xref: /linux/drivers/net/wireless/mediatek/mt76/dma.c (revision 8a922b7728a93d837954315c98b84f6b78de0c4f)
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 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 			mt76_put_page_pool_buf(t->ptr, false);
177 		kfree(t);
178 	}
179 	local_bh_enable();
180 }
181 EXPORT_SYMBOL_GPL(mt76_free_pending_rxwi);
182 
183 static void
184 mt76_dma_sync_idx(struct mt76_dev *dev, struct mt76_queue *q)
185 {
186 	Q_WRITE(dev, q, desc_base, q->desc_dma);
187 	Q_WRITE(dev, q, ring_size, q->ndesc);
188 	q->head = Q_READ(dev, q, dma_idx);
189 	q->tail = q->head;
190 }
191 
192 static void
193 mt76_dma_queue_reset(struct mt76_dev *dev, struct mt76_queue *q)
194 {
195 	int i;
196 
197 	if (!q || !q->ndesc)
198 		return;
199 
200 	/* clear descriptors */
201 	for (i = 0; i < q->ndesc; i++)
202 		q->desc[i].ctrl = cpu_to_le32(MT_DMA_CTL_DMA_DONE);
203 
204 	Q_WRITE(dev, q, cpu_idx, 0);
205 	Q_WRITE(dev, q, dma_idx, 0);
206 	mt76_dma_sync_idx(dev, q);
207 }
208 
209 static int
210 mt76_dma_add_rx_buf(struct mt76_dev *dev, struct mt76_queue *q,
211 		    struct mt76_queue_buf *buf, void *data)
212 {
213 	struct mt76_desc *desc = &q->desc[q->head];
214 	struct mt76_queue_entry *entry = &q->entry[q->head];
215 	struct mt76_txwi_cache *txwi = NULL;
216 	u32 buf1 = 0, ctrl;
217 	int idx = q->head;
218 	int rx_token;
219 
220 	ctrl = FIELD_PREP(MT_DMA_CTL_SD_LEN0, buf[0].len);
221 
222 	if (mt76_queue_is_wed_rx(q)) {
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 (mt76_queue_is_wed_rx(q)) {
405 		u32 token = FIELD_GET(MT_DMA_CTL_TOKEN,
406 				      le32_to_cpu(desc->buf1));
407 		struct mt76_txwi_cache *t = mt76_rx_token_release(dev, token);
408 
409 		if (!t)
410 			return NULL;
411 
412 		dma_sync_single_for_cpu(dev->dma_dev, t->dma_addr,
413 				SKB_WITH_OVERHEAD(q->buf_size),
414 				page_pool_get_dma_dir(q->page_pool));
415 
416 		buf = t->ptr;
417 		t->dma_addr = 0;
418 		t->ptr = NULL;
419 
420 		mt76_put_rxwi(dev, t);
421 
422 		if (drop) {
423 			u32 ctrl = le32_to_cpu(READ_ONCE(desc->ctrl));
424 
425 			*drop = !!(ctrl & (MT_DMA_CTL_TO_HOST_A |
426 					   MT_DMA_CTL_DROP));
427 		}
428 	} else {
429 		buf = e->buf;
430 		e->buf = NULL;
431 		dma_sync_single_for_cpu(dev->dma_dev, e->dma_addr[0],
432 				SKB_WITH_OVERHEAD(q->buf_size),
433 				page_pool_get_dma_dir(q->page_pool));
434 	}
435 
436 	return buf;
437 }
438 
439 static void *
440 mt76_dma_dequeue(struct mt76_dev *dev, struct mt76_queue *q, bool flush,
441 		 int *len, u32 *info, bool *more, bool *drop)
442 {
443 	int idx = q->tail;
444 
445 	*more = false;
446 	if (!q->queued)
447 		return NULL;
448 
449 	if (flush)
450 		q->desc[idx].ctrl |= cpu_to_le32(MT_DMA_CTL_DMA_DONE);
