le16(AR9170_TX_MAC_ENCR_AES);
 952                         break;
 953                 default:
 954                         WARN_ON(1);
 955                         goto err_out;
 956                 }
 957         }
 958
 959         ampdu = !!(info->flags & IEEE80211_TX_CTL_AMPDU);
 960         if (ampdu) {
 961                 unsigned int density, factor;
 962
 963                 if (unlikely(!sta || !cvif))
 964                         goto err_out;
 965
 966                 factor = min_t(unsigned int, 1u, sta->ht_cap.ampdu_factor);
 967                 density = sta->ht_cap.ampdu_density;
 968
 969                 if (density) {
 970                         /*
 971                          * Watch out!
 972                          *
 973                          * Otus uses slightly different density values than
 974                          * those from the 802.11n spec.
 975                          */
 976
 977                         density = max_t(unsigned int, density + 1, 7u);
 978                 }
 979
 980                 SET_VAL(CARL9170_TX_SUPER_AMPDU_DENSITY,
 981                         txc->s.ampdu_settings, density);
 982
 983                 SET_VAL(CARL9170_TX_SUPER_AMPDU_FACTOR,
 984                         txc->s.ampdu_settings, factor);
 985
 986                 for (i = 0; i < CARL9170_TX_MAX_RATES; i++) {
 987                         txrate = &info->control.rates[i];
 988                         if (txrate->idx >= 0) {
 989                                 txc->s.ri[i] =
 990                                         CARL9170_TX_SUPER_RI_AMPDU;
 991
 992                                 if (WARN_ON(!(txrate->flags &
 993                                               IEEE80211_TX_RC_MCS))) {
 994                                         /*
 995                                          * Not sure if it's even possible
 996                                          * to aggregate non-ht rates with
 997                                          * this HW.
 998                                          */
 999                                         goto err_out;
1000                                 }
1001                                 continue;
1002                         }
1003
1004                         txrate->idx = 0;
1005                         txrate->count = ar->hw->max_rate_tries;
1006                 }
1007
1008                 mac_tmp |= cpu_to_le16(AR9170_TX_MAC_AGGR);
1009         }
1010
1011         /*
1012          * NOTE: For the first rate, the ERP & AMPDU flags are directly
1013          * taken from mac_control. For all fallback rate, the firmware
1014          * updates the mac_control flags from the rate info field.
1015          */
1016         for (i = 1; i < CARL9170_TX_MAX_RATES; i++) {
1017                 txrate = &info->control.rates[i];
1018                 if (txrate->idx < 0)
1019                         break;
1020
1021                 SET_VAL(CARL9170_TX_SUPER_RI_TRIES, txc->s.ri[i],
1022                         txrate->count);
1023
1024                 if (carl9170_tx_rts_check(ar, txrate, ampdu, no_ack))
1025                         txc->s.ri[i] |= (AR9170_TX_MAC_PROT_RTS <<
1026                                 CARL9170_TX_SUPER_RI_ERP_PROT_S);
1027                 else if (carl9170_tx_cts_check(ar, txrate))
1028                         txc->s.ri[i] |= (AR9170_TX_MAC_PROT_CTS <<
1029                                 CARL9170_TX_SUPER_RI_ERP_PROT_S);
1030
1031                 txc->s.rr[i - 1] = carl9170_tx_physet(ar, info, txrate);
1032         }
1033
1034         txrate = &info->control.rates[0];
1035         SET_VAL(CARL9170_TX_SUPER_RI_TRIES, txc->s.ri[0], txrate->count);
1036
1037         if (carl9170_tx_rts_check(ar, txrate, ampdu, no_ack))
1038                 mac_tmp |= cpu_to_le16(AR9170_TX_MAC_PROT_RTS);
1039         else if (carl9170_tx_cts_check(ar, txrate))
1040                 mac_tmp |= cpu_to_le16(AR9170_TX_MAC_PROT_CTS);
1041
1042         txc->s.len = cpu_to_le16(skb->len);
1043         txc->f.length = cpu_to_le16(len + FCS_LEN);
1044         txc->f.mac_control = mac_tmp;
1045         txc->f.phy_control = carl9170_tx_physet(ar, info, txrate);
1046
1047         arinfo = (void *)info->rate_driver_data;
1048         arinfo->timeout = jiffies;
1049         arinfo->ar = ar;
1050         kref_init(&arinfo->ref);
1051         return 0;
1052
1053 err_out:
1054         skb_pull(skb, sizeof(*txc));
1055         return -EINVAL;
1056 }
1057
1058 static void carl9170_set_immba(struct ar9170 *ar, struct sk_buff *skb)
1059 {
1060         struct _carl9170_tx_superframe *super;
1061
1062         super = (void *) skb->data;
1063         super->f.mac_control |= cpu_to_le16(AR9170_TX_MAC_IMM_BA);
1064 }
1065
1066 static void carl9170_set_ampdu_params(struct ar9170 *ar, struct sk_buff *skb)
1067 {
1068         struct _carl9170_tx_superframe *super;
1069         int tmp;
1070
1071         super = (void *) skb->data;
1072
1073         tmp = (super->s.ampdu_settings & CARL9170_TX_SUPER_AMPDU_DENSITY) <<
1074                 CARL9170_TX_SUPER_AMPDU_DENSITY_S;
1075
1076         /*
1077          * If you haven't noticed carl9170_tx_prepare has already filled
1078          * in all ampdu spacing & factor parameters.
