carl9170: Fix tx aggregation problems with some clients
[linux-3.10.git] / drivers / net / wireless / ath / carl9170 / tx.c
1 /*
2  * Atheros CARL9170 driver
3  *
4  * 802.11 xmit & status routines
5  *
6  * Copyright 2008, Johannes Berg <johannes@sipsolutions.net>
7  * Copyright 2009, 2010, Christian Lamparter <chunkeey@googlemail.com>
8  *
9  * This program is free software; you can redistribute it and/or modify
10  * it under the terms of the GNU General Public License as published by
11  * the Free Software Foundation; either version 2 of the License, or
12  * (at your option) any later version.
13  *
14  * This program is distributed in the hope that it will be useful,
15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
17  * GNU General Public License for more details.
18  *
19  * You should have received a copy of the GNU General Public License
20  * along with this program; see the file COPYING.  If not, see
21  * http://www.gnu.org/licenses/.
22  *
23  * This file incorporates work covered by the following copyright and
24  * permission notice:
25  *    Copyright (c) 2007-2008 Atheros Communications, Inc.
26  *
27  *    Permission to use, copy, modify, and/or distribute this software for any
28  *    purpose with or without fee is hereby granted, provided that the above
29  *    copyright notice and this permission notice appear in all copies.
30  *
31  *    THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
32  *    WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
33  *    MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
34  *    ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
35  *    WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
36  *    ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
37  *    OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
38  */
39
40 #include <linux/init.h>
41 #include <linux/slab.h>
42 #include <linux/module.h>
43 #include <linux/etherdevice.h>
44 #include <net/mac80211.h>
45 #include "carl9170.h"
46 #include "hw.h"
47 #include "cmd.h"
48
49 static inline unsigned int __carl9170_get_queue(struct ar9170 *ar,
50                                                 unsigned int queue)
51 {
52         if (unlikely(modparam_noht)) {
53                 return queue;
54         } else {
55                 /*
56                  * This is just another workaround, until
57                  * someone figures out how to get QoS and
58                  * AMPDU to play nicely together.
59                  */
60
61                 return 2;               /* AC_BE */
62         }
63 }
64
65 static inline unsigned int carl9170_get_queue(struct ar9170 *ar,
66                                               struct sk_buff *skb)
67 {
68         return __carl9170_get_queue(ar, skb_get_queue_mapping(skb));
69 }
70
71 static bool is_mem_full(struct ar9170 *ar)
72 {
73         return (DIV_ROUND_UP(IEEE80211_MAX_FRAME_LEN, ar->fw.mem_block_size) >
74                 atomic_read(&ar->mem_free_blocks));
75 }
76
77 static void carl9170_tx_accounting(struct ar9170 *ar, struct sk_buff *skb)
78 {
79         int queue, i;
80         bool mem_full;
81
82         atomic_inc(&ar->tx_total_queued);
83
84         queue = skb_get_queue_mapping(skb);
85         spin_lock_bh(&ar->tx_stats_lock);
86
87         /*
88          * The driver has to accept the frame, regardless if the queue is
89          * full to the brim, or not. We have to do the queuing internally,
90          * since mac80211 assumes that a driver which can operate with
91          * aggregated frames does not reject frames for this reason.
92          */
93         ar->tx_stats[queue].len++;
94         ar->tx_stats[queue].count++;
95
96         mem_full = is_mem_full(ar);
97         for (i = 0; i < ar->hw->queues; i++) {
98                 if (mem_full || ar->tx_stats[i].len >= ar->tx_stats[i].limit) {
99                         ieee80211_stop_queue(ar->hw, i);
100                         ar->queue_stop_timeout[i] = jiffies;
101                 }
102         }
103
104         spin_unlock_bh(&ar->tx_stats_lock);
105 }
106
107 static void carl9170_tx_accounting_free(struct ar9170 *ar, struct sk_buff *skb)
108 {
109         struct ieee80211_tx_info *txinfo;
110         int queue;
111
112         txinfo = IEEE80211_SKB_CB(skb);
113         queue = skb_get_queue_mapping(skb);
114
115         spin_lock_bh(&ar->tx_stats_lock);
116
117         ar->tx_stats[queue].len--;
118
119         if (!is_mem_full(ar)) {
120                 unsigned int i;
121                 for (i = 0; i < ar->hw->queues; i++) {
122                         if (ar->tx_stats[i].len >= CARL9170_NUM_TX_LIMIT_SOFT)
123                                 continue;
124
125                         if (ieee80211_queue_stopped(ar->hw, i)) {
126                                 unsigned long tmp;
127
128                                 tmp = jiffies - ar->queue_stop_timeout[i];
129                                 if (tmp > ar->max_queue_stop_timeout[i])
130                                         ar->max_queue_stop_timeout[i] = tmp;
131                         }
132
133                         ieee80211_wake_queue(ar->hw, i);
134                 }
135         }
136
137         spin_unlock_bh(&ar->tx_stats_lock);
138         if (atomic_dec_and_test(&ar->tx_total_queued))
139                 complete(&ar->tx_flush);
140 }
141
142 static int carl9170_alloc_dev_space(struct ar9170 *ar, struct sk_buff *skb)
143 {
144         struct _carl9170_tx_superframe *super = (void *) skb->data;
145         unsigned int chunks;
146         int cookie = -1;
147
148         atomic_inc(&ar->mem_allocs);
149
150         chunks = DIV_ROUND_UP(skb->len, ar->fw.mem_block_size);
151         if (unlikely(atomic_sub_return(chunks, &ar->mem_free_blocks) < 0)) {
152                 atomic_add(chunks, &ar->mem_free_blocks);
153                 return -ENOSPC;
154         }
155
156         spin_lock_bh(&ar->mem_lock);
157         cookie = bitmap_find_free_region(ar->mem_bitmap, ar->fw.mem_blocks, 0);
158         spin_unlock_bh(&ar->mem_lock);
159
160         if (unlikely(cookie < 0)) {
161                 atomic_add(chunks, &ar->mem_free_blocks);
162                 return -ENOSPC;
163         }
164
165         super = (void *) skb->data;
166
167         /*
168          * Cookie #0 serves two special purposes:
169          *  1. The firmware might use it generate BlockACK frames
170          *     in responds of an incoming BlockAckReqs.
171          *
172          *  2. Prevent double-free bugs.
173          */
174         super->s.cookie = (u8) cookie + 1;
175         return 0;
176 }
177
178 static void carl9170_release_dev_space(struct ar9170 *ar, struct sk_buff *skb)
179 {
180         struct _carl9170_tx_superframe *super = (void *) skb->data;
181         int cookie;
182
183         /* make a local copy of the cookie */
184         cookie = super->s.cookie;
185         /* invalidate cookie */
186         super->s.cookie = 0;
187
188         /*
189          * Do a out-of-bounds check on the cookie:
190          *
191          *  * cookie "0" is reserved and won't be assigned to any
192          *    out-going frame. Internally however, it is used to
193          *    mark no longer/un-accounted frames and serves as a
194          *    cheap way of preventing frames from being freed
195          *    twice by _accident_. NB: There is a tiny race...
