xref: /linux/drivers/dma/sh/rcar-dmac.c (revision e9f0878c4b2004ac19581274c1ae4c61ae3ca70e)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Renesas R-Car Gen2 DMA Controller Driver
4  *
5  * Copyright (C) 2014 Renesas Electronics Inc.
6  *
7  * Author: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
8  */
9 
10 #include <linux/delay.h>
11 #include <linux/dma-mapping.h>
12 #include <linux/dmaengine.h>
13 #include <linux/interrupt.h>
14 #include <linux/list.h>
15 #include <linux/module.h>
16 #include <linux/mutex.h>
17 #include <linux/of.h>
18 #include <linux/of_dma.h>
19 #include <linux/of_platform.h>
20 #include <linux/platform_device.h>
21 #include <linux/pm_runtime.h>
22 #include <linux/slab.h>
23 #include <linux/spinlock.h>
24 
25 #include "../dmaengine.h"
26 
27 /*
28  * struct rcar_dmac_xfer_chunk - Descriptor for a hardware transfer
29  * @node: entry in the parent's chunks list
30  * @src_addr: device source address
31  * @dst_addr: device destination address
32  * @size: transfer size in bytes
33  */
34 struct rcar_dmac_xfer_chunk {
35 	struct list_head node;
36 
37 	dma_addr_t src_addr;
38 	dma_addr_t dst_addr;
39 	u32 size;
40 };
41 
42 /*
43  * struct rcar_dmac_hw_desc - Hardware descriptor for a transfer chunk
44  * @sar: value of the SAR register (source address)
45  * @dar: value of the DAR register (destination address)
46  * @tcr: value of the TCR register (transfer count)
47  */
48 struct rcar_dmac_hw_desc {
49 	u32 sar;
50 	u32 dar;
51 	u32 tcr;
52 	u32 reserved;
53 } __attribute__((__packed__));
54 
55 /*
56  * struct rcar_dmac_desc - R-Car Gen2 DMA Transfer Descriptor
57  * @async_tx: base DMA asynchronous transaction descriptor
58  * @direction: direction of the DMA transfer
59  * @xfer_shift: log2 of the transfer size
60  * @chcr: value of the channel configuration register for this transfer
61  * @node: entry in the channel's descriptors lists
62  * @chunks: list of transfer chunks for this transfer
63  * @running: the transfer chunk being currently processed
64  * @nchunks: number of transfer chunks for this transfer
65  * @hwdescs.use: whether the transfer descriptor uses hardware descriptors
66  * @hwdescs.mem: hardware descriptors memory for the transfer
67  * @hwdescs.dma: device address of the hardware descriptors memory
68  * @hwdescs.size: size of the hardware descriptors in bytes
69  * @size: transfer size in bytes
70  * @cyclic: when set indicates that the DMA transfer is cyclic
71  */
72 struct rcar_dmac_desc {
73 	struct dma_async_tx_descriptor async_tx;
74 	enum dma_transfer_direction direction;
75 	unsigned int xfer_shift;
76 	u32 chcr;
77 
78 	struct list_head node;
79 	struct list_head chunks;
80 	struct rcar_dmac_xfer_chunk *running;
81 	unsigned int nchunks;
82 
83 	struct {
84 		bool use;
85 		struct rcar_dmac_hw_desc *mem;
86 		dma_addr_t dma;
87 		size_t size;
88 	} hwdescs;
89 
90 	unsigned int size;
91 	bool cyclic;
92 };
93 
94 #define to_rcar_dmac_desc(d)	container_of(d, struct rcar_dmac_desc, async_tx)
95 
96 /*
97  * struct rcar_dmac_desc_page - One page worth of descriptors
98  * @node: entry in the channel's pages list
99  * @descs: array of DMA descriptors
100  * @chunks: array of transfer chunk descriptors
101  */
102 struct rcar_dmac_desc_page {
103 	struct list_head node;
104 
105 	union {
106 		struct rcar_dmac_desc descs[0];
107 		struct rcar_dmac_xfer_chunk chunks[0];
108 	};
109 };
110 
111 #define RCAR_DMAC_DESCS_PER_PAGE					\
112 	((PAGE_SIZE - offsetof(struct rcar_dmac_desc_page, descs)) /	\
113 	sizeof(struct rcar_dmac_desc))
114 #define RCAR_DMAC_XFER_CHUNKS_PER_PAGE					\
115 	((PAGE_SIZE - offsetof(struct rcar_dmac_desc_page, chunks)) /	\
116 	sizeof(struct rcar_dmac_xfer_chunk))
117 
118 /*
119  * struct rcar_dmac_chan_slave - Slave configuration
120  * @slave_addr: slave memory address
121  * @xfer_size: size (in bytes) of hardware transfers
122  */
123 struct rcar_dmac_chan_slave {
124 	phys_addr_t slave_addr;
125 	unsigned int xfer_size;
126 };
127 
128 /*
129  * struct rcar_dmac_chan_map - Map of slave device phys to dma address
130  * @addr: slave dma address
131  * @dir: direction of mapping
132  * @slave: slave configuration that is mapped
133  */
134 struct rcar_dmac_chan_map {
135 	dma_addr_t addr;
136 	enum dma_data_direction dir;
137 	struct rcar_dmac_chan_slave slave;
138 };
139 
140 /*
141  * struct rcar_dmac_chan - R-Car Gen2 DMA Controller Channel
142  * @chan: base DMA channel object
143  * @iomem: channel I/O memory base
144  * @index: index of this channel in the controller
145  * @irq: channel IRQ
146  * @src: slave memory address and size on the source side
147  * @dst: slave memory address and size on the destination side
148  * @mid_rid: hardware MID/RID for the DMA client using this channel
149  * @lock: protects the channel CHCR register and the desc members
150  * @desc.free: list of free descriptors
151  * @desc.pending: list of pending descriptors (submitted with tx_submit)
152  * @desc.active: list of active descriptors (activated with issue_pending)
153  * @desc.done: list of completed descriptors
154  * @desc.wait: list of descriptors waiting for an ack
155  * @desc.running: the descriptor being processed (a member of the active list)
156  * @desc.chunks_free: list of free transfer chunk descriptors
157  * @desc.pages: list of pages used by allocated descriptors
158  */
159 struct rcar_dmac_chan {
160 	struct dma_chan chan;
161 	void __iomem *iomem;
162 	unsigned int index;
163 	int irq;
164 
165 	struct rcar_dmac_chan_slave src;
166 	struct rcar_dmac_chan_slave dst;
167 	struct rcar_dmac_chan_map map;
168 	int mid_rid;
169 
170 	spinlock_t lock;
171 
172 	struct {
173 		struct list_head free;
174 		struct list_head pending;
175 		struct list_head active;
176 		struct list_head done;
177 		struct list_head wait;
178 		struct rcar_dmac_desc *running;
179 
180 		struct list_head chunks_free;
181 
182 		struct list_head pages;
183 	} desc;
184 };
185 
186 #define to_rcar_dmac_chan(c)	container_of(c, struct rcar_dmac_chan, chan)
187 
188 /*
189  * struct rcar_dmac - R-Car Gen2 DMA Controller
190  * @engine: base DMA engine object
191  * @dev: the hardware device
192  * @iomem: remapped I/O memory base
193  * @n_channels: number of available channels
194  * @channels: array of DMAC channels
195  * @modules: bitmask of client modules in use
196  */
197 struct rcar_dmac {
198 	struct dma_device engine;
199 	struct device *dev;
200 	void __iomem *iomem;
201 
202 	unsigned int n_channels;
203 	struct rcar_dmac_chan *channels;
204 
205 	DECLARE_BITMAP(modules, 256);
206 };
207 
208 #define to_rcar_dmac(d)		container_of(d, struct rcar_dmac, engine)
209 
210 /* -----------------------------------------------------------------------------
211  * Registers
212  */
213 
