xref: /linux/drivers/net/wireless/mediatek/mt76/dma.c (revision ae22a94997b8a03dcb3c922857c203246711f9d4)
1 // SPDX-License-Identifier: ISC
2 /*
3  * Copyright (C) 2016 Felix Fietkau <nbd@nbd.name>
4  */
5 
6 #include <linux/dma-mapping.h>
7 #include "mt76.h"
8 #include "dma.h"
9 
10 #if IS_ENABLED(CONFIG_NET_MEDIATEK_SOC_WED)
11 
12 #define Q_READ(_q, _field) ({						\
13 	u32 _offset = offsetof(struct mt76_queue_regs, _field);		\
14 	u32 _val;							\
15 	if ((_q)->flags & MT_QFLAG_WED)					\
16 		_val = mtk_wed_device_reg_read((_q)->wed,		\
17 					       ((_q)->wed_regs +	\
18 					        _offset));		\
19 	else								\
20 		_val = readl(&(_q)->regs->_field);			\
21 	_val;								\
22 })
23 
24 #define Q_WRITE(_q, _field, _val)	do {				\
25 	u32 _offset = offsetof(struct mt76_queue_regs, _field);		\
26 	if ((_q)->flags & MT_QFLAG_WED)					\
27 		mtk_wed_device_reg_write((_q)->wed,			\
28 					 ((_q)->wed_regs + _offset),	\
29 					 _val);				\
30 	else								\
31 		writel(_val, &(_q)->regs->_field);			\
32 } while (0)
33 
34 #else
35 
36 #define Q_READ(_q, _field)		readl(&(_q)->regs->_field)
37 #define Q_WRITE(_q, _field, _val)	writel(_val, &(_q)->regs->_field)
38 
39 #endif
40 
41 static struct mt76_txwi_cache *
42 mt76_alloc_txwi(struct mt76_dev *dev)
43 {
44 	struct mt76_txwi_cache *t;
45 	dma_addr_t addr;
46 	u8 *txwi;
47 	int size;
48 
49 	size = L1_CACHE_ALIGN(dev->drv->txwi_size + sizeof(*t));
50 	txwi = kzalloc(size, GFP_ATOMIC);
51 	if (!txwi)
52 		return NULL;
53 
54 	addr = dma_map_single(dev->dma_dev, txwi, dev->drv->txwi_size,
55 			      DMA_TO_DEVICE);
56 	if (unlikely(dma_mapping_error(dev->dma_dev, addr))) {
57 		kfree(txwi);
58 		return NULL;
59 	}
60 
61 	t = (struct mt76_txwi_cache *)(txwi + dev->drv->txwi_size);
62 	t->dma_addr = addr;
63 
64 	return t;
65 }
66 
67 static struct mt76_txwi_cache *
68 mt76_alloc_rxwi(struct mt76_dev *dev)
69 {
70 	struct mt76_txwi_cache *t;
71 
72 	t = kzalloc(L1_CACHE_ALIGN(sizeof(*t)), GFP_ATOMIC);
73 	if (!t)
74 		return NULL;
75 
76 	t->ptr = NULL;
77 	return t;
78 }
79 
80 static struct mt76_txwi_cache *
81 __mt76_get_txwi(struct mt76_dev *dev)
82 {
83 	struct mt76_txwi_cache *t = NULL;
84 
85 	spin_lock(&dev->lock);
86 	if (!list_empty(&dev->txwi_cache)) {
87 		t = list_first_entry(&dev->txwi_cache, struct mt76_txwi_cache,
88 				     list);
89 		list_del(&t->list);
90 	}
91 	spin_unlock(&dev->lock);
92 
93 	return t;
94 }
95 
96 static struct mt76_txwi_cache *
97 __mt76_get_rxwi(struct mt76_dev *dev)
98 {
99 	struct mt76_txwi_cache *t = NULL;
100 
101 	spin_lock_bh(&dev->wed_lock);
102 	if (!list_empty(&dev->rxwi_cache)) {
103 		t = list_first_entry(&dev->rxwi_cache, struct mt76_txwi_cache,
104 				     list);
105 		list_del(&t->list);
106 	}
107 	spin_unlock_bh(&dev->wed_lock);
108 
109 	return t;
110 }
111 
112 static struct mt76_txwi_cache *
113 mt76_get_txwi(struct mt76_dev *dev)
114 {
115 	struct mt76_txwi_cache *t = __mt76_get_txwi(dev);
116 
117 	if (t)
118 		return t;
119 
120 	return mt76_alloc_txwi(dev);
121 }
122 
123 struct mt76_txwi_cache *
124 mt76_get_rxwi(struct mt76_dev *dev)
125 {
126 	struct mt76_txwi_cache *t = __mt76_get_rxwi(dev);
127 
128 	if (t)
129 		return t;
130 
131 	return mt76_alloc_rxwi(dev);
132 }
133 EXPORT_SYMBOL_GPL(mt76_get_rxwi);
134 