196          *
197          *  * obviously, cookie number is limited by the amount
198          *    of available memory blocks, so the number can
199          *    never execeed the mem_blocks count.
200          */
201         if (unlikely(WARN_ON_ONCE(cookie == 0) ||
202             WARN_ON_ONCE(cookie > ar->fw.mem_blocks)))
203                 return;
204
205         atomic_add(DIV_ROUND_UP(skb->len, ar->fw.mem_block_size),
206                    &ar->mem_free_blocks);
207
208         spin_lock_bh(&ar->mem_lock);
209         bitmap_release_region(ar->mem_bitmap, cookie - 1, 0);
210         spin_unlock_bh(&ar->mem_lock);
211 }
212
213 /* Called from any context */
214 static void carl9170_tx_release(struct kref *ref)
215 {
216         struct ar9170 *ar;
217         struct carl9170_tx_info *arinfo;
218         struct ieee80211_tx_info *txinfo;
219         struct sk_buff *skb;
220
221         arinfo = container_of(ref, struct carl9170_tx_info, ref);
222         txinfo = container_of((void *) arinfo, struct ieee80211_tx_info,
223                               rate_driver_data);
224         skb = container_of((void *) txinfo, struct sk_buff, cb);
225
226         ar = arinfo->ar;
227         if (WARN_ON_ONCE(!ar))
228                 return;
229
230         BUILD_BUG_ON(
231             offsetof(struct ieee80211_tx_info, status.ampdu_ack_len) != 23);
232
233         memset(&txinfo->status.ampdu_ack_len, 0,
234                sizeof(struct ieee80211_tx_info) -
235                offsetof(struct ieee80211_tx_info, status.ampdu_ack_len));
236
237         if (atomic_read(&ar->tx_total_queued))
238                 ar->tx_schedule = true;
239
240         if (txinfo->flags & IEEE80211_TX_CTL_AMPDU) {
241                 if (!atomic_read(&ar->tx_ampdu_upload))
242                         ar->tx_ampdu_schedule = true;
243
244                 if (txinfo->flags & IEEE80211_TX_STAT_AMPDU) {
245                         struct _carl9170_tx_superframe *super;
246
247                         super = (void *)skb->data;
248                         txinfo->status.ampdu_len = super->s.rix;
249                         txinfo->status.ampdu_ack_len = super->s.cnt;
250                 } else if (txinfo->flags & IEEE80211_TX_STAT_ACK) {
251                         /*
252                          * drop redundant tx_status reports:
253                          *
254                          * 1. ampdu_ack_len of the final tx_status does
255                          *    include the feedback of this particular frame.
256                          *
257                          * 2. tx_status_irqsafe only queues up to 128
258                          *    tx feedback reports and discards the rest.
259                          *
260                          * 3. minstrel_ht is picky, it only accepts
261                          *    reports of frames with the TX_STATUS_AMPDU flag.
262                          */
263
264                         dev_kfree_skb_any(skb);
265                         return;
266                 } else {
267                         /*
268                          * Frame has failed, but we want to keep it in
269                          * case it was lost due to a power-state
270                          * transition.
271                          */
272                 }
273         }
274
275         skb_pull(skb, sizeof(struct _carl9170_tx_superframe));
276         ieee80211_tx_status_irqsafe(ar->hw, skb);
277 }
278
279 void carl9170_tx_get_skb(struct sk_buff *skb)
280 {
281         struct carl9170_tx_info *arinfo = (void *)
282                 (IEEE80211_SKB_CB(skb))->rate_driver_data;
283         kref_get(&arinfo->ref);
284 }
285
286 int carl9170_tx_put_skb(struct sk_buff *skb)
287 {
288         struct carl9170_tx_info *arinfo = (void *)
289                 (IEEE80211_SKB_CB(skb))->rate_driver_data;
290
291         return kref_put(&arinfo->ref, carl9170_tx_release);
292 }
293
294 /* Caller must hold the tid_info->lock & rcu_read_lock */
295 static void carl9170_tx_shift_bm(struct ar9170 *ar,
296         struct carl9170_sta_tid *tid_info, u16 seq)
297 {
298         u16 off;
299
300         off = SEQ_DIFF(seq, tid_info->bsn);
301
302         if (WARN_ON_ONCE(off >= CARL9170_BAW_BITS))
303                 return;
304
305         /*
306          * Sanity check. For each MPDU we set the bit in bitmap and
307          * clear it once we received the tx_status.
308          * But if the bit is already cleared then we've been bitten
309          * by a bug.
310          */
311         WARN_ON_ONCE(!test_and_clear_bit(off, tid_info->bitmap));
312
313         off = SEQ_DIFF(tid_info->snx, tid_info->bsn);
314         if (WARN_ON_ONCE(off >= CARL9170_BAW_BITS))
315                 return;
316
317         if (!bitmap_empty(tid_info->bitmap, off))
318                 off = find_first_bit(tid_info->bitmap, off);
319
320         tid_info->bsn += off;
321         tid_info->bsn &= 0x0fff;
322
323         bitmap_shift_right(tid_info->bitmap, tid_info->bitmap,
324                            off, CARL9170_BAW_BITS);
325 }
326
327 static void carl9170_tx_status_process_ampdu(struct ar9170 *ar,
328         struct sk_buff *skb, struct ieee80211_tx_info *txinfo)
329 {
330         struct _carl9170_tx_superframe *super = (void *) skb->data;
331         struct ieee80211_hdr *hdr = (void *) super->frame_data;
332         struct ieee80211_tx_info *tx_info;
333         struct carl9170_tx_info *ar_info;
334         struct carl9170_sta_info *sta_info;
335         struct ieee80211_sta *sta;
336         struct carl9170_sta_tid *tid_info;
337         struct ieee80211_vif *vif;
338         unsigned int vif_id;
339         u8 tid;
340
341         if (!(txinfo->flags & IEEE80211_TX_CTL_AMPDU) ||
342             txinfo->flags & IEEE80211_TX_CTL_INJECTED ||
343            (!(super->f.mac_control & cpu_to_le16(AR9170_TX_MAC_AGGR))))
344                 return;
345
346         tx_info = IEEE80211_SKB_CB(skb);
347         ar_info = (void *) tx_info->rate_driver_data;
348
349         vif_id = (super->s.misc & CARL9170_TX_SUPER_MISC_VIF_ID) >>
350                  CARL9170_TX_SUPER_MISC_VIF_ID_S;
351
352         if (WARN_ON_ONCE(vif_id >= AR9170_MAX_VIRTUAL_MAC))
353                 return;
354
355         rcu_read_lock();
356         vif = rcu_dereference(ar->vif_priv[vif_id].vif);
357         if (unlikely(!vif))
358                 goto out_rcu;
359
360         /*
361          * Normally we should use wrappers like ieee80211_get_DA to get
362          * the correct peer ieee80211_sta.
