xref: /linux/drivers/net/wireless/mediatek/mt76/dma.c (revision b61104e7a6349bd2c2b3e2fb3260d87f15eda8f4)
1 // SPDX-License-Identifier: BSD-3-Clause-Clear
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 static struct mt76_txwi_cache *
11 mt76_alloc_txwi(struct mt76_dev *dev)
12 {
13 	struct mt76_txwi_cache *t;
14 	dma_addr_t addr;
15 	u8 *txwi;
16 	int size;
17 
18 	size = L1_CACHE_ALIGN(dev->drv->txwi_size + sizeof(*t));
19 	txwi = kzalloc(size, GFP_ATOMIC);
20 	if (!txwi)
21 		return NULL;
22 
23 	addr = dma_map_single(dev->dma_dev, txwi, dev->drv->txwi_size,
24 			      DMA_TO_DEVICE);
25 	if (unlikely(dma_mapping_error(dev->dma_dev, addr))) {
26 		kfree(txwi);
27 		return NULL;
28 	}
29 
30 	t = (struct mt76_txwi_cache *)(txwi + dev->drv->txwi_size);
31 	t->dma_addr = addr;
32 
33 	return t;
34 }
35 
36 static struct mt76_txwi_cache *
37 mt76_alloc_rxwi(struct mt76_dev *dev)
38 {
39 	struct mt76_txwi_cache *t;
40 
41 	t = kzalloc(L1_CACHE_ALIGN(sizeof(*t)), GFP_ATOMIC);
42 	if (!t)
43 		return NULL;
44 
45 	t->ptr = NULL;
46 	return t;
47 }
48 
49 static struct mt76_txwi_cache *
50 __mt76_get_txwi(struct mt76_dev *dev)
51 {
52 	struct mt76_txwi_cache *t = NULL;
53 
54 	spin_lock(&dev->lock);
55 	if (!list_empty(&dev->txwi_cache)) {
56 		t = list_first_entry(&dev->txwi_cache, struct mt76_txwi_cache,
57 				     list);
58 		list_del(&t->list);
59 	}
60 	spin_unlock(&dev->lock);
61 
62 	return t;
63 }
64 
65 static struct mt76_txwi_cache *
66 __mt76_get_rxwi(struct mt76_dev *dev)
67 {
68 	struct mt76_txwi_cache *t = NULL;
69 
70 	spin_lock_bh(&dev->wed_lock);
71 	if (!list_empty(&dev->rxwi_cache)) {
72 		t = list_first_entry(&dev->rxwi_cache, struct mt76_txwi_cache,
73 				     list);
74 		list_del(&t->list);
75 	}
76 	spin_unlock_bh(&dev->wed_lock);
77 
78 	return t;
79 }
80 
81 static struct mt76_txwi_cache *
82 mt76_get_txwi(struct mt76_dev *dev)
83 {
84 	struct mt76_txwi_cache *t = __mt76_get_txwi(dev);
85 
86 	if (t)
87 		return t;
88 
89 	return mt76_alloc_txwi(dev);
90 }
91 
92 struct mt76_txwi_cache *
93 mt76_get_rxwi(struct mt76_dev *dev)
94 {
95 	struct mt76_txwi_cache *t = __mt76_get_rxwi(dev);
96 
97 	if (t)
98 		return t;
99 
100 	return mt76_alloc_rxwi(dev);
101 }
102 EXPORT_SYMBOL_GPL(mt76_get_rxwi);
103 
104 void
105 mt76_put_txwi(struct mt76_dev *dev, struct mt76_txwi_cache *t)
106 {
107 	if (!t)
108 		return;
109 
110 	spin_lock(&dev->lock);
111 	list_add(&t->list, &dev->txwi_cache);
112 	spin_unlock(&dev->lock);
113 }
114 EXPORT_SYMBOL_GPL(mt76_put_txwi);
115 
116 void
117 mt76_put_rxwi(struct mt76_dev *dev, struct mt76_txwi_cache *t)
118 {
119 	if (!t)
120 		return;
121 
122 	spin_lock_bh(&dev->wed_lock);
123 	list_add(&t->list, &dev->rxwi_cache);
124 	spin_unlock_bh(&dev->wed_lock);
125 }
126 EXPORT_SYMBOL_GPL(mt76_put_rxwi);
127 
128 static void
129 mt76_free_pending_txwi(struct mt76_dev *dev)
130 {
131 	struct mt76_txwi_cache *t;
132 
133 	local_bh_disable();
134 	while ((t = __mt76_get_txwi(dev)) != NULL) {
135 		dma_unmap_single(dev->dma_dev, t->dma_addr, dev->drv->txwi_size,
136 				 DMA_TO_DEVICE);
137 		kfree(mt76_get_txwi_ptr(dev, t));
138 	}
139 	local_bh_enable();
140 }
141 
142 void
143 mt76_free_pending_rxwi(struct mt76_dev *dev)
144 {
145 	struct mt76_txwi_cache *t;
146 
147 	local_bh_disable();
148 	while ((t = __mt76_get_rxwi(dev)) != NULL) {
149 		if (t->ptr)
150 			mt76_put_page_pool_buf(t->ptr, false);
151 		kfree(t);
152 	}
153 	local_bh_enable();
154 }
155 EXPORT_SYMBOL_GPL(mt76_free_pending_rxwi);
156 
157 static void
158 mt76_dma_queue_magic_cnt_init(struct mt76_dev *dev, struct mt76_queue *q)
159 {
160 	if (!mt76_queue_is_wed_rro(q))
161 		return;
162 
163 	q->magic_cnt = 0;
164 	if (mt76_queue_is_wed_rro_ind(q)) {
165 		struct mt76_wed_rro_desc *rro_desc;
