xref: /linux/drivers/dma/nbpfaxi.c (revision 4e0ae876f77bc01a7e77724dea57b4b82bd53244)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (C) 2013-2014 Renesas Electronics Europe Ltd.
4  * Author: Guennadi Liakhovetski <g.liakhovetski@gmx.de>
5  */
6 
7 #include <linux/bitmap.h>
8 #include <linux/bitops.h>
9 #include <linux/clk.h>
10 #include <linux/dma-mapping.h>
11 #include <linux/dmaengine.h>
12 #include <linux/err.h>
13 #include <linux/interrupt.h>
14 #include <linux/io.h>
15 #include <linux/log2.h>
16 #include <linux/module.h>
17 #include <linux/of.h>
18 #include <linux/of_device.h>
19 #include <linux/of_dma.h>
20 #include <linux/platform_device.h>
21 #include <linux/slab.h>
22 
23 #include <dt-bindings/dma/nbpfaxi.h>
24 
25 #include "dmaengine.h"
26 
27 #define NBPF_REG_CHAN_OFFSET	0
28 #define NBPF_REG_CHAN_SIZE	0x40
29 
30 /* Channel Current Transaction Byte register */
31 #define NBPF_CHAN_CUR_TR_BYTE	0x20
32 
33 /* Channel Status register */
34 #define NBPF_CHAN_STAT	0x24
35 #define NBPF_CHAN_STAT_EN	1
36 #define NBPF_CHAN_STAT_TACT	4
37 #define NBPF_CHAN_STAT_ERR	0x10
38 #define NBPF_CHAN_STAT_END	0x20
39 #define NBPF_CHAN_STAT_TC	0x40
40 #define NBPF_CHAN_STAT_DER	0x400
41 
42 /* Channel Control register */
43 #define NBPF_CHAN_CTRL	0x28
44 #define NBPF_CHAN_CTRL_SETEN	1
45 #define NBPF_CHAN_CTRL_CLREN	2
46 #define NBPF_CHAN_CTRL_STG	4
47 #define NBPF_CHAN_CTRL_SWRST	8
48 #define NBPF_CHAN_CTRL_CLRRQ	0x10
49 #define NBPF_CHAN_CTRL_CLREND	0x20
50 #define NBPF_CHAN_CTRL_CLRTC	0x40
51 #define NBPF_CHAN_CTRL_SETSUS	0x100
52 #define NBPF_CHAN_CTRL_CLRSUS	0x200
53 
54 /* Channel Configuration register */
55 #define NBPF_CHAN_CFG	0x2c
56 #define NBPF_CHAN_CFG_SEL	7		/* terminal SELect: 0..7 */
57 #define NBPF_CHAN_CFG_REQD	8		/* REQuest Direction: DMAREQ is 0: input, 1: output */
58 #define NBPF_CHAN_CFG_LOEN	0x10		/* LOw ENable: low DMA request line is: 0: inactive, 1: active */
59 #define NBPF_CHAN_CFG_HIEN	0x20		/* HIgh ENable: high DMA request line is: 0: inactive, 1: active */
60 #define NBPF_CHAN_CFG_LVL	0x40		/* LeVeL: DMA request line is sensed as 0: edge, 1: level */
61 #define NBPF_CHAN_CFG_AM	0x700		/* ACK Mode: 0: Pulse mode, 1: Level mode, b'1x: Bus Cycle */
62 #define NBPF_CHAN_CFG_SDS	0xf000		/* Source Data Size: 0: 8 bits,... , 7: 1024 bits */
63 #define NBPF_CHAN_CFG_DDS	0xf0000		/* Destination Data Size: as above */
64 #define NBPF_CHAN_CFG_SAD	0x100000	/* Source ADdress counting: 0: increment, 1: fixed */
65 #define NBPF_CHAN_CFG_DAD	0x200000	/* Destination ADdress counting: 0: increment, 1: fixed */
66 #define NBPF_CHAN_CFG_TM	0x400000	/* Transfer Mode: 0: single, 1: block TM */
67 #define NBPF_CHAN_CFG_DEM	0x1000000	/* DMAEND interrupt Mask */
68 #define NBPF_CHAN_CFG_TCM	0x2000000	/* DMATCO interrupt Mask */
69 #define NBPF_CHAN_CFG_SBE	0x8000000	/* Sweep Buffer Enable */
70 #define NBPF_CHAN_CFG_RSEL	0x10000000	/* RM: Register Set sELect */
71 #define NBPF_CHAN_CFG_RSW	0x20000000	/* RM: Register Select sWitch */
72 #define NBPF_CHAN_CFG_REN	0x40000000	/* RM: Register Set Enable */
73 #define NBPF_CHAN_CFG_DMS	0x80000000	/* 0: register mode (RM), 1: link mode (LM) */
74 
75 #define NBPF_CHAN_NXLA	0x38
76 #define NBPF_CHAN_CRLA	0x3c
77 
78 /* Link Header field */
79 #define NBPF_HEADER_LV	1
80 #define NBPF_HEADER_LE	2
81 #define NBPF_HEADER_WBD	4
82 #define NBPF_HEADER_DIM	8
83 
84 #define NBPF_CTRL	0x300
85 #define NBPF_CTRL_PR	1		/* 0: fixed priority, 1: round robin */
86 #define NBPF_CTRL_LVINT	2		/* DMAEND and DMAERR signalling: 0: pulse, 1: level */
87 
88 #define NBPF_DSTAT_ER	0x314
89 #define NBPF_DSTAT_END	0x318
90 
91 #define NBPF_DMA_BUSWIDTHS \
92 	(BIT(DMA_SLAVE_BUSWIDTH_UNDEFINED) | \
93 	 BIT(DMA_SLAVE_BUSWIDTH_1_BYTE) | \
94 	 BIT(DMA_SLAVE_BUSWIDTH_2_BYTES) | \
95 	 BIT(DMA_SLAVE_BUSWIDTH_4_BYTES) | \
96 	 BIT(DMA_SLAVE_BUSWIDTH_8_BYTES))
97 
98 struct nbpf_config {
99 	int num_channels;
100 	int buffer_size;
101 };
102 
103 /*
104  * We've got 3 types of objects, used to describe DMA transfers:
105  * 1. high-level descriptor, containing a struct dma_async_tx_descriptor object
106  *	in it, used to communicate with the user
107  * 2. hardware DMA link descriptors, that we pass to DMAC for DMA transfer
108  *	queuing, these must be DMAable, using either the streaming DMA API or
109  *	allocated from coherent memory - one per SG segment
110  * 3. one per SG segment descriptors, used to manage HW link descriptors from
111  *	(2). They do not have to be DMAable. They can either be (a) allocated
112  *	together with link descriptors as mixed (DMA / CPU) objects, or (b)
113  *	separately. Even if allocated separately it would be best to link them
114  *	to link descriptors once during channel resource allocation and always
115  *	use them as a single object.
116  * Therefore for both cases (a) and (b) at run-time objects (2) and (3) shall be
117  * treated as a single SG segment descriptor.
