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 	t = (struct mt76_txwi_cache *)(txwi + dev->drv->txwi_size);
57 	t->dma_addr = addr;
58 
59 	return t;
60 }
61 
62 static struct mt76_txwi_cache *
63 mt76_alloc_rxwi(struct mt76_dev *dev)
64 {
65 	struct mt76_txwi_cache *t;
66 
67 	t = kzalloc(L1_CACHE_ALIGN(sizeof(*t)), GFP_ATOMIC);
68 	if (!t)
69 		return NULL;
70 
71 	t->ptr = NULL;
72 	return t;
73 }
74 
75 static struct mt76_txwi_cache *
76 __mt76_get_txwi(struct mt76_dev *dev)
77 {
78 	struct mt76_txwi_cache *t = NULL;
79 
80 	spin_lock(&dev->lock);
81 	if (!list_empty(&dev->txwi_cache)) {
82 		t = list_first_entry(&dev->txwi_cache, struct mt76_txwi_cache,
83 				     list);
84 		list_del(&t->list);
85 	}
86 	spin_unlock(&dev->lock);
87 
88 	return t;
89 }
90 
91 static struct mt76_txwi_cache *
92 __mt76_get_rxwi(struct mt76_dev *dev)
93 {
94 	struct mt76_txwi_cache *t = NULL;
95 
96 	spin_lock(&dev->wed_lock);
97 	if (!list_empty(&dev->rxwi_cache)) {
98 		t = list_first_entry(&dev->rxwi_cache, struct mt76_txwi_cache,
99 				     list);
100 		list_del(&t->list);
101 	}
102 	spin_unlock(&dev->wed_lock);
103 
104 	return t;
105 }
106 
107 static struct mt76_txwi_cache *
108 mt76_get_txwi(struct mt76_dev *dev)
109 {
110 	struct mt76_txwi_cache *t = __mt76_get_txwi(dev);
111 
112 	if (t)
113 		return t;
114 
115 	return mt76_alloc_txwi(dev);
116 }
117 
118 struct mt76_txwi_cache *
119 mt76_get_rxwi(struct mt76_dev *dev)
120 {
121 	struct mt76_txwi_cache *t = __mt76_get_rxwi(dev);
122 
123 	if (t)
124 		return t;
125 
126 	return mt76_alloc_rxwi(dev);
127 }
128 EXPORT_SYMBOL_GPL(mt76_get_rxwi);
129 
130 void
131 mt76_put_txwi(struct mt76_dev *dev, struct mt76_txwi_cache *t)
132 {
133 	if (!t)
134 		return;
135 
136 	spin_lock(&dev->lock);
137 	list_add(&t->list, &dev->txwi_cache);
138 	spin_unlock(&dev->lock);
139 }
140 EXPORT_SYMBOL_GPL(mt76_put_txwi);
141 
142 void
143 mt76_put_rxwi(struct mt76_dev *dev, struct mt76_txwi_cache *t)
144 {
145 	if (!t)
146 		return;
147 
148 	spin_lock(&dev->wed_lock);
149 	list_add(&t->list, &dev->rxwi_cache);
150 	spin_unlock(&dev->wed_lock);
151 }
152 EXPORT_SYMBOL_GPL(mt76_put_rxwi);
153 
154 static void
155 mt76_free_pending_txwi(struct mt76_dev *dev)
156 {
157 	struct mt76_txwi_cache *t;
158 
159 	local_bh_disable();
160 	while ((t = __mt76_get_txwi(dev)) != NULL) {
161 		dma_unmap_single(dev->dma_dev, t->dma_addr, dev->drv->txwi_size,
162 				 DMA_TO_DEVICE);
163 		kfree(mt76_get_txwi_ptr(dev, t));
164 	}
165 	local_bh_enable();
166 }
167 
168 static void
169 mt76_free_pending_rxwi(struct mt76_dev *dev)
170 {
171 	struct mt76_txwi_cache *t;
172 
173 	local_bh_disable();
174 	while ((t = __mt76_get_rxwi(dev)) != NULL) {
175 		if (t->ptr)
176 			skb_free_frag(t->ptr);
177 		kfree(t);
178 	}
179 	local_bh_enable();
180 }
181 
182 static void
183 mt76_dma_sync_idx(struct mt76_dev *dev, struct mt76_queue *q)
