xref: /openbmc/linux/drivers/net/wireless/ath/carl9170/tx.c (revision baa7eb025ab14f3cba2e35c0a8648f9c9f01d24f)
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].ampdu_len = 0;
387 		sta_info->stats[tid].ampdu_ack_len = 0;
388 	}
389 
390 	sta_info->stats[tid].ampdu_len++;
391 	if (txinfo->status.rates[0].count == 1)
392 		sta_info->stats[tid].ampdu_ack_len++;
393 
394 	if (super->f.mac_control & cpu_to_le16(AR9170_TX_MAC_IMM_BA)) {
395 		super->s.rix = sta_info->stats[tid].ampdu_len;
396 		super->s.cnt = sta_info->stats[tid].ampdu_ack_len;
397 		txinfo->flags |= IEEE80211_TX_STAT_AMPDU;
398 		sta_info->stats[tid].clear = true;
399 	}
400 	spin_unlock_bh(&tid_info->lock);
401 
402 out_rcu:
403 	rcu_read_unlock();
404 }
405 
406 void carl9170_tx_status(struct ar9170 *ar, struct sk_buff *skb,
407 			const bool success)
408 {
409 	struct ieee80211_tx_info *txinfo;
410 
411 	carl9170_tx_accounting_free(ar, skb);
412 
413 	txinfo = IEEE80211_SKB_CB(skb);
414 
415 	if (success)
416 		txinfo->flags |= IEEE80211_TX_STAT_ACK;
417 	else
418 		ar->tx_ack_failures++;
419 
420 	if (txinfo->flags & IEEE80211_TX_CTL_AMPDU)
421 		carl9170_tx_status_process_ampdu(ar, skb, txinfo);
422 
423 	carl9170_tx_put_skb(skb);
424 }
425 
426 /* This function may be called form any context */
427 void carl9170_tx_callback(struct ar9170 *ar, struct sk_buff *skb)
428 {
429 	struct ieee80211_tx_info *txinfo = IEEE80211_SKB_CB(skb);
430 
431 	atomic_dec(&ar->tx_total_pending);
432 
433 	if (txinfo->flags & IEEE80211_TX_CTL_AMPDU)
434 		atomic_dec(&ar->tx_ampdu_upload);
435 
436 	if (carl9170_tx_put_skb(skb))
437 		tasklet_hi_schedule(&ar->usb_tasklet);
438 }
439 
440 static struct sk_buff *carl9170_get_queued_skb(struct ar9170 *ar, u8 cookie,
441 					       struct sk_buff_head *queue)
442 {
443 	struct sk_buff *skb;
444 
445 	spin_lock_bh(&queue->lock);
446 	skb_queue_walk(queue, skb) {
447 		struct _carl9170_tx_superframe *txc = (void *) skb->data;
448 
449 		if (txc->s.cookie != cookie)
450 			continue;
451 
452 		__skb_unlink(skb, queue);
453 		spin_unlock_bh(&queue->lock);
454 
455 		carl9170_release_dev_space(ar, skb);
456 		return skb;
457 	}
458 	spin_unlock_bh(&queue->lock);
459 
460 	return NULL;
461 }
462 
463 static void carl9170_tx_fill_rateinfo(struct ar9170 *ar, unsigned int rix,
464 	unsigned int tries, struct ieee80211_tx_info *txinfo)
465 {
466 	unsigned int i;
467 
468 	for (i = 0; i < IEEE80211_TX_MAX_RATES; i++) {
469 		if (txinfo->status.rates[i].idx < 0)
470 			break;
471 
472 		if (i == rix) {
473 			txinfo->status.rates[i].count = tries;
474 			i++;
475 			break;
476 		}
477 	}
478 
479 	for (; i < IEEE80211_TX_MAX_RATES; i++) {
480 		txinfo->status.rates[i].idx = -1;
481 		txinfo->status.rates[i].count = 0;
482 	}
483 }
484 
485 static void carl9170_check_queue_stop_timeout(struct ar9170 *ar)
486 {
487 	int i;
488 	struct sk_buff *skb;
489 	struct ieee80211_tx_info *txinfo;
490 	struct carl9170_tx_info *arinfo;
491 	bool restart = false;
492 
493 	for (i = 0; i < ar->hw->queues; i++) {
494 		spin_lock_bh(&ar->tx_status[i].lock);
495 
496 		skb = skb_peek(&ar->tx_status[i]);
497 
498 		if (!skb)
499 			goto next;
500 
501 		txinfo = IEEE80211_SKB_CB(skb);
502 		arinfo = (void *) txinfo->rate_driver_data;
503 
504 		if (time_is_before_jiffies(arinfo->timeout +
505 		    msecs_to_jiffies(CARL9170_QUEUE_STUCK_TIMEOUT)) == true)
506 			restart = true;
507 
508 next:
509 		spin_unlock_bh(&ar->tx_status[i].lock);
510 	}
511 
512 	if (restart) {
513 		/*
514 		 * At least one queue has been stuck for long enough.
515 		 * Give the device a kick and hope it gets back to
516 		 * work.
