xref: /linux/drivers/dma/sun6i-dma.c (revision edbafe65eef2b58625db1e113fbbfb1fe10c0291)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * Copyright (C) 2013-2014 Allwinner Tech Co., Ltd
4  * Author: Sugar <shuge@allwinnertech.com>
5  *
6  * Copyright (C) 2014 Maxime Ripard
7  * Maxime Ripard <maxime.ripard@free-electrons.com>
8  */
9 
10 #include <linux/clk.h>
11 #include <linux/delay.h>
12 #include <linux/dma-mapping.h>
13 #include <linux/dmaengine.h>
14 #include <linux/dmapool.h>
15 #include <linux/interrupt.h>
16 #include <linux/module.h>
17 #include <linux/of.h>
18 #include <linux/of_dma.h>
19 #include <linux/platform_device.h>
20 #include <linux/reset.h>
21 #include <linux/slab.h>
22 #include <linux/string_choices.h>
23 #include <linux/types.h>
24 
25 #include "virt-dma.h"
26 
27 /*
28  * Common registers
29  */
30 #define DMA_IRQ_EN(x)		((x) * 0x04)
31 #define DMA_IRQ_HALF			BIT(0)
32 #define DMA_IRQ_PKG			BIT(1)
33 #define DMA_IRQ_QUEUE			BIT(2)
34 
35 #define DMA_IRQ_CHAN_NR			8
36 #define DMA_IRQ_CHAN_WIDTH		4
37 
38 
39 #define DMA_IRQ_STAT(x)		((x) * 0x04 + 0x10)
40 
41 #define DMA_STAT		0x30
42 
43 /* Offset between DMA_IRQ_EN and DMA_IRQ_STAT limits number of channels */
44 #define DMA_MAX_CHANNELS	(DMA_IRQ_CHAN_NR * 0x10 / 4)
45 
46 /*
47  * sun8i specific registers
48  */
49 #define SUN8I_DMA_GATE		0x20
50 #define SUN8I_DMA_GATE_ENABLE	0x4
51 
52 #define SUNXI_H3_SECURE_REG		0x20
53 #define SUNXI_H3_DMA_GATE		0x28
54 #define SUNXI_H3_DMA_GATE_ENABLE	0x4
55 /*
56  * Channels specific registers
57  */
58 #define DMA_CHAN_ENABLE		0x00
59 #define DMA_CHAN_ENABLE_START		BIT(0)
60 #define DMA_CHAN_ENABLE_STOP		0
61 
62 #define DMA_CHAN_PAUSE		0x04
63 #define DMA_CHAN_PAUSE_PAUSE		BIT(1)
64 #define DMA_CHAN_PAUSE_RESUME		0
65 
66 #define DMA_CHAN_LLI_ADDR	0x08
67 
68 #define DMA_CHAN_CUR_CFG	0x0c
69 #define DMA_CHAN_MAX_DRQ_A31		0x1f
70 #define DMA_CHAN_MAX_DRQ_H6		0x3f
71 #define DMA_CHAN_CFG_SRC_DRQ_A31(x)	((x) & DMA_CHAN_MAX_DRQ_A31)
72 #define DMA_CHAN_CFG_SRC_DRQ_H6(x)	((x) & DMA_CHAN_MAX_DRQ_H6)
73 #define DMA_CHAN_CFG_SRC_MODE_A31(x)	(((x) & 0x1) << 5)
74 #define DMA_CHAN_CFG_SRC_MODE_H6(x)	(((x) & 0x1) << 8)
75 #define DMA_CHAN_CFG_SRC_BURST_A31(x)	(((x) & 0x3) << 7)
76 #define DMA_CHAN_CFG_SRC_BURST_H3(x)	(((x) & 0x3) << 6)
77 #define DMA_CHAN_CFG_SRC_WIDTH(x)	(((x) & 0x3) << 9)
78 
79 #define DMA_CHAN_CFG_DST_DRQ_A31(x)	(DMA_CHAN_CFG_SRC_DRQ_A31(x) << 16)
80 #define DMA_CHAN_CFG_DST_DRQ_H6(x)	(DMA_CHAN_CFG_SRC_DRQ_H6(x) << 16)
81 #define DMA_CHAN_CFG_DST_MODE_A31(x)	(DMA_CHAN_CFG_SRC_MODE_A31(x) << 16)
82 #define DMA_CHAN_CFG_DST_MODE_H6(x)	(DMA_CHAN_CFG_SRC_MODE_H6(x) << 16)
83 #define DMA_CHAN_CFG_DST_BURST_A31(x)	(DMA_CHAN_CFG_SRC_BURST_A31(x) << 16)
84 #define DMA_CHAN_CFG_DST_BURST_H3(x)	(DMA_CHAN_CFG_SRC_BURST_H3(x) << 16)
85 #define DMA_CHAN_CFG_DST_WIDTH(x)	(DMA_CHAN_CFG_SRC_WIDTH(x) << 16)
86 
87 #define DMA_CHAN_CUR_SRC	0x10
88 
89 #define DMA_CHAN_CUR_DST	0x14
90 
91 #define DMA_CHAN_CUR_CNT	0x18
92 
93 #define DMA_CHAN_CUR_PARA	0x1c
94 
95 /*
96  * LLI address mangling
97  *
98  * The LLI link physical address is also mangled, but we avoid dealing
99  * with that by allocating LLIs from the DMA32 zone.
100  */
101 #define SRC_HIGH_ADDR(x)		(((x) & 0x3U) << 16)
102 #define DST_HIGH_ADDR(x)		(((x) & 0x3U) << 18)
103 
104 /*
105  * Various hardware related defines
106  */
107 #define LLI_LAST_ITEM	0xfffff800
108 #define NORMAL_WAIT	8
109 #define DRQ_SDRAM	1
110 #define LINEAR_MODE     0
111 #define IO_MODE         1
112 
113 /* forward declaration */
114 struct sun6i_dma_dev;
115 
116 /*
117  * Hardware channels / ports representation
118  *
119  * The hardware is used in several SoCs, with differing numbers
120  * of channels and endpoints. This structure ties those numbers
121  * to a certain compatible string.
122  */
123 struct sun6i_dma_config {
124 	u32 nr_max_channels;
125 	u32 nr_max_requests;
126 	u32 nr_max_vchans;
127 	/*
128 	 * In the datasheets/user manuals of newer Allwinner SoCs, a special
129 	 * bit (bit 2 at register 0x20) is present.
130 	 * It's named "DMA MCLK interface circuit auto gating bit" in the
131 	 * documents, and the footnote of this register says that this bit
132 	 * should be set up when initializing the DMA controller.
133 	 * Allwinner A23/A33 user manuals do not have this bit documented,
134 	 * however these SoCs really have and need this bit, as seen in the
135 	 * BSP kernel source code.
136 	 */
137 	void (*clock_autogate_enable)(struct sun6i_dma_dev *);
138 	void (*set_burst_length)(u32 *p_cfg, s8 src_burst, s8 dst_burst);
139 	void (*set_drq)(u32 *p_cfg, s8 src_drq, s8 dst_drq);
140 	void (*set_mode)(u32 *p_cfg, s8 src_mode, s8 dst_mode);
141 	u32 src_burst_lengths;
142 	u32 dst_burst_lengths;
143 	u32 src_addr_widths;
144 	u32 dst_addr_widths;
145 	bool has_high_addr;
146 	bool has_mbus_clk;
147 };
148 
149 /*
150  * Hardware representation of the LLI
151  *
152  * The hardware will be fed the physical address of this structure,
153  * and read its content in order to start the transfer.
154  */
155 struct sun6i_dma_lli {
156 	u32			cfg;
157 	u32			src;
158 	u32			dst;
159 	u32			len;
160 	u32			para;
161 	u32			p_lli_next;
162 
163 	/*
164 	 * This field is not used by the DMA controller, but will be
165 	 * used by the CPU to go through the list (mostly for dumping
166 	 * or freeing it).