363          *
364          * But there is a problem with indirect traffic (broadcasts, or
365          * data which is designated for other stations) in station mode.
366          * The frame will be directed to the AP for distribution and not
367          * to the actual destination.
368          */
369         sta = ieee80211_find_sta(vif, hdr->addr1);
370         if (unlikely(!sta))
371                 goto out_rcu;
372
373         tid = get_tid_h(hdr);
374
375         sta_info = (void *) sta->drv_priv;
376         tid_info = rcu_dereference(sta_info->agg[tid]);
377         if (!tid_info)
378                 goto out_rcu;
379
380         spin_lock_bh(&tid_info->lock);
381         if (likely(tid_info->state >= CARL9170_TID_STATE_IDLE))
382                 carl9170_tx_shift_bm(ar, tid_info, get_seq_h(hdr));
383
384         if (sta_info->stats[tid].clear) {
385                 sta_info->stats[tid].clear = false;
386                 sta_info->stats[tid].req = false;
387                 sta_info->stats[tid].ampdu_len = 0;
388                 sta_info->stats[tid].ampdu_ack_len = 0;
389         }
390
391         sta_info->stats[tid].ampdu_len++;
392         if (txinfo->status.rates[0].count == 1)
393                 sta_info->stats[tid].ampdu_ack_len++;
394
395         if (!(txinfo->flags & IEEE80211_TX_STAT_ACK))
396                 sta_info->stats[tid].req = true;
397
398         if (super->f.mac_control & cpu_to_le16(AR9170_TX_MAC_IMM_BA)) {
399                 super->s.rix = sta_info->stats[tid].ampdu_len;
400                 super->s.cnt = sta_info->stats[tid].ampdu_ack_len;
401                 txinfo->flags |= IEEE80211_TX_STAT_AMPDU;
402                 if (sta_info->stats[tid].req)
403                         txinfo->flags |= IEEE80211_TX_STAT_AMPDU_NO_BACK;
404
405                 sta_info->stats[tid].clear = true;
406         }
407         spin_unlock_bh(&tid_info->lock);
408
409 out_rcu:
410         rcu_read_unlock();
411 }
412
413 void carl9170_tx_status(struct ar9170 *ar, struct sk_buff *skb,
414                         const bool success)
415 {
416         struct ieee80211_tx_info *txinfo;
417
418         carl9170_tx_accounting_free(ar, skb);
419
420         txinfo = IEEE80211_SKB_CB(skb);
421
422         if (success)
423                 txinfo->flags |= IEEE80211_TX_STAT_ACK;
424         else
425                 ar->tx_ack_failures++;
426
427         if (txinfo->flags & IEEE80211_TX_CTL_AMPDU)
428                 carl9170_tx_status_process_ampdu(ar, skb, txinfo);
429
430         carl9170_tx_put_skb(skb);
431 }
432
433 /* This function may be called form any context */
434 void carl9170_tx_callback(struct ar9170 *ar, struct sk_buff *skb)
435 {
436         struct ieee80211_tx_info *txinfo = IEEE80211_SKB_CB(skb);
437
438         atomic_dec(&ar->tx_total_pending);
439
440         if (txinfo->flags & IEEE80211_TX_CTL_AMPDU)
441                 atomic_dec(&ar->tx_ampdu_upload);
442
443         if (carl9170_tx_put_skb(skb))
444                 tasklet_hi_schedule(&ar->usb_tasklet);
445 }
446
447 static struct sk_buff *carl9170_get_queued_skb(struct ar9170 *ar, u8 cookie,
448                                                struct sk_buff_head *queue)
449 {
450         struct sk_buff *skb;
451
452         spin_lock_bh(&queue->lock);
453         skb_queue_walk(queue, skb) {
454                 struct _carl9170_tx_superframe *txc = (void *) skb->data;
455
456                 if (txc->s.cookie != cookie)
457                         continue;
458
459                 __skb_unlink(skb, queue);
460                 spin_unlock_bh(&queue->lock);
461
462                 carl9170_release_dev_space(ar, skb);
463                 return skb;
464         }
465         spin_unlock_bh(&queue->lock);
466
467         return NULL;
468 }
469
470 static void carl9170_tx_fill_rateinfo(struct ar9170 *ar, unsigned int rix,
471         unsigned int tries, struct ieee80211_tx_info *txinfo)
472 {
473         unsigned int i;
474
475         for (i = 0; i < IEEE80211_TX_MAX_RATES; i++) {
476                 if (txinfo->status.rates[i].idx < 0)
477                         break;
478
479                 if (i == rix) {
480                         txinfo->status.rates[i].count = tries;
481                         i++;
482                         break;
483                 }
484         }
485
486         for (; i < IEEE80211_TX_MAX_RATES; i++) {
487                 txinfo->status.rates[i].idx = -1;
488                 txinfo->status.rates[i].count = 0;
489         }
490 }
491
492 static void carl9170_check_queue_stop_timeout(struct ar9170 *ar)
493 {
494         int i;
495         struct sk_buff *skb;
496         struct ieee80211_tx_info *txinfo;
497         struct carl9170_tx_info *arinfo;
498         bool restart = false;
499
500         for (i = 0; i < ar->hw->queues; i++) {
501                 spin_lock_bh(&ar->tx_status[i].lock);
502
503                 skb = skb_peek(&ar->tx_status[i]);
504
505                 if (!skb)
506                         goto next;
507
508                 txinfo = IEEE80211_SKB_CB(skb);
509                 arinfo = (void *) txinfo->rate_driver_data;
510
511                 if (time_is_before_jiffies(arinfo->timeout +
512                     msecs_to_jiffies(CARL9170_QUEUE_STUCK_TIMEOUT)) == true)
513                         restart = true;
514
515 next:
516                 spin_unlock_bh(&ar->tx_status[i].lock);
517         }
518
519         if (restart) {
520                 /*
521                  * At least one queue has been stuck for long enough.
522                  * Give the device a kick and hope it gets back to
523                  * work.
524                  *
525                  * possible reasons may include:
526                  *  - frames got lost/corrupted (bad connection to the device)
527                  *  - stalled rx processing/usb controller hiccups
528                  *  - firmware errors/bugs
529                  *  - every bug you can think of.
530                  *  - all bugs you can't...
531                  *  - ...