517 		 *
518 		 * possible reasons may include:
519 		 *  - frames got lost/corrupted (bad connection to the device)
520 		 *  - stalled rx processing/usb controller hiccups
521 		 *  - firmware errors/bugs
522 		 *  - every bug you can think of.
523 		 *  - all bugs you can't...
524 		 *  - ...
525 		 */
526 		carl9170_restart(ar, CARL9170_RR_STUCK_TX);
527 	}
528 }
529 
530 static void carl9170_tx_ampdu_timeout(struct ar9170 *ar)
531 {
532 	struct carl9170_sta_tid *iter;
533 	struct sk_buff *skb;
534 	struct ieee80211_tx_info *txinfo;
535 	struct carl9170_tx_info *arinfo;
536 	struct _carl9170_tx_superframe *super;
537 	struct ieee80211_sta *sta;
538 	struct ieee80211_vif *vif;
539 	struct ieee80211_hdr *hdr;
540 	unsigned int vif_id;
541 
542 	rcu_read_lock();
543 	list_for_each_entry_rcu(iter, &ar->tx_ampdu_list, list) {
544 		if (iter->state < CARL9170_TID_STATE_IDLE)
545 			continue;
546 
547 		spin_lock_bh(&iter->lock);
548 		skb = skb_peek(&iter->queue);
549 		if (!skb)
550 			goto unlock;
551 
552 		txinfo = IEEE80211_SKB_CB(skb);
553 		arinfo = (void *)txinfo->rate_driver_data;
554 		if (time_is_after_jiffies(arinfo->timeout +
555 		    msecs_to_jiffies(CARL9170_QUEUE_TIMEOUT)))
556 			goto unlock;
557 
558 		super = (void *) skb->data;
559 		hdr = (void *) super->frame_data;
560 
561 		vif_id = (super->s.misc & CARL9170_TX_SUPER_MISC_VIF_ID) >>
562 			 CARL9170_TX_SUPER_MISC_VIF_ID_S;
563 
564 		if (WARN_ON(vif_id >= AR9170_MAX_VIRTUAL_MAC))
565 			goto unlock;
566 
567 		vif = rcu_dereference(ar->vif_priv[vif_id].vif);
568 		if (WARN_ON(!vif))
569 			goto unlock;
570 
571 		sta = ieee80211_find_sta(vif, hdr->addr1);
572 		if (WARN_ON(!sta))
573 			goto unlock;
574 
575 		ieee80211_stop_tx_ba_session(sta, iter->tid);
576 unlock:
577 		spin_unlock_bh(&iter->lock);
578 
579 	}
580 	rcu_read_unlock();
581 }
582 
583 void carl9170_tx_janitor(struct work_struct *work)
584 {
585 	struct ar9170 *ar = container_of(work, struct ar9170,
586 					 tx_janitor.work);
587 	if (!IS_STARTED(ar))
588 		return;
589 
590 	ar->tx_janitor_last_run = jiffies;
591 
592 	carl9170_check_queue_stop_timeout(ar);
593 	carl9170_tx_ampdu_timeout(ar);
594 
595 	if (!atomic_read(&ar->tx_total_queued))
596 		return;
597 
598 	ieee80211_queue_delayed_work(ar->hw, &ar->tx_janitor,
599 		msecs_to_jiffies(CARL9170_TX_TIMEOUT));
600 }
601 
602 static void __carl9170_tx_process_status(struct ar9170 *ar,
603 	const uint8_t cookie, const uint8_t info)
604 {
605 	struct sk_buff *skb;
606 	struct ieee80211_tx_info *txinfo;
607 	struct carl9170_tx_info *arinfo;
608 	unsigned int r, t, q;
609 	bool success = true;
610 
611 	q = ar9170_qmap[info & CARL9170_TX_STATUS_QUEUE];
612 
613 	skb = carl9170_get_queued_skb(ar, cookie, &ar->tx_status[q]);
614 	if (!skb) {
615 		/*
616 		 * We have lost the race to another thread.
