xref: /linux/drivers/net/wireless/ath/carl9170/tx.c (revision 6eb2fb3170549737207974c2c6ad34bcc2f3025e)
1 /*
2  * Atheros CARL9170 driver
3  *
4  * 802.11 xmit & status routines
5  *
6  * Copyright 2008, Johannes Berg <johannes@sipsolutions.net>
7  * Copyright 2009, 2010, Christian Lamparter <chunkeey@googlemail.com>
8  *
9  * This program is free software; you can redistribute it and/or modify
10  * it under the terms of the GNU General Public License as published by
11  * the Free Software Foundation; either version 2 of the License, or
12  * (at your option) any later version.
13  *
14  * This program is distributed in the hope that it will be useful,
15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
17  * GNU General Public License for more details.
18  *
19  * You should have received a copy of the GNU General Public License
20  * along with this program; see the file COPYING.  If not, see
21  * http://www.gnu.org/licenses/.
22  *
23  * This file incorporates work covered by the following copyright and
24  * permission notice:
25  *    Copyright (c) 2007-2008 Atheros Communications, Inc.
26  *
27  *    Permission to use, copy, modify, and/or distribute this software for any
28  *    purpose with or without fee is hereby granted, provided that the above
29  *    copyright notice and this permission notice appear in all copies.
30  *
31  *    THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
32  *    WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
33  *    MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
34  *    ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
35  *    WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
36  *    ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
37  *    OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
38  */
39 
40 #include <linux/init.h>
41 #include <linux/slab.h>
42 #include <linux/module.h>
43 #include <linux/etherdevice.h>
44 #include <net/mac80211.h>
45 #include "carl9170.h"
46 #include "hw.h"
47 #include "cmd.h"
48 
49 static inline unsigned int __carl9170_get_queue(struct ar9170 *ar,
50 						unsigned int queue)
51 {
52 	if (unlikely(modparam_noht)) {
53 		return queue;
54 	} else {
55 		/*
56 		 * This is just another workaround, until
57 		 * someone figures out how to get QoS and
58 		 * AMPDU to play nicely together.
59 		 */
60 
61 		return 2;		/* AC_BE */
62 	}
63 }
64 
65 static inline unsigned int carl9170_get_queue(struct ar9170 *ar,
66 					      struct sk_buff *skb)
67 {
68 	return __carl9170_get_queue(ar, skb_get_queue_mapping(skb));
69 }
70 
71 static bool is_mem_full(struct ar9170 *ar)
72 {
73 	return (DIV_ROUND_UP(IEEE80211_MAX_FRAME_LEN, ar->fw.mem_block_size) >
74 		atomic_read(&ar->mem_free_blocks));
75 }
76 
77 static void carl9170_tx_accounting(struct ar9170 *ar, struct sk_buff *skb)
78 {
79 	int queue, i;
80 	bool mem_full;
81 
82 	atomic_inc(&ar->tx_total_queued);
83 
84 	queue = skb_get_queue_mapping(skb);
85 	spin_lock_bh(&ar->tx_stats_lock);
86 
87 	/*
88 	 * The driver has to accept the frame, regardless if the queue is
89 	 * full to the brim, or not. We have to do the queuing internally,
90 	 * since mac80211 assumes that a driver which can operate with
91 	 * aggregated frames does not reject frames for this reason.
92 	 */
93 	ar->tx_stats[queue].len++;
94 	ar->tx_stats[queue].count++;
95 
96 	mem_full = is_mem_full(ar);
97 	for (i = 0; i < ar->hw->queues; i++) {
98 		if (mem_full || ar->tx_stats[i].len >= ar->tx_stats[i].limit) {
99 			ieee80211_stop_queue(ar->hw, i);
100 			ar->queue_stop_timeout[i] = jiffies;
101 		}
102 	}
103 
104 	spin_unlock_bh(&ar->tx_stats_lock);
105 }
106 
107 /* needs rcu_read_lock */
108 static struct ieee80211_sta *__carl9170_get_tx_sta(struct ar9170 *ar,
109 						   struct sk_buff *skb)
110 {
111 	struct _carl9170_tx_superframe *super = (void *) skb->data;
112 	struct ieee80211_hdr *hdr = (void *) super->frame_data;
113 	struct ieee80211_vif *vif;
114 	unsigned int vif_id;
115 
116 	vif_id = (super->s.misc & CARL9170_TX_SUPER_MISC_VIF_ID) >>
117 		 CARL9170_TX_SUPER_MISC_VIF_ID_S;
118 
119 	if (WARN_ON_ONCE(vif_id >= AR9170_MAX_VIRTUAL_MAC))
120 		return NULL;
121 
122 	vif = rcu_dereference(ar->vif_priv[vif_id].vif);
123 	if (unlikely(!vif))
124 		return NULL;
125 
126 	/*
127 	 * Normally we should use wrappers like ieee80211_get_DA to get
128 	 * the correct peer ieee80211_sta.
129 	 *
130 	 * But there is a problem with indirect traffic (broadcasts, or
131 	 * data which is designated for other stations) in station mode.
132 	 * The frame will be directed to the AP for distribution and not
133 	 * to the actual destination.
134 	 */
135 
136 	return ieee80211_find_sta(vif, hdr->addr1);
137 }
138 
139 static void carl9170_tx_ps_unblock(struct ar9170 *ar, struct sk_buff *skb)
140 {
141 	struct ieee80211_sta *sta;
142 	struct carl9170_sta_info *sta_info;
143 
144 	rcu_read_lock();
145 	sta = __carl9170_get_tx_sta(ar, skb);
146 	if (unlikely(!sta))
147 		goto out_rcu;
148 
149 	sta_info = (struct carl9170_sta_info *) sta->drv_priv;
150 	if (atomic_dec_return(&sta_info->pending_frames) == 0)
151 		ieee80211_sta_block_awake(ar->hw, sta, false);
152 
153 out_rcu:
154 	rcu_read_unlock();
155 }
156 
157 static void carl9170_tx_accounting_free(struct ar9170 *ar, struct sk_buff *skb)
158 {
159 	int queue;
160 
161 	queue = skb_get_queue_mapping(skb);
162 
163 	spin_lock_bh(&ar->tx_stats_lock);
164 
165 	ar->tx_stats[queue].len--;
166 
167 	if (!is_mem_full(ar)) {
168 		unsigned int i;
169 		for (i = 0; i < ar->hw->queues; i++) {
170 			if (ar->tx_stats[i].len >= CARL9170_NUM_TX_LIMIT_SOFT)
171 				continue;
172 
173 			if (ieee80211_queue_stopped(ar->hw, i)) {
174 				unsigned long tmp;
175 
176 				tmp = jiffies - ar->queue_stop_timeout[i];
177 				if (tmp > ar->max_queue_stop_timeout[i])
178 					ar->max_queue_stop_timeout[i] = tmp;
179 			}
180 
181 			ieee80211_wake_queue(ar->hw, i);
182 		}
183 	}
184 
185 	spin_unlock_bh(&ar->tx_stats_lock);
186 
187 	if (atomic_dec_and_test(&ar->tx_total_queued))
188 		complete(&ar->tx_flush);
189 }
190 
191 static int carl9170_alloc_dev_space(struct ar9170 *ar, struct sk_buff *skb)
192 {
193 	struct _carl9170_tx_superframe *super = (void *) skb->data;
194 	unsigned int chunks;
195 	int cookie = -1;
196 
197 	atomic_inc(&ar->mem_allocs);
198 
199 	chunks = DIV_ROUND_UP(skb->len, ar->fw.mem_block_size);
200 	if (unlikely(atomic_sub_return(chunks, &ar->mem_free_blocks) < 0)) {
201 		atomic_add(chunks, &ar->mem_free_blocks);
202 		return -ENOSPC;
203 	}
204 
205 	spin_lock_bh(&ar->mem_lock);
206 	cookie = bitmap_find_free_region(ar->mem_bitmap, ar->fw.mem_blocks, 0);
207 	spin_unlock_bh(&ar->mem_lock);
208 
209 	if (unlikely(cookie < 0)) {
210 		atomic_add(chunks, &ar->mem_free_blocks);
211 		return -ENOSPC;
212 	}
213 
214 	super = (void *) skb->data;
215 
216 	/*
217 	 * Cookie #0 serves two special purposes:
218 	 *  1. The firmware might use it generate BlockACK frames
219 	 *     in responds of an incoming BlockAckReqs.
220 	 *
221 	 *  2. Prevent double-free bugs.
