xref: /freebsd/sys/contrib/dev/mediatek/mt76/dma.c (revision 397e83df75e0fcd0d3fcb95ae4d794cb7600fc89)
1 // SPDX-License-Identifier: ISC
2 /*
3  * Copyright (C) 2016 Felix Fietkau <nbd@nbd.name>
4  */
5 
6 #include <linux/dma-mapping.h>
7 #if defined(__FreeBSD__)
8 #include <linux/cache.h>
9 #include <net/page_pool.h>
10 #endif
11 #include "mt76.h"
12 #include "dma.h"
13 
14 #if IS_ENABLED(CONFIG_NET_MEDIATEK_SOC_WED)
15 
16 #define Q_READ(_dev, _q, _field) ({					\
17 	u32 _offset = offsetof(struct mt76_queue_regs, _field);		\
18 	u32 _val;							\
19 	if ((_q)->flags & MT_QFLAG_WED)					\
20 		_val = mtk_wed_device_reg_read(&(_dev)->mmio.wed,	\
21 					       ((_q)->wed_regs +	\
22 					        _offset));		\
23 	else								\
24 		_val = readl(&(_q)->regs->_field);			\
25 	_val;								\
26 })
27 
28 #define Q_WRITE(_dev, _q, _field, _val)	do {				\
29 	u32 _offset = offsetof(struct mt76_queue_regs, _field);		\
30 	if ((_q)->flags & MT_QFLAG_WED)					\
31 		mtk_wed_device_reg_write(&(_dev)->mmio.wed,		\
32 					 ((_q)->wed_regs + _offset),	\
33 					 _val);				\
34 	else								\
35 		writel(_val, &(_q)->regs->_field);			\
36 } while (0)
37 
38 #else
39 
40 #define Q_READ(_dev, _q, _field)	readl(&(_q)->regs->_field)
41 #define Q_WRITE(_dev, _q, _field, _val)	writel(_val, &(_q)->regs->_field)
42 
43 #endif
44 
45 static struct mt76_txwi_cache *
46 mt76_alloc_txwi(struct mt76_dev *dev)
47 {
48 	struct mt76_txwi_cache *t;
49 	dma_addr_t addr;
50 	u8 *txwi;
51 	int size;
52 
53 	size = L1_CACHE_ALIGN(dev->drv->txwi_size + sizeof(*t));
54 	txwi = kzalloc(size, GFP_ATOMIC);
55 	if (!txwi)
56 		return NULL;
57 
58 	addr = dma_map_single(dev->dma_dev, txwi, dev->drv->txwi_size,
59 			      DMA_TO_DEVICE);
60 	t = (struct mt76_txwi_cache *)(txwi + dev->drv->txwi_size);
61 	t->dma_addr = addr;
62 
63 	return t;
64 }
65 
66 static struct mt76_txwi_cache *
67 mt76_alloc_rxwi(struct mt76_dev *dev)
68 {
69 	struct mt76_txwi_cache *t;
70 
71 	t = kzalloc(L1_CACHE_ALIGN(sizeof(*t)), GFP_ATOMIC);
72 	if (!t)
73 		return NULL;
74 
75 	t->ptr = NULL;
76 	return t;
77 }
78 
79 static struct mt76_txwi_cache *
80 __mt76_get_txwi(struct mt76_dev *dev)
81 {
82 	struct mt76_txwi_cache *t = NULL;
83 
84 	spin_lock(&dev->lock);
85 	if (!list_empty(&dev->txwi_cache)) {
86 		t = list_first_entry(&dev->txwi_cache, struct mt76_txwi_cache,
87 				     list);
88 		list_del(&t->list);
89 	}
90 	spin_unlock(&dev->lock);
91 
92 	return t;
93 }
94 
95 static struct mt76_txwi_cache *
96 __mt76_get_rxwi(struct mt76_dev *dev)
97 {
98 	struct mt76_txwi_cache *t = NULL;
99 
100 	spin_lock(&dev->wed_lock);
101 	if (!list_empty(&dev->rxwi_cache)) {
102 		t = list_first_entry(&dev->rxwi_cache, struct mt76_txwi_cache,
103 				     list);
104 		list_del(&t->list);
105 	}
106 	spin_unlock(&dev->wed_lock);
107 
108 	return t;
109 }
110 
111 static struct mt76_txwi_cache *
112 mt76_get_txwi(struct mt76_dev *dev)
113 {
114 	struct mt76_txwi_cache *t = __mt76_get_txwi(dev);
115 
116 	if (t)
117 		return t;
118 
119 	return mt76_alloc_txwi(dev);
120 }
121 
122 struct mt76_txwi_cache *
123 mt76_get_rxwi(struct mt76_dev *dev)
124 {
125 	struct mt76_txwi_cache *t = __mt76_get_rxwi(dev);
126 
127 	if (t)
128 		return t;
