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/slab.h>
41 #include <linux/module.h>
42 #include <linux/etherdevice.h>
43 #include <net/mac80211.h>
44 #include "carl9170.h"
45 #include "hw.h"
46 #include "cmd.h"
47 
48 static inline unsigned int __carl9170_get_queue(struct ar9170 *ar,
49 						unsigned int queue)
50 {
51 	if (unlikely(modparam_noht)) {
52 		return queue;
53 	} else {
54 		/*
55 		 * This is just another workaround, until
56 		 * someone figures out how to get QoS and
57 		 * AMPDU to play nicely together.
58 		 */
59 
60 		return 2;		/* AC_BE */
61 	}
62 }
63 
64 static inline unsigned int carl9170_get_queue(struct ar9170 *ar,
65 					      struct sk_buff *skb)
66 {
67 	return __carl9170_get_queue(ar, skb_get_queue_mapping(skb));
68 }
69 
70 static bool is_mem_full(struct ar9170 *ar)
71 {
72 	return (DIV_ROUND_UP(IEEE80211_MAX_FRAME_LEN, ar->fw.mem_block_size) >
73 		atomic_read(&ar->mem_free_blocks));
74 }
75 
76 static void carl9170_tx_accounting(struct ar9170 *ar, struct sk_buff *skb)
77 {
78 	int queue, i;
79 	bool mem_full;
80 
81 	atomic_inc(&ar->tx_total_queued);
82 
83 	queue = skb_get_queue_mapping(skb);
84 	spin_lock_bh(&ar->tx_stats_lock);
85 
86 	/*
87 	 * The driver has to accept the frame, regardless if the queue is
88 	 * full to the brim, or not. We have to do the queuing internally,
89 	 * since mac80211 assumes that a driver which can operate with
90 	 * aggregated frames does not reject frames for this reason.
91 	 */
92 	ar->tx_stats[queue].len++;
93 	ar->tx_stats[queue].count++;
94 
95 	mem_full = is_mem_full(ar);
96 	for (i = 0; i < ar->hw->queues; i++) {
97 		if (mem_full || ar->tx_stats[i].len >= ar->tx_stats[i].limit) {
98 			ieee80211_stop_queue(ar->hw, i);
99 			ar->queue_stop_timeout[i] = jiffies;
100 		}
101 	}
102 
103 	spin_unlock_bh(&ar->tx_stats_lock);
104 }
105 
106 /* needs rcu_read_lock */
107 static struct ieee80211_sta *__carl9170_get_tx_sta(struct ar9170 *ar,
108 						   struct sk_buff *skb)
109 {
110 	struct _carl9170_tx_superframe *super = (void *) skb->data;
111 	struct ieee80211_hdr *hdr = (void *) super->frame_data;
112 	struct ieee80211_vif *vif;
113 	unsigned int vif_id;
114 
115 	vif_id = (super->s.misc & CARL9170_TX_SUPER_MISC_VIF_ID) >>
116 		 CARL9170_TX_SUPER_MISC_VIF_ID_S;
117 
118 	if (WARN_ON_ONCE(vif_id >= AR9170_MAX_VIRTUAL_MAC))
119 		return NULL;
120 
121 	vif = rcu_dereference(ar->vif_priv[vif_id].vif);
122 	if (unlikely(!vif))
123 		return NULL;
124 
125 	/*
126 	 * Normally we should use wrappers like ieee80211_get_DA to get
127 	 * the correct peer ieee80211_sta.
128 	 *
129 	 * But there is a problem with indirect traffic (broadcasts, or
130 	 * data which is designated for other stations) in station mode.
131 	 * The frame will be directed to the AP for distribution and not
132 	 * to the actual destination.
133 	 */
134 
135 	return ieee80211_find_sta(vif, hdr->addr1);
136 }
137 
138 static void carl9170_tx_ps_unblock(struct ar9170 *ar, struct sk_buff *skb)
139 {
140 	struct ieee80211_sta *sta;
141 	struct carl9170_sta_info *sta_info;
142 
143 	rcu_read_lock();
144 	sta = __carl9170_get_tx_sta(ar, skb);
145 	if (unlikely(!sta))
146 		goto out_rcu;
147 
148 	sta_info = (struct carl9170_sta_info *) sta->drv_priv;
149 	if (atomic_dec_return(&sta_info->pending_frames) == 0)
150 		ieee80211_sta_block_awake(ar->hw, sta, false);
151 
152 out_rcu:
153 	rcu_read_unlock();
154 }
155 
156 static void carl9170_tx_accounting_free(struct ar9170 *ar, struct sk_buff *skb)
157 {
158 	int queue;
159 
160 	queue = skb_get_queue_mapping(skb);
161 
162 	spin_lock_bh(&ar->tx_stats_lock);
163 
164 	ar->tx_stats[queue].len--;
165 
166 	if (!is_mem_full(ar)) {
167 		unsigned int i;
168 		for (i = 0; i < ar->hw->queues; i++) {
169 			if (ar->tx_stats[i].len >= CARL9170_NUM_TX_LIMIT_SOFT)
170 				continue;
171 
172 			if (ieee80211_queue_stopped(ar->hw, i)) {
173 				unsigned long tmp;
174 
175 				tmp = jiffies - ar->queue_stop_timeout[i];
176 				if (tmp > ar->max_queue_stop_timeout[i])
177 					ar->max_queue_stop_timeout[i] = tmp;
178 			}
179 
180 			ieee80211_wake_queue(ar->hw, i);
181 		}
182 	}
183 
184 	spin_unlock_bh(&ar->tx_stats_lock);
185 
186 	if (atomic_dec_and_test(&ar->tx_total_queued))
187 		complete(&ar->tx_flush);
188 }
189 
190 static int carl9170_alloc_dev_space(struct ar9170 *ar, struct sk_buff *skb)
191 {
192 	struct _carl9170_tx_superframe *super = (void *) skb->data;
193 	unsigned int chunks;
194 	int cookie = -1;
195 
196 	atomic_inc(&ar->mem_allocs);
197 
198 	chunks = DIV_ROUND_UP(skb->len, ar->fw.mem_block_size);
199 	if (unlikely(atomic_sub_return(chunks, &ar->mem_free_blocks) < 0)) {
200 		atomic_add(chunks, &ar->mem_free_blocks);
201 		return -ENOSPC;
202 	}
203 
204 	spin_lock_bh(&ar->mem_lock);
205 	cookie = bitmap_find_free_region(ar->mem_bitmap, ar->fw.mem_blocks, 0);
206 	spin_unlock_bh(&ar->mem_lock);
207 
208 	if (unlikely(cookie < 0)) {
209 		atomic_add(chunks, &ar->mem_free_blocks);
210 		return -ENOSPC;
211 	}
212 
213 	super = (void *) skb->data;
214 
215 	/*
216 	 * Cookie #0 serves two special purposes:
217 	 *  1. The firmware might use it generate BlockACK frames
218 	 *     in responds of an incoming BlockAckReqs.
219 	 *
220 	 *  2. Prevent double-free bugs.
221 	 */
222 	super->s.cookie = (u8) cookie + 1;
223 	return 0;
224 }
225 
226 static void carl9170_release_dev_space(struct ar9170 *ar, struct sk_buff *skb)
227 {
228 	struct _carl9170_tx_superframe *super = (void *) skb->data;
229 	int cookie;
230 
231 	/* make a local copy of the cookie */
232 	cookie = super->s.cookie;
233 	/* invalidate cookie */
234 	super->s.cookie = 0;
235 
236 	/*
237 	 * Do a out-of-bounds check on the cookie:
238 	 *
239 	 *  * cookie "0" is reserved and won't be assigned to any
240 	 *    out-going frame. Internally however, it is used to
241 	 *    mark no longer/un-accounted frames and serves as a
242 	 *    cheap way of preventing frames from being freed
243 	 *    twice by _accident_. NB: There is a tiny race...
244 	 *
245 	 *  * obviously, cookie number is limited by the amount
246 	 *    of available memory blocks, so the number can
247 	 *    never execeed the mem_blocks count.
