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_bh(&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_bh(&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_bh(&dev->wed_lock);
149 	list_add(&t->list, &dev->rxwi_cache);
150 	spin_unlock_bh(&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 	*prev_e = *e;
334 	memset(e, 0, sizeof(*e));
335 }
336 
337 static void
338 mt76_dma_kick_queue(struct mt76_dev *dev, struct mt76_queue *q)
339 {
340 	wmb();
341 	Q_WRITE(dev, q, cpu_idx, q->head);
342 }
343 
344 static void
345 mt76_dma_tx_cleanup(struct mt76_dev *dev, struct mt76_queue *q, bool flush)
346 {
347 	struct mt76_queue_entry entry;
348 	int last;
349 
350 	if (!q || !q->ndesc)
351 		return;
352 
353 	spin_lock_bh(&q->cleanup_lock);
354 	if (flush)
355 		last = -1;
356 	else
357 		last = Q_READ(dev, q, dma_idx);
358 
359 	while (q->queued > 0 && q->tail != last) {
360 		mt76_dma_tx_cleanup_idx(dev, q, q->tail, &entry);
361 		mt76_queue_tx_complete(dev, q, &entry);
362 
363 		if (entry.txwi) {
364 			if (!(dev->drv->drv_flags & MT_DRV_TXWI_NO_FREE))
365 				mt76_put_txwi(dev, entry.txwi);
366 		}
367 
368 		if (!flush && q->tail == last)
369 			last = Q_READ(dev, q, dma_idx);
370 	}
371 	spin_unlock_bh(&q->cleanup_lock);
372 
373 	if (flush) {
374 		spin_lock_bh(&q->lock);
375 		mt76_dma_sync_idx(dev, q);
376 		mt76_dma_kick_queue(dev, q);
377 		spin_unlock_bh(&q->lock);
378 	}
379 
380 	if (!q->queued)
381 		wake_up(&dev->tx_wait);
382 }
383 
384 static void *
385 mt76_dma_get_buf(struct mt76_dev *dev, struct mt76_queue *q, int idx,
386 		 int *len, u32 *info, bool *more, bool *drop)
387 {
388 	struct mt76_queue_entry *e = &q->entry[idx];
389 	struct mt76_desc *desc = &q->desc[idx];
390 	void *buf;
391 
392 	if (len) {
393 		u32 ctrl = le32_to_cpu(READ_ONCE(desc->ctrl));
394 		*len = FIELD_GET(MT_DMA_CTL_SD_LEN0, ctrl);
395 		*more = !(ctrl & MT_DMA_CTL_LAST_SEC0);
396 	}
397 
398 	if (info)
399 		*info = le32_to_cpu(desc->info);
400 
401 	if (mt76_queue_is_wed_rx(q)) {
402 		u32 buf1 = le32_to_cpu(desc->buf1);
403 		u32 token = FIELD_GET(MT_DMA_CTL_TOKEN, buf1);
404 		struct mt76_txwi_cache *t = mt76_rx_token_release(dev, token);
405 
406 		if (!t)
407 			return NULL;
408 
409 		dma_sync_single_for_cpu(dev->dma_dev, t->dma_addr,
410 				SKB_WITH_OVERHEAD(q->buf_size),
411 				page_pool_get_dma_dir(q->page_pool));
412 
413 		buf = t->ptr;
414 		t->dma_addr = 0;
415 		t->ptr = NULL;
416 
417 		mt76_put_rxwi(dev, t);
418 
419 		if (drop) {
420 			u32 ctrl = le32_to_cpu(READ_ONCE(desc->ctrl));
421 
422 			*drop = !!(ctrl & (MT_DMA_CTL_TO_HOST_A |
423 					   MT_DMA_CTL_DROP));
424 
425 			*drop |= !!(buf1 & MT_DMA_CTL_WO_DROP);