129 
130 	return mt76_alloc_rxwi(dev);
131 }
132 EXPORT_SYMBOL_GPL(mt76_get_rxwi);
133 
134 void
135 mt76_put_txwi(struct mt76_dev *dev, struct mt76_txwi_cache *t)
136 {
137 	if (!t)
138 		return;
139 
140 	spin_lock(&dev->lock);
141 	list_add(&t->list, &dev->txwi_cache);
142 	spin_unlock(&dev->lock);
143 }
144 EXPORT_SYMBOL_GPL(mt76_put_txwi);
145 
146 void
147 mt76_put_rxwi(struct mt76_dev *dev, struct mt76_txwi_cache *t)
148 {
149 	if (!t)
150 		return;
151 
152 	spin_lock(&dev->wed_lock);
153 	list_add(&t->list, &dev->rxwi_cache);
154 	spin_unlock(&dev->wed_lock);
155 }
156 EXPORT_SYMBOL_GPL(mt76_put_rxwi);
157 
158 static void
159 mt76_free_pending_txwi(struct mt76_dev *dev)
160 {
161 	struct mt76_txwi_cache *t;
162 
163 	local_bh_disable();
164 	while ((t = __mt76_get_txwi(dev)) != NULL) {
165 		dma_unmap_single(dev->dma_dev, t->dma_addr, dev->drv->txwi_size,
166 				 DMA_TO_DEVICE);
167 		kfree(mt76_get_txwi_ptr(dev, t));
168 	}
169 	local_bh_enable();
170 }
171 
172 void
173 mt76_free_pending_rxwi(struct mt76_dev *dev)
174 {
175 	struct mt76_txwi_cache *t;
176 
177 	local_bh_disable();
178 	while ((t = __mt76_get_rxwi(dev)) != NULL) {
179 		if (t->ptr)
180 			mt76_put_page_pool_buf(t->ptr, false);
181 		kfree(t);
182 	}
183 	local_bh_enable();
184 }
185 EXPORT_SYMBOL_GPL(mt76_free_pending_rxwi);
186 
187 static void
188 mt76_dma_sync_idx(struct mt76_dev *dev, struct mt76_queue *q)
189 {
190 	Q_WRITE(dev, q, desc_base, q->desc_dma);
191 	Q_WRITE(dev, q, ring_size, q->ndesc);
192 	q->head = Q_READ(dev, q, dma_idx);
193 	q->tail = q->head;
194 }
195 
196 static void
197 mt76_dma_queue_reset(struct mt76_dev *dev, struct mt76_queue *q)
198 {
199 	int i;
200 
201 	if (!q || !q->ndesc)
202 		return;
203 
204 	/* clear descriptors */
205 	for (i = 0; i < q->ndesc; i++)
206 		q->desc[i].ctrl = cpu_to_le32(MT_DMA_CTL_DMA_DONE);
207 
208 	Q_WRITE(dev, q, cpu_idx, 0);
209 	Q_WRITE(dev, q, dma_idx, 0);
210 	mt76_dma_sync_idx(dev, q);
211 }
212 
213 static int
214 mt76_dma_add_rx_buf(struct mt76_dev *dev, struct mt76_queue *q,
215 		    struct mt76_queue_buf *buf, void *data)
216 {
217 	struct mt76_desc *desc = &q->desc[q->head];
218 	struct mt76_queue_entry *entry = &q->entry[q->head];
219 	struct mt76_txwi_cache *txwi = NULL;
220 	u32 buf1 = 0, ctrl;
221 	int idx = q->head;
222 	int rx_token;
223 
224 	ctrl = FIELD_PREP(MT_DMA_CTL_SD_LEN0, buf[0].len);
225 
226 	if (mt76_queue_is_wed_rx(q)) {
227 		txwi = mt76_get_rxwi(dev);
228 		if (!txwi)
229 			return -ENOMEM;
230 
231 		rx_token = mt76_rx_token_consume(dev, data, txwi, buf->addr);
232 		if (rx_token < 0) {
233 			mt76_put_rxwi(dev, txwi);
234 			return -ENOMEM;
235 		}
236 
237 		buf1 |= FIELD_PREP(MT_DMA_CTL_TOKEN, rx_token);
238 		ctrl |= MT_DMA_CTL_TO_HOST;
239 	}
240 
241 	WRITE_ONCE(desc->buf0, cpu_to_le32(buf->addr));
242 	WRITE_ONCE(desc->buf1, cpu_to_le32(buf1));
243 	WRITE_ONCE(desc->ctrl, cpu_to_le32(ctrl));
244 	WRITE_ONCE(desc->info, 0);
245 
246 	entry->dma_addr[0] = buf->addr;
247 	entry->dma_len[0] = buf->len;
248 	entry->txwi = txwi;
249 	entry->buf = data;
250 	entry->wcid = 0xffff;
251 	entry->skip_buf1 = true;
252 	q->head = (q->head + 1) % q->ndesc;
253 	q->queued++;
254 
255 	return idx;
256 }
257 