214 #define RCAR_DMAC_CHAN_OFFSET(i)	(0x8000 + 0x80 * (i))
215 
216 #define RCAR_DMAISTA			0x0020
217 #define RCAR_DMASEC			0x0030
218 #define RCAR_DMAOR			0x0060
219 #define RCAR_DMAOR_PRI_FIXED		(0 << 8)
220 #define RCAR_DMAOR_PRI_ROUND_ROBIN	(3 << 8)
221 #define RCAR_DMAOR_AE			(1 << 2)
222 #define RCAR_DMAOR_DME			(1 << 0)
223 #define RCAR_DMACHCLR			0x0080
224 #define RCAR_DMADPSEC			0x00a0
225 
226 #define RCAR_DMASAR			0x0000
227 #define RCAR_DMADAR			0x0004
228 #define RCAR_DMATCR			0x0008
229 #define RCAR_DMATCR_MASK		0x00ffffff
230 #define RCAR_DMATSR			0x0028
231 #define RCAR_DMACHCR			0x000c
232 #define RCAR_DMACHCR_CAE		(1 << 31)
233 #define RCAR_DMACHCR_CAIE		(1 << 30)
234 #define RCAR_DMACHCR_DPM_DISABLED	(0 << 28)
235 #define RCAR_DMACHCR_DPM_ENABLED	(1 << 28)
236 #define RCAR_DMACHCR_DPM_REPEAT		(2 << 28)
237 #define RCAR_DMACHCR_DPM_INFINITE	(3 << 28)
238 #define RCAR_DMACHCR_RPT_SAR		(1 << 27)
239 #define RCAR_DMACHCR_RPT_DAR		(1 << 26)
240 #define RCAR_DMACHCR_RPT_TCR		(1 << 25)
241 #define RCAR_DMACHCR_DPB		(1 << 22)
242 #define RCAR_DMACHCR_DSE		(1 << 19)
243 #define RCAR_DMACHCR_DSIE		(1 << 18)
244 #define RCAR_DMACHCR_TS_1B		((0 << 20) | (0 << 3))
245 #define RCAR_DMACHCR_TS_2B		((0 << 20) | (1 << 3))
246 #define RCAR_DMACHCR_TS_4B		((0 << 20) | (2 << 3))
247 #define RCAR_DMACHCR_TS_16B		((0 << 20) | (3 << 3))
248 #define RCAR_DMACHCR_TS_32B		((1 << 20) | (0 << 3))
249 #define RCAR_DMACHCR_TS_64B		((1 << 20) | (1 << 3))
250 #define RCAR_DMACHCR_TS_8B		((1 << 20) | (3 << 3))
251 #define RCAR_DMACHCR_DM_FIXED		(0 << 14)
252 #define RCAR_DMACHCR_DM_INC		(1 << 14)
253 #define RCAR_DMACHCR_DM_DEC		(2 << 14)
254 #define RCAR_DMACHCR_SM_FIXED		(0 << 12)
255 #define RCAR_DMACHCR_SM_INC		(1 << 12)
256 #define RCAR_DMACHCR_SM_DEC		(2 << 12)
257 #define RCAR_DMACHCR_RS_AUTO		(4 << 8)
258 #define RCAR_DMACHCR_RS_DMARS		(8 << 8)
259 #define RCAR_DMACHCR_IE			(1 << 2)
260 #define RCAR_DMACHCR_TE			(1 << 1)
261 #define RCAR_DMACHCR_DE			(1 << 0)
262 #define RCAR_DMATCRB			0x0018
263 #define RCAR_DMATSRB			0x0038
264 #define RCAR_DMACHCRB			0x001c
265 #define RCAR_DMACHCRB_DCNT(n)		((n) << 24)
266 #define RCAR_DMACHCRB_DPTR_MASK		(0xff << 16)
267 #define RCAR_DMACHCRB_DPTR_SHIFT	16
268 #define RCAR_DMACHCRB_DRST		(1 << 15)
269 #define RCAR_DMACHCRB_DTS		(1 << 8)
270 #define RCAR_DMACHCRB_SLM_NORMAL	(0 << 4)
271 #define RCAR_DMACHCRB_SLM_CLK(n)	((8 | (n)) << 4)
272 #define RCAR_DMACHCRB_PRI(n)		((n) << 0)
273 #define RCAR_DMARS			0x0040
274 #define RCAR_DMABUFCR			0x0048
275 #define RCAR_DMABUFCR_MBU(n)		((n) << 16)
276 #define RCAR_DMABUFCR_ULB(n)		((n) << 0)
277 #define RCAR_DMADPBASE			0x0050
278 #define RCAR_DMADPBASE_MASK		0xfffffff0
279 #define RCAR_DMADPBASE_SEL		(1 << 0)
280 #define RCAR_DMADPCR			0x0054
281 #define RCAR_DMADPCR_DIPT(n)		((n) << 24)
282 #define RCAR_DMAFIXSAR			0x0010
283 #define RCAR_DMAFIXDAR			0x0014
284 #define RCAR_DMAFIXDPBASE		0x0060
285 
286 /* Hardcode the MEMCPY transfer size to 4 bytes. */
287 #define RCAR_DMAC_MEMCPY_XFER_SIZE	4
288 
289 /* -----------------------------------------------------------------------------
290  * Device access
291  */
292 
293 static void rcar_dmac_write(struct rcar_dmac *dmac, u32 reg, u32 data)
294 {
295 	if (reg == RCAR_DMAOR)
296 		writew(data, dmac->iomem + reg);
297 	else
298 		writel(data, dmac->iomem + reg);
299 }
300 
301 static u32 rcar_dmac_read(struct rcar_dmac *dmac, u32 reg)
302 {
303 	if (reg == RCAR_DMAOR)
304 		return readw(dmac->iomem + reg);
305 	else
306 		return readl(dmac->iomem + reg);
307 }
308 
309 static u32 rcar_dmac_chan_read(struct rcar_dmac_chan *chan, u32 reg)
310 {
311 	if (reg == RCAR_DMARS)
312 		return readw(chan->iomem + reg);
313 	else
314 		return readl(chan->iomem + reg);
315 }
316 
317 static void rcar_dmac_chan_write(struct rcar_dmac_chan *chan, u32 reg, u32 data)
318 {
319 	if (reg == RCAR_DMARS)
320 		writew(data, chan->iomem + reg);
321 	else
322 		writel(data, chan->iomem + reg);
323 }
324 
325 /* -----------------------------------------------------------------------------
326  * Initialization and configuration
327  */
328 
329 static bool rcar_dmac_chan_is_busy(struct rcar_dmac_chan *chan)
330 {
331 	u32 chcr = rcar_dmac_chan_read(chan, RCAR_DMACHCR);
332 
333 	return !!(chcr & (RCAR_DMACHCR_DE | RCAR_DMACHCR_TE));
334 }
335 
336 static void rcar_dmac_chan_start_xfer(struct rcar_dmac_chan *chan)
337 {
338 	struct rcar_dmac_desc *desc = chan->desc.running;
339 	u32 chcr = desc->chcr;
340 
341 	WARN_ON_ONCE(rcar_dmac_chan_is_busy(chan));
342 
343 	if (chan->mid_rid >= 0)
344 		rcar_dmac_chan_write(chan, RCAR_DMARS, chan->mid_rid);
345 
346 	if (desc->hwdescs.use) {
347 		struct rcar_dmac_xfer_chunk *chunk =
348 			list_first_entry(&desc->chunks,
349 					 struct rcar_dmac_xfer_chunk, node);
350 
351 		dev_dbg(chan->chan.device->dev,
352 			"chan%u: queue desc %p: %u@%pad\n",
353 			chan->index, desc, desc->nchunks, &desc->hwdescs.dma);
354 
355 #ifdef CONFIG_ARCH_DMA_ADDR_T_64BIT
356 		rcar_dmac_chan_write(chan, RCAR_DMAFIXSAR,
357 				     chunk->src_addr >> 32);
358 		rcar_dmac_chan_write(chan, RCAR_DMAFIXDAR,
359 				     chunk->dst_addr >> 32);
360 		rcar_dmac_chan_write(chan, RCAR_DMAFIXDPBASE,
361 				     desc->hwdescs.dma >> 32);
362 #endif
363 		rcar_dmac_chan_write(chan, RCAR_DMADPBASE,
364 				     (desc->hwdescs.dma & 0xfffffff0) |
365 				     RCAR_DMADPBASE_SEL);
366 		rcar_dmac_chan_write(chan, RCAR_DMACHCRB,
367 				     RCAR_DMACHCRB_DCNT(desc->nchunks - 1) |
368 				     RCAR_DMACHCRB_DRST);
369 
370 		/*
371 		 * Errata: When descriptor memory is accessed through an IOMMU
372 		 * the DMADAR register isn't initialized automatically from the
373 		 * first descriptor at beginning of transfer by the DMAC like it
374 		 * should. Initialize it manually with the destination address
375 		 * of the first chunk.
376 		 */
377 		rcar_dmac_chan_write(chan, RCAR_DMADAR,
378 				     chunk->dst_addr & 0xffffffff);
379 
380 		/*
381 		 * Program the descriptor stage interrupt to occur after the end
382 		 * of the first stage.
383 		 */
384 		rcar_dmac_chan_write(chan, RCAR_DMADPCR, RCAR_DMADPCR_DIPT(1));
385 
386 		chcr |= RCAR_DMACHCR_RPT_SAR | RCAR_DMACHCR_RPT_DAR
387 		     |  RCAR_DMACHCR_RPT_TCR | RCAR_DMACHCR_DPB;
388 
389 		/*
390 		 * If the descriptor isn't cyclic enable normal descriptor mode
391 		 * and the transfer completion interrupt.
392 		 */
393 		if (!desc->cyclic)
394 			chcr |= RCAR_DMACHCR_DPM_ENABLED | RCAR_DMACHCR_IE;
395 		/*
396 		 * If the descriptor is cyclic and has a callback enable the
397 		 * descriptor stage interrupt in infinite repeat mode.
398 		 */
399 		else if (desc->async_tx.callback)
400 			chcr |= RCAR_DMACHCR_DPM_INFINITE | RCAR_DMACHCR_DSIE;
401 		/*
402 		 * Otherwise just select infinite repeat mode without any
403 		 * interrupt.
