xref: /linux/drivers/net/wireless/mediatek/mt76/dma.c (revision 3d0fe49454652117522f60bfbefb978ba0e5300b)
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(_dev, _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(&(_dev)->mmio.wed,	\
17 					       ((_q)->wed_regs +	\
18 					        _offset));		\
19 	else								\
20 		_val = readl(&(_q)->regs->_field);			\
21 	_val;								\
22 })
23 
24 #define Q_WRITE(_dev, _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(&(_dev)->mmio.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(_dev, _q, _field)	readl(&(_q)->regs->_field)
37 #define Q_WRITE(_dev, _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(dev, q, desc_base, q->desc_dma);
192 	Q_WRITE(dev, q, ring_size, q->ndesc);
193 	q->head = Q_READ(dev, q, dma_idx);
194 	q->tail = q->head;
195 }
196 
197 static void
198 mt76_dma_queue_reset(struct mt76_dev *dev, struct mt76_queue *q)
199 {
200 	int i;
201 
202 	if (!q || !q->ndesc)
203 		return;
204 
205 	/* clear descriptors */
206 	for (i = 0; i < q->ndesc; i++)
207 		q->desc[i].ctrl = cpu_to_le32(MT_DMA_CTL_DMA_DONE);
208 
209 	Q_WRITE(dev, q, cpu_idx, 0);
210 	Q_WRITE(dev, q, dma_idx, 0);
211 	mt76_dma_sync_idx(dev, q);
212 }
213 
214 static int
215 mt76_dma_add_rx_buf(struct mt76_dev *dev, struct mt76_queue *q,
216 		    struct mt76_queue_buf *buf, void *data)
217 {
218 	struct mt76_desc *desc = &q->desc[q->head];
219 	struct mt76_queue_entry *entry = &q->entry[q->head];
220 	struct mt76_txwi_cache *txwi = NULL;
221 	u32 buf1 = 0, ctrl;
222 	int idx = q->head;
223 	int rx_token;
224 
225 	ctrl = FIELD_PREP(MT_DMA_CTL_SD_LEN0, buf[0].len);
226 
227 	if (mt76_queue_is_wed_rx(q)) {
228 		txwi = mt76_get_rxwi(dev);
229 		if (!txwi)
230 			return -ENOMEM;
231 
232 		rx_token = mt76_rx_token_consume(dev, data, txwi, buf->addr);
233 		if (rx_token < 0) {
234 			mt76_put_rxwi(dev, txwi);
235 			return -ENOMEM;
236 		}
237 
238 		buf1 |= FIELD_PREP(MT_DMA_CTL_TOKEN, rx_token);
239 		ctrl |= MT_DMA_CTL_TO_HOST;
240 	}
241 
242 	WRITE_ONCE(desc->buf0, cpu_to_le32(buf->addr));
243 	WRITE_ONCE(desc->buf1, cpu_to_le32(buf1));
244 	WRITE_ONCE(desc->ctrl, cpu_to_le32(ctrl));
245 	WRITE_ONCE(desc->info, 0);
246 
247 	entry->dma_addr[0] = buf->addr;
248 	entry->dma_len[0] = buf->len;
249 	entry->txwi = txwi;
250 	entry->buf = data;
251 	entry->wcid = 0xffff;
252 	entry->skip_buf1 = true;
253 	q->head = (q->head + 1) % q->ndesc;
254 	q->queued++;
255 
256 	return idx;
257 }
258 
259 static int
260 mt76_dma_add_buf(struct mt76_dev *dev, struct mt76_queue *q,
261 		 struct mt76_queue_buf *buf, int nbufs, u32 info,
262 		 struct sk_buff *skb, void *txwi)
263 {
264 	struct mt76_queue_entry *entry;