118  */
119 
120 struct nbpf_link_reg {
121 	u32	header;
122 	u32	src_addr;
123 	u32	dst_addr;
124 	u32	transaction_size;
125 	u32	config;
126 	u32	interval;
127 	u32	extension;
128 	u32	next;
129 } __packed;
130 
131 struct nbpf_device;
132 struct nbpf_channel;
133 struct nbpf_desc;
134 
135 struct nbpf_link_desc {
136 	struct nbpf_link_reg *hwdesc;
137 	dma_addr_t hwdesc_dma_addr;
138 	struct nbpf_desc *desc;
139 	struct list_head node;
140 };
141 
142 /**
143  * struct nbpf_desc - DMA transfer descriptor
144  * @async_tx:	dmaengine object
145  * @user_wait:	waiting for a user ack
146  * @length:	total transfer length
147  * @sg:		list of hardware descriptors, represented by struct nbpf_link_desc
148  * @node:	member in channel descriptor lists
149  */
150 struct nbpf_desc {
151 	struct dma_async_tx_descriptor async_tx;
152 	bool user_wait;
153 	size_t length;
154 	struct nbpf_channel *chan;
155 	struct list_head sg;
156 	struct list_head node;
157 };
158 
159 /* Take a wild guess: allocate 4 segments per descriptor */
160 #define NBPF_SEGMENTS_PER_DESC 4
161 #define NBPF_DESCS_PER_PAGE ((PAGE_SIZE - sizeof(struct list_head)) /	\
162 	(sizeof(struct nbpf_desc) +					\
163 	 NBPF_SEGMENTS_PER_DESC *					\
164 	 (sizeof(struct nbpf_link_desc) + sizeof(struct nbpf_link_reg))))
165 #define NBPF_SEGMENTS_PER_PAGE (NBPF_SEGMENTS_PER_DESC * NBPF_DESCS_PER_PAGE)
166 
167 struct nbpf_desc_page {
168 	struct list_head node;
169 	struct nbpf_desc desc[NBPF_DESCS_PER_PAGE];
170 	struct nbpf_link_desc ldesc[NBPF_SEGMENTS_PER_PAGE];
171 	struct nbpf_link_reg hwdesc[NBPF_SEGMENTS_PER_PAGE];
172 };
173 
174 /**
175  * struct nbpf_channel - one DMAC channel
176  * @dma_chan:	standard dmaengine channel object
177  * @base:	register address base
178  * @nbpf:	DMAC
179  * @name:	IRQ name
180  * @irq:	IRQ number
181  * @slave_addr:	address for slave DMA
182  * @slave_width:slave data size in bytes
183  * @slave_burst:maximum slave burst size in bytes
184  * @terminal:	DMA terminal, assigned to this channel
185  * @dmarq_cfg:	DMA request line configuration - high / low, edge / level for NBPF_CHAN_CFG
186  * @flags:	configuration flags from DT
187  * @lock:	protect descriptor lists
188  * @free_links:	list of free link descriptors
189  * @free:	list of free descriptors
190  * @queued:	list of queued descriptors
191  * @active:	list of descriptors, scheduled for processing
192  * @done:	list of completed descriptors, waiting post-processing
193  * @desc_page:	list of additionally allocated descriptor pages - if any
194  */
195 struct nbpf_channel {
196 	struct dma_chan dma_chan;
197 	struct tasklet_struct tasklet;
198 	void __iomem *base;
199 	struct nbpf_device *nbpf;
200 	char name[16];
201 	int irq;
202 	dma_addr_t slave_src_addr;
203 	size_t slave_src_width;
204 	size_t slave_src_burst;
205 	dma_addr_t slave_dst_addr;
206 	size_t slave_dst_width;
207 	size_t slave_dst_burst;
208 	unsigned int terminal;
209 	u32 dmarq_cfg;
210 	unsigned long flags;
211 	spinlock_t lock;
212 	struct list_head free_links;
213 	struct list_head free;
214 	struct list_head queued;
215 	struct list_head active;
216 	struct list_head done;
217 	struct list_head desc_page;
218 	struct nbpf_desc *running;
219 	bool paused;
220 };
221 
222 struct nbpf_device {
223 	struct dma_device dma_dev;
224 	void __iomem *base;
225 	u32 max_burst_mem_read;
226 	u32 max_burst_mem_write;
227 	struct clk *clk;
228 	const struct nbpf_config *config;
229 	unsigned int eirq;
230 	struct nbpf_channel chan[];
231 };
232 
233 enum nbpf_model {
234 	NBPF1B4,
235 	NBPF1B8,
236 	NBPF1B16,
237 	NBPF4B4,
238 	NBPF4B8,
239 	NBPF4B16,
240 	NBPF8B4,
241 	NBPF8B8,
242 	NBPF8B16,
243 };
244 
245 static struct nbpf_config nbpf_cfg[] = {
246 	[NBPF1B4] = {
247 		.num_channels = 1,
248 		.buffer_size = 4,
249 	},
250 	[NBPF1B8] = {
251 		.num_channels = 1,
252 		.buffer_size = 8,
253 	},
254 	[NBPF1B16] = {
255 		.num_channels = 1,
256 		.buffer_size = 16,
257 	},
258 	[NBPF4B4] = {
259 		.num_channels = 4,
260 		.buffer_size = 4,
261 	},
262 	[NBPF4B8] = {
263 		.num_channels = 4,
264 		.buffer_size = 8,
265 	},
266 	[NBPF4B16] = {
267 		.num_channels = 4,
268 		.buffer_size = 16,
269 	},
270 	[NBPF8B4] = {
271 		.num_channels = 8,
272 		.buffer_size = 4,
273 	},
274 	[NBPF8B8] = {
275 		.num_channels = 8,
276 		.buffer_size = 8,
277 	},
278 	[NBPF8B16] = {
279 		.num_channels = 8,
280 		.buffer_size = 16,
281 	},
282 };
283 
284 #define nbpf_to_chan(d) container_of(d, struct nbpf_channel, dma_chan)
285 
286 /*
287  * dmaengine drivers seem to have a lot in common and instead of sharing more
288  * code, they reimplement those common algorithms independently. In this driver
289  * we try to separate the hardware-specific part from the (largely) generic
290  * part. This improves code readability and makes it possible in the future to
291  * reuse the generic code in form of a helper library. That generic code should
292  * be suitable for various DMA controllers, using transfer descriptors in RAM
293  * and pushing one SG list at a time to the DMA controller.
294  */
295 
296 /*		Hardware-specific part		*/
297 
298 static inline u32 nbpf_chan_read(struct nbpf_channel *chan,
299 				 unsigned int offset)
300 {
301 	u32 data = ioread32(chan->base + offset);
302 	dev_dbg(chan->dma_chan.device->dev, "%s(0x%p + 0x%x) = 0x%x\n",
303 		__func__, chan->base, offset, data);
304 	return data;
305 }
306 
307 static inline void nbpf_chan_write(struct nbpf_channel *chan,
308 				   unsigned int offset, u32 data)
309 {
310 	iowrite32(data, chan->base + offset);
311 	dev_dbg(chan->dma_chan.device->dev, "%s(0x%p + 0x%x) = 0x%x\n",
312 		__func__, chan->base, offset, data);
313 }
314 
315 static inline u32 nbpf_read(struct nbpf_device *nbpf,
316 			    unsigned int offset)
317 {
318 	u32 data = ioread32(nbpf->base + offset);
319 	dev_dbg(nbpf->dma_dev.dev, "%s(0x%p + 0x%x) = 0x%x\n",
320 		__func__, nbpf->base, offset, data);
321 	return data;
322 }
323 
324 static inline void nbpf_write(struct nbpf_device *nbpf,
325 			      unsigned int offset, u32 data)
326 {
327 	iowrite32(data, nbpf->base + offset);
328 	dev_dbg(nbpf->dma_dev.dev, "%s(0x%p + 0x%x) = 0x%x\n",
329 		__func__, nbpf->base, offset, data);
330 }
331 
332 static void nbpf_chan_halt(struct nbpf_channel *chan)
333 {
334 	nbpf_chan_write(chan, NBPF_CHAN_CTRL, NBPF_CHAN_CTRL_CLREN);
335 }
336 
337 static bool nbpf_status_get(struct nbpf_channel *chan)
338 {
339 	u32 status = nbpf_read(chan->nbpf, NBPF_DSTAT_END);
340 
341 	return status & BIT(chan - chan->nbpf->chan);
342 }
343 
344 static void nbpf_status_ack(struct nbpf_channel *chan)
345 {
346 	nbpf_chan_write(chan, NBPF_CHAN_CTRL, NBPF_CHAN_CTRL_CLREND);
347 }
348 
349 static u32 nbpf_error_get(struct nbpf_device *nbpf)
350 {
351 	return nbpf_read(nbpf, NBPF_DSTAT_ER);
352 }
353 
354 static struct nbpf_channel *nbpf_error_get_channel(struct nbpf_device *nbpf, u32 error)
355 {
356 	return nbpf->chan + __ffs(error);
357 }
358 
359 static void nbpf_error_clear(struct nbpf_channel *chan)
360 {
361 	u32 status;
362 	int i;
363 
364 	/* Stop the channel, make sure DMA has been aborted */
365 	nbpf_chan_halt(chan);
366 
367 	for (i = 1000; i; i--) {
368 		status = nbpf_chan_read(chan, NBPF_CHAN_STAT);
369 		if (!(status & NBPF_CHAN_STAT_TACT))
370 			break;
371 		cpu_relax();
372 	}
373 
374 	if (!i)
375 		dev_err(chan->dma_chan.device->dev,
376 			"%s(): abort timeout, channel status 0x%x\n", __func__, status);
377 
378 	nbpf_chan_write(chan, NBPF_CHAN_CTRL, NBPF_CHAN_CTRL_SWRST);
379 }
380 
381 static int nbpf_start(struct nbpf_desc *desc)
382 {
383 	struct nbpf_channel *chan = desc->chan;
384 	struct nbpf_link_desc *ldesc = list_first_entry(&desc->sg, struct nbpf_link_desc, node);
385 
386 	nbpf_chan_write(chan, NBPF_CHAN_NXLA, (u32)ldesc->hwdesc_dma_addr);
387 	nbpf_chan_write(chan, NBPF_CHAN_CTRL, NBPF_CHAN_CTRL_SETEN | NBPF_CHAN_CTRL_CLRSUS);
388 	chan->paused = false;
389 
390 	/* Software trigger MEMCPY - only MEMCPY uses the block mode */
391 	if (ldesc->hwdesc->config & NBPF_CHAN_CFG_TM)
392 		nbpf_chan_write(chan, NBPF_CHAN_CTRL, NBPF_CHAN_CTRL_STG);
393 
394 	dev_dbg(chan->nbpf->dma_dev.dev, "%s(): next 0x%x, cur 0x%x\n", __func__,
395 		nbpf_chan_read(chan, NBPF_CHAN_NXLA), nbpf_chan_read(chan, NBPF_CHAN_CRLA));
396 
397 	return 0;
398 }
399 
400 static void nbpf_chan_prepare(struct nbpf_channel *chan)
401 {
402 	chan->dmarq_cfg = (chan->flags & NBPF_SLAVE_RQ_HIGH ? NBPF_CHAN_CFG_HIEN : 0) |
403 		(chan->flags & NBPF_SLAVE_RQ_LOW ? NBPF_CHAN_CFG_LOEN : 0) |
404 		(chan->flags & NBPF_SLAVE_RQ_LEVEL ?
