xref: /linux/drivers/dma/nbpfaxi.c (revision db1ecca22edf27c5a3dd66af406c88b5b5ac7cc1)
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_dma.h>
19 #include <linux/platform_device.h>
20 #include <linux/slab.h>
21 
22 #include <dt-bindings/dma/nbpfaxi.h>
23 
24 #include "dmaengine.h"
25 
26 #define NBPF_REG_CHAN_OFFSET	0
27 #define NBPF_REG_CHAN_SIZE	0x40
28 
29 /* Channel Current Transaction Byte register */
30 #define NBPF_CHAN_CUR_TR_BYTE	0x20
31 
32 /* Channel Status register */
33 #define NBPF_CHAN_STAT	0x24
34 #define NBPF_CHAN_STAT_EN	1
35 #define NBPF_CHAN_STAT_TACT	4
36 #define NBPF_CHAN_STAT_ERR	0x10
37 #define NBPF_CHAN_STAT_END	0x20
38 #define NBPF_CHAN_STAT_TC	0x40
39 #define NBPF_CHAN_STAT_DER	0x400
40 
41 /* Channel Control register */
42 #define NBPF_CHAN_CTRL	0x28
43 #define NBPF_CHAN_CTRL_SETEN	1
44 #define NBPF_CHAN_CTRL_CLREN	2
45 #define NBPF_CHAN_CTRL_STG	4
46 #define NBPF_CHAN_CTRL_SWRST	8
47 #define NBPF_CHAN_CTRL_CLRRQ	0x10
48 #define NBPF_CHAN_CTRL_CLREND	0x20
49 #define NBPF_CHAN_CTRL_CLRTC	0x40
50 #define NBPF_CHAN_CTRL_SETSUS	0x100
51 #define NBPF_CHAN_CTRL_CLRSUS	0x200
52 
53 /* Channel Configuration register */
54 #define NBPF_CHAN_CFG	0x2c
55 #define NBPF_CHAN_CFG_SEL	7		/* terminal SELect: 0..7 */
56 #define NBPF_CHAN_CFG_REQD	8		/* REQuest Direction: DMAREQ is 0: input, 1: output */
57 #define NBPF_CHAN_CFG_LOEN	0x10		/* LOw ENable: low DMA request line is: 0: inactive, 1: active */
58 #define NBPF_CHAN_CFG_HIEN	0x20		/* HIgh ENable: high DMA request line is: 0: inactive, 1: active */
59 #define NBPF_CHAN_CFG_LVL	0x40		/* LeVeL: DMA request line is sensed as 0: edge, 1: level */
60 #define NBPF_CHAN_CFG_AM	0x700		/* ACK Mode: 0: Pulse mode, 1: Level mode, b'1x: Bus Cycle */
61 #define NBPF_CHAN_CFG_SDS	0xf000		/* Source Data Size: 0: 8 bits,... , 7: 1024 bits */
62 #define NBPF_CHAN_CFG_DDS	0xf0000		/* Destination Data Size: as above */
63 #define NBPF_CHAN_CFG_SAD	0x100000	/* Source ADdress counting: 0: increment, 1: fixed */
64 #define NBPF_CHAN_CFG_DAD	0x200000	/* Destination ADdress counting: 0: increment, 1: fixed */
65 #define NBPF_CHAN_CFG_TM	0x400000	/* Transfer Mode: 0: single, 1: block TM */
66 #define NBPF_CHAN_CFG_DEM	0x1000000	/* DMAEND interrupt Mask */
67 #define NBPF_CHAN_CFG_TCM	0x2000000	/* DMATCO interrupt Mask */
68 #define NBPF_CHAN_CFG_SBE	0x8000000	/* Sweep Buffer Enable */
69 #define NBPF_CHAN_CFG_RSEL	0x10000000	/* RM: Register Set sELect */
70 #define NBPF_CHAN_CFG_RSW	0x20000000	/* RM: Register Select sWitch */
71 #define NBPF_CHAN_CFG_REN	0x40000000	/* RM: Register Set Enable */
72 #define NBPF_CHAN_CFG_DMS	0x80000000	/* 0: register mode (RM), 1: link mode (LM) */
73 
74 #define NBPF_CHAN_NXLA	0x38
75 #define NBPF_CHAN_CRLA	0x3c
76 
77 /* Link Header field */
78 #define NBPF_HEADER_LV	1
79 #define NBPF_HEADER_LE	2
80 #define NBPF_HEADER_WBD	4
81 #define NBPF_HEADER_DIM	8
82 
83 #define NBPF_CTRL	0x300
84 #define NBPF_CTRL_PR	1		/* 0: fixed priority, 1: round robin */
85 #define NBPF_CTRL_LVINT	2		/* DMAEND and DMAERR signalling: 0: pulse, 1: level */
86 
87 #define NBPF_DSTAT_ER	0x314
88 #define NBPF_DSTAT_END	0x318
89 
90 #define NBPF_DMA_BUSWIDTHS \