265 	struct mt76_desc *desc;
266 	int i, idx = -1;
267 	u32 ctrl, next;
268 
269 	if (txwi) {
270 		q->entry[q->head].txwi = DMA_DUMMY_DATA;
271 		q->entry[q->head].skip_buf0 = true;
272 	}
273 
274 	for (i = 0; i < nbufs; i += 2, buf += 2) {
275 		u32 buf0 = buf[0].addr, buf1 = 0;
276 
277 		idx = q->head;
278 		next = (q->head + 1) % q->ndesc;
279 
280 		desc = &q->desc[idx];
281 		entry = &q->entry[idx];
282 
283 		if (buf[0].skip_unmap)
284 			entry->skip_buf0 = true;
285 		entry->skip_buf1 = i == nbufs - 1;
286 
287 		entry->dma_addr[0] = buf[0].addr;
288 		entry->dma_len[0] = buf[0].len;
289 
290 		ctrl = FIELD_PREP(MT_DMA_CTL_SD_LEN0, buf[0].len);
291 		if (i < nbufs - 1) {
292 			entry->dma_addr[1] = buf[1].addr;
293 			entry->dma_len[1] = buf[1].len;
294 			buf1 = buf[1].addr;
295 			ctrl |= FIELD_PREP(MT_DMA_CTL_SD_LEN1, buf[1].len);
296 			if (buf[1].skip_unmap)
297 				entry->skip_buf1 = true;
298 		}
299 
300 		if (i == nbufs - 1)
301 			ctrl |= MT_DMA_CTL_LAST_SEC0;
302 		else if (i == nbufs - 2)
303 			ctrl |= MT_DMA_CTL_LAST_SEC1;
304 
305 		WRITE_ONCE(desc->buf0, cpu_to_le32(buf0));
306 		WRITE_ONCE(desc->buf1, cpu_to_le32(buf1));
307 		WRITE_ONCE(desc->info, cpu_to_le32(info));
308 		WRITE_ONCE(desc->ctrl, cpu_to_le32(ctrl));
309 
310 		q->head = next;
311 		q->queued++;
312 	}
313 
314 	q->entry[idx].txwi = txwi;
315 	q->entry[idx].skb = skb;
316 	q->entry[idx].wcid = 0xffff;
317 
318 	return idx;
319 }
320 
321 static void
322 mt76_dma_tx_cleanup_idx(struct mt76_dev *dev, struct mt76_queue *q, int idx,
323 			struct mt76_queue_entry *prev_e)
324 {
325 	struct mt76_queue_entry *e = &q->entry[idx];
326 
327 	if (!e->skip_buf0)
328 		dma_unmap_single(dev->dma_dev, e->dma_addr[0], e->dma_len[0],
329 				 DMA_TO_DEVICE);
330 
331 	if (!e->skip_buf1)
332 		dma_unmap_single(dev->dma_dev, e->dma_addr[1], e->dma_len[1],
333 				 DMA_TO_DEVICE);
334 
335 	if (e->txwi == DMA_DUMMY_DATA)
336 		e->txwi = NULL;
337 
338 	*prev_e = *e;
339 	memset(e, 0, sizeof(*e));
340 }
341 
342 static void
343 mt76_dma_kick_queue(struct mt76_dev *dev, struct mt76_queue *q)
344 {
345 	wmb();
346 	Q_WRITE(dev, q, cpu_idx, q->head);
347 }
348 
349 static void
350 mt76_dma_tx_cleanup(struct mt76_dev *dev, struct mt76_queue *q, bool flush)
351 {
352 	struct mt76_queue_entry entry;
353 	int last;
354 
355 	if (!q || !q->ndesc)
356 		return;
357 
358 	spin_lock_bh(&q->cleanup_lock);
359 	if (flush)
360 		last = -1;
361 	else
362 		last = Q_READ(dev, q, dma_idx);
363 
364 	while (q->queued > 0 && q->tail != last) {
365 		mt76_dma_tx_cleanup_idx(dev, q, q->tail, &entry);
366 		mt76_queue_tx_complete(dev, q, &entry);
367 
368 		if (entry.txwi) {