248 	 */
249 	if (unlikely(WARN_ON_ONCE(cookie == 0) ||
250 	    WARN_ON_ONCE(cookie > ar->fw.mem_blocks)))
251 		return;
252 
253 	atomic_add(DIV_ROUND_UP(skb->len, ar->fw.mem_block_size),
254 		   &ar->mem_free_blocks);
255 
256 	spin_lock_bh(&ar->mem_lock);
257 	bitmap_release_region(ar->mem_bitmap, cookie - 1, 0);
258 	spin_unlock_bh(&ar->mem_lock);
259 }
260 
261 /* Called from any context */
262 static void carl9170_tx_release(struct kref *ref)
263 {
264 	struct ar9170 *ar;
265 	struct carl9170_tx_info *arinfo;
266 	struct ieee80211_tx_info *txinfo;
267 	struct sk_buff *skb;
268 
269 	arinfo = container_of(ref, struct carl9170_tx_info, ref);
270 	txinfo = container_of((void *) arinfo, struct ieee80211_tx_info,
271 			      rate_driver_data);
272 	skb = container_of((void *) txinfo, struct sk_buff, cb);
273 
274 	ar = arinfo->ar;
275 	if (WARN_ON_ONCE(!ar))
276 		return;
277 
278 	BUILD_BUG_ON(
279 	    offsetof(struct ieee80211_tx_info, status.ack_signal) != 20);
280 
281 	memset(&txinfo->status.ack_signal, 0,
282 	       sizeof(struct ieee80211_tx_info) -
283 	       offsetof(struct ieee80211_tx_info, status.ack_signal));
284 
285 	if (atomic_read(&ar->tx_total_queued))
286 		ar->tx_schedule = true;
287 
288 	if (txinfo->flags & IEEE80211_TX_CTL_AMPDU) {
289 		if (!atomic_read(&ar->tx_ampdu_upload))
290 			ar->tx_ampdu_schedule = true;
291 
292 		if (txinfo->flags & IEEE80211_TX_STAT_AMPDU) {
293 			struct _carl9170_tx_superframe *super;
294 
295 			super = (void *)skb->data;
296 			txinfo->status.ampdu_len = super->s.rix;
297 			txinfo->status.ampdu_ack_len = super->s.cnt;
298 		} else if ((txinfo->flags & IEEE80211_TX_STAT_ACK) &&
299 			   !(txinfo->flags & IEEE80211_TX_CTL_REQ_TX_STATUS)) {
300 			/*
301 			 * drop redundant tx_status reports:
302 			 *
303 			 * 1. ampdu_ack_len of the final tx_status does
304 			 *    include the feedback of this particular frame.
305 			 *
306 			 * 2. tx_status_irqsafe only queues up to 128
307 			 *    tx feedback reports and discards the rest.
308 			 *
309 			 * 3. minstrel_ht is picky, it only accepts
310 			 *    reports of frames with the TX_STATUS_AMPDU flag.
311 			 *
312 			 * 4. mac80211 is not particularly interested in
313 			 *    feedback either [CTL_REQ_TX_STATUS not set]
314 			 */
315 
316 			ieee80211_free_txskb(ar->hw, skb);
317 			return;
318 		} else {
319 			/*
320 			 * Either the frame transmission has failed or
321 			 * mac80211 requested tx status.
322 			 */
323 		}
324 	}
325 
326 	skb_pull(skb, sizeof(struct _carl9170_tx_superframe));
327 	ieee80211_tx_status_irqsafe(ar->hw, skb);
328 }
329 
330 void carl9170_tx_get_skb(struct sk_buff *skb)
331 {
332 	struct carl9170_tx_info *arinfo = (void *)
333 		(IEEE80211_SKB_CB(skb))->rate_driver_data;
334 	kref_get(&arinfo->ref);
335 }
336 
337 int carl9170_tx_put_skb(struct sk_buff *skb)
338 {
339 	struct carl9170_tx_info *arinfo = (void *)
340 		(IEEE80211_SKB_CB(skb))->rate_driver_data;
341 
342 	return kref_put(&arinfo->ref, carl9170_tx_release);
343 }
344 
345 /* Caller must hold the tid_info->lock & rcu_read_lock */
346 static void carl9170_tx_shift_bm(struct ar9170 *ar,
347 	struct carl9170_sta_tid *tid_info, u16 seq)
348 {
349 	u16 off;
350 
351 	off = SEQ_DIFF(seq, tid_info->bsn);
352 
353 	if (WARN_ON_ONCE(off >= CARL9170_BAW_BITS))
354 		return;
355 
356 	/*
357 	 * Sanity check. For each MPDU we set the bit in bitmap and
358 	 * clear it once we received the tx_status.
359 	 * But if the bit is already cleared then we've been bitten
360 	 * by a bug.
361 	 */
362 	WARN_ON_ONCE(!test_and_clear_bit(off, tid_info->bitmap));
363 
364 	off = SEQ_DIFF(tid_info->snx, tid_info->bsn);
365 	if (WARN_ON_ONCE(off >= CARL9170_BAW_BITS))
366 		return;
367 
368 	if (!bitmap_empty(tid_info->bitmap, off))
369 		off = find_first_bit(tid_info->bitmap, off);
370 
371 	tid_info->bsn += off;
372 	tid_info->bsn &= 0x0fff;
373 
374 	bitmap_shift_right(tid_info->bitmap, tid_info->bitmap,
375 			   off, CARL9170_BAW_BITS);
376 }
377 
378 static void carl9170_tx_status_process_ampdu(struct ar9170 *ar,
379 	struct sk_buff *skb, struct ieee80211_tx_info *txinfo)
380 {
381 	struct _carl9170_tx_superframe *super = (void *) skb->data;
382 	struct ieee80211_hdr *hdr = (void *) super->frame_data;
383 	struct ieee80211_sta *sta;
384 	struct carl9170_sta_info *sta_info;
385 	struct carl9170_sta_tid *tid_info;
386 	u8 tid;
387 
388 	if (!(txinfo->flags & IEEE80211_TX_CTL_AMPDU) ||
389 	    txinfo->flags & IEEE80211_TX_CTL_INJECTED)
390 		return;
391 
392 	rcu_read_lock();
393 	sta = __carl9170_get_tx_sta(ar, skb);
394 	if (unlikely(!sta))
395 		goto out_rcu;
396 
397 	tid = get_tid_h(hdr);
398 
399 	sta_info = (void *) sta->drv_priv;
400 	tid_info = rcu_dereference(sta_info->agg[tid]);
401 	if (!tid_info)
402 		goto out_rcu;
403 
404 	spin_lock_bh(&tid_info->lock);
405 	if (likely(tid_info->state >= CARL9170_TID_STATE_IDLE))
406 		carl9170_tx_shift_bm(ar, tid_info, get_seq_h(hdr));
407 
408 	if (sta_info->stats[tid].clear) {
409 		sta_info->stats[tid].clear = false;
410 		sta_info->stats[tid].req = false;
411 		sta_info->stats[tid].ampdu_len = 0;
412 		sta_info->stats[tid].ampdu_ack_len = 0;
413 	}
414 
415 	sta_info->stats[tid].ampdu_len++;
416 	if (txinfo->status.rates[0].count == 1)
417 		sta_info->stats[tid].ampdu_ack_len++;
418 
419 	if (!(txinfo->flags & IEEE80211_TX_STAT_ACK))
420 		sta_info->stats[tid].req = true;
421 
422 	if (super->f.mac_control & cpu_to_le16(AR9170_TX_MAC_IMM_BA)) {
423 		super->s.rix = sta_info->stats[tid].ampdu_len;
424 		super->s.cnt = sta_info->stats[tid].ampdu_ack_len;
425 		txinfo->flags |= IEEE80211_TX_STAT_AMPDU;
426 		if (sta_info->stats[tid].req)
427 			txinfo->flags |= IEEE80211_TX_STAT_AMPDU_NO_BACK;
428 
429 		sta_info->stats[tid].clear = true;
430 	}
431 	spin_unlock_bh(&tid_info->lock);
432 
433 out_rcu:
434 	rcu_read_unlock();
435 }
436 
437 static void carl9170_tx_bar_status(struct ar9170 *ar, struct sk_buff *skb,
438 	struct ieee80211_tx_info *tx_info)
439 {
440 	struct _carl9170_tx_superframe *super = (void *) skb->data;
441 	struct ieee80211_bar *bar = (void *) super->frame_data;
442 
443 	/*
444 	 * Unlike all other frames, the status report for BARs does
445 	 * not directly come from the hardware as it is incapable of
446 	 * matching a BA to a previously send BAR.
447 	 * Instead the RX-path will scan for incoming BAs and set the
448 	 * IEEE80211_TX_STAT_ACK if it sees one that was likely
449 	 * caused by a BAR from us.