91 	(BIT(DMA_SLAVE_BUSWIDTH_UNDEFINED) | \
92 	 BIT(DMA_SLAVE_BUSWIDTH_1_BYTE) | \
93 	 BIT(DMA_SLAVE_BUSWIDTH_2_BYTES) | \
94 	 BIT(DMA_SLAVE_BUSWIDTH_4_BYTES) | \
95 	 BIT(DMA_SLAVE_BUSWIDTH_8_BYTES))
96 
97 struct nbpf_config {
98 	int num_channels;
99 	int buffer_size;
100 };
101 
102 /*
103  * We've got 3 types of objects, used to describe DMA transfers:
104  * 1. high-level descriptor, containing a struct dma_async_tx_descriptor object
105  *	in it, used to communicate with the user
106  * 2. hardware DMA link descriptors, that we pass to DMAC for DMA transfer
107  *	queuing, these must be DMAable, using either the streaming DMA API or
108  *	allocated from coherent memory - one per SG segment
109  * 3. one per SG segment descriptors, used to manage HW link descriptors from
110  *	(2). They do not have to be DMAable. They can either be (a) allocated
111  *	together with link descriptors as mixed (DMA / CPU) objects, or (b)
112  *	separately. Even if allocated separately it would be best to link them
113  *	to link descriptors once during channel resource allocation and always
114  *	use them as a single object.
115  * Therefore for both cases (a) and (b) at run-time objects (2) and (3) shall be
116  * treated as a single SG segment descriptor.
117  */
118 
119 struct nbpf_link_reg {
120 	u32	header;
121 	u32	src_addr;
122 	u32	dst_addr;
123 	u32	transaction_size;
124 	u32	config;
125 	u32	interval;
126 	u32	extension;
127 	u32	next;
128 } __packed;
129 
130 struct nbpf_device;
131 struct nbpf_channel;
132 struct nbpf_desc;
133 
134 struct nbpf_link_desc {
135 	struct nbpf_link_reg *hwdesc;
136 	dma_addr_t hwdesc_dma_addr;
137 	struct nbpf_desc *desc;
138 	struct list_head node;
139 };
140 
141 /**
142  * struct nbpf_desc - DMA transfer descriptor
143  * @async_tx:	dmaengine object
144  * @user_wait:	waiting for a user ack
145  * @length:	total transfer length
146  * @chan:	associated DMAC channel
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  * @tasklet:	channel specific tasklet used for callbacks
178  * @base:	register address base
179  * @nbpf:	DMAC
180  * @name:	IRQ name
181  * @irq:	IRQ number
182  * @slave_src_addr:	source address for slave DMA
183  * @slave_src_width:	source slave data size in bytes
184  * @slave_src_burst:	maximum source slave burst size in bytes
185  * @slave_dst_addr:	destination address for slave DMA
186  * @slave_dst_width:	destination slave data size in bytes
187  * @slave_dst_burst:	maximum destination slave burst size in bytes
188  * @terminal:	DMA terminal, assigned to this channel
189  * @dmarq_cfg:	DMA request line configuration - high / low, edge / level for NBPF_CHAN_CFG
190  * @flags:	configuration flags from DT
191  * @lock:	protect descriptor lists
192  * @free_links:	list of free link descriptors
193  * @free:	list of free descriptors
194  * @queued:	list of queued descriptors
195  * @active:	list of descriptors, scheduled for processing
196  * @done:	list of completed descriptors, waiting post-processing
197  * @desc_page:	list of additionally allocated descriptor pages - if any
198  * @running:	linked descriptor of running transaction
199  * @paused:	are translations on this channel paused?