617 		 */
618 
619 		return ;
620 	}
621 
622 	txinfo = IEEE80211_SKB_CB(skb);
623 	arinfo = (void *) txinfo->rate_driver_data;
624 
625 	if (!(info & CARL9170_TX_STATUS_SUCCESS))
626 		success = false;
627 
628 	r = (info & CARL9170_TX_STATUS_RIX) >> CARL9170_TX_STATUS_RIX_S;
629 	t = (info & CARL9170_TX_STATUS_TRIES) >> CARL9170_TX_STATUS_TRIES_S;
630 
631 	carl9170_tx_fill_rateinfo(ar, r, t, txinfo);
632 	carl9170_tx_status(ar, skb, success);
633 }
634 
635 void carl9170_tx_process_status(struct ar9170 *ar,
636 				const struct carl9170_rsp *cmd)
637 {
638 	unsigned int i;
639 
640 	for (i = 0;  i < cmd->hdr.ext; i++) {
641 		if (WARN_ON(i > ((cmd->hdr.len / 2) + 1))) {
642 			print_hex_dump_bytes("UU:", DUMP_PREFIX_NONE,
643 					     (void *) cmd, cmd->hdr.len + 4);
644 			break;
645 		}
646 
647 		__carl9170_tx_process_status(ar, cmd->_tx_status[i].cookie,
648 					     cmd->_tx_status[i].info);
649 	}
650 }
651 
652 static __le32 carl9170_tx_physet(struct ar9170 *ar,
653 	struct ieee80211_tx_info *info, struct ieee80211_tx_rate *txrate)
654 {
655 	struct ieee80211_rate *rate = NULL;
656 	u32 power, chains;
657 	__le32 tmp;
658 
659 	tmp = cpu_to_le32(0);
660 
661 	if (txrate->flags & IEEE80211_TX_RC_40_MHZ_WIDTH)
662 		tmp |= cpu_to_le32(AR9170_TX_PHY_BW_40MHZ <<
663 			AR9170_TX_PHY_BW_S);
664 	/* this works because 40 MHz is 2 and dup is 3 */
665 	if (txrate->flags & IEEE80211_TX_RC_DUP_DATA)
666 		tmp |= cpu_to_le32(AR9170_TX_PHY_BW_40MHZ_DUP <<
667 			AR9170_TX_PHY_BW_S);
668 
669 	if (txrate->flags & IEEE80211_TX_RC_SHORT_GI)
670 		tmp |= cpu_to_le32(AR9170_TX_PHY_SHORT_GI);
671 
672 	if (txrate->flags & IEEE80211_TX_RC_MCS) {
673 		u32 r = txrate->idx;
674 		u8 *txpower;
675 
676 		/* heavy clip control */
677 		tmp |= cpu_to_le32((r & 0x7) <<
678 			AR9170_TX_PHY_TX_HEAVY_CLIP_S);
679 
680 		if (txrate->flags & IEEE80211_TX_RC_40_MHZ_WIDTH) {
681 			if (info->band == IEEE80211_BAND_5GHZ)
682 				txpower = ar->power_5G_ht40;
683 			else
684 				txpower = ar->power_2G_ht40;
685 		} else {
686 			if (info->band == IEEE80211_BAND_5GHZ)
687 				txpower = ar->power_5G_ht20;
688 			else
689 				txpower = ar->power_2G_ht20;
690 		}
691 
692 		power = txpower[r & 7];
693 
694 		/* +1 dBm for HT40 */
695 		if (txrate->flags & IEEE80211_TX_RC_40_MHZ_WIDTH)
696 			power += 2;
697 
698 		r <<= AR9170_TX_PHY_MCS_S;
699 		BUG_ON(r & ~AR9170_TX_PHY_MCS);
700 
701 		tmp |= cpu_to_le32(r & AR9170_TX_PHY_MCS);
702 		tmp |= cpu_to_le32(AR9170_TX_PHY_MOD_HT);
703 
704 		/*
705 		 * green field preamble does not work.
706 		 *
707 		 * if (txrate->flags & IEEE80211_TX_RC_GREEN_FIELD)
708 		 * tmp |= cpu_to_le32(AR9170_TX_PHY_GREENFIELD);
709 		 */
710 	} else {
711 		u8 *txpower;
712 		u32 mod;
713 		u32 phyrate;
714 		u8 idx = txrate->idx;
715 
716 		if (info->band != IEEE80211_BAND_2GHZ) {
717 			idx += 4;
718 			txpower = ar->power_5G_leg;
719 			mod = AR9170_TX_PHY_MOD_OFDM;
720 		} else {
721 			if (idx < 4) {
722 				txpower = ar->power_2G_cck;
723 				mod = AR9170_TX_PHY_MOD_CCK;
724 			} else {
725 				mod = AR9170_TX_PHY_MOD_OFDM;
726 				txpower = ar->power_2G_ofdm;
727 			}
728 		}
729 
730 		rate = &__carl9170_ratetable[idx];
731 
732 		phyrate = rate->hw_value & 0xF;
733 		power = txpower[(rate->hw_value & 0x30) >> 4];
734 		phyrate <<= AR9170_TX_PHY_MCS_S;
735 
736 		tmp |= cpu_to_le32(mod);
737 		tmp |= cpu_to_le32(phyrate);
738 
739 		/*
740 		 * short preamble seems to be broken too.