166 		u32 data1 = FIELD_PREP(RRO_IND_DATA1_MAGIC_CNT_MASK,
167 				       MT_DMA_WED_IND_CMD_CNT - 1);
168 		int i;
169 
170 		rro_desc = (struct mt76_wed_rro_desc *)q->desc;
171 		for (i = 0; i < q->ndesc; i++) {
172 			struct mt76_wed_rro_ind *cmd;
173 
174 			cmd = (struct mt76_wed_rro_ind *)&rro_desc[i];
175 			cmd->data1 = cpu_to_le32(data1);
176 		}
177 	} else if (mt76_queue_is_wed_rro_rxdmad_c(q)) {
178 		struct mt76_rro_rxdmad_c *dmad = (void *)q->desc;
179 		u32 data3 = FIELD_PREP(RRO_RXDMAD_DATA3_MAGIC_CNT_MASK,
180 				       MT_DMA_MAGIC_CNT - 1);
181 		int i;
182 
183 		for (i = 0; i < q->ndesc; i++)
184 			dmad[i].data3 = cpu_to_le32(data3);
185 	}
186 }
187 
188 static void
189 mt76_dma_sync_idx(struct mt76_dev *dev, struct mt76_queue *q)
190 {
191 	Q_WRITE(q, desc_base, q->desc_dma);
192 	if ((q->flags & MT_QFLAG_WED_RRO_EN) && !mt76_npu_device_active(dev))
193 		Q_WRITE(q, ring_size, MT_DMA_RRO_EN | q->ndesc);
194 	else
195 		Q_WRITE(q, ring_size, q->ndesc);
196 
197 	if (mt76_queue_is_npu_tx(q)) {
198 		writel(q->desc_dma, &q->regs->desc_base);
199 		writel(q->ndesc, &q->regs->ring_size);
200 	}
201 	q->head = Q_READ(q, dma_idx);
202 	q->tail = q->head;
203 }
204 
205 void mt76_dma_queue_reset(struct mt76_dev *dev, struct mt76_queue *q,
206 			  bool reset_idx)
207 {
208 	if (!q || !q->ndesc)
209 		return;
210 
211 	if (!mt76_queue_is_wed_rro_ind(q) &&
212 	    !mt76_queue_is_wed_rro_rxdmad_c(q) && !mt76_queue_is_npu(q)) {
213 		int i;
214 
215 		/* clear descriptors */
216 		for (i = 0; i < q->ndesc; i++)
217 			q->desc[i].ctrl = cpu_to_le32(MT_DMA_CTL_DMA_DONE);
218 	}
219 
220 	mt76_dma_queue_magic_cnt_init(dev, q);
221 	if (reset_idx) {
222 		if (mt76_queue_is_emi(q))
223 			*q->emi_cpu_idx = 0;
224 		else
225 			Q_WRITE(q, cpu_idx, 0);
226 		Q_WRITE(q, dma_idx, 0);
227 	}
228 	mt76_dma_sync_idx(dev, q);
229 }
230 
231 static int
232 mt76_dma_add_rx_buf(struct mt76_dev *dev, struct mt76_queue *q,
233 		    struct mt76_queue_buf *buf, void *data)
234 {
235 	struct mt76_queue_entry *entry = &q->entry[q->head];
236 	struct mt76_txwi_cache *txwi = NULL;
237 	u32 buf1 = 0, ctrl, info = 0;
238 	struct mt76_desc *desc;
239 	int idx = q->head;
240 	int rx_token;
241 
242 	if (mt76_queue_is_wed_rro_ind(q)) {
243 		struct mt76_wed_rro_desc *rro_desc;
244 
245 		rro_desc = (struct mt76_wed_rro_desc *)q->desc;
246 		data = &rro_desc[q->head];
247 		goto done;
248 	} else if (mt76_queue_is_wed_rro_rxdmad_c(q)) {
249 		data = &q->desc[q->head];
250 		goto done;
251 	}
252 
253 	desc = &q->desc[q->head];
254 	ctrl = FIELD_PREP(MT_DMA_CTL_SD_LEN0, buf[0].len);
255 #ifdef CONFIG_ARCH_DMA_ADDR_T_64BIT
256 	buf1 = FIELD_PREP(MT_DMA_CTL_SDP0_H, buf->addr >> 32);
257 #endif
258 
259 	if (mt76_queue_is_wed_rx(q) || mt76_queue_is_wed_rro_data(q)) {
260 		txwi = mt76_get_rxwi(dev);
261 		if (!txwi)
262 			return -ENOMEM;
263 
264 		rx_token = mt76_rx_token_consume(dev, data, txwi, buf->addr);
265 		if (rx_token < 0) {
266 			mt76_put_rxwi(dev, txwi);
267 			return -ENOMEM;
268 		}
269 
270 		buf1 |= FIELD_PREP(MT_DMA_CTL_TOKEN, rx_token);
271 		ctrl |= MT_DMA_CTL_TO_HOST;
272 
273 		txwi->qid = q - dev->q_rx;
274 	}
275 
276 	if (mt76_queue_is_wed_rro_msdu_pg(q) &&
277 	    dev->drv->rx_rro_add_msdu_page) {
278 		if (dev->drv->rx_rro_add_msdu_page(dev, q, buf->addr, data))
279 			return -ENOMEM;
280 	}
281 
282 	if (q->flags & MT_QFLAG_WED_RRO_EN) {
283 		info |= FIELD_PREP(MT_DMA_MAGIC_MASK, q->magic_cnt);
284 		if ((q->head + 1) == q->ndesc)
285 			q->magic_cnt = (q->magic_cnt + 1) % MT_DMA_MAGIC_CNT;
286 	}
287 
288 	WRITE_ONCE(desc->buf0, cpu_to_le32(buf->addr));
289 	WRITE_ONCE(desc->buf1, cpu_to_le32(buf1));
290 	WRITE_ONCE(desc->ctrl, cpu_to_le32(ctrl));
291 	WRITE_ONCE(desc->info, cpu_to_le32(info));
292 
293 done:
294 	entry->dma_addr[0] = buf->addr;
