1 /*
2  * Copyright (C) 2016 Felix Fietkau <nbd@nbd.name>
3  *
4  * Permission to use, copy, modify, and/or distribute this software for any
5  * purpose with or without fee is hereby granted, provided that the above
6  * copyright notice and this permission notice appear in all copies.
7  *
8  * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
9  * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
10  * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
11  * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
12  * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
13  * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
14  * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
15  */
16 
17 #include <linux/dma-mapping.h>
18 #include "mt76.h"
19 #include "dma.h"
20 
21 static int
22 mt76_dma_alloc_queue(struct mt76_dev *dev, struct mt76_queue *q,
23 		     int idx, int n_desc, int bufsize,
24 		     u32 ring_base)
25 {
26 	int size;
27 	int i;
28 
29 	spin_lock_init(&q->lock);
30 
31 	q->regs = dev->mmio.regs + ring_base + idx * MT_RING_SIZE;
32 	q->ndesc = n_desc;
33 	q->buf_size = bufsize;
34 	q->hw_idx = idx;
35 
36 	size = q->ndesc * sizeof(struct mt76_desc);
37 	q->desc = dmam_alloc_coherent(dev->dev, size, &q->desc_dma, GFP_KERNEL);
38 	if (!q->desc)
39 		return -ENOMEM;
40 
41 	size = q->ndesc * sizeof(*q->entry);
42 	q->entry = devm_kzalloc(dev->dev, size, GFP_KERNEL);
43 	if (!q->entry)
44 		return -ENOMEM;
45 
46 	/* clear descriptors */
47 	for (i = 0; i < q->ndesc; i++)
48 		q->desc[i].ctrl = cpu_to_le32(MT_DMA_CTL_DMA_DONE);
49 
50 	writel(q->desc_dma, &q->regs->desc_base);
51 	writel(0, &q->regs->cpu_idx);
52 	writel(0, &q->regs->dma_idx);
53 	writel(q->ndesc, &q->regs->ring_size);
54 
55 	return 0;
56 }
57 
58 static int
59 mt76_dma_add_buf(struct mt76_dev *dev, struct mt76_queue *q,
60 		 struct mt76_queue_buf *buf, int nbufs, u32 info,
61 		 struct sk_buff *skb, void *txwi)
62 {
63 	struct mt76_desc *desc;
64 	u32 ctrl;
65 	int i, idx = -1;
66 
67 	if (txwi)
68 		q->entry[q->head].txwi = DMA_DUMMY_DATA;
69 
70 	for (i = 0; i < nbufs; i += 2, buf += 2) {
71 		u32 buf0 = buf[0].addr, buf1 = 0;
72 
73 		ctrl = FIELD_PREP(MT_DMA_CTL_SD_LEN0, buf[0].len);
74 		if (i < nbufs - 1) {
75 			buf1 = buf[1].addr;
76 			ctrl |= FIELD_PREP(MT_DMA_CTL_SD_LEN1, buf[1].len);
77 		}
78 
79 		if (i == nbufs - 1)
80 			ctrl |= MT_DMA_CTL_LAST_SEC0;
81 		else if (i == nbufs - 2)
82 			ctrl |= MT_DMA_CTL_LAST_SEC1;
83 
84 		idx = q->head;
85 		q->head = (q->head + 1) % q->ndesc;
86 
87 		desc = &q->desc[idx];
88 
89 		WRITE_ONCE(desc->buf0, cpu_to_le32(buf0));
90 		WRITE_ONCE(desc->buf1, cpu_to_le32(buf1));
91 		WRITE_ONCE(desc->info, cpu_to_le32(info));
92 		WRITE_ONCE(desc->ctrl, cpu_to_le32(ctrl));
93 
94 		q->queued++;
95 	}
96 
97 	q->entry[idx].txwi = txwi;
98 	q->entry[idx].skb = skb;
99 
100 	return idx;
101 }
102 
103 static void
104 mt76_dma_tx_cleanup_idx(struct mt76_dev *dev, struct mt76_queue *q, int idx,
105 			struct mt76_queue_entry *prev_e)
106 {
107 	struct mt76_queue_entry *e = &q->entry[idx];
108 	__le32 __ctrl = READ_ONCE(q->desc[idx].ctrl);