1079          * Now it's the time to check whenever the settings have to be
1080          * updated by the firmware, or if everything is still the same.
1081          *
1082          * There's no sane way to handle different density values with
1083          * this hardware, so we may as well just do the compare in the
1084          * driver.
1085          */
1086
1087         if (tmp != ar->current_density) {
1088                 ar->current_density = tmp;
1089                 super->s.ampdu_settings |=
1090                         CARL9170_TX_SUPER_AMPDU_COMMIT_DENSITY;
1091         }
1092
1093         tmp = (super->s.ampdu_settings & CARL9170_TX_SUPER_AMPDU_FACTOR) <<
1094                 CARL9170_TX_SUPER_AMPDU_FACTOR_S;
1095
1096         if (tmp != ar->current_factor) {
1097                 ar->current_factor = tmp;
1098                 super->s.ampdu_settings |=
1099                         CARL9170_TX_SUPER_AMPDU_COMMIT_FACTOR;
1100         }
1101 }
1102
1103 static bool carl9170_tx_rate_check(struct ar9170 *ar, struct sk_buff *_dest,
1104                                    struct sk_buff *_src)
1105 {
1106         struct _carl9170_tx_superframe *dest, *src;
1107
1108         dest = (void *) _dest->data;
1109         src = (void *) _src->data;
1110
1111         /*
1112          * The mac80211 rate control algorithm expects that all MPDUs in
1113          * an AMPDU share the same tx vectors.
1114          * This is not really obvious right now, because the hardware
1115          * does the AMPDU setup according to its own rulebook.
1116          * Our nicely assembled, strictly monotonic increasing mpdu
1117          * chains will be broken up, mashed back together...
1118          */
1119
1120         return (dest->f.phy_control == src->f.phy_control);
1121 }
1122
1123 static void carl9170_tx_ampdu(struct ar9170 *ar)
1124 {
1125         struct sk_buff_head agg;
1126         struct carl9170_sta_tid *tid_info;
1127         struct sk_buff *skb, *first;
1128         unsigned int i = 0, done_ampdus = 0;
1129         u16 seq, queue, tmpssn;
1130
1131         atomic_inc(&ar->tx_ampdu_scheduler);
1132         ar->tx_ampdu_schedule = false;
1133
1134         if (atomic_read(&ar->tx_ampdu_upload))
1135                 return;
1136
1137         if (!ar->tx_ampdu_list_len)
1138                 return;
1139
1140         __skb_queue_head_init(&agg);
1141
1142         rcu_read_lock();
1143         tid_info = rcu_dereference(ar->tx_ampdu_iter);
1144         if (WARN_ON_ONCE(!tid_info)) {
1145                 rcu_read_unlock();
1146                 return;
1147         }
1148
1149 retry:
1150         list_for_each_entry_continue_rcu(tid_info, &ar->tx_ampdu_list, list) {
1151                 i++;
1152
1153                 if (tid_info->state < CARL9170_TID_STATE_PROGRESS)
1154                         continue;
1155
1156                 queue = TID_TO_WME_AC(tid_info->tid);
1157
1158                 spin_lock_bh(&tid_info->lock);
1159                 if (tid_info->state != CARL9170_TID_STATE_XMIT)
1160                         goto processed;
1161
1162                 tid_info->counter++;
1163                 first = skb_peek(&tid_info->queue);
1164                 tmpssn = carl9170_get_seq(first);
1165                 seq = tid_info->snx;
1166
1167                 if (unlikely(tmpssn != seq)) {
1168                         tid_info->state = CARL9170_TID_STATE_IDLE;
1169
1170                         goto processed;
1171                 }
1172
1173                 while ((skb = skb_peek(&tid_info->queue))) {