451 	else if (!(q->desc[idx].ctrl & cpu_to_le32(MT_DMA_CTL_DMA_DONE)))
452 		return NULL;
453 
454 	q->tail = (q->tail + 1) % q->ndesc;
455 	q->queued--;
456 
457 	return mt76_dma_get_buf(dev, q, idx, len, info, more, drop);
458 }
459 
460 static int
461 mt76_dma_tx_queue_skb_raw(struct mt76_dev *dev, struct mt76_queue *q,
462 			  struct sk_buff *skb, u32 tx_info)
463 {
464 	struct mt76_queue_buf buf = {};
465 	dma_addr_t addr;
466 
467 	if (q->queued + 1 >= q->ndesc - 1)
468 		goto error;
469 
470 	addr = dma_map_single(dev->dma_dev, skb->data, skb->len,
471 			      DMA_TO_DEVICE);
472 	if (unlikely(dma_mapping_error(dev->dma_dev, addr)))
473 		goto error;
474 
475 	buf.addr = addr;
476 	buf.len = skb->len;
477 
478 	spin_lock_bh(&q->lock);
479 	mt76_dma_add_buf(dev, q, &buf, 1, tx_info, skb, NULL);
480 	mt76_dma_kick_queue(dev, q);
481 	spin_unlock_bh(&q->lock);
482 
483 	return 0;
484 
485 error:
486 	dev_kfree_skb(skb);
487 	return -ENOMEM;
488 }
489 
490 static int
491 mt76_dma_tx_queue_skb(struct mt76_dev *dev, struct mt76_queue *q,
492 		      enum mt76_txq_id qid, struct sk_buff *skb,
493 		      struct mt76_wcid *wcid, struct ieee80211_sta *sta)
494 {
495 	struct ieee80211_tx_status status = {
496 		.sta = sta,
497 	};
498 	struct mt76_tx_info tx_info = {
499 		.skb = skb,
500 	};
501 	struct ieee80211_hw *hw;
502 	int len, n = 0, ret = -ENOMEM;
503 	struct mt76_txwi_cache *t;
504 	struct sk_buff *iter;
505 	dma_addr_t addr;
506 	u8 *txwi;
507 
508 	t = mt76_get_txwi(dev);
509 	if (!t)
510 		goto free_skb;
511 
512 	txwi = mt76_get_txwi_ptr(dev, t);
513 
514 	skb->prev = skb->next = NULL;
515 	if (dev->drv->drv_flags & MT_DRV_TX_ALIGNED4_SKBS)
516 		mt76_insert_hdr_pad(skb);
517 
518 	len = skb_headlen(skb);
519 	addr = dma_map_single(dev->dma_dev, skb->data, len, DMA_TO_DEVICE);
520 	if (unlikely(dma_mapping_error(dev->dma_dev, addr)))
521 		goto free;
522 
523 	tx_info.buf[n].addr = t->dma_addr;
524 	tx_info.buf[n++].len = dev->drv->txwi_size;
525 	tx_info.buf[n].addr = addr;
526 	tx_info.buf[n++].len = len;
527 
528 	skb_walk_frags(skb, iter) {
529 		if (n == ARRAY_SIZE(tx_info.buf))
530 			goto unmap;
531 
532 		addr = dma_map_single(dev->dma_dev, iter->data, iter->len,
533 				      DMA_TO_DEVICE);
534 		if (unlikely(dma_mapping_error(dev->dma_dev, addr)))
535 			goto unmap;
536 
537 		tx_info.buf[n].addr = addr;
538 		tx_info.buf[n++].len = iter->len;
539 	}
540 	tx_info.nbuf = n;
541 
542 	if (q->queued + (tx_info.nbuf + 1) / 2 >= q->ndesc - 1) {
543 		ret = -ENOMEM;
544 		goto unmap;
545 	}
546 
547 	dma_sync_single_for_cpu(dev->dma_dev, t->dma_addr, dev->drv->txwi_size,
548 				DMA_TO_DEVICE);
549 	ret = dev->drv->tx_prepare_skb(dev, txwi, qid, wcid, sta, &tx_info);