167 	 */
168 	struct sun6i_dma_lli	*v_lli_next;
169 };
170 
171 
172 struct sun6i_desc {
173 	struct virt_dma_desc	vd;
174 	dma_addr_t		p_lli;
175 	struct sun6i_dma_lli	*v_lli;
176 };
177 
178 struct sun6i_pchan {
179 	u32			idx;
180 	void __iomem		*base;
181 	struct sun6i_vchan	*vchan;
182 	struct sun6i_desc	*desc;
183 	struct sun6i_desc	*done;
184 };
185 
186 struct sun6i_vchan {
187 	struct virt_dma_chan	vc;
188 	struct list_head	node;
189 	struct dma_slave_config	cfg;
190 	struct sun6i_pchan	*phy;
191 	u8			port;
192 	u8			irq_type;
193 	bool			cyclic;
194 };
195 
196 struct sun6i_dma_dev {
197 	struct dma_device	slave;
198 	void __iomem		*base;
199 	struct clk		*clk;
200 	struct clk		*clk_mbus;
201 	int			irq;
202 	spinlock_t		lock;
203 	struct reset_control	*rstc;
204 	struct tasklet_struct	task;
205 	atomic_t		tasklet_shutdown;
206 	struct list_head	pending;
207 	struct dma_pool		*pool;
208 	struct sun6i_pchan	*pchans;
209 	struct sun6i_vchan	*vchans;
210 	const struct sun6i_dma_config *cfg;
211 	u32			num_pchans;
212 	u32			num_vchans;
213 	u32			max_request;
214 };
215 
216 static struct device *chan2dev(struct dma_chan *chan)
217 {
218 	return &chan->dev->device;
219 }
220 
221 static inline struct sun6i_dma_dev *to_sun6i_dma_dev(struct dma_device *d)
222 {
223 	return container_of(d, struct sun6i_dma_dev, slave);
224 }
225 
226 static inline struct sun6i_vchan *to_sun6i_vchan(struct dma_chan *chan)
227 {
228 	return container_of(chan, struct sun6i_vchan, vc.chan);
229 }
230 
231 static inline struct sun6i_desc *
232 to_sun6i_desc(struct dma_async_tx_descriptor *tx)
233 {
234 	return container_of(tx, struct sun6i_desc, vd.tx);
235 }
236 
237 static inline void sun6i_dma_dump_com_regs(struct sun6i_dma_dev *sdev)
238 {
239 	dev_dbg(sdev->slave.dev, "Common register:\n"
240 		"\tmask0(%04x): 0x%08x\n"
241 		"\tmask1(%04x): 0x%08x\n"
242 		"\tpend0(%04x): 0x%08x\n"
243 		"\tpend1(%04x): 0x%08x\n"
244 		"\tstats(%04x): 0x%08x\n",
245 		DMA_IRQ_EN(0), readl(sdev->base + DMA_IRQ_EN(0)),
246 		DMA_IRQ_EN(1), readl(sdev->base + DMA_IRQ_EN(1)),
247 		DMA_IRQ_STAT(0), readl(sdev->base + DMA_IRQ_STAT(0)),
248 		DMA_IRQ_STAT(1), readl(sdev->base + DMA_IRQ_STAT(1)),
249 		DMA_STAT, readl(sdev->base + DMA_STAT));
250 }
251 
252 static inline void sun6i_dma_dump_chan_regs(struct sun6i_dma_dev *sdev,
253 					    struct sun6i_pchan *pchan)
254 {
255 	dev_dbg(sdev->slave.dev, "Chan %d reg:\n"
256 		"\t___en(%04x): \t0x%08x\n"
257 		"\tpause(%04x): \t0x%08x\n"
258 		"\tstart(%04x): \t0x%08x\n"
259 		"\t__cfg(%04x): \t0x%08x\n"
260 		"\t__src(%04x): \t0x%08x\n"
261 		"\t__dst(%04x): \t0x%08x\n"
262 		"\tcount(%04x): \t0x%08x\n"
263 		"\t_para(%04x): \t0x%08x\n\n",
264 		pchan->idx,
265 		DMA_CHAN_ENABLE,
266 		readl(pchan->base + DMA_CHAN_ENABLE),
267 		DMA_CHAN_PAUSE,
268 		readl(pchan->base + DMA_CHAN_PAUSE),
269 		DMA_CHAN_LLI_ADDR,
270 		readl(pchan->base + DMA_CHAN_LLI_ADDR),
271 		DMA_CHAN_CUR_CFG,
272 		readl(pchan->base + DMA_CHAN_CUR_CFG),
273 		DMA_CHAN_CUR_SRC,
274 		readl(pchan->base + DMA_CHAN_CUR_SRC),
275 		DMA_CHAN_CUR_DST,
276 		readl(pchan->base + DMA_CHAN_CUR_DST),
277 		DMA_CHAN_CUR_CNT,
278 		readl(pchan->base + DMA_CHAN_CUR_CNT),
279 		DMA_CHAN_CUR_PARA,
280 		readl(pchan->base + DMA_CHAN_CUR_PARA));
281 }
282 
283 static inline s8 convert_burst(u32 maxburst)
284 {
285 	switch (maxburst) {
286 	case 1:
287 		return 0;
288 	case 4:
289 		return 1;
290 	case 8:
291 		return 2;
292 	case 16:
293 		return 3;
294 	default:
295 		return -EINVAL;
296 	}
297 }
298 
299 static inline s8 convert_buswidth(enum dma_slave_buswidth addr_width)
300 {
301 	return ilog2(addr_width);
302 }
303 
304 static void sun6i_enable_clock_autogate_a23(struct sun6i_dma_dev *sdev)
305 {
306 	writel(SUN8I_DMA_GATE_ENABLE, sdev->base + SUN8I_DMA_GATE);
307 }
308 
309 static void sun6i_enable_clock_autogate_h3(struct sun6i_dma_dev *sdev)
310 {
311 	writel(SUNXI_H3_DMA_GATE_ENABLE, sdev->base + SUNXI_H3_DMA_GATE);
312 }
313 
314 static void sun6i_set_burst_length_a31(u32 *p_cfg, s8 src_burst, s8 dst_burst)
315 {
316 	*p_cfg |= DMA_CHAN_CFG_SRC_BURST_A31(src_burst) |
317 		  DMA_CHAN_CFG_DST_BURST_A31(dst_burst);
318 }
319 
320 static void sun6i_set_burst_length_h3(u32 *p_cfg, s8 src_burst, s8 dst_burst)
321 {
322 	*p_cfg |= DMA_CHAN_CFG_SRC_BURST_H3(src_burst) |
323 		  DMA_CHAN_CFG_DST_BURST_H3(dst_burst);
324 }
325 
326 static void sun6i_set_drq_a31(u32 *p_cfg, s8 src_drq, s8 dst_drq)
327 {
328 	*p_cfg |= DMA_CHAN_CFG_SRC_DRQ_A31(src_drq) |
329 		  DMA_CHAN_CFG_DST_DRQ_A31(dst_drq);
330 }
331 
332 static void sun6i_set_drq_h6(u32 *p_cfg, s8 src_drq, s8 dst_drq)
333 {
334 	*p_cfg |= DMA_CHAN_CFG_SRC_DRQ_H6(src_drq) |
335 		  DMA_CHAN_CFG_DST_DRQ_H6(dst_drq);
336 }
337 
338 static void sun6i_set_mode_a31(u32 *p_cfg, s8 src_mode, s8 dst_mode)
339 {
340 	*p_cfg |= DMA_CHAN_CFG_SRC_MODE_A31(src_mode) |
341 		  DMA_CHAN_CFG_DST_MODE_A31(dst_mode);
342 }
343 
344 static void sun6i_set_mode_h6(u32 *p_cfg, s8 src_mode, s8 dst_mode)
345 {
346 	*p_cfg |= DMA_CHAN_CFG_SRC_MODE_H6(src_mode) |
347 		  DMA_CHAN_CFG_DST_MODE_H6(dst_mode);
348 }
349 
350 static size_t sun6i_get_chan_size(struct sun6i_pchan *pchan)
351 {
352 	struct sun6i_desc *txd = pchan->desc;
353 	struct sun6i_dma_lli *lli;
354 	size_t bytes;
355 	dma_addr_t pos;
356 
357 	pos = readl(pchan->base + DMA_CHAN_LLI_ADDR);
358 	bytes = readl(pchan->base + DMA_CHAN_CUR_CNT);
359 
360 	if (pos == LLI_LAST_ITEM)
361 		return bytes;
362 
363 	for (lli = txd->v_lli; lli; lli = lli->v_lli_next) {
364 		if (lli->p_lli_next == pos) {
365 			for (lli = lli->v_lli_next; lli; lli = lli->v_lli_next)
366 				bytes += lli->len;
367 			break;
368 		}
369 	}
370 
371 	return bytes;
372 }
373 
374 static void *sun6i_dma_lli_add(struct sun6i_dma_lli *prev,
375 			       struct sun6i_dma_lli *next,
376 			       dma_addr_t next_phy,
377 			       struct sun6i_desc *txd)
378 {
379 	if ((!prev && !txd) || !next)
380 		return NULL;
381 
382 	if (!prev) {
383 		txd->p_lli = next_phy;
384 		txd->v_lli = next;
385 	} else {
386 		prev->p_lli_next = next_phy;
387 		prev->v_lli_next = next;
388 	}
389 
390 	next->p_lli_next = LLI_LAST_ITEM;
391 	next->v_lli_next = NULL;
392 
393 	return next;
394 }
395 
396 static inline void sun6i_dma_dump_lli(struct sun6i_vchan *vchan,
397 				      struct sun6i_dma_lli *v_lli,