222 	 */
223 	super->s.cookie = (u8) cookie + 1;
224 	return 0;
225 }
226 
227 static void carl9170_release_dev_space(struct ar9170 *ar, struct sk_buff *skb)
228 {
229 	struct _carl9170_tx_superframe *super = (void *) skb->data;
230 	int cookie;
231 
232 	/* make a local copy of the cookie */
233 	cookie = super->s.cookie;
234 	/* invalidate cookie */
235 	super->s.cookie = 0;
236 
237 	/*
238 	 * Do a out-of-bounds check on the cookie:
239 	 *
240 	 *  * cookie "0" is reserved and won't be assigned to any
241 	 *    out-going frame. Internally however, it is used to
242 	 *    mark no longer/un-accounted frames and serves as a
243 	 *    cheap way of preventing frames from being freed
244 	 *    twice by _accident_. NB: There is a tiny race...
245 	 *
246 	 *  * obviously, cookie number is limited by the amount
247 	 *    of available memory blocks, so the number can
248 	 *    never execeed the mem_blocks count.
249 	 */
250 	if (unlikely(WARN_ON_ONCE(cookie == 0) ||
251 	    WARN_ON_ONCE(cookie > ar->fw.mem_blocks)))
252 		return;
253 
254 	atomic_add(DIV_ROUND_UP(skb->len, ar->fw.mem_block_size),
255 		   &ar->mem_free_blocks);
256 
257 	spin_lock_bh(&ar->mem_lock);
258 	bitmap_release_region(ar->mem_bitmap, cookie - 1, 0);
259 	spin_unlock_bh(&ar->mem_lock);
260 }
261 
262 /* Called from any context */
263 static void carl9170_tx_release(struct kref *ref)
264 {
265 	struct ar9170 *ar;
266 	struct carl9170_tx_info *arinfo;
267 	struct ieee80211_tx_info *txinfo;
268 	struct sk_buff *skb;
269 
270 	arinfo = container_of(ref, struct carl9170_tx_info, ref);
271 	txinfo = container_of((void *) arinfo, struct ieee80211_tx_info,
272 			      rate_driver_data);
273 	skb = container_of((void *) txinfo, struct sk_buff, cb);
274 
275 	ar = arinfo->ar;
276 	if (WARN_ON_ONCE(!ar))
277 		return;
278 
279 	BUILD_BUG_ON(
280 	    offsetof(struct ieee80211_tx_info, status.ack_signal) != 20);
281 
282 	memset(&txinfo->status.ack_signal, 0,
283 	       sizeof(struct ieee80211_tx_info) -
284 	       offsetof(struct ieee80211_tx_info, status.ack_signal));
285 
286 	if (atomic_read(&ar->tx_total_queued))
287 		ar->tx_schedule = true;
288 
289 	if (txinfo->flags & IEEE80211_TX_CTL_AMPDU) {
290 		if (!atomic_read(&ar->tx_ampdu_upload))
291 			ar->tx_ampdu_schedule = true;
292 
293 		if (txinfo->flags & IEEE80211_TX_STAT_AMPDU) {
294 			struct _carl9170_tx_superframe *super;
295 
296 			super = (void *)skb->data;
297 			txinfo->status.ampdu_len = super->s.rix;
298 			txinfo->status.ampdu_ack_len = super->s.cnt;
299 		} else if ((txinfo->flags & IEEE80211_TX_STAT_ACK) &&
300 			   !(txinfo->flags & IEEE80211_TX_CTL_REQ_TX_STATUS)) {
301 			/*
302 			 * drop redundant tx_status reports:
303 			 *
304 			 * 1. ampdu_ack_len of the final tx_status does
305 			 *    include the feedback of this particular frame.
306 			 *
307 			 * 2. tx_status_irqsafe only queues up to 128
308 			 *    tx feedback reports and discards the rest.
309 			 *
310 			 * 3. minstrel_ht is picky, it only accepts
311 			 *    reports of frames with the TX_STATUS_AMPDU flag.
312 			 *
313 			 * 4. mac80211 is not particularly interested in
314 			 *    feedback either [CTL_REQ_TX_STATUS not set]
315 			 */
316 
317 			ieee80211_free_txskb(ar->hw, skb);
318 			return;
319 		} else {
320 			/*
321 			 * Either the frame transmission has failed or
322 			 * mac80211 requested tx status.
323 			 */
324 		}
325 	}
326 
327 	skb_pull(skb, sizeof(struct _carl9170_tx_superframe));
328 	ieee80211_tx_status_irqsafe(ar->hw, skb);
329 }
330 
331 void carl9170_tx_get_skb(struct sk_buff *skb)
332 {
333 	struct carl9170_tx_info *arinfo = (void *)
334 		(IEEE80211_SKB_CB(skb))->rate_driver_data;
335 	kref_get(&arinfo->ref);
336 }
337 
338 int carl9170_tx_put_skb(struct sk_buff *skb)
339 {
340 	struct carl9170_tx_info *arinfo = (void *)
341 		(IEEE80211_SKB_CB(skb))->rate_driver_data;
342 
343 	return kref_put(&arinfo->ref, carl9170_tx_release);
344 }
345 
346 /* Caller must hold the tid_info->lock & rcu_read_lock */
347 static void carl9170_tx_shift_bm(struct ar9170 *ar,
348 	struct carl9170_sta_tid *tid_info, u16 seq)
349 {
350 	u16 off;
351 
352 	off = SEQ_DIFF(seq, tid_info->bsn);
353 
354 	if (WARN_ON_ONCE(off >= CARL9170_BAW_BITS))
355 		return;
356 
357 	/*
358 	 * Sanity check. For each MPDU we set the bit in bitmap and
359 	 * clear it once we received the tx_status.
360 	 * But if the bit is already cleared then we've been bitten
361 	 * by a bug.
362 	 */
363 	WARN_ON_ONCE(!test_and_clear_bit(off, tid_info->bitmap));
364 
365 	off = SEQ_DIFF(tid_info->snx, tid_info->bsn);
366 	if (WARN_ON_ONCE(off >= CARL9170_BAW_BITS))
367 		return;
368 
369 	if (!bitmap_empty(tid_info->bitmap, off))
370 		off = find_first_bit(tid_info->bitmap, off);
371 
372 	tid_info->bsn += off;
373 	tid_info->bsn &= 0x0fff;
374 
375 	bitmap_shift_right(tid_info->bitmap, tid_info->bitmap,
376 			   off, CARL9170_BAW_BITS);
377 }
378 
379 static void carl9170_tx_status_process_ampdu(struct ar9170 *ar,
380 	struct sk_buff *skb, struct ieee80211_tx_info *txinfo)
381 {
382 	struct _carl9170_tx_superframe *super = (void *) skb->data;
383 	struct ieee80211_hdr *hdr = (void *) super->frame_data;
384 	struct ieee80211_sta *sta;
385 	struct carl9170_sta_info *sta_info;
386 	struct carl9170_sta_tid *tid_info;
387 	u8 tid;
388 
389 	if (!(txinfo->flags & IEEE80211_TX_CTL_AMPDU) ||
390 	    txinfo->flags & IEEE80211_TX_CTL_INJECTED)
391 		return;
392 
393 	rcu_read_lock();
394 	sta = __carl9170_get_tx_sta(ar, skb);
395 	if (unlikely(!sta))
396 		goto out_rcu;
397 
398 	tid = get_tid_h(hdr);
399 
400 	sta_info = (void *) sta->drv_priv;
401 	tid_info = rcu_dereference(sta_info->agg[tid]);
402 	if (!tid_info)
403 		goto out_rcu;
404 
405 	spin_lock_bh(&tid_info->lock);
406 	if (likely(tid_info->state >= CARL9170_TID_STATE_IDLE))
407 		carl9170_tx_shift_bm(ar, tid_info, get_seq_h(hdr));
408 
409 	if (sta_info->stats[tid].clear) {
410 		sta_info->stats[tid].clear = false;
411 		sta_info->stats[tid].req = false;
412 		sta_info->stats[tid].ampdu_len = 0;
413 		sta_info->stats[tid].ampdu_ack_len = 0;
414 	}
415 
416 	sta_info->stats[tid].ampdu_len++;
417 	if (txinfo->status.rates[0].count == 1)
418 		sta_info->stats[tid].ampdu_ack_len++;
419 
420 	if (!(txinfo->flags & IEEE80211_TX_STAT_ACK))
421 		sta_info->stats[tid].req = true;
422 
423 	if (super->f.mac_control & cpu_to_le16(AR9170_TX_MAC_IMM_BA)) {
424 		super->s.rix = sta_info->stats[tid].ampdu_len;
425 		super->s.cnt = sta_info->stats[tid].ampdu_ack_len;
426 		txinfo->flags |= IEEE80211_TX_STAT_AMPDU;
427 		if (sta_info->stats[tid].req)
428 			txinfo->flags |= IEEE80211_TX_STAT_AMPDU_NO_BACK;
429 
430 		sta_info->stats[tid].clear = true;
431 	}
432 	spin_unlock_bh(&tid_info->lock);
433 
434 out_rcu:
435 	rcu_read_unlock();
436 }
437 
438 static void carl9170_tx_bar_status(struct ar9170 *ar, struct sk_buff *skb,
439 	struct ieee80211_tx_info *tx_info)
440 {
441 	struct _carl9170_tx_superframe *super = (void *) skb->data;
442 	struct ieee80211_bar *bar = (void *) super->frame_data;
443 
444 	/*
445 	 * Unlike all other frames, the status report for BARs does
446 	 * not directly come from the hardware as it is incapable of
447 	 * matching a BA to a previously send BAR.