741 		 *
742 		 * if (txrate->flags & IEEE80211_TX_RC_USE_SHORT_PREAMBLE)
743 		 *	tmp |= cpu_to_le32(AR9170_TX_PHY_SHORT_PREAMBLE);
744 		 */
745 	}
746 	power <<= AR9170_TX_PHY_TX_PWR_S;
747 	power &= AR9170_TX_PHY_TX_PWR;
748 	tmp |= cpu_to_le32(power);
749 
750 	/* set TX chains */
751 	if (ar->eeprom.tx_mask == 1) {
752 		chains = AR9170_TX_PHY_TXCHAIN_1;
753 	} else {
754 		chains = AR9170_TX_PHY_TXCHAIN_2;
755 
756 		/* >= 36M legacy OFDM - use only one chain */
757 		if (rate && rate->bitrate >= 360 &&
758 		    !(txrate->flags & IEEE80211_TX_RC_MCS))
759 			chains = AR9170_TX_PHY_TXCHAIN_1;
760 	}
761 	tmp |= cpu_to_le32(chains << AR9170_TX_PHY_TXCHAIN_S);
762 
763 	return tmp;
764 }
765 
766 static bool carl9170_tx_rts_check(struct ar9170 *ar,
767 				  struct ieee80211_tx_rate *rate,
768 				  bool ampdu, bool multi)
769 {
770 	switch (ar->erp_mode) {
771 	case CARL9170_ERP_AUTO:
772 		if (ampdu)
773 			break;
774 
775 	case CARL9170_ERP_MAC80211:
776 		if (!(rate->flags & IEEE80211_TX_RC_USE_RTS_CTS))
777 			break;
778 
779 	case CARL9170_ERP_RTS:
780 		if (likely(!multi))
781 			return true;
782 
783 	default:
784 		break;
785 	}
786 
787 	return false;
788 }
789 
790 static bool carl9170_tx_cts_check(struct ar9170 *ar,
791 				  struct ieee80211_tx_rate *rate)
792 {
793 	switch (ar->erp_mode) {
794 	case CARL9170_ERP_AUTO:
795 	case CARL9170_ERP_MAC80211:
796 		if (!(rate->flags & IEEE80211_TX_RC_USE_CTS_PROTECT))
797 			break;
798 
799 	case CARL9170_ERP_CTS:
800 		return true;
801 
802 	default:
803 		break;
804 	}
805 
806 	return false;
807 }
808 
809 static int carl9170_tx_prepare(struct ar9170 *ar, struct sk_buff *skb)
810 {
811 	struct ieee80211_hdr *hdr;
812 	struct _carl9170_tx_superframe *txc;
813 	struct carl9170_vif_info *cvif;
814 	struct ieee80211_tx_info *info;
815 	struct ieee80211_tx_rate *txrate;
816 	struct ieee80211_sta *sta;
817 	struct carl9170_tx_info *arinfo;
818 	unsigned int hw_queue;
819 	int i;
820 	__le16 mac_tmp;
821 	u16 len;
822 	bool ampdu, no_ack;
823 
824 	BUILD_BUG_ON(sizeof(*arinfo) > sizeof(info->rate_driver_data));
825 	BUILD_BUG_ON(sizeof(struct _carl9170_tx_superdesc) !=
826 		     CARL9170_TX_SUPERDESC_LEN);
827 
828 	BUILD_BUG_ON(sizeof(struct _ar9170_tx_hwdesc) !=
829 		     AR9170_TX_HWDESC_LEN);
830 
831 	BUILD_BUG_ON(IEEE80211_TX_MAX_RATES < CARL9170_TX_MAX_RATES);
832 
833 	BUILD_BUG_ON(AR9170_MAX_VIRTUAL_MAC >
834 		((CARL9170_TX_SUPER_MISC_VIF_ID >>
835 		 CARL9170_TX_SUPER_MISC_VIF_ID_S) + 1));
836 
837 	hw_queue = ar9170_qmap[carl9170_get_queue(ar, skb)];
838 
839 	hdr = (void *)skb->data;
840 	info = IEEE80211_SKB_CB(skb);
841 	len = skb->len;
842 
843 	/*
844 	 * Note: If the frame was sent through a monitor interface,
845 	 * the ieee80211_vif pointer can be NULL.
846 	 */
847 	if (likely(info->control.vif))
848 		cvif = (void *) info->control.vif->drv_priv;
849 	else
850 		cvif = NULL;
851 
852 	sta = info->control.sta;
853 
854 	txc = (void *)skb_push(skb, sizeof(*txc));
855 	memset(txc, 0, sizeof(*txc));
856 
857 	SET_VAL(CARL9170_TX_SUPER_MISC_QUEUE, txc->s.misc, hw_queue);
858 
859 	if (likely(cvif))
860 		SET_VAL(CARL9170_TX_SUPER_MISC_VIF_ID, txc->s.misc, cvif->id);
861 
862 	if (unlikely(info->flags & IEEE80211_TX_CTL_SEND_AFTER_DTIM))
863 		txc->s.misc |= CARL9170_TX_SUPER_MISC_CAB;
864 
865 	if (unlikely(ieee80211_is_probe_resp(hdr->frame_control)))
866 		txc->s.misc |= CARL9170_TX_SUPER_MISC_FILL_IN_TSF;
867 
868 	mac_tmp = cpu_to_le16(AR9170_TX_MAC_HW_DURATION |
869 			      AR9170_TX_MAC_BACKOFF);
870 	mac_tmp |= cpu_to_le16((hw_queue << AR9170_TX_MAC_QOS_S) &
871 			       AR9170_TX_MAC_QOS);
872 
873 	no_ack = !!(info->flags & IEEE80211_TX_CTL_NO_ACK);
874 	if (unlikely(no_ack))
875 		mac_tmp |= cpu_to_le16(AR9170_TX_MAC_NO_ACK);
876 
877 	if (info->control.hw_key) {
878 		len += info->control.hw_key->icv_len;
879 
880 		switch (info->control.hw_key->cipher) {
881 		case WLAN_CIPHER_SUITE_WEP40:
882 		case WLAN_CIPHER_SUITE_WEP104:
883 		case WLAN_CIPHER_SUITE_TKIP:
884 			mac_tmp |= cpu_to_le16(AR9170_TX_MAC_ENCR_RC4);
885 			break;
886 		case WLAN_CIPHER_SUITE_CCMP:
887 			mac_tmp |= cpu_to_le16(AR9170_TX_MAC_ENCR_AES);
888 			break;
889 		default:
890 			WARN_ON(1);
891 			goto err_out;
892 		}
893 	}
894 
895 	ampdu = !!(info->flags & IEEE80211_TX_CTL_AMPDU);
896 	if (ampdu) {
897 		unsigned int density, factor;
898 
899 		if (unlikely(!sta || !cvif))
900 			goto err_out;
901 
902 		factor = min_t(unsigned int, 1u, sta->ht_cap.ampdu_factor);
903 		density = sta->ht_cap.ampdu_density;
904 
905 		if (density) {
906 			/*
907 			 * Watch out!