295 	entry->dma_len[0] = buf->len;
296 	entry->txwi = txwi;
297 	entry->buf = data;
298 	entry->wcid = 0xffff;
299 	entry->skip_buf1 = true;
300 	q->head = (q->head + 1) % q->ndesc;
301 	q->queued++;
302 
303 	return idx;
304 }
305 
306 static int
307 mt76_dma_add_buf(struct mt76_dev *dev, struct mt76_queue *q,
308 		 struct mt76_queue_buf *buf, int nbufs, u32 info,
309 		 struct sk_buff *skb, void *txwi)
310 {
311 	struct mt76_queue_entry *entry;
312 	struct mt76_desc *desc;
313 	int i, idx = -1;
314 	u32 ctrl, next;
315 
316 	if (txwi) {
317 		q->entry[q->head].txwi = DMA_DUMMY_DATA;
318 		q->entry[q->head].skip_buf0 = true;
319 	}
320 
321 	for (i = 0; i < nbufs; i += 2, buf += 2) {
322 		u32 buf0 = buf[0].addr, buf1 = 0;
323 
324 		idx = q->head;
325 		next = (q->head + 1) % q->ndesc;
326 
327 		desc = &q->desc[idx];
328 		entry = &q->entry[idx];
329 
330 		if (buf[0].skip_unmap)
331 			entry->skip_buf0 = true;
332 		entry->skip_buf1 = i == nbufs - 1;
333 
334 		entry->dma_addr[0] = buf[0].addr;
335 		entry->dma_len[0] = buf[0].len;
336 
337 		ctrl = FIELD_PREP(MT_DMA_CTL_SD_LEN0, buf[0].len);
338 #ifdef CONFIG_ARCH_DMA_ADDR_T_64BIT
339 		info |= FIELD_PREP(MT_DMA_CTL_SDP0_H, buf[0].addr >> 32);
340 #endif
341 		if (i < nbufs - 1) {
342 			entry->dma_addr[1] = buf[1].addr;
343 			entry->dma_len[1] = buf[1].len;
344 			buf1 = buf[1].addr;
345 			ctrl |= FIELD_PREP(MT_DMA_CTL_SD_LEN1, buf[1].len);
346 #ifdef CONFIG_ARCH_DMA_ADDR_T_64BIT
347 			info |= FIELD_PREP(MT_DMA_CTL_SDP1_H,
348 					   buf[1].addr >> 32);
349 #endif
350 			if (buf[1].skip_unmap)
351 				entry->skip_buf1 = true;
352 		}
353 
354 		if (i == nbufs - 1)
355 			ctrl |= MT_DMA_CTL_LAST_SEC0;
356 		else if (i == nbufs - 2)
357 			ctrl |= MT_DMA_CTL_LAST_SEC1;
358 
359 		WRITE_ONCE(desc->buf0, cpu_to_le32(buf0));
360 		WRITE_ONCE(desc->buf1, cpu_to_le32(buf1));
361 		WRITE_ONCE(desc->info, cpu_to_le32(info));
362 		WRITE_ONCE(desc->ctrl, cpu_to_le32(ctrl));
363 
364 		q->head = next;
365 		q->queued++;
366 	}
367 
368 	q->entry[idx].txwi = txwi;
369 	q->entry[idx].skb = skb;
370 	q->entry[idx].wcid = 0xffff;
371 
372 	return idx;
373 }
374 
375 static void
376 mt76_dma_tx_cleanup_idx(struct mt76_dev *dev, struct mt76_queue *q, int idx,
377 			struct mt76_queue_entry *prev_e)
378 {
379 	struct mt76_queue_entry *e = &q->entry[idx];
380 
381 	if (!e->skip_buf0)
382 		dma_unmap_single(dev->dma_dev, e->dma_addr[0], e->dma_len[0],
383 				 DMA_TO_DEVICE);
384 
385 	if (!e->skip_buf1)
386 		dma_unmap_single(dev->dma_dev, e->dma_addr[1], e->dma_len[1],
387 				 DMA_TO_DEVICE);
388 
389 	if (e->txwi == DMA_DUMMY_DATA)
390 		e->txwi = NULL;
391 
392 	*prev_e = *e;
393 	memset(e, 0, sizeof(*e));
394 }
395 
396 static void
397 mt76_dma_kick_queue(struct mt76_dev *dev, struct mt76_queue *q)
398 {
399 	wmb();
400 	if (mt76_queue_is_emi(q))
401 		*q->emi_cpu_idx = cpu_to_le16(q->head);
402 	else
403 		Q_WRITE(q, cpu_idx, q->head);
404 }
405 
406 static void
407 mt76_dma_tx_cleanup(struct mt76_dev *dev, struct mt76_queue *q, bool flush)
408 {
409 	struct mt76_queue_entry entry;
410 	int last;
411 
412 	if (!q || !q->ndesc)
413 		return;
414 
415 	spin_lock_bh(&q->cleanup_lock);
416 	if (flush)
417 		last = -1;
418 	else
419 		last = Q_READ(q, dma_idx);
420 
421 	while (q->queued > 0 && q->tail != last) {
422 		mt76_dma_tx_cleanup_idx(dev, q, q->tail, &entry);
423 		mt76_npu_txdesc_cleanup(q, q->tail);
424 		mt76_queue_tx_complete(dev, q, &entry);
425 
426 		if (entry.txwi) {
427 			if (!(dev->drv->drv_flags & MT_DRV_TXWI_NO_FREE))
428 				mt76_put_txwi(dev, entry.txwi);
429 		}
430 
431 		if (!flush && q->tail == last)
432 			last = Q_READ(q, dma_idx);
433 	}
434 	spin_unlock_bh(&q->cleanup_lock);
435 
436 	if (flush) {
437 		spin_lock_bh(&q->lock);
438 		mt76_dma_sync_idx(dev, q);