405 		 NBPF_CHAN_CFG_LVL | (NBPF_CHAN_CFG_AM & 0x200) : 0) |
406 		chan->terminal;
407 }
408 
409 static void nbpf_chan_prepare_default(struct nbpf_channel *chan)
410 {
411 	/* Don't output DMAACK */
412 	chan->dmarq_cfg = NBPF_CHAN_CFG_AM & 0x400;
413 	chan->terminal = 0;
414 	chan->flags = 0;
415 }
416 
417 static void nbpf_chan_configure(struct nbpf_channel *chan)
418 {
419 	/*
420 	 * We assume, that only the link mode and DMA request line configuration
421 	 * have to be set in the configuration register manually. Dynamic
422 	 * per-transfer configuration will be loaded from transfer descriptors.
423 	 */
424 	nbpf_chan_write(chan, NBPF_CHAN_CFG, NBPF_CHAN_CFG_DMS | chan->dmarq_cfg);
425 }
426 
427 static u32 nbpf_xfer_ds(struct nbpf_device *nbpf, size_t size,
428 			enum dma_transfer_direction direction)
429 {
430 	int max_burst = nbpf->config->buffer_size * 8;
431 
432 	if (nbpf->max_burst_mem_read || nbpf->max_burst_mem_write) {
433 		switch (direction) {
434 		case DMA_MEM_TO_MEM:
435 			max_burst = min_not_zero(nbpf->max_burst_mem_read,
436 						 nbpf->max_burst_mem_write);
437 			break;
438 		case DMA_MEM_TO_DEV:
439 			if (nbpf->max_burst_mem_read)
440 				max_burst = nbpf->max_burst_mem_read;
441 			break;
442 		case DMA_DEV_TO_MEM:
443 			if (nbpf->max_burst_mem_write)
444 				max_burst = nbpf->max_burst_mem_write;
445 			break;
446 		case DMA_DEV_TO_DEV:
447 		default:
448 			break;
449 		}
450 	}
451 
452 	/* Maximum supported bursts depend on the buffer size */
453 	return min_t(int, __ffs(size), ilog2(max_burst));
454 }
455 
456 static size_t nbpf_xfer_size(struct nbpf_device *nbpf,
457 			     enum dma_slave_buswidth width, u32 burst)
458 {
459 	size_t size;
460 
461 	if (!burst)
462 		burst = 1;
463 
464 	switch (width) {
465 	case DMA_SLAVE_BUSWIDTH_8_BYTES:
466 		size = 8 * burst;
467 		break;
468 
469 	case DMA_SLAVE_BUSWIDTH_4_BYTES:
470 		size = 4 * burst;
471 		break;
472 
473 	case DMA_SLAVE_BUSWIDTH_2_BYTES:
474 		size = 2 * burst;
475 		break;
476 
477 	default:
478 		pr_warn("%s(): invalid bus width %u\n", __func__, width);
479 		/* fall through */
480 	case DMA_SLAVE_BUSWIDTH_1_BYTE:
481 		size = burst;
482 	}
483 
484 	return nbpf_xfer_ds(nbpf, size, DMA_TRANS_NONE);
485 }
486 
487 /*
488  * We need a way to recognise slaves, whose data is sent "raw" over the bus,
489  * i.e. it isn't known in advance how many bytes will be received. Therefore
490  * the slave driver has to provide a "large enough" buffer and either read the
491  * buffer, when it is full, or detect, that some data has arrived, then wait for
492  * a timeout, if no more data arrives - receive what's already there. We want to
493  * handle such slaves in a special way to allow an optimised mode for other
494  * users, for whom the amount of data is known in advance. So far there's no way
495  * to recognise such slaves. We use a data-width check to distinguish between
496  * the SD host and the PL011 UART.
497  */
498 
499 static int nbpf_prep_one(struct nbpf_link_desc *ldesc,
500 			 enum dma_transfer_direction direction,
501 			 dma_addr_t src, dma_addr_t dst, size_t size, bool last)
502 {
503 	struct nbpf_link_reg *hwdesc = ldesc->hwdesc;
504 	struct nbpf_desc *desc = ldesc->desc;
505 	struct nbpf_channel *chan = desc->chan;
506 	struct device *dev = chan->dma_chan.device->dev;
507 	size_t mem_xfer, slave_xfer;
508 	bool can_burst;
509 
510 	hwdesc->header = NBPF_HEADER_WBD | NBPF_HEADER_LV |
511 		(last ? NBPF_HEADER_LE : 0);
512 
513 	hwdesc->src_addr = src;
514 	hwdesc->dst_addr = dst;
515 	hwdesc->transaction_size = size;
516 
517 	/*
518 	 * set config: SAD, DAD, DDS, SDS, etc.
519 	 * Note on transfer sizes: the DMAC can perform unaligned DMA transfers,
520 	 * but it is important to have transaction size a multiple of both
521 	 * receiver and transmitter transfer sizes. It is also possible to use
522 	 * different RAM and device transfer sizes, and it does work well with
523 	 * some devices, e.g. with V08R07S01E SD host controllers, which can use
524 	 * 128 byte transfers. But this doesn't work with other devices,
525 	 * especially when the transaction size is unknown. This is the case,
526 	 * e.g. with serial drivers like amba-pl011.c. For reception it sets up
527 	 * the transaction size of 4K and if fewer bytes are received, it
528 	 * pauses DMA and reads out data received via DMA as well as those left
529 	 * in the Rx FIFO. For this to work with the RAM side using burst
530 	 * transfers we enable the SBE bit and terminate the transfer in our
531 	 * .device_pause handler.
532 	 */
533 	mem_xfer = nbpf_xfer_ds(chan->nbpf, size, direction);
534 
535 	switch (direction) {
536 	case DMA_DEV_TO_MEM:
537 		can_burst = chan->slave_src_width >= 3;
538 		slave_xfer = min(mem_xfer, can_burst ?
539 				 chan->slave_src_burst : chan->slave_src_width);
540 		/*
541 		 * Is the slave narrower than 64 bits, i.e. isn't using the full
542 		 * bus width and cannot use bursts?