532                  */
533                 carl9170_restart(ar, CARL9170_RR_STUCK_TX);
534         }
535 }
536
537 static void carl9170_tx_ampdu_timeout(struct ar9170 *ar)
538 {
539         struct carl9170_sta_tid *iter;
540         struct sk_buff *skb;
541         struct ieee80211_tx_info *txinfo;
542         struct carl9170_tx_info *arinfo;
543         struct _carl9170_tx_superframe *super;
544         struct ieee80211_sta *sta;
545         struct ieee80211_vif *vif;
546         struct ieee80211_hdr *hdr;
547         unsigned int vif_id;
548
549         rcu_read_lock();
550         list_for_each_entry_rcu(iter, &ar->tx_ampdu_list, list) {
551                 if (iter->state < CARL9170_TID_STATE_IDLE)
552                         continue;
553
554                 spin_lock_bh(&iter->lock);
555                 skb = skb_peek(&iter->queue);
556                 if (!skb)
557                         goto unlock;
558
559                 txinfo = IEEE80211_SKB_CB(skb);
560                 arinfo = (void *)txinfo->rate_driver_data;
561                 if (time_is_after_jiffies(arinfo->timeout +
562                     msecs_to_jiffies(CARL9170_QUEUE_TIMEOUT)))
563                         goto unlock;
564
565                 super = (void *) skb->data;
566                 hdr = (void *) super->frame_data;
567
568                 vif_id = (super->s.misc & CARL9170_TX_SUPER_MISC_VIF_ID) >>
569                          CARL9170_TX_SUPER_MISC_VIF_ID_S;
570
571                 if (WARN_ON(vif_id >= AR9170_MAX_VIRTUAL_MAC))
572                         goto unlock;
573
574                 vif = rcu_dereference(ar->vif_priv[vif_id].vif);
575                 if (WARN_ON(!vif))
576                         goto unlock;
577
578                 sta = ieee80211_find_sta(vif, hdr->addr1);
579                 if (WARN_ON(!sta))
580                         goto unlock;
581
582                 ieee80211_stop_tx_ba_session(sta, iter->tid);
583 unlock:
584                 spin_unlock_bh(&iter->lock);
585
586         }
587         rcu_read_unlock();
588 }
589
590 void carl9170_tx_janitor(struct work_struct *work)
591 {
592         struct ar9170 *ar = container_of(work, struct ar9170,
593                                          tx_janitor.work);
594         if (!IS_STARTED(ar))
595                 return;
596
597         ar->tx_janitor_last_run = jiffies;
598
599         carl9170_check_queue_stop_timeout(ar);
600         carl9170_tx_ampdu_timeout(ar);
601
602         if (!atomic_read(&ar->tx_total_queued))
603                 return;
604
605         ieee80211_queue_delayed_work(ar->hw, &ar->tx_janitor,
606                 msecs_to_jiffies(CARL9170_TX_TIMEOUT));
607 }
608
609 static void __carl9170_tx_process_status(struct ar9170 *ar,
610         const uint8_t cookie, const uint8_t info)
611 {
612         struct sk_buff *skb;
613         struct ieee80211_tx_info *txinfo;
614         struct carl9170_tx_info *arinfo;
615         unsigned int r, t, q;
616         bool success = true;
617
618         q = ar9170_qmap[info & CARL9170_TX_STATUS_QUEUE];
619
620         skb = carl9170_get_queued_skb(ar, cookie, &ar->tx_status[q]);
621         if (!skb) {
622                 /*
623                  * We have lost the race to another thread.
624                  */
625
626                 return ;
627         }
628
629         txinfo = IEEE80211_SKB_CB(skb);
630         arinfo = (void *) txinfo->rate_driver_data;
631
632         if (!(info & CARL9170_TX_STATUS_SUCCESS))
633                 success = false;
634
635         r = (info & CARL9170_TX_STATUS_RIX) >> CARL9170_TX_STATUS_RIX_S;
636         t = (info & CARL9170_TX_STATUS_TRIES) >> CARL9170_TX_STATUS_TRIES_S;
637
638         carl9170_tx_fill_rateinfo(ar, r, t, txinfo);
639         carl9170_tx_status(ar, skb, success);
640 }
641
642 void carl9170_tx_process_status(struct ar9170 *ar,
643                                 const struct carl9170_rsp *cmd)
644 {
645         unsigned int i;
646
647         for (i = 0;  i < cmd->hdr.ext; i++) {
648                 if (WARN_ON(i > ((cmd->hdr.len / 2) + 1))) {
649                         print_hex_dump_bytes("UU:", DUMP_PREFIX_NONE,
650                                              (void *) cmd, cmd->hdr.len + 4);
651                         break;
652                 }
653
654                 __carl9170_tx_process_status(ar, cmd->_tx_status[i].cookie,
655                                              cmd->_tx_status[i].info);
656         }
657 }
658
659 static __le32 carl9170_tx_physet(struct ar9170 *ar,
660         struct ieee80211_tx_info *info, struct ieee80211_tx_rate *txrate)
661 {
662         struct ieee80211_rate *rate = NULL;
663         u32 power, chains;
664         __le32 tmp;
665
666         tmp = cpu_to_le32(0);
667
668         if (txrate->flags & IEEE80211_TX_RC_40_MHZ_WIDTH)
669                 tmp |= cpu_to_le32(AR9170_TX_PHY_BW_40MHZ <<
670                         AR9170_TX_PHY_BW_S);
671         /* this works because 40 MHz is 2 and dup is 3 */
672         if (txrate->flags & IEEE80211_TX_RC_DUP_DATA)
673                 tmp |= cpu_to_le32(AR9170_TX_PHY_BW_40MHZ_DUP <<
674                         AR9170_TX_PHY_BW_S);
675
676         if (txrate->flags & IEEE80211_TX_RC_SHORT_GI)
677                 tmp |= cpu_to_le32(AR9170_TX_PHY_SHORT_GI);
678
679         if (txrate->flags & IEEE80211_TX_RC_MCS) {
680                 u32 r = txrate->idx;
681                 u8 *txpower;
682
683                 /* heavy clip control */
684                 tmp |= cpu_to_le32((r & 0x7) <<
685                         AR9170_TX_PHY_TX_HEAVY_CLIP_S);
686
687                 if (txrate->flags & IEEE80211_TX_RC_40_MHZ_WIDTH) {
688                         if (info->band == IEEE80211_BAND_5GHZ)
689                                 txpower = ar->power_5G_ht40;
690                         else
691                                 txpower = ar->power_2G_ht40;
692                 } else {
693                         if (info->band == IEEE80211_BAND_5GHZ)
694                                 txpower = ar->power_5G_ht20;
695                         else
696                                 txpower = ar->power_2G_ht20;
697                 }
698
699                 power = txpower[r & 7];
700
701                 /* +1 dBm for HT40 */
702                 if (txrate->flags & IEEE80211_TX_RC_40_MHZ_WIDTH)
703                         power += 2;
704
705                 r <<= AR9170_TX_PHY_MCS_S;
706                 BUG_ON(r & ~AR9170_TX_PHY_MCS);
707
708                 tmp |= cpu_to_le32(r & AR9170_TX_PHY_MCS);
709                 tmp |= cpu_to_le32(AR9170_TX_PHY_MOD_HT);
710
711                 /*
712                  * green field preamble does not work.