450 	 */
451 
452 	if (unlikely(ieee80211_is_back_req(bar->frame_control)) &&
453 	   !(tx_info->flags & IEEE80211_TX_STAT_ACK)) {
454 		struct carl9170_bar_list_entry *entry;
455 		int queue = skb_get_queue_mapping(skb);
456 
457 		rcu_read_lock();
458 		list_for_each_entry_rcu(entry, &ar->bar_list[queue], list) {
459 			if (entry->skb == skb) {
460 				spin_lock_bh(&ar->bar_list_lock[queue]);
461 				list_del_rcu(&entry->list);
462 				spin_unlock_bh(&ar->bar_list_lock[queue]);
463 				kfree_rcu(entry, head);
464 				goto out;
465 			}
466 		}
467 
468 		WARN(1, "bar not found in %d - ra:%pM ta:%pM c:%x ssn:%x\n",
469 		       queue, bar->ra, bar->ta, bar->control,
470 			bar->start_seq_num);
471 out:
472 		rcu_read_unlock();
473 	}
474 }
475 
476 void carl9170_tx_status(struct ar9170 *ar, struct sk_buff *skb,
477 			const bool success)
478 {
479 	struct ieee80211_tx_info *txinfo;
480 
481 	carl9170_tx_accounting_free(ar, skb);
482 
483 	txinfo = IEEE80211_SKB_CB(skb);
484 
485 	carl9170_tx_bar_status(ar, skb, txinfo);
486 
487 	if (success)
488 		txinfo->flags |= IEEE80211_TX_STAT_ACK;
489 	else
490 		ar->tx_ack_failures++;
491 
492 	if (txinfo->flags & IEEE80211_TX_CTL_AMPDU)
493 		carl9170_tx_status_process_ampdu(ar, skb, txinfo);
494 
495 	carl9170_tx_ps_unblock(ar, skb);
496 	carl9170_tx_put_skb(skb);
497 }
498 
499 /* This function may be called form any context */
500 void carl9170_tx_callback(struct ar9170 *ar, struct sk_buff *skb)
501 {
502 	struct ieee80211_tx_info *txinfo = IEEE80211_SKB_CB(skb);
503 
504 	atomic_dec(&ar->tx_total_pending);
505 
506 	if (txinfo->flags & IEEE80211_TX_CTL_AMPDU)
507 		atomic_dec(&ar->tx_ampdu_upload);
508 
509 	if (carl9170_tx_put_skb(skb))
510 		tasklet_hi_schedule(&ar->usb_tasklet);
511 }
512 
513 static struct sk_buff *carl9170_get_queued_skb(struct ar9170 *ar, u8 cookie,
514 					       struct sk_buff_head *queue)
515 {
516 	struct sk_buff *skb;
517 
518 	spin_lock_bh(&queue->lock);
519 	skb_queue_walk(queue, skb) {
520 		struct _carl9170_tx_superframe *txc = (void *) skb->data;
521 
522 		if (txc->s.cookie != cookie)
523 			continue;
524 
525 		__skb_unlink(skb, queue);
526 		spin_unlock_bh(&queue->lock);
527 
528 		carl9170_release_dev_space(ar, skb);
529 		return skb;
530 	}
531 	spin_unlock_bh(&queue->lock);
532 
533 	return NULL;
534 }
535 
536 static void carl9170_tx_fill_rateinfo(struct ar9170 *ar, unsigned int rix,
537 	unsigned int tries, struct ieee80211_tx_info *txinfo)
538 {
539 	unsigned int i;
540 
541 	for (i = 0; i < IEEE80211_TX_MAX_RATES; i++) {
542 		if (txinfo->status.rates[i].idx < 0)
543 			break;
544 
545 		if (i == rix) {
546 			txinfo->status.rates[i].count = tries;
547 			i++;
548 			break;
549 		}
550 	}
551 
552 	for (; i < IEEE80211_TX_MAX_RATES; i++) {
553 		txinfo->status.rates[i].idx = -1;
554 		txinfo->status.rates[i].count = 0;
555 	}
556 }
557 
558 static void carl9170_check_queue_stop_timeout(struct ar9170 *ar)
559 {
560 	int i;
561 	struct sk_buff *skb;
562 	struct ieee80211_tx_info *txinfo;
563 	struct carl9170_tx_info *arinfo;
564 	bool restart = false;
565 
566 	for (i = 0; i < ar->hw->queues; i++) {
567 		spin_lock_bh(&ar->tx_status[i].lock);
568 
569 		skb = skb_peek(&ar->tx_status[i]);
570 
571 		if (!skb)
572 			goto next;
573 
574 		txinfo = IEEE80211_SKB_CB(skb);
575 		arinfo = (void *) txinfo->rate_driver_data;
576 
577 		if (time_is_before_jiffies(arinfo->timeout +
578 		    msecs_to_jiffies(CARL9170_QUEUE_STUCK_TIMEOUT)) == true)
579 			restart = true;
580 
581 next:
582 		spin_unlock_bh(&ar->tx_status[i].lock);
583 	}
584 
585 	if (restart) {
586 		/*
587 		 * At least one queue has been stuck for long enough.
588 		 * Give the device a kick and hope it gets back to
589 		 * work.
590 		 *
591 		 * possible reasons may include:
592 		 *  - frames got lost/corrupted (bad connection to the device)
593 		 *  - stalled rx processing/usb controller hiccups
594 		 *  - firmware errors/bugs
595 		 *  - every bug you can think of.
596 		 *  - all bugs you can't...
597 		 *  - ...
598 		 */
599 		carl9170_restart(ar, CARL9170_RR_STUCK_TX);
600 	}
601 }
602 
603 static void carl9170_tx_ampdu_timeout(struct ar9170 *ar)
604 {
605 	struct carl9170_sta_tid *iter;
606 	struct sk_buff *skb;
607 	struct ieee80211_tx_info *txinfo;
608 	struct carl9170_tx_info *arinfo;
609 	struct ieee80211_sta *sta;
610 
611 	rcu_read_lock();
612 	list_for_each_entry_rcu(iter, &ar->tx_ampdu_list, list) {
613 		if (iter->state < CARL9170_TID_STATE_IDLE)
614 			continue;
615 
616 		spin_lock_bh(&iter->lock);
617 		skb = skb_peek(&iter->queue);
618 		if (!skb)
619 			goto unlock;
620 
621 		txinfo = IEEE80211_SKB_CB(skb);
622 		arinfo = (void *)txinfo->rate_driver_data;
623 		if (time_is_after_jiffies(arinfo->timeout +
624 		    msecs_to_jiffies(CARL9170_QUEUE_TIMEOUT)))
625 			goto unlock;
626 
627 		sta = iter->sta;
628 		if (WARN_ON(!sta))
629 			goto unlock;
630 
631 		ieee80211_stop_tx_ba_session(sta, iter->tid);
632 unlock:
633 		spin_unlock_bh(&iter->lock);
634 
635 	}
636 	rcu_read_unlock();
637 }
638 
639 void carl9170_tx_janitor(struct work_struct *work)
640 {
641 	struct ar9170 *ar = container_of(work, struct ar9170,
642 					 tx_janitor.work);
643 	if (!IS_STARTED(ar))
644 		return;
645 
646 	ar->tx_janitor_last_run = jiffies;
647 
648 	carl9170_check_queue_stop_timeout(ar);
649 	carl9170_tx_ampdu_timeout(ar);
650 
651 	if (!atomic_read(&ar->tx_total_queued))
652 		return;
653 
654 	ieee80211_queue_delayed_work(ar->hw, &ar->tx_janitor,
655 		msecs_to_jiffies(CARL9170_TX_TIMEOUT));
656 }
657 
658 static void __carl9170_tx_process_status(struct ar9170 *ar,
659 	const uint8_t cookie, const uint8_t info)
660 {
661 	struct sk_buff *skb;
662 	struct ieee80211_tx_info *txinfo;
663 	unsigned int r, t, q;
664 	bool success = true;
665 
666 	q = ar9170_qmap[info & CARL9170_TX_STATUS_QUEUE];
667 
668 	skb = carl9170_get_queued_skb(ar, cookie, &ar->tx_status[q]);
669 	if (!skb) {
670 		/*
671 		 * We have lost the race to another thread.