543 		 */
544 		if (mem_xfer > chan->slave_src_burst && !can_burst)
545 			mem_xfer = chan->slave_src_burst;
546 		/* Device-to-RAM DMA is unreliable without REQD set */
547 		hwdesc->config = NBPF_CHAN_CFG_SAD | (NBPF_CHAN_CFG_DDS & (mem_xfer << 16)) |
548 			(NBPF_CHAN_CFG_SDS & (slave_xfer << 12)) | NBPF_CHAN_CFG_REQD |
549 			NBPF_CHAN_CFG_SBE;
550 		break;
551 
552 	case DMA_MEM_TO_DEV:
553 		slave_xfer = min(mem_xfer, chan->slave_dst_width >= 3 ?
554 				 chan->slave_dst_burst : chan->slave_dst_width);
555 		hwdesc->config = NBPF_CHAN_CFG_DAD | (NBPF_CHAN_CFG_SDS & (mem_xfer << 12)) |
556 			(NBPF_CHAN_CFG_DDS & (slave_xfer << 16)) | NBPF_CHAN_CFG_REQD;
557 		break;
558 
559 	case DMA_MEM_TO_MEM:
560 		hwdesc->config = NBPF_CHAN_CFG_TCM | NBPF_CHAN_CFG_TM |
561 			(NBPF_CHAN_CFG_SDS & (mem_xfer << 12)) |
562 			(NBPF_CHAN_CFG_DDS & (mem_xfer << 16));
563 		break;
564 
565 	default:
566 		return -EINVAL;
567 	}
568 
569 	hwdesc->config |= chan->dmarq_cfg | (last ? 0 : NBPF_CHAN_CFG_DEM) |
570 		NBPF_CHAN_CFG_DMS;
571 
572 	dev_dbg(dev, "%s(): desc @ %pad: hdr 0x%x, cfg 0x%x, %zu @ %pad -> %pad\n",
573 		__func__, &ldesc->hwdesc_dma_addr, hwdesc->header,
574 		hwdesc->config, size, &src, &dst);
575 
576 	dma_sync_single_for_device(dev, ldesc->hwdesc_dma_addr, sizeof(*hwdesc),
577 				   DMA_TO_DEVICE);
578 
579 	return 0;
580 }
581 
582 static size_t nbpf_bytes_left(struct nbpf_channel *chan)
583 {
584 	return nbpf_chan_read(chan, NBPF_CHAN_CUR_TR_BYTE);
585 }
586 
587 static void nbpf_configure(struct nbpf_device *nbpf)
588 {
589 	nbpf_write(nbpf, NBPF_CTRL, NBPF_CTRL_LVINT);
590 }
591 
592 /*		Generic part			*/
593 
594 /* DMA ENGINE functions */
595 static void nbpf_issue_pending(struct dma_chan *dchan)
596 {
597 	struct nbpf_channel *chan = nbpf_to_chan(dchan);
598 	unsigned long flags;
599 
600 	dev_dbg(dchan->device->dev, "Entry %s()\n", __func__);
601 
602 	spin_lock_irqsave(&chan->lock, flags);
603 	if (list_empty(&chan->queued))
604 		goto unlock;
605 
606 	list_splice_tail_init(&chan->queued, &chan->active);
607 
608 	if (!chan->running) {
609 		struct nbpf_desc *desc = list_first_entry(&chan->active,
610 						struct nbpf_desc, node);
611 		if (!nbpf_start(desc))
612 			chan->running = desc;
613 	}
614 
615 unlock:
616 	spin_unlock_irqrestore(&chan->lock, flags);
617 }
618 
619 static enum dma_status nbpf_tx_status(struct dma_chan *dchan,
620 		dma_cookie_t cookie, struct dma_tx_state *state)
621 {
622 	struct nbpf_channel *chan = nbpf_to_chan(dchan);
623 	enum dma_status status = dma_cookie_status(dchan, cookie, state);
624 
625 	if (state) {
626 		dma_cookie_t running;
627 		unsigned long flags;
628 
629 		spin_lock_irqsave(&chan->lock, flags);
630 		running = chan->running ? chan->running->async_tx.cookie : -EINVAL;
631 
632 		if (cookie == running) {
633 			state->residue = nbpf_bytes_left(chan);
634 			dev_dbg(dchan->device->dev, "%s(): residue %u\n", __func__,
635 				state->residue);
636 		} else if (status == DMA_IN_PROGRESS) {
637 			struct nbpf_desc *desc;
638 			bool found = false;
639 
640 			list_for_each_entry(desc, &chan->active, node)
641 				if (desc->async_tx.cookie == cookie) {
642 					found = true;
643 					break;
644 				}
645 
646 			if (!found)
647 				list_for_each_entry(desc, &chan->queued, node)
648 					if (desc->async_tx.cookie == cookie) {
649 						found = true;
650 						break;
651 
652 					}
653 
654 			state->residue = found ? desc->length : 0;
655 		}
656 
657 		spin_unlock_irqrestore(&chan->lock, flags);
658 	}
659 
660 	if (chan->paused)
661 		status = DMA_PAUSED;
662 
663 	return status;
664 }
665 
666 static dma_cookie_t nbpf_tx_submit(struct dma_async_tx_descriptor *tx)
667 {
668 	struct nbpf_desc *desc = container_of(tx, struct nbpf_desc, async_tx);
669 	struct nbpf_channel *chan = desc->chan;
670 	unsigned long flags;
671 	dma_cookie_t cookie;
672 
673 	spin_lock_irqsave(&chan->lock, flags);
674 	cookie = dma_cookie_assign(tx);
675 	list_add_tail(&desc->node, &chan->queued);
676 	spin_unlock_irqrestore(&chan->lock, flags);
677 
678 	dev_dbg(chan->dma_chan.device->dev, "Entry %s(%d)\n", __func__, cookie);
679 
680 	return cookie;
681 }
682 
683 static int nbpf_desc_page_alloc(struct nbpf_channel *chan)
684 {
685 	struct dma_chan *dchan = &chan->dma_chan;
686 	struct nbpf_desc_page *dpage = (void *)get_zeroed_page(GFP_KERNEL | GFP_DMA);
687 	struct nbpf_link_desc *ldesc;
688 	struct nbpf_link_reg *hwdesc;
689 	struct nbpf_desc *desc;
690 	LIST_HEAD(head);
691 	LIST_HEAD(lhead);
692 	int i;
693 	struct device *dev = dchan->device->dev;
694 
695 	if (!dpage)
696 		return -ENOMEM;
697 
698 	dev_dbg(dev, "%s(): alloc %lu descriptors, %lu segments, total alloc %zu\n",
699 		__func__, NBPF_DESCS_PER_PAGE, NBPF_SEGMENTS_PER_PAGE, sizeof(*dpage));
700 
701 	for (i = 0, ldesc = dpage->ldesc, hwdesc = dpage->hwdesc;
702 	     i < ARRAY_SIZE(dpage->ldesc);
703 	     i++, ldesc++, hwdesc++) {
704 		ldesc->hwdesc = hwdesc;
705 		list_add_tail(&ldesc->node, &lhead);
706 		ldesc->hwdesc_dma_addr = dma_map_single(dchan->device->dev,
707 					hwdesc, sizeof(*hwdesc), DMA_TO_DEVICE);
708 
709 		dev_dbg(dev, "%s(): mapped 0x%p to %pad\n", __func__,
710 			hwdesc, &ldesc->hwdesc_dma_addr);
711 	}
712 
713 	for (i = 0, desc = dpage->desc;
714 	     i < ARRAY_SIZE(dpage->desc);
715 	     i++, desc++) {
716 		dma_async_tx_descriptor_init(&desc->async_tx, dchan);
717 		desc->async_tx.tx_submit = nbpf_tx_submit;
718 		desc->chan = chan;
719 		INIT_LIST_HEAD(&desc->sg);
720 		list_add_tail(&desc->node, &head);
721 	}
722 
723 	/*
724 	 * This function cannot be called from interrupt context, so, no need to
725 	 * save flags
726 	 */
727 	spin_lock_irq(&chan->lock);
728 	list_splice_tail(&lhead, &chan->free_links);
729 	list_splice_tail(&head, &chan->free);
730 	list_add(&dpage->node, &chan->desc_page);
731 	spin_unlock_irq(&chan->lock);
732 
733 	return ARRAY_SIZE(dpage->desc);
734 }
735 
736 static void nbpf_desc_put(struct nbpf_desc *desc)
737 {
738 	struct nbpf_channel *chan = desc->chan;
739 	struct nbpf_link_desc *ldesc, *tmp;
740 	unsigned long flags;
741 
742 	spin_lock_irqsave(&chan->lock, flags);
743 	list_for_each_entry_safe(ldesc, tmp, &desc->sg, node)
744 		list_move(&ldesc->node, &chan->free_links);
745 
746 	list_add(&desc->node, &chan->free);
747 	spin_unlock_irqrestore(&chan->lock, flags);
748 }
749 
750 static void nbpf_scan_acked(struct nbpf_channel *chan)
751 {
752 	struct nbpf_desc *desc, *tmp;
753 	unsigned long flags;
754 	LIST_HEAD(head);
755 
756 	spin_lock_irqsave(&chan->lock, flags);
757 	list_for_each_entry_safe(desc, tmp, &chan->done, node)
758 		if (async_tx_test_ack(&desc->async_tx) && desc->user_wait) {
759 			list_move(&desc->node, &head);
760 			desc->user_wait = false;
761 		}
762 	spin_unlock_irqrestore(&chan->lock, flags);
763 
764 	list_for_each_entry_safe(desc, tmp, &head, node) {
765 		list_del(&desc->node);
766 		nbpf_desc_put(desc);
767 	}
768 }
769 
770 /*
771  * We have to allocate descriptors with the channel lock dropped. This means,
772  * before we re-acquire the lock buffers can be taken already, so we have to
773  * re-check after re-acquiring the lock and possibly retry, if buffers are gone
774  * again.