398 				      dma_addr_t p_lli)
399 {
400 	dev_dbg(chan2dev(&vchan->vc.chan),
401 		"\n\tdesc:\tp - %pad v - 0x%p\n"
402 		"\t\tc - 0x%08x s - 0x%08x d - 0x%08x\n"
403 		"\t\tl - 0x%08x p - 0x%08x n - 0x%08x\n",
404 		&p_lli, v_lli,
405 		v_lli->cfg, v_lli->src, v_lli->dst,
406 		v_lli->len, v_lli->para, v_lli->p_lli_next);
407 }
408 
409 static void sun6i_dma_free_desc(struct virt_dma_desc *vd)
410 {
411 	struct sun6i_desc *txd = to_sun6i_desc(&vd->tx);
412 	struct sun6i_dma_dev *sdev = to_sun6i_dma_dev(vd->tx.chan->device);
413 	struct sun6i_dma_lli *v_lli, *v_next;
414 	dma_addr_t p_lli, p_next;
415 
416 	if (unlikely(!txd))
417 		return;
418 
419 	p_lli = txd->p_lli;
420 	v_lli = txd->v_lli;
421 
422 	while (v_lli) {
423 		v_next = v_lli->v_lli_next;
424 		p_next = v_lli->p_lli_next;
425 
426 		dma_pool_free(sdev->pool, v_lli, p_lli);
427 
428 		v_lli = v_next;
429 		p_lli = p_next;
430 	}
431 
432 	kfree(txd);
433 }
434 
435 static int sun6i_dma_start_desc(struct sun6i_vchan *vchan)
436 {
437 	struct sun6i_dma_dev *sdev = to_sun6i_dma_dev(vchan->vc.chan.device);
438 	struct virt_dma_desc *desc = vchan_next_desc(&vchan->vc);
439 	struct sun6i_pchan *pchan = vchan->phy;
440 	u32 irq_val, irq_reg, irq_offset;
441 
442 	if (!pchan)
443 		return -EAGAIN;
444 
445 	if (!desc) {
446 		pchan->desc = NULL;
447 		pchan->done = NULL;
448 		return -EAGAIN;
449 	}
450 
451 	list_del(&desc->node);
452 
453 	pchan->desc = to_sun6i_desc(&desc->tx);
454 	pchan->done = NULL;
455 
456 	sun6i_dma_dump_lli(vchan, pchan->desc->v_lli, pchan->desc->p_lli);
457 
458 	irq_reg = pchan->idx / DMA_IRQ_CHAN_NR;
459 	irq_offset = pchan->idx % DMA_IRQ_CHAN_NR;
460 
461 	vchan->irq_type = vchan->cyclic ? DMA_IRQ_PKG : DMA_IRQ_QUEUE;
462 
463 	irq_val = readl(sdev->base + DMA_IRQ_EN(irq_reg));
464 	irq_val &= ~((DMA_IRQ_HALF | DMA_IRQ_PKG | DMA_IRQ_QUEUE) <<
465 			(irq_offset * DMA_IRQ_CHAN_WIDTH));
466 	irq_val |= vchan->irq_type << (irq_offset * DMA_IRQ_CHAN_WIDTH);
467 	writel(irq_val, sdev->base + DMA_IRQ_EN(irq_reg));
468 
469 	writel(pchan->desc->p_lli, pchan->base + DMA_CHAN_LLI_ADDR);
470 	writel(DMA_CHAN_ENABLE_START, pchan->base + DMA_CHAN_ENABLE);
471 
472 	sun6i_dma_dump_com_regs(sdev);
473 	sun6i_dma_dump_chan_regs(sdev, pchan);
474 
475 	return 0;
476 }
477 
478 static void sun6i_dma_tasklet(struct tasklet_struct *t)
479 {
480 	struct sun6i_dma_dev *sdev = from_tasklet(sdev, t, task);
481 	struct sun6i_vchan *vchan;
482 	struct sun6i_pchan *pchan;
483 	unsigned int pchan_alloc = 0;
484 	unsigned int pchan_idx;
485 
486 	list_for_each_entry(vchan, &sdev->slave.channels, vc.chan.device_node) {
487 		spin_lock_irq(&vchan->vc.lock);
488 
489 		pchan = vchan->phy;
490 
491 		if (pchan && pchan->done) {
492 			if (sun6i_dma_start_desc(vchan)) {
493 				/*
494 				 * No current txd associated with this channel
495 				 */
496 				dev_dbg(sdev->slave.dev, "pchan %u: free\n",
497 					pchan->idx);
498 
499 				/* Mark this channel free */
500 				vchan->phy = NULL;
501 				pchan->vchan = NULL;
502 			}
503 		}
504 		spin_unlock_irq(&vchan->vc.lock);
505 	}
506 
507 	spin_lock_irq(&sdev->lock);
508 	for (pchan_idx = 0; pchan_idx < sdev->num_pchans; pchan_idx++) {
509 		pchan = &sdev->pchans[pchan_idx];
510 
511 		if (pchan->vchan || list_empty(&sdev->pending))
512 			continue;
513 
514 		vchan = list_first_entry(&sdev->pending,
515 					 struct sun6i_vchan, node);
516 
517 		/* Remove from pending channels */
518 		list_del_init(&vchan->node);
519 		pchan_alloc |= BIT(pchan_idx);
520 
521 		/* Mark this channel allocated */
522 		pchan->vchan = vchan;
523 		vchan->phy = pchan;
524 		dev_dbg(sdev->slave.dev, "pchan %u: alloc vchan %p\n",
525 			pchan->idx, &vchan->vc);
526 	}
527 	spin_unlock_irq(&sdev->lock);
528 
529 	for (pchan_idx = 0; pchan_idx < sdev->num_pchans; pchan_idx++) {
530 		if (!(pchan_alloc & BIT(pchan_idx)))
531 			continue;
532 
533 		pchan = sdev->pchans + pchan_idx;
534 		vchan = pchan->vchan;
535 		if (vchan) {
536 			spin_lock_irq(&vchan->vc.lock);
537 			sun6i_dma_start_desc(vchan);
538 			spin_unlock_irq(&vchan->vc.lock);
539 		}
540 	}
541 }
542 
543 static irqreturn_t sun6i_dma_interrupt(int irq, void *dev_id)
544 {
545 	struct sun6i_dma_dev *sdev = dev_id;
546 	struct sun6i_vchan *vchan;
547 	struct sun6i_pchan *pchan;
548 	int i, j, ret = IRQ_NONE;
549 	u32 status;
550 
551 	for (i = 0; i < sdev->num_pchans / DMA_IRQ_CHAN_NR; i++) {
552 		status = readl(sdev->base + DMA_IRQ_STAT(i));
553 		if (!status)
554 			continue;
555 
556 		dev_dbg(sdev->slave.dev, "DMA irq status %s: 0x%x\n",
557 			str_high_low(i), status);
558 
559 		writel(status, sdev->base + DMA_IRQ_STAT(i));
560 
561 		for (j = 0; (j < DMA_IRQ_CHAN_NR) && status; j++) {
562 			pchan = sdev->pchans + j;
563 			vchan = pchan->vchan;
564 			if (vchan && (status & vchan->irq_type)) {
565 				if (vchan->cyclic) {
566 					vchan_cyclic_callback(&pchan->desc->vd);
567 				} else {
568 					spin_lock(&vchan->vc.lock);
569 					vchan_cookie_complete(&pchan->desc->vd);
570 					pchan->done = pchan->desc;
571 					spin_unlock(&vchan->vc.lock);
572 				}
573 			}
574 
575 			status = status >> DMA_IRQ_CHAN_WIDTH;
576 		}
577 
578 		if (!atomic_read(&sdev->tasklet_shutdown))
579 			tasklet_schedule(&sdev->task);
580 		ret = IRQ_HANDLED;
581 	}
582 
583 	return ret;
584 }
585 
586 static u32 find_burst_size(const u32 burst_lengths, u32 maxburst)
587 {
588 	if (!maxburst)
589 		return 1;
590 
591 	if (BIT(maxburst) & burst_lengths)
592 		return maxburst;
593 
594 	/* Hardware only does power-of-two bursts. */
595 	for (u32 burst = rounddown_pow_of_two(maxburst); burst > 0; burst /= 2)
596 		if (BIT(burst) & burst_lengths)
597 			return burst;
598 
599 	return 1;
600 }
601 
602 static int set_config(struct sun6i_dma_dev *sdev,
603 			struct dma_slave_config *sconfig,
604 			enum dma_transfer_direction direction,
605 			u32 *p_cfg)
606 {
607 	enum dma_slave_buswidth src_addr_width, dst_addr_width;
608 	u32 src_maxburst, dst_maxburst;
609 	s8 src_width, dst_width, src_burst, dst_burst;
610 
611 	src_addr_width = sconfig->src_addr_width;
612 	dst_addr_width = sconfig->dst_addr_width;
613 	src_maxburst = sconfig->src_maxburst;
614 	dst_maxburst = sconfig->dst_maxburst;
615 
616 	switch (direction) {
617 	case DMA_MEM_TO_DEV:
618 		if (src_addr_width == DMA_SLAVE_BUSWIDTH_UNDEFINED)