404 		 */
405 		else
406 			chcr |= RCAR_DMACHCR_DPM_INFINITE;
407 	} else {
408 		struct rcar_dmac_xfer_chunk *chunk = desc->running;
409 
410 		dev_dbg(chan->chan.device->dev,
411 			"chan%u: queue chunk %p: %u@%pad -> %pad\n",
412 			chan->index, chunk, chunk->size, &chunk->src_addr,
413 			&chunk->dst_addr);
414 
415 #ifdef CONFIG_ARCH_DMA_ADDR_T_64BIT
416 		rcar_dmac_chan_write(chan, RCAR_DMAFIXSAR,
417 				     chunk->src_addr >> 32);
418 		rcar_dmac_chan_write(chan, RCAR_DMAFIXDAR,
419 				     chunk->dst_addr >> 32);
420 #endif
421 		rcar_dmac_chan_write(chan, RCAR_DMASAR,
422 				     chunk->src_addr & 0xffffffff);
423 		rcar_dmac_chan_write(chan, RCAR_DMADAR,
424 				     chunk->dst_addr & 0xffffffff);
425 		rcar_dmac_chan_write(chan, RCAR_DMATCR,
426 				     chunk->size >> desc->xfer_shift);
427 
428 		chcr |= RCAR_DMACHCR_DPM_DISABLED | RCAR_DMACHCR_IE;
429 	}
430 
431 	rcar_dmac_chan_write(chan, RCAR_DMACHCR,
432 			     chcr | RCAR_DMACHCR_DE | RCAR_DMACHCR_CAIE);
433 }
434 
435 static int rcar_dmac_init(struct rcar_dmac *dmac)
436 {
437 	u16 dmaor;
438 
439 	/* Clear all channels and enable the DMAC globally. */
440 	rcar_dmac_write(dmac, RCAR_DMACHCLR, GENMASK(dmac->n_channels - 1, 0));
441 	rcar_dmac_write(dmac, RCAR_DMAOR,
442 			RCAR_DMAOR_PRI_FIXED | RCAR_DMAOR_DME);
443 
444 	dmaor = rcar_dmac_read(dmac, RCAR_DMAOR);
445 	if ((dmaor & (RCAR_DMAOR_AE | RCAR_DMAOR_DME)) != RCAR_DMAOR_DME) {
446 		dev_warn(dmac->dev, "DMAOR initialization failed.\n");
447 		return -EIO;
448 	}
449 
450 	return 0;
451 }
452 
453 /* -----------------------------------------------------------------------------
454  * Descriptors submission
455  */
456 
457 static dma_cookie_t rcar_dmac_tx_submit(struct dma_async_tx_descriptor *tx)
458 {
459 	struct rcar_dmac_chan *chan = to_rcar_dmac_chan(tx->chan);
460 	struct rcar_dmac_desc *desc = to_rcar_dmac_desc(tx);
461 	unsigned long flags;
462 	dma_cookie_t cookie;
463 
464 	spin_lock_irqsave(&chan->lock, flags);
465 
466 	cookie = dma_cookie_assign(tx);
467 
468 	dev_dbg(chan->chan.device->dev, "chan%u: submit #%d@%p\n",
469 		chan->index, tx->cookie, desc);
470 
471 	list_add_tail(&desc->node, &chan->desc.pending);
472 	desc->running = list_first_entry(&desc->chunks,
473 					 struct rcar_dmac_xfer_chunk, node);
474 
475 	spin_unlock_irqrestore(&chan->lock, flags);
476 
477 	return cookie;
478 }
479 
480 /* -----------------------------------------------------------------------------
481  * Descriptors allocation and free
482  */
483 
484 /*
485  * rcar_dmac_desc_alloc - Allocate a page worth of DMA descriptors
486  * @chan: the DMA channel
487  * @gfp: allocation flags
488  */
489 static int rcar_dmac_desc_alloc(struct rcar_dmac_chan *chan, gfp_t gfp)
490 {
491 	struct rcar_dmac_desc_page *page;
492 	unsigned long flags;
493 	LIST_HEAD(list);
494 	unsigned int i;
495 
496 	page = (void *)get_zeroed_page(gfp);
497 	if (!page)
498 		return -ENOMEM;
499 
500 	for (i = 0; i < RCAR_DMAC_DESCS_PER_PAGE; ++i) {
501 		struct rcar_dmac_desc *desc = &page->descs[i];
502 
503 		dma_async_tx_descriptor_init(&desc->async_tx, &chan->chan);
504 		desc->async_tx.tx_submit = rcar_dmac_tx_submit;
505 		INIT_LIST_HEAD(&desc->chunks);
506 
507 		list_add_tail(&desc->node, &list);
508 	}
509 
510 	spin_lock_irqsave(&chan->lock, flags);
511 	list_splice_tail(&list, &chan->desc.free);
512 	list_add_tail(&page->node, &chan->desc.pages);
513 	spin_unlock_irqrestore(&chan->lock, flags);
514 
515 	return 0;
516 }
517 
518 /*
519  * rcar_dmac_desc_put - Release a DMA transfer descriptor
520  * @chan: the DMA channel
521  * @desc: the descriptor
522  *
523  * Put the descriptor and its transfer chunk descriptors back in the channel's
524  * free descriptors lists. The descriptor's chunks list will be reinitialized to
525  * an empty list as a result.
526  *
527  * The descriptor must have been removed from the channel's lists before calling
528  * this function.
529  */
530 static void rcar_dmac_desc_put(struct rcar_dmac_chan *chan,
531 			       struct rcar_dmac_desc *desc)
532 {
533 	unsigned long flags;
534 
535 	spin_lock_irqsave(&chan->lock, flags);
536 	list_splice_tail_init(&desc->chunks, &chan->desc.chunks_free);
537 	list_add(&desc->node, &chan->desc.free);
538 	spin_unlock_irqrestore(&chan->lock, flags);
539 }
540 
541 static void rcar_dmac_desc_recycle_acked(struct rcar_dmac_chan *chan)
542 {
543 	struct rcar_dmac_desc *desc, *_desc;
544 	unsigned long flags;
545 	LIST_HEAD(list);
546 
547 	/*
548 	 * We have to temporarily move all descriptors from the wait list to a
549 	 * local list as iterating over the wait list, even with
550 	 * list_for_each_entry_safe, isn't safe if we release the channel lock
551 	 * around the rcar_dmac_desc_put() call.
552 	 */
553 	spin_lock_irqsave(&chan->lock, flags);
554 	list_splice_init(&chan->desc.wait, &list);
555 	spin_unlock_irqrestore(&chan->lock, flags);
556 
557 	list_for_each_entry_safe(desc, _desc, &list, node) {
558 		if (async_tx_test_ack(&desc->async_tx)) {
559 			list_del(&desc->node);
560 			rcar_dmac_desc_put(chan, desc);
561 		}
562 	}
563 
564 	if (list_empty(&list))
565 		return;
566 
567 	/* Put the remaining descriptors back in the wait list. */
568 	spin_lock_irqsave(&chan->lock, flags);
569 	list_splice(&list, &chan->desc.wait);
570 	spin_unlock_irqrestore(&chan->lock, flags);
571 }
572 
573 /*
574  * rcar_dmac_desc_get - Allocate a descriptor for a DMA transfer
575  * @chan: the DMA channel
576  *
577  * Locking: This function must be called in a non-atomic context.
578  *
579  * Return: A pointer to the allocated descriptor or NULL if no descriptor can
580  * be allocated.
581  */
582 static struct rcar_dmac_desc *rcar_dmac_desc_get(struct rcar_dmac_chan *chan)
583 {
584 	struct rcar_dmac_desc *desc;
585 	unsigned long flags;
586 	int ret;
587 
588 	/* Recycle acked descriptors before attempting allocation. */
589 	rcar_dmac_desc_recycle_acked(chan);
590 
591 	spin_lock_irqsave(&chan->lock, flags);
592 
593 	while (list_empty(&chan->desc.free)) {
594 		/*
595 		 * No free descriptors, allocate a page worth of them and try
596 		 * again, as someone else could race us to get the newly
597 		 * allocated descriptors. If the allocation fails return an
598 		 * error.
599 		 */
600 		spin_unlock_irqrestore(&chan->lock, flags);
601 		ret = rcar_dmac_desc_alloc(chan, GFP_NOWAIT);
602 		if (ret < 0)
603 			return NULL;
604 		spin_lock_irqsave(&chan->lock, flags);
605 	}
606 
607 	desc = list_first_entry(&chan->desc.free, struct rcar_dmac_desc, node);
608 	list_del(&desc->node);
609 
610 	spin_unlock_irqrestore(&chan->lock, flags);
611 
612 	return desc;
613 }
614 
615 /*
616  * rcar_dmac_xfer_chunk_alloc - Allocate a page worth of transfer chunks
617  * @chan: the DMA channel
618  * @gfp: allocation flags
619  */
620 static int rcar_dmac_xfer_chunk_alloc(struct rcar_dmac_chan *chan, gfp_t gfp)
621 {
622 	struct rcar_dmac_desc_page *page;
623 	unsigned long flags;
624 	LIST_HEAD(list);
625 	unsigned int i;
626 
627 	page = (void *)get_zeroed_page(gfp);
628 	if (!page)
629 		return -ENOMEM;
630 
631 	for (i = 0; i < RCAR_DMAC_XFER_CHUNKS_PER_PAGE; ++i) {
632 		struct rcar_dmac_xfer_chunk *chunk = &page->chunks[i];
633 
634 		list_add_tail(&chunk->node, &list);
635 	}
636 
637 	spin_lock_irqsave(&chan->lock, flags);
638 	list_splice_tail(&list, &chan->desc.chunks_free);
639 	list_add_tail(&page->node, &chan->desc.pages);
640 	spin_unlock_irqrestore(&chan->lock, flags);
641 
642 	return 0;
643 }
644 
645 /*
646  * rcar_dmac_xfer_chunk_get - Allocate a transfer chunk for a DMA transfer
647  * @chan: the DMA channel
648  *
649  * Locking: This function must be called in a non-atomic context.
650  *
651  * Return: A pointer to the allocated transfer chunk descriptor or NULL if no
652  * descriptor can be allocated.
653  */
654 static struct rcar_dmac_xfer_chunk *
655 rcar_dmac_xfer_chunk_get(struct rcar_dmac_chan *chan)
656 {
657 	struct rcar_dmac_xfer_chunk *chunk;
658 	unsigned long flags;
659 	int ret;
660 
661 	spin_lock_irqsave(&chan->lock, flags);
662 
663 	while (list_empty(&chan->desc.chunks_free)) {
664 		/*
665 		 * No free descriptors, allocate a page worth of them and try
666 		 * again, as someone else could race us to get the newly
667 		 * allocated descriptors. If the allocation fails return an
668 		 * error.
669 		 */
670 		spin_unlock_irqrestore(&chan->lock, flags);
671 		ret = rcar_dmac_xfer_chunk_alloc(chan, GFP_NOWAIT);
672 		if (ret < 0)
673 			return NULL;
674 		spin_lock_irqsave(&chan->lock, flags);
675 	}
676 
677 	chunk = list_first_entry(&chan->desc.chunks_free,
678 				 struct rcar_dmac_xfer_chunk, node);
679 	list_del(&chunk->node);
680 
681 	spin_unlock_irqrestore(&chan->lock, flags);
682 
683 	return chunk;
684 }
685 
686 static void rcar_dmac_realloc_hwdesc(struct rcar_dmac_chan *chan,
687 				     struct rcar_dmac_desc *desc, size_t size)
688 {
689 	/*
690 	 * dma_alloc_coherent() allocates memory in page size increments. To
691 	 * avoid reallocating the hardware descriptors when the allocated size
692 	 * wouldn't change align the requested size to a multiple of the page
693 	 * size.