109 	u32 ctrl = le32_to_cpu(__ctrl);
110 
111 	if (!e->txwi || !e->skb) {
112 		__le32 addr = READ_ONCE(q->desc[idx].buf0);
113 		u32 len = FIELD_GET(MT_DMA_CTL_SD_LEN0, ctrl);
114 
115 		dma_unmap_single(dev->dev, le32_to_cpu(addr), len,
116 				 DMA_TO_DEVICE);
117 	}
118 
119 	if (!(ctrl & MT_DMA_CTL_LAST_SEC0)) {
120 		__le32 addr = READ_ONCE(q->desc[idx].buf1);
121 		u32 len = FIELD_GET(MT_DMA_CTL_SD_LEN1, ctrl);
122 
123 		dma_unmap_single(dev->dev, le32_to_cpu(addr), len,
124 				 DMA_TO_DEVICE);
125 	}
126 
127 	if (e->txwi == DMA_DUMMY_DATA)
128 		e->txwi = NULL;
129 
130 	if (e->skb == DMA_DUMMY_DATA)
131 		e->skb = NULL;
132 
133 	*prev_e = *e;
134 	memset(e, 0, sizeof(*e));
135 }
136 
137 static void
138 mt76_dma_sync_idx(struct mt76_dev *dev, struct mt76_queue *q)
139 {
140 	writel(q->desc_dma, &q->regs->desc_base);
141 	writel(q->ndesc, &q->regs->ring_size);
142 	q->head = readl(&q->regs->dma_idx);
143 	q->tail = q->head;
144 	writel(q->head, &q->regs->cpu_idx);
145 }
146 
147 static void
148 mt76_dma_tx_cleanup(struct mt76_dev *dev, enum mt76_txq_id qid, bool flush)
149 {
150 	struct mt76_sw_queue *sq = &dev->q_tx[qid];
151 	struct mt76_queue *q = sq->q;
152 	struct mt76_queue_entry entry;
153 	unsigned int n_swq_queued[4] = {};
154 	unsigned int n_queued = 0;
155 	bool wake = false;
156 	int i, last;
157 
158 	if (!q)
159 		return;
160 
161 	if (flush)
162 		last = -1;
163 	else
164 		last = readl(&q->regs->dma_idx);
165 
166 	while ((q->queued > n_queued) && q->tail != last) {
167 		mt76_dma_tx_cleanup_idx(dev, q, q->tail, &entry);
168 		if (entry.schedule)
169 			n_swq_queued[entry.qid]++;
170 
171 		q->tail = (q->tail + 1) % q->ndesc;
172 		n_queued++;
173 
174 		if (entry.skb)
175 			dev->drv->tx_complete_skb(dev, qid, &entry);
176 
177 		if (entry.txwi) {
178 			if (!(dev->drv->txwi_flags & MT_TXWI_NO_FREE))
179 				mt76_put_txwi(dev, entry.txwi);
180 			wake = !flush;
181 		}
182 
183 		if (!flush && q->tail == last)
184 			last = readl(&q->regs->dma_idx);
185 	}
186 
187 	spin_lock_bh(&q->lock);
188 
189 	q->queued -= n_queued;
190 	for (i = 0; i < ARRAY_SIZE(n_swq_queued); i++) {
191 		if (!n_swq_queued[i])
192 			continue;
193 
194 		dev->q_tx[i].swq_queued -= n_swq_queued[i];
195 	}
196 
197 	if (flush)
198 		mt76_dma_sync_idx(dev, q);
199 
200 	wake = wake && q->stopped &&
201 	       qid < IEEE80211_NUM_ACS && q->queued < q->ndesc - 8;
202 	if (wake)
203 		q->stopped = false;
204 
205 	if (!q->queued)
206 		wake_up(&dev->tx_wait);
207 
208 	spin_unlock_bh(&q->lock);
209 
210 	if (wake)
211 		ieee80211_wake_queue(dev->hw, qid);
212 }
213 
214 static void *
215 mt76_dma_get_buf(struct mt76_dev *dev, struct mt76_queue *q, int idx,
216 		 int *len, u32 *info, bool *more)
217 {
218 	struct mt76_queue_entry *e = &q->entry[idx];
219 	struct mt76_desc *desc = &q->desc[idx];
220 	dma_addr_t buf_addr;
221 	void *buf = e->buf;
222 	int buf_len = SKB_WITH_OVERHEAD(q->buf_size);
223 
224 	buf_addr = le32_to_cpu(READ_ONCE(desc->buf0));
225 	if (len) {
226 		u32 ctl = le32_to_cpu(READ_ONCE(desc->ctrl));