550 	dma_sync_single_for_device(dev->dma_dev, t->dma_addr, dev->drv->txwi_size,
551 				   DMA_TO_DEVICE);
552 	if (ret < 0)
553 		goto unmap;
554 
555 	return mt76_dma_add_buf(dev, q, tx_info.buf, tx_info.nbuf,
556 				tx_info.info, tx_info.skb, t);
557 
558 unmap:
559 	for (n--; n > 0; n--)
560 		dma_unmap_single(dev->dma_dev, tx_info.buf[n].addr,
561 				 tx_info.buf[n].len, DMA_TO_DEVICE);
562 
563 free:
564 #ifdef CONFIG_NL80211_TESTMODE
565 	/* fix tx_done accounting on queue overflow */
566 	if (mt76_is_testmode_skb(dev, skb, &hw)) {
567 		struct mt76_phy *phy = hw->priv;
568 
569 		if (tx_info.skb == phy->test.tx_skb)
570 			phy->test.tx_done--;
571 	}
572 #endif
573 
574 	mt76_put_txwi(dev, t);
575 
576 free_skb:
577 	status.skb = tx_info.skb;
578 	hw = mt76_tx_status_get_hw(dev, tx_info.skb);
579 	ieee80211_tx_status_ext(hw, &status);
580 
581 	return ret;
582 }
583 
584 static int
585 mt76_dma_rx_fill(struct mt76_dev *dev, struct mt76_queue *q,
586 		 bool allow_direct)
587 {
588 	int len = SKB_WITH_OVERHEAD(q->buf_size);
589 	int frames = 0;
590 
591 	if (!q->ndesc)
592 		return 0;
593 
594 	spin_lock_bh(&q->lock);
595 
596 	while (q->queued < q->ndesc - 1) {
597 		enum dma_data_direction dir;
598 		struct mt76_queue_buf qbuf;
599 		dma_addr_t addr;
600 		int offset;
601 		void *buf;
602 
603 		buf = mt76_get_page_pool_buf(q, &offset, q->buf_size);
604 		if (!buf)
605 			break;
606 
607 		addr = page_pool_get_dma_addr(virt_to_head_page(buf)) + offset;
608 		dir = page_pool_get_dma_dir(q->page_pool);
609 		dma_sync_single_for_device(dev->dma_dev, addr, len, dir);
610 
611 		qbuf.addr = addr + q->buf_offset;
612 		qbuf.len = len - q->buf_offset;
613 		qbuf.skip_unmap = false;
614 		if (mt76_dma_add_rx_buf(dev, q, &qbuf, buf) < 0) {
615 			mt76_put_page_pool_buf(buf, allow_direct);
616 			break;
617 		}
618 		frames++;
619 	}
620 
621 	if (frames)
622 		mt76_dma_kick_queue(dev, q);
623 
624 	spin_unlock_bh(&q->lock);
625 
626 	return frames;
627 }
628 
629 int mt76_dma_wed_setup(struct mt76_dev *dev, struct mt76_queue *q, bool reset)
630 {
631 #ifdef CONFIG_NET_MEDIATEK_SOC_WED
632 	struct mtk_wed_device *wed = &dev->mmio.wed;
633 	int ret, type, ring;
634 	u8 flags;
635 
636 	if (!q || !q->ndesc)
637 		return -EINVAL;
638 
639 	flags = q->flags;
640 	if (!mtk_wed_device_active(wed))
641 		q->flags &= ~MT_QFLAG_WED;
642 
643 	if (!(q->flags & MT_QFLAG_WED))
644 		return 0;
645 
646 	type = FIELD_GET(MT_QFLAG_WED_TYPE, q->flags);
647 	ring = FIELD_GET(MT_QFLAG_WED_RING, q->flags);
648 
649 	switch (type) {
650 	case MT76_WED_Q_TX:
651 		ret = mtk_wed_device_tx_ring_setup(wed, ring, q->regs, reset);
652 		if (!ret)
653 			q->wed_regs = wed->tx_ring[ring].reg_base;
