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 static int
11 mt76_dma_alloc_queue(struct mt76_dev *dev, struct mt76_queue *q,
12 		     int idx, int n_desc, int bufsize,
13 		     u32 ring_base)
14 {
15 	int size;
16 	int i;
17 
18 	spin_lock_init(&q->lock);
19 
20 	q->regs = dev->mmio.regs + ring_base + idx * MT_RING_SIZE;
21 	q->ndesc = n_desc;
22 	q->buf_size = bufsize;
23 	q->hw_idx = idx;
24 
25 	size = q->ndesc * sizeof(struct mt76_desc);
26 	q->desc = dmam_alloc_coherent(dev->dev, size, &q->desc_dma, GFP_KERNEL);
27 	if (!q->desc)
28 		return -ENOMEM;
29 
30 	size = q->ndesc * sizeof(*q->entry);
31 	q->entry = devm_kzalloc(dev->dev, size, GFP_KERNEL);
32 	if (!q->entry)
33 		return -ENOMEM;
34 
35 	/* clear descriptors */
36 	for (i = 0; i < q->ndesc; i++)
37 		q->desc[i].ctrl = cpu_to_le32(MT_DMA_CTL_DMA_DONE);
38 
39 	writel(q->desc_dma, &q->regs->desc_base);
40 	writel(0, &q->regs->cpu_idx);
41 	writel(0, &q->regs->dma_idx);
42 	writel(q->ndesc, &q->regs->ring_size);
43 
44 	return 0;
45 }
46 
47 static int
48 mt76_dma_add_buf(struct mt76_dev *dev, struct mt76_queue *q,
49 		 struct mt76_queue_buf *buf, int nbufs, u32 info,
50 		 struct sk_buff *skb, void *txwi)
51 {
52 	struct mt76_desc *desc;
53 	u32 ctrl;
54 	int i, idx = -1;
55 
56 	if (txwi) {
57 		q->entry[q->head].txwi = DMA_DUMMY_DATA;
58 		q->entry[q->head].skip_buf0 = true;
59 	}
60 
61 	for (i = 0; i < nbufs; i += 2, buf += 2) {
62 		u32 buf0 = buf[0].addr, buf1 = 0;
63 
64 		ctrl = FIELD_PREP(MT_DMA_CTL_SD_LEN0, buf[0].len);
65 		if (i < nbufs - 1) {
66 			buf1 = buf[1].addr;
67 			ctrl |= FIELD_PREP(MT_DMA_CTL_SD_LEN1, buf[1].len);
68 		}
69 
70 		if (i == nbufs - 1)
71 			ctrl |= MT_DMA_CTL_LAST_SEC0;
72 		else if (i == nbufs - 2)
73 			ctrl |= MT_DMA_CTL_LAST_SEC1;
74 
75 		idx = q->head;
76 		q->head = (q->head + 1) % q->ndesc;
77 
78 		desc = &q->desc[idx];
79 
80 		WRITE_ONCE(desc->buf0, cpu_to_le32(buf0));
81 		WRITE_ONCE(desc->buf1, cpu_to_le32(buf1));
82 		WRITE_ONCE(desc->info, cpu_to_le32(info));
83 		WRITE_ONCE(desc->ctrl, cpu_to_le32(ctrl));
84 
85 		q->queued++;
86 	}
87 
88 	q->entry[idx].txwi = txwi;
89 	q->entry[idx].skb = skb;
90 
91 	return idx;
92 }
93 
94 static void
95 mt76_dma_tx_cleanup_idx(struct mt76_dev *dev, struct mt76_queue *q, int idx,
96 			struct mt76_queue_entry *prev_e)
97 {
98 	struct mt76_queue_entry *e = &q->entry[idx];
99 	__le32 __ctrl = READ_ONCE(q->desc[idx].ctrl);
100 	u32 ctrl = le32_to_cpu(__ctrl);
101 
102 	if (!e->skip_buf0) {
103 		__le32 addr = READ_ONCE(q->desc[idx].buf0);
104 		u32 len = FIELD_GET(MT_DMA_CTL_SD_LEN0, ctrl);
105 
106 		dma_unmap_single(dev->dev, le32_to_cpu(addr), len,
107 				 DMA_TO_DEVICE);
108 	}
109 
110 	if (!(ctrl & MT_DMA_CTL_LAST_SEC0)) {
111 		__le32 addr = READ_ONCE(q->desc[idx].buf1);
112 		u32 len = FIELD_GET(MT_DMA_CTL_SD_LEN1, ctrl);
113 
114 		dma_unmap_single(dev->dev, le32_to_cpu(addr), len,
115 				 DMA_TO_DEVICE);
116 	}
117 
118 	if (e->txwi == DMA_DUMMY_DATA)