619 			src_addr_width = DMA_SLAVE_BUSWIDTH_4_BYTES;
620 		src_maxburst = src_maxburst ? src_maxburst : 8;
621 		break;
622 	case DMA_DEV_TO_MEM:
623 		if (dst_addr_width == DMA_SLAVE_BUSWIDTH_UNDEFINED)
624 			dst_addr_width = DMA_SLAVE_BUSWIDTH_4_BYTES;
625 		dst_maxburst = dst_maxburst ? dst_maxburst : 8;
626 		break;
627 	default:
628 		return -EINVAL;
629 	}
630 
631 	if (!(BIT(src_addr_width) & sdev->slave.src_addr_widths))
632 		return -EINVAL;
633 	if (!(BIT(dst_addr_width) & sdev->slave.dst_addr_widths))
634 		return -EINVAL;
635 
636 	src_width = convert_buswidth(src_addr_width);
637 	dst_width = convert_buswidth(dst_addr_width);
638 	src_burst = find_burst_size(sdev->cfg->src_burst_lengths, src_maxburst);
639 	dst_burst = find_burst_size(sdev->cfg->dst_burst_lengths, dst_maxburst);
640 	dst_burst = convert_burst(dst_burst);
641 	src_burst = convert_burst(src_burst);
642 
643 	*p_cfg = DMA_CHAN_CFG_SRC_WIDTH(src_width) |
644 		DMA_CHAN_CFG_DST_WIDTH(dst_width);
645 
646 	sdev->cfg->set_burst_length(p_cfg, src_burst, dst_burst);
647 
648 	return 0;
649 }
650 
651 static inline void sun6i_dma_set_addr(struct sun6i_dma_dev *sdev,
652 				      struct sun6i_dma_lli *v_lli,
653 				      dma_addr_t src, dma_addr_t dst)
654 {
655 	v_lli->src = lower_32_bits(src);
656 	v_lli->dst = lower_32_bits(dst);
657 
658 	if (sdev->cfg->has_high_addr)
659 		v_lli->para |= SRC_HIGH_ADDR(upper_32_bits(src)) |
660 			       DST_HIGH_ADDR(upper_32_bits(dst));
661 }
662 
663 static struct dma_async_tx_descriptor *sun6i_dma_prep_dma_memcpy(
664 		struct dma_chan *chan, dma_addr_t dest, dma_addr_t src,
665 		size_t len, unsigned long flags)
666 {
667 	struct sun6i_dma_dev *sdev = to_sun6i_dma_dev(chan->device);
668 	struct sun6i_vchan *vchan = to_sun6i_vchan(chan);
669 	struct sun6i_dma_lli *v_lli;
670 	struct sun6i_desc *txd;
671 	dma_addr_t p_lli;
672 	s8 burst, width;
673 
674 	dev_dbg(chan2dev(chan),
675 		"%s; chan: %d, dest: %pad, src: %pad, len: %zu. flags: 0x%08lx\n",
676 		__func__, vchan->vc.chan.chan_id, &dest, &src, len, flags);
677 
678 	if (!len)
679 		return NULL;
680 
681 	txd = kzalloc_obj(*txd, GFP_NOWAIT);
682 	if (!txd)
683 		return NULL;
684 
685 	v_lli = dma_pool_alloc(sdev->pool, GFP_DMA32 | GFP_NOWAIT, &p_lli);
686 	if (!v_lli) {
687 		dev_err(sdev->slave.dev, "Failed to alloc lli memory\n");
688 		goto err_txd_free;
689 	}
690 
691 	v_lli->len = len;
692 	v_lli->para = NORMAL_WAIT;
693 	sun6i_dma_set_addr(sdev, v_lli, src, dest);
694 
695 	burst = convert_burst(8);
696 	width = convert_buswidth(DMA_SLAVE_BUSWIDTH_4_BYTES);
697 	v_lli->cfg = DMA_CHAN_CFG_SRC_WIDTH(width) |
698 		DMA_CHAN_CFG_DST_WIDTH(width);
699 
700 	sdev->cfg->set_burst_length(&v_lli->cfg, burst, burst);
701 	sdev->cfg->set_drq(&v_lli->cfg, DRQ_SDRAM, DRQ_SDRAM);
702 	sdev->cfg->set_mode(&v_lli->cfg, LINEAR_MODE, LINEAR_MODE);
703 
704 	sun6i_dma_lli_add(NULL, v_lli, p_lli, txd);
705 
706 	sun6i_dma_dump_lli(vchan, v_lli, p_lli);
707 
708 	return vchan_tx_prep(&vchan->vc, &txd->vd, flags);
709 
710 err_txd_free:
711 	kfree(txd);
712 	return NULL;
713 }
714 
715 static struct dma_async_tx_descriptor *sun6i_dma_prep_slave_sg(
716 		struct dma_chan *chan, struct scatterlist *sgl,
717 		unsigned int sg_len, enum dma_transfer_direction dir,
718 		unsigned long flags, void *context)
719 {
720 	struct sun6i_dma_dev *sdev = to_sun6i_dma_dev(chan->device);
721 	struct sun6i_vchan *vchan = to_sun6i_vchan(chan);
722 	struct dma_slave_config *sconfig = &vchan->cfg;
723 	struct sun6i_dma_lli *v_lli, *prev = NULL;
724 	struct sun6i_desc *txd;
725 	struct scatterlist *sg;
726 	dma_addr_t p_lli;
727 	u32 lli_cfg;
728 	int i, ret;
729 
730 	if (!sgl)
731 		return NULL;
732 
733 	ret = set_config(sdev, sconfig, dir, &lli_cfg);
734 	if (ret) {
735 		dev_err(chan2dev(chan), "Invalid DMA configuration\n");
736 		return NULL;
737 	}
738 
739 	txd = kzalloc_obj(*txd, GFP_NOWAIT);
740 	if (!txd)
741 		return NULL;
742 
743 	for_each_sg(sgl, sg, sg_len, i) {
744 		v_lli = dma_pool_alloc(sdev->pool, GFP_DMA32 | GFP_NOWAIT, &p_lli);
745 		if (!v_lli)
746 			goto err_lli_free;
747 
748 		v_lli->len = sg_dma_len(sg);
749 		v_lli->para = NORMAL_WAIT;
750 
751 		if (dir == DMA_MEM_TO_DEV) {
752 			sun6i_dma_set_addr(sdev, v_lli,
753 					   sg_dma_address(sg),
754 					   sconfig->dst_addr);
755 			v_lli->cfg = lli_cfg;
756 			sdev->cfg->set_drq(&v_lli->cfg, DRQ_SDRAM, vchan->port);
757 			sdev->cfg->set_mode(&v_lli->cfg, LINEAR_MODE, IO_MODE);
758 
759 			dev_dbg(chan2dev(chan),
760 				"%s; chan: %d, dest: %pad, src: %pad, len: %u. flags: 0x%08lx\n",
761 				__func__, vchan->vc.chan.chan_id,
762 				&sconfig->dst_addr, &sg_dma_address(sg),
763 				sg_dma_len(sg), flags);
764 
765 		} else {
766 			sun6i_dma_set_addr(sdev, v_lli,
767 					   sconfig->src_addr,
768 					   sg_dma_address(sg));
769 			v_lli->cfg = lli_cfg;
770 			sdev->cfg->set_drq(&v_lli->cfg, vchan->port, DRQ_SDRAM);
771 			sdev->cfg->set_mode(&v_lli->cfg, IO_MODE, LINEAR_MODE);
772 
773 			dev_dbg(chan2dev(chan),
774 				"%s; chan: %d, dest: %pad, src: %pad, len: %u. flags: 0x%08lx\n",
775 				__func__, vchan->vc.chan.chan_id,
776 				&sg_dma_address(sg), &sconfig->src_addr,
777 				sg_dma_len(sg), flags);
778 		}
779 
780 		prev = sun6i_dma_lli_add(prev, v_lli, p_lli, txd);
781 	}
782 
783 	dev_dbg(chan2dev(chan), "First: %pad\n", &txd->p_lli);
784 	for (p_lli = txd->p_lli, v_lli = txd->v_lli; v_lli;
785 	     p_lli = v_lli->p_lli_next, v_lli = v_lli->v_lli_next)
786 		sun6i_dma_dump_lli(vchan, v_lli, p_lli);
787 
788 	return vchan_tx_prep(&vchan->vc, &txd->vd, flags);
789 
790 err_lli_free:
791 	for (p_lli = txd->p_lli, v_lli = txd->v_lli; v_lli;
792 	     p_lli = v_lli->p_lli_next, v_lli = v_lli->v_lli_next)
793 		dma_pool_free(sdev->pool, v_lli, p_lli);
794 	kfree(txd);
795 	return NULL;
796 }
797 
798 static struct dma_async_tx_descriptor *sun6i_dma_prep_dma_cyclic(
799 					struct dma_chan *chan,
800 					dma_addr_t buf_addr,
801 					size_t buf_len,
802 					size_t period_len,
803 					enum dma_transfer_direction dir,
804 					unsigned long flags)
805 {
806 	struct sun6i_dma_dev *sdev = to_sun6i_dma_dev(chan->device);
807 	struct sun6i_vchan *vchan = to_sun6i_vchan(chan);
808 	struct dma_slave_config *sconfig = &vchan->cfg;
809 	struct sun6i_dma_lli *v_lli, *prev = NULL;
810 	struct sun6i_desc *txd;