713                  *
714                  * if (txrate->flags & IEEE80211_TX_RC_GREEN_FIELD)
715                  * tmp |= cpu_to_le32(AR9170_TX_PHY_GREENFIELD);
716                  */
717         } else {
718                 u8 *txpower;
719                 u32 mod;
720                 u32 phyrate;
721                 u8 idx = txrate->idx;
722
723                 if (info->band != IEEE80211_BAND_2GHZ) {
724                         idx += 4;
725                         txpower = ar->power_5G_leg;
726                         mod = AR9170_TX_PHY_MOD_OFDM;
727                 } else {
728                         if (idx < 4) {
729                                 txpower = ar->power_2G_cck;
730                                 mod = AR9170_TX_PHY_MOD_CCK;
731                         } else {
732                                 mod = AR9170_TX_PHY_MOD_OFDM;
733                                 txpower = ar->power_2G_ofdm;
734                         }
735                 }
736
737                 rate = &__carl9170_ratetable[idx];
738
739                 phyrate = rate->hw_value & 0xF;
740                 power = txpower[(rate->hw_value & 0x30) >> 4];
741                 phyrate <<= AR9170_TX_PHY_MCS_S;
742
743                 tmp |= cpu_to_le32(mod);
744                 tmp |= cpu_to_le32(phyrate);
745
746                 /*
747                  * short preamble seems to be broken too.
748                  *
749                  * if (txrate->flags & IEEE80211_TX_RC_USE_SHORT_PREAMBLE)
750                  *      tmp |= cpu_to_le32(AR9170_TX_PHY_SHORT_PREAMBLE);
751                  */
752         }
753         power <<= AR9170_TX_PHY_TX_PWR_S;
754         power &= AR9170_TX_PHY_TX_PWR;
755         tmp |= cpu_to_le32(power);
756
757         /* set TX chains */
758         if (ar->eeprom.tx_mask == 1) {
759                 chains = AR9170_TX_PHY_TXCHAIN_1;
760         } else {
761                 chains = AR9170_TX_PHY_TXCHAIN_2;
762
763                 /* >= 36M legacy OFDM - use only one chain */
764                 if (rate && rate->bitrate >= 360 &&
765                     !(txrate->flags & IEEE80211_TX_RC_MCS))
766                         chains = AR9170_TX_PHY_TXCHAIN_1;
767         }
768         tmp |= cpu_to_le32(chains << AR9170_TX_PHY_TXCHAIN_S);
769
770         return tmp;
771 }
772
773 static bool carl9170_tx_rts_check(struct ar9170 *ar,
774                                   struct ieee80211_tx_rate *rate,
775                                   bool ampdu, bool multi)
776 {
777         switch (ar->erp_mode) {
778         case CARL9170_ERP_AUTO:
779                 if (ampdu)
780                         break;
781
782         case CARL9170_ERP_MAC80211:
783                 if (!(rate->flags & IEEE80211_TX_RC_USE_RTS_CTS))
784                         break;
785
786         case CARL9170_ERP_RTS:
787                 if (likely(!multi))
788                         return true;
789
790         default:
791                 break;
792         }
793
794         return false;
795 }
796
797 static bool carl9170_tx_cts_check(struct ar9170 *ar,
798                                   struct ieee80211_tx_rate *rate)
799 {
800         switch (ar->erp_mode) {
801         case CARL9170_ERP_AUTO:
802         case CARL9170_ERP_MAC80211:
803                 if (!(rate->flags & IEEE80211_TX_RC_USE_CTS_PROTECT))
804                         break;
805
806         case CARL9170_ERP_CTS:
807                 return true;
808
809         default:
810                 break;
811         }
812
813         return false;
814 }
815
816 static int carl9170_tx_prepare(struct ar9170 *ar, struct sk_buff *skb)
817 {
818         struct ieee80211_hdr *hdr;
819         struct _carl9170_tx_superframe *txc;
820         struct carl9170_vif_info *cvif;
821         struct ieee80211_tx_info *info;
822         struct ieee80211_tx_rate *txrate;
823         struct ieee80211_sta *sta;
824         struct carl9170_tx_info *arinfo;
825         unsigned int hw_queue;
826         int i;
827         __le16 mac_tmp;
828         u16 len;
829         bool ampdu, no_ack;
830
831         BUILD_BUG_ON(sizeof(*arinfo) > sizeof(info->rate_driver_data));
832         BUILD_BUG_ON(sizeof(struct _carl9170_tx_superdesc) !=
833                      CARL9170_TX_SUPERDESC_LEN);
834
835         BUILD_BUG_ON(sizeof(struct _ar9170_tx_hwdesc) !=
836                      AR9170_TX_HWDESC_LEN);
837
838         BUILD_BUG_ON(IEEE80211_TX_MAX_RATES < CARL9170_TX_MAX_RATES);
839
840         BUILD_BUG_ON(AR9170_MAX_VIRTUAL_MAC >
841                 ((CARL9170_TX_SUPER_MISC_VIF_ID >>
842                  CARL9170_TX_SUPER_MISC_VIF_ID_S) + 1));
843
844         hw_queue = ar9170_qmap[carl9170_get_queue(ar, skb)];
845
846         hdr = (void *)skb->data;
847         info = IEEE80211_SKB_CB(skb);
848         len = skb->len;
849
850         /*
851          * Note: If the frame was sent through a monitor interface,
852          * the ieee80211_vif pointer can be NULL.