448 	 * Instead the RX-path will scan for incoming BAs and set the
449 	 * IEEE80211_TX_STAT_ACK if it sees one that was likely
450 	 * caused by a BAR from us.
451 	 */
452 
453 	if (unlikely(ieee80211_is_back_req(bar->frame_control)) &&
454 	   !(tx_info->flags & IEEE80211_TX_STAT_ACK)) {
455 		struct carl9170_bar_list_entry *entry;
456 		int queue = skb_get_queue_mapping(skb);
457 
458 		rcu_read_lock();
459 		list_for_each_entry_rcu(entry, &ar->bar_list[queue], list) {
460 			if (entry->skb == skb) {
461 				spin_lock_bh(&ar->bar_list_lock[queue]);
462 				list_del_rcu(&entry->list);
463 				spin_unlock_bh(&ar->bar_list_lock[queue]);
464 				kfree_rcu(entry, head);
465 				goto out;
466 			}
467 		}
468 
469 		WARN(1, "bar not found in %d - ra:%pM ta:%pM c:%x ssn:%x\n",
470 		       queue, bar->ra, bar->ta, bar->control,
471 			bar->start_seq_num);
472 out:
473 		rcu_read_unlock();
474 	}
475 }
476 
477 void carl9170_tx_status(struct ar9170 *ar, struct sk_buff *skb,
478 			const bool success)
479 {
480 	struct ieee80211_tx_info *txinfo;
481 
482 	carl9170_tx_accounting_free(ar, skb);
483 
484 	txinfo = IEEE80211_SKB_CB(skb);
485 
486 	carl9170_tx_bar_status(ar, skb, txinfo);
487 
488 	if (success)
489 		txinfo->flags |= IEEE80211_TX_STAT_ACK;
490 	else
491 		ar->tx_ack_failures++;
492 
493 	if (txinfo->flags & IEEE80211_TX_CTL_AMPDU)
494 		carl9170_tx_status_process_ampdu(ar, skb, txinfo);
495 
496 	carl9170_tx_ps_unblock(ar, skb);
497 	carl9170_tx_put_skb(skb);
498 }
499 
500 /* This function may be called form any context */
501 void carl9170_tx_callback(struct ar9170 *ar, struct sk_buff *skb)
502 {
503 	struct ieee80211_tx_info *txinfo = IEEE80211_SKB_CB(skb);
504 
505 	atomic_dec(&ar->tx_total_pending);
506 
507 	if (txinfo->flags & IEEE80211_TX_CTL_AMPDU)
508 		atomic_dec(&ar->tx_ampdu_upload);
509 
510 	if (carl9170_tx_put_skb(skb))
511 		tasklet_hi_schedule(&ar->usb_tasklet);
512 }
513 
514 static struct sk_buff *carl9170_get_queued_skb(struct ar9170 *ar, u8 cookie,
515 					       struct sk_buff_head *queue)
516 {
517 	struct sk_buff *skb;
518 
519 	spin_lock_bh(&queue->lock);
520 	skb_queue_walk(queue, skb) {
521 		struct _carl9170_tx_superframe *txc = (void *) skb->data;
522 
523 		if (txc->s.cookie != cookie)
524 			continue;
525 
526 		__skb_unlink(skb, queue);
527 		spin_unlock_bh(&queue->lock);
528 
529 		carl9170_release_dev_space(ar, skb);
530 		return skb;
531 	}
532 	spin_unlock_bh(&queue->lock);
533 
534 	return NULL;
535 }
536 
537 static void carl9170_tx_fill_rateinfo(struct ar9170 *ar, unsigned int rix,
538 	unsigned int tries, struct ieee80211_tx_info *txinfo)
539 {
540 	unsigned int i;
541 
542 	for (i = 0; i < IEEE80211_TX_MAX_RATES; i++) {
543 		if (txinfo->status.rates[i].idx < 0)
544 			break;
545 
546 		if (i == rix) {
547 			txinfo->status.rates[i].count = tries;
548 			i++;
549 			break;
550 		}
551 	}
552 
553 	for (; i < IEEE80211_TX_MAX_RATES; i++) {
554 		txinfo->status.rates[i].idx = -1;
555 		txinfo->status.rates[i].count = 0;
556 	}
557 }
558 
559 static void carl9170_check_queue_stop_timeout(struct ar9170 *ar)
560 {
561 	int i;
562 	struct sk_buff *skb;
563 	struct ieee80211_tx_info *txinfo;
564 	struct carl9170_tx_info *arinfo;
565 	bool restart = false;
566 
567 	for (i = 0; i < ar->hw->queues; i++) {
568 		spin_lock_bh(&ar->tx_status[i].lock);
569 
570 		skb = skb_peek(&ar->tx_status[i]);
571 
572 		if (!skb)
573 			goto next;
574 
575 		txinfo = IEEE80211_SKB_CB(skb);
576 		arinfo = (void *) txinfo->rate_driver_data;
577 
578 		if (time_is_before_jiffies(arinfo->timeout +
579 		    msecs_to_jiffies(CARL9170_QUEUE_STUCK_TIMEOUT)) == true)
580 			restart = true;
581 
582 next:
583 		spin_unlock_bh(&ar->tx_status[i].lock);
584 	}
585 
586 	if (restart) {
587 		/*
588 		 * At least one queue has been stuck for long enough.
589 		 * Give the device a kick and hope it gets back to
590 		 * work.
591 		 *
592 		 * possible reasons may include:
593 		 *  - frames got lost/corrupted (bad connection to the device)
594 		 *  - stalled rx processing/usb controller hiccups
595 		 *  - firmware errors/bugs
596 		 *  - every bug you can think of.
597 		 *  - all bugs you can't...
598 		 *  - ...
599 		 */
600 		carl9170_restart(ar, CARL9170_RR_STUCK_TX);
601 	}
602 }
603 
604 static void carl9170_tx_ampdu_timeout(struct ar9170 *ar)
605 {
606 	struct carl9170_sta_tid *iter;
607 	struct sk_buff *skb;
608 	struct ieee80211_tx_info *txinfo;
609 	struct carl9170_tx_info *arinfo;
610 	struct ieee80211_sta *sta;
611 
612 	rcu_read_lock();
613 	list_for_each_entry_rcu(iter, &ar->tx_ampdu_list, list) {
614 		if (iter->state < CARL9170_TID_STATE_IDLE)
615 			continue;
616 
617 		spin_lock_bh(&iter->lock);
618 		skb = skb_peek(&iter->queue);
619 		if (!skb)
620 			goto unlock;
621 
622 		txinfo = IEEE80211_SKB_CB(skb);
623 		arinfo = (void *)txinfo->rate_driver_data;
624 		if (time_is_after_jiffies(arinfo->timeout +
625 		    msecs_to_jiffies(CARL9170_QUEUE_TIMEOUT)))
626 			goto unlock;
627 
628 		sta = __carl9170_get_tx_sta(ar, skb);
629 		if (WARN_ON(!sta))
630 			goto unlock;
631 
632 		ieee80211_stop_tx_ba_session(sta, iter->tid);
633 unlock:
634 		spin_unlock_bh(&iter->lock);
635 
636 	}
637 	rcu_read_unlock();
638 }
639 
640 void carl9170_tx_janitor(struct work_struct *work)
641 {
642 	struct ar9170 *ar = container_of(work, struct ar9170,
643 					 tx_janitor.work);
644 	if (!IS_STARTED(ar))
645 		return;
646 
647 	ar->tx_janitor_last_run = jiffies;
648 
649 	carl9170_check_queue_stop_timeout(ar);
650 	carl9170_tx_ampdu_timeout(ar);
651 
652 	if (!atomic_read(&ar->tx_total_queued))
653 		return;
654 
655 	ieee80211_queue_delayed_work(ar->hw, &ar->tx_janitor,
656 		msecs_to_jiffies(CARL9170_TX_TIMEOUT));
657 }
658 
659 static void __carl9170_tx_process_status(struct ar9170 *ar,
660 	const uint8_t cookie, const uint8_t info)
661 {
662 	struct sk_buff *skb;
663 	struct ieee80211_tx_info *txinfo;
664 	unsigned int r, t, q;
665 	bool success = true;
666 
667 	q = ar9170_qmap[info & CARL9170_TX_STATUS_QUEUE];
668 
669 	skb = carl9170_get_queued_skb(ar, cookie, &ar->tx_status[q]);
670 	if (!skb) {
671 		/*
672 		 * We have lost the race to another thread.