426 		}
427 	} else {
428 		buf = e->buf;
429 		e->buf = NULL;
430 		dma_sync_single_for_cpu(dev->dma_dev, e->dma_addr[0],
431 				SKB_WITH_OVERHEAD(q->buf_size),
432 				page_pool_get_dma_dir(q->page_pool));
433 	}
434 
435 	return buf;
436 }
437 
438 static void *
439 mt76_dma_dequeue(struct mt76_dev *dev, struct mt76_queue *q, bool flush,
440 		 int *len, u32 *info, bool *more, bool *drop)
441 {
442 	int idx = q->tail;
443 
444 	*more = false;
445 	if (!q->queued)
446 		return NULL;
447 
448 	if (flush)
449 		q->desc[idx].ctrl |= cpu_to_le32(MT_DMA_CTL_DMA_DONE);
450 	else if (!(q->desc[idx].ctrl & cpu_to_le32(MT_DMA_CTL_DMA_DONE)))
451 		return NULL;
452 
453 	q->tail = (q->tail + 1) % q->ndesc;
454 	q->queued--;
455 
456 	return mt76_dma_get_buf(dev, q, idx, len, info, more, drop);
457 }
458 
459 static int
460 mt76_dma_tx_queue_skb_raw(struct mt76_dev *dev, struct mt76_queue *q,
461 			  struct sk_buff *skb, u32 tx_info)
462 {
463 	struct mt76_queue_buf buf = {};
464 	dma_addr_t addr;
465 
466 	if (test_bit(MT76_MCU_RESET, &dev->phy.state))
467 		goto error;
468 
469 	if (q->queued + 1 >= q->ndesc - 1)
470 		goto error;
471 
472 	addr = dma_map_single(dev->dma_dev, skb->data, skb->len,
473 			      DMA_TO_DEVICE);
474 	if (unlikely(dma_mapping_error(dev->dma_dev, addr)))
475 		goto error;
476 
477 	buf.addr = addr;
478 	buf.len = skb->len;
479 
480 	spin_lock_bh(&q->lock);
481 	mt76_dma_add_buf(dev, q, &buf, 1, tx_info, skb, NULL);
482 	mt76_dma_kick_queue(dev, q);
483 	spin_unlock_bh(&q->lock);
484 
485 	return 0;
486 
487 error:
488 	dev_kfree_skb(skb);
489 	return -ENOMEM;
490 }
491 
492 static int
493 mt76_dma_tx_queue_skb(struct mt76_dev *dev, struct mt76_queue *q,
494 		      enum mt76_txq_id qid, struct sk_buff *skb,
495 		      struct mt76_wcid *wcid, struct ieee80211_sta *sta)
496 {
497 	struct ieee80211_tx_status status = {
498 		.sta = sta,
499 	};
500 	struct mt76_tx_info tx_info = {
501 		.skb = skb,
502 	};
503 	struct ieee80211_hw *hw;
504 	int len, n = 0, ret = -ENOMEM;
505 	struct mt76_txwi_cache *t;
506 	struct sk_buff *iter;
507 	dma_addr_t addr;
508 	u8 *txwi;
509 
510 	if (test_bit(MT76_RESET, &dev->phy.state))
511 		goto free_skb;
512 
513 	t = mt76_get_txwi(dev);
514 	if (!t)
515 		goto free_skb;
516 
517 	txwi = mt76_get_txwi_ptr(dev, t);
518 
519 	skb->prev = skb->next = NULL;
520 	if (dev->drv->drv_flags & MT_DRV_TX_ALIGNED4_SKBS)
521 		mt76_insert_hdr_pad(skb);
522 
523 	len = skb_headlen(skb);
524 	addr = dma_map_single(dev->dma_dev, skb->data, len, DMA_TO_DEVICE);
525 	if (unlikely(dma_mapping_error(dev->dma_dev, addr)))
526 		goto free;
527 
528 	tx_info.buf[n].addr = t->dma_addr;