369 			if (!(dev->drv->drv_flags & MT_DRV_TXWI_NO_FREE))
370 				mt76_put_txwi(dev, entry.txwi);
371 		}
372 
373 		if (!flush && q->tail == last)
374 			last = Q_READ(dev, q, dma_idx);
375 	}
376 	spin_unlock_bh(&q->cleanup_lock);
377 
378 	if (flush) {
379 		spin_lock_bh(&q->lock);
380 		mt76_dma_sync_idx(dev, q);
381 		mt76_dma_kick_queue(dev, q);
382 		spin_unlock_bh(&q->lock);
383 	}
384 
385 	if (!q->queued)
386 		wake_up(&dev->tx_wait);
387 }
388 
389 static void *
390 mt76_dma_get_buf(struct mt76_dev *dev, struct mt76_queue *q, int idx,
391 		 int *len, u32 *info, bool *more, bool *drop)
392 {
393 	struct mt76_queue_entry *e = &q->entry[idx];
394 	struct mt76_desc *desc = &q->desc[idx];
395 	void *buf;
396 
397 	if (len) {
398 		u32 ctrl = le32_to_cpu(READ_ONCE(desc->ctrl));
399 		*len = FIELD_GET(MT_DMA_CTL_SD_LEN0, ctrl);
400 		*more = !(ctrl & MT_DMA_CTL_LAST_SEC0);
401 	}
402 
403 	if (info)
404 		*info = le32_to_cpu(desc->info);
405 
406 	if (mt76_queue_is_wed_rx(q)) {
407 		u32 buf1 = le32_to_cpu(desc->buf1);
408 		u32 token = FIELD_GET(MT_DMA_CTL_TOKEN, buf1);
409 		struct mt76_txwi_cache *t = mt76_rx_token_release(dev, token);
410 
411 		if (!t)
412 			return NULL;
413 
414 		dma_sync_single_for_cpu(dev->dma_dev, t->dma_addr,
415 				SKB_WITH_OVERHEAD(q->buf_size),
416 				page_pool_get_dma_dir(q->page_pool));
417 
418 		buf = t->ptr;
419 		t->dma_addr = 0;
420 		t->ptr = NULL;
421 
422 		mt76_put_rxwi(dev, t);
423 
424 		if (drop) {
425 			u32 ctrl = le32_to_cpu(READ_ONCE(desc->ctrl));
426 
427 			*drop = !!(ctrl & (MT_DMA_CTL_TO_HOST_A |
428 					   MT_DMA_CTL_DROP));
429 
430 			*drop |= !!(buf1 & MT_DMA_CTL_WO_DROP);
431 		}
432 	} else {
433 		buf = e->buf;
434 		e->buf = NULL;
435 		dma_sync_single_for_cpu(dev->dma_dev, e->dma_addr[0],
436 				SKB_WITH_OVERHEAD(q->buf_size),
437 				page_pool_get_dma_dir(q->page_pool));
438 	}
439 
440 	return buf;
441 }
442 
443 static void *
444 mt76_dma_dequeue(struct mt76_dev *dev, struct mt76_queue *q, bool flush,
445 		 int *len, u32 *info, bool *more, bool *drop)
446 {
447 	int idx = q->tail;
448 
449 	*more = false;
450 	if (!q->queued)
451 		return NULL;
452 
453 	if (flush)
454 		q->desc[idx].ctrl |= cpu_to_le32(MT_DMA_CTL_DMA_DONE);
455 	else if (!(q->desc[idx].ctrl & cpu_to_le32(MT_DMA_CTL_DMA_DONE)))
456 		return NULL;
457 
458 	q->tail = (q->tail + 1) % q->ndesc;
459 	q->queued--;
460 
461 	return mt76_dma_get_buf(dev, q, idx, len, info, more, drop);
462 }
463 
464 static int
465 mt76_dma_tx_queue_skb_raw(struct mt76_dev *dev, struct mt76_queue *q,
466 			  struct sk_buff *skb, u32 tx_info)
467 {
468 	struct mt76_queue_buf buf = {};
469 	dma_addr_t addr;
470 
471 	if (test_bit(MT76_MCU_RESET, &dev->phy.state))