908 			 *
909 			 * Otus uses slightly different density values than
910 			 * those from the 802.11n spec.
911 			 */
912 
913 			density = max_t(unsigned int, density + 1, 7u);
914 		}
915 
916 		SET_VAL(CARL9170_TX_SUPER_AMPDU_DENSITY,
917 			txc->s.ampdu_settings, density);
918 
919 		SET_VAL(CARL9170_TX_SUPER_AMPDU_FACTOR,
920 			txc->s.ampdu_settings, factor);
921 
922 		for (i = 0; i < CARL9170_TX_MAX_RATES; i++) {
923 			txrate = &info->control.rates[i];
924 			if (txrate->idx >= 0) {
925 				txc->s.ri[i] =
926 					CARL9170_TX_SUPER_RI_AMPDU;
927 
928 				if (WARN_ON(!(txrate->flags &
929 					      IEEE80211_TX_RC_MCS))) {
930 					/*
931 					 * Not sure if it's even possible
932 					 * to aggregate non-ht rates with
933 					 * this HW.
934 					 */
935 					goto err_out;
936 				}
937 				continue;
938 			}
939 
940 			txrate->idx = 0;
941 			txrate->count = ar->hw->max_rate_tries;
942 		}
943 
944 		mac_tmp |= cpu_to_le16(AR9170_TX_MAC_AGGR);
945 	}
946 
947 	/*
948 	 * NOTE: For the first rate, the ERP & AMPDU flags are directly
949 	 * taken from mac_control. For all fallback rate, the firmware
950 	 * updates the mac_control flags from the rate info field.
951 	 */
952 	for (i = 1; i < CARL9170_TX_MAX_RATES; i++) {
953 		txrate = &info->control.rates[i];
954 		if (txrate->idx < 0)
955 			break;
956 
957 		SET_VAL(CARL9170_TX_SUPER_RI_TRIES, txc->s.ri[i],
958 			txrate->count);
959 
960 		if (carl9170_tx_rts_check(ar, txrate, ampdu, no_ack))
961 			txc->s.ri[i] |= (AR9170_TX_MAC_PROT_RTS <<
962 				CARL9170_TX_SUPER_RI_ERP_PROT_S);
963 		else if (carl9170_tx_cts_check(ar, txrate))
964 			txc->s.ri[i] |= (AR9170_TX_MAC_PROT_CTS <<
965 				CARL9170_TX_SUPER_RI_ERP_PROT_S);
966 
967 		txc->s.rr[i - 1] = carl9170_tx_physet(ar, info, txrate);
968 	}
969 
970 	txrate = &info->control.rates[0];
971 	SET_VAL(CARL9170_TX_SUPER_RI_TRIES, txc->s.ri[0], txrate->count);
972 
973 	if (carl9170_tx_rts_check(ar, txrate, ampdu, no_ack))
974 		mac_tmp |= cpu_to_le16(AR9170_TX_MAC_PROT_RTS);
975 	else if (carl9170_tx_cts_check(ar, txrate))
976 		mac_tmp |= cpu_to_le16(AR9170_TX_MAC_PROT_CTS);
977 
978 	txc->s.len = cpu_to_le16(skb->len);
979 	txc->f.length = cpu_to_le16(len + FCS_LEN);
980 	txc->f.mac_control = mac_tmp;
981 	txc->f.phy_control = carl9170_tx_physet(ar, info, txrate);
982 
983 	arinfo = (void *)info->rate_driver_data;
984 	arinfo->timeout = jiffies;
985 	arinfo->ar = ar;
986 	kref_init(&arinfo->ref);
987 	return 0;
988 
989 err_out:
990 	skb_pull(skb, sizeof(*txc));
991 	return -EINVAL;
992 }
993 
994 static void carl9170_set_immba(struct ar9170 *ar, struct sk_buff *skb)
995 {
996 	struct _carl9170_tx_superframe *super;
997 
998 	super = (void *) skb->data;
999 	super->f.mac_control |= cpu_to_le16(AR9170_TX_MAC_IMM_BA);
1000 }
1001 
1002 static void carl9170_set_ampdu_params(struct ar9170 *ar, struct sk_buff *skb)
1003 {
1004 	struct _carl9170_tx_superframe *super;
1005 	int tmp;
1006 
1007 	super = (void *) skb->data;
1008 
1009 	tmp = (super->s.ampdu_settings & CARL9170_TX_SUPER_AMPDU_DENSITY) <<
1010 		CARL9170_TX_SUPER_AMPDU_DENSITY_S;
1011 
1012 	/*
1013 	 * If you haven't noticed carl9170_tx_prepare has already filled
1014 	 * in all ampdu spacing & factor parameters.