135 void
136 mt76_put_txwi(struct mt76_dev *dev, struct mt76_txwi_cache *t)
137 {
138 	if (!t)
139 		return;
140 
141 	spin_lock(&dev->lock);
142 	list_add(&t->list, &dev->txwi_cache);
143 	spin_unlock(&dev->lock);
144 }
145 EXPORT_SYMBOL_GPL(mt76_put_txwi);
146 
147 void
148 mt76_put_rxwi(struct mt76_dev *dev, struct mt76_txwi_cache *t)
149 {
150 	if (!t)
151 		return;
152 
153 	spin_lock_bh(&dev->wed_lock);
154 	list_add(&t->list, &dev->rxwi_cache);
155 	spin_unlock_bh(&dev->wed_lock);
156 }
157 EXPORT_SYMBOL_GPL(mt76_put_rxwi);
158 
159 static void
160 mt76_free_pending_txwi(struct mt76_dev *dev)
161 {
162 	struct mt76_txwi_cache *t;
163 
164 	local_bh_disable();
165 	while ((t = __mt76_get_txwi(dev)) != NULL) {
166 		dma_unmap_single(dev->dma_dev, t->dma_addr, dev->drv->txwi_size,
167 				 DMA_TO_DEVICE);
168 		kfree(mt76_get_txwi_ptr(dev, t));
169 	}
170 	local_bh_enable();
171 }
172 
173 void
174 mt76_free_pending_rxwi(struct mt76_dev *dev)
175 {
176 	struct mt76_txwi_cache *t;
177 
178 	local_bh_disable();
179 	while ((t = __mt76_get_rxwi(dev)) != NULL) {
180 		if (t->ptr)
181 			mt76_put_page_pool_buf(t->ptr, false);
182 		kfree(t);
183 	}
184 	local_bh_enable();
185 }
186 EXPORT_SYMBOL_GPL(mt76_free_pending_rxwi);
187 
188 static void
189 mt76_dma_sync_idx(struct mt76_dev *dev, struct mt76_queue *q)
190 {
191 	Q_WRITE(q, desc_base, q->desc_dma);
192 	if (q->flags & MT_QFLAG_WED_RRO_EN)
193 		Q_WRITE(q, ring_size, MT_DMA_RRO_EN | q->ndesc);
194 	else
195 		Q_WRITE(q, ring_size, q->ndesc);
196 	q->head = Q_READ(q, dma_idx);
197 	q->tail = q->head;
198 }
199 
200 void __mt76_dma_queue_reset(struct mt76_dev *dev, struct mt76_queue *q,
201 			    bool reset_idx)
202 {
203 	if (!q || !q->ndesc)
204 		return;
205 
206 	if (!mt76_queue_is_wed_rro_ind(q)) {
207 		int i;
208 
209 		/* clear descriptors */
210 		for (i = 0; i < q->ndesc; i++)
211 			q->desc[i].ctrl = cpu_to_le32(MT_DMA_CTL_DMA_DONE);
212 	}
213 
214 	if (reset_idx) {
215 		Q_WRITE(q, cpu_idx, 0);
216 		Q_WRITE(q, dma_idx, 0);
217 	}
218 	mt76_dma_sync_idx(dev, q);
219 }
220 
221 void mt76_dma_queue_reset(struct mt76_dev *dev, struct mt76_queue *q)
222 {
223 	__mt76_dma_queue_reset(dev, q, true);
224 }
225 
226 static int
227 mt76_dma_add_rx_buf(struct mt76_dev *dev, struct mt76_queue *q,
228 		    struct mt76_queue_buf *buf, void *data)
229 {
230 	struct mt76_queue_entry *entry = &q->entry[q->head];
231 	struct mt76_txwi_cache *txwi = NULL;
232 	struct mt76_desc *desc;
233 	int idx = q->head;
234 	u32 buf1 = 0, ctrl;
235 	int rx_token;
236 
237 	if (mt76_queue_is_wed_rro_ind(q)) {
238 		struct mt76_wed_rro_desc *rro_desc;
239 
240 		rro_desc = (struct mt76_wed_rro_desc *)q->desc;
241 		data = &rro_desc[q->head];
242 		goto done;
243 	}
244 
245 	desc = &q->desc[q->head];
246 	ctrl = FIELD_PREP(MT_DMA_CTL_SD_LEN0, buf[0].len);
247 #ifdef CONFIG_ARCH_DMA_ADDR_T_64BIT
248 	buf1 = FIELD_PREP(MT_DMA_CTL_SDP0_H, buf->addr >> 32);
249 #endif
250 
251 	if (mt76_queue_is_wed_rx(q)) {
252 		txwi = mt76_get_rxwi(dev);
253 		if (!txwi)
254 			return -ENOMEM;
255 
256 		rx_token = mt76_rx_token_consume(dev, data, txwi, buf->addr);
257 		if (rx_token < 0) {
258 			mt76_put_rxwi(dev, txwi);
259 			return -ENOMEM;
260 		}
261 
262 		buf1 |= FIELD_PREP(MT_DMA_CTL_TOKEN, rx_token);