439 		mt76_dma_kick_queue(dev, q);
440 		spin_unlock_bh(&q->lock);
441 	}
442 
443 	if (!q->queued)
444 		wake_up(&dev->tx_wait);
445 }
446 
447 static void *
448 mt76_dma_get_rxdmad_c_buf(struct mt76_dev *dev, struct mt76_queue *q,
449 			  int idx, int *len, bool *more)
450 {
451 	struct mt76_queue_entry *e = &q->entry[idx];
452 	struct mt76_rro_rxdmad_c *dmad = e->buf;
453 	u32 data1 = le32_to_cpu(dmad->data1);
454 	u32 data2 = le32_to_cpu(dmad->data2);
455 	struct mt76_txwi_cache *t;
456 	u16 rx_token_id;
457 	u8 ind_reason;
458 	void *buf;
459 
460 	rx_token_id = FIELD_GET(RRO_RXDMAD_DATA2_RX_TOKEN_ID_MASK, data2);
461 	t = mt76_rx_token_release(dev, rx_token_id);
462 	if (!t)
463 		return ERR_PTR(-EAGAIN);
464 
465 	q = &dev->q_rx[t->qid];
466 	dma_sync_single_for_cpu(dev->dma_dev, t->dma_addr,
467 				SKB_WITH_OVERHEAD(q->buf_size),
468 				page_pool_get_dma_dir(q->page_pool));
469 
470 	if (len)
471 		*len = FIELD_GET(RRO_RXDMAD_DATA1_SDL0_MASK, data1);
472 	if (more)
473 		*more = !FIELD_GET(RRO_RXDMAD_DATA1_LS_MASK, data1);
474 
475 	buf = t->ptr;
476 	ind_reason = FIELD_GET(RRO_RXDMAD_DATA2_IND_REASON_MASK, data2);
477 	if (ind_reason == MT_DMA_WED_IND_REASON_REPEAT ||
478 	    ind_reason == MT_DMA_WED_IND_REASON_OLDPKT) {
479 		mt76_put_page_pool_buf(buf, false);
480 		buf = ERR_PTR(-EAGAIN);
481 	}
482 	t->ptr = NULL;
483 	t->dma_addr = 0;
484 
485 	mt76_put_rxwi(dev, t);
486 
487 	return buf;
488 }
489 
490 static void *
491 mt76_dma_get_buf(struct mt76_dev *dev, struct mt76_queue *q, int idx,
492 		 int *len, u32 *info, bool *more, bool *drop, bool flush)
493 {
494 	struct mt76_queue_entry *e = &q->entry[idx];
495 	struct mt76_desc *desc = &q->desc[idx];
496 	u32 ctrl, desc_info, buf1;
497 	void *buf = e->buf;
498 
499 	if (mt76_queue_is_wed_rro_rxdmad_c(q) && !flush)
500 		buf = mt76_dma_get_rxdmad_c_buf(dev, q, idx, len, more);
501 
502 	if (mt76_queue_is_wed_rro(q))
503 		goto done;
504 
505 	ctrl = le32_to_cpu(READ_ONCE(desc->ctrl));
506 	if (len) {
507 		*len = FIELD_GET(MT_DMA_CTL_SD_LEN0, ctrl);
508 		*more = !(ctrl & MT_DMA_CTL_LAST_SEC0);
509 	}
510 
511 	desc_info = le32_to_cpu(desc->info);
512 	if (info)
513 		*info = desc_info;
514 
515 	buf1 = le32_to_cpu(desc->buf1);
516 	mt76_dma_should_drop_buf(drop, ctrl, buf1, desc_info);
517 
518 	if (mt76_queue_is_wed_rx(q)) {
519 		u32 token = FIELD_GET(MT_DMA_CTL_TOKEN, buf1);
520 		struct mt76_txwi_cache *t = mt76_rx_token_release(dev, token);
521 
522 		if (!t)
523 			return NULL;
524 
525 		dma_sync_single_for_cpu(dev->dma_dev, t->dma_addr,
526 				SKB_WITH_OVERHEAD(q->buf_size),
527 				page_pool_get_dma_dir(q->page_pool));
528 
529 		buf = t->ptr;
530 		t->dma_addr = 0;
531 		t->ptr = NULL;
532 
533 		mt76_put_rxwi(dev, t);
534 		if (drop)
535 			*drop |= !!(buf1 & MT_DMA_CTL_WO_DROP);
536 	} else {
537 		dma_sync_single_for_cpu(dev->dma_dev, e->dma_addr[0],
538 				SKB_WITH_OVERHEAD(q->buf_size),
539 				page_pool_get_dma_dir(q->page_pool));
540 	}
541 
542 done:
543 	e->buf = NULL;
544 	return buf;
545 }
546 
547 static void *
548 mt76_dma_dequeue(struct mt76_dev *dev, struct mt76_queue *q, bool flush,
549 		 int *len, u32 *info, bool *more, bool *drop)
550 {
551 	int idx = q->tail;
552 
553 	*more = false;
554 	if (!q->queued)
555 		return NULL;
556 
557 	if (mt76_queue_is_wed_rro_data(q) || mt76_queue_is_wed_rro_msdu_pg(q))
558 		goto done;
559 
560 	if (mt76_queue_is_wed_rro_ind(q)) {
561 		struct mt76_wed_rro_ind *cmd;
562 		u8 magic_cnt;
563 
564 		if (flush)
565 			goto done;
566 
567 		cmd = q->entry[idx].buf;
568 		magic_cnt = FIELD_GET(RRO_IND_DATA1_MAGIC_CNT_MASK,
569 				      le32_to_cpu(cmd->data1));
570 		if (magic_cnt != q->magic_cnt)
571 			return NULL;
572 
573 		if (q->tail == q->ndesc - 1)
574 			q->magic_cnt = (q->magic_cnt + 1) % MT_DMA_WED_IND_CMD_CNT;