654 		break;
655 	case MT76_WED_Q_TXFREE:
656 		/* WED txfree queue needs ring to be initialized before setup */
657 		q->flags = 0;
658 		mt76_dma_queue_reset(dev, q);
659 		mt76_dma_rx_fill(dev, q, false);
660 		q->flags = flags;
661 
662 		ret = mtk_wed_device_txfree_ring_setup(wed, q->regs);
663 		if (!ret)
664 			q->wed_regs = wed->txfree_ring.reg_base;
665 		break;
666 	case MT76_WED_Q_RX:
667 		ret = mtk_wed_device_rx_ring_setup(wed, ring, q->regs, reset);
668 		if (!ret)
669 			q->wed_regs = wed->rx_ring[ring].reg_base;
670 		break;
671 	default:
672 		ret = -EINVAL;
673 	}
674 
675 	return ret;
676 #else
677 	return 0;
678 #endif
679 }
680 EXPORT_SYMBOL_GPL(mt76_dma_wed_setup);
681 
682 static int
683 mt76_dma_alloc_queue(struct mt76_dev *dev, struct mt76_queue *q,
684 		     int idx, int n_desc, int bufsize,
685 		     u32 ring_base)
686 {
687 	int ret, size;
688 
689 	spin_lock_init(&q->lock);
690 	spin_lock_init(&q->cleanup_lock);
691 
692 	q->regs = dev->mmio.regs + ring_base + idx * MT_RING_SIZE;
693 	q->ndesc = n_desc;
694 	q->buf_size = bufsize;
695 	q->hw_idx = idx;
696 
697 	size = q->ndesc * sizeof(struct mt76_desc);
698 	q->desc = dmam_alloc_coherent(dev->dma_dev, size, &q->desc_dma, GFP_KERNEL);
699 	if (!q->desc)
700 		return -ENOMEM;
701 
702 	size = q->ndesc * sizeof(*q->entry);
703 	q->entry = devm_kzalloc(dev->dev, size, GFP_KERNEL);
704 	if (!q->entry)
705 		return -ENOMEM;
706 
707 	ret = mt76_create_page_pool(dev, q);
708 	if (ret)
709 		return ret;
710 
711 	ret = mt76_dma_wed_setup(dev, q, false);
712 	if (ret)
713 		return ret;
714 
715 	if (q->flags != MT_WED_Q_TXFREE)
716 		mt76_dma_queue_reset(dev, q);
717 
718 	return 0;
719 }
720 
721 static void
722 mt76_dma_rx_cleanup(struct mt76_dev *dev, struct mt76_queue *q)
723 {
724 	void *buf;
725 	bool more;
726 
727 	if (!q->ndesc)
728 		return;
729 
730 	spin_lock_bh(&q->lock);
731 
732 	do {
733 		buf = mt76_dma_dequeue(dev, q, true, NULL, NULL, &more, NULL);
734 		if (!buf)
735 			break;
736 
737 		mt76_put_page_pool_buf(buf, false);
738 	} while (1);
739 
740 	if (q->rx_head) {
741 		dev_kfree_skb(q->rx_head);
742 		q->rx_head = NULL;
743 	}
744 
745 	spin_unlock_bh(&q->lock);
746 }
747 
748 static void
749 mt76_dma_rx_reset(struct mt76_dev *dev, enum mt76_rxq_id qid)
750 {
751 	struct mt76_queue *q = &dev->q_rx[qid];
752 	int i;
753 
754 	if (!q->ndesc)
755 		return;
756 
757 	for (i = 0; i < q->ndesc; i++)
758 		q->desc[i].ctrl = cpu_to_le32(MT_DMA_CTL_DMA_DONE);
759 
760 	mt76_dma_rx_cleanup(dev, q);
761 
762 	/* reset WED rx queues */
763 	mt76_dma_wed_setup(dev, q, true);
764 	if (q->flags != MT_WED_Q_TXFREE) {
765 		mt76_dma_sync_idx(dev, q);
766 		mt76_dma_rx_fill(dev, q, false);
767 	}
768 }