472 		goto error;
473 
474 	if (q->queued + 1 >= q->ndesc - 1)
475 		goto error;
476 
477 	addr = dma_map_single(dev->dma_dev, skb->data, skb->len,
478 			      DMA_TO_DEVICE);
479 	if (unlikely(dma_mapping_error(dev->dma_dev, addr)))
480 		goto error;
481 
482 	buf.addr = addr;
483 	buf.len = skb->len;
484 
485 	spin_lock_bh(&q->lock);
486 	mt76_dma_add_buf(dev, q, &buf, 1, tx_info, skb, NULL);
487 	mt76_dma_kick_queue(dev, q);
488 	spin_unlock_bh(&q->lock);
489 
490 	return 0;
491 
492 error:
493 	dev_kfree_skb(skb);
494 	return -ENOMEM;
495 }
496 
497 static int
498 mt76_dma_tx_queue_skb(struct mt76_dev *dev, struct mt76_queue *q,
499 		      enum mt76_txq_id qid, struct sk_buff *skb,
500 		      struct mt76_wcid *wcid, struct ieee80211_sta *sta)
501 {
502 	struct ieee80211_tx_status status = {
503 		.sta = sta,
504 	};
505 	struct mt76_tx_info tx_info = {
506 		.skb = skb,
507 	};
508 	struct ieee80211_hw *hw;
509 	int len, n = 0, ret = -ENOMEM;
510 	struct mt76_txwi_cache *t;
511 	struct sk_buff *iter;
512 	dma_addr_t addr;
513 	u8 *txwi;
514 
515 	if (test_bit(MT76_RESET, &dev->phy.state))
516 		goto free_skb;
517 
518 	t = mt76_get_txwi(dev);
519 	if (!t)
520 		goto free_skb;
521 
522 	txwi = mt76_get_txwi_ptr(dev, t);
523 
524 	skb->prev = skb->next = NULL;
525 	if (dev->drv->drv_flags & MT_DRV_TX_ALIGNED4_SKBS)
526 		mt76_insert_hdr_pad(skb);
527 
528 	len = skb_headlen(skb);
529 	addr = dma_map_single(dev->dma_dev, skb->data, len, DMA_TO_DEVICE);
530 	if (unlikely(dma_mapping_error(dev->dma_dev, addr)))
531 		goto free;
532 
533 	tx_info.buf[n].addr = t->dma_addr;
534 	tx_info.buf[n++].len = dev->drv->txwi_size;
535 	tx_info.buf[n].addr = addr;
536 	tx_info.buf[n++].len = len;
537 
538 	skb_walk_frags(skb, iter) {
539 		if (n == ARRAY_SIZE(tx_info.buf))
540 			goto unmap;
541 
542 		addr = dma_map_single(dev->dma_dev, iter->data, iter->len,
543 				      DMA_TO_DEVICE);
544 		if (unlikely(dma_mapping_error(dev->dma_dev, addr)))
545 			goto unmap;
546 
547 		tx_info.buf[n].addr = addr;
548 		tx_info.buf[n++].len = iter->len;
549 	}
550 	tx_info.nbuf = n;
551 
552 	if (q->queued + (tx_info.nbuf + 1) / 2 >= q->ndesc - 1) {
553 		ret = -ENOMEM;
554 		goto unmap;
555 	}
556 
557 	dma_sync_single_for_cpu(dev->dma_dev, t->dma_addr, dev->drv->txwi_size,
558 				DMA_TO_DEVICE);
559 	ret = dev->drv->tx_prepare_skb(dev, txwi, qid, wcid, sta, &tx_info);
560 	dma_sync_single_for_device(dev->dma_dev, t->dma_addr, dev->drv->txwi_size,
561 				   DMA_TO_DEVICE);
562 	if (ret < 0)
563 		goto unmap;
564 
565 	return mt76_dma_add_buf(dev, q, tx_info.buf, tx_info.nbuf,
566 				tx_info.info, tx_info.skb, t);
567 
568 unmap:
569 	for (n--; n > 0; n--)
570 		dma_unmap_single(dev->dma_dev, tx_info.buf[n].addr,
571 				 tx_info.buf[n].len, DMA_TO_DEVICE);
572 
573 free:
574 #ifdef CONFIG_NL80211_TESTMODE
575 	/* fix tx_done accounting on queue overflow */
576 	if (mt76_is_testmode_skb(dev, skb, &hw)) {
577 		struct mt76_phy *phy = hw->priv;
578 
579 		if (tx_info.skb == phy->test.tx_skb)
580 			phy->test.tx_done--;
581 	}
582 #endif
583 
584 	mt76_put_txwi(dev, t);
585 
586 free_skb:
587 	status.skb = tx_info.skb;
588 	hw = mt76_tx_status_get_hw(dev, tx_info.skb);
589 	spin_lock_bh(&dev->rx_lock);
590 	ieee80211_tx_status_ext(hw, &status);
591 	spin_unlock_bh(&dev->rx_lock);
592 
593 	return ret;
594 }
595 
596 static int
597 mt76_dma_rx_fill(struct mt76_dev *dev, struct mt76_queue *q,
598 		 bool allow_direct)
599 {
600 	int len = SKB_WITH_OVERHEAD(q->buf_size);
601 	int frames = 0;
602 
603 	if (!q->ndesc)
604 		return 0;
605 
606 	spin_lock_bh(&q->lock);
607 
608 	while (q->queued < q->ndesc - 1) {
609 		enum dma_data_direction dir;
610 		struct mt76_queue_buf qbuf;
611 		dma_addr_t addr;
612 		int offset;
613 		void *buf;
614 
615 		buf = mt76_get_page_pool_buf(q, &offset, q->buf_size);
616 		if (!buf)
617 			break;
618 
619 		addr = page_pool_get_dma_addr(virt_to_head_page(buf)) + offset;
620 		dir = page_pool_get_dma_dir(q->page_pool);
621 		dma_sync_single_for_device(dev->dma_dev, addr, len, dir);
622 
623 		qbuf.addr = addr + q->buf_offset;
624 		qbuf.len = len - q->buf_offset;
625 		qbuf.skip_unmap = false;
626 		if (mt76_dma_add_rx_buf(dev, q, &qbuf, buf) < 0) {
627 			mt76_put_page_pool_buf(buf, allow_direct);
628 			break;
629 		}
630 		frames++;
631 	}
632 
633 	if (frames)
634 		mt76_dma_kick_queue(dev, q);
635 
636 	spin_unlock_bh(&q->lock);
637 
638 	return frames;
639 }
640 
641 int mt76_dma_wed_setup(struct mt76_dev *dev, struct mt76_queue *q, bool reset)
642 {
643 #ifdef CONFIG_NET_MEDIATEK_SOC_WED
644 	struct mtk_wed_device *wed = &dev->mmio.wed;
645 	int ret, type, ring;
646 	u8 flags;
647 
648 	if (!q || !q->ndesc)
649 		return -EINVAL;
650 
651 	flags = q->flags;
652 	if (!mtk_wed_device_active(wed))
653 		q->flags &= ~MT_QFLAG_WED;
654 
655 	if (!(q->flags & MT_QFLAG_WED))
656 		return 0;
657 
658 	type = FIELD_GET(MT_QFLAG_WED_TYPE, q->flags);
659 	ring = FIELD_GET(MT_QFLAG_WED_RING, q->flags);
660 
661 	switch (type) {
662 	case MT76_WED_Q_TX:
663 		ret = mtk_wed_device_tx_ring_setup(wed, ring, q->regs, reset);
664 		if (!ret)
665 			q->wed_regs = wed->tx_ring[ring].reg_base;
666 		break;
667 	case MT76_WED_Q_TXFREE:
668 		/* WED txfree queue needs ring to be initialized before setup */
669 		q->flags = 0;
670 		mt76_dma_queue_reset(dev, q);