575 	} else if (mt76_queue_is_wed_rro_rxdmad_c(q)) {
576 		struct mt76_rro_rxdmad_c *dmad;
577 		u16 magic_cnt;
578 
579 		if (flush)
580 			goto done;
581 
582 		dmad = q->entry[idx].buf;
583 		magic_cnt = FIELD_GET(RRO_RXDMAD_DATA3_MAGIC_CNT_MASK,
584 				      le32_to_cpu(dmad->data3));
585 		if (magic_cnt != q->magic_cnt)
586 			return NULL;
587 
588 		if (q->tail == q->ndesc - 1)
589 			q->magic_cnt = (q->magic_cnt + 1) % MT_DMA_MAGIC_CNT;
590 	} else {
591 		if (flush)
592 			q->desc[idx].ctrl |= cpu_to_le32(MT_DMA_CTL_DMA_DONE);
593 		else if (!(q->desc[idx].ctrl & cpu_to_le32(MT_DMA_CTL_DMA_DONE)))
594 			return NULL;
595 	}
596 done:
597 	q->tail = (q->tail + 1) % q->ndesc;
598 	q->queued--;
599 
600 	return mt76_dma_get_buf(dev, q, idx, len, info, more, drop, flush);
601 }
602 
603 static int
604 mt76_dma_tx_queue_skb_raw(struct mt76_dev *dev, struct mt76_queue *q,
605 			  struct sk_buff *skb, u32 tx_info)
606 {
607 	struct mt76_queue_buf buf = {};
608 	dma_addr_t addr;
609 
610 	if (test_bit(MT76_MCU_RESET, &dev->phy.state))
611 		goto error;
612 
613 	if (q->queued + 1 >= q->ndesc - 1)
614 		goto error;
615 
616 	addr = dma_map_single(dev->dma_dev, skb->data, skb->len,
617 			      DMA_TO_DEVICE);
618 	if (unlikely(dma_mapping_error(dev->dma_dev, addr)))
619 		goto error;
620 
621 	buf.addr = addr;
622 	buf.len = skb->len;
623 
624 	spin_lock_bh(&q->lock);
625 	mt76_dma_add_buf(dev, q, &buf, 1, tx_info, skb, NULL);
626 	mt76_dma_kick_queue(dev, q);
627 	spin_unlock_bh(&q->lock);
628 
629 	return 0;
630 
631 error:
632 	dev_kfree_skb(skb);
633 	return -ENOMEM;
634 }
635 
636 static int
637 mt76_dma_tx_queue_skb(struct mt76_phy *phy, struct mt76_queue *q,
638 		      enum mt76_txq_id qid, struct sk_buff *skb,
639 		      struct mt76_wcid *wcid, struct ieee80211_sta *sta)
640 {
641 	struct ieee80211_tx_status status = {
642 		.sta = sta,
643 	};
644 	struct mt76_tx_info tx_info = {
645 		.skb = skb,
646 	};
647 	struct mt76_dev *dev = phy->dev;
648 	struct ieee80211_hw *hw;
649 	int len, n = 0, ret = -ENOMEM;
650 	struct mt76_txwi_cache *t;
651 	struct sk_buff *iter;
652 	dma_addr_t addr;
653 	u8 *txwi;
654 
655 	if (test_bit(MT76_RESET, &phy->state))
656 		goto free_skb;
657 
658 	/* TODO: Take into account unlinear skbs */
659 	if (mt76_npu_device_active(dev) && skb_linearize(skb))
660 		goto free_skb;
661 
662 	t = mt76_get_txwi(dev);
663 	if (!t)
664 		goto free_skb;
665 
666 	txwi = mt76_get_txwi_ptr(dev, t);
667 
668 	skb->prev = skb->next = NULL;
669 	if (dev->drv->drv_flags & MT_DRV_TX_ALIGNED4_SKBS)
670 		mt76_insert_hdr_pad(skb);
671 
672 	len = skb_headlen(skb);
673 	addr = dma_map_single(dev->dma_dev, skb->data, len, DMA_TO_DEVICE);
674 	if (unlikely(dma_mapping_error(dev->dma_dev, addr)))
675 		goto free;
676 
677 	tx_info.buf[n].addr = t->dma_addr;
678 	tx_info.buf[n++].len = dev->drv->txwi_size;
679 	tx_info.buf[n].addr = addr;
680 	tx_info.buf[n++].len = len;
681 
682 	skb_walk_frags(skb, iter) {
683 		if (n == ARRAY_SIZE(tx_info.buf))
684 			goto unmap;
685 
686 		addr = dma_map_single(dev->dma_dev, iter->data, iter->len,
687 				      DMA_TO_DEVICE);
688 		if (unlikely(dma_mapping_error(dev->dma_dev, addr)))
689 			goto unmap;
690 
691 		tx_info.buf[n].addr = addr;
692 		tx_info.buf[n++].len = iter->len;
693 	}
694 	tx_info.nbuf = n;
695 
696 	if (q->queued + (tx_info.nbuf + 1) / 2 >= q->ndesc - 1) {
697 		ret = -ENOMEM;
698 		goto unmap;
699 	}
700 
701 	dma_sync_single_for_cpu(dev->dma_dev, t->dma_addr, dev->drv->txwi_size,
702 				DMA_TO_DEVICE);
703 	ret = dev->drv->tx_prepare_skb(dev, txwi, qid, wcid, sta, &tx_info);
704 	dma_sync_single_for_device(dev->dma_dev, t->dma_addr, dev->drv->txwi_size,
705 				   DMA_TO_DEVICE);
706 	if (ret < 0)
707 		goto unmap;
708 
709 	if (mt76_npu_device_active(dev))