811 	dma_addr_t p_lli;
812 	u32 lli_cfg;
813 	unsigned int i, periods = buf_len / period_len;
814 	int ret;
815 
816 	ret = set_config(sdev, sconfig, dir, &lli_cfg);
817 	if (ret) {
818 		dev_err(chan2dev(chan), "Invalid DMA configuration\n");
819 		return NULL;
820 	}
821 
822 	txd = kzalloc_obj(*txd, GFP_NOWAIT);
823 	if (!txd)
824 		return NULL;
825 
826 	for (i = 0; i < periods; i++) {
827 		v_lli = dma_pool_alloc(sdev->pool, GFP_DMA32 | GFP_NOWAIT, &p_lli);
828 		if (!v_lli) {
829 			dev_err(sdev->slave.dev, "Failed to alloc lli memory\n");
830 			goto err_lli_free;
831 		}
832 
833 		v_lli->len = period_len;
834 		v_lli->para = NORMAL_WAIT;
835 
836 		if (dir == DMA_MEM_TO_DEV) {
837 			sun6i_dma_set_addr(sdev, v_lli,
838 					   buf_addr + period_len * i,
839 					   sconfig->dst_addr);
840 			v_lli->cfg = lli_cfg;
841 			sdev->cfg->set_drq(&v_lli->cfg, DRQ_SDRAM, vchan->port);
842 			sdev->cfg->set_mode(&v_lli->cfg, LINEAR_MODE, IO_MODE);
843 			dev_dbg(chan2dev(chan),
844 				"%s; chan: %d, dest: %pad, src: %pad, len: %zu. flags: 0x%08lx\n",
845 				__func__, vchan->vc.chan.chan_id,
846 				&sconfig->dst_addr, &buf_addr,
847 				buf_len, flags);
848 		} else {
849 			sun6i_dma_set_addr(sdev, v_lli,
850 					   sconfig->src_addr,
851 					   buf_addr + period_len * i);
852 			v_lli->cfg = lli_cfg;
853 			sdev->cfg->set_drq(&v_lli->cfg, vchan->port, DRQ_SDRAM);
854 			sdev->cfg->set_mode(&v_lli->cfg, IO_MODE, LINEAR_MODE);
855 			dev_dbg(chan2dev(chan),
856 				"%s; chan: %d, dest: %pad, src: %pad, len: %zu. flags: 0x%08lx\n",
857 				__func__, vchan->vc.chan.chan_id,
858 				&buf_addr, &sconfig->src_addr,
859 				buf_len, flags);
860 		}
861 
862 		prev = sun6i_dma_lli_add(prev, v_lli, p_lli, txd);
863 	}
864 
865 	prev->p_lli_next = txd->p_lli;		/* cyclic list */
866 
867 	vchan->cyclic = true;
868 
869 	return vchan_tx_prep(&vchan->vc, &txd->vd, flags);
870 
871 err_lli_free:
872 	for (p_lli = txd->p_lli, v_lli = txd->v_lli; v_lli;
873 	     p_lli = v_lli->p_lli_next, v_lli = v_lli->v_lli_next)
874 		dma_pool_free(sdev->pool, v_lli, p_lli);
875 	kfree(txd);
876 	return NULL;
877 }
878 
879 static int sun6i_dma_config(struct dma_chan *chan,
880 			    struct dma_slave_config *config)
881 {
882 	struct sun6i_vchan *vchan = to_sun6i_vchan(chan);
883 
884 	memcpy(&vchan->cfg, config, sizeof(*config));
885 
886 	return 0;
887 }
888 
889 static int sun6i_dma_pause(struct dma_chan *chan)
890 {
891 	struct sun6i_dma_dev *sdev = to_sun6i_dma_dev(chan->device);
892 	struct sun6i_vchan *vchan = to_sun6i_vchan(chan);
893 	struct sun6i_pchan *pchan = vchan->phy;
894 
895 	dev_dbg(chan2dev(chan), "vchan %p: pause\n", &vchan->vc);
896 
897 	if (pchan) {
898 		writel(DMA_CHAN_PAUSE_PAUSE,
899 		       pchan->base + DMA_CHAN_PAUSE);
900 	} else {
901 		spin_lock(&sdev->lock);
902 		list_del_init(&vchan->node);
903 		spin_unlock(&sdev->lock);
904 	}
905 
906 	return 0;
907 }
908 
909 static int sun6i_dma_resume(struct dma_chan *chan)
910 {
911 	struct sun6i_dma_dev *sdev = to_sun6i_dma_dev(chan->device);
912 	struct sun6i_vchan *vchan = to_sun6i_vchan(chan);
913 	struct sun6i_pchan *pchan = vchan->phy;
914 	unsigned long flags;
915 
916 	dev_dbg(chan2dev(chan), "vchan %p: resume\n", &vchan->vc);
917 
918 	spin_lock_irqsave(&vchan->vc.lock, flags);
919 
920 	if (pchan) {
921 		writel(DMA_CHAN_PAUSE_RESUME,
922 		       pchan->base + DMA_CHAN_PAUSE);
923 	} else if (!list_empty(&vchan->vc.desc_issued)) {
924 		spin_lock(&sdev->lock);
925 		list_add_tail(&vchan->node, &sdev->pending);
926 		spin_unlock(&sdev->lock);
927 	}
928 
929 	spin_unlock_irqrestore(&vchan->vc.lock, flags);
930 
931 	return 0;
932 }
933 
934 static int sun6i_dma_terminate_all(struct dma_chan *chan)
935 {
936 	struct sun6i_dma_dev *sdev = to_sun6i_dma_dev(chan->device);
937 	struct sun6i_vchan *vchan = to_sun6i_vchan(chan);
938 	struct sun6i_pchan *pchan = vchan->phy;
939 	unsigned long flags;
940 	LIST_HEAD(head);
941 
942 	spin_lock(&sdev->lock);
943 	list_del_init(&vchan->node);
944 	spin_unlock(&sdev->lock);
945 
946 	spin_lock_irqsave(&vchan->vc.lock, flags);
947 
948 	if (pchan && pchan->desc && pchan->desc != pchan->done) {
949 		struct virt_dma_desc *vd = &pchan->desc->vd;
950 
951 		vchan_terminate_vdesc(vd);
952 	}
953 
954 	vchan->cyclic = false;
955 	vchan_get_all_descriptors(&vchan->vc, &head);
956 
957 	if (pchan) {
958 		writel(DMA_CHAN_ENABLE_STOP, pchan->base + DMA_CHAN_ENABLE);
959 		writel(DMA_CHAN_PAUSE_RESUME, pchan->base + DMA_CHAN_PAUSE);
960 
961 		vchan->phy = NULL;
962 		pchan->vchan = NULL;
963 		pchan->desc = NULL;
964 		pchan->done = NULL;
965 	}
966 
967 	spin_unlock_irqrestore(&vchan->vc.lock, flags);
968 
969 	vchan_dma_desc_free_list(&vchan->vc, &head);
970 
971 	return 0;
972 }
973 
974 static enum dma_status sun6i_dma_tx_status(struct dma_chan *chan,
975 					   dma_cookie_t cookie,
976 					   struct dma_tx_state *state)
977 {
978 	struct sun6i_vchan *vchan = to_sun6i_vchan(chan);
979 	struct sun6i_pchan *pchan = vchan->phy;
980 	struct sun6i_dma_lli *lli;
981 	struct virt_dma_desc *vd;
982 	struct sun6i_desc *txd;
983 	enum dma_status ret;
984 	unsigned long flags;
985 	size_t bytes = 0;
986 
987 	ret = dma_cookie_status(chan, cookie, state);
988 	if (ret == DMA_COMPLETE || !state)
989 		return ret;
990 
991 	spin_lock_irqsave(&vchan->vc.lock, flags);
992 
993 	vd = vchan_find_desc(&vchan->vc, cookie);
994 	txd = to_sun6i_desc(&vd->tx);
995 
996 	if (vd) {
997 		for (lli = txd->v_lli; lli != NULL; lli = lli->v_lli_next)
998 			bytes += lli->len;
999 	} else if (!pchan || !pchan->desc) {
1000 		bytes = 0;
1001 	} else {
1002 		bytes = sun6i_get_chan_size(pchan);
1003 	}
1004 
1005 	spin_unlock_irqrestore(&vchan->vc.lock, flags);
1006 
1007 	dma_set_residue(state, bytes);
1008 
1009 	return ret;
1010 }
1011 
1012 static void sun6i_dma_issue_pending(struct dma_chan *chan)
1013 {
1014 	struct sun6i_dma_dev *sdev = to_sun6i_dma_dev(chan->device);
1015 	struct sun6i_vchan *vchan = to_sun6i_vchan(chan);
1016 	unsigned long flags;
1017 
1018 	spin_lock_irqsave(&vchan->vc.lock, flags);
1019 
1020 	if (vchan_issue_pending(&vchan->vc)) {
1021 		spin_lock(&sdev->lock);
1022 
1023 		if (!vchan->phy && list_empty(&vchan->node)) {
1024 			list_add_tail(&vchan->node, &sdev->pending);
1025 			tasklet_schedule(&sdev->task);