672 		 */
673 
674 		return ;
675 	}
676 
677 	txinfo = IEEE80211_SKB_CB(skb);
678 
679 	if (!(info & CARL9170_TX_STATUS_SUCCESS))
680 		success = false;
681 
682 	r = (info & CARL9170_TX_STATUS_RIX) >> CARL9170_TX_STATUS_RIX_S;
683 	t = (info & CARL9170_TX_STATUS_TRIES) >> CARL9170_TX_STATUS_TRIES_S;
684 
685 	carl9170_tx_fill_rateinfo(ar, r, t, txinfo);
686 	carl9170_tx_status(ar, skb, success);
687 }
688 
689 void carl9170_tx_process_status(struct ar9170 *ar,
690 				const struct carl9170_rsp *cmd)
691 {
692 	unsigned int i;
693 
694 	for (i = 0;  i < cmd->hdr.ext; i++) {
695 		if (WARN_ON(i > ((cmd->hdr.len / 2) + 1))) {
696 			print_hex_dump_bytes("UU:", DUMP_PREFIX_NONE,
697 					     (void *) cmd, cmd->hdr.len + 4);
698 			break;
699 		}
700 
701 		__carl9170_tx_process_status(ar, cmd->_tx_status[i].cookie,
702 					     cmd->_tx_status[i].info);
703 	}
704 }
705 
706 static void carl9170_tx_rate_tpc_chains(struct ar9170 *ar,
707 	struct ieee80211_tx_info *info,	struct ieee80211_tx_rate *txrate,
708 	unsigned int *phyrate, unsigned int *tpc, unsigned int *chains)
709 {
710 	struct ieee80211_rate *rate = NULL;
711 	u8 *txpower;
712 	unsigned int idx;
713 
714 	idx = txrate->idx;
715 	*tpc = 0;
716 	*phyrate = 0;
717 
718 	if (txrate->flags & IEEE80211_TX_RC_MCS) {
719 		if (txrate->flags & IEEE80211_TX_RC_40_MHZ_WIDTH) {
720 			/* +1 dBm for HT40 */
721 			*tpc += 2;
722 
723 			if (info->band == IEEE80211_BAND_2GHZ)
724 				txpower = ar->power_2G_ht40;
725 			else
726 				txpower = ar->power_5G_ht40;
727 		} else {
728 			if (info->band == IEEE80211_BAND_2GHZ)
729 				txpower = ar->power_2G_ht20;
730 			else
731 				txpower = ar->power_5G_ht20;
732 		}
733 
734 		*phyrate = txrate->idx;
735 		*tpc += txpower[idx & 7];
736 	} else {
737 		if (info->band == IEEE80211_BAND_2GHZ) {
738 			if (idx < 4)
739 				txpower = ar->power_2G_cck;
740 			else
741 				txpower = ar->power_2G_ofdm;
742 		} else {
743 			txpower = ar->power_5G_leg;
744 			idx += 4;
745 		}
746 
747 		rate = &__carl9170_ratetable[idx];
748 		*tpc += txpower[(rate->hw_value & 0x30) >> 4];
749 		*phyrate = rate->hw_value & 0xf;
750 	}
751 
752 	if (ar->eeprom.tx_mask == 1) {
753 		*chains = AR9170_TX_PHY_TXCHAIN_1;
754 	} else {
755 		if (!(txrate->flags & IEEE80211_TX_RC_MCS) &&
756 		    rate && rate->bitrate >= 360)
757 			*chains = AR9170_TX_PHY_TXCHAIN_1;
758 		else
759 			*chains = AR9170_TX_PHY_TXCHAIN_2;
760 	}
761 
762 	*tpc = min_t(unsigned int, *tpc, ar->hw->conf.power_level * 2);
763 }
764 
765 static __le32 carl9170_tx_physet(struct ar9170 *ar,
766 	struct ieee80211_tx_info *info, struct ieee80211_tx_rate *txrate)
767 {
768 	unsigned int power = 0, chains = 0, phyrate = 0;
769 	__le32 tmp;
770 
771 	tmp = cpu_to_le32(0);
772 
773 	if (txrate->flags & IEEE80211_TX_RC_40_MHZ_WIDTH)
774 		tmp |= cpu_to_le32(AR9170_TX_PHY_BW_40MHZ <<
775 			AR9170_TX_PHY_BW_S);
776 	/* this works because 40 MHz is 2 and dup is 3 */
777 	if (txrate->flags & IEEE80211_TX_RC_DUP_DATA)
778 		tmp |= cpu_to_le32(AR9170_TX_PHY_BW_40MHZ_DUP <<
779 			AR9170_TX_PHY_BW_S);
780 
781 	if (txrate->flags & IEEE80211_TX_RC_SHORT_GI)
782 		tmp |= cpu_to_le32(AR9170_TX_PHY_SHORT_GI);
783 
784 	if (txrate->flags & IEEE80211_TX_RC_MCS) {
785 		SET_VAL(AR9170_TX_PHY_MCS, phyrate, txrate->idx);
786 
787 		/* heavy clip control */
788 		tmp |= cpu_to_le32((txrate->idx & 0x7) <<
789 			AR9170_TX_PHY_TX_HEAVY_CLIP_S);
790 
791 		tmp |= cpu_to_le32(AR9170_TX_PHY_MOD_HT);
792 
793 		/*
794 		 * green field preamble does not work.
795 		 *
796 		 * if (txrate->flags & IEEE80211_TX_RC_GREEN_FIELD)
797 		 * tmp |= cpu_to_le32(AR9170_TX_PHY_GREENFIELD);
798 		 */
799 	} else {
800 		if (info->band == IEEE80211_BAND_2GHZ) {
801 			if (txrate->idx <= AR9170_TX_PHY_RATE_CCK_11M)
802 				tmp |= cpu_to_le32(AR9170_TX_PHY_MOD_CCK);
803 			else
804 				tmp |= cpu_to_le32(AR9170_TX_PHY_MOD_OFDM);
805 		} else {
806 			tmp |= cpu_to_le32(AR9170_TX_PHY_MOD_OFDM);
807 		}
808 
809 		/*
810 		 * short preamble seems to be broken too.
811 		 *
812 		 * if (txrate->flags & IEEE80211_TX_RC_USE_SHORT_PREAMBLE)
813 		 *	tmp |= cpu_to_le32(AR9170_TX_PHY_SHORT_PREAMBLE);
814 		 */
815 	}
816 	carl9170_tx_rate_tpc_chains(ar, info, txrate,
817 				    &phyrate, &power, &chains);
818 
819 	tmp |= cpu_to_le32(SET_CONSTVAL(AR9170_TX_PHY_MCS, phyrate));
820 	tmp |= cpu_to_le32(SET_CONSTVAL(AR9170_TX_PHY_TX_PWR, power));
821 	tmp |= cpu_to_le32(SET_CONSTVAL(AR9170_TX_PHY_TXCHAIN, chains));
822 	return tmp;
823 }
824 
825 static bool carl9170_tx_rts_check(struct ar9170 *ar,
826 				  struct ieee80211_tx_rate *rate,
827 				  bool ampdu, bool multi)
828 {
829 	switch (ar->erp_mode) {
830 	case CARL9170_ERP_AUTO:
831 		if (ampdu)
832 			break;
833 
834 	case CARL9170_ERP_MAC80211:
835 		if (!(rate->flags & IEEE80211_TX_RC_USE_RTS_CTS))
836 			break;
837 
838 	case CARL9170_ERP_RTS:
839 		if (likely(!multi))
840 			return true;
841 
842 	default:
843 		break;
844 	}
845 
846 	return false;
847 }
848 
849 static bool carl9170_tx_cts_check(struct ar9170 *ar,
850 				  struct ieee80211_tx_rate *rate)
851 {
852 	switch (ar->erp_mode) {
853 	case CARL9170_ERP_AUTO:
854 	case CARL9170_ERP_MAC80211:
855 		if (!(rate->flags & IEEE80211_TX_RC_USE_CTS_PROTECT))
856 			break;
857 
858 	case CARL9170_ERP_CTS:
859 		return true;
860 
861 	default:
862 		break;
863 	}
864 
865 	return false;
866 }
867 
868 static void carl9170_tx_get_rates(struct ar9170 *ar,
869 				  struct ieee80211_vif *vif,
870 				  struct ieee80211_sta *sta,
871 				  struct sk_buff *skb)
872 {
873 	struct ieee80211_tx_info *info;
874 
875 	BUILD_BUG_ON(IEEE80211_TX_MAX_RATES < CARL9170_TX_MAX_RATES);
876 	BUILD_BUG_ON(IEEE80211_TX_MAX_RATES > IEEE80211_TX_RATE_TABLE_SIZE);
877 
878 	info = IEEE80211_SKB_CB(skb);
879 
880 	ieee80211_get_tx_rates(vif, sta, skb,
881 			       info->control.rates,
882 			       IEEE80211_TX_MAX_RATES);
883 }
884 
885 static void carl9170_tx_apply_rateset(struct ar9170 *ar,
886 				      struct ieee80211_tx_info *sinfo,
887 				      struct sk_buff *skb)
888 {
889 	struct ieee80211_tx_rate *txrate;
890 	struct ieee80211_tx_info *info;
891 	struct _carl9170_tx_superframe *txc = (void *) skb->data;
892 	int i;
893 	bool ampdu;
894 	bool no_ack;
895 
896 	info = IEEE80211_SKB_CB(skb);
897 	ampdu = !!(info->flags & IEEE80211_TX_CTL_AMPDU);
898 	no_ack = !!(info->flags & IEEE80211_TX_CTL_NO_ACK);
899 
900 	/* Set the rate control probe flag for all (sub-) frames.
901 	 * This is because the TX_STATS_AMPDU flag is only set on
902 	 * the last frame, so it has to be inherited.