673 		 */
674 
675 		return ;
676 	}
677 
678 	txinfo = IEEE80211_SKB_CB(skb);
679 
680 	if (!(info & CARL9170_TX_STATUS_SUCCESS))
681 		success = false;
682 
683 	r = (info & CARL9170_TX_STATUS_RIX) >> CARL9170_TX_STATUS_RIX_S;
684 	t = (info & CARL9170_TX_STATUS_TRIES) >> CARL9170_TX_STATUS_TRIES_S;
685 
686 	carl9170_tx_fill_rateinfo(ar, r, t, txinfo);
687 	carl9170_tx_status(ar, skb, success);
688 }
689 
690 void carl9170_tx_process_status(struct ar9170 *ar,
691 				const struct carl9170_rsp *cmd)
692 {
693 	unsigned int i;
694 
695 	for (i = 0;  i < cmd->hdr.ext; i++) {
696 		if (WARN_ON(i > ((cmd->hdr.len / 2) + 1))) {
697 			print_hex_dump_bytes("UU:", DUMP_PREFIX_NONE,
698 					     (void *) cmd, cmd->hdr.len + 4);
699 			break;
700 		}
701 
702 		__carl9170_tx_process_status(ar, cmd->_tx_status[i].cookie,
703 					     cmd->_tx_status[i].info);
704 	}
705 }
706 
707 static void carl9170_tx_rate_tpc_chains(struct ar9170 *ar,
708 	struct ieee80211_tx_info *info,	struct ieee80211_tx_rate *txrate,
709 	unsigned int *phyrate, unsigned int *tpc, unsigned int *chains)
710 {
711 	struct ieee80211_rate *rate = NULL;
712 	u8 *txpower;
713 	unsigned int idx;
714 
715 	idx = txrate->idx;
716 	*tpc = 0;
717 	*phyrate = 0;
718 
719 	if (txrate->flags & IEEE80211_TX_RC_MCS) {
720 		if (txrate->flags & IEEE80211_TX_RC_40_MHZ_WIDTH) {
721 			/* +1 dBm for HT40 */
722 			*tpc += 2;
723 
724 			if (info->band == IEEE80211_BAND_2GHZ)
725 				txpower = ar->power_2G_ht40;
726 			else
727 				txpower = ar->power_5G_ht40;
728 		} else {
729 			if (info->band == IEEE80211_BAND_2GHZ)
730 				txpower = ar->power_2G_ht20;
731 			else
732 				txpower = ar->power_5G_ht20;
733 		}
734 
735 		*phyrate = txrate->idx;
736 		*tpc += txpower[idx & 7];
737 	} else {
738 		if (info->band == IEEE80211_BAND_2GHZ) {
739 			if (idx < 4)
740 				txpower = ar->power_2G_cck;
741 			else
742 				txpower = ar->power_2G_ofdm;
743 		} else {
744 			txpower = ar->power_5G_leg;
745 			idx += 4;
746 		}
747 
748 		rate = &__carl9170_ratetable[idx];
749 		*tpc += txpower[(rate->hw_value & 0x30) >> 4];
750 		*phyrate = rate->hw_value & 0xf;
751 	}
752 
753 	if (ar->eeprom.tx_mask == 1) {
754 		*chains = AR9170_TX_PHY_TXCHAIN_1;
755 	} else {
756 		if (!(txrate->flags & IEEE80211_TX_RC_MCS) &&
757 		    rate && rate->bitrate >= 360)
758 			*chains = AR9170_TX_PHY_TXCHAIN_1;
759 		else
760 			*chains = AR9170_TX_PHY_TXCHAIN_2;
761 	}
762 
763 	*tpc = min_t(unsigned int, *tpc, ar->hw->conf.power_level * 2);
764 }
765 
766 static __le32 carl9170_tx_physet(struct ar9170 *ar,
767 	struct ieee80211_tx_info *info, struct ieee80211_tx_rate *txrate)
768 {
769 	unsigned int power = 0, chains = 0, phyrate = 0;
770 	__le32 tmp;
771 
772 	tmp = cpu_to_le32(0);
773 
774 	if (txrate->flags & IEEE80211_TX_RC_40_MHZ_WIDTH)
775 		tmp |= cpu_to_le32(AR9170_TX_PHY_BW_40MHZ <<
776 			AR9170_TX_PHY_BW_S);
777 	/* this works because 40 MHz is 2 and dup is 3 */
778 	if (txrate->flags & IEEE80211_TX_RC_DUP_DATA)
779 		tmp |= cpu_to_le32(AR9170_TX_PHY_BW_40MHZ_DUP <<
780 			AR9170_TX_PHY_BW_S);
781 
782 	if (txrate->flags & IEEE80211_TX_RC_SHORT_GI)
783 		tmp |= cpu_to_le32(AR9170_TX_PHY_SHORT_GI);
784 
785 	if (txrate->flags & IEEE80211_TX_RC_MCS) {
786 		SET_VAL(AR9170_TX_PHY_MCS, phyrate, txrate->idx);
787 
788 		/* heavy clip control */
789 		tmp |= cpu_to_le32((txrate->idx & 0x7) <<
790 			AR9170_TX_PHY_TX_HEAVY_CLIP_S);
791 
792 		tmp |= cpu_to_le32(AR9170_TX_PHY_MOD_HT);
793 
794 		/*
795 		 * green field preamble does not work.
796 		 *
797 		 * if (txrate->flags & IEEE80211_TX_RC_GREEN_FIELD)
798 		 * tmp |= cpu_to_le32(AR9170_TX_PHY_GREENFIELD);
799 		 */
800 	} else {
801 		if (info->band == IEEE80211_BAND_2GHZ) {
802 			if (txrate->idx <= AR9170_TX_PHY_RATE_CCK_11M)
803 				tmp |= cpu_to_le32(AR9170_TX_PHY_MOD_CCK);
804 			else
805 				tmp |= cpu_to_le32(AR9170_TX_PHY_MOD_OFDM);
806 		} else {
807 			tmp |= cpu_to_le32(AR9170_TX_PHY_MOD_OFDM);
808 		}
809 
810 		/*
811 		 * short preamble seems to be broken too.
812 		 *
813 		 * if (txrate->flags & IEEE80211_TX_RC_USE_SHORT_PREAMBLE)
814 		 *	tmp |= cpu_to_le32(AR9170_TX_PHY_SHORT_PREAMBLE);
815 		 */
816 	}
817 	carl9170_tx_rate_tpc_chains(ar, info, txrate,
818 				    &phyrate, &power, &chains);
819 
820 	tmp |= cpu_to_le32(SET_CONSTVAL(AR9170_TX_PHY_MCS, phyrate));
821 	tmp |= cpu_to_le32(SET_CONSTVAL(AR9170_TX_PHY_TX_PWR, power));
822 	tmp |= cpu_to_le32(SET_CONSTVAL(AR9170_TX_PHY_TXCHAIN, chains));
823 	return tmp;
824 }
825 
826 static bool carl9170_tx_rts_check(struct ar9170 *ar,
827 				  struct ieee80211_tx_rate *rate,
828 				  bool ampdu, bool multi)
829 {
830 	switch (ar->erp_mode) {
831 	case CARL9170_ERP_AUTO:
832 		if (ampdu)
833 			break;
834 
835 	case CARL9170_ERP_MAC80211:
836 		if (!(rate->flags & IEEE80211_TX_RC_USE_RTS_CTS))
837 			break;
838 
839 	case CARL9170_ERP_RTS:
840 		if (likely(!multi))
841 			return true;
842 
843 	default:
844 		break;
845 	}
846 
847 	return false;
848 }
849 
850 static bool carl9170_tx_cts_check(struct ar9170 *ar,
851 				  struct ieee80211_tx_rate *rate)
852 {
853 	switch (ar->erp_mode) {
854 	case CARL9170_ERP_AUTO:
855 	case CARL9170_ERP_MAC80211:
856 		if (!(rate->flags & IEEE80211_TX_RC_USE_CTS_PROTECT))
857 			break;
858 
859 	case CARL9170_ERP_CTS:
860 		return true;
861 
862 	default:
863 		break;
864 	}
865 
866 	return false;
867 }
868 
869 static int carl9170_tx_prepare(struct ar9170 *ar,
870 			       struct ieee80211_sta *sta,
871 			       struct sk_buff *skb)
872 {
873 	struct ieee80211_hdr *hdr;
874 	struct _carl9170_tx_superframe *txc;
875 	struct carl9170_vif_info *cvif;
876 	struct ieee80211_tx_info *info;
877 	struct ieee80211_tx_rate *txrate;
878 	struct carl9170_tx_info *arinfo;
879 	unsigned int hw_queue;
880 	int i;
881 	__le16 mac_tmp;
882 	u16 len;
883 	bool ampdu, no_ack;
884 
885 	BUILD_BUG_ON(sizeof(*arinfo) > sizeof(info->rate_driver_data));
886 	BUILD_BUG_ON(sizeof(struct _carl9170_tx_superdesc) !=
887 		     CARL9170_TX_SUPERDESC_LEN);
888 
889 	BUILD_BUG_ON(sizeof(struct _ar9170_tx_hwdesc) !=
890 		     AR9170_TX_HWDESC_LEN);
891 
892 	BUILD_BUG_ON(IEEE80211_TX_MAX_RATES < CARL9170_TX_MAX_RATES);
893 
894 	BUILD_BUG_ON(AR9170_MAX_VIRTUAL_MAC >
895 		((CARL9170_TX_SUPER_MISC_VIF_ID >>
896 		 CARL9170_TX_SUPER_MISC_VIF_ID_S) + 1));
897 
898 	hw_queue = ar9170_qmap[carl9170_get_queue(ar, skb)];
899 
900 	hdr = (void *)skb->data;
901 	info = IEEE80211_SKB_CB(skb);
902 	len = skb->len;
903 
904 	/*
905 	 * Note: If the frame was sent through a monitor interface,
906 	 * the ieee80211_vif pointer can be NULL.