529 	tx_info.buf[n++].len = dev->drv->txwi_size;
530 	tx_info.buf[n].addr = addr;
531 	tx_info.buf[n++].len = len;
532 
533 	skb_walk_frags(skb, iter) {
534 		if (n == ARRAY_SIZE(tx_info.buf))
535 			goto unmap;
536 
537 		addr = dma_map_single(dev->dma_dev, iter->data, iter->len,
538 				      DMA_TO_DEVICE);
539 		if (unlikely(dma_mapping_error(dev->dma_dev, addr)))
540 			goto unmap;
541 
542 		tx_info.buf[n].addr = addr;
543 		tx_info.buf[n++].len = iter->len;
544 	}
545 	tx_info.nbuf = n;
546 
547 	if (q->queued + (tx_info.nbuf + 1) / 2 >= q->ndesc - 1) {
548 		ret = -ENOMEM;
549 		goto unmap;
550 	}
551 
552 	dma_sync_single_for_cpu(dev->dma_dev, t->dma_addr, dev->drv->txwi_size,
553 				DMA_TO_DEVICE);
554 	ret = dev->drv->tx_prepare_skb(dev, txwi, qid, wcid, sta, &tx_info);
555 	dma_sync_single_for_device(dev->dma_dev, t->dma_addr, dev->drv->txwi_size,
556 				   DMA_TO_DEVICE);
557 	if (ret < 0)
558 		goto unmap;
559 
560 	return mt76_dma_add_buf(dev, q, tx_info.buf, tx_info.nbuf,
561 				tx_info.info, tx_info.skb, t);
562 
563 unmap:
564 	for (n--; n > 0; n--)
565 		dma_unmap_single(dev->dma_dev, tx_info.buf[n].addr,
566 				 tx_info.buf[n].len, DMA_TO_DEVICE);
567 
568 free:
569 #ifdef CONFIG_NL80211_TESTMODE
570 	/* fix tx_done accounting on queue overflow */
571 	if (mt76_is_testmode_skb(dev, skb, &hw)) {
572 		struct mt76_phy *phy = hw->priv;
573 
574 		if (tx_info.skb == phy->test.tx_skb)
575 			phy->test.tx_done--;
576 	}
577 #endif
578 
579 	mt76_put_txwi(dev, t);
580 
581 free_skb:
582 	status.skb = tx_info.skb;
583 	hw = mt76_tx_status_get_hw(dev, tx_info.skb);
584 	spin_lock_bh(&dev->rx_lock);
585 	ieee80211_tx_status_ext(hw, &status);
586 	spin_unlock_bh(&dev->rx_lock);
587 
588 	return ret;
589 }
590 
591 static int
592 mt76_dma_rx_fill(struct mt76_dev *dev, struct mt76_queue *q,
593 		 bool allow_direct)
594 {
595 	int len = SKB_WITH_OVERHEAD(q->buf_size);
596 	int frames = 0;
597 
598 	if (!q->ndesc)
599 		return 0;
600 
601 	spin_lock_bh(&q->lock);
602 
603 	while (q->queued < q->ndesc - 1) {
604 		enum dma_data_direction dir;
605 		struct mt76_queue_buf qbuf;
606 		dma_addr_t addr;
607 		int offset;
608 		void *buf;
609 
610 		buf = mt76_get_page_pool_buf(q, &offset, q->buf_size);
611 		if (!buf)
612 			break;
613 
614 		addr = page_pool_get_dma_addr(virt_to_head_page(buf)) + offset;
615 		dir = page_pool_get_dma_dir(q->page_pool);
616 		dma_sync_single_for_device(dev->dma_dev, addr, len, dir);
617 
618 		qbuf.addr = addr + q->buf_offset;
619 		qbuf.len = len - q->buf_offset;
620 		qbuf.skip_unmap = false;
621 		if (mt76_dma_add_rx_buf(dev, q, &qbuf, buf) < 0) {
622 			mt76_put_page_pool_buf(buf, allow_direct);