769 
770 static void
771 mt76_add_fragment(struct mt76_dev *dev, struct mt76_queue *q, void *data,
772 		  int len, bool more, u32 info)
773 {
774 	struct sk_buff *skb = q->rx_head;
775 	struct skb_shared_info *shinfo = skb_shinfo(skb);
776 	int nr_frags = shinfo->nr_frags;
777 
778 	if (nr_frags < ARRAY_SIZE(shinfo->frags)) {
779 		struct page *page = virt_to_head_page(data);
780 		int offset = data - page_address(page) + q->buf_offset;
781 
782 		skb_add_rx_frag(skb, nr_frags, page, offset, len, q->buf_size);
783 	} else {
784 		mt76_put_page_pool_buf(data, true);
785 	}
786 
787 	if (more)
788 		return;
789 
790 	q->rx_head = NULL;
791 	if (nr_frags < ARRAY_SIZE(shinfo->frags))
792 		dev->drv->rx_skb(dev, q - dev->q_rx, skb, &info);
793 	else
794 		dev_kfree_skb(skb);
795 }
796 
797 static int
798 mt76_dma_rx_process(struct mt76_dev *dev, struct mt76_queue *q, int budget)
799 {
800 	int len, data_len, done = 0, dma_idx;
801 	struct sk_buff *skb;
802 	unsigned char *data;
803 	bool check_ddone = false;
804 	bool more;
805 
806 	if (IS_ENABLED(CONFIG_NET_MEDIATEK_SOC_WED) &&
807 	    q->flags == MT_WED_Q_TXFREE) {
808 		dma_idx = Q_READ(dev, q, dma_idx);
809 		check_ddone = true;
810 	}
811 
812 	while (done < budget) {
813 		bool drop = false;
814 		u32 info;
815 
816 		if (check_ddone) {
817 			if (q->tail == dma_idx)
818 				dma_idx = Q_READ(dev, q, dma_idx);
819 
820 			if (q->tail == dma_idx)
821 				break;
822 		}
823 
824 		data = mt76_dma_dequeue(dev, q, false, &len, &info, &more,
825 					&drop);
826 		if (!data)
827 			break;
828 
829 		if (drop)
830 			goto free_frag;
831 
832 		if (q->rx_head)
833 			data_len = q->buf_size;
834 		else
835 			data_len = SKB_WITH_OVERHEAD(q->buf_size);
836 
837 		if (data_len < len + q->buf_offset) {
838 			dev_kfree_skb(q->rx_head);
839 			q->rx_head = NULL;
840 			goto free_frag;
841 		}
842 
843 		if (q->rx_head) {
844 			mt76_add_fragment(dev, q, data, len, more, info);
845 			continue;
846 		}
847 
848 		if (!more && dev->drv->rx_check &&
849 		    !(dev->drv->rx_check(dev, data, len)))
850 			goto free_frag;
851 
852 		skb = build_skb(data, q->buf_size);
853 		if (!skb)
854 			goto free_frag;
855 
856 		skb_reserve(skb, q->buf_offset);
857 		skb_mark_for_recycle(skb);
858 
859 		*(u32 *)skb->cb = info;
860 
861 		__skb_put(skb, len);
862 		done++;
863 
864 		if (more) {
865 			q->rx_head = skb;
866 			continue;
867 		}
868 
869 		dev->drv->rx_skb(dev, q - dev->q_rx, skb, &info);
870 		continue;
871 
872 free_frag:
873 		mt76_put_page_pool_buf(data, true);
874 	}
875 
876 	mt76_dma_rx_fill(dev, q, true);
877 	return done;
878 }
879 
880 int mt76_dma_rx_poll(struct napi_struct *napi, int budget)
881 {
882 	struct mt76_dev *dev;