694 	 */
695 	size = PAGE_ALIGN(size);
696 
697 	if (desc->hwdescs.size == size)
698 		return;
699 
700 	if (desc->hwdescs.mem) {
701 		dma_free_coherent(chan->chan.device->dev, desc->hwdescs.size,
702 				  desc->hwdescs.mem, desc->hwdescs.dma);
703 		desc->hwdescs.mem = NULL;
704 		desc->hwdescs.size = 0;
705 	}
706 
707 	if (!size)
708 		return;
709 
710 	desc->hwdescs.mem = dma_alloc_coherent(chan->chan.device->dev, size,
711 					       &desc->hwdescs.dma, GFP_NOWAIT);
712 	if (!desc->hwdescs.mem)
713 		return;
714 
715 	desc->hwdescs.size = size;
716 }
717 
718 static int rcar_dmac_fill_hwdesc(struct rcar_dmac_chan *chan,
719 				 struct rcar_dmac_desc *desc)
720 {
721 	struct rcar_dmac_xfer_chunk *chunk;
722 	struct rcar_dmac_hw_desc *hwdesc;
723 
724 	rcar_dmac_realloc_hwdesc(chan, desc, desc->nchunks * sizeof(*hwdesc));
725 
726 	hwdesc = desc->hwdescs.mem;
727 	if (!hwdesc)
728 		return -ENOMEM;
729 
730 	list_for_each_entry(chunk, &desc->chunks, node) {
731 		hwdesc->sar = chunk->src_addr;
732 		hwdesc->dar = chunk->dst_addr;
733 		hwdesc->tcr = chunk->size >> desc->xfer_shift;
734 		hwdesc++;
735 	}
736 
737 	return 0;
738 }
739 
740 /* -----------------------------------------------------------------------------
741  * Stop and reset
742  */
743 static void rcar_dmac_chcr_de_barrier(struct rcar_dmac_chan *chan)
744 {
745 	u32 chcr;
746 	unsigned int i;
747 
748 	/*
749 	 * Ensure that the setting of the DE bit is actually 0 after
750 	 * clearing it.
751 	 */
752 	for (i = 0; i < 1024; i++) {
753 		chcr = rcar_dmac_chan_read(chan, RCAR_DMACHCR);
754 		if (!(chcr & RCAR_DMACHCR_DE))
755 			return;
756 		udelay(1);
757 	}
758 
759 	dev_err(chan->chan.device->dev, "CHCR DE check error\n");
760 }
761 
762 static void rcar_dmac_clear_chcr_de(struct rcar_dmac_chan *chan)
763 {
764 	u32 chcr = rcar_dmac_chan_read(chan, RCAR_DMACHCR);
765 
766 	/* set DE=0 and flush remaining data */
767 	rcar_dmac_chan_write(chan, RCAR_DMACHCR, (chcr & ~RCAR_DMACHCR_DE));
768 
769 	/* make sure all remaining data was flushed */
770 	rcar_dmac_chcr_de_barrier(chan);
771 }
772 
773 static void rcar_dmac_chan_halt(struct rcar_dmac_chan *chan)
774 {
775 	u32 chcr = rcar_dmac_chan_read(chan, RCAR_DMACHCR);
776 
777 	chcr &= ~(RCAR_DMACHCR_DSE | RCAR_DMACHCR_DSIE | RCAR_DMACHCR_IE |
778 		  RCAR_DMACHCR_TE | RCAR_DMACHCR_DE |
779 		  RCAR_DMACHCR_CAE | RCAR_DMACHCR_CAIE);
780 	rcar_dmac_chan_write(chan, RCAR_DMACHCR, chcr);
781 	rcar_dmac_chcr_de_barrier(chan);
782 }
783 
784 static void rcar_dmac_chan_reinit(struct rcar_dmac_chan *chan)
785 {
786 	struct rcar_dmac_desc *desc, *_desc;
787 	unsigned long flags;
788 	LIST_HEAD(descs);
789 
790 	spin_lock_irqsave(&chan->lock, flags);
791 
792 	/* Move all non-free descriptors to the local lists. */
793 	list_splice_init(&chan->desc.pending, &descs);
794 	list_splice_init(&chan->desc.active, &descs);
795 	list_splice_init(&chan->desc.done, &descs);
796 	list_splice_init(&chan->desc.wait, &descs);
797 
798 	chan->desc.running = NULL;
799 
800 	spin_unlock_irqrestore(&chan->lock, flags);
801 
802 	list_for_each_entry_safe(desc, _desc, &descs, node) {
803 		list_del(&desc->node);
804 		rcar_dmac_desc_put(chan, desc);
805 	}
806 }
807 
808 static void rcar_dmac_stop_all_chan(struct rcar_dmac *dmac)
809 {
810 	unsigned int i;
811 
812 	/* Stop all channels. */
813 	for (i = 0; i < dmac->n_channels; ++i) {
814 		struct rcar_dmac_chan *chan = &dmac->channels[i];
815 
816 		/* Stop and reinitialize the channel. */
817 		spin_lock_irq(&chan->lock);
818 		rcar_dmac_chan_halt(chan);
819 		spin_unlock_irq(&chan->lock);
820 	}
821 }
822 
823 static int rcar_dmac_chan_pause(struct dma_chan *chan)
824 {
825 	unsigned long flags;
826 	struct rcar_dmac_chan *rchan = to_rcar_dmac_chan(chan);
827 
828 	spin_lock_irqsave(&rchan->lock, flags);
829 	rcar_dmac_clear_chcr_de(rchan);
830 	spin_unlock_irqrestore(&rchan->lock, flags);
831 
832 	return 0;
833 }
834 
835 /* -----------------------------------------------------------------------------
836  * Descriptors preparation
837  */
838 
839 static void rcar_dmac_chan_configure_desc(struct rcar_dmac_chan *chan,
840 					  struct rcar_dmac_desc *desc)
841 {
842 	static const u32 chcr_ts[] = {
843 		RCAR_DMACHCR_TS_1B, RCAR_DMACHCR_TS_2B,
844 		RCAR_DMACHCR_TS_4B, RCAR_DMACHCR_TS_8B,
845 		RCAR_DMACHCR_TS_16B, RCAR_DMACHCR_TS_32B,
846 		RCAR_DMACHCR_TS_64B,
847 	};
848 
849 	unsigned int xfer_size;
850 	u32 chcr;
851 
852 	switch (desc->direction) {
853 	case DMA_DEV_TO_MEM:
854 		chcr = RCAR_DMACHCR_DM_INC | RCAR_DMACHCR_SM_FIXED
855 		     | RCAR_DMACHCR_RS_DMARS;
856 		xfer_size = chan->src.xfer_size;
857 		break;
858 
859 	case DMA_MEM_TO_DEV:
860 		chcr = RCAR_DMACHCR_DM_FIXED | RCAR_DMACHCR_SM_INC
861 		     | RCAR_DMACHCR_RS_DMARS;
862 		xfer_size = chan->dst.xfer_size;
863 		break;
864 
865 	case DMA_MEM_TO_MEM:
866 	default:
867 		chcr = RCAR_DMACHCR_DM_INC | RCAR_DMACHCR_SM_INC
868 		     | RCAR_DMACHCR_RS_AUTO;
869 		xfer_size = RCAR_DMAC_MEMCPY_XFER_SIZE;
870 		break;
871 	}
872 
873 	desc->xfer_shift = ilog2(xfer_size);
874 	desc->chcr = chcr | chcr_ts[desc->xfer_shift];
875 }
876 
877 /*
878  * rcar_dmac_chan_prep_sg - prepare transfer descriptors from an SG list
879  *
880  * Common routine for public (MEMCPY) and slave DMA. The MEMCPY case is also
881  * converted to scatter-gather to guarantee consistent locking and a correct
882  * list manipulation. For slave DMA direction carries the usual meaning, and,
883  * logically, the SG list is RAM and the addr variable contains slave address,
884  * e.g., the FIFO I/O register. For MEMCPY direction equals DMA_MEM_TO_MEM
885  * and the SG list contains only one element and points at the source buffer.
886  */
887 static struct dma_async_tx_descriptor *
888 rcar_dmac_chan_prep_sg(struct rcar_dmac_chan *chan, struct scatterlist *sgl,
889 		       unsigned int sg_len, dma_addr_t dev_addr,
890 		       enum dma_transfer_direction dir, unsigned long dma_flags,
891 		       bool cyclic)
892 {
893 	struct rcar_dmac_xfer_chunk *chunk;
894 	struct rcar_dmac_desc *desc;
895 	struct scatterlist *sg;
896 	unsigned int nchunks = 0;
897 	unsigned int max_chunk_size;
898 	unsigned int full_size = 0;
899 	bool cross_boundary = false;
900 	unsigned int i;
901 #ifdef CONFIG_ARCH_DMA_ADDR_T_64BIT
902 	u32 high_dev_addr;
903 	u32 high_mem_addr;
904 #endif
905 
906 	desc = rcar_dmac_desc_get(chan);
907 	if (!desc)
908 		return NULL;
909 
910 	desc->async_tx.flags = dma_flags;
911 	desc->async_tx.cookie = -EBUSY;
912 
913 	desc->cyclic = cyclic;
914 	desc->direction = dir;
915 
916 	rcar_dmac_chan_configure_desc(chan, desc);
917 
918 	max_chunk_size = RCAR_DMATCR_MASK << desc->xfer_shift;
919 
920 	/*
921 	 * Allocate and fill the transfer chunk descriptors. We own the only
922 	 * reference to the DMA descriptor, there's no need for locking.
923 	 */
924 	for_each_sg(sgl, sg, sg_len, i) {
925 		dma_addr_t mem_addr = sg_dma_address(sg);
926 		unsigned int len = sg_dma_len(sg);
927 
928 		full_size += len;
929 
930 #ifdef CONFIG_ARCH_DMA_ADDR_T_64BIT
931 		if (i == 0) {
932 			high_dev_addr = dev_addr >> 32;
933 			high_mem_addr = mem_addr >> 32;
934 		}
935 
936 		if ((dev_addr >> 32 != high_dev_addr) ||
937 		    (mem_addr >> 32 != high_mem_addr))
938 			cross_boundary = true;
939 #endif
940 		while (len) {
941 			unsigned int size = min(len, max_chunk_size);
942 
943 #ifdef CONFIG_ARCH_DMA_ADDR_T_64BIT
944 			/*
945 			 * Prevent individual transfers from crossing 4GB
946 			 * boundaries.