227 		*len = FIELD_GET(MT_DMA_CTL_SD_LEN0, ctl);
228 		*more = !(ctl & MT_DMA_CTL_LAST_SEC0);
229 	}
230 
231 	if (info)
232 		*info = le32_to_cpu(desc->info);
233 
234 	dma_unmap_single(dev->dev, buf_addr, buf_len, DMA_FROM_DEVICE);
235 	e->buf = NULL;
236 
237 	return buf;
238 }
239 
240 static void *
241 mt76_dma_dequeue(struct mt76_dev *dev, struct mt76_queue *q, bool flush,
242 		 int *len, u32 *info, bool *more)
243 {
244 	int idx = q->tail;
245 
246 	*more = false;
247 	if (!q->queued)
248 		return NULL;
249 
250 	if (!flush && !(q->desc[idx].ctrl & cpu_to_le32(MT_DMA_CTL_DMA_DONE)))
251 		return NULL;
252 
253 	q->tail = (q->tail + 1) % q->ndesc;
254 	q->queued--;
255 
256 	return mt76_dma_get_buf(dev, q, idx, len, info, more);
257 }
258 
259 static void
260 mt76_dma_kick_queue(struct mt76_dev *dev, struct mt76_queue *q)
261 {
262 	writel(q->head, &q->regs->cpu_idx);
263 }
264 
265 static int
266 mt76_dma_tx_queue_skb_raw(struct mt76_dev *dev, enum mt76_txq_id qid,
267 			  struct sk_buff *skb, u32 tx_info)
268 {
269 	struct mt76_queue *q = dev->q_tx[qid].q;
270 	struct mt76_queue_buf buf;
271 	dma_addr_t addr;
272 
273 	addr = dma_map_single(dev->dev, skb->data, skb->len,
274 			      DMA_TO_DEVICE);
275 	if (unlikely(dma_mapping_error(dev->dev, addr)))
276 		return -ENOMEM;
277 
278 	buf.addr = addr;
279 	buf.len = skb->len;
280 
281 	spin_lock_bh(&q->lock);
282 	mt76_dma_add_buf(dev, q, &buf, 1, tx_info, skb, NULL);
283 	mt76_dma_kick_queue(dev, q);
284 	spin_unlock_bh(&q->lock);
285 
286 	return 0;
287 }
288 
289 static int
290 mt76_dma_tx_queue_skb(struct mt76_dev *dev, enum mt76_txq_id qid,
291 		      struct sk_buff *skb, struct mt76_wcid *wcid,
292 		      struct ieee80211_sta *sta)
293 {
294 	struct mt76_queue *q = dev->q_tx[qid].q;
295 	struct mt76_tx_info tx_info = {
296 		.skb = skb,
297 	};
298 	int len, n = 0, ret = -ENOMEM;
299 	struct mt76_queue_entry e;
300 	struct mt76_txwi_cache *t;
301 	struct sk_buff *iter;
302 	dma_addr_t addr;
303 	u8 *txwi;
304 
305 	t = mt76_get_txwi(dev);
306 	if (!t) {
307 		ieee80211_free_txskb(dev->hw, skb);
308 		return -ENOMEM;
309 	}
310 	txwi = mt76_get_txwi_ptr(dev, t);
311 
312 	skb->prev = skb->next = NULL;
313 	if (dev->drv->tx_aligned4_skbs)
314 		mt76_insert_hdr_pad(skb);
315 
316 	len = skb_headlen(skb);
317 	addr = dma_map_single(dev->dev, skb->data, len, DMA_TO_DEVICE);
318 	if (unlikely(dma_mapping_error(dev->dev, addr)))
319 		goto free;
320 
321 	tx_info.buf[n].addr = t->dma_addr;
322 	tx_info.buf[n++].len = dev->drv->txwi_size;
323 	tx_info.buf[n].addr = addr;
324 	tx_info.buf[n++].len = len;
325 
326 	skb_walk_frags(skb, iter) {
327 		if (n == ARRAY_SIZE(tx_info.buf))
328 			goto unmap;
329 
330 		addr = dma_map_single(dev->dev, iter->data, iter->len,
331 				      DMA_TO_DEVICE);
332 		if (unlikely(dma_mapping_error(dev->dev, addr)))
333 			goto unmap;
334 
335 		tx_info.buf[n].addr = addr;
336 		tx_info.buf[n++].len = iter->len;
337 	}
338 	tx_info.nbuf = n;
339 
340 	dma_sync_single_for_cpu(dev->dev, t->dma_addr, dev->drv->txwi_size,
341 				DMA_TO_DEVICE);
342 	ret = dev->drv->tx_prepare_skb(dev, txwi, qid, wcid, sta, &tx_info);