184 {
185 	Q_WRITE(dev, q, desc_base, q->desc_dma);
186 	Q_WRITE(dev, q, ring_size, q->ndesc);
187 	q->head = Q_READ(dev, q, dma_idx);
188 	q->tail = q->head;
189 }
190 
191 static void
192 mt76_dma_queue_reset(struct mt76_dev *dev, struct mt76_queue *q)
193 {
194 	int i;
195 
196 	if (!q || !q->ndesc)
197 		return;
198 
199 	/* clear descriptors */
200 	for (i = 0; i < q->ndesc; i++)
201 		q->desc[i].ctrl = cpu_to_le32(MT_DMA_CTL_DMA_DONE);
202 
203 	Q_WRITE(dev, q, cpu_idx, 0);
204 	Q_WRITE(dev, q, dma_idx, 0);
205 	mt76_dma_sync_idx(dev, q);
206 }
207 
208 static int
209 mt76_dma_add_rx_buf(struct mt76_dev *dev, struct mt76_queue *q,
210 		    struct mt76_queue_buf *buf, void *data)
211 {
212 	struct mt76_desc *desc = &q->desc[q->head];
213 	struct mt76_queue_entry *entry = &q->entry[q->head];
214 	struct mt76_txwi_cache *txwi = NULL;
215 	u32 buf1 = 0, ctrl;
216 	int idx = q->head;
217 	int rx_token;
218 
219 	ctrl = FIELD_PREP(MT_DMA_CTL_SD_LEN0, buf[0].len);
220 
221 	if ((q->flags & MT_QFLAG_WED) &&
222 	    FIELD_GET(MT_QFLAG_WED_TYPE, q->flags) == MT76_WED_Q_RX) {
223 		txwi = mt76_get_rxwi(dev);
224 		if (!txwi)
225 			return -ENOMEM;
226 
227 		rx_token = mt76_rx_token_consume(dev, data, txwi, buf->addr);
228 		if (rx_token < 0) {
229 			mt76_put_rxwi(dev, txwi);
230 			return -ENOMEM;
231 		}
232 
233 		buf1 |= FIELD_PREP(MT_DMA_CTL_TOKEN, rx_token);
234 		ctrl |= MT_DMA_CTL_TO_HOST;
235 	}
236 
237 	WRITE_ONCE(desc->buf0, cpu_to_le32(buf->addr));
238 	WRITE_ONCE(desc->buf1, cpu_to_le32(buf1));
239 	WRITE_ONCE(desc->ctrl, cpu_to_le32(ctrl));
240 	WRITE_ONCE(desc->info, 0);
241 
242 	entry->dma_addr[0] = buf->addr;
243 	entry->dma_len[0] = buf->len;
244 	entry->txwi = txwi;
245 	entry->buf = data;
246 	entry->wcid = 0xffff;
247 	entry->skip_buf1 = true;
248 	q->head = (q->head + 1) % q->ndesc;
249 	q->queued++;
250 
251 	return idx;
252 }
253 
254 static int
255 mt76_dma_add_buf(struct mt76_dev *dev, struct mt76_queue *q,
256 		 struct mt76_queue_buf *buf, int nbufs, u32 info,
257 		 struct sk_buff *skb, void *txwi)
258 {
259 	struct mt76_queue_entry *entry;
260 	struct mt76_desc *desc;
261 	int i, idx = -1;
262 	u32 ctrl, next;
263 
264 	if (txwi) {
265 		q->entry[q->head].txwi = DMA_DUMMY_DATA;
266 		q->entry[q->head].skip_buf0 = true;
267 	}
268 
269 	for (i = 0; i < nbufs; i += 2, buf += 2) {
270 		u32 buf0 = buf[0].addr, buf1 = 0;
271 
272 		idx = q->head;
273 		next = (q->head + 1) % q->ndesc;
274 
275 		desc = &q->desc[idx];
276 		entry = &q->entry[idx];
277 
278 		if (buf[0].skip_unmap)
279 			entry->skip_buf0 = true;
280 		entry->skip_buf1 = i == nbufs - 1;
281 
282 		entry->dma_addr[0] = buf[0].addr;
283 		entry->dma_len[0] = buf[0].len;
284 
285 		ctrl = FIELD_PREP(MT_DMA_CTL_SD_LEN0, buf[0].len);
286 		if (i < nbufs - 1) {
287 			entry->dma_addr[1] = buf[1].addr;
288 			entry->dma_len[1] = buf[1].len;