258 static int
259 mt76_dma_add_buf(struct mt76_dev *dev, struct mt76_queue *q,
260 		 struct mt76_queue_buf *buf, int nbufs, u32 info,
261 		 struct sk_buff *skb, void *txwi)
262 {
263 	struct mt76_queue_entry *entry;
264 	struct mt76_desc *desc;
265 	int i, idx = -1;
266 	u32 ctrl, next;
267 
268 	if (txwi) {
269 		q->entry[q->head].txwi = DMA_DUMMY_DATA;
270 		q->entry[q->head].skip_buf0 = true;
271 	}
272 
273 	for (i = 0; i < nbufs; i += 2, buf += 2) {
274 		u32 buf0 = buf[0].addr, buf1 = 0;
275 
276 		idx = q->head;
277 		next = (q->head + 1) % q->ndesc;
278 
279 		desc = &q->desc[idx];
280 		entry = &q->entry[idx];
281 
282 		if (buf[0].skip_unmap)
283 			entry->skip_buf0 = true;
284 		entry->skip_buf1 = i == nbufs - 1;
285 
286 		entry->dma_addr[0] = buf[0].addr;
287 		entry->dma_len[0] = buf[0].len;
288 
289 		ctrl = FIELD_PREP(MT_DMA_CTL_SD_LEN0, buf[0].len);
290 		if (i < nbufs - 1) {
291 			entry->dma_addr[1] = buf[1].addr;
292 			entry->dma_len[1] = buf[1].len;
293 			buf1 = buf[1].addr;
294 			ctrl |= FIELD_PREP(MT_DMA_CTL_SD_LEN1, buf[1].len);
295 			if (buf[1].skip_unmap)
296 				entry->skip_buf1 = true;
297 		}
298 
299 		if (i == nbufs - 1)
300 			ctrl |= MT_DMA_CTL_LAST_SEC0;
301 		else if (i == nbufs - 2)
302 			ctrl |= MT_DMA_CTL_LAST_SEC1;
303 
304 		WRITE_ONCE(desc->buf0, cpu_to_le32(buf0));
305 		WRITE_ONCE(desc->buf1, cpu_to_le32(buf1));
306 		WRITE_ONCE(desc->info, cpu_to_le32(info));
307 		WRITE_ONCE(desc->ctrl, cpu_to_le32(ctrl));
308 
309 		q->head = next;
310 		q->queued++;
311 	}
312 
313 	q->entry[idx].txwi = txwi;
314 	q->entry[idx].skb = skb;
315 	q->entry[idx].wcid = 0xffff;
316 
317 	return idx;
318 }
319 
320 static void
321 mt76_dma_tx_cleanup_idx(struct mt76_dev *dev, struct mt76_queue *q, int idx,
322 			struct mt76_queue_entry *prev_e)
323 {
324 	struct mt76_queue_entry *e = &q->entry[idx];
325 
326 	if (!e->skip_buf0)
327 		dma_unmap_single(dev->dma_dev, e->dma_addr[0], e->dma_len[0],
328 				 DMA_TO_DEVICE);
329 
330 	if (!e->skip_buf1)
331 		dma_unmap_single(dev->dma_dev, e->dma_addr[1], e->dma_len[1],
332 				 DMA_TO_DEVICE);
333 
334 	if (e->txwi == DMA_DUMMY_DATA)
335 		e->txwi = NULL;
336 
337 	if (e->skb == DMA_DUMMY_DATA)
338 		e->skb = NULL;
339 
340 	*prev_e = *e;
341 	memset(e, 0, sizeof(*e));
342 }
343 
344 static void
345 mt76_dma_kick_queue(struct mt76_dev *dev, struct mt76_queue *q)
346 {
347 	wmb();
348 	Q_WRITE(dev, q, cpu_idx, q->head);
349 }
350 
351 static void
352 mt76_dma_tx_cleanup(struct mt76_dev *dev, struct mt76_queue *q, bool flush)
353 {
354 	struct mt76_queue_entry entry;
355 	int last;
356 
357 	if (!q || !q->ndesc)
358 		return;
359 
360 	spin_lock_bh(&q->cleanup_lock);
361 	if (flush)
362 		last = -1;
363 	else
364 		last = Q_READ(dev, q, dma_idx);
365 
366 	while (q->queued > 0 && q->tail != last) {
367 		mt76_dma_tx_cleanup_idx(dev, q, q->tail, &entry);
368 		mt76_queue_tx_complete(dev, q, &entry);
369 
370 		if (entry.txwi) {
371 			if (!(dev->drv->drv_flags & MT_DRV_TXWI_NO_FREE))
372 				mt76_put_txwi(dev, entry.txwi);
373 		}
374 
375 		if (!flush && q->tail == last)
376 			last = Q_READ(dev, q, dma_idx);
377 	}
378 	spin_unlock_bh(&q->cleanup_lock);
379 
380 	if (flush) {