903 	 */
904 	info->flags |= (sinfo->flags & IEEE80211_TX_CTL_RATE_CTRL_PROBE);
905 
906 	/* NOTE: For the first rate, the ERP & AMPDU flags are directly
907 	 * taken from mac_control. For all fallback rate, the firmware
908 	 * updates the mac_control flags from the rate info field.
909 	 */
910 	for (i = 0; i < CARL9170_TX_MAX_RATES; i++) {
911 		__le32 phy_set;
912 
913 		txrate = &sinfo->control.rates[i];
914 		if (txrate->idx < 0)
915 			break;
916 
917 		phy_set = carl9170_tx_physet(ar, info, txrate);
918 		if (i == 0) {
919 			__le16 mac_tmp = cpu_to_le16(0);
920 
921 			/* first rate - part of the hw's frame header */
922 			txc->f.phy_control = phy_set;
923 
924 			if (ampdu && txrate->flags & IEEE80211_TX_RC_MCS)
925 				mac_tmp |= cpu_to_le16(AR9170_TX_MAC_AGGR);
926 
927 			if (carl9170_tx_rts_check(ar, txrate, ampdu, no_ack))
928 				mac_tmp |= cpu_to_le16(AR9170_TX_MAC_PROT_RTS);
929 			else if (carl9170_tx_cts_check(ar, txrate))
930 				mac_tmp |= cpu_to_le16(AR9170_TX_MAC_PROT_CTS);
931 
932 			txc->f.mac_control |= mac_tmp;
933 		} else {
934 			/* fallback rates are stored in the firmware's
935 			 * retry rate set array.
936 			 */
937 			txc->s.rr[i - 1] = phy_set;
938 		}
939 
940 		SET_VAL(CARL9170_TX_SUPER_RI_TRIES, txc->s.ri[i],
941 			txrate->count);
942 
943 		if (carl9170_tx_rts_check(ar, txrate, ampdu, no_ack))
944 			txc->s.ri[i] |= (AR9170_TX_MAC_PROT_RTS <<
945 				CARL9170_TX_SUPER_RI_ERP_PROT_S);
946 		else if (carl9170_tx_cts_check(ar, txrate))
947 			txc->s.ri[i] |= (AR9170_TX_MAC_PROT_CTS <<
948 				CARL9170_TX_SUPER_RI_ERP_PROT_S);
949 
950 		if (ampdu && (txrate->flags & IEEE80211_TX_RC_MCS))
951 			txc->s.ri[i] |= CARL9170_TX_SUPER_RI_AMPDU;
952 	}
953 }
954 
955 static int carl9170_tx_prepare(struct ar9170 *ar,
956 			       struct ieee80211_sta *sta,
957 			       struct sk_buff *skb)
958 {
959 	struct ieee80211_hdr *hdr;
960 	struct _carl9170_tx_superframe *txc;
961 	struct carl9170_vif_info *cvif;
962 	struct ieee80211_tx_info *info;
963 	struct carl9170_tx_info *arinfo;
964 	unsigned int hw_queue;
965 	__le16 mac_tmp;
966 	u16 len;
967 
968 	BUILD_BUG_ON(sizeof(*arinfo) > sizeof(info->rate_driver_data));
969 	BUILD_BUG_ON(sizeof(struct _carl9170_tx_superdesc) !=
970 		     CARL9170_TX_SUPERDESC_LEN);
971 
972 	BUILD_BUG_ON(sizeof(struct _ar9170_tx_hwdesc) !=
973 		     AR9170_TX_HWDESC_LEN);
974 
975 	BUILD_BUG_ON(AR9170_MAX_VIRTUAL_MAC >
976 		((CARL9170_TX_SUPER_MISC_VIF_ID >>
977 		 CARL9170_TX_SUPER_MISC_VIF_ID_S) + 1));
978 
979 	hw_queue = ar9170_qmap[carl9170_get_queue(ar, skb)];
980 
981 	hdr = (void *)skb->data;
982 	info = IEEE80211_SKB_CB(skb);
983 	len = skb->len;
984 
985 	/*
986 	 * Note: If the frame was sent through a monitor interface,
987 	 * the ieee80211_vif pointer can be NULL.
988 	 */
989 	if (likely(info->control.vif))
990 		cvif = (void *) info->control.vif->drv_priv;
991 	else
992 		cvif = NULL;
993 
994 	txc = (void *)skb_push(skb, sizeof(*txc));
995 	memset(txc, 0, sizeof(*txc));
996 
997 	SET_VAL(CARL9170_TX_SUPER_MISC_QUEUE, txc->s.misc, hw_queue);
998 
999 	if (likely(cvif))
1000 		SET_VAL(CARL9170_TX_SUPER_MISC_VIF_ID, txc->s.misc, cvif->id);
1001 
1002 	if (unlikely(info->flags & IEEE80211_TX_CTL_SEND_AFTER_DTIM))
1003 		txc->s.misc |= CARL9170_TX_SUPER_MISC_CAB;
1004 
1005 	if (unlikely(info->flags & IEEE80211_TX_CTL_ASSIGN_SEQ))
1006 		txc->s.misc |= CARL9170_TX_SUPER_MISC_ASSIGN_SEQ;
1007 
1008 	if (unlikely(ieee80211_is_probe_resp(hdr->frame_control)))
1009 		txc->s.misc |= CARL9170_TX_SUPER_MISC_FILL_IN_TSF;
1010 
1011 	mac_tmp = cpu_to_le16(AR9170_TX_MAC_HW_DURATION |
1012 			      AR9170_TX_MAC_BACKOFF);
1013 	mac_tmp |= cpu_to_le16((hw_queue << AR9170_TX_MAC_QOS_S) &
1014 			       AR9170_TX_MAC_QOS);
1015 
1016 	if (unlikely(info->flags & IEEE80211_TX_CTL_NO_ACK))
1017 		mac_tmp |= cpu_to_le16(AR9170_TX_MAC_NO_ACK);
1018 
1019 	if (info->control.hw_key) {
1020 		len += info->control.hw_key->icv_len;
1021 
1022 		switch (info->control.hw_key->cipher) {
1023 		case WLAN_CIPHER_SUITE_WEP40:
1024 		case WLAN_CIPHER_SUITE_WEP104:
1025 		case WLAN_CIPHER_SUITE_TKIP:
1026 			mac_tmp |= cpu_to_le16(AR9170_TX_MAC_ENCR_RC4);
1027 			break;
1028 		case WLAN_CIPHER_SUITE_CCMP:
1029 			mac_tmp |= cpu_to_le16(AR9170_TX_MAC_ENCR_AES);
1030 			break;
1031 		default:
1032 			WARN_ON(1);
1033 			goto err_out;
1034 		}
1035 	}
1036 
1037 	if (info->flags & IEEE80211_TX_CTL_AMPDU) {
1038 		unsigned int density, factor;
1039 
1040 		if (unlikely(!sta || !cvif))
1041 			goto err_out;
1042 
1043 		factor = min_t(unsigned int, 1u, sta->ht_cap.ampdu_factor);
1044 		density = sta->ht_cap.ampdu_density;
1045 
1046 		if (density) {
1047 			/*
1048 			 * Watch out!
1049 			 *
1050 			 * Otus uses slightly different density values than
1051 			 * those from the 802.11n spec.
1052 			 */
1053 
1054 			density = max_t(unsigned int, density + 1, 7u);
1055 		}
1056 
1057 		SET_VAL(CARL9170_TX_SUPER_AMPDU_DENSITY,
1058 			txc->s.ampdu_settings, density);
1059 
1060 		SET_VAL(CARL9170_TX_SUPER_AMPDU_FACTOR,
1061 			txc->s.ampdu_settings, factor);
1062 	}
1063 
1064 	txc->s.len = cpu_to_le16(skb->len);
1065 	txc->f.length = cpu_to_le16(len + FCS_LEN);
1066 	txc->f.mac_control = mac_tmp;
1067 
1068 	arinfo = (void *)info->rate_driver_data;
1069 	arinfo->timeout = jiffies;
1070 	arinfo->ar = ar;
1071 	kref_init(&arinfo->ref);
1072 	return 0;
1073 
1074 err_out:
1075 	skb_pull(skb, sizeof(*txc));
1076 	return -EINVAL;
1077 }
1078 
1079 static void carl9170_set_immba(struct ar9170 *ar, struct sk_buff *skb)
1080 {
1081 	struct _carl9170_tx_superframe *super;
1082 
1083 	super = (void *) skb->data;
1084 	super->f.mac_control |= cpu_to_le16(AR9170_TX_MAC_IMM_BA);
1085 }
1086 
1087 static void carl9170_set_ampdu_params(struct ar9170 *ar, struct sk_buff *skb)
1088 {
1089 	struct _carl9170_tx_superframe *super;
1090 	int tmp;
1091 
1092 	super = (void *) skb->data;
1093 
1094 	tmp = (super->s.ampdu_settings & CARL9170_TX_SUPER_AMPDU_DENSITY) <<
1095 		CARL9170_TX_SUPER_AMPDU_DENSITY_S;
1096 
1097 	/*
1098 	 * If you haven't noticed carl9170_tx_prepare has already filled
1099 	 * in all ampdu spacing & factor parameters.