775  */
776 static struct nbpf_desc *nbpf_desc_get(struct nbpf_channel *chan, size_t len)
777 {
778 	struct nbpf_desc *desc = NULL;
779 	struct nbpf_link_desc *ldesc, *prev = NULL;
780 
781 	nbpf_scan_acked(chan);
782 
783 	spin_lock_irq(&chan->lock);
784 
785 	do {
786 		int i = 0, ret;
787 
788 		if (list_empty(&chan->free)) {
789 			/* No more free descriptors */
790 			spin_unlock_irq(&chan->lock);
791 			ret = nbpf_desc_page_alloc(chan);
792 			if (ret < 0)
793 				return NULL;
794 			spin_lock_irq(&chan->lock);
795 			continue;
796 		}
797 		desc = list_first_entry(&chan->free, struct nbpf_desc, node);
798 		list_del(&desc->node);
799 
800 		do {
801 			if (list_empty(&chan->free_links)) {
802 				/* No more free link descriptors */
803 				spin_unlock_irq(&chan->lock);
804 				ret = nbpf_desc_page_alloc(chan);
805 				if (ret < 0) {
806 					nbpf_desc_put(desc);
807 					return NULL;
808 				}
809 				spin_lock_irq(&chan->lock);
810 				continue;
811 			}
812 
813 			ldesc = list_first_entry(&chan->free_links,
814 						 struct nbpf_link_desc, node);
815 			ldesc->desc = desc;
816 			if (prev)
817 				prev->hwdesc->next = (u32)ldesc->hwdesc_dma_addr;
818 
819 			prev = ldesc;
820 			list_move_tail(&ldesc->node, &desc->sg);
821 
822 			i++;
823 		} while (i < len);
824 	} while (!desc);
825 
826 	prev->hwdesc->next = 0;
827 
828 	spin_unlock_irq(&chan->lock);
829 
830 	return desc;
831 }
832 
833 static void nbpf_chan_idle(struct nbpf_channel *chan)
834 {
835 	struct nbpf_desc *desc, *tmp;
836 	unsigned long flags;
837 	LIST_HEAD(head);
838 
839 	spin_lock_irqsave(&chan->lock, flags);
840 
841 	list_splice_init(&chan->done, &head);
842 	list_splice_init(&chan->active, &head);
843 	list_splice_init(&chan->queued, &head);
844 
845 	chan->running = NULL;
846 
847 	spin_unlock_irqrestore(&chan->lock, flags);
848 
849 	list_for_each_entry_safe(desc, tmp, &head, node) {
850 		dev_dbg(chan->nbpf->dma_dev.dev, "%s(): force-free desc %p cookie %d\n",
851 			__func__, desc, desc->async_tx.cookie);
852 		list_del(&desc->node);
853 		nbpf_desc_put(desc);
854 	}
855 }
856 
857 static int nbpf_pause(struct dma_chan *dchan)
858 {
859 	struct nbpf_channel *chan = nbpf_to_chan(dchan);
860 
861 	dev_dbg(dchan->device->dev, "Entry %s\n", __func__);
862 
863 	chan->paused = true;
864 	nbpf_chan_write(chan, NBPF_CHAN_CTRL, NBPF_CHAN_CTRL_SETSUS);
865 	/* See comment in nbpf_prep_one() */
866 	nbpf_chan_write(chan, NBPF_CHAN_CTRL, NBPF_CHAN_CTRL_CLREN);
867 
868 	return 0;
869 }
870 
871 static int nbpf_terminate_all(struct dma_chan *dchan)
872 {
873 	struct nbpf_channel *chan = nbpf_to_chan(dchan);
874 
875 	dev_dbg(dchan->device->dev, "Entry %s\n", __func__);
876 	dev_dbg(dchan->device->dev, "Terminating\n");
877 
878 	nbpf_chan_halt(chan);
879 	nbpf_chan_idle(chan);
880 
881 	return 0;
882 }
883 
884 static int nbpf_config(struct dma_chan *dchan,
885 		       struct dma_slave_config *config)
886 {
887 	struct nbpf_channel *chan = nbpf_to_chan(dchan);
888 
889 	dev_dbg(dchan->device->dev, "Entry %s\n", __func__);
890 
891 	/*
892 	 * We could check config->slave_id to match chan->terminal here,
893 	 * but with DT they would be coming from the same source, so
894 	 * such a check would be superflous
895 	 */
896 
897 	chan->slave_dst_addr = config->dst_addr;
898 	chan->slave_dst_width = nbpf_xfer_size(chan->nbpf,
899 					       config->dst_addr_width, 1);
900 	chan->slave_dst_burst = nbpf_xfer_size(chan->nbpf,
901 					       config->dst_addr_width,
902 					       config->dst_maxburst);
903 	chan->slave_src_addr = config->src_addr;
904 	chan->slave_src_width = nbpf_xfer_size(chan->nbpf,
905 					       config->src_addr_width, 1);
906 	chan->slave_src_burst = nbpf_xfer_size(chan->nbpf,
907 					       config->src_addr_width,
908 					       config->src_maxburst);
909 
910 	return 0;
911 }
912 
913 static struct dma_async_tx_descriptor *nbpf_prep_sg(struct nbpf_channel *chan,
914 		struct scatterlist *src_sg, struct scatterlist *dst_sg,
915 		size_t len, enum dma_transfer_direction direction,
916 		unsigned long flags)
917 {
918 	struct nbpf_link_desc *ldesc;
919 	struct scatterlist *mem_sg;
920 	struct nbpf_desc *desc;
921 	bool inc_src, inc_dst;
922 	size_t data_len = 0;
923 	int i = 0;
924 
925 	switch (direction) {
926 	case DMA_DEV_TO_MEM:
927 		mem_sg = dst_sg;
928 		inc_src = false;
929 		inc_dst = true;
930 		break;
931 
932 	case DMA_MEM_TO_DEV:
933 		mem_sg = src_sg;
934 		inc_src = true;
935 		inc_dst = false;
936 		break;
937 
938 	default:
939 	case DMA_MEM_TO_MEM:
940 		mem_sg = src_sg;
941 		inc_src = true;
942 		inc_dst = true;
943 	}
944 
945 	desc = nbpf_desc_get(chan, len);
946 	if (!desc)
947 		return NULL;
948 
949 	desc->async_tx.flags = flags;
950 	desc->async_tx.cookie = -EBUSY;
951 	desc->user_wait = false;
952 
953 	/*
954 	 * This is a private descriptor list, and we own the descriptor. No need
955 	 * to lock.