1015 	 * Now it's the time to check whenever the settings have to be
1016 	 * updated by the firmware, or if everything is still the same.
1017 	 *
1018 	 * There's no sane way to handle different density values with
1019 	 * this hardware, so we may as well just do the compare in the
1020 	 * driver.
1021 	 */
1022 
1023 	if (tmp != ar->current_density) {
1024 		ar->current_density = tmp;
1025 		super->s.ampdu_settings |=
1026 			CARL9170_TX_SUPER_AMPDU_COMMIT_DENSITY;
1027 	}
1028 
1029 	tmp = (super->s.ampdu_settings & CARL9170_TX_SUPER_AMPDU_FACTOR) <<
1030 		CARL9170_TX_SUPER_AMPDU_FACTOR_S;
1031 
1032 	if (tmp != ar->current_factor) {
1033 		ar->current_factor = tmp;
1034 		super->s.ampdu_settings |=
1035 			CARL9170_TX_SUPER_AMPDU_COMMIT_FACTOR;
1036 	}
1037 }
1038 
1039 static bool carl9170_tx_rate_check(struct ar9170 *ar, struct sk_buff *_dest,
1040 				   struct sk_buff *_src)
1041 {
1042 	struct _carl9170_tx_superframe *dest, *src;
1043 
1044 	dest = (void *) _dest->data;
1045 	src = (void *) _src->data;
1046 
1047 	/*
1048 	 * The mac80211 rate control algorithm expects that all MPDUs in
1049 	 * an AMPDU share the same tx vectors.
1050 	 * This is not really obvious right now, because the hardware
1051 	 * does the AMPDU setup according to its own rulebook.
1052 	 * Our nicely assembled, strictly monotonic increasing mpdu
1053 	 * chains will be broken up, mashed back together...
1054 	 */
1055 
1056 	return (dest->f.phy_control == src->f.phy_control);
1057 }
1058 
1059 static void carl9170_tx_ampdu(struct ar9170 *ar)
1060 {
1061 	struct sk_buff_head agg;
1062 	struct carl9170_sta_tid *tid_info;
1063 	struct sk_buff *skb, *first;
1064 	unsigned int i = 0, done_ampdus = 0;
1065 	u16 seq, queue, tmpssn;
1066 
1067 	atomic_inc(&ar->tx_ampdu_scheduler);
1068 	ar->tx_ampdu_schedule = false;
1069 
1070 	if (atomic_read(&ar->tx_ampdu_upload))
1071 		return;
1072 
1073 	if (!ar->tx_ampdu_list_len)
1074 		return;
1075 
1076 	__skb_queue_head_init(&agg);
1077 
1078 	rcu_read_lock();
1079 	tid_info = rcu_dereference(ar->tx_ampdu_iter);
1080 	if (WARN_ON_ONCE(!tid_info)) {
1081 		rcu_read_unlock();
1082 		return;
1083 	}
1084 
1085 retry:
1086 	list_for_each_entry_continue_rcu(tid_info, &ar->tx_ampdu_list, list) {
1087 		i++;
1088 
1089 		if (tid_info->state < CARL9170_TID_STATE_PROGRESS)
1090 			continue;
1091 
1092 		queue = TID_TO_WME_AC(tid_info->tid);
1093 
1094 		spin_lock_bh(&tid_info->lock);
1095 		if (tid_info->state != CARL9170_TID_STATE_XMIT)
1096 			goto processed;
1097 
1098 		tid_info->counter++;
1099 		first = skb_peek(&tid_info->queue);
1100 		tmpssn = carl9170_get_seq(first);
1101 		seq = tid_info->snx;
1102 
1103 		if (unlikely(tmpssn != seq)) {
1104 			tid_info->state = CARL9170_TID_STATE_IDLE;
1105 
1106 			goto processed;
1107 		}
1108 
1109 		while ((skb = skb_peek(&tid_info->queue))) {
1110 			/* strict 0, 1, ..., n - 1, n frame sequence order */
1111 			if (unlikely(carl9170_get_seq(skb) != seq))
1112 				break;
1113 
1114 			/* don't upload more than AMPDU FACTOR allows. */
1115 			if (unlikely(SEQ_DIFF(tid_info->snx, tid_info->bsn) >=
1116 			    (tid_info->max - 1)))
1117 				break;
1118 
1119 			if (!carl9170_tx_rate_check(ar, skb, first))
1120 				break;
1121 
1122 			atomic_inc(&ar->tx_ampdu_upload);
1123 			tid_info->snx = seq = SEQ_NEXT(seq);
1124 			__skb_unlink(skb, &tid_info->queue);
1125 
1126 			__skb_queue_tail(&agg, skb);
1127 
1128 			if (skb_queue_len(&agg) >= CARL9170_NUM_TX_AGG_MAX)
1129 				break;
1130 		}
1131 
1132 		if (skb_queue_empty(&tid_info->queue) ||
1133 		    carl9170_get_seq(skb_peek(&tid_info->queue)) !=
1134 		    tid_info->snx) {
1135 			/*
1136 			 * stop TID, if A-MPDU frames are still missing,
1137 			 * or whenever the queue is empty.