263 		ctrl |= MT_DMA_CTL_TO_HOST;
264 	}
265 
266 	WRITE_ONCE(desc->buf0, cpu_to_le32(buf->addr));
267 	WRITE_ONCE(desc->buf1, cpu_to_le32(buf1));
268 	WRITE_ONCE(desc->ctrl, cpu_to_le32(ctrl));
269 	WRITE_ONCE(desc->info, 0);
270 
271 done:
272 	entry->dma_addr[0] = buf->addr;
273 	entry->dma_len[0] = buf->len;
274 	entry->txwi = txwi;
275 	entry->buf = data;
276 	entry->wcid = 0xffff;
277 	entry->skip_buf1 = true;
278 	q->head = (q->head + 1) % q->ndesc;
279 	q->queued++;
280 
281 	return idx;
282 }
283 
284 static int
285 mt76_dma_add_buf(struct mt76_dev *dev, struct mt76_queue *q,
286 		 struct mt76_queue_buf *buf, int nbufs, u32 info,
287 		 struct sk_buff *skb, void *txwi)
288 {
289 	struct mt76_queue_entry *entry;
290 	struct mt76_desc *desc;
291 	int i, idx = -1;
292 	u32 ctrl, next;
293 
294 	if (txwi) {
295 		q->entry[q->head].txwi = DMA_DUMMY_DATA;
296 		q->entry[q->head].skip_buf0 = true;
297 	}
298 
299 	for (i = 0; i < nbufs; i += 2, buf += 2) {
300 		u32 buf0 = buf[0].addr, buf1 = 0;
301 
302 		idx = q->head;
303 		next = (q->head + 1) % q->ndesc;
304 
305 		desc = &q->desc[idx];
306 		entry = &q->entry[idx];
307 
308 		if (buf[0].skip_unmap)
309 			entry->skip_buf0 = true;
310 		entry->skip_buf1 = i == nbufs - 1;
311 
312 		entry->dma_addr[0] = buf[0].addr;
313 		entry->dma_len[0] = buf[0].len;
314 
315 		ctrl = FIELD_PREP(MT_DMA_CTL_SD_LEN0, buf[0].len);
316 #ifdef CONFIG_ARCH_DMA_ADDR_T_64BIT
317 		info |= FIELD_PREP(MT_DMA_CTL_SDP0_H, buf[0].addr >> 32);
318 #endif
319 		if (i < nbufs - 1) {
320 			entry->dma_addr[1] = buf[1].addr;
321 			entry->dma_len[1] = buf[1].len;
322 			buf1 = buf[1].addr;
323 			ctrl |= FIELD_PREP(MT_DMA_CTL_SD_LEN1, buf[1].len);
324 #ifdef CONFIG_ARCH_DMA_ADDR_T_64BIT
325 			info |= FIELD_PREP(MT_DMA_CTL_SDP1_H,
326 					   buf[1].addr >> 32);
327 #endif
328 			if (buf[1].skip_unmap)
329 				entry->skip_buf1 = true;
330 		}
331 
332 		if (i == nbufs - 1)
333 			ctrl |= MT_DMA_CTL_LAST_SEC0;
334 		else if (i == nbufs - 2)
335 			ctrl |= MT_DMA_CTL_LAST_SEC1;
336 
337 		WRITE_ONCE(desc->buf0, cpu_to_le32(buf0));
338 		WRITE_ONCE(desc->buf1, cpu_to_le32(buf1));
339 		WRITE_ONCE(desc->info, cpu_to_le32(info));
340 		WRITE_ONCE(desc->ctrl, cpu_to_le32(ctrl));
341 
342 		q->head = next;
343 		q->queued++;
344 	}
345 
346 	q->entry[idx].txwi = txwi;
347 	q->entry[idx].skb = skb;
348 	q->entry[idx].wcid = 0xffff;
349 
350 	return idx;
351 }
352 
353 static void
354 mt76_dma_tx_cleanup_idx(struct mt76_dev *dev, struct mt76_queue *q, int idx,
355 			struct mt76_queue_entry *prev_e)
356 {
357 	struct mt76_queue_entry *e = &q->entry[idx];
358 
359 	if (!e->skip_buf0)
360 		dma_unmap_single(dev->dma_dev, e->dma_addr[0], e->dma_len[0],
361 				 DMA_TO_DEVICE);
362 
363 	if (!e->skip_buf1)
364 		dma_unmap_single(dev->dma_dev, e->dma_addr[1], e->dma_len[1],
365 				 DMA_TO_DEVICE);
366 
367 	if (e->txwi == DMA_DUMMY_DATA)
368 		e->txwi = NULL;
369 
370 	*prev_e = *e;
371 	memset(e, 0, sizeof(*e));
372 }
373 
374 static void
375 mt76_dma_kick_queue(struct mt76_dev *dev, struct mt76_queue *q)
376 {
377 	wmb();
378 	Q_WRITE(q, cpu_idx, q->head);
379 }
380 
381 static void
382 mt76_dma_tx_cleanup(struct mt76_dev *dev, struct mt76_queue *q, bool flush)
383 {
384 	struct mt76_queue_entry entry;
385 	int last;
386 