381 		spin_lock_bh(&q->lock);
382 		mt76_dma_sync_idx(dev, q);
383 		mt76_dma_kick_queue(dev, q);
384 		spin_unlock_bh(&q->lock);
385 	}
386 
387 	if (!q->queued)
388 		wake_up(&dev->tx_wait);
389 }
390 
391 static void *
392 mt76_dma_get_buf(struct mt76_dev *dev, struct mt76_queue *q, int idx,
393 		 int *len, u32 *info, bool *more, bool *drop)
394 {
395 	struct mt76_queue_entry *e = &q->entry[idx];
396 	struct mt76_desc *desc = &q->desc[idx];
397 	void *buf;
398 
399 	if (len) {
400 		u32 ctrl = le32_to_cpu(READ_ONCE(desc->ctrl));
401 		*len = FIELD_GET(MT_DMA_CTL_SD_LEN0, ctrl);
402 		*more = !(ctrl & MT_DMA_CTL_LAST_SEC0);
403 	}
404 
405 	if (info)
406 		*info = le32_to_cpu(desc->info);
407 
408 	if (mt76_queue_is_wed_rx(q)) {
409 		u32 buf1 = le32_to_cpu(desc->buf1);
410 		u32 token = FIELD_GET(MT_DMA_CTL_TOKEN, buf1);
411 		struct mt76_txwi_cache *t = mt76_rx_token_release(dev, token);
412 
413 		if (!t)
414 			return NULL;
415 
416 		dma_sync_single_for_cpu(dev->dma_dev, t->dma_addr,
417 				SKB_WITH_OVERHEAD(q->buf_size),
418 				page_pool_get_dma_dir(q->page_pool));
419 
420 		buf = t->ptr;
421 		t->dma_addr = 0;
422 		t->ptr = NULL;
423 
424 		mt76_put_rxwi(dev, t);
425 
426 		if (drop) {
427 			u32 ctrl = le32_to_cpu(READ_ONCE(desc->ctrl));
428 
429 			*drop = !!(ctrl & (MT_DMA_CTL_TO_HOST_A |
430 					   MT_DMA_CTL_DROP));
431 
432 			*drop |= !!(buf1 & MT_DMA_CTL_WO_DROP);
433 		}
434 	} else {
435 		buf = e->buf;
436 		e->buf = NULL;
437 		dma_sync_single_for_cpu(dev->dma_dev, e->dma_addr[0],
438 				SKB_WITH_OVERHEAD(q->buf_size),
439 				page_pool_get_dma_dir(q->page_pool));
440 	}
441 
442 	return buf;
443 }
444 
445 static void *
446 mt76_dma_dequeue(struct mt76_dev *dev, struct mt76_queue *q, bool flush,
447 		 int *len, u32 *info, bool *more, bool *drop)
448 {
449 	int idx = q->tail;
450 
451 	*more = false;
452 	if (!q->queued)
453 		return NULL;
454 
455 	if (flush)
456 		q->desc[idx].ctrl |= cpu_to_le32(MT_DMA_CTL_DMA_DONE);
457 	else if (!(q->desc[idx].ctrl & cpu_to_le32(MT_DMA_CTL_DMA_DONE)))
458 		return NULL;
459 
460 	q->tail = (q->tail + 1) % q->ndesc;
461 	q->queued--;
462 
463 	return mt76_dma_get_buf(dev, q, idx, len, info, more, drop);
464 }
465 
466 static int
467 mt76_dma_tx_queue_skb_raw(struct mt76_dev *dev, struct mt76_queue *q,
468 			  struct sk_buff *skb, u32 tx_info)
469 {
470 	struct mt76_queue_buf buf = {};
471 	dma_addr_t addr;
472 
473 	if (test_bit(MT76_MCU_RESET, &dev->phy.state))
474 		goto error;
475 
476 	if (q->queued + 1 >= q->ndesc - 1)
477 		goto error;
478 
479 	addr = dma_map_single(dev->dma_dev, skb->data, skb->len,
480 			      DMA_TO_DEVICE);
481 	if (unlikely(dma_mapping_error(dev->dma_dev, addr)))
482 		goto error;
483 
484 	buf.addr = addr;
485 	buf.len = skb->len;
486 
487 	spin_lock_bh(&q->lock);
488 	mt76_dma_add_buf(dev, q, &buf, 1, tx_info, skb, NULL);
489 	mt76_dma_kick_queue(dev, q);
490 	spin_unlock_bh(&q->lock);
491 
492 	return 0;
493 
494 error:
495 	dev_kfree_skb(skb);
496 	return -ENOMEM;
497 }
498 
499 static int
500 mt76_dma_tx_queue_skb(struct mt76_dev *dev, struct mt76_queue *q,
501 		      enum mt76_txq_id qid, struct sk_buff *skb,
502 		      struct mt76_wcid *wcid, struct ieee80211_sta *sta)
503 {