671 		mt76_dma_rx_fill(dev, q, false);
672 		q->flags = flags;
673 
674 		ret = mtk_wed_device_txfree_ring_setup(wed, q->regs);
675 		if (!ret)
676 			q->wed_regs = wed->txfree_ring.reg_base;
677 		break;
678 	case MT76_WED_Q_RX:
679 		ret = mtk_wed_device_rx_ring_setup(wed, ring, q->regs, reset);
680 		if (!ret)
681 			q->wed_regs = wed->rx_ring[ring].reg_base;
682 		break;
683 	default:
684 		ret = -EINVAL;
685 	}
686 
687 	return ret;
688 #else
689 	return 0;
690 #endif
691 }
692 EXPORT_SYMBOL_GPL(mt76_dma_wed_setup);
693 
694 static int
695 mt76_dma_alloc_queue(struct mt76_dev *dev, struct mt76_queue *q,
696 		     int idx, int n_desc, int bufsize,
697 		     u32 ring_base)
698 {
699 	int ret, size;
700 
701 	spin_lock_init(&q->lock);
702 	spin_lock_init(&q->cleanup_lock);
703 
704 	q->regs = dev->mmio.regs + ring_base + idx * MT_RING_SIZE;
705 	q->ndesc = n_desc;
706 	q->buf_size = bufsize;
707 	q->hw_idx = idx;
708 
709 	size = q->ndesc * sizeof(struct mt76_desc);
710 	q->desc = dmam_alloc_coherent(dev->dma_dev, size, &q->desc_dma, GFP_KERNEL);
711 	if (!q->desc)
712 		return -ENOMEM;
713 
714 	size = q->ndesc * sizeof(*q->entry);
715 	q->entry = devm_kzalloc(dev->dev, size, GFP_KERNEL);
716 	if (!q->entry)
717 		return -ENOMEM;
718 
719 	ret = mt76_create_page_pool(dev, q);
720 	if (ret)
721 		return ret;
722 
723 	ret = mt76_dma_wed_setup(dev, q, false);
724 	if (ret)
725 		return ret;
726 
727 	if (q->flags != MT_WED_Q_TXFREE)
728 		mt76_dma_queue_reset(dev, q);
729 
730 	return 0;
731 }
732 
733 static void
734 mt76_dma_rx_cleanup(struct mt76_dev *dev, struct mt76_queue *q)
735 {
736 	void *buf;
737 	bool more;
738 
739 	if (!q->ndesc)
740 		return;
741 
742 	do {
743 		spin_lock_bh(&q->lock);
744 		buf = mt76_dma_dequeue(dev, q, true, NULL, NULL, &more, NULL);
745 		spin_unlock_bh(&q->lock);
746 
747 		if (!buf)
748 			break;
749 
750 		mt76_put_page_pool_buf(buf, false);
751 	} while (1);
752 
753 	spin_lock_bh(&q->lock);
754 	if (q->rx_head) {
755 		dev_kfree_skb(q->rx_head);
756 		q->rx_head = NULL;
757 	}
758 
759 	spin_unlock_bh(&q->lock);
760 }
761 
762 static void
763 mt76_dma_rx_reset(struct mt76_dev *dev, enum mt76_rxq_id qid)
764 {
765 	struct mt76_queue *q = &dev->q_rx[qid];
766 	int i;
767 
768 	if (!q->ndesc)
769 		return;
770 
771 	for (i = 0; i < q->ndesc; i++)
772 		q->desc[i].ctrl = cpu_to_le32(MT_DMA_CTL_DMA_DONE);
773 
774 	mt76_dma_rx_cleanup(dev, q);
775 
776 	/* reset WED rx queues */
777 	mt76_dma_wed_setup(dev, q, true);
778 	if (q->flags != MT_WED_Q_TXFREE) {
779 		mt76_dma_sync_idx(dev, q);
780 		mt76_dma_rx_fill(dev, q, false);
781 	}
782 }
783 
784 static void
785 mt76_add_fragment(struct mt76_dev *dev, struct mt76_queue *q, void *data,