289 			buf1 = buf[1].addr;
290 			ctrl |= FIELD_PREP(MT_DMA_CTL_SD_LEN1, buf[1].len);
291 			if (buf[1].skip_unmap)
292 				entry->skip_buf1 = true;
293 		}
294 
295 		if (i == nbufs - 1)
296 			ctrl |= MT_DMA_CTL_LAST_SEC0;
297 		else if (i == nbufs - 2)
298 			ctrl |= MT_DMA_CTL_LAST_SEC1;
299 
300 		WRITE_ONCE(desc->buf0, cpu_to_le32(buf0));
301 		WRITE_ONCE(desc->buf1, cpu_to_le32(buf1));
302 		WRITE_ONCE(desc->info, cpu_to_le32(info));
303 		WRITE_ONCE(desc->ctrl, cpu_to_le32(ctrl));
304 
305 		q->head = next;
306 		q->queued++;
307 	}
308 
309 	q->entry[idx].txwi = txwi;
310 	q->entry[idx].skb = skb;
311 	q->entry[idx].wcid = 0xffff;
312 
313 	return idx;
314 }
315 
316 static void
317 mt76_dma_tx_cleanup_idx(struct mt76_dev *dev, struct mt76_queue *q, int idx,
318 			struct mt76_queue_entry *prev_e)
319 {
320 	struct mt76_queue_entry *e = &q->entry[idx];
321 
322 	if (!e->skip_buf0)
323 		dma_unmap_single(dev->dma_dev, e->dma_addr[0], e->dma_len[0],
324 				 DMA_TO_DEVICE);
325 
326 	if (!e->skip_buf1)
327 		dma_unmap_single(dev->dma_dev, e->dma_addr[1], e->dma_len[1],
328 				 DMA_TO_DEVICE);
329 
330 	if (e->txwi == DMA_DUMMY_DATA)
331 		e->txwi = NULL;
332 
333 	if (e->skb == DMA_DUMMY_DATA)
334 		e->skb = NULL;
335 
336 	*prev_e = *e;
337 	memset(e, 0, sizeof(*e));
338 }
339 
340 static void
341 mt76_dma_kick_queue(struct mt76_dev *dev, struct mt76_queue *q)
342 {
343 	wmb();
344 	Q_WRITE(dev, q, cpu_idx, q->head);
345 }
346 
347 static void
348 mt76_dma_tx_cleanup(struct mt76_dev *dev, struct mt76_queue *q, bool flush)
349 {
350 	struct mt76_queue_entry entry;
351 	int last;
352 
353 	if (!q || !q->ndesc)
354 		return;
355 
356 	spin_lock_bh(&q->cleanup_lock);
357 	if (flush)
358 		last = -1;
359 	else
360 		last = Q_READ(dev, q, dma_idx);
361 
362 	while (q->queued > 0 && q->tail != last) {
363 		mt76_dma_tx_cleanup_idx(dev, q, q->tail, &entry);
364 		mt76_queue_tx_complete(dev, q, &entry);
365 
366 		if (entry.txwi) {
367 			if (!(dev->drv->drv_flags & MT_DRV_TXWI_NO_FREE))
368 				mt76_put_txwi(dev, entry.txwi);
369 		}
370 
371 		if (!flush && q->tail == last)
372 			last = Q_READ(dev, q, dma_idx);
373 	}
374 	spin_unlock_bh(&q->cleanup_lock);
375 
376 	if (flush) {
377 		spin_lock_bh(&q->lock);
378 		mt76_dma_sync_idx(dev, q);
379 		mt76_dma_kick_queue(dev, q);
380 		spin_unlock_bh(&q->lock);
381 	}
382 
383 	if (!q->queued)
384 		wake_up(&dev->tx_wait);
385 }
386 
387 static void *
388 mt76_dma_get_buf(struct mt76_dev *dev, struct mt76_queue *q, int idx,
389 		 int *len, u32 *info, bool *more, bool *drop)
390 {
391 	struct mt76_queue_entry *e = &q->entry[idx];
392 	struct mt76_desc *desc = &q->desc[idx];
393 	void *buf;
394 
395 	if (len) {
396 		u32 ctrl = le32_to_cpu(READ_ONCE(desc->ctrl));
397 		*len = FIELD_GET(MT_DMA_CTL_SD_LEN0, ctrl);
398 		*more = !(ctrl & MT_DMA_CTL_LAST_SEC0);
399 	}
400 
401 	if (info)