1174                         /* strict 0, 1, ..., n - 1, n frame sequence order */
1175                         if (unlikely(carl9170_get_seq(skb) != seq))
1176                                 break;
1177
1178                         /* don't upload more than AMPDU FACTOR allows. */
1179                         if (unlikely(SEQ_DIFF(tid_info->snx, tid_info->bsn) >=
1180                             (tid_info->max - 1)))
1181                                 break;
1182
1183                         if (!carl9170_tx_rate_check(ar, skb, first))
1184                                 break;
1185
1186                         atomic_inc(&ar->tx_ampdu_upload);
1187                         tid_info->snx = seq = SEQ_NEXT(seq);
1188                         __skb_unlink(skb, &tid_info->queue);
1189
1190                         __skb_queue_tail(&agg, skb);
1191
1192                         if (skb_queue_len(&agg) >= CARL9170_NUM_TX_AGG_MAX)
1193                                 break;
1194                 }
1195
1196                 if (skb_queue_empty(&tid_info->queue) ||
1197                     carl9170_get_seq(skb_peek(&tid_info->queue)) !=
1198                     tid_info->snx) {
1199                         /*
1200                          * stop TID, if A-MPDU frames are still missing,
1201                          * or whenever the queue is empty.
1202                          */
1203
1204                         tid_info->state = CARL9170_TID_STATE_IDLE;
1205                 }
1206                 done_ampdus++;
1207
1208 processed:
1209                 spin_unlock_bh(&tid_info->lock);
1210
1211                 if (skb_queue_empty(&agg))
1212                         continue;
1213
1214                 /* apply ampdu spacing & factor settings */
1215                 carl9170_set_ampdu_params(ar, skb_peek(&agg));
1216
1217                 /* set aggregation push bit */
1218                 carl9170_set_immba(ar, skb_peek_tail(&agg));
1219
1220                 spin_lock_bh(&ar->tx_pending[queue].lock);
1221                 skb_queue_splice_tail_init(&agg, &ar->tx_pending[queue]);
1222                 spin_unlock_bh(&ar->tx_pending[queue].lock);
1223                 ar->tx_schedule = true;
1224         }
1225         if ((done_ampdus++ == 0) && (i++ == 0))
1226                 goto retry;
1227
1228         rcu_assign_pointer(ar->tx_ampdu_iter, tid_info);
1229         rcu_read_unlock();
1230 }
1231
1232 static struct sk_buff *carl9170_tx_pick_skb(struct ar9170 *ar,
1233                                             struct sk_buff_head *queue)
1234 {
1235         struct sk_buff *skb;
1236         struct ieee80211_tx_info *info;
1237         struct carl9170_tx_info *arinfo;
1238
1239         BUILD_BUG_ON(sizeof(*arinfo) > sizeof(info->rate_driver_data));
1240
1241         spin_lock_bh(&queue->lock);
1242         skb = skb_peek(queue);
1243         if (unlikely(!skb))
1244                 goto err_unlock;
1245
1246         if (carl9170_alloc_dev_space(ar, skb))
1247                 goto err_unlock;
1248
1249         __skb_unlink(skb, queue);
1250         spin_unlock_bh(&queue->lock);
1251
1252         info = IEEE80211_SKB_CB(skb);
1253         arinfo = (void *) info->rate_driver_data;
1254
1255         arinfo->timeout = jiffies;
1256         return skb;
1257
1258 err_unlock:
1259         spin_unlock_bh(&queue->lock);
1260         return NULL;
1261 }
1262
1263 void carl9170_tx_drop(struct ar9170 *ar, struct sk_buff *skb)
1264 {
1265         struct _carl9170_tx_superframe *super;