907 	 */
908 	if (likely(info->control.vif))
909 		cvif = (void *) info->control.vif->drv_priv;
910 	else
911 		cvif = NULL;
912 
913 	txc = (void *)skb_push(skb, sizeof(*txc));
914 	memset(txc, 0, sizeof(*txc));
915 
916 	SET_VAL(CARL9170_TX_SUPER_MISC_QUEUE, txc->s.misc, hw_queue);
917 
918 	if (likely(cvif))
919 		SET_VAL(CARL9170_TX_SUPER_MISC_VIF_ID, txc->s.misc, cvif->id);
920 
921 	if (unlikely(info->flags & IEEE80211_TX_CTL_SEND_AFTER_DTIM))
922 		txc->s.misc |= CARL9170_TX_SUPER_MISC_CAB;
923 
924 	if (unlikely(info->flags & IEEE80211_TX_CTL_ASSIGN_SEQ))
925 		txc->s.misc |= CARL9170_TX_SUPER_MISC_ASSIGN_SEQ;
926 
927 	if (unlikely(ieee80211_is_probe_resp(hdr->frame_control)))
928 		txc->s.misc |= CARL9170_TX_SUPER_MISC_FILL_IN_TSF;
929 
930 	mac_tmp = cpu_to_le16(AR9170_TX_MAC_HW_DURATION |
931 			      AR9170_TX_MAC_BACKOFF);
932 	mac_tmp |= cpu_to_le16((hw_queue << AR9170_TX_MAC_QOS_S) &
933 			       AR9170_TX_MAC_QOS);
934 
935 	no_ack = !!(info->flags & IEEE80211_TX_CTL_NO_ACK);
936 	if (unlikely(no_ack))
937 		mac_tmp |= cpu_to_le16(AR9170_TX_MAC_NO_ACK);
938 
939 	if (info->control.hw_key) {
940 		len += info->control.hw_key->icv_len;
941 
942 		switch (info->control.hw_key->cipher) {
943 		case WLAN_CIPHER_SUITE_WEP40:
944 		case WLAN_CIPHER_SUITE_WEP104:
945 		case WLAN_CIPHER_SUITE_TKIP:
946 			mac_tmp |= cpu_to_le16(AR9170_TX_MAC_ENCR_RC4);
947 			break;
948 		case WLAN_CIPHER_SUITE_CCMP:
949 			mac_tmp |= cpu_to_le16(AR9170_TX_MAC_ENCR_AES);
950 			break;
951 		default:
952 			WARN_ON(1);
953 			goto err_out;
954 		}
955 	}
956 
957 	ampdu = !!(info->flags & IEEE80211_TX_CTL_AMPDU);
958 	if (ampdu) {
959 		unsigned int density, factor;
960 
961 		if (unlikely(!sta || !cvif))
962 			goto err_out;
963 
964 		factor = min_t(unsigned int, 1u, sta->ht_cap.ampdu_factor);
965 		density = sta->ht_cap.ampdu_density;
966 
967 		if (density) {
968 			/*
969 			 * Watch out!
970 			 *
971 			 * Otus uses slightly different density values than
972 			 * those from the 802.11n spec.
973 			 */
974 
975 			density = max_t(unsigned int, density + 1, 7u);
976 		}
977 
978 		SET_VAL(CARL9170_TX_SUPER_AMPDU_DENSITY,
979 			txc->s.ampdu_settings, density);
980 
981 		SET_VAL(CARL9170_TX_SUPER_AMPDU_FACTOR,
982 			txc->s.ampdu_settings, factor);
983 	}
984 
985 	/*
986 	 * NOTE: For the first rate, the ERP & AMPDU flags are directly
987 	 * taken from mac_control. For all fallback rate, the firmware
988 	 * updates the mac_control flags from the rate info field.
989 	 */
990 	for (i = 0; i < CARL9170_TX_MAX_RATES; i++) {
991 		__le32 phy_set;
992 		txrate = &info->control.rates[i];
993 		if (txrate->idx < 0)
994 			break;
995 
996 		phy_set = carl9170_tx_physet(ar, info, txrate);
997 		if (i == 0) {
998 			/* first rate - part of the hw's frame header */
999 			txc->f.phy_control = phy_set;
1000 
1001 			if (ampdu && txrate->flags & IEEE80211_TX_RC_MCS)
1002 				mac_tmp |= cpu_to_le16(AR9170_TX_MAC_AGGR);
1003 			if (carl9170_tx_rts_check(ar, txrate, ampdu, no_ack))
1004 				mac_tmp |= cpu_to_le16(AR9170_TX_MAC_PROT_RTS);
1005 			else if (carl9170_tx_cts_check(ar, txrate))
1006 				mac_tmp |= cpu_to_le16(AR9170_TX_MAC_PROT_CTS);
1007 
1008 		} else {
1009 			/* fallback rates are stored in the firmware's
1010 			 * retry rate set array.
1011 			 */
1012 			txc->s.rr[i - 1] = phy_set;
1013 		}
1014 
1015 		SET_VAL(CARL9170_TX_SUPER_RI_TRIES, txc->s.ri[i],
1016 			txrate->count);
1017 
1018 		if (carl9170_tx_rts_check(ar, txrate, ampdu, no_ack))
1019 			txc->s.ri[i] |= (AR9170_TX_MAC_PROT_RTS <<
1020 				CARL9170_TX_SUPER_RI_ERP_PROT_S);
1021 		else if (carl9170_tx_cts_check(ar, txrate))
1022 			txc->s.ri[i] |= (AR9170_TX_MAC_PROT_CTS <<
1023 				CARL9170_TX_SUPER_RI_ERP_PROT_S);
1024 
1025 		if (ampdu && (txrate->flags & IEEE80211_TX_RC_MCS))
1026 			txc->s.ri[i] |= CARL9170_TX_SUPER_RI_AMPDU;
1027 	}
1028 
1029 	txc->s.len = cpu_to_le16(skb->len);
1030 	txc->f.length = cpu_to_le16(len + FCS_LEN);
1031 	txc->f.mac_control = mac_tmp;
1032 
1033 	arinfo = (void *)info->rate_driver_data;
1034 	arinfo->timeout = jiffies;
1035 	arinfo->ar = ar;
1036 	kref_init(&arinfo->ref);
1037 	return 0;
1038 
1039 err_out:
1040 	skb_pull(skb, sizeof(*txc));
1041 	return -EINVAL;
1042 }
1043 
1044 static void carl9170_set_immba(struct ar9170 *ar, struct sk_buff *skb)
1045 {
1046 	struct _carl9170_tx_superframe *super;
1047 
1048 	super = (void *) skb->data;
1049 	super->f.mac_control |= cpu_to_le16(AR9170_TX_MAC_IMM_BA);
1050 }
1051 
1052 static void carl9170_set_ampdu_params(struct ar9170 *ar, struct sk_buff *skb)
1053 {
1054 	struct _carl9170_tx_superframe *super;
1055 	int tmp;
1056 
1057 	super = (void *) skb->data;
1058 
1059 	tmp = (super->s.ampdu_settings & CARL9170_TX_SUPER_AMPDU_DENSITY) <<
1060 		CARL9170_TX_SUPER_AMPDU_DENSITY_S;
1061 
1062 	/*
1063 	 * If you haven't noticed carl9170_tx_prepare has already filled
1064 	 * in all ampdu spacing & factor parameters.