387 	if (!q || !q->ndesc)
388 		return;
389 
390 	spin_lock_bh(&q->cleanup_lock);
391 	if (flush)
392 		last = -1;
393 	else
394 		last = Q_READ(q, dma_idx);
395 
396 	while (q->queued > 0 && q->tail != last) {
397 		mt76_dma_tx_cleanup_idx(dev, q, q->tail, &entry);
398 		mt76_queue_tx_complete(dev, q, &entry);
399 
400 		if (entry.txwi) {
401 			if (!(dev->drv->drv_flags & MT_DRV_TXWI_NO_FREE))
402 				mt76_put_txwi(dev, entry.txwi);
403 		}
404 
405 		if (!flush && q->tail == last)
406 			last = Q_READ(q, dma_idx);
407 	}
408 	spin_unlock_bh(&q->cleanup_lock);
409 
410 	if (flush) {
411 		spin_lock_bh(&q->lock);
412 		mt76_dma_sync_idx(dev, q);
413 		mt76_dma_kick_queue(dev, q);
414 		spin_unlock_bh(&q->lock);
415 	}
416 
417 	if (!q->queued)
418 		wake_up(&dev->tx_wait);
419 }
420 
421 static void *
422 mt76_dma_get_buf(struct mt76_dev *dev, struct mt76_queue *q, int idx,
423 		 int *len, u32 *info, bool *more, bool *drop)
424 {
425 	struct mt76_queue_entry *e = &q->entry[idx];
426 	struct mt76_desc *desc = &q->desc[idx];
427 	u32 ctrl, desc_info, buf1;
428 	void *buf = e->buf;
429 
430 	if (mt76_queue_is_wed_rro_ind(q))
431 		goto done;
432 
433 	ctrl = le32_to_cpu(READ_ONCE(desc->ctrl));
434 	if (len) {
435 		*len = FIELD_GET(MT_DMA_CTL_SD_LEN0, ctrl);
436 		*more = !(ctrl & MT_DMA_CTL_LAST_SEC0);
437 	}
438 
439 	desc_info = le32_to_cpu(desc->info);
440 	if (info)
441 		*info = desc_info;
442 
443 	buf1 = le32_to_cpu(desc->buf1);
444 	mt76_dma_should_drop_buf(drop, ctrl, buf1, desc_info);
445 
446 	if (mt76_queue_is_wed_rx(q)) {
447 		u32 token = FIELD_GET(MT_DMA_CTL_TOKEN, buf1);
448 		struct mt76_txwi_cache *t = mt76_rx_token_release(dev, token);
449 
450 		if (!t)
451 			return NULL;
452 
453 		dma_sync_single_for_cpu(dev->dma_dev, t->dma_addr,
454 				SKB_WITH_OVERHEAD(q->buf_size),
455 				page_pool_get_dma_dir(q->page_pool));
456 
457 		buf = t->ptr;
458 		t->dma_addr = 0;
459 		t->ptr = NULL;
460 
461 		mt76_put_rxwi(dev, t);
462 		if (drop)
463 			*drop |= !!(buf1 & MT_DMA_CTL_WO_DROP);
464 	} else {
465 		dma_sync_single_for_cpu(dev->dma_dev, e->dma_addr[0],
466 				SKB_WITH_OVERHEAD(q->buf_size),
467 				page_pool_get_dma_dir(q->page_pool));
468 	}
469 
470 done:
471 	e->buf = NULL;
472 	return buf;
473 }
474 
475 static void *
476 mt76_dma_dequeue(struct mt76_dev *dev, struct mt76_queue *q, bool flush,
477 		 int *len, u32 *info, bool *more, bool *drop)
478 {
479 	int idx = q->tail;
480 
481 	*more = false;
482 	if (!q->queued)
483 		return NULL;
484 
485 	if (mt76_queue_is_wed_rro_data(q))
486 		return NULL;
487 
488 	if (!mt76_queue_is_wed_rro_ind(q)) {
489 		if (flush)
490 			q->desc[idx].ctrl |= cpu_to_le32(MT_DMA_CTL_DMA_DONE);
491 		else if (!(q->desc[idx].ctrl & cpu_to_le32(MT_DMA_CTL_DMA_DONE)))
492 			return NULL;
493 	}
494 
495 	q->tail = (q->tail + 1) % q->ndesc;
496 	q->queued--;
497 
498 	return mt76_dma_get_buf(dev, q, idx, len, info, more, drop);
499 }
500 
501 static int
502 mt76_dma_tx_queue_skb_raw(struct mt76_dev *dev, struct mt76_queue *q,
503 			  struct sk_buff *skb, u32 tx_info)
504 {
505 	struct mt76_queue_buf buf = {};
506 	dma_addr_t addr;
507 
508 	if (test_bit(MT76_MCU_RESET, &dev->phy.state))
509 		goto error;
510 
511 	if (q->queued + 1 >= q->ndesc - 1)
512 		goto error;
513 