1266         uint8_t q = 0;
1267
1268         ar->tx_dropped++;
1269
1270         super = (void *)skb->data;
1271         SET_VAL(CARL9170_TX_SUPER_MISC_QUEUE, q,
1272                 ar9170_qmap[carl9170_get_queue(ar, skb)]);
1273         __carl9170_tx_process_status(ar, super->s.cookie, q);
1274 }
1275
1276 static bool carl9170_tx_ps_drop(struct ar9170 *ar, struct sk_buff *skb)
1277 {
1278         struct ieee80211_sta *sta;
1279         struct carl9170_sta_info *sta_info;
1280         struct ieee80211_tx_info *tx_info;
1281
1282         rcu_read_lock();
1283         sta = __carl9170_get_tx_sta(ar, skb);
1284         if (!sta)
1285                 goto out_rcu;
1286
1287         sta_info = (void *) sta->drv_priv;
1288         tx_info = IEEE80211_SKB_CB(skb);
1289
1290         if (unlikely(sta_info->sleeping) &&
1291             !(tx_info->flags & (IEEE80211_TX_CTL_NO_PS_BUFFER |
1292                                 IEEE80211_TX_CTL_CLEAR_PS_FILT))) {
1293                 rcu_read_unlock();
1294
1295                 if (tx_info->flags & IEEE80211_TX_CTL_AMPDU)
1296                         atomic_dec(&ar->tx_ampdu_upload);
1297
1298                 tx_info->flags |= IEEE80211_TX_STAT_TX_FILTERED;
1299                 carl9170_release_dev_space(ar, skb);
1300                 carl9170_tx_status(ar, skb, false);
1301                 return true;
1302         }
1303
1304 out_rcu:
1305         rcu_read_unlock();
1306         return false;
1307 }
1308
1309 static void carl9170_bar_check(struct ar9170 *ar, struct sk_buff *skb)
1310 {
1311         struct _carl9170_tx_superframe *super = (void *) skb->data;
1312         struct ieee80211_bar *bar = (void *) super->frame_data;
1313
1314         if (unlikely(ieee80211_is_back_req(bar->frame_control)) &&
1315             skb->len >= sizeof(struct ieee80211_bar)) {
1316                 struct carl9170_bar_list_entry *entry;
1317                 unsigned int queue = skb_get_queue_mapping(skb);
1318
1319                 entry = kmalloc(sizeof(*entry), GFP_ATOMIC);
1320                 if (!WARN_ON_ONCE(!entry)) {
1321                         entry->skb = skb;
1322                         spin_lock_bh(&ar->bar_list_lock[queue]);
1323                         list_add_tail_rcu(&entry->list, &ar->bar_list[queue]);
1324                         spin_unlock_bh(&ar->bar_list_lock[queue]);
1325                 }
1326         }
1327 }
1328
1329 static void carl9170_tx(struct ar9170 *ar)
1330 {
1331         struct sk_buff *skb;
1332         unsigned int i, q;
1333         bool schedule_garbagecollector = false;
1334
1335         ar->tx_schedule = false;
1336
1337         if (unlikely(!IS_STARTED(ar)))
1338                 return;
1339
1340         carl9170_usb_handle_tx_err(ar);
1341
1342         for (i = 0; i < ar->hw->queues; i++) {
1343                 while (!skb_queue_empty(&ar->tx_pending[i])) {
1344                         skb = carl9170_tx_pick_skb(ar, &ar->tx_pending[i]);
1345                         if (unlikely(!skb))
1346                                 break;
1347
1348                         if (unlikely(carl9170_tx_ps_drop(ar, skb)))
1349                                 continue;
1350
1351                         carl9170_bar_check(ar, skb);
1352
1353                         atomic_inc(&ar->tx_total_pending);
1354
1355                         q = __carl9170_get_queue(ar, i);
1356                         /*
1357                          * NB: tx_status[i] vs. tx_status[q],
1358                          * TODO: Move into pick_skb or alloc_dev_space.