853          */
854         if (likely(info->control.vif))
855                 cvif = (void *) info->control.vif->drv_priv;
856         else
857                 cvif = NULL;
858
859         sta = info->control.sta;
860
861         txc = (void *)skb_push(skb, sizeof(*txc));
862         memset(txc, 0, sizeof(*txc));
863
864         SET_VAL(CARL9170_TX_SUPER_MISC_QUEUE, txc->s.misc, hw_queue);
865
866         if (likely(cvif))
867                 SET_VAL(CARL9170_TX_SUPER_MISC_VIF_ID, txc->s.misc, cvif->id);
868
869         if (unlikely(info->flags & IEEE80211_TX_CTL_SEND_AFTER_DTIM))
870                 txc->s.misc |= CARL9170_TX_SUPER_MISC_CAB;
871
872         if (unlikely(info->flags & IEEE80211_TX_CTL_ASSIGN_SEQ))
873                 txc->s.misc |= CARL9170_TX_SUPER_MISC_ASSIGN_SEQ;
874
875         if (unlikely(ieee80211_is_probe_resp(hdr->frame_control)))
876                 txc->s.misc |= CARL9170_TX_SUPER_MISC_FILL_IN_TSF;
877
878         mac_tmp = cpu_to_le16(AR9170_TX_MAC_HW_DURATION |
879                               AR9170_TX_MAC_BACKOFF);
880         mac_tmp |= cpu_to_le16((hw_queue << AR9170_TX_MAC_QOS_S) &
881                                AR9170_TX_MAC_QOS);
882
883         no_ack = !!(info->flags & IEEE80211_TX_CTL_NO_ACK);
884         if (unlikely(no_ack))
885                 mac_tmp |= cpu_to_le16(AR9170_TX_MAC_NO_ACK);
886
887         if (info->control.hw_key) {
888                 len += info->control.hw_key->icv_len;
889
890                 switch (info->control.hw_key->cipher) {
891                 case WLAN_CIPHER_SUITE_WEP40:
892                 case WLAN_CIPHER_SUITE_WEP104:
893                 case WLAN_CIPHER_SUITE_TKIP:
894                         mac_tmp |= cpu_to_le16(AR9170_TX_MAC_ENCR_RC4);
895                         break;
896                 case WLAN_CIPHER_SUITE_CCMP:
897                         mac_tmp |= cpu_to_le16(AR9170_TX_MAC_ENCR_AES);
898                         break;
899                 default:
900                         WARN_ON(1);
901                         goto err_out;
902                 }
903         }
904
905         ampdu = !!(info->flags & IEEE80211_TX_CTL_AMPDU);
906         if (ampdu) {
907                 unsigned int density, factor;
908
909                 if (unlikely(!sta || !cvif))
910                         goto err_out;
911
912                 factor = min_t(unsigned int, 1u, sta->ht_cap.ampdu_factor);
913                 density = sta->ht_cap.ampdu_density;
914
915                 if (density) {
916                         /*
917                          * Watch out!
918                          *
919                          * Otus uses slightly different density values than
920                          * those from the 802.11n spec.
921                          */
922
923                         density = max_t(unsigned int, density + 1, 7u);
924                 }
925
926                 SET_VAL(CARL9170_TX_SUPER_AMPDU_DENSITY,
927                         txc->s.ampdu_settings, density);
928
929                 SET_VAL(CARL9170_TX_SUPER_AMPDU_FACTOR,
930                         txc->s.ampdu_settings, factor);
931
932                 for (i = 0; i < CARL9170_TX_MAX_RATES; i++) {
933                         txrate = &info->control.rates[i];
934                         if (txrate->idx >= 0) {
935                                 txc->s.ri[i] =
936                                         CARL9170_TX_SUPER_RI_AMPDU;
937
938                                 if (WARN_ON(!(txrate->flags &
939                                               IEEE80211_TX_RC_MCS))) {
940                                         /*
941                                          * Not sure if it's even possible
942                                          * to aggregate non-ht rates with
943                                          * this HW.
944                                          */
945                                         goto err_out;
946                                 }
947                                 continue;
948                         }
949
950                         txrate->idx = 0;
951                         txrate->count = ar->hw->max_rate_tries;
952                 }
953
954                 mac_tmp |= cpu_to_le16(AR9170_TX_MAC_AGGR);
955         }
956
957         /*
958          * NOTE: For the first rate, the ERP & AMPDU flags are directly
959          * taken from mac_control. For all fallback rate, the firmware
960          * updates the mac_control flags from the rate info field.
961          */
962         for (i = 1; i < CARL9170_TX_MAX_RATES; i++) {
963                 txrate = &info->control.rates[i];
964                 if (txrate->idx < 0)
965                         break;
966
967                 SET_VAL(CARL9170_TX_SUPER_RI_TRIES, txc->s.ri[i],
968                         txrate->count);
969
970                 if (carl9170_tx_rts_check(ar, txrate, ampdu, no_ack))
971                         txc->s.ri[i] |= (AR9170_TX_MAC_PROT_RTS <<
972                                 CARL9170_TX_SUPER_RI_ERP_PROT_S);
973                 else if (carl9170_tx_cts_check(ar, txrate))
974                         txc->s.ri[i] |= (AR9170_TX_MAC_PROT_CTS <<
975                                 CARL9170_TX_SUPER_RI_ERP_PROT_S);
976
977                 txc->s.rr[i - 1] = carl9170_tx_physet(ar, info, txrate);
978         }
979
980         txrate = &info->control.rates[0];
981         SET_VAL(CARL9170_TX_SUPER_RI_TRIES, txc->s.ri[0], txrate->count);
982
983         if (carl9170_tx_rts_check(ar, txrate, ampdu, no_ack))
984                 mac_tmp |= cpu_to_le16(AR9170_TX_MAC_PROT_RTS);
985         else if (carl9170_tx_cts_check(ar, txrate))
986                 mac_tmp |= cpu_to_le16(AR9170_TX_MAC_PROT_CTS);
987
988         txc->s.len = cpu_to_le16(skb->len);
989         txc->f.length = cpu_to_le16(len + FCS_LEN);
990         txc->f.mac_control = mac_tmp;
991         txc->f.phy_control = carl9170_tx_physet(ar, info, txrate);
992
993         arinfo = (void *)info->rate_driver_data;
994         arinfo->timeout = jiffies;
995         arinfo->ar = ar;
996         kref_init(&arinfo->ref);
997         return 0;
998
999 err_out:
1000         skb_pull(skb, sizeof(*txc));
1001         return -EINVAL;
1002 }
1003
1004 static void carl9170_set_immba(struct ar9170 *ar, struct sk_buff *skb)
1005 {
1006         struct _carl9170_tx_superframe *super;
1007
1008         super = (void *) skb->data;
1009         super->f.mac_control |= cpu_to_le16(AR9170_TX_MAC_IMM_BA);
1010 }
1011
1012 static void carl9170_set_ampdu_params(struct ar9170 *ar, struct sk_buff *skb)
1013 {
1014         struct _carl9170_tx_superframe *super;
1015         int tmp;
1016
1017         super = (void *) skb->data;
1018
1019         tmp = (super->s.ampdu_settings & CARL9170_TX_SUPER_AMPDU_DENSITY) <<
1020                 CARL9170_TX_SUPER_AMPDU_DENSITY_S;
1021
1022         /*
1023          * If you haven't noticed carl9170_tx_prepare has already filled
1024          * in all ampdu spacing & factor parameters.
1025          * Now it's the time to check whenever the settings have to be
1026          * updated by the firmware, or if everything is still the same.
1027          *
1028          * There's no sane way to handle different density values with
1029          * this hardware, so we may as well just do the compare in the
1030          * driver.