956 	 */
957 	list_for_each_entry(ldesc, &desc->sg, node) {
958 		int ret = nbpf_prep_one(ldesc, direction,
959 					sg_dma_address(src_sg),
960 					sg_dma_address(dst_sg),
961 					sg_dma_len(mem_sg),
962 					i == len - 1);
963 		if (ret < 0) {
964 			nbpf_desc_put(desc);
965 			return NULL;
966 		}
967 		data_len += sg_dma_len(mem_sg);
968 		if (inc_src)
969 			src_sg = sg_next(src_sg);
970 		if (inc_dst)
971 			dst_sg = sg_next(dst_sg);
972 		mem_sg = direction == DMA_DEV_TO_MEM ? dst_sg : src_sg;
973 		i++;
974 	}
975 
976 	desc->length = data_len;
977 
978 	/* The user has to return the descriptor to us ASAP via .tx_submit() */
979 	return &desc->async_tx;
980 }
981 
982 static struct dma_async_tx_descriptor *nbpf_prep_memcpy(
983 	struct dma_chan *dchan, dma_addr_t dst, dma_addr_t src,
984 	size_t len, unsigned long flags)
985 {
986 	struct nbpf_channel *chan = nbpf_to_chan(dchan);
987 	struct scatterlist dst_sg;
988 	struct scatterlist src_sg;
989 
990 	sg_init_table(&dst_sg, 1);
991 	sg_init_table(&src_sg, 1);
992 
993 	sg_dma_address(&dst_sg) = dst;
994 	sg_dma_address(&src_sg) = src;
995 
996 	sg_dma_len(&dst_sg) = len;
997 	sg_dma_len(&src_sg) = len;
998 
999 	dev_dbg(dchan->device->dev, "%s(): %zu @ %pad -> %pad\n",
1000 		__func__, len, &src, &dst);
1001 
1002 	return nbpf_prep_sg(chan, &src_sg, &dst_sg, 1,
1003 			    DMA_MEM_TO_MEM, flags);
1004 }
1005 
1006 static struct dma_async_tx_descriptor *nbpf_prep_slave_sg(
1007 	struct dma_chan *dchan, struct scatterlist *sgl, unsigned int sg_len,
1008 	enum dma_transfer_direction direction, unsigned long flags, void *context)
1009 {
1010 	struct nbpf_channel *chan = nbpf_to_chan(dchan);
1011 	struct scatterlist slave_sg;
1012 
1013 	dev_dbg(dchan->device->dev, "Entry %s()\n", __func__);
1014 
1015 	sg_init_table(&slave_sg, 1);
1016 
1017 	switch (direction) {
1018 	case DMA_MEM_TO_DEV:
1019 		sg_dma_address(&slave_sg) = chan->slave_dst_addr;
1020 		return nbpf_prep_sg(chan, sgl, &slave_sg, sg_len,
1021 				    direction, flags);
1022 
1023 	case DMA_DEV_TO_MEM:
1024 		sg_dma_address(&slave_sg) = chan->slave_src_addr;
1025 		return nbpf_prep_sg(chan, &slave_sg, sgl, sg_len,
1026 				    direction, flags);
1027 
1028 	default:
1029 		return NULL;
1030 	}
1031 }
1032 
1033 static int nbpf_alloc_chan_resources(struct dma_chan *dchan)
1034 {
1035 	struct nbpf_channel *chan = nbpf_to_chan(dchan);
1036 	int ret;
1037 
1038 	INIT_LIST_HEAD(&chan->free);
1039 	INIT_LIST_HEAD(&chan->free_links);
1040 	INIT_LIST_HEAD(&chan->queued);
1041 	INIT_LIST_HEAD(&chan->active);
1042 	INIT_LIST_HEAD(&chan->done);
1043 
1044 	ret = nbpf_desc_page_alloc(chan);
1045 	if (ret < 0)
1046 		return ret;
1047 
1048 	dev_dbg(dchan->device->dev, "Entry %s(): terminal %u\n", __func__,
1049 		chan->terminal);
1050 
1051 	nbpf_chan_configure(chan);
1052 
1053 	return ret;
1054 }
1055 
1056 static void nbpf_free_chan_resources(struct dma_chan *dchan)
1057 {
1058 	struct nbpf_channel *chan = nbpf_to_chan(dchan);
1059 	struct nbpf_desc_page *dpage, *tmp;
1060 
1061 	dev_dbg(dchan->device->dev, "Entry %s()\n", __func__);
1062 
1063 	nbpf_chan_halt(chan);
1064 	nbpf_chan_idle(chan);
1065 	/* Clean up for if a channel is re-used for MEMCPY after slave DMA */
1066 	nbpf_chan_prepare_default(chan);
1067 
1068 	list_for_each_entry_safe(dpage, tmp, &chan->desc_page, node) {
1069 		struct nbpf_link_desc *ldesc;
1070 		int i;
1071 		list_del(&dpage->node);
1072 		for (i = 0, ldesc = dpage->ldesc;
1073 		     i < ARRAY_SIZE(dpage->ldesc);
1074 		     i++, ldesc++)
1075 			dma_unmap_single(dchan->device->dev, ldesc->hwdesc_dma_addr,
1076 					 sizeof(*ldesc->hwdesc), DMA_TO_DEVICE);
1077 		free_page((unsigned long)dpage);
1078 	}
1079 }
1080 
1081 static struct dma_chan *nbpf_of_xlate(struct of_phandle_args *dma_spec,
1082 				      struct of_dma *ofdma)
1083 {
1084 	struct nbpf_device *nbpf = ofdma->of_dma_data;
1085 	struct dma_chan *dchan;
1086 	struct nbpf_channel *chan;
1087 
1088 	if (dma_spec->args_count != 2)
1089 		return NULL;
1090 
1091 	dchan = dma_get_any_slave_channel(&nbpf->dma_dev);
1092 	if (!dchan)
1093 		return NULL;
1094 
1095 	dev_dbg(dchan->device->dev, "Entry %s(%pOFn)\n", __func__,
1096 		dma_spec->np);
1097 
1098 	chan = nbpf_to_chan(dchan);
1099 
1100 	chan->terminal = dma_spec->args[0];
1101 	chan->flags = dma_spec->args[1];
1102 
1103 	nbpf_chan_prepare(chan);
1104 	nbpf_chan_configure(chan);
1105 
1106 	return dchan;
1107 }
1108 
1109 static void nbpf_chan_tasklet(unsigned long data)
1110 {
1111 	struct nbpf_channel *chan = (struct nbpf_channel *)data;
1112 	struct nbpf_desc *desc, *tmp;
1113 	struct dmaengine_desc_callback cb;
1114 
1115 	while (!list_empty(&chan->done)) {
1116 		bool found = false, must_put, recycling = false;
1117 
1118 		spin_lock_irq(&chan->lock);
1119 
1120 		list_for_each_entry_safe(desc, tmp, &chan->done, node) {
1121 			if (!desc->user_wait) {
1122 				/* Newly completed descriptor, have to process */
1123 				found = true;
1124 				break;
1125 			} else if (async_tx_test_ack(&desc->async_tx)) {
1126 				/*
1127 				 * This descriptor was waiting for a user ACK,
1128 				 * it can be recycled now.