947 			 */
948 			if (dev_addr >> 32 != (dev_addr + size - 1) >> 32) {
949 				size = ALIGN(dev_addr, 1ULL << 32) - dev_addr;
950 				cross_boundary = true;
951 			}
952 			if (mem_addr >> 32 != (mem_addr + size - 1) >> 32) {
953 				size = ALIGN(mem_addr, 1ULL << 32) - mem_addr;
954 				cross_boundary = true;
955 			}
956 #endif
957 
958 			chunk = rcar_dmac_xfer_chunk_get(chan);
959 			if (!chunk) {
960 				rcar_dmac_desc_put(chan, desc);
961 				return NULL;
962 			}
963 
964 			if (dir == DMA_DEV_TO_MEM) {
965 				chunk->src_addr = dev_addr;
966 				chunk->dst_addr = mem_addr;
967 			} else {
968 				chunk->src_addr = mem_addr;
969 				chunk->dst_addr = dev_addr;
970 			}
971 
972 			chunk->size = size;
973 
974 			dev_dbg(chan->chan.device->dev,
975 				"chan%u: chunk %p/%p sgl %u@%p, %u/%u %pad -> %pad\n",
976 				chan->index, chunk, desc, i, sg, size, len,
977 				&chunk->src_addr, &chunk->dst_addr);
978 
979 			mem_addr += size;
980 			if (dir == DMA_MEM_TO_MEM)
981 				dev_addr += size;
982 
983 			len -= size;
984 
985 			list_add_tail(&chunk->node, &desc->chunks);
986 			nchunks++;
987 		}
988 	}
989 
990 	desc->nchunks = nchunks;
991 	desc->size = full_size;
992 
993 	/*
994 	 * Use hardware descriptor lists if possible when more than one chunk
995 	 * needs to be transferred (otherwise they don't make much sense).
996 	 *
997 	 * Source/Destination address should be located in same 4GiB region
998 	 * in the 40bit address space when it uses Hardware descriptor,
999 	 * and cross_boundary is checking it.
1000 	 */
1001 	desc->hwdescs.use = !cross_boundary && nchunks > 1;
1002 	if (desc->hwdescs.use) {
1003 		if (rcar_dmac_fill_hwdesc(chan, desc) < 0)
1004 			desc->hwdescs.use = false;
1005 	}
1006 
1007 	return &desc->async_tx;
1008 }
1009 
1010 /* -----------------------------------------------------------------------------
1011  * DMA engine operations
1012  */
1013 
1014 static int rcar_dmac_alloc_chan_resources(struct dma_chan *chan)
1015 {
1016 	struct rcar_dmac_chan *rchan = to_rcar_dmac_chan(chan);
1017 	int ret;
1018 
1019 	INIT_LIST_HEAD(&rchan->desc.chunks_free);
1020 	INIT_LIST_HEAD(&rchan->desc.pages);
1021 
1022 	/* Preallocate descriptors. */
1023 	ret = rcar_dmac_xfer_chunk_alloc(rchan, GFP_KERNEL);
1024 	if (ret < 0)
1025 		return -ENOMEM;
1026 
1027 	ret = rcar_dmac_desc_alloc(rchan, GFP_KERNEL);
1028 	if (ret < 0)
1029 		return -ENOMEM;
1030 
1031 	return pm_runtime_get_sync(chan->device->dev);
1032 }
1033 
1034 static void rcar_dmac_free_chan_resources(struct dma_chan *chan)
1035 {
1036 	struct rcar_dmac_chan *rchan = to_rcar_dmac_chan(chan);
1037 	struct rcar_dmac *dmac = to_rcar_dmac(chan->device);
1038 	struct rcar_dmac_chan_map *map = &rchan->map;
1039 	struct rcar_dmac_desc_page *page, *_page;
1040 	struct rcar_dmac_desc *desc;
1041 	LIST_HEAD(list);
1042 
1043 	/* Protect against ISR */
1044 	spin_lock_irq(&rchan->lock);
1045 	rcar_dmac_chan_halt(rchan);
1046 	spin_unlock_irq(&rchan->lock);
1047 
1048 	/*
1049 	 * Now no new interrupts will occur, but one might already be
1050 	 * running. Wait for it to finish before freeing resources.
1051 	 */
1052 	synchronize_irq(rchan->irq);
1053 
1054 	if (rchan->mid_rid >= 0) {
1055 		/* The caller is holding dma_list_mutex */
1056 		clear_bit(rchan->mid_rid, dmac->modules);
1057 		rchan->mid_rid = -EINVAL;
1058 	}
1059 
1060 	list_splice_init(&rchan->desc.free, &list);
1061 	list_splice_init(&rchan->desc.pending, &list);
1062 	list_splice_init(&rchan->desc.active, &list);
1063 	list_splice_init(&rchan->desc.done, &list);
1064 	list_splice_init(&rchan->desc.wait, &list);
1065 
1066 	rchan->desc.running = NULL;
1067 
1068 	list_for_each_entry(desc, &list, node)
1069 		rcar_dmac_realloc_hwdesc(rchan, desc, 0);
1070 
1071 	list_for_each_entry_safe(page, _page, &rchan->desc.pages, node) {
1072 		list_del(&page->node);
1073 		free_page((unsigned long)page);
1074 	}
1075 
1076 	/* Remove slave mapping if present. */
1077 	if (map->slave.xfer_size) {
1078 		dma_unmap_resource(chan->device->dev, map->addr,
1079 				   map->slave.xfer_size, map->dir, 0);
1080 		map->slave.xfer_size = 0;
1081 	}
1082 
1083 	pm_runtime_put(chan->device->dev);
1084 }
1085 
1086 static struct dma_async_tx_descriptor *
1087 rcar_dmac_prep_dma_memcpy(struct dma_chan *chan, dma_addr_t dma_dest,
1088 			  dma_addr_t dma_src, size_t len, unsigned long flags)
1089 {
1090 	struct rcar_dmac_chan *rchan = to_rcar_dmac_chan(chan);
1091 	struct scatterlist sgl;
1092 
1093 	if (!len)
1094 		return NULL;
1095 
1096 	sg_init_table(&sgl, 1);
1097 	sg_set_page(&sgl, pfn_to_page(PFN_DOWN(dma_src)), len,
1098 		    offset_in_page(dma_src));
1099 	sg_dma_address(&sgl) = dma_src;
1100 	sg_dma_len(&sgl) = len;
1101 
1102 	return rcar_dmac_chan_prep_sg(rchan, &sgl, 1, dma_dest,
1103 				      DMA_MEM_TO_MEM, flags, false);
1104 }
1105 
1106 static int rcar_dmac_map_slave_addr(struct dma_chan *chan,
1107 				    enum dma_transfer_direction dir)
1108 {
1109 	struct rcar_dmac_chan *rchan = to_rcar_dmac_chan(chan);
1110 	struct rcar_dmac_chan_map *map = &rchan->map;
1111 	phys_addr_t dev_addr;
1112 	size_t dev_size;
1113 	enum dma_data_direction dev_dir;
1114 
1115 	if (dir == DMA_DEV_TO_MEM) {
1116 		dev_addr = rchan->src.slave_addr;
1117 		dev_size = rchan->src.xfer_size;
1118 		dev_dir = DMA_TO_DEVICE;
1119 	} else {
1120 		dev_addr = rchan->dst.slave_addr;
1121 		dev_size = rchan->dst.xfer_size;
1122 		dev_dir = DMA_FROM_DEVICE;
1123 	}
1124 
1125 	/* Reuse current map if possible. */
1126 	if (dev_addr == map->slave.slave_addr &&
1127 	    dev_size == map->slave.xfer_size &&
1128 	    dev_dir == map->dir)
1129 		return 0;
1130 
1131 	/* Remove old mapping if present. */
1132 	if (map->slave.xfer_size)
1133 		dma_unmap_resource(chan->device->dev, map->addr,
1134 				   map->slave.xfer_size, map->dir, 0);
1135 	map->slave.xfer_size = 0;
1136 
1137 	/* Create new slave address map. */
1138 	map->addr = dma_map_resource(chan->device->dev, dev_addr, dev_size,
1139 				     dev_dir, 0);
1140 
1141 	if (dma_mapping_error(chan->device->dev, map->addr)) {
1142 		dev_err(chan->device->dev,
1143 			"chan%u: failed to map %zx@%pap", rchan->index,
1144 			dev_size, &dev_addr);
1145 		return -EIO;
1146 	}
1147 
1148 	dev_dbg(chan->device->dev, "chan%u: map %zx@%pap to %pad dir: %s\n",
1149 		rchan->index, dev_size, &dev_addr, &map->addr,
1150 		dev_dir == DMA_TO_DEVICE ? "DMA_TO_DEVICE" : "DMA_FROM_DEVICE");
1151 
1152 	map->slave.slave_addr = dev_addr;
1153 	map->slave.xfer_size = dev_size;
1154 	map->dir = dev_dir;
1155 
1156 	return 0;
1157 }
1158 
1159 static struct dma_async_tx_descriptor *
1160 rcar_dmac_prep_slave_sg(struct dma_chan *chan, struct scatterlist *sgl,
1161 			unsigned int sg_len, enum dma_transfer_direction dir,
1162 			unsigned long flags, void *context)
1163 {
1164 	struct rcar_dmac_chan *rchan = to_rcar_dmac_chan(chan);
1165 
1166 	/* Someone calling slave DMA on a generic channel? */
1167 	if (rchan->mid_rid < 0 || !sg_len) {
1168 		dev_warn(chan->device->dev,
1169 			 "%s: bad parameter: len=%d, id=%d\n",
1170 			 __func__, sg_len, rchan->mid_rid);
1171 		return NULL;
1172 	}
1173 
1174 	if (rcar_dmac_map_slave_addr(chan, dir))
1175 		return NULL;
1176 
1177 	return rcar_dmac_chan_prep_sg(rchan, sgl, sg_len, rchan->map.addr,
1178 				      dir, flags, false);
1179 }
1180 
1181 #define RCAR_DMAC_MAX_SG_LEN	32
1182 
1183 static struct dma_async_tx_descriptor *
1184 rcar_dmac_prep_dma_cyclic(struct dma_chan *chan, dma_addr_t buf_addr,
1185 			  size_t buf_len, size_t period_len,
1186 			  enum dma_transfer_direction dir, unsigned long flags)
1187 {
1188 	struct rcar_dmac_chan *rchan = to_rcar_dmac_chan(chan);
1189 	struct dma_async_tx_descriptor *desc;
1190 	struct scatterlist *sgl;
1191 	unsigned int sg_len;
1192 	unsigned int i;
1193 
1194 	/* Someone calling slave DMA on a generic channel? */
1195 	if (rchan->mid_rid < 0 || buf_len < period_len) {
1196 		dev_warn(chan->device->dev,
1197 			"%s: bad parameter: buf_len=%zu, period_len=%zu, id=%d\n",
1198 			__func__, buf_len, period_len, rchan->mid_rid);
1199 		return NULL;
1200 	}
1201 
1202 	if (rcar_dmac_map_slave_addr(chan, dir))
1203 		return NULL;
1204 
1205 	sg_len = buf_len / period_len;
1206 	if (sg_len > RCAR_DMAC_MAX_SG_LEN) {
1207 		dev_err(chan->device->dev,
1208 			"chan%u: sg length %d exceds limit %d",
1209 			rchan->index, sg_len, RCAR_DMAC_MAX_SG_LEN);
1210 		return NULL;
1211 	}
1212 
1213 	/*
1214 	 * Allocate the sg list dynamically as it would consume too much stack
1215 	 * space.