1031          */
1032
1033         if (tmp != ar->current_density) {
1034                 ar->current_density = tmp;
1035                 super->s.ampdu_settings |=
1036                         CARL9170_TX_SUPER_AMPDU_COMMIT_DENSITY;
1037         }
1038
1039         tmp = (super->s.ampdu_settings & CARL9170_TX_SUPER_AMPDU_FACTOR) <<
1040                 CARL9170_TX_SUPER_AMPDU_FACTOR_S;
1041
1042         if (tmp != ar->current_factor) {
1043                 ar->current_factor = tmp;
1044                 super->s.ampdu_settings |=
1045                         CARL9170_TX_SUPER_AMPDU_COMMIT_FACTOR;
1046         }
1047 }
1048
1049 static bool carl9170_tx_rate_check(struct ar9170 *ar, struct sk_buff *_dest,
1050                                    struct sk_buff *_src)
1051 {
1052         struct _carl9170_tx_superframe *dest, *src;
1053
1054         dest = (void *) _dest->data;
1055         src = (void *) _src->data;
1056
1057         /*
1058          * The mac80211 rate control algorithm expects that all MPDUs in
1059          * an AMPDU share the same tx vectors.
1060          * This is not really obvious right now, because the hardware
1061          * does the AMPDU setup according to its own rulebook.
1062          * Our nicely assembled, strictly monotonic increasing mpdu
1063          * chains will be broken up, mashed back together...
1064          */
1065
1066         return (dest->f.phy_control == src->f.phy_control);
1067 }
1068
1069 static void carl9170_tx_ampdu(struct ar9170 *ar)
1070 {
1071         struct sk_buff_head agg;
1072         struct carl9170_sta_tid *tid_info;
1073         struct sk_buff *skb, *first;
1074         unsigned int i = 0, done_ampdus = 0;
1075         u16 seq, queue, tmpssn;
1076
1077         atomic_inc(&ar->tx_ampdu_scheduler);
1078         ar->tx_ampdu_schedule = false;
1079
1080         if (atomic_read(&ar->tx_ampdu_upload))
1081                 return;
1082
1083         if (!ar->tx_ampdu_list_len)
1084                 return;
1085
1086         __skb_queue_head_init(&agg);
1087
1088         rcu_read_lock();
1089         tid_info = rcu_dereference(ar->tx_ampdu_iter);
1090         if (WARN_ON_ONCE(!tid_info)) {
1091                 rcu_read_unlock();
1092                 return;
1093         }
1094
1095 retry:
1096         list_for_each_entry_continue_rcu(tid_info, &ar->tx_ampdu_list, list) {
1097                 i++;
1098
1099                 if (tid_info->state < CARL9170_TID_STATE_PROGRESS)
1100                         continue;
1101
1102                 queue = TID_TO_WME_AC(tid_info->tid);
1103
1104                 spin_lock_bh(&tid_info->lock);
1105                 if (tid_info->state != CARL9170_TID_STATE_XMIT)
1106                         goto processed;
1107
1108                 tid_info->counter++;
1109                 first = skb_peek(&tid_info->queue);
1110                 tmpssn = carl9170_get_seq(first);
1111                 seq = tid_info->snx;
1112
1113                 if (unlikely(tmpssn != seq)) {
1114                         tid_info->state = CARL9170_TID_STATE_IDLE;
1115
1116                         goto processed;
1117                 }
1118
1119                 while ((skb = skb_peek(&tid_info->queue))) {
1120                         /* strict 0, 1, ..., n - 1, n frame sequence order */
1121                         if (unlikely(carl9170_get_seq(skb) != seq))
1122                                 break;
1123
1124                         /* don't upload more than AMPDU FACTOR allows. */
1125                         if (unlikely(SEQ_DIFF(tid_info->snx, tid_info->bsn) >=
1126                             (tid_info->max - 1)))
1127                                 break;
1128
1129                         if (!carl9170_tx_rate_check(ar, skb, first))
1130                                 break;
1131
1132                         atomic_inc(&ar->tx_ampdu_upload);
1133                         tid_info->snx = seq = SEQ_NEXT(seq);
1134                         __skb_unlink(skb, &tid_info->queue);
1135
1136                         __skb_queue_tail(&agg, skb);
1137
1138                         if (skb_queue_len(&agg) >= CARL9170_NUM_TX_AGG_MAX)
1139                                 break;
1140                 }
1141
1142                 if (skb_queue_empty(&tid_info->queue) ||
1143                     carl9170_get_seq(skb_peek(&tid_info->queue)) !=
1144                     tid_info->snx) {
1145                         /*
1146                          * stop TID, if A-MPDU frames are still missing,
1147                          * or whenever the queue is empty.
1148                          */
1149
1150                         tid_info->state = CARL9170_TID_STATE_IDLE;
1151                 }
1152                 done_ampdus++;
1153
1154 processed:
1155                 spin_unlock_bh(&tid_info->lock);
1156
1157                 if (skb_queue_empty(&agg))
1158                         continue;
1159
1160                 /* apply ampdu spacing & factor settings */
1161                 carl9170_set_ampdu_params(ar, skb_peek(&agg));
1162
1163                 /* set aggregation push bit */
1164                 carl9170_set_immba(ar, skb_peek_tail(&agg));
1165
1166                 spin_lock_bh(&ar->tx_pending[queue].lock);
1167                 skb_queue_splice_tail_init(&agg, &ar->tx_pending[queue]);
1168                 spin_unlock_bh(&ar->tx_pending[queue].lock);
1169                 ar->tx_schedule = true;
1170         }
1171         if ((done_ampdus++ == 0) && (i++ == 0))
1172                 goto retry;
1173
1174         rcu_assign_pointer(ar->tx_ampdu_iter, tid_info);
1175         rcu_read_unlock();
1176 }
1177
1178 static struct sk_buff *carl9170_tx_pick_skb(struct ar9170 *ar,
1179                                             struct sk_buff_head *queue)
1180 {
1181         struct sk_buff *skb;
1182         struct ieee80211_tx_info *info;
1183         struct carl9170_tx_info *arinfo;
1184
1185         BUILD_BUG_ON(sizeof(*arinfo) > sizeof(info->rate_driver_data));
1186
1187         spin_lock_bh(&queue->lock);
1188         skb = skb_peek(queue);
1189         if (unlikely(!skb))
1190                 goto err_unlock;
1191
1192         if (carl9170_alloc_dev_space(ar, skb))
1193                 goto err_unlock;
1194
1195         __skb_unlink(skb, queue);
1196         spin_unlock_bh(&queue->lock);
1197
1198         info = IEEE80211_SKB_CB(skb);
1199         arinfo = (void *) info->rate_driver_data;
1200
1201         arinfo->timeout = jiffies;
1202
1203         /*
1204          * increase ref count to "2".
1205          * Ref counting is the easiest way to solve the race between
1206          * the the urb's completion routine: carl9170_tx_callback and
1207          * wlan tx status functions: carl9170_tx_status/janitor.