1129 				 */
1130 				list_del(&desc->node);
1131 				spin_unlock_irq(&chan->lock);
1132 				nbpf_desc_put(desc);
1133 				recycling = true;
1134 				break;
1135 			}
1136 		}
1137 
1138 		if (recycling)
1139 			continue;
1140 
1141 		if (!found) {
1142 			/* This can happen if TERMINATE_ALL has been called */
1143 			spin_unlock_irq(&chan->lock);
1144 			break;
1145 		}
1146 
1147 		dma_cookie_complete(&desc->async_tx);
1148 
1149 		/*
1150 		 * With released lock we cannot dereference desc, maybe it's
1151 		 * still on the "done" list
1152 		 */
1153 		if (async_tx_test_ack(&desc->async_tx)) {
1154 			list_del(&desc->node);
1155 			must_put = true;
1156 		} else {
1157 			desc->user_wait = true;
1158 			must_put = false;
1159 		}
1160 
1161 		dmaengine_desc_get_callback(&desc->async_tx, &cb);
1162 
1163 		/* ack and callback completed descriptor */
1164 		spin_unlock_irq(&chan->lock);
1165 
1166 		dmaengine_desc_callback_invoke(&cb, NULL);
1167 
1168 		if (must_put)
1169 			nbpf_desc_put(desc);
1170 	}
1171 }
1172 
1173 static irqreturn_t nbpf_chan_irq(int irq, void *dev)
1174 {
1175 	struct nbpf_channel *chan = dev;
1176 	bool done = nbpf_status_get(chan);
1177 	struct nbpf_desc *desc;
1178 	irqreturn_t ret;
1179 	bool bh = false;
1180 
1181 	if (!done)
1182 		return IRQ_NONE;
1183 
1184 	nbpf_status_ack(chan);
1185 
1186 	dev_dbg(&chan->dma_chan.dev->device, "%s()\n", __func__);
1187 
1188 	spin_lock(&chan->lock);
1189 	desc = chan->running;
1190 	if (WARN_ON(!desc)) {
1191 		ret = IRQ_NONE;
1192 		goto unlock;
1193 	} else {
1194 		ret = IRQ_HANDLED;
1195 		bh = true;
1196 	}
1197 
1198 	list_move_tail(&desc->node, &chan->done);
1199 	chan->running = NULL;
1200 
1201 	if (!list_empty(&chan->active)) {
1202 		desc = list_first_entry(&chan->active,
1203 					struct nbpf_desc, node);
1204 		if (!nbpf_start(desc))
1205 			chan->running = desc;
1206 	}
1207 
1208 unlock:
1209 	spin_unlock(&chan->lock);
1210 
1211 	if (bh)
1212 		tasklet_schedule(&chan->tasklet);
1213 
1214 	return ret;
1215 }
1216 
1217 static irqreturn_t nbpf_err_irq(int irq, void *dev)
1218 {
1219 	struct nbpf_device *nbpf = dev;
1220 	u32 error = nbpf_error_get(nbpf);
1221 
1222 	dev_warn(nbpf->dma_dev.dev, "DMA error IRQ %u\n", irq);
1223 
1224 	if (!error)
1225 		return IRQ_NONE;
1226 
1227 	do {
1228 		struct nbpf_channel *chan = nbpf_error_get_channel(nbpf, error);
1229 		/* On error: abort all queued transfers, no callback */
1230 		nbpf_error_clear(chan);
1231 		nbpf_chan_idle(chan);
1232 		error = nbpf_error_get(nbpf);
1233 	} while (error);
1234 
1235 	return IRQ_HANDLED;
1236 }
1237 
1238 static int nbpf_chan_probe(struct nbpf_device *nbpf, int n)
1239 {
1240 	struct dma_device *dma_dev = &nbpf->dma_dev;
1241 	struct nbpf_channel *chan = nbpf->chan + n;
1242 	int ret;
1243 
1244 	chan->nbpf = nbpf;
1245 	chan->base = nbpf->base + NBPF_REG_CHAN_OFFSET + NBPF_REG_CHAN_SIZE * n;
1246 	INIT_LIST_HEAD(&chan->desc_page);
1247 	spin_lock_init(&chan->lock);
1248 	chan->dma_chan.device = dma_dev;
1249 	dma_cookie_init(&chan->dma_chan);
1250 	nbpf_chan_prepare_default(chan);
1251 
1252 	dev_dbg(dma_dev->dev, "%s(): channel %d: -> %p\n", __func__, n, chan->base);
1253 
1254 	snprintf(chan->name, sizeof(chan->name), "nbpf %d", n);
1255 
1256 	tasklet_init(&chan->tasklet, nbpf_chan_tasklet, (unsigned long)chan);
1257 	ret = devm_request_irq(dma_dev->dev, chan->irq,
1258 			nbpf_chan_irq, IRQF_SHARED,
1259 			chan->name, chan);
1260 	if (ret < 0)
1261 		return ret;
1262 
1263 	/* Add the channel to DMA device channel list */
1264 	list_add_tail(&chan->dma_chan.device_node,
1265 		      &dma_dev->channels);
1266 
1267 	return 0;
1268 }
1269 
1270 static const struct of_device_id nbpf_match[] = {
1271 	{.compatible = "renesas,nbpfaxi64dmac1b4",	.data = &nbpf_cfg[NBPF1B4]},
1272 	{.compatible = "renesas,nbpfaxi64dmac1b8",	.data = &nbpf_cfg[NBPF1B8]},
1273 	{.compatible = "renesas,nbpfaxi64dmac1b16",	.data = &nbpf_cfg[NBPF1B16]},
1274 	{.compatible = "renesas,nbpfaxi64dmac4b4",	.data = &nbpf_cfg[NBPF4B4]},
1275 	{.compatible = "renesas,nbpfaxi64dmac4b8",	.data = &nbpf_cfg[NBPF4B8]},
1276 	{.compatible = "renesas,nbpfaxi64dmac4b16",	.data = &nbpf_cfg[NBPF4B16]},
1277 	{.compatible = "renesas,nbpfaxi64dmac8b4",	.data = &nbpf_cfg[NBPF8B4]},
1278 	{.compatible = "renesas,nbpfaxi64dmac8b8",	.data = &nbpf_cfg[NBPF8B8]},
1279 	{.compatible = "renesas,nbpfaxi64dmac8b16",	.data = &nbpf_cfg[NBPF8B16]},
1280 	{}
1281 };
1282 MODULE_DEVICE_TABLE(of, nbpf_match);
1283 
1284 static int nbpf_probe(struct platform_device *pdev)
1285 {
1286 	struct device *dev = &pdev->dev;
1287 	struct device_node *np = dev->of_node;
1288 	struct nbpf_device *nbpf;
1289 	struct dma_device *dma_dev;
1290 	struct resource *iomem, *irq_res;
1291 	const struct nbpf_config *cfg;
1292 	int num_channels;
1293 	int ret, irq, eirq, i;
1294 	int irqbuf[9] /* maximum 8 channels + error IRQ */;
1295 	unsigned int irqs = 0;
1296 
1297 	BUILD_BUG_ON(sizeof(struct nbpf_desc_page) > PAGE_SIZE);
1298 
1299 	/* DT only */
1300 	if (!np)
1301 		return -ENODEV;
1302 
1303 	cfg = of_device_get_match_data(dev);
1304 	num_channels = cfg->num_channels;
1305 
1306 	nbpf = devm_kzalloc(dev, struct_size(nbpf, chan, num_channels),
1307 			    GFP_KERNEL);
1308 	if (!nbpf)
1309 		return -ENOMEM;
1310 
1311 	dma_dev = &nbpf->dma_dev;
1312 	dma_dev->dev = dev;
1313 
1314 	iomem = platform_get_resource(pdev, IORESOURCE_MEM, 0);
1315 	nbpf->base = devm_ioremap_resource(dev, iomem);
1316 	if (IS_ERR(nbpf->base))
1317 		return PTR_ERR(nbpf->base);
1318 
1319 	nbpf->clk = devm_clk_get(dev, NULL);
1320 	if (IS_ERR(nbpf->clk))
1321 		return PTR_ERR(nbpf->clk);
1322 
1323 	of_property_read_u32(np, "max-burst-mem-read",
1324 			     &nbpf->max_burst_mem_read);
1325 	of_property_read_u32(np, "max-burst-mem-write",
1326 			     &nbpf->max_burst_mem_write);
1327 
1328 	nbpf->config = cfg;
1329 
1330 	for (i = 0; irqs < ARRAY_SIZE(irqbuf); i++) {