1065 	 * Now it's the time to check whenever the settings have to be
1066 	 * updated by the firmware, or if everything is still the same.
1067 	 *
1068 	 * There's no sane way to handle different density values with
1069 	 * this hardware, so we may as well just do the compare in the
1070 	 * driver.
1071 	 */
1072 
1073 	if (tmp != ar->current_density) {
1074 		ar->current_density = tmp;
1075 		super->s.ampdu_settings |=
1076 			CARL9170_TX_SUPER_AMPDU_COMMIT_DENSITY;
1077 	}
1078 
1079 	tmp = (super->s.ampdu_settings & CARL9170_TX_SUPER_AMPDU_FACTOR) <<
1080 		CARL9170_TX_SUPER_AMPDU_FACTOR_S;
1081 
1082 	if (tmp != ar->current_factor) {
1083 		ar->current_factor = tmp;
1084 		super->s.ampdu_settings |=
1085 			CARL9170_TX_SUPER_AMPDU_COMMIT_FACTOR;
1086 	}
1087 }
1088 
1089 static bool carl9170_tx_rate_check(struct ar9170 *ar, struct sk_buff *_dest,
1090 				   struct sk_buff *_src)
1091 {
1092 	struct _carl9170_tx_superframe *dest, *src;
1093 
1094 	dest = (void *) _dest->data;
1095 	src = (void *) _src->data;
1096 
1097 	/*
1098 	 * The mac80211 rate control algorithm expects that all MPDUs in
1099 	 * an AMPDU share the same tx vectors.
1100 	 * This is not really obvious right now, because the hardware
1101 	 * does the AMPDU setup according to its own rulebook.
1102 	 * Our nicely assembled, strictly monotonic increasing mpdu
1103 	 * chains will be broken up, mashed back together...
1104 	 */
1105 
1106 	return (dest->f.phy_control == src->f.phy_control);
1107 }
1108 
1109 static void carl9170_tx_ampdu(struct ar9170 *ar)
1110 {
1111 	struct sk_buff_head agg;
1112 	struct carl9170_sta_tid *tid_info;
1113 	struct sk_buff *skb, *first;
1114 	unsigned int i = 0, done_ampdus = 0;
1115 	u16 seq, queue, tmpssn;
1116 
1117 	atomic_inc(&ar->tx_ampdu_scheduler);
1118 	ar->tx_ampdu_schedule = false;
1119 
1120 	if (atomic_read(&ar->tx_ampdu_upload))
1121 		return;
1122 
1123 	if (!ar->tx_ampdu_list_len)
1124 		return;
1125 
1126 	__skb_queue_head_init(&agg);
1127 
1128 	rcu_read_lock();
1129 	tid_info = rcu_dereference(ar->tx_ampdu_iter);
1130 	if (WARN_ON_ONCE(!tid_info)) {
1131 		rcu_read_unlock();
1132 		return;
1133 	}
1134 
1135 retry:
1136 	list_for_each_entry_continue_rcu(tid_info, &ar->tx_ampdu_list, list) {
1137 		i++;
1138 
1139 		if (tid_info->state < CARL9170_TID_STATE_PROGRESS)
1140 			continue;
1141 
1142 		queue = TID_TO_WME_AC(tid_info->tid);
1143 
1144 		spin_lock_bh(&tid_info->lock);
1145 		if (tid_info->state != CARL9170_TID_STATE_XMIT)
1146 			goto processed;
1147 
1148 		tid_info->counter++;
1149 		first = skb_peek(&tid_info->queue);
1150 		tmpssn = carl9170_get_seq(first);
1151 		seq = tid_info->snx;
1152 
1153 		if (unlikely(tmpssn != seq)) {
1154 			tid_info->state = CARL9170_TID_STATE_IDLE;
1155 
1156 			goto processed;
1157 		}
1158 
1159 		while ((skb = skb_peek(&tid_info->queue))) {
1160 			/* strict 0, 1, ..., n - 1, n frame sequence order */
1161 			if (unlikely(carl9170_get_seq(skb) != seq))
1162 				break;
1163 
1164 			/* don't upload more than AMPDU FACTOR allows. */
1165 			if (unlikely(SEQ_DIFF(tid_info->snx, tid_info->bsn) >=
1166 			    (tid_info->max - 1)))
1167 				break;
1168 
1169 			if (!carl9170_tx_rate_check(ar, skb, first))
1170 				break;
1171 
1172 			atomic_inc(&ar->tx_ampdu_upload);
1173 			tid_info->snx = seq = SEQ_NEXT(seq);
1174 			__skb_unlink(skb, &tid_info->queue);
1175 
1176 			__skb_queue_tail(&agg, skb);
1177 
1178 			if (skb_queue_len(&agg) >= CARL9170_NUM_TX_AGG_MAX)
1179 				break;
1180 		}
1181 
1182 		if (skb_queue_empty(&tid_info->queue) ||
1183 		    carl9170_get_seq(skb_peek(&tid_info->queue)) !=
1184 		    tid_info->snx) {
1185 			/*
1186 			 * stop TID, if A-MPDU frames are still missing,
1187 			 * or whenever the queue is empty.
1188 			 */
1189 
1190 			tid_info->state = CARL9170_TID_STATE_IDLE;
1191 		}
1192 		done_ampdus++;
1193 
1194 processed:
1195 		spin_unlock_bh(&tid_info->lock);
1196 
1197 		if (skb_queue_empty(&agg))
1198 			continue;
1199 
1200 		/* apply ampdu spacing & factor settings */
1201 		carl9170_set_ampdu_params(ar, skb_peek(&agg));
1202 
1203 		/* set aggregation push bit */
1204 		carl9170_set_immba(ar, skb_peek_tail(&agg));
1205 
1206 		spin_lock_bh(&ar->tx_pending[queue].lock);
1207 		skb_queue_splice_tail_init(&agg, &ar->tx_pending[queue]);
1208 		spin_unlock_bh(&ar->tx_pending[queue].lock);
1209 		ar->tx_schedule = true;
1210 	}
1211 	if ((done_ampdus++ == 0) && (i++ == 0))
1212 		goto retry;
1213 
1214 	rcu_assign_pointer(ar->tx_ampdu_iter, tid_info);
1215 	rcu_read_unlock();
1216 }
1217 
1218 static struct sk_buff *carl9170_tx_pick_skb(struct ar9170 *ar,
1219 					    struct sk_buff_head *queue)
1220 {
1221 	struct sk_buff *skb;
1222 	struct ieee80211_tx_info *info;
1223 	struct carl9170_tx_info *arinfo;
1224 
1225 	BUILD_BUG_ON(sizeof(*arinfo) > sizeof(info->rate_driver_data));
1226 
1227 	spin_lock_bh(&queue->lock);
1228 	skb = skb_peek(queue);
1229 	if (unlikely(!skb))
1230 		goto err_unlock;
1231 
1232 	if (carl9170_alloc_dev_space(ar, skb))
1233 		goto err_unlock;
1234 
1235 	__skb_unlink(skb, queue);
1236 	spin_unlock_bh(&queue->lock);
1237 
1238 	info = IEEE80211_SKB_CB(skb);
1239 	arinfo = (void *) info->rate_driver_data;
1240 
1241 	arinfo->timeout = jiffies;
1242 	return skb;
1243 
1244 err_unlock:
1245 	spin_unlock_bh(&queue->lock);
1246 	return NULL;
1247 }
1248 
1249 void carl9170_tx_drop(struct ar9170 *ar, struct sk_buff *skb)
1250 {
1251 	struct _carl9170_tx_superframe *super;
1252 	uint8_t q = 0;
1253 
1254 	ar->tx_dropped++;
1255 
1256 	super = (void *)skb->data;
1257 	SET_VAL(CARL9170_TX_SUPER_MISC_QUEUE, q,
1258 		ar9170_qmap[carl9170_get_queue(ar, skb)]);
1259 	__carl9170_tx_process_status(ar, super->s.cookie, q);
1260 }
1261 
1262 static bool carl9170_tx_ps_drop(struct ar9170 *ar, struct sk_buff *skb)
1263 {
1264 	struct ieee80211_sta *sta;
1265 	struct carl9170_sta_info *sta_info;
1266 	struct ieee80211_tx_info *tx_info;
1267 
1268 	rcu_read_lock();
1269 	sta = __carl9170_get_tx_sta(ar, skb);
1270 	if (!sta)
1271 		goto out_rcu;
1272 