504 	struct ieee80211_tx_status status = {
505 		.sta = sta,
506 	};
507 	struct mt76_tx_info tx_info = {
508 		.skb = skb,
509 	};
510 	struct ieee80211_hw *hw;
511 	int len, n = 0, ret = -ENOMEM;
512 	struct mt76_txwi_cache *t;
513 	struct sk_buff *iter;
514 	dma_addr_t addr;
515 	u8 *txwi;
516 
517 	if (test_bit(MT76_RESET, &dev->phy.state))
518 		goto free_skb;
519 
520 	t = mt76_get_txwi(dev);
521 	if (!t)
522 		goto free_skb;
523 
524 	txwi = mt76_get_txwi_ptr(dev, t);
525 
526 	skb->prev = skb->next = NULL;
527 	if (dev->drv->drv_flags & MT_DRV_TX_ALIGNED4_SKBS)
528 		mt76_insert_hdr_pad(skb);
529 
530 	len = skb_headlen(skb);
531 	addr = dma_map_single(dev->dma_dev, skb->data, len, DMA_TO_DEVICE);
532 	if (unlikely(dma_mapping_error(dev->dma_dev, addr)))
533 		goto free;
534 
535 	tx_info.buf[n].addr = t->dma_addr;
536 	tx_info.buf[n++].len = dev->drv->txwi_size;
537 	tx_info.buf[n].addr = addr;
538 	tx_info.buf[n++].len = len;
539 
540 	skb_walk_frags(skb, iter) {
541 		if (n == ARRAY_SIZE(tx_info.buf))
542 			goto unmap;
543 
544 		addr = dma_map_single(dev->dma_dev, iter->data, iter->len,
545 				      DMA_TO_DEVICE);
546 		if (unlikely(dma_mapping_error(dev->dma_dev, addr)))
547 			goto unmap;
548 
549 		tx_info.buf[n].addr = addr;
550 		tx_info.buf[n++].len = iter->len;
551 	}
552 	tx_info.nbuf = n;
553 
554 	if (q->queued + (tx_info.nbuf + 1) / 2 >= q->ndesc - 1) {
555 		ret = -ENOMEM;
556 		goto unmap;
557 	}
558 
559 	dma_sync_single_for_cpu(dev->dma_dev, t->dma_addr, dev->drv->txwi_size,
560 				DMA_TO_DEVICE);
561 	ret = dev->drv->tx_prepare_skb(dev, txwi, qid, wcid, sta, &tx_info);
562 	dma_sync_single_for_device(dev->dma_dev, t->dma_addr, dev->drv->txwi_size,
563 				   DMA_TO_DEVICE);
564 	if (ret < 0)
565 		goto unmap;
566 
567 	return mt76_dma_add_buf(dev, q, tx_info.buf, tx_info.nbuf,
568 				tx_info.info, tx_info.skb, t);
569 
570 unmap:
571 	for (n--; n > 0; n--)
572 		dma_unmap_single(dev->dma_dev, tx_info.buf[n].addr,
573 				 tx_info.buf[n].len, DMA_TO_DEVICE);
574 
575 free:
576 #ifdef CONFIG_NL80211_TESTMODE
577 	/* fix tx_done accounting on queue overflow */
578 	if (mt76_is_testmode_skb(dev, skb, &hw)) {
579 		struct mt76_phy *phy = hw->priv;
580 
581 		if (tx_info.skb == phy->test.tx_skb)
582 			phy->test.tx_done--;
583 	}
584 #endif
585 
586 	mt76_put_txwi(dev, t);
587 
588 free_skb:
589 	status.skb = tx_info.skb;
590 	hw = mt76_tx_status_get_hw(dev, tx_info.skb);
591 	spin_lock_bh(&dev->rx_lock);
592 	ieee80211_tx_status_ext(hw, &status);
593 	spin_unlock_bh(&dev->rx_lock);
594 
595 	return ret;
596 }
597 
598 static int
599 mt76_dma_rx_fill(struct mt76_dev *dev, struct mt76_queue *q,
600 		 bool allow_direct)
601 {
602 	int len = SKB_WITH_OVERHEAD(q->buf_size);
603 	int frames = 0;
604 
605 	if (!q->ndesc)
606 		return 0;
607 
608 	spin_lock_bh(&q->lock);
609 
610 	while (q->queued < q->ndesc - 1) {
611 		enum dma_data_direction dir;
612 		struct mt76_queue_buf qbuf;
613 		dma_addr_t addr;
614 		int offset;
615 		void *buf;
616 
617 		buf = mt76_get_page_pool_buf(q, &offset, q->buf_size);
618 		if (!buf)
619 			break;
620 
621 		addr = page_pool_get_dma_addr(virt_to_head_page(buf)) + offset;
622 		dir = page_pool_get_dma_dir(q->page_pool);