514 	addr = dma_map_single(dev->dma_dev, skb->data, skb->len,
515 			      DMA_TO_DEVICE);
516 	if (unlikely(dma_mapping_error(dev->dma_dev, addr)))
517 		goto error;
518 
519 	buf.addr = addr;
520 	buf.len = skb->len;
521 
522 	spin_lock_bh(&q->lock);
523 	mt76_dma_add_buf(dev, q, &buf, 1, tx_info, skb, NULL);
524 	mt76_dma_kick_queue(dev, q);
525 	spin_unlock_bh(&q->lock);
526 
527 	return 0;
528 
529 error:
530 	dev_kfree_skb(skb);
531 	return -ENOMEM;
532 }
533 
534 static int
535 mt76_dma_tx_queue_skb(struct mt76_dev *dev, struct mt76_queue *q,
536 		      enum mt76_txq_id qid, struct sk_buff *skb,
537 		      struct mt76_wcid *wcid, struct ieee80211_sta *sta)
538 {
539 	struct ieee80211_tx_status status = {
540 		.sta = sta,
541 	};
542 	struct mt76_tx_info tx_info = {
543 		.skb = skb,
544 	};
545 	struct ieee80211_hw *hw;
546 	int len, n = 0, ret = -ENOMEM;
547 	struct mt76_txwi_cache *t;
548 	struct sk_buff *iter;
549 	dma_addr_t addr;
550 	u8 *txwi;
551 
552 	if (test_bit(MT76_RESET, &dev->phy.state))
553 		goto free_skb;
554 
555 	t = mt76_get_txwi(dev);
556 	if (!t)
557 		goto free_skb;
558 
559 	txwi = mt76_get_txwi_ptr(dev, t);
560 
561 	skb->prev = skb->next = NULL;
562 	if (dev->drv->drv_flags & MT_DRV_TX_ALIGNED4_SKBS)
563 		mt76_insert_hdr_pad(skb);
564 
565 	len = skb_headlen(skb);
566 	addr = dma_map_single(dev->dma_dev, skb->data, len, DMA_TO_DEVICE);
567 	if (unlikely(dma_mapping_error(dev->dma_dev, addr)))
568 		goto free;
569 
570 	tx_info.buf[n].addr = t->dma_addr;
571 	tx_info.buf[n++].len = dev->drv->txwi_size;
572 	tx_info.buf[n].addr = addr;
573 	tx_info.buf[n++].len = len;
574 
575 	skb_walk_frags(skb, iter) {
576 		if (n == ARRAY_SIZE(tx_info.buf))
577 			goto unmap;
578 
579 		addr = dma_map_single(dev->dma_dev, iter->data, iter->len,
580 				      DMA_TO_DEVICE);
581 		if (unlikely(dma_mapping_error(dev->dma_dev, addr)))
582 			goto unmap;
583 
584 		tx_info.buf[n].addr = addr;
585 		tx_info.buf[n++].len = iter->len;
586 	}
587 	tx_info.nbuf = n;
588 
589 	if (q->queued + (tx_info.nbuf + 1) / 2 >= q->ndesc - 1) {
590 		ret = -ENOMEM;
591 		goto unmap;
592 	}
593 
594 	dma_sync_single_for_cpu(dev->dma_dev, t->dma_addr, dev->drv->txwi_size,
595 				DMA_TO_DEVICE);
596 	ret = dev->drv->tx_prepare_skb(dev, txwi, qid, wcid, sta, &tx_info);
597 	dma_sync_single_for_device(dev->dma_dev, t->dma_addr, dev->drv->txwi_size,
598 				   DMA_TO_DEVICE);
599 	if (ret < 0)
600 		goto unmap;
601 
602 	return mt76_dma_add_buf(dev, q, tx_info.buf, tx_info.nbuf,
603 				tx_info.info, tx_info.skb, t);
604 
605 unmap:
606 	for (n--; n > 0; n--)
607 		dma_unmap_single(dev->dma_dev, tx_info.buf[n].addr,
608 				 tx_info.buf[n].len, DMA_TO_DEVICE);
609 
610 free:
611 #ifdef CONFIG_NL80211_TESTMODE
612 	/* fix tx_done accounting on queue overflow */
613 	if (mt76_is_testmode_skb(dev, skb, &hw)) {
614 		struct mt76_phy *phy = hw->priv;
615 
616 		if (tx_info.skb == phy->test.tx_skb)
617 			phy->test.tx_done--;
618 	}
619 #endif
620 
621 	mt76_put_txwi(dev, t);
622 
623 free_skb:
624 	status.skb = tx_info.skb;
625 	hw = mt76_tx_status_get_hw(dev, tx_info.skb);
626 	spin_lock_bh(&dev->rx_lock);
627 	ieee80211_tx_status_ext(hw, &status);
628 	spin_unlock_bh(&dev->rx_lock);
629 
630 	return ret;
631 }
632 
633 int mt76_dma_rx_fill(struct mt76_dev *dev, struct mt76_queue *q,