1273 	sta_info = (void *) sta->drv_priv;
1274 	tx_info = IEEE80211_SKB_CB(skb);
1275 
1276 	if (unlikely(sta_info->sleeping) &&
1277 	    !(tx_info->flags & (IEEE80211_TX_CTL_NO_PS_BUFFER |
1278 				IEEE80211_TX_CTL_CLEAR_PS_FILT))) {
1279 		rcu_read_unlock();
1280 
1281 		if (tx_info->flags & IEEE80211_TX_CTL_AMPDU)
1282 			atomic_dec(&ar->tx_ampdu_upload);
1283 
1284 		tx_info->flags |= IEEE80211_TX_STAT_TX_FILTERED;
1285 		carl9170_release_dev_space(ar, skb);
1286 		carl9170_tx_status(ar, skb, false);
1287 		return true;
1288 	}
1289 
1290 out_rcu:
1291 	rcu_read_unlock();
1292 	return false;
1293 }
1294 
1295 static void carl9170_bar_check(struct ar9170 *ar, struct sk_buff *skb)
1296 {
1297 	struct _carl9170_tx_superframe *super = (void *) skb->data;
1298 	struct ieee80211_bar *bar = (void *) super->frame_data;
1299 
1300 	if (unlikely(ieee80211_is_back_req(bar->frame_control)) &&
1301 	    skb->len >= sizeof(struct ieee80211_bar)) {
1302 		struct carl9170_bar_list_entry *entry;
1303 		unsigned int queue = skb_get_queue_mapping(skb);
1304 
1305 		entry = kmalloc(sizeof(*entry), GFP_ATOMIC);
1306 		if (!WARN_ON_ONCE(!entry)) {
1307 			entry->skb = skb;
1308 			spin_lock_bh(&ar->bar_list_lock[queue]);
1309 			list_add_tail_rcu(&entry->list, &ar->bar_list[queue]);
1310 			spin_unlock_bh(&ar->bar_list_lock[queue]);
1311 		}
1312 	}
1313 }
1314 
1315 static void carl9170_tx(struct ar9170 *ar)
1316 {
1317 	struct sk_buff *skb;
1318 	unsigned int i, q;
1319 	bool schedule_garbagecollector = false;
1320 
1321 	ar->tx_schedule = false;
1322 
1323 	if (unlikely(!IS_STARTED(ar)))
1324 		return;
1325 
1326 	carl9170_usb_handle_tx_err(ar);
1327 
1328 	for (i = 0; i < ar->hw->queues; i++) {
1329 		while (!skb_queue_empty(&ar->tx_pending[i])) {
1330 			skb = carl9170_tx_pick_skb(ar, &ar->tx_pending[i]);
1331 			if (unlikely(!skb))
1332 				break;
1333 
1334 			if (unlikely(carl9170_tx_ps_drop(ar, skb)))
1335 				continue;
1336 
1337 			carl9170_bar_check(ar, skb);
1338 
1339 			atomic_inc(&ar->tx_total_pending);
1340 
1341 			q = __carl9170_get_queue(ar, i);
1342 			/*
1343 			 * NB: tx_status[i] vs. tx_status[q],
1344 			 * TODO: Move into pick_skb or alloc_dev_space.
1345 			 */
1346 			skb_queue_tail(&ar->tx_status[q], skb);
1347 
1348 			/*
1349 			 * increase ref count to "2".
1350 			 * Ref counting is the easiest way to solve the
1351 			 * race between the urb's completion routine:
1352 			 *	carl9170_tx_callback
1353 			 * and wlan tx status functions:
1354 			 *	carl9170_tx_status/janitor.
1355 			 */
1356 			carl9170_tx_get_skb(skb);
1357 
1358 			carl9170_usb_tx(ar, skb);
1359 			schedule_garbagecollector = true;
1360 		}
1361 	}
1362 
1363 	if (!schedule_garbagecollector)
1364 		return;
1365 
1366 	ieee80211_queue_delayed_work(ar->hw, &ar->tx_janitor,
1367 		msecs_to_jiffies(CARL9170_TX_TIMEOUT));
1368 }
1369 
1370 static bool carl9170_tx_ampdu_queue(struct ar9170 *ar,
1371 	struct ieee80211_sta *sta, struct sk_buff *skb,
1372 	struct ieee80211_tx_info *txinfo)
1373 {
1374 	struct carl9170_sta_info *sta_info;
1375 	struct carl9170_sta_tid *agg;
1376 	struct sk_buff *iter;
1377 	u16 tid, seq, qseq, off;
1378 	bool run = false;
1379 
1380 	tid = carl9170_get_tid(skb);
1381 	seq = carl9170_get_seq(skb);
1382 	sta_info = (void *) sta->drv_priv;
1383 
1384 	rcu_read_lock();
1385 	agg = rcu_dereference(sta_info->agg[tid]);
1386 
1387 	if (!agg)
1388 		goto err_unlock_rcu;
1389 
1390 	spin_lock_bh(&agg->lock);
1391 	if (unlikely(agg->state < CARL9170_TID_STATE_IDLE))
1392 		goto err_unlock;
1393 
1394 	/* check if sequence is within the BA window */
1395 	if (unlikely(!BAW_WITHIN(agg->bsn, CARL9170_BAW_BITS, seq)))
1396 		goto err_unlock;
1397 
1398 	if (WARN_ON_ONCE(!BAW_WITHIN(agg->snx, CARL9170_BAW_BITS, seq)))
1399 		goto err_unlock;
1400 
1401 	off = SEQ_DIFF(seq, agg->bsn);
1402 	if (WARN_ON_ONCE(test_and_set_bit(off, agg->bitmap)))
1403 		goto err_unlock;
1404 
1405 	if (likely(BAW_WITHIN(agg->hsn, CARL9170_BAW_BITS, seq))) {
1406 		__skb_queue_tail(&agg->queue, skb);
1407 		agg->hsn = seq;
1408 		goto queued;
1409 	}
1410 
1411 	skb_queue_reverse_walk(&agg->queue, iter) {
1412 		qseq = carl9170_get_seq(iter);
1413 
1414 		if (BAW_WITHIN(qseq, CARL9170_BAW_BITS, seq)) {
1415 			__skb_queue_after(&agg->queue, iter, skb);
1416 			goto queued;
1417 		}
1418 	}
1419 
1420 	__skb_queue_head(&agg->queue, skb);
1421 queued:
1422 
1423 	if (unlikely(agg->state != CARL9170_TID_STATE_XMIT)) {
1424 		if (agg->snx == carl9170_get_seq(skb_peek(&agg->queue))) {
1425 			agg->state = CARL9170_TID_STATE_XMIT;
1426 			run = true;
1427 		}
1428 	}
1429 
1430 	spin_unlock_bh(&agg->lock);
1431 	rcu_read_unlock();
1432 
1433 	return run;
1434 
1435 err_unlock:
1436 	spin_unlock_bh(&agg->lock);
1437 
1438 err_unlock_rcu:
1439 	rcu_read_unlock();
1440 	txinfo->flags &= ~IEEE80211_TX_CTL_AMPDU;
1441 	carl9170_tx_status(ar, skb, false);
1442 	ar->tx_dropped++;
1443 	return false;
1444 }
1445 
1446 void carl9170_op_tx(struct ieee80211_hw *hw,
1447 		    struct ieee80211_tx_control *control,
1448 		    struct sk_buff *skb)
1449 {
1450 	struct ar9170 *ar = hw->priv;
1451 	struct ieee80211_tx_info *info;
1452 	struct ieee80211_sta *sta = control->sta;
1453 	bool run;
1454 
1455 	if (unlikely(!IS_STARTED(ar)))
1456 		goto err_free;
1457 
1458 	info = IEEE80211_SKB_CB(skb);
1459 
1460 	if (unlikely(carl9170_tx_prepare(ar, sta, skb)))
1461 		goto err_free;
1462 
1463 	carl9170_tx_accounting(ar, skb);
1464 	/*
1465 	 * from now on, one has to use carl9170_tx_status to free
1466 	 * all ressouces which are associated with the frame.
1467 	 */
1468 
1469 	if (sta) {
1470 		struct carl9170_sta_info *stai = (void *) sta->drv_priv;
1471 		atomic_inc(&stai->pending_frames);
1472 	}
1473 
1474 	if (info->flags & IEEE80211_TX_CTL_AMPDU) {
1475 		/* to static code analyzers and reviewers:
1476 		 * mac80211 guarantees that a valid "sta"
1477 		 * reference is present, if a frame is to
1478 		 * be part of an ampdu. Hence any extra
1479 		 * sta == NULL checks are redundant in this
1480 		 * special case.