1100 	 * Now it's the time to check whenever the settings have to be
1101 	 * updated by the firmware, or if everything is still the same.
1102 	 *
1103 	 * There's no sane way to handle different density values with
1104 	 * this hardware, so we may as well just do the compare in the
1105 	 * driver.
1106 	 */
1107 
1108 	if (tmp != ar->current_density) {
1109 		ar->current_density = tmp;
1110 		super->s.ampdu_settings |=
1111 			CARL9170_TX_SUPER_AMPDU_COMMIT_DENSITY;
1112 	}
1113 
1114 	tmp = (super->s.ampdu_settings & CARL9170_TX_SUPER_AMPDU_FACTOR) <<
1115 		CARL9170_TX_SUPER_AMPDU_FACTOR_S;
1116 
1117 	if (tmp != ar->current_factor) {
1118 		ar->current_factor = tmp;
1119 		super->s.ampdu_settings |=
1120 			CARL9170_TX_SUPER_AMPDU_COMMIT_FACTOR;
1121 	}
1122 }
1123 
1124 static void carl9170_tx_ampdu(struct ar9170 *ar)
1125 {
1126 	struct sk_buff_head agg;
1127 	struct carl9170_sta_tid *tid_info;
1128 	struct sk_buff *skb, *first;
1129 	struct ieee80211_tx_info *tx_info_first;
1130 	unsigned int i = 0, done_ampdus = 0;
1131 	u16 seq, queue, tmpssn;
1132 
1133 	atomic_inc(&ar->tx_ampdu_scheduler);
1134 	ar->tx_ampdu_schedule = false;
1135 
1136 	if (atomic_read(&ar->tx_ampdu_upload))
1137 		return;
1138 
1139 	if (!ar->tx_ampdu_list_len)
1140 		return;
1141 
1142 	__skb_queue_head_init(&agg);
1143 
1144 	rcu_read_lock();
1145 	tid_info = rcu_dereference(ar->tx_ampdu_iter);
1146 	if (WARN_ON_ONCE(!tid_info)) {
1147 		rcu_read_unlock();
1148 		return;
1149 	}
1150 
1151 retry:
1152 	list_for_each_entry_continue_rcu(tid_info, &ar->tx_ampdu_list, list) {
1153 		i++;
1154 
1155 		if (tid_info->state < CARL9170_TID_STATE_PROGRESS)
1156 			continue;
1157 
1158 		queue = TID_TO_WME_AC(tid_info->tid);
1159 
1160 		spin_lock_bh(&tid_info->lock);
1161 		if (tid_info->state != CARL9170_TID_STATE_XMIT)
1162 			goto processed;
1163 
1164 		tid_info->counter++;
1165 		first = skb_peek(&tid_info->queue);
1166 		tmpssn = carl9170_get_seq(first);
1167 		seq = tid_info->snx;
1168 
1169 		if (unlikely(tmpssn != seq)) {
1170 			tid_info->state = CARL9170_TID_STATE_IDLE;
1171 
1172 			goto processed;
1173 		}
1174 
1175 		tx_info_first = NULL;
1176 		while ((skb = skb_peek(&tid_info->queue))) {
1177 			/* strict 0, 1, ..., n - 1, n frame sequence order */
1178 			if (unlikely(carl9170_get_seq(skb) != seq))
1179 				break;
1180 
1181 			/* don't upload more than AMPDU FACTOR allows. */
1182 			if (unlikely(SEQ_DIFF(tid_info->snx, tid_info->bsn) >=
1183 			    (tid_info->max - 1)))
1184 				break;
1185 
1186 			if (!tx_info_first) {
1187 				carl9170_tx_get_rates(ar, tid_info->vif,
1188 						      tid_info->sta, first);
1189 				tx_info_first = IEEE80211_SKB_CB(first);
1190 			}
1191 
1192 			carl9170_tx_apply_rateset(ar, tx_info_first, skb);
1193 
1194 			atomic_inc(&ar->tx_ampdu_upload);
1195 			tid_info->snx = seq = SEQ_NEXT(seq);
1196 			__skb_unlink(skb, &tid_info->queue);
1197 
1198 			__skb_queue_tail(&agg, skb);
1199 
1200 			if (skb_queue_len(&agg) >= CARL9170_NUM_TX_AGG_MAX)
1201 				break;
1202 		}
1203 
1204 		if (skb_queue_empty(&tid_info->queue) ||
1205 		    carl9170_get_seq(skb_peek(&tid_info->queue)) !=
1206 		    tid_info->snx) {
1207 			/* stop TID, if A-MPDU frames are still missing,
1208 			 * or whenever the queue is empty.
1209 			 */
1210 
1211 			tid_info->state = CARL9170_TID_STATE_IDLE;
1212 		}
1213 		done_ampdus++;
1214 
1215 processed:
1216 		spin_unlock_bh(&tid_info->lock);
1217 
1218 		if (skb_queue_empty(&agg))
1219 			continue;
1220 
1221 		/* apply ampdu spacing & factor settings */
1222 		carl9170_set_ampdu_params(ar, skb_peek(&agg));
1223 
1224 		/* set aggregation push bit */
1225 		carl9170_set_immba(ar, skb_peek_tail(&agg));
1226 
1227 		spin_lock_bh(&ar->tx_pending[queue].lock);
1228 		skb_queue_splice_tail_init(&agg, &ar->tx_pending[queue]);
1229 		spin_unlock_bh(&ar->tx_pending[queue].lock);
1230 		ar->tx_schedule = true;
1231 	}
1232 	if ((done_ampdus++ == 0) && (i++ == 0))
1233 		goto retry;
1234 
1235 	rcu_assign_pointer(ar->tx_ampdu_iter, tid_info);
1236 	rcu_read_unlock();
1237 }
1238 
1239 static struct sk_buff *carl9170_tx_pick_skb(struct ar9170 *ar,
1240 					    struct sk_buff_head *queue)
1241 {
1242 	struct sk_buff *skb;
1243 	struct ieee80211_tx_info *info;
1244 	struct carl9170_tx_info *arinfo;
1245 
1246 	BUILD_BUG_ON(sizeof(*arinfo) > sizeof(info->rate_driver_data));
1247 
1248 	spin_lock_bh(&queue->lock);
1249 	skb = skb_peek(queue);
1250 	if (unlikely(!skb))
1251 		goto err_unlock;
1252 
1253 	if (carl9170_alloc_dev_space(ar, skb))
1254 		goto err_unlock;
1255 
1256 	__skb_unlink(skb, queue);
1257 	spin_unlock_bh(&queue->lock);
1258 
1259 	info = IEEE80211_SKB_CB(skb);
1260 	arinfo = (void *) info->rate_driver_data;
1261 
1262 	arinfo->timeout = jiffies;
1263 	return skb;
1264 
1265 err_unlock:
1266 	spin_unlock_bh(&queue->lock);
1267 	return NULL;
1268 }
1269 
1270 void carl9170_tx_drop(struct ar9170 *ar, struct sk_buff *skb)
1271 {
1272 	struct _carl9170_tx_superframe *super;
1273 	uint8_t q = 0;
1274 
1275 	ar->tx_dropped++;
1276 
1277 	super = (void *)skb->data;
1278 	SET_VAL(CARL9170_TX_SUPER_MISC_QUEUE, q,
1279 		ar9170_qmap[carl9170_get_queue(ar, skb)]);
1280 	__carl9170_tx_process_status(ar, super->s.cookie, q);
1281 }
1282 
1283 static bool carl9170_tx_ps_drop(struct ar9170 *ar, struct sk_buff *skb)
1284 {
1285 	struct ieee80211_sta *sta;
1286 	struct carl9170_sta_info *sta_info;
1287 	struct ieee80211_tx_info *tx_info;
1288 
1289 	rcu_read_lock();
1290 	sta = __carl9170_get_tx_sta(ar, skb);
1291 	if (!sta)
1292 		goto out_rcu;
1293 
1294 	sta_info = (void *) sta->drv_priv;
1295 	tx_info = IEEE80211_SKB_CB(skb);
1296 
1297 	if (unlikely(sta_info->sleeping) &&
1298 	    !(tx_info->flags & (IEEE80211_TX_CTL_NO_PS_BUFFER |
1299 				IEEE80211_TX_CTL_CLEAR_PS_FILT))) {
1300 		rcu_read_unlock();
1301 