343 	dma_sync_single_for_device(dev->dev, t->dma_addr, dev->drv->txwi_size,
344 				   DMA_TO_DEVICE);
345 	if (ret < 0)
346 		goto unmap;
347 
348 	if (q->queued + (tx_info.nbuf + 1) / 2 >= q->ndesc - 1) {
349 		ret = -ENOMEM;
350 		goto unmap;
351 	}
352 
353 	return mt76_dma_add_buf(dev, q, tx_info.buf, tx_info.nbuf,
354 				tx_info.info, tx_info.skb, t);
355 
356 unmap:
357 	for (n--; n > 0; n--)
358 		dma_unmap_single(dev->dev, tx_info.buf[n].addr,
359 				 tx_info.buf[n].len, DMA_TO_DEVICE);
360 
361 free:
362 	e.skb = tx_info.skb;
363 	e.txwi = t;
364 	dev->drv->tx_complete_skb(dev, qid, &e);
365 	mt76_put_txwi(dev, t);
366 	return ret;
367 }
368 
369 static int
370 mt76_dma_rx_fill(struct mt76_dev *dev, struct mt76_queue *q)
371 {
372 	dma_addr_t addr;
373 	void *buf;
374 	int frames = 0;
375 	int len = SKB_WITH_OVERHEAD(q->buf_size);
376 	int offset = q->buf_offset;
377 	int idx;
378 
379 	spin_lock_bh(&q->lock);
380 
381 	while (q->queued < q->ndesc - 1) {
382 		struct mt76_queue_buf qbuf;
383 
384 		buf = page_frag_alloc(&q->rx_page, q->buf_size, GFP_ATOMIC);
385 		if (!buf)
386 			break;
387 
388 		addr = dma_map_single(dev->dev, buf, len, DMA_FROM_DEVICE);
389 		if (unlikely(dma_mapping_error(dev->dev, addr))) {
390 			skb_free_frag(buf);
391 			break;
392 		}
393 
394 		qbuf.addr = addr + offset;
395 		qbuf.len = len - offset;
396 		idx = mt76_dma_add_buf(dev, q, &qbuf, 1, 0, buf, NULL);
397 		frames++;
398 	}
399 
400 	if (frames)
401 		mt76_dma_kick_queue(dev, q);
402 
403 	spin_unlock_bh(&q->lock);
404 
405 	return frames;
406 }
407 
408 static void
409 mt76_dma_rx_cleanup(struct mt76_dev *dev, struct mt76_queue *q)
410 {
411 	struct page *page;
412 	void *buf;
413 	bool more;
414 
415 	spin_lock_bh(&q->lock);
416 	do {
417 		buf = mt76_dma_dequeue(dev, q, true, NULL, NULL, &more);
418 		if (!buf)
419 			break;
420 
421 		skb_free_frag(buf);
422 	} while (1);
423 	spin_unlock_bh(&q->lock);
424 
425 	if (!q->rx_page.va)
426 		return;
427 
428 	page = virt_to_page(q->rx_page.va);
429 	__page_frag_cache_drain(page, q->rx_page.pagecnt_bias);
430 	memset(&q->rx_page, 0, sizeof(q->rx_page));
431 }
432 
433 static void
434 mt76_dma_rx_reset(struct mt76_dev *dev, enum mt76_rxq_id qid)
435 {
436 	struct mt76_queue *q = &dev->q_rx[qid];
437 	int i;
438 
439 	for (i = 0; i < q->ndesc; i++)
440 		q->desc[i].ctrl &= ~cpu_to_le32(MT_DMA_CTL_DMA_DONE);
441 
442 	mt76_dma_rx_cleanup(dev, q);
443 	mt76_dma_sync_idx(dev, q);
444 	mt76_dma_rx_fill(dev, q);
445 }
446 
447 static void
448 mt76_add_fragment(struct mt76_dev *dev, struct mt76_queue *q, void *data,
449 		  int len, bool more)
450 {
451 	struct page *page = virt_to_head_page(data);
452 	int offset = data - page_address(page);
453 	struct sk_buff *skb = q->rx_head;
454 
455 	offset += q->buf_offset;
456 	skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset, len,
457 			q->buf_size);
458 
459 	if (more)
460 		return;
461 
462 	q->rx_head = NULL;
463 	dev->drv->rx_skb(dev, q - dev->q_rx, skb);
464 }
465 
466 static int
467 mt76_dma_rx_process(struct mt76_dev *dev, struct mt76_queue *q, int budget)