710 		return mt76_npu_dma_add_buf(phy, q, skb, &tx_info.buf[1], txwi);
711 
712 	return mt76_dma_add_buf(dev, q, tx_info.buf, tx_info.nbuf,
713 				tx_info.info, tx_info.skb, t);
714 
715 unmap:
716 	for (n--; n > 0; n--)
717 		dma_unmap_single(dev->dma_dev, tx_info.buf[n].addr,
718 				 tx_info.buf[n].len, DMA_TO_DEVICE);
719 
720 free:
721 #ifdef CONFIG_NL80211_TESTMODE
722 	/* fix tx_done accounting on queue overflow */
723 	if (mt76_is_testmode_skb(dev, skb, &hw)) {
724 		struct mt76_phy *phy = hw->priv;
725 
726 		if (tx_info.skb == phy->test.tx_skb)
727 			phy->test.tx_done--;
728 	}
729 #endif
730 
731 	mt76_put_txwi(dev, t);
732 
733 free_skb:
734 	status.skb = tx_info.skb;
735 	hw = mt76_tx_status_get_hw(dev, tx_info.skb);
736 	spin_lock_bh(&dev->rx_lock);
737 	ieee80211_tx_status_ext(hw, &status);
738 	spin_unlock_bh(&dev->rx_lock);
739 
740 	return ret;
741 }
742 
743 static int
744 mt76_dma_rx_fill_buf(struct mt76_dev *dev, struct mt76_queue *q,
745 		     bool allow_direct)
746 {
747 	int len = SKB_WITH_OVERHEAD(q->buf_size);
748 	int frames = 0;
749 
750 	if (!q->ndesc)
751 		return 0;
752 
753 	while (q->queued < q->ndesc - 1) {
754 		struct mt76_queue_buf qbuf = {};
755 		void *buf = NULL;
756 		int offset;
757 
758 		if (mt76_queue_is_wed_rro_ind(q) ||
759 		    mt76_queue_is_wed_rro_rxdmad_c(q))
760 			goto done;
761 
762 		buf = mt76_get_page_pool_buf(q, &offset, q->buf_size);
763 		if (!buf)
764 			break;
765 
766 		qbuf.addr = page_pool_get_dma_addr(virt_to_head_page(buf)) +
767 			    offset + q->buf_offset;
768 done:
769 		qbuf.len = len - q->buf_offset;
770 		qbuf.skip_unmap = false;
771 		if (mt76_dma_add_rx_buf(dev, q, &qbuf, buf) < 0) {
772 			mt76_put_page_pool_buf(buf, allow_direct);
773 			break;
774 		}
775 		frames++;
776 	}
777 
778 	if (frames || mt76_queue_is_wed_rx(q))
779 		mt76_dma_kick_queue(dev, q);
780 
781 	return frames;
782 }
783 
784 int mt76_dma_rx_fill(struct mt76_dev *dev, struct mt76_queue *q,
785 		     bool allow_direct)
786 {
787 	int frames;
788 
789 	spin_lock_bh(&q->lock);
790 	frames = mt76_dma_rx_fill_buf(dev, q, allow_direct);
791 	spin_unlock_bh(&q->lock);
792 
793 	return frames;
794 }
795 
796 static int
797 mt76_dma_alloc_queue(struct mt76_dev *dev, struct mt76_queue *q,
798 		     int idx, int n_desc, int bufsize,
799 		     u32 ring_base)
800 {
801 	int ret, size;
802 
803 	spin_lock_init(&q->lock);
804 	spin_lock_init(&q->cleanup_lock);
805 
806 	q->regs = dev->mmio.regs + ring_base + idx * MT_RING_SIZE;
807 	q->ndesc = n_desc;
808 	q->buf_size = bufsize;
809 	q->hw_idx = idx;
810 	q->dev = dev;
811 
812 	if (mt76_queue_is_wed_rro_ind(q))
813 		size = sizeof(struct mt76_wed_rro_desc);
814 	else if (mt76_queue_is_npu_tx(q))
815 		size = sizeof(struct airoha_npu_tx_dma_desc);
816 	else if (mt76_queue_is_npu_rx(q))
817 		size = sizeof(struct airoha_npu_rx_dma_desc);
818 	else
819 		size = sizeof(struct mt76_desc);
820 
821 	q->desc = dmam_alloc_coherent(dev->dma_dev, q->ndesc * size,
822 				      &q->desc_dma, GFP_KERNEL);
823 	if (!q->desc)
824 		return -ENOMEM;
825 
826 	mt76_dma_queue_magic_cnt_init(dev, q);
827 	size = q->ndesc * sizeof(*q->entry);
828 	q->entry = devm_kzalloc(dev->dev, size, GFP_KERNEL);
829 	if (!q->entry)
830 		return -ENOMEM;
831 
832 	ret = mt76_create_page_pool(dev, q);
833 	if (ret)
834 		return ret;
835 
836 	mt76_npu_queue_setup(dev, q);
837 	ret = mt76_wed_dma_setup(dev, q, false);
838 	if (ret)
839 		return ret;
840 
841 	if (mtk_wed_device_active(&dev->mmio.wed)) {
842 		if ((mtk_wed_get_rx_capa(&dev->mmio.wed) && mt76_queue_is_wed_rro(q)) ||
843 		    mt76_queue_is_wed_tx_free(q))
844 			return 0;
845 	}
846 
847 	/* HW specific driver is supposed to reset brand-new EMI queues since
848 	 * it needs to set cpu index pointer.