402 		*info = le32_to_cpu(desc->info);
403 
404 	if ((q->flags & MT_QFLAG_WED) &&
405 	    FIELD_GET(MT_QFLAG_WED_TYPE, q->flags) == MT76_WED_Q_RX) {
406 		u32 token = FIELD_GET(MT_DMA_CTL_TOKEN,
407 				      le32_to_cpu(desc->buf1));
408 		struct mt76_txwi_cache *t = mt76_rx_token_release(dev, token);
409 
410 		if (!t)
411 			return NULL;
412 
413 		dma_unmap_single(dev->dma_dev, t->dma_addr,
414 				 SKB_WITH_OVERHEAD(q->buf_size),
415 				 DMA_FROM_DEVICE);
416 
417 		buf = t->ptr;
418 		t->dma_addr = 0;
419 		t->ptr = NULL;
420 
421 		mt76_put_rxwi(dev, t);
422 
423 		if (drop) {
424 			u32 ctrl = le32_to_cpu(READ_ONCE(desc->ctrl));
425 
426 			*drop = !!(ctrl & (MT_DMA_CTL_TO_HOST_A |
427 					   MT_DMA_CTL_DROP));
428 		}
429 	} else {
430 		buf = e->buf;
431 		e->buf = NULL;
432 		dma_unmap_single(dev->dma_dev, e->dma_addr[0],
433 				 SKB_WITH_OVERHEAD(q->buf_size),
434 				 DMA_FROM_DEVICE);
435 	}
436 
437 	return buf;
438 }
439 
440 static void *
441 mt76_dma_dequeue(struct mt76_dev *dev, struct mt76_queue *q, bool flush,
442 		 int *len, u32 *info, bool *more, bool *drop)
443 {
444 	int idx = q->tail;
445 
446 	*more = false;
447 	if (!q->queued)
448 		return NULL;
449 
450 	if (flush)
451 		q->desc[idx].ctrl |= cpu_to_le32(MT_DMA_CTL_DMA_DONE);
452 	else if (!(q->desc[idx].ctrl & cpu_to_le32(MT_DMA_CTL_DMA_DONE)))
453 		return NULL;
454 
455 	q->tail = (q->tail + 1) % q->ndesc;
456 	q->queued--;
457 
458 	return mt76_dma_get_buf(dev, q, idx, len, info, more, drop);
459 }
460 
461 static int
462 mt76_dma_tx_queue_skb_raw(struct mt76_dev *dev, struct mt76_queue *q,
463 			  struct sk_buff *skb, u32 tx_info)
464 {
465 	struct mt76_queue_buf buf = {};
466 	dma_addr_t addr;
467 
468 	if (q->queued + 1 >= q->ndesc - 1)
469 		goto error;
470 
471 	addr = dma_map_single(dev->dma_dev, skb->data, skb->len,
472 			      DMA_TO_DEVICE);
473 	if (unlikely(dma_mapping_error(dev->dma_dev, addr)))
474 		goto error;
475 
476 	buf.addr = addr;
477 	buf.len = skb->len;
478 
479 	spin_lock_bh(&q->lock);
480 	mt76_dma_add_buf(dev, q, &buf, 1, tx_info, skb, NULL);
481 	mt76_dma_kick_queue(dev, q);
482 	spin_unlock_bh(&q->lock);
483 
484 	return 0;
485 
486 error:
487 	dev_kfree_skb(skb);
488 	return -ENOMEM;
489 }
490 
491 static int
492 mt76_dma_tx_queue_skb(struct mt76_dev *dev, struct mt76_queue *q,
493 		      enum mt76_txq_id qid, struct sk_buff *skb,
494 		      struct mt76_wcid *wcid, struct ieee80211_sta *sta)
495 {
496 	struct ieee80211_tx_status status = {
497 		.sta = sta,
498 	};
499 	struct mt76_tx_info tx_info = {
500 		.skb = skb,
501 	};
502 	struct ieee80211_hw *hw;
503 	int len, n = 0, ret = -ENOMEM;
504 	struct mt76_txwi_cache *t;
505 	struct sk_buff *iter;
506 	dma_addr_t addr;
507 	u8 *txwi;
508 
509 	t = mt76_get_txwi(dev);
510 	if (!t)
511 		goto free_skb;
512 
513 	txwi = mt76_get_txwi_ptr(dev, t);
514 