1138 			 */
1139 
1140 			tid_info->state = CARL9170_TID_STATE_IDLE;
1141 		}
1142 		done_ampdus++;
1143 
1144 processed:
1145 		spin_unlock_bh(&tid_info->lock);
1146 
1147 		if (skb_queue_empty(&agg))
1148 			continue;
1149 
1150 		/* apply ampdu spacing & factor settings */
1151 		carl9170_set_ampdu_params(ar, skb_peek(&agg));
1152 
1153 		/* set aggregation push bit */
1154 		carl9170_set_immba(ar, skb_peek_tail(&agg));
1155 
1156 		spin_lock_bh(&ar->tx_pending[queue].lock);
1157 		skb_queue_splice_tail_init(&agg, &ar->tx_pending[queue]);
1158 		spin_unlock_bh(&ar->tx_pending[queue].lock);
1159 		ar->tx_schedule = true;
1160 	}
1161 	if ((done_ampdus++ == 0) && (i++ == 0))
1162 		goto retry;
1163 
1164 	rcu_assign_pointer(ar->tx_ampdu_iter, tid_info);
1165 	rcu_read_unlock();
1166 }
1167 
1168 static struct sk_buff *carl9170_tx_pick_skb(struct ar9170 *ar,
1169 					    struct sk_buff_head *queue)
1170 {
1171 	struct sk_buff *skb;
1172 	struct ieee80211_tx_info *info;
1173 	struct carl9170_tx_info *arinfo;
1174 
1175 	BUILD_BUG_ON(sizeof(*arinfo) > sizeof(info->rate_driver_data));
1176 
1177 	spin_lock_bh(&queue->lock);
1178 	skb = skb_peek(queue);
1179 	if (unlikely(!skb))
1180 		goto err_unlock;
1181 
1182 	if (carl9170_alloc_dev_space(ar, skb))
1183 		goto err_unlock;
1184 
1185 	__skb_unlink(skb, queue);
1186 	spin_unlock_bh(&queue->lock);
1187 
1188 	info = IEEE80211_SKB_CB(skb);
1189 	arinfo = (void *) info->rate_driver_data;
1190 
1191 	arinfo->timeout = jiffies;
1192 
1193 	/*
1194 	 * increase ref count to "2".
1195 	 * Ref counting is the easiest way to solve the race between
1196 	 * the the urb's completion routine: carl9170_tx_callback and
1197 	 * wlan tx status functions: carl9170_tx_status/janitor.
1198 	 */
1199 	carl9170_tx_get_skb(skb);
1200 
1201 	return skb;
1202 
1203 err_unlock:
1204 	spin_unlock_bh(&queue->lock);
1205 	return NULL;
1206 }
1207 
1208 void carl9170_tx_drop(struct ar9170 *ar, struct sk_buff *skb)
1209 {
1210 	struct _carl9170_tx_superframe *super;
1211 	uint8_t q = 0;
1212 
1213 	ar->tx_dropped++;
1214 
1215 	super = (void *)skb->data;
1216 	SET_VAL(CARL9170_TX_SUPER_MISC_QUEUE, q,
1217 		ar9170_qmap[carl9170_get_queue(ar, skb)]);
1218 	__carl9170_tx_process_status(ar, super->s.cookie, q);
1219 }
1220 
1221 static void carl9170_tx(struct ar9170 *ar)
1222 {
1223 	struct sk_buff *skb;
1224 	unsigned int i, q;
1225 	bool schedule_garbagecollector = false;
1226 
1227 	ar->tx_schedule = false;
1228 
1229 	if (unlikely(!IS_STARTED(ar)))
1230 		return;
1231 
1232 	carl9170_usb_handle_tx_err(ar);
1233 
1234 	for (i = 0; i < ar->hw->queues; i++) {
1235 		while (!skb_queue_empty(&ar->tx_pending[i])) {
1236 			skb = carl9170_tx_pick_skb(ar, &ar->tx_pending[i]);
1237 			if (unlikely(!skb))
1238 				break;
1239 
1240 			atomic_inc(&ar->tx_total_pending);
1241 
1242 			q = __carl9170_get_queue(ar, i);
1243 			/*
1244 			 * NB: tx_status[i] vs. tx_status[q],
1245 			 * TODO: Move into pick_skb or alloc_dev_space.