119 		e->txwi = NULL;
120 
121 	if (e->skb == DMA_DUMMY_DATA)
122 		e->skb = NULL;
123 
124 	*prev_e = *e;
125 	memset(e, 0, sizeof(*e));
126 }
127 
128 static void
129 mt76_dma_sync_idx(struct mt76_dev *dev, struct mt76_queue *q)
130 {
131 	writel(q->desc_dma, &q->regs->desc_base);
132 	writel(q->ndesc, &q->regs->ring_size);
133 	q->head = readl(&q->regs->dma_idx);
134 	q->tail = q->head;
135 	writel(q->head, &q->regs->cpu_idx);
136 }
137 
138 static void
139 mt76_dma_tx_cleanup(struct mt76_dev *dev, enum mt76_txq_id qid, bool flush)
140 {
141 	struct mt76_sw_queue *sq = &dev->q_tx[qid];
142 	struct mt76_queue *q = sq->q;
143 	struct mt76_queue_entry entry;
144 	unsigned int n_swq_queued[4] = {};
145 	unsigned int n_queued = 0;
146 	bool wake = false;
147 	int i, last;
148 
149 	if (!q)
150 		return;
151 
152 	if (flush)
153 		last = -1;
154 	else
155 		last = readl(&q->regs->dma_idx);
156 
157 	while ((q->queued > n_queued) && q->tail != last) {
158 		mt76_dma_tx_cleanup_idx(dev, q, q->tail, &entry);
159 		if (entry.schedule)
160 			n_swq_queued[entry.qid]++;
161 
162 		q->tail = (q->tail + 1) % q->ndesc;
163 		n_queued++;
164 
165 		if (entry.skb)
166 			dev->drv->tx_complete_skb(dev, qid, &entry);
167 
168 		if (entry.txwi) {
169 			if (!(dev->drv->drv_flags & MT_DRV_TXWI_NO_FREE))
170 				mt76_put_txwi(dev, entry.txwi);
171 			wake = !flush;
172 		}
173 
174 		if (!flush && q->tail == last)
175 			last = readl(&q->regs->dma_idx);
176 	}
177 
178 	spin_lock_bh(&q->lock);
179 
180 	q->queued -= n_queued;
181 	for (i = 0; i < ARRAY_SIZE(n_swq_queued); i++) {
182 		if (!n_swq_queued[i])
183 			continue;
184 
185 		dev->q_tx[i].swq_queued -= n_swq_queued[i];
186 	}
187 
188 	if (flush)
189 		mt76_dma_sync_idx(dev, q);
190 
191 	wake = wake && q->stopped &&
192 	       qid < IEEE80211_NUM_ACS && q->queued < q->ndesc - 8;
193 	if (wake)
194 		q->stopped = false;
195 
196 	if (!q->queued)
197 		wake_up(&dev->tx_wait);
198 
199 	spin_unlock_bh(&q->lock);
200 
201 	if (wake)
202 		ieee80211_wake_queue(dev->hw, qid);
203 }
204 
205 static void *
206 mt76_dma_get_buf(struct mt76_dev *dev, struct mt76_queue *q, int idx,
207 		 int *len, u32 *info, bool *more)
208 {
209 	struct mt76_queue_entry *e = &q->entry[idx];
210 	struct mt76_desc *desc = &q->desc[idx];
211 	dma_addr_t buf_addr;
212 	void *buf = e->buf;
213 	int buf_len = SKB_WITH_OVERHEAD(q->buf_size);
214 
215 	buf_addr = le32_to_cpu(READ_ONCE(desc->buf0));
216 	if (len) {
217 		u32 ctl = le32_to_cpu(READ_ONCE(desc->ctrl));
218 		*len = FIELD_GET(MT_DMA_CTL_SD_LEN0, ctl);
219 		*more = !(ctl & MT_DMA_CTL_LAST_SEC0);
220 	}
221 
222 	if (info)
223 		*info = le32_to_cpu(desc->info);
224 
225 	dma_unmap_single(dev->dev, buf_addr, buf_len, DMA_FROM_DEVICE);
226 	e->buf = NULL;
227 
228 	return buf;
229 }
230 
231 static void *
232 mt76_dma_dequeue(struct mt76_dev *dev, struct mt76_queue *q, bool flush,
233 		 int *len, u32 *info, bool *more)
234 {
235 	int idx = q->tail;
236 
237 	*more = false;