515 	skb->prev = skb->next = NULL;
516 	if (dev->drv->drv_flags & MT_DRV_TX_ALIGNED4_SKBS)
517 		mt76_insert_hdr_pad(skb);
518 
519 	len = skb_headlen(skb);
520 	addr = dma_map_single(dev->dma_dev, skb->data, len, DMA_TO_DEVICE);
521 	if (unlikely(dma_mapping_error(dev->dma_dev, addr)))
522 		goto free;
523 
524 	tx_info.buf[n].addr = t->dma_addr;
525 	tx_info.buf[n++].len = dev->drv->txwi_size;
526 	tx_info.buf[n].addr = addr;
527 	tx_info.buf[n++].len = len;
528 
529 	skb_walk_frags(skb, iter) {
530 		if (n == ARRAY_SIZE(tx_info.buf))
531 			goto unmap;
532 
533 		addr = dma_map_single(dev->dma_dev, iter->data, iter->len,
534 				      DMA_TO_DEVICE);
535 		if (unlikely(dma_mapping_error(dev->dma_dev, addr)))
536 			goto unmap;
537 
538 		tx_info.buf[n].addr = addr;
539 		tx_info.buf[n++].len = iter->len;
540 	}
541 	tx_info.nbuf = n;
542 
543 	if (q->queued + (tx_info.nbuf + 1) / 2 >= q->ndesc - 1) {
544 		ret = -ENOMEM;
545 		goto unmap;
546 	}
547 
548 	dma_sync_single_for_cpu(dev->dma_dev, t->dma_addr, dev->drv->txwi_size,
549 				DMA_TO_DEVICE);
550 	ret = dev->drv->tx_prepare_skb(dev, txwi, qid, wcid, sta, &tx_info);
551 	dma_sync_single_for_device(dev->dma_dev, t->dma_addr, dev->drv->txwi_size,
552 				   DMA_TO_DEVICE);
553 	if (ret < 0)
554 		goto unmap;
555 
556 	return mt76_dma_add_buf(dev, q, tx_info.buf, tx_info.nbuf,
557 				tx_info.info, tx_info.skb, t);
558 
559 unmap:
560 	for (n--; n > 0; n--)
561 		dma_unmap_single(dev->dma_dev, tx_info.buf[n].addr,
562 				 tx_info.buf[n].len, DMA_TO_DEVICE);
563 
564 free:
565 #ifdef CONFIG_NL80211_TESTMODE
566 	/* fix tx_done accounting on queue overflow */
567 	if (mt76_is_testmode_skb(dev, skb, &hw)) {
568 		struct mt76_phy *phy = hw->priv;
569 
570 		if (tx_info.skb == phy->test.tx_skb)
571 			phy->test.tx_done--;
572 	}
573 #endif
574 
575 	mt76_put_txwi(dev, t);
576 
577 free_skb:
578 	status.skb = tx_info.skb;
579 	hw = mt76_tx_status_get_hw(dev, tx_info.skb);
580 	ieee80211_tx_status_ext(hw, &status);
581 
582 	return ret;
583 }
584 
585 static struct page_frag_cache *
586 mt76_dma_rx_get_frag_cache(struct mt76_dev *dev, struct mt76_queue *q)
587 {
588 	struct page_frag_cache *rx_page = &q->rx_page;
589 
590 #ifdef CONFIG_NET_MEDIATEK_SOC_WED
591 	if ((q->flags & MT_QFLAG_WED) &&
592 	    FIELD_GET(MT_QFLAG_WED_TYPE, q->flags) == MT76_WED_Q_RX)
593 		rx_page = &dev->mmio.wed.rx_buf_ring.rx_page;
594 #endif
595 	return rx_page;
596 }
597 
598 static int
599 mt76_dma_rx_fill(struct mt76_dev *dev, struct mt76_queue *q)
600 {
601 	struct page_frag_cache *rx_page = mt76_dma_rx_get_frag_cache(dev, q);
602 	int len = SKB_WITH_OVERHEAD(q->buf_size);
603 	int frames = 0, offset = q->buf_offset;
604 	dma_addr_t addr;
605 
606 	if (!q->ndesc)
607 		return 0;
608 
609 	spin_lock_bh(&q->lock);
610 
611 	while (q->queued < q->ndesc - 1) {
612 		struct mt76_queue_buf qbuf;
613 		void *buf = NULL;