849 	 */
850 	mt76_dma_queue_reset(dev, q, !mt76_queue_is_emi(q));
851 
852 	return 0;
853 }
854 
855 static void
856 mt76_dma_rx_cleanup(struct mt76_dev *dev, struct mt76_queue *q)
857 {
858 	void *buf;
859 	bool more;
860 
861 	if (!q->ndesc)
862 		return;
863 
864 	if (mt76_queue_is_npu(q)) {
865 		mt76_npu_queue_cleanup(dev, q);
866 		return;
867 	}
868 
869 	do {
870 		spin_lock_bh(&q->lock);
871 		buf = mt76_dma_dequeue(dev, q, true, NULL, NULL, &more, NULL);
872 		spin_unlock_bh(&q->lock);
873 
874 		if (!buf)
875 			break;
876 
877 		if (!mt76_queue_is_wed_rro(q))
878 			mt76_put_page_pool_buf(buf, false);
879 	} while (1);
880 
881 	spin_lock_bh(&q->lock);
882 	if (q->rx_head) {
883 		dev_kfree_skb(q->rx_head);
884 		q->rx_head = NULL;
885 	}
886 
887 	spin_unlock_bh(&q->lock);
888 }
889 
890 static void
891 mt76_dma_rx_reset(struct mt76_dev *dev, enum mt76_rxq_id qid)
892 {
893 	struct mt76_queue *q = &dev->q_rx[qid];
894 
895 	if (!q->ndesc)
896 		return;
897 
898 	if (!mt76_queue_is_wed_rro_ind(q) &&
899 	    !mt76_queue_is_wed_rro_rxdmad_c(q) && !mt76_queue_is_npu(q)) {
900 		int i;
901 
902 		for (i = 0; i < q->ndesc; i++)
903 			q->desc[i].ctrl = cpu_to_le32(MT_DMA_CTL_DMA_DONE);
904 	}
905 
906 	mt76_dma_rx_cleanup(dev, q);
907 
908 	/* reset WED rx queues */
909 	mt76_wed_dma_setup(dev, q, true);
910 
911 	if (mt76_queue_is_wed_tx_free(q))
912 		return;
913 
914 	if (mtk_wed_device_active(&dev->mmio.wed) &&
915 	    mt76_queue_is_wed_rro(q))
916 		return;
917 
918 	mt76_dma_sync_idx(dev, q);
919 	if (mt76_queue_is_npu(q))
920 		mt76_npu_fill_rx_queue(dev, q);
921 	else
922 		mt76_dma_rx_fill(dev, q, false);
923 }
924 
925 static void
926 mt76_add_fragment(struct mt76_dev *dev, struct mt76_queue *q, void *data,
927 		  int len, bool more, u32 info, bool allow_direct)
928 {
929 	struct sk_buff *skb = q->rx_head;
930 	struct skb_shared_info *shinfo = skb_shinfo(skb);
931 	int nr_frags = shinfo->nr_frags;
932 
933 	if (nr_frags < ARRAY_SIZE(shinfo->frags)) {
934 		struct page *page = virt_to_head_page(data);
935 		int offset = data - page_address(page) + q->buf_offset;
936 
937 		skb_add_rx_frag(skb, nr_frags, page, offset, len, q->buf_size);
938 	} else {
939 		mt76_put_page_pool_buf(data, allow_direct);
940 	}
941 
942 	if (more)
943 		return;
944 
945 	q->rx_head = NULL;
946 	if (nr_frags < ARRAY_SIZE(shinfo->frags))
947 		dev->drv->rx_skb(dev, q - dev->q_rx, skb, &info);
948 	else
949 		dev_kfree_skb(skb);
950 }
951 
952 static int
953 mt76_dma_rx_process(struct mt76_dev *dev, struct mt76_queue *q, int budget)
954 {
955 	int len, data_len, done = 0, dma_idx;
956 	struct sk_buff *skb;
957 	unsigned char *data;
958 	bool check_ddone = false;
959 	bool allow_direct = !mt76_queue_is_wed_rx(q);
960 	bool more;
961 
962 	if ((q->flags & MT_QFLAG_WED_RRO_EN) ||
963 	    (IS_ENABLED(CONFIG_NET_MEDIATEK_SOC_WED) &&
964 	     mt76_queue_is_wed_tx_free(q))) {
965 		dma_idx = Q_READ(q, dma_idx);
966 		check_ddone = true;
967 	}
968 
969 	while (done < budget) {
970 		bool drop = false;
971 		u32 info;
972 
973 		if (check_ddone) {
974 			if (q->tail == dma_idx)
975 				dma_idx = Q_READ(q, dma_idx);
976 
977 			if (q->tail == dma_idx)
978 				break;
979 		}
980 
981 		data = mt76_dma_dequeue(dev, q, false, &len, &info, &more,
982 					&drop);
983 		if (!data)
984 			break;
985 
986 		if (PTR_ERR(data) == -EAGAIN) {
987 			done++;
988 			continue;
989 		}
990 
991 		if (mt76_queue_is_wed_rro_ind(q) && dev->drv->rx_rro_ind_process)
992 			dev->drv->rx_rro_ind_process(dev, data);
993 
994 		if (mt76_queue_is_wed_rro(q) &&
995 		    !mt76_queue_is_wed_rro_rxdmad_c(q)) {
996 			done++;
997 			continue;
998 		}
999 
1000 		if (drop)
1001 			goto free_frag;
1002 
1003 		if (q->rx_head)
1004 			data_len = q->buf_size;
1005 		else
1006 			data_len = SKB_WITH_OVERHEAD(q->buf_size);
1007 
1008 		if (data_len < len + q->buf_offset) {
1009 			dev_kfree_skb(q->rx_head);
1010 			q->rx_head = NULL;
1011 			goto free_frag;
1012 		}
1013 
1014 		if (q->rx_head) {
1015 			mt76_add_fragment(dev, q, data, len, more, info,
1016 					  allow_direct);
1017 			continue;
1018 		}
1019 
1020 		if (!more && dev->drv->rx_check &&
1021 		    !(dev->drv->rx_check(dev, data, len)))
1022 			goto free_frag;
1023 
1024 		skb = napi_build_skb(data, q->buf_size);
1025 		if (!skb)
1026 			goto free_frag;
1027 
1028 		skb_reserve(skb, q->buf_offset);
1029 		skb_mark_for_recycle(skb);
1030 