238 	if (!q->queued)
239 		return NULL;
240 
241 	if (!flush && !(q->desc[idx].ctrl & cpu_to_le32(MT_DMA_CTL_DMA_DONE)))
242 		return NULL;
243 
244 	q->tail = (q->tail + 1) % q->ndesc;
245 	q->queued--;
246 
247 	return mt76_dma_get_buf(dev, q, idx, len, info, more);
248 }
249 
250 static void
251 mt76_dma_kick_queue(struct mt76_dev *dev, struct mt76_queue *q)
252 {
253 	writel(q->head, &q->regs->cpu_idx);
254 }
255 
256 static int
257 mt76_dma_tx_queue_skb_raw(struct mt76_dev *dev, enum mt76_txq_id qid,
258 			  struct sk_buff *skb, u32 tx_info)
259 {
260 	struct mt76_queue *q = dev->q_tx[qid].q;
261 	struct mt76_queue_buf buf;
262 	dma_addr_t addr;
263 
264 	addr = dma_map_single(dev->dev, skb->data, skb->len,
265 			      DMA_TO_DEVICE);
266 	if (unlikely(dma_mapping_error(dev->dev, addr)))
267 		return -ENOMEM;
268 
269 	buf.addr = addr;
270 	buf.len = skb->len;
271 
272 	spin_lock_bh(&q->lock);
273 	mt76_dma_add_buf(dev, q, &buf, 1, tx_info, skb, NULL);
274 	mt76_dma_kick_queue(dev, q);
275 	spin_unlock_bh(&q->lock);
276 
277 	return 0;
278 }
279 
280 static int
281 mt76_dma_tx_queue_skb(struct mt76_dev *dev, enum mt76_txq_id qid,
282 		      struct sk_buff *skb, struct mt76_wcid *wcid,
283 		      struct ieee80211_sta *sta)
284 {
285 	struct mt76_queue *q = dev->q_tx[qid].q;
286 	struct mt76_tx_info tx_info = {
287 		.skb = skb,
288 	};
289 	int len, n = 0, ret = -ENOMEM;
290 	struct mt76_queue_entry e;
291 	struct mt76_txwi_cache *t;
292 	struct sk_buff *iter;
293 	dma_addr_t addr;
294 	u8 *txwi;
295 
296 	t = mt76_get_txwi(dev);
297 	if (!t) {
298 		ieee80211_free_txskb(dev->hw, skb);
299 		return -ENOMEM;
300 	}
301 	txwi = mt76_get_txwi_ptr(dev, t);
302 
303 	skb->prev = skb->next = NULL;
304 	if (dev->drv->drv_flags & MT_DRV_TX_ALIGNED4_SKBS)
305 		mt76_insert_hdr_pad(skb);
306 
307 	len = skb_headlen(skb);
308 	addr = dma_map_single(dev->dev, skb->data, len, DMA_TO_DEVICE);
309 	if (unlikely(dma_mapping_error(dev->dev, addr)))
310 		goto free;
311 
312 	tx_info.buf[n].addr = t->dma_addr;
313 	tx_info.buf[n++].len = dev->drv->txwi_size;
314 	tx_info.buf[n].addr = addr;
315 	tx_info.buf[n++].len = len;
316 
317 	skb_walk_frags(skb, iter) {
318 		if (n == ARRAY_SIZE(tx_info.buf))
319 			goto unmap;
320 
321 		addr = dma_map_single(dev->dev, iter->data, iter->len,
322 				      DMA_TO_DEVICE);
323 		if (unlikely(dma_mapping_error(dev->dev, addr)))
324 			goto unmap;
325 
326 		tx_info.buf[n].addr = addr;
327 		tx_info.buf[n++].len = iter->len;
328 	}
329 	tx_info.nbuf = n;
330 
331 	dma_sync_single_for_cpu(dev->dev, t->dma_addr, dev->drv->txwi_size,
332 				DMA_TO_DEVICE);
333 	ret = dev->drv->tx_prepare_skb(dev, txwi, qid, wcid, sta, &tx_info);
334 	dma_sync_single_for_device(dev->dev, t->dma_addr, dev->drv->txwi_size,
335 				   DMA_TO_DEVICE);
336 	if (ret < 0)
337 		goto unmap;
338 
339 	if (q->queued + (tx_info.nbuf + 1) / 2 >= q->ndesc - 1) {
340 		ret = -ENOMEM;
341 		goto unmap;
342 	}
343 