786 		  int len, bool more, u32 info)
787 {
788 	struct sk_buff *skb = q->rx_head;
789 	struct skb_shared_info *shinfo = skb_shinfo(skb);
790 	int nr_frags = shinfo->nr_frags;
791 
792 	if (nr_frags < ARRAY_SIZE(shinfo->frags)) {
793 		struct page *page = virt_to_head_page(data);
794 		int offset = data - page_address(page) + q->buf_offset;
795 
796 		skb_add_rx_frag(skb, nr_frags, page, offset, len, q->buf_size);
797 	} else {
798 		mt76_put_page_pool_buf(data, true);
799 	}
800 
801 	if (more)
802 		return;
803 
804 	q->rx_head = NULL;
805 	if (nr_frags < ARRAY_SIZE(shinfo->frags))
806 		dev->drv->rx_skb(dev, q - dev->q_rx, skb, &info);
807 	else
808 		dev_kfree_skb(skb);
809 }
810 
811 static int
812 mt76_dma_rx_process(struct mt76_dev *dev, struct mt76_queue *q, int budget)
813 {
814 	int len, data_len, done = 0, dma_idx;
815 	struct sk_buff *skb;
816 	unsigned char *data;
817 	bool check_ddone = false;
818 	bool more;
819 
820 	if (IS_ENABLED(CONFIG_NET_MEDIATEK_SOC_WED) &&
821 	    q->flags == MT_WED_Q_TXFREE) {
822 		dma_idx = Q_READ(dev, q, dma_idx);
823 		check_ddone = true;
824 	}
825 
826 	while (done < budget) {
827 		bool drop = false;
828 		u32 info;
829 
830 		if (check_ddone) {
831 			if (q->tail == dma_idx)
832 				dma_idx = Q_READ(dev, q, dma_idx);
833 
834 			if (q->tail == dma_idx)
835 				break;
836 		}
837 
838 		data = mt76_dma_dequeue(dev, q, false, &len, &info, &more,
839 					&drop);
840 		if (!data)
841 			break;
842 
843 		if (drop)
844 			goto free_frag;
845 
846 		if (q->rx_head)
847 			data_len = q->buf_size;
848 		else
849 			data_len = SKB_WITH_OVERHEAD(q->buf_size);
850 
851 		if (data_len < len + q->buf_offset) {
852 			dev_kfree_skb(q->rx_head);
853 			q->rx_head = NULL;
854 			goto free_frag;
855 		}
856 
857 		if (q->rx_head) {
858 			mt76_add_fragment(dev, q, data, len, more, info);
859 			continue;
860 		}
861 
862 		if (!more && dev->drv->rx_check &&
863 		    !(dev->drv->rx_check(dev, data, len)))
864 			goto free_frag;
865 
866 		skb = napi_build_skb(data, q->buf_size);
867 		if (!skb)
868 			goto free_frag;
869 
870 		skb_reserve(skb, q->buf_offset);
871 		skb_mark_for_recycle(skb);
872 
873 		*(u32 *)skb->cb = info;
874 
875 		__skb_put(skb, len);
876 		done++;
877 
878 		if (more) {
879 			q->rx_head = skb;
880 			continue;
881 		}
882 
883 		dev->drv->rx_skb(dev, q - dev->q_rx, skb, &info);
884 		continue;
885 
886 free_frag:
887 		mt76_put_page_pool_buf(data, true);
888 	}
889 
890 	mt76_dma_rx_fill(dev, q, true);
891 	return done;
892 }
893 
894 int mt76_dma_rx_poll(struct napi_struct *napi, int budget)
895 {
896 	struct mt76_dev *dev;
897 	int qid, done = 0, cur;
898 