1331 		irq_res = platform_get_resource(pdev, IORESOURCE_IRQ, i);
1332 		if (!irq_res)
1333 			break;
1334 
1335 		for (irq = irq_res->start; irq <= irq_res->end;
1336 		     irq++, irqs++)
1337 			irqbuf[irqs] = irq;
1338 	}
1339 
1340 	/*
1341 	 * 3 IRQ resource schemes are supported:
1342 	 * 1. 1 shared IRQ for error and all channels
1343 	 * 2. 2 IRQs: one for error and one shared for all channels
1344 	 * 3. 1 IRQ for error and an own IRQ for each channel
1345 	 */
1346 	if (irqs != 1 && irqs != 2 && irqs != num_channels + 1)
1347 		return -ENXIO;
1348 
1349 	if (irqs == 1) {
1350 		eirq = irqbuf[0];
1351 
1352 		for (i = 0; i <= num_channels; i++)
1353 			nbpf->chan[i].irq = irqbuf[0];
1354 	} else {
1355 		eirq = platform_get_irq_byname(pdev, "error");
1356 		if (eirq < 0)
1357 			return eirq;
1358 
1359 		if (irqs == num_channels + 1) {
1360 			struct nbpf_channel *chan;
1361 
1362 			for (i = 0, chan = nbpf->chan; i <= num_channels;
1363 			     i++, chan++) {
1364 				/* Skip the error IRQ */
1365 				if (irqbuf[i] == eirq)
1366 					i++;
1367 				chan->irq = irqbuf[i];
1368 			}
1369 
1370 			if (chan != nbpf->chan + num_channels)
1371 				return -EINVAL;
1372 		} else {
1373 			/* 2 IRQs and more than one channel */
1374 			if (irqbuf[0] == eirq)
1375 				irq = irqbuf[1];
1376 			else
1377 				irq = irqbuf[0];
1378 
1379 			for (i = 0; i <= num_channels; i++)
1380 				nbpf->chan[i].irq = irq;
1381 		}
1382 	}
1383 
1384 	ret = devm_request_irq(dev, eirq, nbpf_err_irq,
1385 			       IRQF_SHARED, "dma error", nbpf);
1386 	if (ret < 0)
1387 		return ret;
1388 	nbpf->eirq = eirq;
1389 
1390 	INIT_LIST_HEAD(&dma_dev->channels);
1391 
1392 	/* Create DMA Channel */
1393 	for (i = 0; i < num_channels; i++) {
1394 		ret = nbpf_chan_probe(nbpf, i);
1395 		if (ret < 0)
1396 			return ret;
1397 	}
1398 
1399 	dma_cap_set(DMA_MEMCPY, dma_dev->cap_mask);
1400 	dma_cap_set(DMA_SLAVE, dma_dev->cap_mask);
1401 	dma_cap_set(DMA_PRIVATE, dma_dev->cap_mask);
1402 
1403 	/* Common and MEMCPY operations */
1404 	dma_dev->device_alloc_chan_resources
1405 		= nbpf_alloc_chan_resources;
1406 	dma_dev->device_free_chan_resources = nbpf_free_chan_resources;
1407 	dma_dev->device_prep_dma_memcpy = nbpf_prep_memcpy;
1408 	dma_dev->device_tx_status = nbpf_tx_status;
1409 	dma_dev->device_issue_pending = nbpf_issue_pending;
1410 
1411 	/*
1412 	 * If we drop support for unaligned MEMCPY buffer addresses and / or
1413 	 * lengths by setting
1414 	 * dma_dev->copy_align = 4;
1415 	 * then we can set transfer length to 4 bytes in nbpf_prep_one() for
1416 	 * DMA_MEM_TO_MEM
1417 	 */
1418 
1419 	/* Compulsory for DMA_SLAVE fields */
1420 	dma_dev->device_prep_slave_sg = nbpf_prep_slave_sg;
1421 	dma_dev->device_config = nbpf_config;
1422 	dma_dev->device_pause = nbpf_pause;
1423 	dma_dev->device_terminate_all = nbpf_terminate_all;
1424 
1425 	dma_dev->src_addr_widths = NBPF_DMA_BUSWIDTHS;
1426 	dma_dev->dst_addr_widths = NBPF_DMA_BUSWIDTHS;
1427 	dma_dev->directions = BIT(DMA_DEV_TO_MEM) | BIT(DMA_MEM_TO_DEV);
1428 
1429 	platform_set_drvdata(pdev, nbpf);
1430 
1431 	ret = clk_prepare_enable(nbpf->clk);
1432 	if (ret < 0)
1433 		return ret;
1434 
1435 	nbpf_configure(nbpf);
1436 
1437 	ret = dma_async_device_register(dma_dev);
1438 	if (ret < 0)
1439 		goto e_clk_off;
1440 
1441 	ret = of_dma_controller_register(np, nbpf_of_xlate, nbpf);
1442 	if (ret < 0)
1443 		goto e_dma_dev_unreg;
1444 
1445 	return 0;
1446 
1447 e_dma_dev_unreg:
1448 	dma_async_device_unregister(dma_dev);
1449 e_clk_off:
1450 	clk_disable_unprepare(nbpf->clk);
1451 
1452 	return ret;
1453 }
1454 
1455 static int nbpf_remove(struct platform_device *pdev)
1456 {
1457 	struct nbpf_device *nbpf = platform_get_drvdata(pdev);
1458 	int i;
1459 
1460 	devm_free_irq(&pdev->dev, nbpf->eirq, nbpf);
1461 
1462 	for (i = 0; i < nbpf->config->num_channels; i++) {
1463 		struct nbpf_channel *chan = nbpf->chan + i;
1464 
1465 		devm_free_irq(&pdev->dev, chan->irq, chan);
1466 
1467 		tasklet_kill(&chan->tasklet);
1468 	}
1469 
1470 	of_dma_controller_free(pdev->dev.of_node);
1471 	dma_async_device_unregister(&nbpf->dma_dev);
1472 	clk_disable_unprepare(nbpf->clk);
1473 
1474 	return 0;
1475 }
1476 
1477 static const struct platform_device_id nbpf_ids[] = {
1478 	{"nbpfaxi64dmac1b4",	(kernel_ulong_t)&nbpf_cfg[NBPF1B4]},
1479 	{"nbpfaxi64dmac1b8",	(kernel_ulong_t)&nbpf_cfg[NBPF1B8]},
1480 	{"nbpfaxi64dmac1b16",	(kernel_ulong_t)&nbpf_cfg[NBPF1B16]},
1481 	{"nbpfaxi64dmac4b4",	(kernel_ulong_t)&nbpf_cfg[NBPF4B4]},
1482 	{"nbpfaxi64dmac4b8",	(kernel_ulong_t)&nbpf_cfg[NBPF4B8]},
1483 	{"nbpfaxi64dmac4b16",	(kernel_ulong_t)&nbpf_cfg[NBPF4B16]},
1484 	{"nbpfaxi64dmac8b4",	(kernel_ulong_t)&nbpf_cfg[NBPF8B4]},
1485 	{"nbpfaxi64dmac8b8",	(kernel_ulong_t)&nbpf_cfg[NBPF8B8]},
1486 	{"nbpfaxi64dmac8b16",	(kernel_ulong_t)&nbpf_cfg[NBPF8B16]},
1487 	{},
1488 };
1489 MODULE_DEVICE_TABLE(platform, nbpf_ids);
1490 
1491 #ifdef CONFIG_PM
1492 static int nbpf_runtime_suspend(struct device *dev)
1493 {
1494 	struct nbpf_device *nbpf = platform_get_drvdata(to_platform_device(dev));
1495 	clk_disable_unprepare(nbpf->clk);
1496 	return 0;
1497 }
1498 
1499 static int nbpf_runtime_resume(struct device *dev)
1500 {
1501 	struct nbpf_device *nbpf = platform_get_drvdata(to_platform_device(dev));
1502 	return clk_prepare_enable(nbpf->clk);
1503 }
1504 #endif
1505 
1506 static const struct dev_pm_ops nbpf_pm_ops = {
1507 	SET_RUNTIME_PM_OPS(nbpf_runtime_suspend, nbpf_runtime_resume, NULL)
1508 };
1509 
1510 static struct platform_driver nbpf_driver = {
1511 	.driver = {
1512 		.name = "dma-nbpf",
1513 		.of_match_table = nbpf_match,
1514 		.pm = &nbpf_pm_ops,
1515 	},
1516 	.id_table = nbpf_ids,
1517 	.probe = nbpf_probe,
1518 	.remove = nbpf_remove,
1519 };
1520 
1521 module_platform_driver(nbpf_driver);
1522 
1523 MODULE_AUTHOR("Guennadi Liakhovetski <g.liakhovetski@gmx.de>");
1524 MODULE_DESCRIPTION("dmaengine driver for NBPFAXI64* DMACs");
1525 MODULE_LICENSE("GPL v2");
1526