1302 		if (tx_info->flags & IEEE80211_TX_CTL_AMPDU)
1303 			atomic_dec(&ar->tx_ampdu_upload);
1304 
1305 		tx_info->flags |= IEEE80211_TX_STAT_TX_FILTERED;
1306 		carl9170_release_dev_space(ar, skb);
1307 		carl9170_tx_status(ar, skb, false);
1308 		return true;
1309 	}
1310 
1311 out_rcu:
1312 	rcu_read_unlock();
1313 	return false;
1314 }
1315 
1316 static void carl9170_bar_check(struct ar9170 *ar, struct sk_buff *skb)
1317 {
1318 	struct _carl9170_tx_superframe *super = (void *) skb->data;
1319 	struct ieee80211_bar *bar = (void *) super->frame_data;
1320 
1321 	if (unlikely(ieee80211_is_back_req(bar->frame_control)) &&
1322 	    skb->len >= sizeof(struct ieee80211_bar)) {
1323 		struct carl9170_bar_list_entry *entry;
1324 		unsigned int queue = skb_get_queue_mapping(skb);
1325 
1326 		entry = kmalloc(sizeof(*entry), GFP_ATOMIC);
1327 		if (!WARN_ON_ONCE(!entry)) {
1328 			entry->skb = skb;
1329 			spin_lock_bh(&ar->bar_list_lock[queue]);
1330 			list_add_tail_rcu(&entry->list, &ar->bar_list[queue]);
1331 			spin_unlock_bh(&ar->bar_list_lock[queue]);
1332 		}
1333 	}
1334 }
1335 
1336 static void carl9170_tx(struct ar9170 *ar)
1337 {
1338 	struct sk_buff *skb;
1339 	unsigned int i, q;
1340 	bool schedule_garbagecollector = false;
1341 
1342 	ar->tx_schedule = false;
1343 
1344 	if (unlikely(!IS_STARTED(ar)))
1345 		return;
1346 
1347 	carl9170_usb_handle_tx_err(ar);
1348 
1349 	for (i = 0; i < ar->hw->queues; i++) {
1350 		while (!skb_queue_empty(&ar->tx_pending[i])) {
1351 			skb = carl9170_tx_pick_skb(ar, &ar->tx_pending[i]);
1352 			if (unlikely(!skb))
1353 				break;
1354 
1355 			if (unlikely(carl9170_tx_ps_drop(ar, skb)))
1356 				continue;
1357 
1358 			carl9170_bar_check(ar, skb);
1359 
1360 			atomic_inc(&ar->tx_total_pending);
1361 
1362 			q = __carl9170_get_queue(ar, i);
1363 			/*
1364 			 * NB: tx_status[i] vs. tx_status[q],
1365 			 * TODO: Move into pick_skb or alloc_dev_space.
1366 			 */
1367 			skb_queue_tail(&ar->tx_status[q], skb);
1368 
1369 			/*
1370 			 * increase ref count to "2".
1371 			 * Ref counting is the easiest way to solve the
1372 			 * race between the urb's completion routine:
1373 			 *	carl9170_tx_callback
1374 			 * and wlan tx status functions:
1375 			 *	carl9170_tx_status/janitor.
1376 			 */
1377 			carl9170_tx_get_skb(skb);
1378 
1379 			carl9170_usb_tx(ar, skb);
1380 			schedule_garbagecollector = true;
1381 		}
1382 	}
1383 
1384 	if (!schedule_garbagecollector)
1385 		return;
1386 
1387 	ieee80211_queue_delayed_work(ar->hw, &ar->tx_janitor,
1388 		msecs_to_jiffies(CARL9170_TX_TIMEOUT));
1389 }
1390 
1391 static bool carl9170_tx_ampdu_queue(struct ar9170 *ar,
1392 	struct ieee80211_sta *sta, struct sk_buff *skb,
1393 	struct ieee80211_tx_info *txinfo)
1394 {
1395 	struct carl9170_sta_info *sta_info;
1396 	struct carl9170_sta_tid *agg;
1397 	struct sk_buff *iter;
1398 	u16 tid, seq, qseq, off;
1399 	bool run = false;
1400 
1401 	tid = carl9170_get_tid(skb);
1402 	seq = carl9170_get_seq(skb);
1403 	sta_info = (void *) sta->drv_priv;
1404 
1405 	rcu_read_lock();
1406 	agg = rcu_dereference(sta_info->agg[tid]);
1407 
1408 	if (!agg)
1409 		goto err_unlock_rcu;
1410 
1411 	spin_lock_bh(&agg->lock);
1412 	if (unlikely(agg->state < CARL9170_TID_STATE_IDLE))
1413 		goto err_unlock;
1414 
1415 	/* check if sequence is within the BA window */
1416 	if (unlikely(!BAW_WITHIN(agg->bsn, CARL9170_BAW_BITS, seq)))
1417 		goto err_unlock;
1418 
1419 	if (WARN_ON_ONCE(!BAW_WITHIN(agg->snx, CARL9170_BAW_BITS, seq)))
1420 		goto err_unlock;
1421 
1422 	off = SEQ_DIFF(seq, agg->bsn);
1423 	if (WARN_ON_ONCE(test_and_set_bit(off, agg->bitmap)))
1424 		goto err_unlock;
1425 
1426 	if (likely(BAW_WITHIN(agg->hsn, CARL9170_BAW_BITS, seq))) {
1427 		__skb_queue_tail(&agg->queue, skb);
1428 		agg->hsn = seq;
1429 		goto queued;
1430 	}
1431 
1432 	skb_queue_reverse_walk(&agg->queue, iter) {
1433 		qseq = carl9170_get_seq(iter);
1434 
1435 		if (BAW_WITHIN(qseq, CARL9170_BAW_BITS, seq)) {
1436 			__skb_queue_after(&agg->queue, iter, skb);
1437 			goto queued;
1438 		}
1439 	}
1440 
1441 	__skb_queue_head(&agg->queue, skb);
1442 queued:
1443 
1444 	if (unlikely(agg->state != CARL9170_TID_STATE_XMIT)) {
1445 		if (agg->snx == carl9170_get_seq(skb_peek(&agg->queue))) {
1446 			agg->state = CARL9170_TID_STATE_XMIT;
1447 			run = true;
1448 		}
1449 	}
1450 
1451 	spin_unlock_bh(&agg->lock);
1452 	rcu_read_unlock();
1453 
1454 	return run;
1455 
1456 err_unlock:
1457 	spin_unlock_bh(&agg->lock);
1458 
1459 err_unlock_rcu:
1460 	rcu_read_unlock();
1461 	txinfo->flags &= ~IEEE80211_TX_CTL_AMPDU;
1462 	carl9170_tx_status(ar, skb, false);
1463 	ar->tx_dropped++;
1464 	return false;
1465 }
1466 
1467 void carl9170_op_tx(struct ieee80211_hw *hw,
1468 		    struct ieee80211_tx_control *control,
1469 		    struct sk_buff *skb)
1470 {
1471 	struct ar9170 *ar = hw->priv;
1472 	struct ieee80211_tx_info *info;
1473 	struct ieee80211_sta *sta = control->sta;
1474 	struct ieee80211_vif *vif;
1475 	bool run;
1476 
1477 	if (unlikely(!IS_STARTED(ar)))
1478 		goto err_free;
1479 
1480 	info = IEEE80211_SKB_CB(skb);
1481 	vif = info->control.vif;
1482 
1483 	if (unlikely(carl9170_tx_prepare(ar, sta, skb)))
1484 		goto err_free;
1485 
1486 	carl9170_tx_accounting(ar, skb);
1487 	/*
1488 	 * from now on, one has to use carl9170_tx_status to free
1489 	 * all ressouces which are associated with the frame.
1490 	 */
1491 
1492 	if (sta) {
1493 		struct carl9170_sta_info *stai = (void *) sta->drv_priv;
1494 		atomic_inc(&stai->pending_frames);
1495 	}
1496 
1497 	if (info->flags & IEEE80211_TX_CTL_AMPDU) {
1498 		/* to static code analyzers and reviewers:
1499 		 * mac80211 guarantees that a valid "sta"
1500 		 * reference is present, if a frame is to
1501 		 * be part of an ampdu. Hence any extra
1502 		 * sta == NULL checks are redundant in this
1503 		 * special case.