1359                          */
1360                         skb_queue_tail(&ar->tx_status[q], skb);
1361
1362                         /*
1363                          * increase ref count to "2".
1364                          * Ref counting is the easiest way to solve the
1365                          * race between the urb's completion routine:
1366                          *      carl9170_tx_callback
1367                          * and wlan tx status functions:
1368                          *      carl9170_tx_status/janitor.
1369                          */
1370                         carl9170_tx_get_skb(skb);
1371
1372                         carl9170_usb_tx(ar, skb);
1373                         schedule_garbagecollector = true;
1374                 }
1375         }
1376
1377         if (!schedule_garbagecollector)
1378                 return;
1379
1380         ieee80211_queue_delayed_work(ar->hw, &ar->tx_janitor,
1381                 msecs_to_jiffies(CARL9170_TX_TIMEOUT));
1382 }
1383
1384 static bool carl9170_tx_ampdu_queue(struct ar9170 *ar,
1385         struct ieee80211_sta *sta, struct sk_buff *skb)
1386 {
1387         struct _carl9170_tx_superframe *super = (void *) skb->data;
1388         struct carl9170_sta_info *sta_info;
1389         struct carl9170_sta_tid *agg;
1390         struct sk_buff *iter;
1391         u16 tid, seq, qseq, off;
1392         bool run = false;
1393
1394         tid = carl9170_get_tid(skb);
1395         seq = carl9170_get_seq(skb);
1396         sta_info = (void *) sta->drv_priv;
1397
1398         rcu_read_lock();
1399         agg = rcu_dereference(sta_info->agg[tid]);
1400
1401         if (!agg)
1402                 goto err_unlock_rcu;
1403
1404         spin_lock_bh(&agg->lock);
1405         if (unlikely(agg->state < CARL9170_TID_STATE_IDLE))
1406                 goto err_unlock;
1407
1408         /* check if sequence is within the BA window */
1409         if (unlikely(!BAW_WITHIN(agg->bsn, CARL9170_BAW_BITS, seq)))
1410                 goto err_unlock;
1411
1412         if (WARN_ON_ONCE(!BAW_WITHIN(agg->snx, CARL9170_BAW_BITS, seq)))
1413                 goto err_unlock;
1414
1415         off = SEQ_DIFF(seq, agg->bsn);
1416         if (WARN_ON_ONCE(test_and_set_bit(off, agg->bitmap)))
1417                 goto err_unlock;
1418
1419         if (likely(BAW_WITHIN(agg->hsn, CARL9170_BAW_BITS, seq))) {
1420                 __skb_queue_tail(&agg->queue, skb);
1421                 agg->hsn = seq;
1422                 goto queued;
1423         }
1424
1425         skb_queue_reverse_walk(&agg->queue, iter) {
1426                 qseq = carl9170_get_seq(iter);
1427
1428                 if (BAW_WITHIN(qseq, CARL9170_BAW_BITS, seq)) {
1429                         __skb_queue_after(&agg->queue, iter, skb);
1430                         goto queued;
1431                 }
1432         }
1433
1434         __skb_queue_head(&agg->queue, skb);
1435 queued:
1436
1437         if (unlikely(agg->state != CARL9170_TID_STATE_XMIT)) {
1438                 if (agg->snx == carl9170_get_seq(skb_peek(&agg->queue))) {
1439                         agg->state = CARL9170_TID_STATE_XMIT;
1440                         run = true;
1441                 }
1442         }
1443
1444         spin_unlock_bh(&agg->lock);
1445         rcu_read_unlock();
1446
1447         return run;
1448
1449 err_unlock:
1450         spin_unlock_bh(&agg->lock);
1451
1452 err_unlock_rcu:
1453         rcu_read_unlock();
1454         super->f.mac_control &= ~cpu_to_le16(AR9170_TX_MAC_AGGR);
1455         carl9170_tx_status(ar, skb, false);
1456         ar->tx_dropped++;
1457         return false;
1458 }
1459
1460 void carl9170_op_tx(struct ieee80211_hw *hw, struct sk_buff *skb)
1461 {
1462         struct ar9170 *ar = hw->priv;
1463         struct ieee80211_tx_info *info;
1464         struct ieee80211_sta *sta;
1465         bool run;
1466
1467         if (unlikely(!IS_STARTED(ar)))
1468                 goto err_free;
1469
1470         info = IEEE80211_SKB_CB(skb);
1471         sta = info->control.sta;
1472
1473         if (unlikely(carl9170_tx_prepare(ar, skb)))
1474                 goto err_free;
1475
1476         carl9170_tx_accounting(ar, skb);
1477         /*
1478          * from now on, one has to use carl9170_tx_status to free
1479          * all ressouces which are associated with the frame.