634 		     bool allow_direct)
635 {
636 	int len = SKB_WITH_OVERHEAD(q->buf_size);
637 	int frames = 0;
638 
639 	if (!q->ndesc)
640 		return 0;
641 
642 	spin_lock_bh(&q->lock);
643 
644 	while (q->queued < q->ndesc - 1) {
645 		struct mt76_queue_buf qbuf = {};
646 		enum dma_data_direction dir;
647 		dma_addr_t addr;
648 		int offset;
649 		void *buf = NULL;
650 
651 		if (mt76_queue_is_wed_rro_ind(q))
652 			goto done;
653 
654 		buf = mt76_get_page_pool_buf(q, &offset, q->buf_size);
655 		if (!buf)
656 			break;
657 
658 		addr = page_pool_get_dma_addr(virt_to_head_page(buf)) + offset;
659 		dir = page_pool_get_dma_dir(q->page_pool);
660 		dma_sync_single_for_device(dev->dma_dev, addr, len, dir);
661 
662 		qbuf.addr = addr + q->buf_offset;
663 done:
664 		qbuf.len = len - q->buf_offset;
665 		qbuf.skip_unmap = false;
666 		if (mt76_dma_add_rx_buf(dev, q, &qbuf, buf) < 0) {
667 			mt76_put_page_pool_buf(buf, allow_direct);
668 			break;
669 		}
670 		frames++;
671 	}
672 
673 	if (frames || mt76_queue_is_wed_rx(q))
674 		mt76_dma_kick_queue(dev, q);
675 
676 	spin_unlock_bh(&q->lock);
677 
678 	return frames;
679 }
680 
681 static int
682 mt76_dma_alloc_queue(struct mt76_dev *dev, struct mt76_queue *q,
683 		     int idx, int n_desc, int bufsize,
684 		     u32 ring_base)
685 {
686 	int ret, size;
687 
688 	spin_lock_init(&q->lock);
689 	spin_lock_init(&q->cleanup_lock);
690 
691 	q->regs = dev->mmio.regs + ring_base + idx * MT_RING_SIZE;
692 	q->ndesc = n_desc;
693 	q->buf_size = bufsize;
694 	q->hw_idx = idx;
695 
696 	size = mt76_queue_is_wed_rro_ind(q) ? sizeof(struct mt76_wed_rro_desc)
697 					    : sizeof(struct mt76_desc);
698 	q->desc = dmam_alloc_coherent(dev->dma_dev, q->ndesc * size,
699 				      &q->desc_dma, GFP_KERNEL);
700 	if (!q->desc)
701 		return -ENOMEM;
702 
703 	if (mt76_queue_is_wed_rro_ind(q)) {
704 		struct mt76_wed_rro_desc *rro_desc;
705 		int i;
706 
707 		rro_desc = (struct mt76_wed_rro_desc *)q->desc;
708 		for (i = 0; i < q->ndesc; i++) {
709 			struct mt76_wed_rro_ind *cmd;
710 
711 			cmd = (struct mt76_wed_rro_ind *)&rro_desc[i];
712 			cmd->magic_cnt = MT_DMA_WED_IND_CMD_CNT - 1;
713 		}
714 	}
715 
716 	size = q->ndesc * sizeof(*q->entry);
717 	q->entry = devm_kzalloc(dev->dev, size, GFP_KERNEL);
718 	if (!q->entry)
719 		return -ENOMEM;
720 
721 	ret = mt76_create_page_pool(dev, q);
722 	if (ret)
723 		return ret;
724 
725 	ret = mt76_wed_dma_setup(dev, q, false);
726 	if (ret)
727 		return ret;
728 
729 	if (mtk_wed_device_active(&dev->mmio.wed)) {
730 		if ((mtk_wed_get_rx_capa(&dev->mmio.wed) && mt76_queue_is_wed_rro(q)) ||
731 		    mt76_queue_is_wed_tx_free(q))
732 			return 0;
733 	}
734 
735 	mt76_dma_queue_reset(dev, q);
736 
737 	return 0;
738 }
739 
740 static void
741 mt76_dma_rx_cleanup(struct mt76_dev *dev, struct mt76_queue *q)
742 {
743 	void *buf;
744 	bool more;
745 
746 	if (!q->ndesc)
747 		return;
748 
749 	do {
750 		spin_lock_bh(&q->lock);
751 		buf = mt76_dma_dequeue(dev, q, true, NULL, NULL, &more, NULL);
752 		spin_unlock_bh(&q->lock);
753 
754 		if (!buf)
755 			break;
756 
757 		if (!mt76_queue_is_wed_rro(q))
758 			mt76_put_page_pool_buf(buf, false);
759 	} while (1);
760 
761 	spin_lock_bh(&q->lock);
762 	if (q->rx_head) {
763 		dev_kfree_skb(q->rx_head);