200  */
201 struct nbpf_channel {
202 	struct dma_chan dma_chan;
203 	struct tasklet_struct tasklet;
204 	void __iomem *base;
205 	struct nbpf_device *nbpf;
206 	char name[16];
207 	int irq;
208 	dma_addr_t slave_src_addr;
209 	size_t slave_src_width;
210 	size_t slave_src_burst;
211 	dma_addr_t slave_dst_addr;
212 	size_t slave_dst_width;
213 	size_t slave_dst_burst;
214 	unsigned int terminal;
215 	u32 dmarq_cfg;
216 	unsigned long flags;
217 	spinlock_t lock;
218 	struct list_head free_links;
219 	struct list_head free;
220 	struct list_head queued;
221 	struct list_head active;
222 	struct list_head done;
223 	struct list_head desc_page;
224 	struct nbpf_desc *running;
225 	bool paused;
226 };
227 
228 struct nbpf_device {
229 	struct dma_device dma_dev;
230 	void __iomem *base;
231 	u32 max_burst_mem_read;
232 	u32 max_burst_mem_write;
233 	struct clk *clk;
234 	const struct nbpf_config *config;
235 	unsigned int eirq;
236 	struct nbpf_channel chan[];
237 };
238 
239 enum nbpf_model {
240 	NBPF1B4,
241 	NBPF1B8,
242 	NBPF1B16,
243 	NBPF4B4,
244 	NBPF4B8,
245 	NBPF4B16,
246 	NBPF8B4,
247 	NBPF8B8,
248 	NBPF8B16,
249 };
250 
251 static struct nbpf_config nbpf_cfg[] = {
252 	[NBPF1B4] = {
253 		.num_channels = 1,
254 		.buffer_size = 4,
255 	},
256 	[NBPF1B8] = {
257 		.num_channels = 1,
258 		.buffer_size = 8,
259 	},
260 	[NBPF1B16] = {
261 		.num_channels = 1,
262 		.buffer_size = 16,
263 	},
264 	[NBPF4B4] = {
265 		.num_channels = 4,
266 		.buffer_size = 4,
267 	},
268 	[NBPF4B8] = {
269 		.num_channels = 4,
270 		.buffer_size = 8,
271 	},
272 	[NBPF4B16] = {
273 		.num_channels = 4,
274 		.buffer_size = 16,
275 	},
276 	[NBPF8B4] = {
277 		.num_channels = 8,
278 		.buffer_size = 4,
279 	},
280 	[NBPF8B8] = {
281 		.num_channels = 8,
282 		.buffer_size = 8,
283 	},
284 	[NBPF8B16] = {
285 		.num_channels = 8,
286 		.buffer_size = 16,
287 	},
288 };
289 
290 #define nbpf_to_chan(d) container_of(d, struct nbpf_channel, dma_chan)
291 
292 /*
293  * dmaengine drivers seem to have a lot in common and instead of sharing more
294  * code, they reimplement those common algorithms independently. In this driver
295  * we try to separate the hardware-specific part from the (largely) generic
296  * part. This improves code readability and makes it possible in the future to
297  * reuse the generic code in form of a helper library. That generic code should
298  * be suitable for various DMA controllers, using transfer descriptors in RAM
299  * and pushing one SG list at a time to the DMA controller.