1216 	 */
1217 	sgl = kcalloc(sg_len, sizeof(*sgl), GFP_NOWAIT);
1218 	if (!sgl)
1219 		return NULL;
1220 
1221 	sg_init_table(sgl, sg_len);
1222 
1223 	for (i = 0; i < sg_len; ++i) {
1224 		dma_addr_t src = buf_addr + (period_len * i);
1225 
1226 		sg_set_page(&sgl[i], pfn_to_page(PFN_DOWN(src)), period_len,
1227 			    offset_in_page(src));
1228 		sg_dma_address(&sgl[i]) = src;
1229 		sg_dma_len(&sgl[i]) = period_len;
1230 	}
1231 
1232 	desc = rcar_dmac_chan_prep_sg(rchan, sgl, sg_len, rchan->map.addr,
1233 				      dir, flags, true);
1234 
1235 	kfree(sgl);
1236 	return desc;
1237 }
1238 
1239 static int rcar_dmac_device_config(struct dma_chan *chan,
1240 				   struct dma_slave_config *cfg)
1241 {
1242 	struct rcar_dmac_chan *rchan = to_rcar_dmac_chan(chan);
1243 
1244 	/*
1245 	 * We could lock this, but you shouldn't be configuring the
1246 	 * channel, while using it...
1247 	 */
1248 	rchan->src.slave_addr = cfg->src_addr;
1249 	rchan->dst.slave_addr = cfg->dst_addr;
1250 	rchan->src.xfer_size = cfg->src_addr_width;
1251 	rchan->dst.xfer_size = cfg->dst_addr_width;
1252 
1253 	return 0;
1254 }
1255 
1256 static int rcar_dmac_chan_terminate_all(struct dma_chan *chan)
1257 {
1258 	struct rcar_dmac_chan *rchan = to_rcar_dmac_chan(chan);
1259 	unsigned long flags;
1260 
1261 	spin_lock_irqsave(&rchan->lock, flags);
1262 	rcar_dmac_chan_halt(rchan);
1263 	spin_unlock_irqrestore(&rchan->lock, flags);
1264 
1265 	/*
1266 	 * FIXME: No new interrupt can occur now, but the IRQ thread might still
1267 	 * be running.
1268 	 */
1269 
1270 	rcar_dmac_chan_reinit(rchan);
1271 
1272 	return 0;
1273 }
1274 
1275 static unsigned int rcar_dmac_chan_get_residue(struct rcar_dmac_chan *chan,
1276 					       dma_cookie_t cookie)
1277 {
1278 	struct rcar_dmac_desc *desc = chan->desc.running;
1279 	struct rcar_dmac_xfer_chunk *running = NULL;
1280 	struct rcar_dmac_xfer_chunk *chunk;
1281 	enum dma_status status;
1282 	unsigned int residue = 0;
1283 	unsigned int dptr = 0;
1284 
1285 	if (!desc)
1286 		return 0;
1287 
1288 	/*
1289 	 * If the cookie corresponds to a descriptor that has been completed
1290 	 * there is no residue. The same check has already been performed by the
1291 	 * caller but without holding the channel lock, so the descriptor could
1292 	 * now be complete.
1293 	 */
1294 	status = dma_cookie_status(&chan->chan, cookie, NULL);
1295 	if (status == DMA_COMPLETE)
1296 		return 0;
1297 
1298 	/*
1299 	 * If the cookie doesn't correspond to the currently running transfer
1300 	 * then the descriptor hasn't been processed yet, and the residue is
1301 	 * equal to the full descriptor size.
1302 	 * Also, a client driver is possible to call this function before
1303 	 * rcar_dmac_isr_channel_thread() runs. In this case, the "desc.running"
1304 	 * will be the next descriptor, and the done list will appear. So, if
1305 	 * the argument cookie matches the done list's cookie, we can assume
1306 	 * the residue is zero.
1307 	 */
1308 	if (cookie != desc->async_tx.cookie) {
1309 		list_for_each_entry(desc, &chan->desc.done, node) {
1310 			if (cookie == desc->async_tx.cookie)
1311 				return 0;
1312 		}
1313 		list_for_each_entry(desc, &chan->desc.pending, node) {
1314 			if (cookie == desc->async_tx.cookie)
1315 				return desc->size;
1316 		}
1317 		list_for_each_entry(desc, &chan->desc.active, node) {
1318 			if (cookie == desc->async_tx.cookie)
1319 				return desc->size;
1320 		}
1321 
1322 		/*
1323 		 * No descriptor found for the cookie, there's thus no residue.
1324 		 * This shouldn't happen if the calling driver passes a correct
1325 		 * cookie value.
1326 		 */
1327 		WARN(1, "No descriptor for cookie!");
1328 		return 0;
1329 	}
1330 
1331 	/*
1332 	 * In descriptor mode the descriptor running pointer is not maintained
1333 	 * by the interrupt handler, find the running descriptor from the
1334 	 * descriptor pointer field in the CHCRB register. In non-descriptor
1335 	 * mode just use the running descriptor pointer.
1336 	 */
1337 	if (desc->hwdescs.use) {
1338 		dptr = (rcar_dmac_chan_read(chan, RCAR_DMACHCRB) &
1339 			RCAR_DMACHCRB_DPTR_MASK) >> RCAR_DMACHCRB_DPTR_SHIFT;
1340 		if (dptr == 0)
1341 			dptr = desc->nchunks;
1342 		dptr--;
1343 		WARN_ON(dptr >= desc->nchunks);
1344 	} else {
1345 		running = desc->running;
1346 	}
1347 
1348 	/* Compute the size of all chunks still to be transferred. */
1349 	list_for_each_entry_reverse(chunk, &desc->chunks, node) {
1350 		if (chunk == running || ++dptr == desc->nchunks)
1351 			break;
1352 
1353 		residue += chunk->size;
1354 	}
1355 
1356 	/* Add the residue for the current chunk. */
1357 	residue += rcar_dmac_chan_read(chan, RCAR_DMATCRB) << desc->xfer_shift;
1358 
1359 	return residue;
1360 }
1361 
1362 static enum dma_status rcar_dmac_tx_status(struct dma_chan *chan,
1363 					   dma_cookie_t cookie,
1364 					   struct dma_tx_state *txstate)
1365 {
1366 	struct rcar_dmac_chan *rchan = to_rcar_dmac_chan(chan);
1367 	enum dma_status status;
1368 	unsigned long flags;
1369 	unsigned int residue;
1370 
1371 	status = dma_cookie_status(chan, cookie, txstate);
1372 	if (status == DMA_COMPLETE || !txstate)
1373 		return status;
1374 
1375 	spin_lock_irqsave(&rchan->lock, flags);
1376 	residue = rcar_dmac_chan_get_residue(rchan, cookie);
1377 	spin_unlock_irqrestore(&rchan->lock, flags);
1378 
1379 	/* if there's no residue, the cookie is complete */
1380 	if (!residue)
1381 		return DMA_COMPLETE;
1382 
1383 	dma_set_residue(txstate, residue);
1384 
1385 	return status;
1386 }
1387 
1388 static void rcar_dmac_issue_pending(struct dma_chan *chan)
1389 {
1390 	struct rcar_dmac_chan *rchan = to_rcar_dmac_chan(chan);
1391 	unsigned long flags;
1392 
1393 	spin_lock_irqsave(&rchan->lock, flags);
1394 
1395 	if (list_empty(&rchan->desc.pending))
1396 		goto done;
1397 
1398 	/* Append the pending list to the active list. */
1399 	list_splice_tail_init(&rchan->desc.pending, &rchan->desc.active);
1400 
1401 	/*
1402 	 * If no transfer is running pick the first descriptor from the active
1403 	 * list and start the transfer.
1404 	 */
1405 	if (!rchan->desc.running) {
1406 		struct rcar_dmac_desc *desc;
1407 
1408 		desc = list_first_entry(&rchan->desc.active,
1409 					struct rcar_dmac_desc, node);
1410 		rchan->desc.running = desc;
1411 
1412 		rcar_dmac_chan_start_xfer(rchan);
1413 	}
1414 
1415 done:
1416 	spin_unlock_irqrestore(&rchan->lock, flags);
1417 }
1418 
1419 static void rcar_dmac_device_synchronize(struct dma_chan *chan)
1420 {
1421 	struct rcar_dmac_chan *rchan = to_rcar_dmac_chan(chan);
1422 
1423 	synchronize_irq(rchan->irq);
1424 }
1425 
1426 /* -----------------------------------------------------------------------------
1427  * IRQ handling
1428  */
1429 
1430 static irqreturn_t rcar_dmac_isr_desc_stage_end(struct rcar_dmac_chan *chan)
1431 {
1432 	struct rcar_dmac_desc *desc = chan->desc.running;
1433 	unsigned int stage;
1434 
1435 	if (WARN_ON(!desc || !desc->cyclic)) {
1436 		/*
1437 		 * This should never happen, there should always be a running
1438 		 * cyclic descriptor when a descriptor stage end interrupt is
1439 		 * triggered. Warn and return.
1440 		 */
1441 		return IRQ_NONE;
1442 	}
1443 
1444 	/* Program the interrupt pointer to the next stage. */
1445 	stage = (rcar_dmac_chan_read(chan, RCAR_DMACHCRB) &
1446 		 RCAR_DMACHCRB_DPTR_MASK) >> RCAR_DMACHCRB_DPTR_SHIFT;
1447 	rcar_dmac_chan_write(chan, RCAR_DMADPCR, RCAR_DMADPCR_DIPT(stage));
1448 
1449 	return IRQ_WAKE_THREAD;
1450 }
1451 
1452 static irqreturn_t rcar_dmac_isr_transfer_end(struct rcar_dmac_chan *chan)
1453 {
1454 	struct rcar_dmac_desc *desc = chan->desc.running;
1455 	irqreturn_t ret = IRQ_WAKE_THREAD;
1456 
1457 	if (WARN_ON_ONCE(!desc)) {
1458 		/*
1459 		 * This should never happen, there should always be a running
1460 		 * descriptor when a transfer end interrupt is triggered. Warn
1461 		 * and return.