623 			break;
624 		}
625 		frames++;
626 	}
627 
628 	if (frames)
629 		mt76_dma_kick_queue(dev, q);
630 
631 	spin_unlock_bh(&q->lock);
632 
633 	return frames;
634 }
635 
636 int mt76_dma_wed_setup(struct mt76_dev *dev, struct mt76_queue *q, bool reset)
637 {
638 #ifdef CONFIG_NET_MEDIATEK_SOC_WED
639 	struct mtk_wed_device *wed = &dev->mmio.wed;
640 	int ret, type, ring;
641 	u8 flags;
642 
643 	if (!q || !q->ndesc)
644 		return -EINVAL;
645 
646 	flags = q->flags;
647 	if (!mtk_wed_device_active(wed))
648 		q->flags &= ~MT_QFLAG_WED;
649 
650 	if (!(q->flags & MT_QFLAG_WED))
651 		return 0;
652 
653 	type = FIELD_GET(MT_QFLAG_WED_TYPE, q->flags);
654 	ring = FIELD_GET(MT_QFLAG_WED_RING, q->flags);
655 
656 	switch (type) {
657 	case MT76_WED_Q_TX:
658 		ret = mtk_wed_device_tx_ring_setup(wed, ring, q->regs, reset);
659 		if (!ret)
660 			q->wed_regs = wed->tx_ring[ring].reg_base;
661 		break;
662 	case MT76_WED_Q_TXFREE:
663 		/* WED txfree queue needs ring to be initialized before setup */
664 		q->flags = 0;
665 		mt76_dma_queue_reset(dev, q);
666 		mt76_dma_rx_fill(dev, q, false);
667 		q->flags = flags;
668 
669 		ret = mtk_wed_device_txfree_ring_setup(wed, q->regs);
670 		if (!ret)
671 			q->wed_regs = wed->txfree_ring.reg_base;
672 		break;
673 	case MT76_WED_Q_RX:
674 		ret = mtk_wed_device_rx_ring_setup(wed, ring, q->regs, reset);
675 		if (!ret)
676 			q->wed_regs = wed->rx_ring[ring].reg_base;
677 		break;
678 	default:
679 		ret = -EINVAL;
680 	}
681 
682 	return ret;
683 #else
684 	return 0;
685 #endif
686 }
687 EXPORT_SYMBOL_GPL(mt76_dma_wed_setup);
688 
689 static int
690 mt76_dma_alloc_queue(struct mt76_dev *dev, struct mt76_queue *q,
691 		     int idx, int n_desc, int bufsize,
692 		     u32 ring_base)
693 {
694 	int ret, size;
695 
696 	spin_lock_init(&q->lock);
697 	spin_lock_init(&q->cleanup_lock);
698 
699 	q->regs = dev->mmio.regs + ring_base + idx * MT_RING_SIZE;
700 	q->ndesc = n_desc;
701 	q->buf_size = bufsize;
702 	q->hw_idx = idx;
703 
704 	size = q->ndesc * sizeof(struct mt76_desc);
705 	q->desc = dmam_alloc_coherent(dev->dma_dev, size, &q->desc_dma, GFP_KERNEL);
706 	if (!q->desc)
707 		return -ENOMEM;
708 
709 	size = q->ndesc * sizeof(*q->entry);
710 	q->entry = devm_kzalloc(dev->dev, size, GFP_KERNEL);
711 	if (!q->entry)
712 		return -ENOMEM;
713 
714 	ret = mt76_create_page_pool(dev, q);
715 	if (ret)
716 		return ret;
717 
718 	ret = mt76_dma_wed_setup(dev, q, false);
719 	if (ret)
720 		return ret;
721 
722 	if (q->flags != MT_WED_Q_TXFREE)
723 		mt76_dma_queue_reset(dev, q);
724 
725 	return 0;
726 }
727 
728 static void
729 mt76_dma_rx_cleanup(struct mt76_dev *dev, struct mt76_queue *q)