614 
615 		buf = page_frag_alloc(rx_page, q->buf_size, GFP_ATOMIC);
616 		if (!buf)
617 			break;
618 
619 		addr = dma_map_single(dev->dma_dev, buf, len, DMA_FROM_DEVICE);
620 		if (unlikely(dma_mapping_error(dev->dma_dev, addr))) {
621 			skb_free_frag(buf);
622 			break;
623 		}
624 
625 		qbuf.addr = addr + offset;
626 		qbuf.len = len - offset;
627 		qbuf.skip_unmap = false;
628 		if (mt76_dma_add_rx_buf(dev, q, &qbuf, buf) < 0) {
629 			dma_unmap_single(dev->dma_dev, addr, len,
630 					 DMA_FROM_DEVICE);
631 			skb_free_frag(buf);
632 			break;
633 		}
634 		frames++;
635 	}
636 
637 	if (frames)
638 		mt76_dma_kick_queue(dev, q);
639 
640 	spin_unlock_bh(&q->lock);
641 
642 	return frames;
643 }
644 
645 static int
646 mt76_dma_wed_setup(struct mt76_dev *dev, struct mt76_queue *q)
647 {
648 #ifdef CONFIG_NET_MEDIATEK_SOC_WED
649 	struct mtk_wed_device *wed = &dev->mmio.wed;
650 	int ret, type, ring;
651 	u8 flags = q->flags;
652 
653 	if (!mtk_wed_device_active(wed))
654 		q->flags &= ~MT_QFLAG_WED;
655 
656 	if (!(q->flags & MT_QFLAG_WED))
657 		return 0;
658 
659 	type = FIELD_GET(MT_QFLAG_WED_TYPE, q->flags);
660 	ring = FIELD_GET(MT_QFLAG_WED_RING, q->flags);
661 
662 	switch (type) {
663 	case MT76_WED_Q_TX:
664 		ret = mtk_wed_device_tx_ring_setup(wed, ring, q->regs, false);
665 		if (!ret)
666 			q->wed_regs = wed->tx_ring[ring].reg_base;
667 		break;
668 	case MT76_WED_Q_TXFREE:
669 		/* WED txfree queue needs ring to be initialized before setup */
670 		q->flags = 0;
671 		mt76_dma_queue_reset(dev, q);
672 		mt76_dma_rx_fill(dev, q);
673 		q->flags = flags;
674 
675 		ret = mtk_wed_device_txfree_ring_setup(wed, q->regs);
676 		if (!ret)
677 			q->wed_regs = wed->txfree_ring.reg_base;
678 		break;
679 	case MT76_WED_Q_RX:
680 		ret = mtk_wed_device_rx_ring_setup(wed, ring, q->regs, false);
681 		if (!ret)
682 			q->wed_regs = wed->rx_ring[ring].reg_base;
683 		break;
684 	default:
685 		ret = -EINVAL;
686 	}
687 
688 	return ret;
689 #else
690 	return 0;
691 #endif
692 }
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_dma_wed_setup(dev, q);
720 	if (ret)
721 		return ret;
722 
723 	if (q->flags != MT_WED_Q_TXFREE)
724 		mt76_dma_queue_reset(dev, q);
725 
726 	return 0;
727 }
728 
729 static void
730 mt76_dma_rx_cleanup(struct mt76_dev *dev, struct mt76_queue *q)
731 {
732 	struct page *page;
733 	void *buf;
734 	bool more;
735 
736 	if (!q->ndesc)
737 		return;
738 
739 	spin_lock_bh(&q->lock);
740 	do {
741 		buf = mt76_dma_dequeue(dev, q, true, NULL, NULL, &more, NULL);
742 		if (!buf)
743 			break;
744 
745 		skb_free_frag(buf);
746 	} while (1);
747 	spin_unlock_bh(&q->lock);
748 
749 	if (!q->rx_page.va)
750 		return;
751 
752 	page = virt_to_page(q->rx_page.va);
753 	__page_frag_cache_drain(page, q->rx_page.pagecnt_bias);
754 	memset(&q->rx_page, 0, sizeof(q->rx_page));
755 }
756 
757 static void