1031 		*(u32 *)skb->cb = info;
1032 
1033 		__skb_put(skb, len);
1034 		done++;
1035 
1036 		if (more) {
1037 			q->rx_head = skb;
1038 			continue;
1039 		}
1040 
1041 		dev->drv->rx_skb(dev, q - dev->q_rx, skb, &info);
1042 		continue;
1043 
1044 free_frag:
1045 		mt76_put_page_pool_buf(data, allow_direct);
1046 	}
1047 
1048 	mt76_dma_rx_fill(dev, q, true);
1049 	return done;
1050 }
1051 
1052 int mt76_dma_rx_poll(struct napi_struct *napi, int budget)
1053 {
1054 	struct mt76_dev *dev;
1055 	int qid, done = 0, cur;
1056 
1057 	dev = mt76_priv(napi->dev);
1058 	qid = napi - dev->napi;
1059 
1060 	rcu_read_lock();
1061 
1062 	do {
1063 		cur = mt76_dma_rx_process(dev, &dev->q_rx[qid], budget - done);
1064 		mt76_rx_poll_complete(dev, qid, napi);
1065 		done += cur;
1066 	} while (cur && done < budget);
1067 
1068 	rcu_read_unlock();
1069 
1070 	if (done < budget && napi_complete(napi))
1071 		dev->drv->rx_poll_complete(dev, qid);
1072 
1073 	return done;
1074 }
1075 EXPORT_SYMBOL_GPL(mt76_dma_rx_poll);
1076 
1077 static void
1078 mt76_dma_rx_queue_init(struct mt76_dev *dev, enum mt76_rxq_id qid,
1079 		       int (*poll)(struct napi_struct *napi, int budget))
1080 {
1081 	netif_napi_add(dev->napi_dev, &dev->napi[qid], poll);
1082 	mt76_dma_rx_fill_buf(dev, &dev->q_rx[qid], false);
1083 	napi_enable(&dev->napi[qid]);
1084 }
1085 
1086 static int
1087 mt76_dma_init(struct mt76_dev *dev,
1088 	      int (*poll)(struct napi_struct *napi, int budget))
1089 {
1090 	struct mt76_dev **priv;
1091 	int i;
1092 
1093 	dev->napi_dev = alloc_netdev_dummy(sizeof(struct mt76_dev *));
1094 	if (!dev->napi_dev)
1095 		return -ENOMEM;
1096 
1097 	/* napi_dev private data points to mt76_dev parent, so, mt76_dev
1098 	 * can be retrieved given napi_dev
1099 	 */
1100 	priv = netdev_priv(dev->napi_dev);
1101 	*priv = dev;
1102 
1103 	dev->tx_napi_dev = alloc_netdev_dummy(sizeof(struct mt76_dev *));
1104 	if (!dev->tx_napi_dev) {
1105 		free_netdev(dev->napi_dev);
1106 		return -ENOMEM;
1107 	}
1108 	priv = netdev_priv(dev->tx_napi_dev);
1109 	*priv = dev;
1110 
1111 	snprintf(dev->napi_dev->name, sizeof(dev->napi_dev->name), "%s",
1112 		 wiphy_name(dev->hw->wiphy));
1113 	dev->napi_dev->threaded = 1;
1114 	init_completion(&dev->mmio.wed_reset);
1115 	init_completion(&dev->mmio.wed_reset_complete);
1116 
1117 	mt76_for_each_q_rx(dev, i) {
1118 		if (mt76_queue_is_wed_rro(&dev->q_rx[i]))
1119 			continue;
1120 
1121 		mt76_dma_rx_queue_init(dev, i, poll);
1122 	}
1123 
1124 	return 0;
1125 }
1126 
1127 static const struct mt76_queue_ops mt76_dma_ops = {
1128 	.init = mt76_dma_init,
1129 	.alloc = mt76_dma_alloc_queue,
1130 	.reset_q = mt76_dma_queue_reset,
1131 	.tx_queue_skb_raw = mt76_dma_tx_queue_skb_raw,
1132 	.tx_queue_skb = mt76_dma_tx_queue_skb,
1133 	.tx_cleanup = mt76_dma_tx_cleanup,
1134 	.rx_queue_init = mt76_dma_rx_queue_init,
1135 	.rx_cleanup = mt76_dma_rx_cleanup,
1136 	.rx_reset = mt76_dma_rx_reset,
1137 	.kick = mt76_dma_kick_queue,
1138 };
1139 
1140 void mt76_dma_attach(struct mt76_dev *dev)
1141 {
1142 	dev->queue_ops = &mt76_dma_ops;
1143 }
1144 EXPORT_SYMBOL_GPL(mt76_dma_attach);
1145 
1146 void mt76_dma_cleanup(struct mt76_dev *dev)
1147 {
1148 	int i;
1149 
1150 	mt76_worker_disable(&dev->tx_worker);
1151 	napi_disable(&dev->tx_napi);
1152 	netif_napi_del(&dev->tx_napi);
1153 
1154 	for (i = 0; i < ARRAY_SIZE(dev->phys); i++) {
1155 		struct mt76_phy *phy = dev->phys[i];
1156 		int j;
1157 
1158 		if (!phy)
1159 			continue;
1160 
1161 		for (j = 0; j < ARRAY_SIZE(phy->q_tx); j++)
1162 			mt76_dma_tx_cleanup(dev, phy->q_tx[j], true);
1163 	}
1164 
1165 	for (i = 0; i < ARRAY_SIZE(dev->q_mcu); i++)
1166 		mt76_dma_tx_cleanup(dev, dev->q_mcu[i], true);
1167 
1168 	mt76_for_each_q_rx(dev, i) {
1169 		struct mt76_queue *q = &dev->q_rx[i];
1170 
1171 		if (mtk_wed_device_active(&dev->mmio.wed) &&
1172 		    mt76_queue_is_wed_rro(q))
1173 			continue;
1174 
1175 		netif_napi_del(&dev->napi[i]);
1176 		mt76_dma_rx_cleanup(dev, q);
1177 
1178 		page_pool_destroy(q->page_pool);
1179 	}
1180 
1181 	if (mtk_wed_device_active(&dev->mmio.wed))
1182 		mtk_wed_device_detach(&dev->mmio.wed);
1183 
1184 	if (mtk_wed_device_active(&dev->mmio.wed_hif2))
1185 		mtk_wed_device_detach(&dev->mmio.wed_hif2);
1186 
1187 	mt76_free_pending_txwi(dev);
1188 	mt76_free_pending_rxwi(dev);
1189 	free_netdev(dev->napi_dev);
1190 	free_netdev(dev->tx_napi_dev);
1191 }
1192 EXPORT_SYMBOL_GPL(mt76_dma_cleanup);
1193