623 		dma_sync_single_for_device(dev->dma_dev, addr, len, dir);
624 
625 		qbuf.addr = addr + q->buf_offset;
626 		qbuf.len = len - q->buf_offset;
627 		qbuf.skip_unmap = false;
628 		if (mt76_dma_add_rx_buf(dev, q, &qbuf, buf) < 0) {
629 			mt76_put_page_pool_buf(buf, allow_direct);
630 			break;
631 		}
632 		frames++;
633 	}
634 
635 	if (frames)
636 		mt76_dma_kick_queue(dev, q);
637 
638 	spin_unlock_bh(&q->lock);
639 
640 	return frames;
641 }
642 
643 int mt76_dma_wed_setup(struct mt76_dev *dev, struct mt76_queue *q, bool reset)
644 {
645 #ifdef CONFIG_NET_MEDIATEK_SOC_WED
646 	struct mtk_wed_device *wed = &dev->mmio.wed;
647 	int ret, type, ring;
648 	u8 flags;
649 
650 	if (!q || !q->ndesc)
651 		return -EINVAL;
652 
653 	flags = q->flags;
654 	if (!mtk_wed_device_active(wed))
655 		q->flags &= ~MT_QFLAG_WED;
656 
657 	if (!(q->flags & MT_QFLAG_WED))
658 		return 0;
659 
660 	type = FIELD_GET(MT_QFLAG_WED_TYPE, q->flags);
661 	ring = FIELD_GET(MT_QFLAG_WED_RING, q->flags);
662 
663 	switch (type) {
664 	case MT76_WED_Q_TX:
665 		ret = mtk_wed_device_tx_ring_setup(wed, ring, q->regs, reset);
666 		if (!ret)
667 			q->wed_regs = wed->tx_ring[ring].reg_base;
668 		break;
669 	case MT76_WED_Q_TXFREE:
670 		/* WED txfree queue needs ring to be initialized before setup */
671 		q->flags = 0;
672 		mt76_dma_queue_reset(dev, q);
673 		mt76_dma_rx_fill(dev, q, false);
674 		q->flags = flags;
675 
676 		ret = mtk_wed_device_txfree_ring_setup(wed, q->regs);
677 		if (!ret)
678 			q->wed_regs = wed->txfree_ring.reg_base;
679 		break;
680 	case MT76_WED_Q_RX:
681 		ret = mtk_wed_device_rx_ring_setup(wed, ring, q->regs, reset);
682 		if (!ret)
683 			q->wed_regs = wed->rx_ring[ring].reg_base;
684 		break;
685 	default:
686 		ret = -EINVAL;
687 	}
688 
689 	return ret;
690 #else
691 	return 0;
692 #endif
693 }
694 EXPORT_SYMBOL_GPL(mt76_dma_wed_setup);
695 
696 static int
697 mt76_dma_alloc_queue(struct mt76_dev *dev, struct mt76_queue *q,
698 		     int idx, int n_desc, int bufsize,
699 		     u32 ring_base)
700 {
701 	int ret, size;
702 
703 	spin_lock_init(&q->lock);
704 	spin_lock_init(&q->cleanup_lock);
705 
706 #if defined(__linux__)
707 	q->regs = dev->mmio.regs + ring_base + idx * MT_RING_SIZE;
708 #elif defined(__FreeBSD__)
709 	q->regs = (void *)((u8 *)dev->mmio.regs + ring_base + idx * MT_RING_SIZE);
710 #endif
711 	q->ndesc = n_desc;
712 	q->buf_size = bufsize;
713 	q->hw_idx = idx;
714 
715 	size = q->ndesc * sizeof(struct mt76_desc);
716 	q->desc = dmam_alloc_coherent(dev->dma_dev, size, &q->desc_dma, GFP_KERNEL);
717 	if (!q->desc)
718 		return -ENOMEM;
719 
720 	size = q->ndesc * sizeof(*q->entry);
721 	q->entry = devm_kzalloc(dev->dev, size, GFP_KERNEL);
722 	if (!q->entry)
723 		return -ENOMEM;
724 
725 	ret = mt76_create_page_pool(dev, q);
726 	if (ret)
727 		return ret;
728 
729 	ret = mt76_dma_wed_setup(dev, q, false);
730 	if (ret)
731 		return ret;
732 
733 	if (q->flags != MT_WED_Q_TXFREE)
734 		mt76_dma_queue_reset(dev, q);
735 
736 	return 0;
737 }
738 
739 static void
740 mt76_dma_rx_cleanup(struct mt76_dev *dev, struct mt76_queue *q)
741 {
742 	void *buf;
743 	bool more;
744 
745 	if (!q->ndesc)
746 		return;
747 
748 	spin_lock_bh(&q->lock);