1246 			 */
1247 			skb_queue_tail(&ar->tx_status[q], skb);
1248 
1249 			carl9170_usb_tx(ar, skb);
1250 			schedule_garbagecollector = true;
1251 		}
1252 	}
1253 
1254 	if (!schedule_garbagecollector)
1255 		return;
1256 
1257 	ieee80211_queue_delayed_work(ar->hw, &ar->tx_janitor,
1258 		msecs_to_jiffies(CARL9170_TX_TIMEOUT));
1259 }
1260 
1261 static bool carl9170_tx_ampdu_queue(struct ar9170 *ar,
1262 	struct ieee80211_sta *sta, struct sk_buff *skb)
1263 {
1264 	struct _carl9170_tx_superframe *super = (void *) skb->data;
1265 	struct carl9170_sta_info *sta_info;
1266 	struct carl9170_sta_tid *agg;
1267 	struct sk_buff *iter;
1268 	unsigned int max;
1269 	u16 tid, seq, qseq, off;
1270 	bool run = false;
1271 
1272 	tid = carl9170_get_tid(skb);
1273 	seq = carl9170_get_seq(skb);
1274 	sta_info = (void *) sta->drv_priv;
1275 
1276 	rcu_read_lock();
1277 	agg = rcu_dereference(sta_info->agg[tid]);
1278 	max = sta_info->ampdu_max_len;
1279 
1280 	if (!agg)
1281 		goto err_unlock_rcu;
1282 
1283 	spin_lock_bh(&agg->lock);
1284 	if (unlikely(agg->state < CARL9170_TID_STATE_IDLE))
1285 		goto err_unlock;
1286 
1287 	/* check if sequence is within the BA window */
1288 	if (unlikely(!BAW_WITHIN(agg->bsn, CARL9170_BAW_BITS, seq)))
1289 		goto err_unlock;
1290 
1291 	if (WARN_ON_ONCE(!BAW_WITHIN(agg->snx, CARL9170_BAW_BITS, seq)))
1292 		goto err_unlock;
1293 
1294 	off = SEQ_DIFF(seq, agg->bsn);
1295 	if (WARN_ON_ONCE(test_and_set_bit(off, agg->bitmap)))
1296 		goto err_unlock;
1297 
1298 	if (likely(BAW_WITHIN(agg->hsn, CARL9170_BAW_BITS, seq))) {
1299 		__skb_queue_tail(&agg->queue, skb);
1300 		agg->hsn = seq;
1301 		goto queued;
1302 	}
1303 
1304 	skb_queue_reverse_walk(&agg->queue, iter) {
1305 		qseq = carl9170_get_seq(iter);
1306 
1307 		if (BAW_WITHIN(qseq, CARL9170_BAW_BITS, seq)) {
1308 			__skb_queue_after(&agg->queue, iter, skb);
1309 			goto queued;
1310 		}
1311 	}
1312 
1313 	__skb_queue_head(&agg->queue, skb);
1314 queued:
1315 
1316 	if (unlikely(agg->state != CARL9170_TID_STATE_XMIT)) {
1317 		if (agg->snx == carl9170_get_seq(skb_peek(&agg->queue))) {
1318 			agg->state = CARL9170_TID_STATE_XMIT;
1319 			run = true;
1320 		}
1321 	}
1322 
1323 	spin_unlock_bh(&agg->lock);
1324 	rcu_read_unlock();
1325 
1326 	return run;
1327 
1328 err_unlock:
1329 	spin_unlock_bh(&agg->lock);
1330 
1331 err_unlock_rcu:
1332 	rcu_read_unlock();
1333 	super->f.mac_control &= ~cpu_to_le16(AR9170_TX_MAC_AGGR);
1334 	carl9170_tx_status(ar, skb, false);
1335 	ar->tx_dropped++;
1336 	return false;
1337 }
1338 
1339 int carl9170_op_tx(struct ieee80211_hw *hw, struct sk_buff *skb)
1340 {
1341 	struct ar9170 *ar = hw->priv;
1342 	struct ieee80211_tx_info *info;
1343 	struct ieee80211_sta *sta;
1344 	bool run;
1345 
1346 	if (unlikely(!IS_STARTED(ar)))
1347 		goto err_free;
1348 
1349 	info = IEEE80211_SKB_CB(skb);
1350 	sta = info->control.sta;
1351 
1352 	if (unlikely(carl9170_tx_prepare(ar, skb)))
1353 		goto err_free;
1354 
1355 	carl9170_tx_accounting(ar, skb);
1356 	/*
1357 	 * from now on, one has to use carl9170_tx_status to free
1358 	 * all ressouces which are associated with the frame.
1359 	 */
1360 
1361 	if (info->flags & IEEE80211_TX_CTL_AMPDU) {
1362 		run = carl9170_tx_ampdu_queue(ar, sta, skb);
1363 		if (run)
1364 			carl9170_tx_ampdu(ar);
1365 
1366 	} else {
1367 		unsigned int queue = skb_get_queue_mapping(skb);
1368 
1369 		skb_queue_tail(&ar->tx_pending[queue], skb);
1370 	}
1371 
1372 	carl9170_tx(ar);
1373 	return NETDEV_TX_OK;
1374 
1375 err_free:
1376 	ar->tx_dropped++;
1377 	dev_kfree_skb_any(skb);
1378 	return NETDEV_TX_OK;
1379 }
1380 
1381 void carl9170_tx_scheduler(struct ar9170 *ar)
1382 {
1383 
1384 	if (ar->tx_ampdu_schedule)
1385 		carl9170_tx_ampdu(ar);
1386 
1387 	if (ar->tx_schedule)
1388 		carl9170_tx(ar);
1389 }
1390