883 	int qid, done = 0, cur;
884 
885 	dev = container_of(napi->dev, struct mt76_dev, napi_dev);
886 	qid = napi - dev->napi;
887 
888 	rcu_read_lock();
889 
890 	do {
891 		cur = mt76_dma_rx_process(dev, &dev->q_rx[qid], budget - done);
892 		mt76_rx_poll_complete(dev, qid, napi);
893 		done += cur;
894 	} while (cur && done < budget);
895 
896 	rcu_read_unlock();
897 
898 	if (done < budget && napi_complete(napi))
899 		dev->drv->rx_poll_complete(dev, qid);
900 
901 	return done;
902 }
903 EXPORT_SYMBOL_GPL(mt76_dma_rx_poll);
904 
905 static int
906 mt76_dma_init(struct mt76_dev *dev,
907 	      int (*poll)(struct napi_struct *napi, int budget))
908 {
909 	int i;
910 
911 	init_dummy_netdev(&dev->napi_dev);
912 	init_dummy_netdev(&dev->tx_napi_dev);
913 	snprintf(dev->napi_dev.name, sizeof(dev->napi_dev.name), "%s",
914 		 wiphy_name(dev->hw->wiphy));
915 	dev->napi_dev.threaded = 1;
916 	init_completion(&dev->mmio.wed_reset);
917 	init_completion(&dev->mmio.wed_reset_complete);
918 
919 	mt76_for_each_q_rx(dev, i) {
920 		netif_napi_add(&dev->napi_dev, &dev->napi[i], poll);
921 		mt76_dma_rx_fill(dev, &dev->q_rx[i], false);
922 		napi_enable(&dev->napi[i]);
923 	}
924 
925 	return 0;
926 }
927 
928 static const struct mt76_queue_ops mt76_dma_ops = {
929 	.init = mt76_dma_init,
930 	.alloc = mt76_dma_alloc_queue,
931 	.reset_q = mt76_dma_queue_reset,
932 	.tx_queue_skb_raw = mt76_dma_tx_queue_skb_raw,
933 	.tx_queue_skb = mt76_dma_tx_queue_skb,
934 	.tx_cleanup = mt76_dma_tx_cleanup,
935 	.rx_cleanup = mt76_dma_rx_cleanup,
936 	.rx_reset = mt76_dma_rx_reset,
937 	.kick = mt76_dma_kick_queue,
938 };
939 
940 void mt76_dma_attach(struct mt76_dev *dev)
941 {
942 	dev->queue_ops = &mt76_dma_ops;
943 }
944 EXPORT_SYMBOL_GPL(mt76_dma_attach);
945 
946 void mt76_dma_cleanup(struct mt76_dev *dev)
947 {
948 	int i;
949 
950 	mt76_worker_disable(&dev->tx_worker);
951 	netif_napi_del(&dev->tx_napi);
952 
953 	for (i = 0; i < ARRAY_SIZE(dev->phys); i++) {
954 		struct mt76_phy *phy = dev->phys[i];
955 		int j;
956 
957 		if (!phy)
958 			continue;
959 
960 		for (j = 0; j < ARRAY_SIZE(phy->q_tx); j++)
961 			mt76_dma_tx_cleanup(dev, phy->q_tx[j], true);
962 	}
963 
964 	for (i = 0; i < ARRAY_SIZE(dev->q_mcu); i++)
965 		mt76_dma_tx_cleanup(dev, dev->q_mcu[i], true);
966 
967 	mt76_for_each_q_rx(dev, i) {
968 		struct mt76_queue *q = &dev->q_rx[i];
969 
970 		netif_napi_del(&dev->napi[i]);
971 		mt76_dma_rx_cleanup(dev, q);
972 
973 		page_pool_destroy(q->page_pool);
974 	}
975 
976 	mt76_free_pending_txwi(dev);
977 	mt76_free_pending_rxwi(dev);
978 
979 	if (mtk_wed_device_active(&dev->mmio.wed))
980 		mtk_wed_device_detach(&dev->mmio.wed);
981 }
982 EXPORT_SYMBOL_GPL(mt76_dma_cleanup);
983