1026 			dev_dbg(chan2dev(chan), "vchan %p: issued\n",
1027 				&vchan->vc);
1028 		}
1029 
1030 		spin_unlock(&sdev->lock);
1031 	} else {
1032 		dev_dbg(chan2dev(chan), "vchan %p: nothing to issue\n",
1033 			&vchan->vc);
1034 	}
1035 
1036 	spin_unlock_irqrestore(&vchan->vc.lock, flags);
1037 }
1038 
1039 static void sun6i_dma_free_chan_resources(struct dma_chan *chan)
1040 {
1041 	struct sun6i_dma_dev *sdev = to_sun6i_dma_dev(chan->device);
1042 	struct sun6i_vchan *vchan = to_sun6i_vchan(chan);
1043 	unsigned long flags;
1044 
1045 	spin_lock_irqsave(&sdev->lock, flags);
1046 	list_del_init(&vchan->node);
1047 	spin_unlock_irqrestore(&sdev->lock, flags);
1048 
1049 	vchan_free_chan_resources(&vchan->vc);
1050 }
1051 
1052 static struct dma_chan *sun6i_dma_of_xlate(struct of_phandle_args *dma_spec,
1053 					   struct of_dma *ofdma)
1054 {
1055 	struct sun6i_dma_dev *sdev = ofdma->of_dma_data;
1056 	struct sun6i_vchan *vchan;
1057 	struct dma_chan *chan;
1058 	u8 port = dma_spec->args[0];
1059 
1060 	if (port > sdev->max_request)
1061 		return NULL;
1062 
1063 	chan = dma_get_any_slave_channel(&sdev->slave);
1064 	if (!chan)
1065 		return NULL;
1066 
1067 	vchan = to_sun6i_vchan(chan);
1068 	vchan->port = port;
1069 
1070 	return chan;
1071 }
1072 
1073 static inline void sun6i_kill_tasklet(struct sun6i_dma_dev *sdev)
1074 {
1075 	/* Disable all interrupts from DMA */
1076 	writel(0, sdev->base + DMA_IRQ_EN(0));
1077 	writel(0, sdev->base + DMA_IRQ_EN(1));
1078 
1079 	/* Prevent spurious interrupts from scheduling the tasklet */
1080 	atomic_inc(&sdev->tasklet_shutdown);
1081 
1082 	/* Make sure we won't have any further interrupts */
1083 	devm_free_irq(sdev->slave.dev, sdev->irq, sdev);
1084 
1085 	/* Actually prevent the tasklet from being scheduled */
1086 	tasklet_kill(&sdev->task);
1087 }
1088 
1089 static inline void sun6i_dma_free(struct sun6i_dma_dev *sdev)
1090 {
1091 	int i;
1092 
1093 	for (i = 0; i < sdev->num_vchans; i++) {
1094 		struct sun6i_vchan *vchan = &sdev->vchans[i];
1095 
1096 		list_del(&vchan->vc.chan.device_node);
1097 		tasklet_kill(&vchan->vc.task);
1098 	}
1099 }
1100 
1101 /*
1102  * For A31:
1103  *
1104  * There's 16 physical channels that can work in parallel.
1105  *
1106  * However we have 30 different endpoints for our requests.
1107  *
1108  * Since the channels are able to handle only an unidirectional
1109  * transfer, we need to allocate more virtual channels so that
1110  * everyone can grab one channel.
1111  *
1112  * Some devices can't work in both direction (mostly because it
1113  * wouldn't make sense), so we have a bit fewer virtual channels than
1114  * 2 channels per endpoints.
1115  */
1116 
1117 static struct sun6i_dma_config sun6i_a31_dma_cfg = {
1118 	.nr_max_channels = 16,
1119 	.nr_max_requests = 30,
1120 	.nr_max_vchans   = 53,
1121 	.set_burst_length = sun6i_set_burst_length_a31,
1122 	.set_drq          = sun6i_set_drq_a31,
1123 	.set_mode         = sun6i_set_mode_a31,
1124 	.src_burst_lengths = BIT(1) | BIT(8),
1125 	.dst_burst_lengths = BIT(1) | BIT(8),
1126 	.src_addr_widths   = BIT(DMA_SLAVE_BUSWIDTH_1_BYTE) |
1127 			     BIT(DMA_SLAVE_BUSWIDTH_2_BYTES) |
1128 			     BIT(DMA_SLAVE_BUSWIDTH_4_BYTES),
1129 	.dst_addr_widths   = BIT(DMA_SLAVE_BUSWIDTH_1_BYTE) |
1130 			     BIT(DMA_SLAVE_BUSWIDTH_2_BYTES) |
1131 			     BIT(DMA_SLAVE_BUSWIDTH_4_BYTES),
1132 };
1133 
1134 /*
1135  * The A23 only has 8 physical channels, a maximum DRQ port id of 24,
1136  * and a total of 37 usable source and destination endpoints.
1137  */
1138 
1139 static struct sun6i_dma_config sun8i_a23_dma_cfg = {
1140 	.nr_max_channels = 8,
1141 	.nr_max_requests = 24,
1142 	.nr_max_vchans   = 37,
1143 	.clock_autogate_enable = sun6i_enable_clock_autogate_a23,
1144 	.set_burst_length = sun6i_set_burst_length_a31,
1145 	.set_drq          = sun6i_set_drq_a31,
1146 	.set_mode         = sun6i_set_mode_a31,
1147 	.src_burst_lengths = BIT(1) | BIT(8),
1148 	.dst_burst_lengths = BIT(1) | BIT(8),
1149 	.src_addr_widths   = BIT(DMA_SLAVE_BUSWIDTH_1_BYTE) |
1150 			     BIT(DMA_SLAVE_BUSWIDTH_2_BYTES) |
1151 			     BIT(DMA_SLAVE_BUSWIDTH_4_BYTES),
1152 	.dst_addr_widths   = BIT(DMA_SLAVE_BUSWIDTH_1_BYTE) |
1153 			     BIT(DMA_SLAVE_BUSWIDTH_2_BYTES) |
1154 			     BIT(DMA_SLAVE_BUSWIDTH_4_BYTES),
1155 };
1156 
1157 static struct sun6i_dma_config sun8i_a83t_dma_cfg = {
1158 	.nr_max_channels = 8,
1159 	.nr_max_requests = 28,
1160 	.nr_max_vchans   = 39,
1161 	.clock_autogate_enable = sun6i_enable_clock_autogate_a23,
1162 	.set_burst_length = sun6i_set_burst_length_a31,
1163 	.set_drq          = sun6i_set_drq_a31,
1164 	.set_mode         = sun6i_set_mode_a31,
1165 	.src_burst_lengths = BIT(1) | BIT(8),
1166 	.dst_burst_lengths = BIT(1) | BIT(8),
1167 	.src_addr_widths   = BIT(DMA_SLAVE_BUSWIDTH_1_BYTE) |
1168 			     BIT(DMA_SLAVE_BUSWIDTH_2_BYTES) |
1169 			     BIT(DMA_SLAVE_BUSWIDTH_4_BYTES),
1170 	.dst_addr_widths   = BIT(DMA_SLAVE_BUSWIDTH_1_BYTE) |
1171 			     BIT(DMA_SLAVE_BUSWIDTH_2_BYTES) |
1172 			     BIT(DMA_SLAVE_BUSWIDTH_4_BYTES),
1173 };
1174 
1175 /*
1176  * The H3 has 12 physical channels, a maximum DRQ port id of 27,
1177  * and a total of 34 usable source and destination endpoints.
1178  * It also supports additional burst lengths and bus widths,
1179  * and the burst length fields have different offsets.
1180  */
1181 
1182 static struct sun6i_dma_config sun8i_h3_dma_cfg = {
1183 	.nr_max_channels = 12,
1184 	.nr_max_requests = 27,
1185 	.nr_max_vchans   = 34,
1186 	.clock_autogate_enable = sun6i_enable_clock_autogate_h3,
1187 	.set_burst_length = sun6i_set_burst_length_h3,
1188 	.set_drq          = sun6i_set_drq_a31,
1189 	.set_mode         = sun6i_set_mode_a31,
1190 	.src_burst_lengths = BIT(1) | BIT(4) | BIT(8) | BIT(16),
1191 	.dst_burst_lengths = BIT(1) | BIT(4) | BIT(8) | BIT(16),
1192 	.src_addr_widths   = BIT(DMA_SLAVE_BUSWIDTH_1_BYTE) |
1193 			     BIT(DMA_SLAVE_BUSWIDTH_2_BYTES) |
1194 			     BIT(DMA_SLAVE_BUSWIDTH_4_BYTES) |
1195 			     BIT(DMA_SLAVE_BUSWIDTH_8_BYTES),
1196 	.dst_addr_widths   = BIT(DMA_SLAVE_BUSWIDTH_1_BYTE) |
1197 			     BIT(DMA_SLAVE_BUSWIDTH_2_BYTES) |
1198 			     BIT(DMA_SLAVE_BUSWIDTH_4_BYTES) |
1199 			     BIT(DMA_SLAVE_BUSWIDTH_8_BYTES),
1200 };
1201 
1202 /*
1203  * The A64 binding uses the number of dma channels from the
1204  * device tree node.