758 mt76_dma_rx_reset(struct mt76_dev *dev, enum mt76_rxq_id qid)
759 {
760 	struct mt76_queue *q = &dev->q_rx[qid];
761 	int i;
762 
763 	if (!q->ndesc)
764 		return;
765 
766 	for (i = 0; i < q->ndesc; i++)
767 		q->desc[i].ctrl = cpu_to_le32(MT_DMA_CTL_DMA_DONE);
768 
769 	mt76_dma_rx_cleanup(dev, q);
770 	mt76_dma_sync_idx(dev, q);
771 	mt76_dma_rx_fill(dev, q);
772 
773 	if (!q->rx_head)
774 		return;
775 
776 	dev_kfree_skb(q->rx_head);
777 	q->rx_head = NULL;
778 }
779 
780 static void
781 mt76_add_fragment(struct mt76_dev *dev, struct mt76_queue *q, void *data,
782 		  int len, bool more, u32 info)
783 {
784 	struct sk_buff *skb = q->rx_head;
785 	struct skb_shared_info *shinfo = skb_shinfo(skb);
786 	int nr_frags = shinfo->nr_frags;
787 
788 	if (nr_frags < ARRAY_SIZE(shinfo->frags)) {
789 		struct page *page = virt_to_head_page(data);
790 		int offset = data - page_address(page) + q->buf_offset;
791 
792 		skb_add_rx_frag(skb, nr_frags, page, offset, len, q->buf_size);
793 	} else {
794 		skb_free_frag(data);
795 	}
796 
797 	if (more)
798 		return;
799 
800 	q->rx_head = NULL;
801 	if (nr_frags < ARRAY_SIZE(shinfo->frags))
802 		dev->drv->rx_skb(dev, q - dev->q_rx, skb, &info);
803 	else
804 		dev_kfree_skb(skb);
805 }
806 
807 static int
808 mt76_dma_rx_process(struct mt76_dev *dev, struct mt76_queue *q, int budget)
809 {
810 	int len, data_len, done = 0, dma_idx;
811 	struct sk_buff *skb;
812 	unsigned char *data;
813 	bool check_ddone = false;
814 	bool more;
815 
816 	if (IS_ENABLED(CONFIG_NET_MEDIATEK_SOC_WED) &&
817 	    q->flags == MT_WED_Q_TXFREE) {
818 		dma_idx = Q_READ(dev, q, dma_idx);
819 		check_ddone = true;
820 	}
821 
822 	while (done < budget) {
823 		bool drop = false;
824 		u32 info;
825 
826 		if (check_ddone) {
827 			if (q->tail == dma_idx)
828 				dma_idx = Q_READ(dev, q, dma_idx);
829 
830 			if (q->tail == dma_idx)
831 				break;
832 		}
833 
834 		data = mt76_dma_dequeue(dev, q, false, &len, &info, &more,
835 					&drop);
836 		if (!data)
837 			break;
838 
839 		if (drop)
840 			goto free_frag;
841 
842 		if (q->rx_head)
843 			data_len = q->buf_size;
844 		else
845 			data_len = SKB_WITH_OVERHEAD(q->buf_size);
846 
847 		if (data_len < len + q->buf_offset) {
848 			dev_kfree_skb(q->rx_head);
849 			q->rx_head = NULL;
850 			goto free_frag;
851 		}
852 
853 		if (q->rx_head) {
854 			mt76_add_fragment(dev, q, data, len, more, info);
855 			continue;
856 		}
857 
858 		if (!more && dev->drv->rx_check &&
859 		    !(dev->drv->rx_check(dev, data, len)))
860 			goto free_frag;
861 
862 		skb = build_skb(data, q->buf_size);
863 		if (!skb)
864 			goto free_frag;
865 
866 		skb_reserve(skb, q->buf_offset);
867 
868 		*(u32 *)skb->cb = info;
869 
870 		__skb_put(skb, len);
871 		done++;
872 
873 		if (more) {
874 			q->rx_head = skb;
875 			continue;
876 		}
877 
878 		dev->drv->rx_skb(dev, q - dev->q_rx, skb, &info);