300  */
301 
302 /*		Hardware-specific part		*/
303 
304 static inline u32 nbpf_chan_read(struct nbpf_channel *chan,
305 				 unsigned int offset)
306 {
307 	u32 data = ioread32(chan->base + offset);
308 	dev_dbg(chan->dma_chan.device->dev, "%s(0x%p + 0x%x) = 0x%x\n",
309 		__func__, chan->base, offset, data);
310 	return data;
311 }
312 
313 static inline void nbpf_chan_write(struct nbpf_channel *chan,
314 				   unsigned int offset, u32 data)
315 {
316 	iowrite32(data, chan->base + offset);
317 	dev_dbg(chan->dma_chan.device->dev, "%s(0x%p + 0x%x) = 0x%x\n",
318 		__func__, chan->base, offset, data);
319 }
320 
321 static inline u32 nbpf_read(struct nbpf_device *nbpf,
322 			    unsigned int offset)
323 {
324 	u32 data = ioread32(nbpf->base + offset);
325 	dev_dbg(nbpf->dma_dev.dev, "%s(0x%p + 0x%x) = 0x%x\n",
326 		__func__, nbpf->base, offset, data);
327 	return data;
328 }
329 
330 static inline void nbpf_write(struct nbpf_device *nbpf,
331 			      unsigned int offset, u32 data)
332 {
333 	iowrite32(data, nbpf->base + offset);
334 	dev_dbg(nbpf->dma_dev.dev, "%s(0x%p + 0x%x) = 0x%x\n",
335 		__func__, nbpf->base, offset, data);
336 }
337 
338 static void nbpf_chan_halt(struct nbpf_channel *chan)
339 {
340 	nbpf_chan_write(chan, NBPF_CHAN_CTRL, NBPF_CHAN_CTRL_CLREN);
341 }
342 
343 static bool nbpf_status_get(struct nbpf_channel *chan)
344 {
345 	u32 status = nbpf_read(chan->nbpf, NBPF_DSTAT_END);
346 
347 	return status & BIT(chan - chan->nbpf->chan);
348 }
349 
350 static void nbpf_status_ack(struct nbpf_channel *chan)
351 {
352 	nbpf_chan_write(chan, NBPF_CHAN_CTRL, NBPF_CHAN_CTRL_CLREND);
353 }
354 
355 static u32 nbpf_error_get(struct nbpf_device *nbpf)
356 {
357 	return nbpf_read(nbpf, NBPF_DSTAT_ER);
358 }
359 
360 static struct nbpf_channel *nbpf_error_get_channel(struct nbpf_device *nbpf, u32 error)
361 {
362 	return nbpf->chan + __ffs(error);
363 }
364 
365 static void nbpf_error_clear(struct nbpf_channel *chan)
366 {
367 	u32 status;
368 	int i;
369 
370 	/* Stop the channel, make sure DMA has been aborted */
371 	nbpf_chan_halt(chan);
372 
373 	for (i = 1000; i; i--) {
374 		status = nbpf_chan_read(chan, NBPF_CHAN_STAT);
375 		if (!(status & NBPF_CHAN_STAT_TACT))
376 			break;
377 		cpu_relax();
378 	}
379 
380 	if (!i)
381 		dev_err(chan->dma_chan.device->dev,
382 			"%s(): abort timeout, channel status 0x%x\n", __func__, status);
383 
384 	nbpf_chan_write(chan, NBPF_CHAN_CTRL, NBPF_CHAN_CTRL_SWRST);
385 }
386 
387 static int nbpf_start(struct nbpf_desc *desc)
388 {
389 	struct nbpf_channel *chan = desc->chan;
390 	struct nbpf_link_desc *ldesc = list_first_entry(&desc->sg, struct nbpf_link_desc, node);
391 
392 	nbpf_chan_write(chan, NBPF_CHAN_NXLA, (u32)ldesc->hwdesc_dma_addr);
393 	nbpf_chan_write(chan, NBPF_CHAN_CTRL, NBPF_CHAN_CTRL_SETEN | NBPF_CHAN_CTRL_CLRSUS);
394 	chan->paused = false;
395 
396 	/* Software trigger MEMCPY - only MEMCPY uses the block mode */
397 	if (ldesc->hwdesc->config & NBPF_CHAN_CFG_TM)
398 		nbpf_chan_write(chan, NBPF_CHAN_CTRL, NBPF_CHAN_CTRL_STG);
399 
400 	dev_dbg(chan->nbpf->dma_dev.dev, "%s(): next 0x%x, cur 0x%x\n", __func__,
401 		nbpf_chan_read(chan, NBPF_CHAN_NXLA), nbpf_chan_read(chan, NBPF_CHAN_CRLA));
402 
403 	return 0;
404 }
405 
406 static void nbpf_chan_prepare(struct nbpf_channel *chan)
407 {
408 	chan->dmarq_cfg = (chan->flags & NBPF_SLAVE_RQ_HIGH ? NBPF_CHAN_CFG_HIEN : 0) |
409 		(chan->flags & NBPF_SLAVE_RQ_LOW ? NBPF_CHAN_CFG_LOEN : 0) |
410 		(chan->flags & NBPF_SLAVE_RQ_LEVEL ?