1205  */
1206 static struct sun6i_dma_config sun50i_a64_dma_cfg = {
1207 	.clock_autogate_enable = sun6i_enable_clock_autogate_h3,
1208 	.set_burst_length = sun6i_set_burst_length_h3,
1209 	.set_drq          = sun6i_set_drq_a31,
1210 	.set_mode         = sun6i_set_mode_a31,
1211 	.src_burst_lengths = BIT(1) | BIT(4) | BIT(8) | BIT(16),
1212 	.dst_burst_lengths = BIT(1) | BIT(4) | BIT(8) | BIT(16),
1213 	.src_addr_widths   = BIT(DMA_SLAVE_BUSWIDTH_1_BYTE) |
1214 			     BIT(DMA_SLAVE_BUSWIDTH_2_BYTES) |
1215 			     BIT(DMA_SLAVE_BUSWIDTH_4_BYTES) |
1216 			     BIT(DMA_SLAVE_BUSWIDTH_8_BYTES),
1217 	.dst_addr_widths   = BIT(DMA_SLAVE_BUSWIDTH_1_BYTE) |
1218 			     BIT(DMA_SLAVE_BUSWIDTH_2_BYTES) |
1219 			     BIT(DMA_SLAVE_BUSWIDTH_4_BYTES) |
1220 			     BIT(DMA_SLAVE_BUSWIDTH_8_BYTES),
1221 };
1222 
1223 /*
1224  * The A100 binding uses the number of dma channels from the
1225  * device tree node.
1226  */
1227 static struct sun6i_dma_config sun50i_a100_dma_cfg = {
1228 	.clock_autogate_enable = sun6i_enable_clock_autogate_h3,
1229 	.set_burst_length = sun6i_set_burst_length_h3,
1230 	.set_drq          = sun6i_set_drq_h6,
1231 	.set_mode         = sun6i_set_mode_h6,
1232 	.src_burst_lengths = BIT(1) | BIT(4) | BIT(8) | BIT(16),
1233 	.dst_burst_lengths = BIT(1) | BIT(4) | BIT(8) | BIT(16),
1234 	.src_addr_widths   = BIT(DMA_SLAVE_BUSWIDTH_1_BYTE) |
1235 			     BIT(DMA_SLAVE_BUSWIDTH_2_BYTES) |
1236 			     BIT(DMA_SLAVE_BUSWIDTH_4_BYTES) |
1237 			     BIT(DMA_SLAVE_BUSWIDTH_8_BYTES),
1238 	.dst_addr_widths   = BIT(DMA_SLAVE_BUSWIDTH_1_BYTE) |
1239 			     BIT(DMA_SLAVE_BUSWIDTH_2_BYTES) |
1240 			     BIT(DMA_SLAVE_BUSWIDTH_4_BYTES) |
1241 			     BIT(DMA_SLAVE_BUSWIDTH_8_BYTES),
1242 	.has_high_addr = true,
1243 	.has_mbus_clk = true,
1244 };
1245 
1246 /*
1247  * The H6 binding uses the number of dma channels from the
1248  * device tree node.
1249  */
1250 static struct sun6i_dma_config sun50i_h6_dma_cfg = {
1251 	.clock_autogate_enable = sun6i_enable_clock_autogate_h3,
1252 	.set_burst_length = sun6i_set_burst_length_h3,
1253 	.set_drq          = sun6i_set_drq_h6,
1254 	.set_mode         = sun6i_set_mode_h6,
1255 	.src_burst_lengths = BIT(1) | BIT(4) | BIT(8) | BIT(16),
1256 	.dst_burst_lengths = BIT(1) | BIT(4) | BIT(8) | BIT(16),
1257 	.src_addr_widths   = BIT(DMA_SLAVE_BUSWIDTH_1_BYTE) |
1258 			     BIT(DMA_SLAVE_BUSWIDTH_2_BYTES) |
1259 			     BIT(DMA_SLAVE_BUSWIDTH_4_BYTES) |
1260 			     BIT(DMA_SLAVE_BUSWIDTH_8_BYTES),
1261 	.dst_addr_widths   = BIT(DMA_SLAVE_BUSWIDTH_1_BYTE) |
1262 			     BIT(DMA_SLAVE_BUSWIDTH_2_BYTES) |
1263 			     BIT(DMA_SLAVE_BUSWIDTH_4_BYTES) |
1264 			     BIT(DMA_SLAVE_BUSWIDTH_8_BYTES),
1265 	.has_mbus_clk = true,
1266 };
1267 
1268 /*
1269  * The V3s have only 8 physical channels, a maximum DRQ port id of 23,
1270  * and a total of 24 usable source and destination endpoints.
1271  */
1272 
1273 static struct sun6i_dma_config sun8i_v3s_dma_cfg = {
1274 	.nr_max_channels = 8,
1275 	.nr_max_requests = 23,
1276 	.nr_max_vchans   = 24,
1277 	.clock_autogate_enable = sun6i_enable_clock_autogate_a23,
1278 	.set_burst_length = sun6i_set_burst_length_a31,
1279 	.set_drq          = sun6i_set_drq_a31,
1280 	.set_mode         = sun6i_set_mode_a31,
1281 	.src_burst_lengths = BIT(1) | BIT(8),
1282 	.dst_burst_lengths = BIT(1) | BIT(8),
1283 	.src_addr_widths   = BIT(DMA_SLAVE_BUSWIDTH_1_BYTE) |
1284 			     BIT(DMA_SLAVE_BUSWIDTH_2_BYTES) |
1285 			     BIT(DMA_SLAVE_BUSWIDTH_4_BYTES),
1286 	.dst_addr_widths   = BIT(DMA_SLAVE_BUSWIDTH_1_BYTE) |
1287 			     BIT(DMA_SLAVE_BUSWIDTH_2_BYTES) |
1288 			     BIT(DMA_SLAVE_BUSWIDTH_4_BYTES),
1289 };
1290 
1291 static const struct of_device_id sun6i_dma_match[] = {
1292 	{ .compatible = "allwinner,sun6i-a31-dma", .data = &sun6i_a31_dma_cfg },
1293 	{ .compatible = "allwinner,sun8i-a23-dma", .data = &sun8i_a23_dma_cfg },
1294 	{ .compatible = "allwinner,sun8i-a83t-dma", .data = &sun8i_a83t_dma_cfg },
1295 	{ .compatible = "allwinner,sun8i-h3-dma", .data = &sun8i_h3_dma_cfg },
1296 	{ .compatible = "allwinner,sun8i-v3s-dma", .data = &sun8i_v3s_dma_cfg },
1297 	{ .compatible = "allwinner,sun20i-d1-dma", .data = &sun50i_a100_dma_cfg },
1298 	{ .compatible = "allwinner,sun50i-a64-dma", .data = &sun50i_a64_dma_cfg },
1299 	{ .compatible = "allwinner,sun50i-a100-dma", .data = &sun50i_a100_dma_cfg },
1300 	{ .compatible = "allwinner,sun50i-h6-dma", .data = &sun50i_h6_dma_cfg },
1301 	{ /* sentinel */ }
1302 };
1303 MODULE_DEVICE_TABLE(of, sun6i_dma_match);
1304 
1305 static int sun6i_dma_probe(struct platform_device *pdev)
1306 {
1307 	struct device_node *np = pdev->dev.of_node;
1308 	struct sun6i_dma_dev *sdc;
1309 	int ret, i;
1310 
1311 	sdc = devm_kzalloc(&pdev->dev, sizeof(*sdc), GFP_KERNEL);
1312 	if (!sdc)
1313 		return -ENOMEM;
1314 
1315 	sdc->cfg = of_device_get_match_data(&pdev->dev);
1316 	if (!sdc->cfg)
1317 		return -ENODEV;
1318 
1319 	sdc->base = devm_platform_ioremap_resource(pdev, 0);
1320 	if (IS_ERR(sdc->base))
1321 		return PTR_ERR(sdc->base);
1322 
1323 	sdc->irq = platform_get_irq(pdev, 0);
1324 	if (sdc->irq < 0)
1325 		return sdc->irq;
1326 
1327 	sdc->clk = devm_clk_get(&pdev->dev, NULL);
1328 	if (IS_ERR(sdc->clk)) {
1329 		dev_err(&pdev->dev, "No clock specified\n");
1330 		return PTR_ERR(sdc->clk);
1331 	}
1332 
1333 	if (sdc->cfg->has_mbus_clk) {
1334 		sdc->clk_mbus = devm_clk_get(&pdev->dev, "mbus");
1335 		if (IS_ERR(sdc->clk_mbus)) {
1336 			dev_err(&pdev->dev, "No mbus clock specified\n");
1337 			return PTR_ERR(sdc->clk_mbus);
1338 		}
1339 	}
1340 
1341 	sdc->rstc = devm_reset_control_get(&pdev->dev, NULL);
1342 	if (IS_ERR(sdc->rstc)) {
1343 		dev_err(&pdev->dev, "No reset controller specified\n");
1344 		return PTR_ERR(sdc->rstc);
1345 	}
1346 
1347 	sdc->pool = dmam_pool_create(dev_name(&pdev->dev), &pdev->dev,
1348 				     sizeof(struct sun6i_dma_lli), 4, 0);
1349 	if (!sdc->pool) {