749 
750 	do {
751 		buf = mt76_dma_dequeue(dev, q, true, NULL, NULL, &more, NULL);
752 		if (!buf)
753 			break;
754 
755 		mt76_put_page_pool_buf(buf, false);
756 	} while (1);
757 
758 	if (q->rx_head) {
759 		dev_kfree_skb(q->rx_head);
760 		q->rx_head = NULL;
761 	}
762 
763 	spin_unlock_bh(&q->lock);
764 }
765 
766 static void
767 mt76_dma_rx_reset(struct mt76_dev *dev, enum mt76_rxq_id qid)
768 {
769 	struct mt76_queue *q = &dev->q_rx[qid];
770 	int i;
771 
772 	if (!q->ndesc)
773 		return;
774 
775 	for (i = 0; i < q->ndesc; i++)
776 		q->desc[i].ctrl = cpu_to_le32(MT_DMA_CTL_DMA_DONE);
777 
778 	mt76_dma_rx_cleanup(dev, q);
779 
780 	/* reset WED rx queues */
781 	mt76_dma_wed_setup(dev, q, true);
782 	if (q->flags != MT_WED_Q_TXFREE) {
783 		mt76_dma_sync_idx(dev, q);
784 		mt76_dma_rx_fill(dev, q, false);
785 	}
786 }
787 
788 static void
789 mt76_add_fragment(struct mt76_dev *dev, struct mt76_queue *q, void *data,
790 		  int len, bool more, u32 info)
791 {
792 	struct sk_buff *skb = q->rx_head;
793 	struct skb_shared_info *shinfo = skb_shinfo(skb);
794 	int nr_frags = shinfo->nr_frags;
795 
796 	if (nr_frags < ARRAY_SIZE(shinfo->frags)) {
797 		struct page *page = virt_to_head_page(data);
798 #if defined(__linux__)
799 		int offset = data - page_address(page) + q->buf_offset;
800 #elif defined(__FreeBSD__)
801 		int offset = (u8 *)data - (u8 *)page_address(page) + q->buf_offset;
802 #endif
803 
804 		skb_add_rx_frag(skb, nr_frags, page, offset, len, q->buf_size);
805 	} else {
806 		mt76_put_page_pool_buf(data, true);
807 	}
808 
809 	if (more)
810 		return;
811 
812 	q->rx_head = NULL;
813 	if (nr_frags < ARRAY_SIZE(shinfo->frags))
814 		dev->drv->rx_skb(dev, q - dev->q_rx, skb, &info);
815 	else
816 		dev_kfree_skb(skb);
817 }
818 
819 static int
820 mt76_dma_rx_process(struct mt76_dev *dev, struct mt76_queue *q, int budget)
821 {
822 	int len, data_len, done = 0, dma_idx;
823 	struct sk_buff *skb;
824 	unsigned char *data;
825 	bool check_ddone = false;
826 	bool more;
827 
828 	if (IS_ENABLED(CONFIG_NET_MEDIATEK_SOC_WED) &&
829 	    q->flags == MT_WED_Q_TXFREE) {
830 		dma_idx = Q_READ(dev, q, dma_idx);
831 		check_ddone = true;
832 	}
833 
834 	while (done < budget) {
835 		bool drop = false;
836 		u32 info;
837 
838 		if (check_ddone) {
839 			if (q->tail == dma_idx)
840 				dma_idx = Q_READ(dev, q, dma_idx);
841 
842 			if (q->tail == dma_idx)
843 				break;
844 		}
845 
846 		data = mt76_dma_dequeue(dev, q, false, &len, &info, &more,
847 					&drop);
848 		if (!data)
849 			break;
850 
851 		if (drop)
852 			goto free_frag;
853 
854 		if (q->rx_head)
855 			data_len = q->buf_size;
856 		else
857 			data_len = SKB_WITH_OVERHEAD(q->buf_size);
858 
859 		if (data_len < len + q->buf_offset) {
860 			dev_kfree_skb(q->rx_head);
861 			q->rx_head = NULL;
862 			goto free_frag;
863 		}
864 
865 		if (q->rx_head) {
866 			mt76_add_fragment(dev, q, data, len, more, info);
867 			continue;
868 		}
869 
870 		if (!more && dev->drv->rx_check &&
871 		    !(dev->drv->rx_check(dev, data, len)))
872 			goto free_frag;
873 
874 		skb = napi_build_skb(data, q->buf_size);
875 		if (!skb)
876 			goto free_frag;
877 
878 		skb_reserve(skb, q->buf_offset);