1208          */
1209         carl9170_tx_get_skb(skb);
1210
1211         return skb;
1212
1213 err_unlock:
1214         spin_unlock_bh(&queue->lock);
1215         return NULL;
1216 }
1217
1218 void carl9170_tx_drop(struct ar9170 *ar, struct sk_buff *skb)
1219 {
1220         struct _carl9170_tx_superframe *super;
1221         uint8_t q = 0;
1222
1223         ar->tx_dropped++;
1224
1225         super = (void *)skb->data;
1226         SET_VAL(CARL9170_TX_SUPER_MISC_QUEUE, q,
1227                 ar9170_qmap[carl9170_get_queue(ar, skb)]);
1228         __carl9170_tx_process_status(ar, super->s.cookie, q);
1229 }
1230
1231 static void carl9170_tx(struct ar9170 *ar)
1232 {
1233         struct sk_buff *skb;
1234         unsigned int i, q;
1235         bool schedule_garbagecollector = false;
1236
1237         ar->tx_schedule = false;
1238
1239         if (unlikely(!IS_STARTED(ar)))
1240                 return;
1241
1242         carl9170_usb_handle_tx_err(ar);
1243
1244         for (i = 0; i < ar->hw->queues; i++) {
1245                 while (!skb_queue_empty(&ar->tx_pending[i])) {
1246                         skb = carl9170_tx_pick_skb(ar, &ar->tx_pending[i]);
1247                         if (unlikely(!skb))
1248                                 break;
1249
1250                         atomic_inc(&ar->tx_total_pending);
1251
1252                         q = __carl9170_get_queue(ar, i);
1253                         /*
1254                          * NB: tx_status[i] vs. tx_status[q],
1255                          * TODO: Move into pick_skb or alloc_dev_space.
1256                          */
1257                         skb_queue_tail(&ar->tx_status[q], skb);
1258
1259                         carl9170_usb_tx(ar, skb);
1260                         schedule_garbagecollector = true;
1261                 }
1262         }
1263
1264         if (!schedule_garbagecollector)
1265                 return;
1266
1267         ieee80211_queue_delayed_work(ar->hw, &ar->tx_janitor,
1268                 msecs_to_jiffies(CARL9170_TX_TIMEOUT));
1269 }
1270
1271 static bool carl9170_tx_ampdu_queue(struct ar9170 *ar,
1272         struct ieee80211_sta *sta, struct sk_buff *skb)
1273 {
1274         struct _carl9170_tx_superframe *super = (void *) skb->data;
1275         struct carl9170_sta_info *sta_info;
1276         struct carl9170_sta_tid *agg;
1277         struct sk_buff *iter;
1278         unsigned int max;
1279         u16 tid, seq, qseq, off;
1280         bool run = false;
1281
1282         tid = carl9170_get_tid(skb);
1283         seq = carl9170_get_seq(skb);
1284         sta_info = (void *) sta->drv_priv;
1285
1286         rcu_read_lock();
1287         agg = rcu_dereference(sta_info->agg[tid]);
1288         max = sta_info->ampdu_max_len;
1289
1290         if (!agg)
1291                 goto err_unlock_rcu;
1292
1293         spin_lock_bh(&agg->lock);
1294         if (unlikely(agg->state < CARL9170_TID_STATE_IDLE))
1295                 goto err_unlock;
1296
1297         /* check if sequence is within the BA window */
1298         if (unlikely(!BAW_WITHIN(agg->bsn, CARL9170_BAW_BITS, seq)))
1299                 goto err_unlock;
1300
1301         if (WARN_ON_ONCE(!BAW_WITHIN(agg->snx, CARL9170_BAW_BITS, seq)))
1302                 goto err_unlock;
1303
1304         off = SEQ_DIFF(seq, agg->bsn);
1305         if (WARN_ON_ONCE(test_and_set_bit(off, agg->bitmap)))
1306                 goto err_unlock;
1307
1308         if (likely(BAW_WITHIN(agg->hsn, CARL9170_BAW_BITS, seq))) {
1309                 __skb_queue_tail(&agg->queue, skb);
1310                 agg->hsn = seq;
1311                 goto queued;
1312         }
1313
1314         skb_queue_reverse_walk(&agg->queue, iter) {
1315                 qseq = carl9170_get_seq(iter);
1316
1317                 if (BAW_WITHIN(qseq, CARL9170_BAW_BITS, seq)) {
1318                         __skb_queue_after(&agg->queue, iter, skb);
1319                         goto queued;
1320                 }
1321         }
1322
1323         __skb_queue_head(&agg->queue, skb);
1324 queued:
1325
1326         if (unlikely(agg->state != CARL9170_TID_STATE_XMIT)) {
1327                 if (agg->snx == carl9170_get_seq(skb_peek(&agg->queue))) {
1328                         agg->state = CARL9170_TID_STATE_XMIT;
1329                         run = true;
1330                 }
1331         }
1332
1333         spin_unlock_bh(&agg->lock);
1334         rcu_read_unlock();
1335
1336         return run;
1337
1338 err_unlock:
1339         spin_unlock_bh(&agg->lock);
1340
1341 err_unlock_rcu:
1342         rcu_read_unlock();
1343         super->f.mac_control &= ~cpu_to_le16(AR9170_TX_MAC_AGGR);
1344         carl9170_tx_status(ar, skb, false);
1345         ar->tx_dropped++;
1346         return false;
1347 }
1348
1349 void carl9170_op_tx(struct ieee80211_hw *hw, struct sk_buff *skb)
1350 {
1351         struct ar9170 *ar = hw->priv;
1352         struct ieee80211_tx_info *info;
1353         struct ieee80211_sta *sta;
1354         bool run;
1355
1356         if (unlikely(!IS_STARTED(ar)))
1357                 goto err_free;
1358
1359         info = IEEE80211_SKB_CB(skb);
1360         sta = info->control.sta;
1361
1362         if (unlikely(carl9170_tx_prepare(ar, skb)))
1363                 goto err_free;
1364
1365         carl9170_tx_accounting(ar, skb);
1366         /*
1367          * from now on, one has to use carl9170_tx_status to free
1368          * all ressouces which are associated with the frame.
1369          */
1370
1371         if (info->flags & IEEE80211_TX_CTL_AMPDU) {
1372                 run = carl9170_tx_ampdu_queue(ar, sta, skb);
1373                 if (run)
1374                         carl9170_tx_ampdu(ar);
1375
1376         } else {
1377                 unsigned int queue = skb_get_queue_mapping(skb);
1378
1379                 skb_queue_tail(&ar->tx_pending[queue], skb);
1380         }
1381
1382         carl9170_tx(ar);
1383         return;
1384
1385 err_free:
1386         ar->tx_dropped++;
1387         dev_kfree_skb_any(skb);
1388 }
1389
1390 void carl9170_tx_scheduler(struct ar9170 *ar)
1391 {
1392
1393         if (ar->tx_ampdu_schedule)
1394                 carl9170_tx_ampdu(ar);
1395
1396         if (ar->tx_schedule)
1397                 carl9170_tx(ar);
1398 }