344 	return mt76_dma_add_buf(dev, q, tx_info.buf, tx_info.nbuf,
345 				tx_info.info, tx_info.skb, t);
346 
347 unmap:
348 	for (n--; n > 0; n--)
349 		dma_unmap_single(dev->dev, tx_info.buf[n].addr,
350 				 tx_info.buf[n].len, DMA_TO_DEVICE);
351 
352 free:
353 	e.skb = tx_info.skb;
354 	e.txwi = t;
355 	dev->drv->tx_complete_skb(dev, qid, &e);
356 	mt76_put_txwi(dev, t);
357 	return ret;
358 }
359 
360 static int
361 mt76_dma_rx_fill(struct mt76_dev *dev, struct mt76_queue *q)
362 {
363 	dma_addr_t addr;
364 	void *buf;
365 	int frames = 0;
366 	int len = SKB_WITH_OVERHEAD(q->buf_size);
367 	int offset = q->buf_offset;
368 
369 	spin_lock_bh(&q->lock);
370 
371 	while (q->queued < q->ndesc - 1) {
372 		struct mt76_queue_buf qbuf;
373 
374 		buf = page_frag_alloc(&q->rx_page, q->buf_size, GFP_ATOMIC);
375 		if (!buf)
376 			break;
377 
378 		addr = dma_map_single(dev->dev, buf, len, DMA_FROM_DEVICE);
379 		if (unlikely(dma_mapping_error(dev->dev, addr))) {
380 			skb_free_frag(buf);
381 			break;
382 		}
383 
384 		qbuf.addr = addr + offset;
385 		qbuf.len = len - offset;
386 		mt76_dma_add_buf(dev, q, &qbuf, 1, 0, buf, NULL);
387 		frames++;
388 	}
389 
390 	if (frames)
391 		mt76_dma_kick_queue(dev, q);
392 
393 	spin_unlock_bh(&q->lock);
394 
395 	return frames;
396 }
397 
398 static void
399 mt76_dma_rx_cleanup(struct mt76_dev *dev, struct mt76_queue *q)
400 {
401 	struct page *page;
402 	void *buf;
403 	bool more;
404 
405 	spin_lock_bh(&q->lock);
406 	do {
407 		buf = mt76_dma_dequeue(dev, q, true, NULL, NULL, &more);
408 		if (!buf)
409 			break;
410 
411 		skb_free_frag(buf);
412 	} while (1);
413 	spin_unlock_bh(&q->lock);
414 
415 	if (!q->rx_page.va)
416 		return;
417 
418 	page = virt_to_page(q->rx_page.va);
419 	__page_frag_cache_drain(page, q->rx_page.pagecnt_bias);
420 	memset(&q->rx_page, 0, sizeof(q->rx_page));
421 }
422 
423 static void
424 mt76_dma_rx_reset(struct mt76_dev *dev, enum mt76_rxq_id qid)
425 {
426 	struct mt76_queue *q = &dev->q_rx[qid];
427 	int i;
428 
429 	for (i = 0; i < q->ndesc; i++)
430 		q->desc[i].ctrl &= ~cpu_to_le32(MT_DMA_CTL_DMA_DONE);
431 
432 	mt76_dma_rx_cleanup(dev, q);
433 	mt76_dma_sync_idx(dev, q);
434 	mt76_dma_rx_fill(dev, q);
435 
436 	if (!q->rx_head)
437 		return;
438 
439 	dev_kfree_skb(q->rx_head);
440 	q->rx_head = NULL;
441 }
442 
443 static void
444 mt76_add_fragment(struct mt76_dev *dev, struct mt76_queue *q, void *data,
445 		  int len, bool more)
446 {
447 	struct page *page = virt_to_head_page(data);
448 	int offset = data - page_address(page);
449 	struct sk_buff *skb = q->rx_head;
450 
451 	offset += q->buf_offset;
452 	skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset, len,
453 			q->buf_size);
454 
455 	if (more)
456 		return;
457 
458 	q->rx_head = NULL;
459 	dev->drv->rx_skb(dev, q - dev->q_rx, skb);
460 }
461 
462 static int
463 mt76_dma_rx_process(struct mt76_dev *dev, struct mt76_queue *q, int budget)
464 {
465 	int len, data_len, done = 0;
466 	struct sk_buff *skb;