879 		skb_mark_for_recycle(skb);
880 
881 		*(u32 *)skb->cb = info;
882 
883 		__skb_put(skb, len);
884 		done++;
885 
886 		if (more) {
887 			q->rx_head = skb;
888 			continue;
889 		}
890 
891 		dev->drv->rx_skb(dev, q - dev->q_rx, skb, &info);
892 		continue;
893 
894 free_frag:
895 		mt76_put_page_pool_buf(data, true);
896 	}
897 
898 	mt76_dma_rx_fill(dev, q, true);
899 	return done;
900 }
901 
902 int mt76_dma_rx_poll(struct napi_struct *napi, int budget)
903 {
904 	struct mt76_dev *dev;
905 	int qid, done = 0, cur;
906 
907 	dev = container_of(napi->dev, struct mt76_dev, napi_dev);
908 	qid = napi - dev->napi;
909 
910 	rcu_read_lock();
911 
912 	do {
913 		cur = mt76_dma_rx_process(dev, &dev->q_rx[qid], budget - done);
914 		mt76_rx_poll_complete(dev, qid, napi);
915 		done += cur;
916 	} while (cur && done < budget);
917 
918 	rcu_read_unlock();
919 
920 	if (done < budget && napi_complete(napi))
921 		dev->drv->rx_poll_complete(dev, qid);
922 
923 	return done;
924 }
925 EXPORT_SYMBOL_GPL(mt76_dma_rx_poll);
926 
927 static int
928 mt76_dma_init(struct mt76_dev *dev,
929 	      int (*poll)(struct napi_struct *napi, int budget))
930 {
931 	int i;
932 
933 	init_dummy_netdev(&dev->napi_dev);
934 	init_dummy_netdev(&dev->tx_napi_dev);
935 	snprintf(dev->napi_dev.name, sizeof(dev->napi_dev.name), "%s",
936 		 wiphy_name(dev->hw->wiphy));
937 	dev->napi_dev.threaded = 1;
938 	init_completion(&dev->mmio.wed_reset);
939 	init_completion(&dev->mmio.wed_reset_complete);
940 
941 	mt76_for_each_q_rx(dev, i) {
942 		netif_napi_add(&dev->napi_dev, &dev->napi[i], poll);
943 		mt76_dma_rx_fill(dev, &dev->q_rx[i], false);
944 		napi_enable(&dev->napi[i]);
945 	}
946 
947 	return 0;
948 }
949 
950 static const struct mt76_queue_ops mt76_dma_ops = {
951 	.init = mt76_dma_init,
952 	.alloc = mt76_dma_alloc_queue,
953 	.reset_q = mt76_dma_queue_reset,
954 	.tx_queue_skb_raw = mt76_dma_tx_queue_skb_raw,
955 	.tx_queue_skb = mt76_dma_tx_queue_skb,
956 	.tx_cleanup = mt76_dma_tx_cleanup,
957 	.rx_cleanup = mt76_dma_rx_cleanup,
958 	.rx_reset = mt76_dma_rx_reset,
959 	.kick = mt76_dma_kick_queue,
960 };
961 
962 void mt76_dma_attach(struct mt76_dev *dev)
963 {
964 	dev->queue_ops = &mt76_dma_ops;
965 }
966 EXPORT_SYMBOL_GPL(mt76_dma_attach);
967 
968 void mt76_dma_cleanup(struct mt76_dev *dev)
969 {
970 	int i;
971 
972 	mt76_worker_disable(&dev->tx_worker);
973 	netif_napi_del(&dev->tx_napi);
974 
975 	for (i = 0; i < ARRAY_SIZE(dev->phys); i++) {
976 		struct mt76_phy *phy = dev->phys[i];
977 		int j;
978 
979 		if (!phy)
980 			continue;
981 
982 		for (j = 0; j < ARRAY_SIZE(phy->q_tx); j++)
983 			mt76_dma_tx_cleanup(dev, phy->q_tx[j], true);
984 	}
985 
986 	for (i = 0; i < ARRAY_SIZE(dev->q_mcu); i++)
987 		mt76_dma_tx_cleanup(dev, dev->q_mcu[i], true);
988 
989 	mt76_for_each_q_rx(dev, i) {
990 		struct mt76_queue *q = &dev->q_rx[i];
991 
992 		netif_napi_del(&dev->napi[i]);
993 		mt76_dma_rx_cleanup(dev, q);
994 
995 		page_pool_destroy(q->page_pool);
996 	}
997 
998 	mt76_free_pending_txwi(dev);
999 	mt76_free_pending_rxwi(dev);
1000 
1001 	if (mtk_wed_device_active(&dev->mmio.wed))
1002 		mtk_wed_device_detach(&dev->mmio.wed);
1003 }
1004 EXPORT_SYMBOL_GPL(mt76_dma_cleanup);
1005