468 {
469 	int len, data_len, done = 0;
470 	struct sk_buff *skb;
471 	unsigned char *data;
472 	bool more;
473 
474 	while (done < budget) {
475 		u32 info;
476 
477 		data = mt76_dma_dequeue(dev, q, false, &len, &info, &more);
478 		if (!data)
479 			break;
480 
481 		if (q->rx_head)
482 			data_len = q->buf_size;
483 		else
484 			data_len = SKB_WITH_OVERHEAD(q->buf_size);
485 
486 		if (data_len < len + q->buf_offset) {
487 			dev_kfree_skb(q->rx_head);
488 			q->rx_head = NULL;
489 
490 			skb_free_frag(data);
491 			continue;
492 		}
493 
494 		if (q->rx_head) {
495 			mt76_add_fragment(dev, q, data, len, more);
496 			continue;
497 		}
498 
499 		skb = build_skb(data, q->buf_size);
500 		if (!skb) {
501 			skb_free_frag(data);
502 			continue;
503 		}
504 		skb_reserve(skb, q->buf_offset);
505 
506 		if (q == &dev->q_rx[MT_RXQ_MCU]) {
507 			u32 *rxfce = (u32 *) skb->cb;
508 			*rxfce = info;
509 		}
510 
511 		__skb_put(skb, len);
512 		done++;
513 
514 		if (more) {
515 			q->rx_head = skb;
516 			continue;
517 		}
518 
519 		dev->drv->rx_skb(dev, q - dev->q_rx, skb);
520 	}
521 
522 	mt76_dma_rx_fill(dev, q);
523 	return done;
524 }
525 
526 static int
527 mt76_dma_rx_poll(struct napi_struct *napi, int budget)
528 {
529 	struct mt76_dev *dev;
530 	int qid, done = 0, cur;
531 
532 	dev = container_of(napi->dev, struct mt76_dev, napi_dev);
533 	qid = napi - dev->napi;
534 
535 	rcu_read_lock();
536 
537 	do {
538 		cur = mt76_dma_rx_process(dev, &dev->q_rx[qid], budget - done);
539 		mt76_rx_poll_complete(dev, qid, napi);
540 		done += cur;
541 	} while (cur && done < budget);
542 
543 	rcu_read_unlock();
544 
545 	if (done < budget) {
546 		napi_complete(napi);
547 		dev->drv->rx_poll_complete(dev, qid);
548 	}
549 
550 	return done;
551 }
552 
553 static int
554 mt76_dma_init(struct mt76_dev *dev)
555 {
556 	int i;
557 
558 	init_dummy_netdev(&dev->napi_dev);
559 
560 	for (i = 0; i < ARRAY_SIZE(dev->q_rx); i++) {
561 		netif_napi_add(&dev->napi_dev, &dev->napi[i], mt76_dma_rx_poll,
562 			       64);
563 		mt76_dma_rx_fill(dev, &dev->q_rx[i]);
564 		skb_queue_head_init(&dev->rx_skb[i]);
565 		napi_enable(&dev->napi[i]);
566 	}
567 
568 	return 0;
569 }
570 
571 static const struct mt76_queue_ops mt76_dma_ops = {
572 	.init = mt76_dma_init,
573 	.alloc = mt76_dma_alloc_queue,
574 	.tx_queue_skb_raw = mt76_dma_tx_queue_skb_raw,
575 	.tx_queue_skb = mt76_dma_tx_queue_skb,
576 	.tx_cleanup = mt76_dma_tx_cleanup,
577 	.rx_reset = mt76_dma_rx_reset,
578 	.kick = mt76_dma_kick_queue,
579 };
580 
581 void mt76_dma_attach(struct mt76_dev *dev)
582 {
583 	dev->queue_ops = &mt76_dma_ops;
584 }
585 EXPORT_SYMBOL_GPL(mt76_dma_attach);
586 
587 void mt76_dma_cleanup(struct mt76_dev *dev)
588 {
589 	int i;
590 
591 	netif_napi_del(&dev->tx_napi);
592 	for (i = 0; i < ARRAY_SIZE(dev->q_tx); i++)
593 		mt76_dma_tx_cleanup(dev, i, true);
594 
595 	for (i = 0; i < ARRAY_SIZE(dev->q_rx); i++) {
596 		netif_napi_del(&dev->napi[i]);
597 		mt76_dma_rx_cleanup(dev, &dev->q_rx[i]);
598 	}
599 }
600 EXPORT_SYMBOL_GPL(mt76_dma_cleanup);
601