411 		 NBPF_CHAN_CFG_LVL | (NBPF_CHAN_CFG_AM & 0x200) : 0) |
412 		chan->terminal;
413 }
414 
415 static void nbpf_chan_prepare_default(struct nbpf_channel *chan)
416 {
417 	/* Don't output DMAACK */
418 	chan->dmarq_cfg = NBPF_CHAN_CFG_AM & 0x400;
419 	chan->terminal = 0;
420 	chan->flags = 0;
421 }
422 
423 static void nbpf_chan_configure(struct nbpf_channel *chan)
424 {
425 	/*
426 	 * We assume, that only the link mode and DMA request line configuration
427 	 * have to be set in the configuration register manually. Dynamic
428 	 * per-transfer configuration will be loaded from transfer descriptors.
429 	 */
430 	nbpf_chan_write(chan, NBPF_CHAN_CFG, NBPF_CHAN_CFG_DMS | chan->dmarq_cfg);
431 }
432 
433 static u32 nbpf_xfer_ds(struct nbpf_device *nbpf, size_t size,
434 			enum dma_transfer_direction direction)
435 {
436 	int max_burst = nbpf->config->buffer_size * 8;
437 
438 	if (nbpf->max_burst_mem_read || nbpf->max_burst_mem_write) {
439 		switch (direction) {
440 		case DMA_MEM_TO_MEM:
441 			max_burst = min_not_zero(nbpf->max_burst_mem_read,
442 						 nbpf->max_burst_mem_write);
443 			break;
444 		case DMA_MEM_TO_DEV:
445 			if (nbpf->max_burst_mem_read)
446 				max_burst = nbpf->max_burst_mem_read;
447 			break;
448 		case DMA_DEV_TO_MEM:
449 			if (nbpf->max_burst_mem_write)
450 				max_burst = nbpf->max_burst_mem_write;
451 			break;
452 		case DMA_DEV_TO_DEV:
453 		default:
454 			break;
455 		}
456 	}
457 
458 	/* Maximum supported bursts depend on the buffer size */
459 	return min_t(int, __ffs(size), ilog2(max_burst));
460 }
461 
462 static size_t nbpf_xfer_size(struct nbpf_device *nbpf,
463 			     enum dma_slave_buswidth width, u32 burst)
464 {
465 	size_t size;
466 
467 	if (!burst)
468 		burst = 1;
469 
470 	switch (width) {
471 	case DMA_SLAVE_BUSWIDTH_8_BYTES:
472 		size = 8 * burst;
473 		break;
474 
475 	case DMA_SLAVE_BUSWIDTH_4_BYTES:
476 		size = 4 * burst;
477 		break;
478 
479 	case DMA_SLAVE_BUSWIDTH_2_BYTES:
480 		size = 2 * burst;
481 		break;
482 
483 	default:
484 		pr_warn("%s(): invalid bus width %u\n", __func__, width);
485 		fallthrough;
486 	case DMA_SLAVE_BUSWIDTH_1_BYTE:
487 		size = burst;
488 	}
489 
490 	return nbpf_xfer_ds(nbpf, size, DMA_TRANS_NONE);
491 }
492 
493 /*
494  * We need a way to recognise slaves, whose data is sent "raw" over the bus,
495  * i.e. it isn't known in advance how many bytes will be received. Therefore
496  * the slave driver has to provide a "large enough" buffer and either read the
497  * buffer, when it is full, or detect, that some data has arrived, then wait for
498  * a timeout, if no more data arrives - receive what's already there. We want to
499  * handle such slaves in a special way to allow an optimised mode for other
500  * users, for whom the amount of data is known in advance. So far there's no way
501  * to recognise such slaves. We use a data-width check to distinguish between
502  * the SD host and the PL011 UART.