1462 		 */
1463 		return IRQ_NONE;
1464 	}
1465 
1466 	/*
1467 	 * The transfer end interrupt isn't generated for each chunk when using
1468 	 * descriptor mode. Only update the running chunk pointer in
1469 	 * non-descriptor mode.
1470 	 */
1471 	if (!desc->hwdescs.use) {
1472 		/*
1473 		 * If we haven't completed the last transfer chunk simply move
1474 		 * to the next one. Only wake the IRQ thread if the transfer is
1475 		 * cyclic.
1476 		 */
1477 		if (!list_is_last(&desc->running->node, &desc->chunks)) {
1478 			desc->running = list_next_entry(desc->running, node);
1479 			if (!desc->cyclic)
1480 				ret = IRQ_HANDLED;
1481 			goto done;
1482 		}
1483 
1484 		/*
1485 		 * We've completed the last transfer chunk. If the transfer is
1486 		 * cyclic, move back to the first one.
1487 		 */
1488 		if (desc->cyclic) {
1489 			desc->running =
1490 				list_first_entry(&desc->chunks,
1491 						 struct rcar_dmac_xfer_chunk,
1492 						 node);
1493 			goto done;
1494 		}
1495 	}
1496 
1497 	/* The descriptor is complete, move it to the done list. */
1498 	list_move_tail(&desc->node, &chan->desc.done);
1499 
1500 	/* Queue the next descriptor, if any. */
1501 	if (!list_empty(&chan->desc.active))
1502 		chan->desc.running = list_first_entry(&chan->desc.active,
1503 						      struct rcar_dmac_desc,
1504 						      node);
1505 	else
1506 		chan->desc.running = NULL;
1507 
1508 done:
1509 	if (chan->desc.running)
1510 		rcar_dmac_chan_start_xfer(chan);
1511 
1512 	return ret;
1513 }
1514 
1515 static irqreturn_t rcar_dmac_isr_channel(int irq, void *dev)
1516 {
1517 	u32 mask = RCAR_DMACHCR_DSE | RCAR_DMACHCR_TE;
1518 	struct rcar_dmac_chan *chan = dev;
1519 	irqreturn_t ret = IRQ_NONE;
1520 	bool reinit = false;
1521 	u32 chcr;
1522 
1523 	spin_lock(&chan->lock);
1524 
1525 	chcr = rcar_dmac_chan_read(chan, RCAR_DMACHCR);
1526 	if (chcr & RCAR_DMACHCR_CAE) {
1527 		struct rcar_dmac *dmac = to_rcar_dmac(chan->chan.device);
1528 
1529 		/*
1530 		 * We don't need to call rcar_dmac_chan_halt()
1531 		 * because channel is already stopped in error case.
1532 		 * We need to clear register and check DE bit as recovery.
1533 		 */
1534 		rcar_dmac_write(dmac, RCAR_DMACHCLR, 1 << chan->index);
1535 		rcar_dmac_chcr_de_barrier(chan);
1536 		reinit = true;
1537 		goto spin_lock_end;
1538 	}
1539 
1540 	if (chcr & RCAR_DMACHCR_TE)
1541 		mask |= RCAR_DMACHCR_DE;
1542 	rcar_dmac_chan_write(chan, RCAR_DMACHCR, chcr & ~mask);
1543 	if (mask & RCAR_DMACHCR_DE)
1544 		rcar_dmac_chcr_de_barrier(chan);
1545 
1546 	if (chcr & RCAR_DMACHCR_DSE)
1547 		ret |= rcar_dmac_isr_desc_stage_end(chan);
1548 
1549 	if (chcr & RCAR_DMACHCR_TE)
1550 		ret |= rcar_dmac_isr_transfer_end(chan);
1551 
1552 spin_lock_end:
1553 	spin_unlock(&chan->lock);
1554 
1555 	if (reinit) {
1556 		dev_err(chan->chan.device->dev, "Channel Address Error\n");
1557 
1558 		rcar_dmac_chan_reinit(chan);
1559 		ret = IRQ_HANDLED;
1560 	}
1561 
1562 	return ret;
1563 }
1564 
1565 static irqreturn_t rcar_dmac_isr_channel_thread(int irq, void *dev)
1566 {
1567 	struct rcar_dmac_chan *chan = dev;
1568 	struct rcar_dmac_desc *desc;
1569 	struct dmaengine_desc_callback cb;
1570 
1571 	spin_lock_irq(&chan->lock);
1572 
1573 	/* For cyclic transfers notify the user after every chunk. */
1574 	if (chan->desc.running && chan->desc.running->cyclic) {
1575 		desc = chan->desc.running;
1576 		dmaengine_desc_get_callback(&desc->async_tx, &cb);
1577 
1578 		if (dmaengine_desc_callback_valid(&cb)) {
1579 			spin_unlock_irq(&chan->lock);
1580 			dmaengine_desc_callback_invoke(&cb, NULL);
1581 			spin_lock_irq(&chan->lock);
1582 		}
1583 	}
1584 
1585 	/*
1586 	 * Call the callback function for all descriptors on the done list and
1587 	 * move them to the ack wait list.
1588 	 */
1589 	while (!list_empty(&chan->desc.done)) {
1590 		desc = list_first_entry(&chan->desc.done, struct rcar_dmac_desc,
1591 					node);
1592 		dma_cookie_complete(&desc->async_tx);
1593 		list_del(&desc->node);
1594 
1595 		dmaengine_desc_get_callback(&desc->async_tx, &cb);
1596 		if (dmaengine_desc_callback_valid(&cb)) {
1597 			spin_unlock_irq(&chan->lock);
1598 			/*
1599 			 * We own the only reference to this descriptor, we can
1600 			 * safely dereference it without holding the channel
1601 			 * lock.
1602 			 */
1603 			dmaengine_desc_callback_invoke(&cb, NULL);
1604 			spin_lock_irq(&chan->lock);
1605 		}
1606 
1607 		list_add_tail(&desc->node, &chan->desc.wait);
1608 	}
1609 
1610 	spin_unlock_irq(&chan->lock);
1611 
1612 	/* Recycle all acked descriptors. */
1613 	rcar_dmac_desc_recycle_acked(chan);
1614 
1615 	return IRQ_HANDLED;
1616 }
1617 
1618 /* -----------------------------------------------------------------------------
1619  * OF xlate and channel filter
1620  */
1621 
1622 static bool rcar_dmac_chan_filter(struct dma_chan *chan, void *arg)
1623 {
1624 	struct rcar_dmac *dmac = to_rcar_dmac(chan->device);
1625 	struct of_phandle_args *dma_spec = arg;
1626 
1627 	/*
1628 	 * FIXME: Using a filter on OF platforms is a nonsense. The OF xlate
1629 	 * function knows from which device it wants to allocate a channel from,
1630 	 * and would be perfectly capable of selecting the channel it wants.
1631 	 * Forcing it to call dma_request_channel() and iterate through all
1632 	 * channels from all controllers is just pointless.
1633 	 */
1634 	if (chan->device->device_config != rcar_dmac_device_config ||
1635 	    dma_spec->np != chan->device->dev->of_node)
1636 		return false;
1637 
1638 	return !test_and_set_bit(dma_spec->args[0], dmac->modules);
1639 }
1640 
1641 static struct dma_chan *rcar_dmac_of_xlate(struct of_phandle_args *dma_spec,
1642 					   struct of_dma *ofdma)
1643 {
1644 	struct rcar_dmac_chan *rchan;
1645 	struct dma_chan *chan;
1646 	dma_cap_mask_t mask;
1647 
1648 	if (dma_spec->args_count != 1)
1649 		return NULL;
1650 
1651 	/* Only slave DMA channels can be allocated via DT */
1652 	dma_cap_zero(mask);
1653 	dma_cap_set(DMA_SLAVE, mask);
1654 
1655 	chan = dma_request_channel(mask, rcar_dmac_chan_filter, dma_spec);
1656 	if (!chan)
1657 		return NULL;
1658 
1659 	rchan = to_rcar_dmac_chan(chan);
1660 	rchan->mid_rid = dma_spec->args[0];
1661 
1662 	return chan;
1663 }
1664 
1665 /* -----------------------------------------------------------------------------
1666  * Power management
1667  */
1668 
1669 #ifdef CONFIG_PM
1670 static int rcar_dmac_runtime_suspend(struct device *dev)
1671 {
1672 	return 0;
1673 }
1674 
1675 static int rcar_dmac_runtime_resume(struct device *dev)
1676 {
1677 	struct rcar_dmac *dmac = dev_get_drvdata(dev);
1678 
1679 	return rcar_dmac_init(dmac);
1680 }
1681 #endif
1682 
1683 static const struct dev_pm_ops rcar_dmac_pm = {
1684 	/*
1685 	 * TODO for system sleep/resume:
1686 	 *   - Wait for the current transfer to complete and stop the device,
1687 	 *   - Resume transfers, if any.