879 		continue;
880 
881 free_frag:
882 		skb_free_frag(data);
883 	}
884 
885 	mt76_dma_rx_fill(dev, q);
886 	return done;
887 }
888 
889 int mt76_dma_rx_poll(struct napi_struct *napi, int budget)
890 {
891 	struct mt76_dev *dev;
892 	int qid, done = 0, cur;
893 
894 	dev = container_of(napi->dev, struct mt76_dev, napi_dev);
895 	qid = napi - dev->napi;
896 
897 	rcu_read_lock();
898 
899 	do {
900 		cur = mt76_dma_rx_process(dev, &dev->q_rx[qid], budget - done);
901 		mt76_rx_poll_complete(dev, qid, napi);
902 		done += cur;
903 	} while (cur && done < budget);
904 
905 	rcu_read_unlock();
906 
907 	if (done < budget && napi_complete(napi))
908 		dev->drv->rx_poll_complete(dev, qid);
909 
910 	return done;
911 }
912 EXPORT_SYMBOL_GPL(mt76_dma_rx_poll);
913 
914 static int
915 mt76_dma_init(struct mt76_dev *dev,
916 	      int (*poll)(struct napi_struct *napi, int budget))
917 {
918 	int i;
919 
920 	init_dummy_netdev(&dev->napi_dev);
921 	init_dummy_netdev(&dev->tx_napi_dev);
922 	snprintf(dev->napi_dev.name, sizeof(dev->napi_dev.name), "%s",
923 		 wiphy_name(dev->hw->wiphy));
924 	dev->napi_dev.threaded = 1;
925 
926 	mt76_for_each_q_rx(dev, i) {
927 		netif_napi_add(&dev->napi_dev, &dev->napi[i], poll);
928 		mt76_dma_rx_fill(dev, &dev->q_rx[i]);
929 		napi_enable(&dev->napi[i]);
930 	}
931 
932 	return 0;
933 }
934 
935 static const struct mt76_queue_ops mt76_dma_ops = {
936 	.init = mt76_dma_init,
937 	.alloc = mt76_dma_alloc_queue,
938 	.reset_q = mt76_dma_queue_reset,
939 	.tx_queue_skb_raw = mt76_dma_tx_queue_skb_raw,
940 	.tx_queue_skb = mt76_dma_tx_queue_skb,
941 	.tx_cleanup = mt76_dma_tx_cleanup,
942 	.rx_cleanup = mt76_dma_rx_cleanup,
943 	.rx_reset = mt76_dma_rx_reset,
944 	.kick = mt76_dma_kick_queue,
945 };
946 
947 void mt76_dma_attach(struct mt76_dev *dev)
948 {
949 	dev->queue_ops = &mt76_dma_ops;
950 }
951 EXPORT_SYMBOL_GPL(mt76_dma_attach);
952 
953 void mt76_dma_cleanup(struct mt76_dev *dev)
954 {
955 	int i;
956 
957 	mt76_worker_disable(&dev->tx_worker);
958 	netif_napi_del(&dev->tx_napi);
959 
960 	for (i = 0; i < ARRAY_SIZE(dev->phys); i++) {
961 		struct mt76_phy *phy = dev->phys[i];
962 		int j;
963 
964 		if (!phy)
965 			continue;
966 
967 		for (j = 0; j < ARRAY_SIZE(phy->q_tx); j++)
968 			mt76_dma_tx_cleanup(dev, phy->q_tx[j], true);
969 	}
970 
971 	for (i = 0; i < ARRAY_SIZE(dev->q_mcu); i++)
972 		mt76_dma_tx_cleanup(dev, dev->q_mcu[i], true);
973 
974 	mt76_for_each_q_rx(dev, i) {
975 		struct mt76_queue *q = &dev->q_rx[i];
976 
977 		netif_napi_del(&dev->napi[i]);
978 		if (FIELD_GET(MT_QFLAG_WED_TYPE, q->flags))
979 			mt76_dma_rx_cleanup(dev, q);
980 	}
981 
982 	mt76_free_pending_txwi(dev);
983 	mt76_free_pending_rxwi(dev);
984 
985 	if (mtk_wed_device_active(&dev->mmio.wed))
986 		mtk_wed_device_detach(&dev->mmio.wed);
987 }
988 EXPORT_SYMBOL_GPL(mt76_dma_cleanup);
989