1350 		dev_err(&pdev->dev, "No memory for descriptors dma pool\n");
1351 		return -ENOMEM;
1352 	}
1353 
1354 	platform_set_drvdata(pdev, sdc);
1355 	INIT_LIST_HEAD(&sdc->pending);
1356 	spin_lock_init(&sdc->lock);
1357 
1358 	dma_set_max_seg_size(&pdev->dev, SZ_32M - 1);
1359 
1360 	dma_cap_set(DMA_PRIVATE, sdc->slave.cap_mask);
1361 	dma_cap_set(DMA_MEMCPY, sdc->slave.cap_mask);
1362 	dma_cap_set(DMA_SLAVE, sdc->slave.cap_mask);
1363 	dma_cap_set(DMA_CYCLIC, sdc->slave.cap_mask);
1364 
1365 	INIT_LIST_HEAD(&sdc->slave.channels);
1366 	sdc->slave.device_free_chan_resources	= sun6i_dma_free_chan_resources;
1367 	sdc->slave.device_tx_status		= sun6i_dma_tx_status;
1368 	sdc->slave.device_issue_pending		= sun6i_dma_issue_pending;
1369 	sdc->slave.device_prep_slave_sg		= sun6i_dma_prep_slave_sg;
1370 	sdc->slave.device_prep_dma_memcpy	= sun6i_dma_prep_dma_memcpy;
1371 	sdc->slave.device_prep_dma_cyclic	= sun6i_dma_prep_dma_cyclic;
1372 	sdc->slave.copy_align			= DMAENGINE_ALIGN_4_BYTES;
1373 	sdc->slave.device_config		= sun6i_dma_config;
1374 	sdc->slave.device_pause			= sun6i_dma_pause;
1375 	sdc->slave.device_resume		= sun6i_dma_resume;
1376 	sdc->slave.device_terminate_all		= sun6i_dma_terminate_all;
1377 	sdc->slave.src_addr_widths		= sdc->cfg->src_addr_widths;
1378 	sdc->slave.dst_addr_widths		= sdc->cfg->dst_addr_widths;
1379 	sdc->slave.directions			= BIT(DMA_DEV_TO_MEM) |
1380 						  BIT(DMA_MEM_TO_DEV);
1381 	sdc->slave.residue_granularity		= DMA_RESIDUE_GRANULARITY_BURST;
1382 	sdc->slave.dev = &pdev->dev;
1383 
1384 	sdc->num_pchans = sdc->cfg->nr_max_channels;
1385 	sdc->num_vchans = sdc->cfg->nr_max_vchans;
1386 	sdc->max_request = sdc->cfg->nr_max_requests;
1387 
1388 	ret = of_property_read_u32(np, "dma-channels", &sdc->num_pchans);
1389 	if (ret && !sdc->num_pchans) {
1390 		dev_err(&pdev->dev, "Can't get dma-channels.\n");
1391 		return ret;
1392 	}
1393 
1394 	ret = of_property_read_u32(np, "dma-requests", &sdc->max_request);
1395 	if (ret && !sdc->max_request) {
1396 		dev_info(&pdev->dev, "Missing dma-requests, using %u.\n",
1397 			 DMA_CHAN_MAX_DRQ_A31);
1398 		sdc->max_request = DMA_CHAN_MAX_DRQ_A31;
1399 	}
1400 
1401 	/*
1402 	 * If the number of vchans is not specified, derive it from the
1403 	 * highest port number, at most one channel per port and direction.
1404 	 */
1405 	if (!sdc->num_vchans)
1406 		sdc->num_vchans = 2 * (sdc->max_request + 1);
1407 
1408 	sdc->pchans = devm_kcalloc(&pdev->dev, sdc->num_pchans,
1409 				   sizeof(struct sun6i_pchan), GFP_KERNEL);
1410 	if (!sdc->pchans)
1411 		return -ENOMEM;
1412 
1413 	sdc->vchans = devm_kcalloc(&pdev->dev, sdc->num_vchans,
1414 				   sizeof(struct sun6i_vchan), GFP_KERNEL);
1415 	if (!sdc->vchans)
1416 		return -ENOMEM;
1417 
1418 	tasklet_setup(&sdc->task, sun6i_dma_tasklet);
1419 
1420 	for (i = 0; i < sdc->num_pchans; i++) {
1421 		struct sun6i_pchan *pchan = &sdc->pchans[i];
1422 
1423 		pchan->idx = i;
1424 		pchan->base = sdc->base + 0x100 + i * 0x40;
1425 	}
1426 
1427 	for (i = 0; i < sdc->num_vchans; i++) {
1428 		struct sun6i_vchan *vchan = &sdc->vchans[i];
1429 
1430 		INIT_LIST_HEAD(&vchan->node);
1431 		vchan->vc.desc_free = sun6i_dma_free_desc;
1432 		vchan_init(&vchan->vc, &sdc->slave);
1433 	}
1434 
1435 	ret = reset_control_deassert(sdc->rstc);
1436 	if (ret) {
1437 		dev_err(&pdev->dev, "Couldn't deassert the device from reset\n");
1438 		goto err_chan_free;
1439 	}
1440 
1441 	ret = clk_prepare_enable(sdc->clk);
1442 	if (ret) {
1443 		dev_err(&pdev->dev, "Couldn't enable the clock\n");
1444 		goto err_reset_assert;
1445 	}
1446 
1447 	if (sdc->cfg->has_mbus_clk) {
1448 		ret = clk_prepare_enable(sdc->clk_mbus);
1449 		if (ret) {
1450 			dev_err(&pdev->dev, "Couldn't enable mbus clock\n");
1451 			goto err_clk_disable;
1452 		}
1453 	}
1454 
1455 	ret = devm_request_irq(&pdev->dev, sdc->irq, sun6i_dma_interrupt, 0,
1456 			       dev_name(&pdev->dev), sdc);
1457 	if (ret) {
1458 		dev_err(&pdev->dev, "Cannot request IRQ\n");
1459 		goto err_mbus_clk_disable;
1460 	}
1461 
1462 	ret = dma_async_device_register(&sdc->slave);
1463 	if (ret) {
1464 		dev_warn(&pdev->dev, "Failed to register DMA engine device\n");
1465 		goto err_irq_disable;
1466 	}
1467 
1468 	ret = of_dma_controller_register(pdev->dev.of_node, sun6i_dma_of_xlate,
1469 					 sdc);
1470 	if (ret) {
1471 		dev_err(&pdev->dev, "of_dma_controller_register failed\n");
1472 		goto err_dma_unregister;
1473 	}
1474 
1475 	if (sdc->cfg->clock_autogate_enable)
1476 		sdc->cfg->clock_autogate_enable(sdc);
1477 
1478 	return 0;
1479 
1480 err_dma_unregister:
1481 	dma_async_device_unregister(&sdc->slave);
1482 err_irq_disable:
1483 	sun6i_kill_tasklet(sdc);
1484 err_mbus_clk_disable:
1485 	clk_disable_unprepare(sdc->clk_mbus);
1486 err_clk_disable:
1487 	clk_disable_unprepare(sdc->clk);
1488 err_reset_assert:
1489 	reset_control_assert(sdc->rstc);
1490 err_chan_free:
1491 	sun6i_dma_free(sdc);
1492 	return ret;
1493 }
1494 
1495 static void sun6i_dma_remove(struct platform_device *pdev)
1496 {
1497 	struct sun6i_dma_dev *sdc = platform_get_drvdata(pdev);
1498 
1499 	of_dma_controller_free(pdev->dev.of_node);
1500 	dma_async_device_unregister(&sdc->slave);
1501 
1502 	sun6i_kill_tasklet(sdc);
1503 
1504 	clk_disable_unprepare(sdc->clk_mbus);
1505 	clk_disable_unprepare(sdc->clk);
1506 	reset_control_assert(sdc->rstc);
1507 
1508 	sun6i_dma_free(sdc);
1509 }
1510 
1511 static struct platform_driver sun6i_dma_driver = {
1512 	.probe		= sun6i_dma_probe,
1513 	.remove		= sun6i_dma_remove,
1514 	.driver = {
1515 		.name		= "sun6i-dma",
1516 		.of_match_table	= sun6i_dma_match,
1517 	},
1518 };
1519 module_platform_driver(sun6i_dma_driver);
1520 
1521 MODULE_DESCRIPTION("Allwinner A31 DMA Controller Driver");
1522 MODULE_AUTHOR("Sugar <shuge@allwinnertech.com>");
1523 MODULE_AUTHOR("Maxime Ripard <maxime.ripard@free-electrons.com>");
1524 MODULE_LICENSE("GPL");
1525