1688 	 */
1689 	SET_NOIRQ_SYSTEM_SLEEP_PM_OPS(pm_runtime_force_suspend,
1690 				      pm_runtime_force_resume)
1691 	SET_RUNTIME_PM_OPS(rcar_dmac_runtime_suspend, rcar_dmac_runtime_resume,
1692 			   NULL)
1693 };
1694 
1695 /* -----------------------------------------------------------------------------
1696  * Probe and remove
1697  */
1698 
1699 static int rcar_dmac_chan_probe(struct rcar_dmac *dmac,
1700 				struct rcar_dmac_chan *rchan,
1701 				unsigned int index)
1702 {
1703 	struct platform_device *pdev = to_platform_device(dmac->dev);
1704 	struct dma_chan *chan = &rchan->chan;
1705 	char pdev_irqname[5];
1706 	char *irqname;
1707 	int ret;
1708 
1709 	rchan->index = index;
1710 	rchan->iomem = dmac->iomem + RCAR_DMAC_CHAN_OFFSET(index);
1711 	rchan->mid_rid = -EINVAL;
1712 
1713 	spin_lock_init(&rchan->lock);
1714 
1715 	INIT_LIST_HEAD(&rchan->desc.free);
1716 	INIT_LIST_HEAD(&rchan->desc.pending);
1717 	INIT_LIST_HEAD(&rchan->desc.active);
1718 	INIT_LIST_HEAD(&rchan->desc.done);
1719 	INIT_LIST_HEAD(&rchan->desc.wait);
1720 
1721 	/* Request the channel interrupt. */
1722 	sprintf(pdev_irqname, "ch%u", index);
1723 	rchan->irq = platform_get_irq_byname(pdev, pdev_irqname);
1724 	if (rchan->irq < 0) {
1725 		dev_err(dmac->dev, "no IRQ specified for channel %u\n", index);
1726 		return -ENODEV;
1727 	}
1728 
1729 	irqname = devm_kasprintf(dmac->dev, GFP_KERNEL, "%s:%u",
1730 				 dev_name(dmac->dev), index);
1731 	if (!irqname)
1732 		return -ENOMEM;
1733 
1734 	/*
1735 	 * Initialize the DMA engine channel and add it to the DMA engine
1736 	 * channels list.
1737 	 */
1738 	chan->device = &dmac->engine;
1739 	dma_cookie_init(chan);
1740 
1741 	list_add_tail(&chan->device_node, &dmac->engine.channels);
1742 
1743 	ret = devm_request_threaded_irq(dmac->dev, rchan->irq,
1744 					rcar_dmac_isr_channel,
1745 					rcar_dmac_isr_channel_thread, 0,
1746 					irqname, rchan);
1747 	if (ret) {
1748 		dev_err(dmac->dev, "failed to request IRQ %u (%d)\n",
1749 			rchan->irq, ret);
1750 		return ret;
1751 	}
1752 
1753 	return 0;
1754 }
1755 
1756 static int rcar_dmac_parse_of(struct device *dev, struct rcar_dmac *dmac)
1757 {
1758 	struct device_node *np = dev->of_node;
1759 	int ret;
1760 
1761 	ret = of_property_read_u32(np, "dma-channels", &dmac->n_channels);
1762 	if (ret < 0) {
1763 		dev_err(dev, "unable to read dma-channels property\n");
1764 		return ret;
1765 	}
1766 
1767 	if (dmac->n_channels <= 0 || dmac->n_channels >= 100) {
1768 		dev_err(dev, "invalid number of channels %u\n",
1769 			dmac->n_channels);
1770 		return -EINVAL;
1771 	}
1772 
1773 	return 0;
1774 }
1775 
1776 static int rcar_dmac_probe(struct platform_device *pdev)
1777 {
1778 	const enum dma_slave_buswidth widths = DMA_SLAVE_BUSWIDTH_1_BYTE |
1779 		DMA_SLAVE_BUSWIDTH_2_BYTES | DMA_SLAVE_BUSWIDTH_4_BYTES |
1780 		DMA_SLAVE_BUSWIDTH_8_BYTES | DMA_SLAVE_BUSWIDTH_16_BYTES |
1781 		DMA_SLAVE_BUSWIDTH_32_BYTES | DMA_SLAVE_BUSWIDTH_64_BYTES;
1782 	unsigned int channels_offset = 0;
1783 	struct dma_device *engine;
1784 	struct rcar_dmac *dmac;
1785 	struct resource *mem;
1786 	unsigned int i;
1787 	int ret;
1788 
1789 	dmac = devm_kzalloc(&pdev->dev, sizeof(*dmac), GFP_KERNEL);
1790 	if (!dmac)
1791 		return -ENOMEM;
1792 
1793 	dmac->dev = &pdev->dev;
1794 	platform_set_drvdata(pdev, dmac);
1795 	dma_set_mask_and_coherent(dmac->dev, DMA_BIT_MASK(40));
1796 
1797 	ret = rcar_dmac_parse_of(&pdev->dev, dmac);
1798 	if (ret < 0)
1799 		return ret;
1800 
1801 	/*
1802 	 * A still unconfirmed hardware bug prevents the IPMMU microTLB 0 to be
1803 	 * flushed correctly, resulting in memory corruption. DMAC 0 channel 0
1804 	 * is connected to microTLB 0 on currently supported platforms, so we
1805 	 * can't use it with the IPMMU. As the IOMMU API operates at the device
1806 	 * level we can't disable it selectively, so ignore channel 0 for now if
1807 	 * the device is part of an IOMMU group.
1808 	 */
1809 	if (pdev->dev.iommu_group) {
1810 		dmac->n_channels--;
1811 		channels_offset = 1;
1812 	}
1813 
1814 	dmac->channels = devm_kcalloc(&pdev->dev, dmac->n_channels,
1815 				      sizeof(*dmac->channels), GFP_KERNEL);
1816 	if (!dmac->channels)
1817 		return -ENOMEM;
1818 
1819 	/* Request resources. */
1820 	mem = platform_get_resource(pdev, IORESOURCE_MEM, 0);
1821 	dmac->iomem = devm_ioremap_resource(&pdev->dev, mem);
1822 	if (IS_ERR(dmac->iomem))
1823 		return PTR_ERR(dmac->iomem);
1824 
1825 	/* Enable runtime PM and initialize the device. */
1826 	pm_runtime_enable(&pdev->dev);
1827 	ret = pm_runtime_get_sync(&pdev->dev);
1828 	if (ret < 0) {
1829 		dev_err(&pdev->dev, "runtime PM get sync failed (%d)\n", ret);
1830 		return ret;
1831 	}
1832 
1833 	ret = rcar_dmac_init(dmac);
1834 	pm_runtime_put(&pdev->dev);
1835 
1836 	if (ret) {
1837 		dev_err(&pdev->dev, "failed to reset device\n");
1838 		goto error;
1839 	}
1840 
1841 	/* Initialize engine */
1842 	engine = &dmac->engine;
1843 
1844 	dma_cap_set(DMA_MEMCPY, engine->cap_mask);
1845 	dma_cap_set(DMA_SLAVE, engine->cap_mask);
1846 
1847 	engine->dev		= &pdev->dev;
1848 	engine->copy_align	= ilog2(RCAR_DMAC_MEMCPY_XFER_SIZE);
1849 
1850 	engine->src_addr_widths	= widths;
1851 	engine->dst_addr_widths	= widths;
1852 	engine->directions	= BIT(DMA_MEM_TO_DEV) | BIT(DMA_DEV_TO_MEM);
1853 	engine->residue_granularity = DMA_RESIDUE_GRANULARITY_BURST;
1854 
1855 	engine->device_alloc_chan_resources	= rcar_dmac_alloc_chan_resources;
1856 	engine->device_free_chan_resources	= rcar_dmac_free_chan_resources;
1857 	engine->device_prep_dma_memcpy		= rcar_dmac_prep_dma_memcpy;
1858 	engine->device_prep_slave_sg		= rcar_dmac_prep_slave_sg;
1859 	engine->device_prep_dma_cyclic		= rcar_dmac_prep_dma_cyclic;
1860 	engine->device_config			= rcar_dmac_device_config;
1861 	engine->device_pause			= rcar_dmac_chan_pause;
1862 	engine->device_terminate_all		= rcar_dmac_chan_terminate_all;
1863 	engine->device_tx_status		= rcar_dmac_tx_status;
1864 	engine->device_issue_pending		= rcar_dmac_issue_pending;
1865 	engine->device_synchronize		= rcar_dmac_device_synchronize;
1866 
1867 	INIT_LIST_HEAD(&engine->channels);
1868 
1869 	for (i = 0; i < dmac->n_channels; ++i) {
1870 		ret = rcar_dmac_chan_probe(dmac, &dmac->channels[i],
1871 					   i + channels_offset);
1872 		if (ret < 0)
1873 			goto error;
1874 	}
1875 
1876 	/* Register the DMAC as a DMA provider for DT. */
1877 	ret = of_dma_controller_register(pdev->dev.of_node, rcar_dmac_of_xlate,
1878 					 NULL);
1879 	if (ret < 0)
1880 		goto error;
1881 
1882 	/*
1883 	 * Register the DMA engine device.
1884 	 *
1885 	 * Default transfer size of 32 bytes requires 32-byte alignment.
1886 	 */
1887 	ret = dma_async_device_register(engine);
1888 	if (ret < 0)
1889 		goto error;
1890 
1891 	return 0;
1892 
1893 error:
1894 	of_dma_controller_free(pdev->dev.of_node);
1895 	pm_runtime_disable(&pdev->dev);
1896 	return ret;
1897 }
1898 
1899 static int rcar_dmac_remove(struct platform_device *pdev)
1900 {
1901 	struct rcar_dmac *dmac = platform_get_drvdata(pdev);
1902 
1903 	of_dma_controller_free(pdev->dev.of_node);
1904 	dma_async_device_unregister(&dmac->engine);
1905 
1906 	pm_runtime_disable(&pdev->dev);
1907 
1908 	return 0;
1909 }
1910 
1911 static void rcar_dmac_shutdown(struct platform_device *pdev)
1912 {
1913 	struct rcar_dmac *dmac = platform_get_drvdata(pdev);
1914 
1915 	rcar_dmac_stop_all_chan(dmac);
1916 }
1917 
1918 static const struct of_device_id rcar_dmac_of_ids[] = {
1919 	{ .compatible = "renesas,rcar-dmac", },
1920 	{ /* Sentinel */ }
1921 };
1922 MODULE_DEVICE_TABLE(of, rcar_dmac_of_ids);
1923 
1924 static struct platform_driver rcar_dmac_driver = {
1925 	.driver		= {
1926 		.pm	= &rcar_dmac_pm,
1927 		.name	= "rcar-dmac",
1928 		.of_match_table = rcar_dmac_of_ids,
1929 	},
1930 	.probe		= rcar_dmac_probe,
1931 	.remove		= rcar_dmac_remove,
1932 	.shutdown	= rcar_dmac_shutdown,
1933 };
1934 
1935 module_platform_driver(rcar_dmac_driver);
1936 
1937 MODULE_DESCRIPTION("R-Car Gen2 DMA Controller Driver");
1938 MODULE_AUTHOR("Laurent Pinchart <laurent.pinchart@ideasonboard.com>");
1939 MODULE_LICENSE("GPL v2");
1940