764 		q->rx_head = NULL;
765 	}
766 
767 	spin_unlock_bh(&q->lock);
768 }
769 
770 static void
771 mt76_dma_rx_reset(struct mt76_dev *dev, enum mt76_rxq_id qid)
772 {
773 	struct mt76_queue *q = &dev->q_rx[qid];
774 
775 	if (!q->ndesc)
776 		return;
777 
778 	if (!mt76_queue_is_wed_rro_ind(q)) {
779 		int i;
780 
781 		for (i = 0; i < q->ndesc; i++)
782 			q->desc[i].ctrl = cpu_to_le32(MT_DMA_CTL_DMA_DONE);
783 	}
784 
785 	mt76_dma_rx_cleanup(dev, q);
786 
787 	/* reset WED rx queues */
788 	mt76_wed_dma_setup(dev, q, true);
789 
790 	if (mt76_queue_is_wed_tx_free(q))
791 		return;
792 
793 	if (mtk_wed_device_active(&dev->mmio.wed) &&
794 	    mt76_queue_is_wed_rro(q))
795 		return;
796 
797 	mt76_dma_sync_idx(dev, q);
798 	mt76_dma_rx_fill(dev, q, false);
799 }
800 
801 static void
802 mt76_add_fragment(struct mt76_dev *dev, struct mt76_queue *q, void *data,
803 		  int len, bool more, u32 info, bool allow_direct)
804 {
805 	struct sk_buff *skb = q->rx_head;
806 	struct skb_shared_info *shinfo = skb_shinfo(skb);
807 	int nr_frags = shinfo->nr_frags;
808 
809 	if (nr_frags < ARRAY_SIZE(shinfo->frags)) {
810 		struct page *page = virt_to_head_page(data);
811 		int offset = data - page_address(page) + q->buf_offset;
812 
813 		skb_add_rx_frag(skb, nr_frags, page, offset, len, q->buf_size);
814 	} else {
815 		mt76_put_page_pool_buf(data, allow_direct);
816 	}
817 
818 	if (more)
819 		return;
820 
821 	q->rx_head = NULL;
822 	if (nr_frags < ARRAY_SIZE(shinfo->frags))
823 		dev->drv->rx_skb(dev, q - dev->q_rx, skb, &info);
824 	else
825 		dev_kfree_skb(skb);
826 }
827 
828 static int
829 mt76_dma_rx_process(struct mt76_dev *dev, struct mt76_queue *q, int budget)
830 {
831 	int len, data_len, done = 0, dma_idx;
832 	struct sk_buff *skb;
833 	unsigned char *data;
834 	bool check_ddone = false;
835 	bool allow_direct = !mt76_queue_is_wed_rx(q);
836 	bool more;
837 
838 	if (IS_ENABLED(CONFIG_NET_MEDIATEK_SOC_WED) &&
839 	    mt76_queue_is_wed_tx_free(q)) {
840 		dma_idx = Q_READ(q, dma_idx);
841 		check_ddone = true;
842 	}
843 
844 	while (done < budget) {
845 		bool drop = false;
846 		u32 info;
847 
848 		if (check_ddone) {
849 			if (q->tail == dma_idx)
850 				dma_idx = Q_READ(q, dma_idx);
851 
852 			if (q->tail == dma_idx)
853 				break;
854 		}
855 
856 		data = mt76_dma_dequeue(dev, q, false, &len, &info, &more,
857 					&drop);
858 		if (!data)
859 			break;
860 
861 		if (drop)
862 			goto free_frag;
863 
864 		if (q->rx_head)
865 			data_len = q->buf_size;
866 		else
867 			data_len = SKB_WITH_OVERHEAD(q->buf_size);
868 
869 		if (data_len < len + q->buf_offset) {
870 			dev_kfree_skb(q->rx_head);
871 			q->rx_head = NULL;
872 			goto free_frag;
873 		}
874 
875 		if (q->rx_head) {
876 			mt76_add_fragment(dev, q, data, len, more, info,
877 					  allow_direct);
878 			continue;
879 		}
880 
881 		if (!more && dev->drv->rx_check &&
882 		    !(dev->drv->rx_check(dev, data, len)))
883 			goto free_frag;
884 
885 		skb = napi_build_skb(data, q->buf_size);
886 		if (!skb)
887 			goto free_frag;
888 
889 		skb_reserve(skb, q->buf_offset);
890 		skb_mark_for_recycle(skb);
891 
892 		*(u32 *)skb->cb = info;
893 
894 		__skb_put(skb, len);
895 		done++;
896 
897 		if (more) {
898 			q->rx_head = skb;
899 			continue;