899 	dev = container_of(napi->dev, struct mt76_dev, napi_dev);
900 	qid = napi - dev->napi;
901 
902 	rcu_read_lock();
903 
904 	do {
905 		cur = mt76_dma_rx_process(dev, &dev->q_rx[qid], budget - done);
906 		mt76_rx_poll_complete(dev, qid, napi);
907 		done += cur;
908 	} while (cur && done < budget);
909 
910 	rcu_read_unlock();
911 
912 	if (done < budget && napi_complete(napi))
913 		dev->drv->rx_poll_complete(dev, qid);
914 
915 	return done;
916 }
917 EXPORT_SYMBOL_GPL(mt76_dma_rx_poll);
918 
919 static int
920 mt76_dma_init(struct mt76_dev *dev,
921 	      int (*poll)(struct napi_struct *napi, int budget))
922 {
923 	int i;
924 
925 	init_dummy_netdev(&dev->napi_dev);
926 	init_dummy_netdev(&dev->tx_napi_dev);
927 	snprintf(dev->napi_dev.name, sizeof(dev->napi_dev.name), "%s",
928 		 wiphy_name(dev->hw->wiphy));
929 	dev->napi_dev.threaded = 1;
930 	init_completion(&dev->mmio.wed_reset);
931 	init_completion(&dev->mmio.wed_reset_complete);
932 
933 	mt76_for_each_q_rx(dev, i) {
934 		netif_napi_add(&dev->napi_dev, &dev->napi[i], poll);
935 		mt76_dma_rx_fill(dev, &dev->q_rx[i], false);
936 		napi_enable(&dev->napi[i]);
937 	}
938 
939 	return 0;
940 }
941 
942 static const struct mt76_queue_ops mt76_dma_ops = {
943 	.init = mt76_dma_init,
944 	.alloc = mt76_dma_alloc_queue,
945 	.reset_q = mt76_dma_queue_reset,
946 	.tx_queue_skb_raw = mt76_dma_tx_queue_skb_raw,
947 	.tx_queue_skb = mt76_dma_tx_queue_skb,
948 	.tx_cleanup = mt76_dma_tx_cleanup,
949 	.rx_cleanup = mt76_dma_rx_cleanup,
950 	.rx_reset = mt76_dma_rx_reset,
951 	.kick = mt76_dma_kick_queue,
952 };
953 
954 void mt76_dma_attach(struct mt76_dev *dev)
955 {
956 	dev->queue_ops = &mt76_dma_ops;
957 }
958 EXPORT_SYMBOL_GPL(mt76_dma_attach);
959 
960 void mt76_dma_cleanup(struct mt76_dev *dev)
961 {
962 	int i;
963 
964 	mt76_worker_disable(&dev->tx_worker);
965 	netif_napi_del(&dev->tx_napi);
966 
967 	for (i = 0; i < ARRAY_SIZE(dev->phys); i++) {
968 		struct mt76_phy *phy = dev->phys[i];
969 		int j;
970 
971 		if (!phy)
972 			continue;
973 
974 		for (j = 0; j < ARRAY_SIZE(phy->q_tx); j++)
975 			mt76_dma_tx_cleanup(dev, phy->q_tx[j], true);
976 	}
977 
978 	for (i = 0; i < ARRAY_SIZE(dev->q_mcu); i++)
979 		mt76_dma_tx_cleanup(dev, dev->q_mcu[i], true);
980 
981 	mt76_for_each_q_rx(dev, i) {
982 		struct mt76_queue *q = &dev->q_rx[i];
983 
984 		netif_napi_del(&dev->napi[i]);
985 		mt76_dma_rx_cleanup(dev, q);
986 
987 		page_pool_destroy(q->page_pool);
988 	}
989 
990 	mt76_free_pending_txwi(dev);
991 	mt76_free_pending_rxwi(dev);
992 
993 	if (mtk_wed_device_active(&dev->mmio.wed))
994 		mtk_wed_device_detach(&dev->mmio.wed);
995 }
996 EXPORT_SYMBOL_GPL(mt76_dma_cleanup);
997