503  */
504 
505 static int nbpf_prep_one(struct nbpf_link_desc *ldesc,
506 			 enum dma_transfer_direction direction,
507 			 dma_addr_t src, dma_addr_t dst, size_t size, bool last)
508 {
509 	struct nbpf_link_reg *hwdesc = ldesc->hwdesc;
510 	struct nbpf_desc *desc = ldesc->desc;
511 	struct nbpf_channel *chan = desc->chan;
512 	struct device *dev = chan->dma_chan.device->dev;
513 	size_t mem_xfer, slave_xfer;
514 	bool can_burst;
515 
516 	hwdesc->header = NBPF_HEADER_WBD | NBPF_HEADER_LV |
517 		(last ? NBPF_HEADER_LE : 0);
518 
519 	hwdesc->src_addr = src;
520 	hwdesc->dst_addr = dst;
521 	hwdesc->transaction_size = size;
522 
523 	/*
524 	 * set config: SAD, DAD, DDS, SDS, etc.
525 	 * Note on transfer sizes: the DMAC can perform unaligned DMA transfers,
526 	 * but it is important to have transaction size a multiple of both
527 	 * receiver and transmitter transfer sizes. It is also possible to use
528 	 * different RAM and device transfer sizes, and it does work well with
529 	 * some devices, e.g. with V08R07S01E SD host controllers, which can use
530 	 * 128 byte transfers. But this doesn't work with other devices,
531 	 * especially when the transaction size is unknown. This is the case,
532 	 * e.g. with serial drivers like amba-pl011.c. For reception it sets up
533 	 * the transaction size of 4K and if fewer bytes are received, it
534 	 * pauses DMA and reads out data received via DMA as well as those left
535 	 * in the Rx FIFO. For this to work with the RAM side using burst
536 	 * transfers we enable the SBE bit and terminate the transfer in our
537 	 * .device_pause handler.
538 	 */
539 	mem_xfer = nbpf_xfer_ds(chan->nbpf, size, direction);
540 
541 	switch (direction) {
542 	case DMA_DEV_TO_MEM:
543 		can_burst = chan->slave_src_width >= 3;
544 		slave_xfer = min(mem_xfer, can_burst ?
545 				 chan->slave_src_burst : chan->slave_src_width);
546 		/*
547 		 * Is the slave narrower than 64 bits, i.e. isn't using the full
548 		 * bus width and cannot use bursts?
549 		 */
550 		if (mem_xfer > chan->slave_src_burst && !can_burst)
551 			mem_xfer = chan->slave_src_burst;
552 		/* Device-to-RAM DMA is unreliable without REQD set */
553 		hwdesc->config = NBPF_CHAN_CFG_SAD | (NBPF_CHAN_CFG_DDS & (mem_xfer << 16)) |
554 			(NBPF_CHAN_CFG_SDS & (slave_xfer << 12)) | NBPF_CHAN_CFG_REQD |
555 			NBPF_CHAN_CFG_SBE;
556 		break;
557 
558 	case DMA_MEM_TO_DEV:
559 		slave_xfer = min(mem_xfer, chan->slave_dst_width >= 3 ?