1480          */
1481
1482         if (sta) {
1483                 struct carl9170_sta_info *stai = (void *) sta->drv_priv;
1484                 atomic_inc(&stai->pending_frames);
1485         }
1486
1487         if (info->flags & IEEE80211_TX_CTL_AMPDU) {
1488                 run = carl9170_tx_ampdu_queue(ar, sta, skb);
1489                 if (run)
1490                         carl9170_tx_ampdu(ar);
1491
1492         } else {
1493                 unsigned int queue = skb_get_queue_mapping(skb);
1494
1495                 skb_queue_tail(&ar->tx_pending[queue], skb);
1496         }
1497
1498         carl9170_tx(ar);
1499         return;
1500
1501 err_free:
1502         ar->tx_dropped++;
1503         ieee80211_free_txskb(ar->hw, skb);
1504 }
1505
1506 void carl9170_tx_scheduler(struct ar9170 *ar)
1507 {
1508
1509         if (ar->tx_ampdu_schedule)
1510                 carl9170_tx_ampdu(ar);
1511
1512         if (ar->tx_schedule)
1513                 carl9170_tx(ar);
1514 }
1515
1516 int carl9170_update_beacon(struct ar9170 *ar, const bool submit)
1517 {
1518         struct sk_buff *skb = NULL;
1519         struct carl9170_vif_info *cvif;
1520         struct ieee80211_tx_info *txinfo;
1521         struct ieee80211_tx_rate *rate;
1522         __le32 *data, *old = NULL;
1523         unsigned int plcp, power, chains;
1524         u32 word, ht1, off, addr, len;
1525         int i = 0, err = 0;
1526
1527         rcu_read_lock();
1528         cvif = rcu_dereference(ar->beacon_iter);
1529 retry:
1530         if (ar->vifs == 0 || !cvif)
1531                 goto out_unlock;
1532
1533         list_for_each_entry_continue_rcu(cvif, &ar->vif_list, list) {
1534                 if (cvif->active && cvif->enable_beacon)
1535                         goto found;
1536         }
1537
1538         if (!ar->beacon_enabled || i++)
1539                 goto out_unlock;
1540
1541         goto retry;
1542
1543 found:
1544         rcu_assign_pointer(ar->beacon_iter, cvif);
1545
1546         skb = ieee80211_beacon_get_tim(ar->hw, carl9170_get_vif(cvif),
1547                 NULL, NULL);
1548
1549         if (!skb) {
1550                 err = -ENOMEM;
1551                 goto err_free;
1552         }
1553
1554         txinfo = IEEE80211_SKB_CB(skb);
1555         spin_lock_bh(&ar->beacon_lock);
1556         data = (__le32 *)skb->data;
1557         if (cvif->beacon)
1558                 old = (__le32 *)cvif->beacon->data;
1559
1560         off = cvif->id * AR9170_MAC_BCN_LENGTH_MAX;
1561         addr = ar->fw.beacon_addr + off;
1562         len = roundup(skb->len + FCS_LEN, 4);
1563
1564         if ((off + len) > ar->fw.beacon_max_len) {
1565                 if (net_ratelimit()) {
1566                         wiphy_err(ar->hw->wiphy, "beacon does not "
1567                                   "fit into device memory!\n");
1568                 }
1569                 err = -EINVAL;
1570                 goto err_unlock;
1571         }
1572
1573         if (len > AR9170_MAC_BCN_LENGTH_MAX) {
1574                 if (net_ratelimit()) {
1575                         wiphy_err(ar->hw->wiphy, "no support for beacons "
1576                                 "bigger than %d (yours:%d).\n",
1577                                  AR9170_MAC_BCN_LENGTH_MAX, len);
1578                 }
1579
1580                 err = -EMSGSIZE;
1581                 goto err_unlock;