730 {
731 	void *buf;
732 	bool more;
733 
734 	if (!q->ndesc)
735 		return;
736 
737 	spin_lock_bh(&q->lock);
738 
739 	do {
740 		buf = mt76_dma_dequeue(dev, q, true, NULL, NULL, &more, NULL);
741 		if (!buf)
742 			break;
743 
744 		mt76_put_page_pool_buf(buf, false);
745 	} while (1);
746 
747 	if (q->rx_head) {
748 		dev_kfree_skb(q->rx_head);
749 		q->rx_head = NULL;
750 	}
751 
752 	spin_unlock_bh(&q->lock);
753 }
754 
755 static void
756 mt76_dma_rx_reset(struct mt76_dev *dev, enum mt76_rxq_id qid)
757 {
758 	struct mt76_queue *q = &dev->q_rx[qid];
759 	int i;
760 
761 	if (!q->ndesc)
762 		return;
763 
764 	for (i = 0; i < q->ndesc; i++)
765 		q->desc[i].ctrl = cpu_to_le32(MT_DMA_CTL_DMA_DONE);
766 
767 	mt76_dma_rx_cleanup(dev, q);
768 
769 	/* reset WED rx queues */
770 	mt76_dma_wed_setup(dev, q, true);
771 	if (q->flags != MT_WED_Q_TXFREE) {
772 		mt76_dma_sync_idx(dev, q);
773 		mt76_dma_rx_fill(dev, q, false);
774 	}
775 }
776 
777 static void
778 mt76_add_fragment(struct mt76_dev *dev, struct mt76_queue *q, void *data,
779 		  int len, bool more, u32 info, bool allow_direct)
780 {
781 	struct sk_buff *skb = q->rx_head;
782 	struct skb_shared_info *shinfo = skb_shinfo(skb);
783 	int nr_frags = shinfo->nr_frags;
784 
785 	if (nr_frags < ARRAY_SIZE(shinfo->frags)) {
786 		struct page *page = virt_to_head_page(data);
787 		int offset = data - page_address(page) + q->buf_offset;
788 
789 		skb_add_rx_frag(skb, nr_frags, page, offset, len, q->buf_size);
790 	} else {
791 		mt76_put_page_pool_buf(data, allow_direct);
792 	}
793 
794 	if (more)
795 		return;
796 
797 	q->rx_head = NULL;
798 	if (nr_frags < ARRAY_SIZE(shinfo->frags))
799 		dev->drv->rx_skb(dev, q - dev->q_rx, skb, &info);
800 	else
801 		dev_kfree_skb(skb);
802 }
803 
804 static int
805 mt76_dma_rx_process(struct mt76_dev *dev, struct mt76_queue *q, int budget)
806 {
807 	int len, data_len, done = 0, dma_idx;
808 	struct sk_buff *skb;
809 	unsigned char *data;
810 	bool check_ddone = false;
811 	bool allow_direct = !mt76_queue_is_wed_rx(q);
812 	bool more;
813 
814 	if (IS_ENABLED(CONFIG_NET_MEDIATEK_SOC_WED) &&
815 	    q->flags == MT_WED_Q_TXFREE) {
816 		dma_idx = Q_READ(dev, q, dma_idx);
817 		check_ddone = true;
818 	}
819 
820 	while (done < budget) {
821 		bool drop = false;
822 		u32 info;
823 
824 		if (check_ddone) {
825 			if (q->tail == dma_idx)
826 				dma_idx = Q_READ(dev, q, dma_idx);
827 
828 			if (q->tail == dma_idx)
829 				break;
830 		}
831 
832 		data = mt76_dma_dequeue(dev, q, false, &len, &info, &more,
833 					&drop);
834 		if (!data)
835 			break;
836 
837 		if (drop)
838 			goto free_frag;
839 
840 		if (q->rx_head)