1481 		 */
1482 		run = carl9170_tx_ampdu_queue(ar, sta, skb, info);
1483 		if (run)
1484 			carl9170_tx_ampdu(ar);
1485 
1486 	} else {
1487 		unsigned int queue = skb_get_queue_mapping(skb);
1488 
1489 		skb_queue_tail(&ar->tx_pending[queue], skb);
1490 	}
1491 
1492 	carl9170_tx(ar);
1493 	return;
1494 
1495 err_free:
1496 	ar->tx_dropped++;
1497 	ieee80211_free_txskb(ar->hw, skb);
1498 }
1499 
1500 void carl9170_tx_scheduler(struct ar9170 *ar)
1501 {
1502 
1503 	if (ar->tx_ampdu_schedule)
1504 		carl9170_tx_ampdu(ar);
1505 
1506 	if (ar->tx_schedule)
1507 		carl9170_tx(ar);
1508 }
1509 
1510 /* caller has to take rcu_read_lock */
1511 static struct carl9170_vif_info *carl9170_pick_beaconing_vif(struct ar9170 *ar)
1512 {
1513 	struct carl9170_vif_info *cvif;
1514 	int i = 1;
1515 
1516 	/* The AR9170 hardware has no fancy beacon queue or some
1517 	 * other scheduling mechanism. So, the driver has to make
1518 	 * due by setting the two beacon timers (pretbtt and tbtt)
1519 	 * once and then swapping the beacon address in the HW's
1520 	 * register file each time the pretbtt fires.
1521 	 */
1522 
1523 	cvif = rcu_dereference(ar->beacon_iter);
1524 	if (ar->vifs > 0 && cvif) {
1525 		do {
1526 			list_for_each_entry_continue_rcu(cvif, &ar->vif_list,
1527 							 list) {
1528 				if (cvif->active && cvif->enable_beacon)
1529 					goto out;
1530 			}
1531 		} while (ar->beacon_enabled && i--);
1532 	}
1533 
1534 out:
1535 	rcu_assign_pointer(ar->beacon_iter, cvif);
1536 	return cvif;
1537 }
1538 
1539 static bool carl9170_tx_beacon_physet(struct ar9170 *ar, struct sk_buff *skb,
1540 				      u32 *ht1, u32 *plcp)
1541 {
1542 	struct ieee80211_tx_info *txinfo;
1543 	struct ieee80211_tx_rate *rate;
1544 	unsigned int power, chains;
1545 	bool ht_rate;
1546 
1547 	txinfo = IEEE80211_SKB_CB(skb);
1548 	rate = &txinfo->control.rates[0];
1549 	ht_rate = !!(txinfo->control.rates[0].flags & IEEE80211_TX_RC_MCS);
1550 	carl9170_tx_rate_tpc_chains(ar, txinfo, rate, plcp, &power, &chains);
1551 
1552 	*ht1 = AR9170_MAC_BCN_HT1_TX_ANT0;
1553 	if (chains == AR9170_TX_PHY_TXCHAIN_2)
1554 		*ht1 |= AR9170_MAC_BCN_HT1_TX_ANT1;
1555 	SET_VAL(AR9170_MAC_BCN_HT1_PWR_CTRL, *ht1, 7);
1556 	SET_VAL(AR9170_MAC_BCN_HT1_TPC, *ht1, power);
1557 	SET_VAL(AR9170_MAC_BCN_HT1_CHAIN_MASK, *ht1, chains);
1558 
1559 	if (ht_rate) {
1560 		*ht1 |= AR9170_MAC_BCN_HT1_HT_EN;
1561 		if (rate->flags & IEEE80211_TX_RC_SHORT_GI)
1562 			*plcp |= AR9170_MAC_BCN_HT2_SGI;
1563 
1564 		if (rate->flags & IEEE80211_TX_RC_40_MHZ_WIDTH) {
1565 			*ht1 |= AR9170_MAC_BCN_HT1_BWC_40M_SHARED;
1566 			*plcp |= AR9170_MAC_BCN_HT2_BW40;
1567 		} else if (rate->flags & IEEE80211_TX_RC_DUP_DATA) {
1568 			*ht1 |= AR9170_MAC_BCN_HT1_BWC_40M_DUP;
1569 			*plcp |= AR9170_MAC_BCN_HT2_BW40;
1570 		}
1571 
1572 		SET_VAL(AR9170_MAC_BCN_HT2_LEN, *plcp, skb->len + FCS_LEN);
1573 	} else {
1574 		if (*plcp <= AR9170_TX_PHY_RATE_CCK_11M)
1575 			*plcp |= ((skb->len + FCS_LEN) << (3 + 16)) + 0x0400;
1576 		else
1577 			*plcp |= ((skb->len + FCS_LEN) << 16) + 0x0010;
1578 	}
1579 
1580 	return ht_rate;
1581 }
1582 
1583 int carl9170_update_beacon(struct ar9170 *ar, const bool submit)
1584 {
1585 	struct sk_buff *skb = NULL;
1586 	struct carl9170_vif_info *cvif;
1587 	__le32 *data, *old = NULL;
1588 	u32 word, ht1, plcp, off, addr, len;
1589 	int i = 0, err = 0;
1590 	bool ht_rate;
1591 
1592 	rcu_read_lock();
1593 	cvif = carl9170_pick_beaconing_vif(ar);
1594 	if (!cvif)
1595 		goto out_unlock;
1596 
1597 	skb = ieee80211_beacon_get_tim(ar->hw, carl9170_get_vif(cvif),
1598 		NULL, NULL);
1599 
1600 	if (!skb) {
1601 		err = -ENOMEM;
1602 		goto err_free;
1603 	}
1604 
1605 	spin_lock_bh(&ar->beacon_lock);
1606 	data = (__le32 *)skb->data;
1607 	if (cvif->beacon)
1608 		old = (__le32 *)cvif->beacon->data;
1609 
1610 	off = cvif->id * AR9170_MAC_BCN_LENGTH_MAX;
1611 	addr = ar->fw.beacon_addr + off;
1612 	len = roundup(skb->len + FCS_LEN, 4);
1613 
1614 	if ((off + len) > ar->fw.beacon_max_len) {
1615 		if (net_ratelimit()) {
1616 			wiphy_err(ar->hw->wiphy, "beacon does not "
1617 				  "fit into device memory!\n");
1618 		}
1619 		err = -EINVAL;
1620 		goto err_unlock;
1621 	}
1622 
1623 	if (len > AR9170_MAC_BCN_LENGTH_MAX) {
1624 		if (net_ratelimit()) {
1625 			wiphy_err(ar->hw->wiphy, "no support for beacons "
1626 				"bigger than %d (yours:%d).\n",
1627 				 AR9170_MAC_BCN_LENGTH_MAX, len);
1628 		}
1629 
1630 		err = -EMSGSIZE;
1631 		goto err_unlock;
1632 	}
1633 
1634 	ht_rate = carl9170_tx_beacon_physet(ar, skb, &ht1, &plcp);
1635 
1636 	carl9170_async_regwrite_begin(ar);
1637 	carl9170_async_regwrite(AR9170_MAC_REG_BCN_HT1, ht1);
1638 	if (ht_rate)
1639 		carl9170_async_regwrite(AR9170_MAC_REG_BCN_HT2, plcp);
1640 	else
1641 		carl9170_async_regwrite(AR9170_MAC_REG_BCN_PLCP, plcp);
1642 
1643 	for (i = 0; i < DIV_ROUND_UP(skb->len, 4); i++) {
1644 		/*
1645 		 * XXX: This accesses beyond skb data for up
1646 		 *	to the last 3 bytes!!
1647 		 */
1648 
1649 		if (old && (data[i] == old[i]))
1650 			continue;
1651 
1652 		word = le32_to_cpu(data[i]);
1653 		carl9170_async_regwrite(addr + 4 * i, word);
1654 	}
1655 	carl9170_async_regwrite_finish();
1656 
1657 	dev_kfree_skb_any(cvif->beacon);
1658 	cvif->beacon = NULL;
1659 
1660 	err = carl9170_async_regwrite_result();
1661 	if (!err)
1662 		cvif->beacon = skb;
1663 	spin_unlock_bh(&ar->beacon_lock);
1664 	if (err)
1665 		goto err_free;
1666 
1667 	if (submit) {
1668 		err = carl9170_bcn_ctrl(ar, cvif->id,
1669 					CARL9170_BCN_CTRL_CAB_TRIGGER,
1670 					addr, skb->len + FCS_LEN);
1671 
1672 		if (err)
1673 			goto err_free;
1674 	}
1675 out_unlock:
1676 	rcu_read_unlock();
1677 	return 0;
1678 
1679 err_unlock:
1680 	spin_unlock_bh(&ar->beacon_lock);
1681 
1682 err_free:
1683 	rcu_read_unlock();
1684 	dev_kfree_skb_any(skb);
1685 	return err;
1686 }
1687