1582         }
1583
1584         ht1 = AR9170_MAC_BCN_HT1_TX_ANT0;
1585         rate = &txinfo->control.rates[0];
1586         carl9170_tx_rate_tpc_chains(ar, txinfo, rate, &plcp, &power, &chains);
1587         if (!(txinfo->control.rates[0].flags & IEEE80211_TX_RC_MCS)) {
1588                 if (plcp <= AR9170_TX_PHY_RATE_CCK_11M)
1589                         plcp |= ((skb->len + FCS_LEN) << (3 + 16)) + 0x0400;
1590                 else
1591                         plcp |= ((skb->len + FCS_LEN) << 16) + 0x0010;
1592         } else {
1593                 ht1 |= AR9170_MAC_BCN_HT1_HT_EN;
1594                 if (rate->flags & IEEE80211_TX_RC_SHORT_GI)
1595                         plcp |= AR9170_MAC_BCN_HT2_SGI;
1596
1597                 if (rate->flags & IEEE80211_TX_RC_40_MHZ_WIDTH) {
1598                         ht1 |= AR9170_MAC_BCN_HT1_BWC_40M_SHARED;
1599                         plcp |= AR9170_MAC_BCN_HT2_BW40;
1600                 }
1601                 if (rate->flags & IEEE80211_TX_RC_DUP_DATA) {
1602                         ht1 |= AR9170_MAC_BCN_HT1_BWC_40M_DUP;
1603                         plcp |= AR9170_MAC_BCN_HT2_BW40;
1604                 }
1605
1606                 SET_VAL(AR9170_MAC_BCN_HT2_LEN, plcp, skb->len + FCS_LEN);
1607         }
1608
1609         SET_VAL(AR9170_MAC_BCN_HT1_PWR_CTRL, ht1, 7);
1610         SET_VAL(AR9170_MAC_BCN_HT1_TPC, ht1, power);
1611         SET_VAL(AR9170_MAC_BCN_HT1_CHAIN_MASK, ht1, chains);
1612         if (chains == AR9170_TX_PHY_TXCHAIN_2)
1613                 ht1 |= AR9170_MAC_BCN_HT1_TX_ANT1;
1614
1615         carl9170_async_regwrite_begin(ar);
1616         carl9170_async_regwrite(AR9170_MAC_REG_BCN_HT1, ht1);
1617         if (!(txinfo->control.rates[0].flags & IEEE80211_TX_RC_MCS))
1618                 carl9170_async_regwrite(AR9170_MAC_REG_BCN_PLCP, plcp);
1619         else
1620                 carl9170_async_regwrite(AR9170_MAC_REG_BCN_HT2, plcp);
1621
1622         for (i = 0; i < DIV_ROUND_UP(skb->len, 4); i++) {
1623                 /*
1624                  * XXX: This accesses beyond skb data for up
1625                  *      to the last 3 bytes!!
1626                  */
1627
1628                 if (old && (data[i] == old[i]))
1629                         continue;
1630
1631                 word = le32_to_cpu(data[i]);
1632                 carl9170_async_regwrite(addr + 4 * i, word);
1633         }
1634         carl9170_async_regwrite_finish();
1635
1636         dev_kfree_skb_any(cvif->beacon);
1637         cvif->beacon = NULL;
1638
1639         err = carl9170_async_regwrite_result();
1640         if (!err)
1641                 cvif->beacon = skb;
1642         spin_unlock_bh(&ar->beacon_lock);
1643         if (err)
1644                 goto err_free;
1645
1646         if (submit) {
1647                 err = carl9170_bcn_ctrl(ar, cvif->id,
1648                                         CARL9170_BCN_CTRL_CAB_TRIGGER,
1649                                         addr, skb->len + FCS_LEN);
1650
1651                 if (err)
1652                         goto err_free;
1653         }
1654 out_unlock:
1655         rcu_read_unlock();
1656         return 0;
1657
1658 err_unlock:
1659         spin_unlock_bh(&ar->beacon_lock);
1660
1661 err_free:
1662         rcu_read_unlock();
1663         dev_kfree_skb_any(skb);
1664         return err;
1665 }