467 	unsigned char *data;
468 	bool more;
469 
470 	while (done < budget) {
471 		u32 info;
472 
473 		data = mt76_dma_dequeue(dev, q, false, &len, &info, &more);
474 		if (!data)
475 			break;
476 
477 		if (q->rx_head)
478 			data_len = q->buf_size;
479 		else
480 			data_len = SKB_WITH_OVERHEAD(q->buf_size);
481 
482 		if (data_len < len + q->buf_offset) {
483 			dev_kfree_skb(q->rx_head);
484 			q->rx_head = NULL;
485 
486 			skb_free_frag(data);
487 			continue;
488 		}
489 
490 		if (q->rx_head) {
491 			mt76_add_fragment(dev, q, data, len, more);
492 			continue;
493 		}
494 
495 		skb = build_skb(data, q->buf_size);
496 		if (!skb) {
497 			skb_free_frag(data);
498 			continue;
499 		}
500 		skb_reserve(skb, q->buf_offset);
501 
502 		if (q == &dev->q_rx[MT_RXQ_MCU]) {
503 			u32 *rxfce = (u32 *)skb->cb;
504 			*rxfce = info;
505 		}
506 
507 		__skb_put(skb, len);
508 		done++;
509 
510 		if (more) {
511 			q->rx_head = skb;
512 			continue;
513 		}
514 
515 		dev->drv->rx_skb(dev, q - dev->q_rx, skb);
516 	}
517 
518 	mt76_dma_rx_fill(dev, q);
519 	return done;
520 }
521 
522 static int
523 mt76_dma_rx_poll(struct napi_struct *napi, int budget)
524 {
525 	struct mt76_dev *dev;
526 	int qid, done = 0, cur;
527 
528 	dev = container_of(napi->dev, struct mt76_dev, napi_dev);
529 	qid = napi - dev->napi;
530 
531 	rcu_read_lock();
532 
533 	do {
534 		cur = mt76_dma_rx_process(dev, &dev->q_rx[qid], budget - done);
535 		mt76_rx_poll_complete(dev, qid, napi);
536 		done += cur;
537 	} while (cur && done < budget);
538 
539 	rcu_read_unlock();
540 
541 	if (done < budget && napi_complete(napi))
542 		dev->drv->rx_poll_complete(dev, qid);
543 
544 	return done;
545 }
546 
547 static int
548 mt76_dma_init(struct mt76_dev *dev)
549 {
550 	int i;
551 
552 	init_dummy_netdev(&dev->napi_dev);
553 
554 	for (i = 0; i < ARRAY_SIZE(dev->q_rx); i++) {
555 		netif_napi_add(&dev->napi_dev, &dev->napi[i], mt76_dma_rx_poll,
556 			       64);
557 		mt76_dma_rx_fill(dev, &dev->q_rx[i]);
558 		skb_queue_head_init(&dev->rx_skb[i]);
559 		napi_enable(&dev->napi[i]);
560 	}
561 
562 	return 0;
563 }
564 
565 static const struct mt76_queue_ops mt76_dma_ops = {
566 	.init = mt76_dma_init,
567 	.alloc = mt76_dma_alloc_queue,
568 	.tx_queue_skb_raw = mt76_dma_tx_queue_skb_raw,
569 	.tx_queue_skb = mt76_dma_tx_queue_skb,
570 	.tx_cleanup = mt76_dma_tx_cleanup,
571 	.rx_reset = mt76_dma_rx_reset,
572 	.kick = mt76_dma_kick_queue,
573 };
574 
575 void mt76_dma_attach(struct mt76_dev *dev)
576 {
577 	dev->queue_ops = &mt76_dma_ops;
578 }
579 EXPORT_SYMBOL_GPL(mt76_dma_attach);
580 
581 void mt76_dma_cleanup(struct mt76_dev *dev)
582 {
583 	int i;
584 
585 	netif_napi_del(&dev->tx_napi);
586 	for (i = 0; i < ARRAY_SIZE(dev->q_tx); i++)
587 		mt76_dma_tx_cleanup(dev, i, true);
588 
589 	for (i = 0; i < ARRAY_SIZE(dev->q_rx); i++) {
590 		netif_napi_del(&dev->napi[i]);
591 		mt76_dma_rx_cleanup(dev, &dev->q_rx[i]);
592 	}
593 }
594 EXPORT_SYMBOL_GPL(mt76_dma_cleanup);
595