841 			data_len = q->buf_size;
842 		else
843 			data_len = SKB_WITH_OVERHEAD(q->buf_size);
844 
845 		if (data_len < len + q->buf_offset) {
846 			dev_kfree_skb(q->rx_head);
847 			q->rx_head = NULL;
848 			goto free_frag;
849 		}
850 
851 		if (q->rx_head) {
852 			mt76_add_fragment(dev, q, data, len, more, info,
853 					  allow_direct);
854 			continue;
855 		}
856 
857 		if (!more && dev->drv->rx_check &&
858 		    !(dev->drv->rx_check(dev, data, len)))
859 			goto free_frag;
860 
861 		skb = napi_build_skb(data, q->buf_size);
862 		if (!skb)
863 			goto free_frag;
864 
865 		skb_reserve(skb, q->buf_offset);
866 		skb_mark_for_recycle(skb);
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 		mt76_put_page_pool_buf(data, allow_direct);
883 	}
884 
885 	mt76_dma_rx_fill(dev, q, true);
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 	init_completion(&dev->mmio.wed_reset);
926 	init_completion(&dev->mmio.wed_reset_complete);
927 
928 	mt76_for_each_q_rx(dev, i) {
929 		netif_napi_add(&dev->napi_dev, &dev->napi[i], poll);
930 		mt76_dma_rx_fill(dev, &dev->q_rx[i], false);
931 		napi_enable(&dev->napi[i]);
932 	}
933 
934 	return 0;
935 }
936 
937 static const struct mt76_queue_ops mt76_dma_ops = {
938 	.init = mt76_dma_init,
939 	.alloc = mt76_dma_alloc_queue,
940 	.reset_q = mt76_dma_queue_reset,
941 	.tx_queue_skb_raw = mt76_dma_tx_queue_skb_raw,
942 	.tx_queue_skb = mt76_dma_tx_queue_skb,
943 	.tx_cleanup = mt76_dma_tx_cleanup,
944 	.rx_cleanup = mt76_dma_rx_cleanup,
945 	.rx_reset = mt76_dma_rx_reset,
946 	.kick = mt76_dma_kick_queue,
947 };
948 
949 void mt76_dma_attach(struct mt76_dev *dev)
950 {
951 	dev->queue_ops = &mt76_dma_ops;
952 }
953 EXPORT_SYMBOL_GPL(mt76_dma_attach);
954 
955 void mt76_dma_cleanup(struct mt76_dev *dev)
956 {
957 	int i;
958 
959 	mt76_worker_disable(&dev->tx_worker);
960 	netif_napi_del(&dev->tx_napi);
961 
962 	for (i = 0; i < ARRAY_SIZE(dev->phys); i++) {
963 		struct mt76_phy *phy = dev->phys[i];
964 		int j;
965 
966 		if (!phy)
967 			continue;
968 
969 		for (j = 0; j < ARRAY_SIZE(phy->q_tx); j++)
970 			mt76_dma_tx_cleanup(dev, phy->q_tx[j], true);
971 	}
972 
973 	for (i = 0; i < ARRAY_SIZE(dev->q_mcu); i++)
974 		mt76_dma_tx_cleanup(dev, dev->q_mcu[i], true);
975 
976 	mt76_for_each_q_rx(dev, i) {
977 		struct mt76_queue *q = &dev->q_rx[i];
978 
979 		netif_napi_del(&dev->napi[i]);
980 		mt76_dma_rx_cleanup(dev, q);
981 
982 		page_pool_destroy(q->page_pool);
983 	}
984 
985 	mt76_free_pending_txwi(dev);
986 	mt76_free_pending_rxwi(dev);
987 
988 	if (mtk_wed_device_active(&dev->mmio.wed))
989 		mtk_wed_device_detach(&dev->mmio.wed);
990 }
991 EXPORT_SYMBOL_GPL(mt76_dma_cleanup);
992