900 		}
901 
902 		dev->drv->rx_skb(dev, q - dev->q_rx, skb, &info);
903 		continue;
904 
905 free_frag:
906 		mt76_put_page_pool_buf(data, allow_direct);
907 	}
908 
909 	mt76_dma_rx_fill(dev, q, true);
910 	return done;
911 }
912 
913 int mt76_dma_rx_poll(struct napi_struct *napi, int budget)
914 {
915 	struct mt76_dev *dev;
916 	int qid, done = 0, cur;
917 
918 	dev = container_of(napi->dev, struct mt76_dev, napi_dev);
919 	qid = napi - dev->napi;
920 
921 	rcu_read_lock();
922 
923 	do {
924 		cur = mt76_dma_rx_process(dev, &dev->q_rx[qid], budget - done);
925 		mt76_rx_poll_complete(dev, qid, napi);
926 		done += cur;
927 	} while (cur && done < budget);
928 
929 	rcu_read_unlock();
930 
931 	if (done < budget && napi_complete(napi))
932 		dev->drv->rx_poll_complete(dev, qid);
933 
934 	return done;
935 }
936 EXPORT_SYMBOL_GPL(mt76_dma_rx_poll);
937 
938 static int
939 mt76_dma_init(struct mt76_dev *dev,
940 	      int (*poll)(struct napi_struct *napi, int budget))
941 {
942 	int i;
943 
944 	init_dummy_netdev(&dev->napi_dev);
945 	init_dummy_netdev(&dev->tx_napi_dev);
946 	snprintf(dev->napi_dev.name, sizeof(dev->napi_dev.name), "%s",
947 		 wiphy_name(dev->hw->wiphy));
948 	dev->napi_dev.threaded = 1;
949 	init_completion(&dev->mmio.wed_reset);
950 	init_completion(&dev->mmio.wed_reset_complete);
951 
952 	mt76_for_each_q_rx(dev, i) {
953 		netif_napi_add(&dev->napi_dev, &dev->napi[i], poll);
954 		mt76_dma_rx_fill(dev, &dev->q_rx[i], false);
955 		napi_enable(&dev->napi[i]);
956 	}
957 
958 	return 0;
959 }
960 
961 static const struct mt76_queue_ops mt76_dma_ops = {
962 	.init = mt76_dma_init,
963 	.alloc = mt76_dma_alloc_queue,
964 	.reset_q = mt76_dma_queue_reset,
965 	.tx_queue_skb_raw = mt76_dma_tx_queue_skb_raw,
966 	.tx_queue_skb = mt76_dma_tx_queue_skb,
967 	.tx_cleanup = mt76_dma_tx_cleanup,
968 	.rx_cleanup = mt76_dma_rx_cleanup,
969 	.rx_reset = mt76_dma_rx_reset,
970 	.kick = mt76_dma_kick_queue,
971 };
972 
973 void mt76_dma_attach(struct mt76_dev *dev)
974 {
975 	dev->queue_ops = &mt76_dma_ops;
976 }
977 EXPORT_SYMBOL_GPL(mt76_dma_attach);
978 
979 void mt76_dma_cleanup(struct mt76_dev *dev)
980 {
981 	int i;
982 
983 	mt76_worker_disable(&dev->tx_worker);
984 	netif_napi_del(&dev->tx_napi);
985 
986 	for (i = 0; i < ARRAY_SIZE(dev->phys); i++) {
987 		struct mt76_phy *phy = dev->phys[i];
988 		int j;
989 
990 		if (!phy)
991 			continue;
992 
993 		for (j = 0; j < ARRAY_SIZE(phy->q_tx); j++)
994 			mt76_dma_tx_cleanup(dev, phy->q_tx[j], true);
995 	}
996 
997 	for (i = 0; i < ARRAY_SIZE(dev->q_mcu); i++)
998 		mt76_dma_tx_cleanup(dev, dev->q_mcu[i], true);
999 
1000 	mt76_for_each_q_rx(dev, i) {
1001 		struct mt76_queue *q = &dev->q_rx[i];
1002 
1003 		if (mtk_wed_device_active(&dev->mmio.wed) &&
1004 		    mt76_queue_is_wed_rro(q))
1005 			continue;
1006 
1007 		netif_napi_del(&dev->napi[i]);
1008 		mt76_dma_rx_cleanup(dev, q);
1009 
1010 		page_pool_destroy(q->page_pool);
1011 	}
1012 
1013 	if (mtk_wed_device_active(&dev->mmio.wed))
1014 		mtk_wed_device_detach(&dev->mmio.wed);
1015 
1016 	if (mtk_wed_device_active(&dev->mmio.wed_hif2))
1017 		mtk_wed_device_detach(&dev->mmio.wed_hif2);
1018 
1019 	mt76_free_pending_txwi(dev);
1020 	mt76_free_pending_rxwi(dev);
1021 }
1022 EXPORT_SYMBOL_GPL(mt76_dma_cleanup);
1023