1504 		 */
1505 		run = carl9170_tx_ampdu_queue(ar, sta, skb, info);
1506 		if (run)
1507 			carl9170_tx_ampdu(ar);
1508 
1509 	} else {
1510 		unsigned int queue = skb_get_queue_mapping(skb);
1511 
1512 		carl9170_tx_get_rates(ar, vif, sta, skb);
1513 		carl9170_tx_apply_rateset(ar, info, skb);
1514 		skb_queue_tail(&ar->tx_pending[queue], skb);
1515 	}
1516 
1517 	carl9170_tx(ar);
1518 	return;
1519 
1520 err_free:
1521 	ar->tx_dropped++;
1522 	ieee80211_free_txskb(ar->hw, skb);
1523 }
1524 
1525 void carl9170_tx_scheduler(struct ar9170 *ar)
1526 {
1527 
1528 	if (ar->tx_ampdu_schedule)
1529 		carl9170_tx_ampdu(ar);
1530 
1531 	if (ar->tx_schedule)
1532 		carl9170_tx(ar);
1533 }
1534 
1535 /* caller has to take rcu_read_lock */
1536 static struct carl9170_vif_info *carl9170_pick_beaconing_vif(struct ar9170 *ar)
1537 {
1538 	struct carl9170_vif_info *cvif;
1539 	int i = 1;
1540 
1541 	/* The AR9170 hardware has no fancy beacon queue or some
1542 	 * other scheduling mechanism. So, the driver has to make
1543 	 * due by setting the two beacon timers (pretbtt and tbtt)
1544 	 * once and then swapping the beacon address in the HW's
1545 	 * register file each time the pretbtt fires.
1546 	 */
1547 
1548 	cvif = rcu_dereference(ar->beacon_iter);
1549 	if (ar->vifs > 0 && cvif) {
1550 		do {
1551 			list_for_each_entry_continue_rcu(cvif, &ar->vif_list,
1552 							 list) {
1553 				if (cvif->active && cvif->enable_beacon)
1554 					goto out;
1555 			}
1556 		} while (ar->beacon_enabled && i--);
1557 	}
1558 
1559 out:
1560 	RCU_INIT_POINTER(ar->beacon_iter, cvif);
1561 	return cvif;
1562 }
1563 
1564 static bool carl9170_tx_beacon_physet(struct ar9170 *ar, struct sk_buff *skb,
1565 				      u32 *ht1, u32 *plcp)
1566 {
1567 	struct ieee80211_tx_info *txinfo;
1568 	struct ieee80211_tx_rate *rate;
1569 	unsigned int power, chains;
1570 	bool ht_rate;
1571 
1572 	txinfo = IEEE80211_SKB_CB(skb);
1573 	rate = &txinfo->control.rates[0];
1574 	ht_rate = !!(txinfo->control.rates[0].flags & IEEE80211_TX_RC_MCS);
1575 	carl9170_tx_rate_tpc_chains(ar, txinfo, rate, plcp, &power, &chains);
1576 
1577 	*ht1 = AR9170_MAC_BCN_HT1_TX_ANT0;
1578 	if (chains == AR9170_TX_PHY_TXCHAIN_2)
1579 		*ht1 |= AR9170_MAC_BCN_HT1_TX_ANT1;
1580 	SET_VAL(AR9170_MAC_BCN_HT1_PWR_CTRL, *ht1, 7);
1581 	SET_VAL(AR9170_MAC_BCN_HT1_TPC, *ht1, power);
1582 	SET_VAL(AR9170_MAC_BCN_HT1_CHAIN_MASK, *ht1, chains);
1583 
1584 	if (ht_rate) {
1585 		*ht1 |= AR9170_MAC_BCN_HT1_HT_EN;
1586 		if (rate->flags & IEEE80211_TX_RC_SHORT_GI)
1587 			*plcp |= AR9170_MAC_BCN_HT2_SGI;
1588 
1589 		if (rate->flags & IEEE80211_TX_RC_40_MHZ_WIDTH) {
1590 			*ht1 |= AR9170_MAC_BCN_HT1_BWC_40M_SHARED;
1591 			*plcp |= AR9170_MAC_BCN_HT2_BW40;
1592 		} else if (rate->flags & IEEE80211_TX_RC_DUP_DATA) {
1593 			*ht1 |= AR9170_MAC_BCN_HT1_BWC_40M_DUP;
1594 			*plcp |= AR9170_MAC_BCN_HT2_BW40;
1595 		}
1596 
1597 		SET_VAL(AR9170_MAC_BCN_HT2_LEN, *plcp, skb->len + FCS_LEN);
1598 	} else {
1599 		if (*plcp <= AR9170_TX_PHY_RATE_CCK_11M)
1600 			*plcp |= ((skb->len + FCS_LEN) << (3 + 16)) + 0x0400;
1601 		else
1602 			*plcp |= ((skb->len + FCS_LEN) << 16) + 0x0010;
1603 	}
1604 
1605 	return ht_rate;
1606 }
1607 
1608 int carl9170_update_beacon(struct ar9170 *ar, const bool submit)
1609 {
1610 	struct sk_buff *skb = NULL;
1611 	struct carl9170_vif_info *cvif;
1612 	__le32 *data, *old = NULL;
1613 	u32 word, ht1, plcp, off, addr, len;
1614 	int i = 0, err = 0;
1615 	bool ht_rate;
1616 
1617 	rcu_read_lock();
1618 	cvif = carl9170_pick_beaconing_vif(ar);
1619 	if (!cvif)
1620 		goto out_unlock;
1621 
1622 	skb = ieee80211_beacon_get_tim(ar->hw, carl9170_get_vif(cvif),
1623 		NULL, NULL);
1624 
1625 	if (!skb) {
1626 		err = -ENOMEM;
1627 		goto err_free;
1628 	}
1629 
1630 	spin_lock_bh(&ar->beacon_lock);
1631 	data = (__le32 *)skb->data;
1632 	if (cvif->beacon)
1633 		old = (__le32 *)cvif->beacon->data;
1634 
1635 	off = cvif->id * AR9170_MAC_BCN_LENGTH_MAX;
1636 	addr = ar->fw.beacon_addr + off;
1637 	len = roundup(skb->len + FCS_LEN, 4);
1638 
1639 	if ((off + len) > ar->fw.beacon_max_len) {
1640 		if (net_ratelimit()) {
1641 			wiphy_err(ar->hw->wiphy, "beacon does not "
1642 				  "fit into device memory!\n");
1643 		}
1644 		err = -EINVAL;
1645 		goto err_unlock;
1646 	}
1647 
1648 	if (len > AR9170_MAC_BCN_LENGTH_MAX) {
1649 		if (net_ratelimit()) {
1650 			wiphy_err(ar->hw->wiphy, "no support for beacons "
1651 				"bigger than %d (yours:%d).\n",
1652 				 AR9170_MAC_BCN_LENGTH_MAX, len);
1653 		}
1654 
1655 		err = -EMSGSIZE;
1656 		goto err_unlock;
1657 	}
1658 
1659 	ht_rate = carl9170_tx_beacon_physet(ar, skb, &ht1, &plcp);
1660 
1661 	carl9170_async_regwrite_begin(ar);
1662 	carl9170_async_regwrite(AR9170_MAC_REG_BCN_HT1, ht1);
1663 	if (ht_rate)
1664 		carl9170_async_regwrite(AR9170_MAC_REG_BCN_HT2, plcp);
1665 	else
1666 		carl9170_async_regwrite(AR9170_MAC_REG_BCN_PLCP, plcp);
1667 
1668 	for (i = 0; i < DIV_ROUND_UP(skb->len, 4); i++) {
1669 		/*
1670 		 * XXX: This accesses beyond skb data for up
1671 		 *	to the last 3 bytes!!
1672 		 */
1673 
1674 		if (old && (data[i] == old[i]))
1675 			continue;
1676 
1677 		word = le32_to_cpu(data[i]);
1678 		carl9170_async_regwrite(addr + 4 * i, word);
1679 	}
1680 	carl9170_async_regwrite_finish();
1681 
1682 	dev_kfree_skb_any(cvif->beacon);
1683 	cvif->beacon = NULL;
1684 
1685 	err = carl9170_async_regwrite_result();
1686 	if (!err)
1687 		cvif->beacon = skb;
1688 	spin_unlock_bh(&ar->beacon_lock);
1689 	if (err)
1690 		goto err_free;
1691 
1692 	if (submit) {
1693 		err = carl9170_bcn_ctrl(ar, cvif->id,
1694 					CARL9170_BCN_CTRL_CAB_TRIGGER,
1695 					addr, skb->len + FCS_LEN);
1696 
1697 		if (err)
1698 			goto err_free;
1699 	}
1700 out_unlock:
1701 	rcu_read_unlock();
1702 	return 0;
1703 
1704 err_unlock:
1705 	spin_unlock_bh(&ar->beacon_lock);
1706 
1707 err_free:
1708 	rcu_read_unlock();
1709 	dev_kfree_skb_any(skb);
1710 	return err;
1711 }
1712