560 				 chan->slave_dst_burst : chan->slave_dst_width);
561 		hwdesc->config = NBPF_CHAN_CFG_DAD | (NBPF_CHAN_CFG_SDS & (mem_xfer << 12)) |
562 			(NBPF_CHAN_CFG_DDS & (slave_xfer << 16)) | NBPF_CHAN_CFG_REQD;
563 		break;
564 
565 	case DMA_MEM_TO_MEM:
566 		hwdesc->config = NBPF_CHAN_CFG_TCM | NBPF_CHAN_CFG_TM |
567 			(NBPF_CHAN_CFG_SDS & (mem_xfer << 12)) |
568 			(NBPF_CHAN_CFG_DDS & (mem_xfer << 16));
569 		break;
570 
571 	default:
572 		return -EINVAL;
573 	}
574 
575 	hwdesc->config |= chan->dmarq_cfg | (last ? 0 : NBPF_CHAN_CFG_DEM) |
576 		NBPF_CHAN_CFG_DMS;
577 
578 	dev_dbg(dev, "%s(): desc @ %pad: hdr 0x%x, cfg 0x%x, %zu @ %pad -> %pad\n",
579 		__func__, &ldesc->hwdesc_dma_addr, hwdesc->header,
580 		hwdesc->config, size, &src, &dst);
581 
582 	dma_sync_single_for_device(dev, ldesc->hwdesc_dma_addr, sizeof(*hwdesc),
583 				   DMA_TO_DEVICE);
584 
585 	return 0;
586 }
587 
588 static size_t nbpf_bytes_left(struct nbpf_channel *chan)
589 {
590 	return nbpf_chan_read(chan, NBPF_CHAN_CUR_TR_BYTE);
591 }
592 
593 static void nbpf_configure(struct nbpf_device *nbpf)
594 {
595 	nbpf_write(nbpf, NBPF_CTRL, NBPF_CTRL_LVINT);
596 }
597 
598 /*		Generic part			*/
599 
600 /* DMA ENGINE functions */
601 static void nbpf_issue_pending(struct dma_chan *dchan)
602 {
603 	struct nbpf_channel *chan = nbpf_to_chan(dchan);
604 	unsigned long flags;
605 
606 	dev_dbg(dchan->device->dev, "Entry %s()\n", __func__);
607 
608 	spin_lock_irqsave(&chan->lock, flags);
609 	if (list_empty(&chan->queued))
610 		goto unlock;
611 
612 	list_splice_tail_init(&chan->queued, &chan->active);
613 
614 	if (!chan->running) {
615 		struct nbpf_desc *desc = list_first_entry(&chan->active,
616 						struct nbpf_desc, node);
617 		if (!nbpf_start(desc))
618 			chan->running = desc;
619 	}
620 
621 unlock:
622 	spin_unlock_irqrestore(&chan->lock, flags);
623 }
624 
625 static enum dma_status nbpf_tx_status(struct dma_chan *dchan,
626 		dma_cookie_t cookie, struct dma_tx_state *state)
627 {
628 	struct nbpf_channel *chan = nbpf_to_chan(dchan);
629 	enum dma_status status = dma_cookie_status(dchan, cookie, state);
630 
631 	if (state) {
632 		dma_cookie_t running;
633 		unsigned long flags;
634 
635 		spin_lock_irqsave(&chan->lock, flags);
636 		running = chan->running ? chan->running->async_tx.cookie : -EINVAL;
637 
638 		if (cookie == running) {
639 			state->residue = nbpf_bytes_left(chan);
640 			dev_dbg(dchan->device->dev, "%s(): residue %u\n", __func__,
641 				state->residue);
642 		} else if (status == DMA_IN_PROGRESS) {
643 			struct nbpf_desc *desc;
644 			bool found = false;
645 
646 			list_for_each_entry(desc, &chan->active, node)
647 				if (desc->async_tx.cookie == cookie) {
648 					found = true;
649 					break;
650 				}
651 
652 			if (!found)
653 				list_for_each_entry(desc, &chan->queued, node)
654 					if (desc->async_tx.cookie == cookie) {
655 						found = true;
656 						break;
657 
658 					}
659 
660 			state->residue = found ? desc->length : 0;
661 		}
662 
663 		spin_unlock_irqrestore(&chan->lock, flags);
664 	}
665 
666 	if (chan->paused)
667 		status = DMA_PAUSED;
668 
669 	return status;
670 }
671 
672 static dma_cookie_t nbpf_tx_submit(struct dma_async_tx_descriptor *tx)
673 {
674 	struct nbpf_desc *desc = container_of(tx, struct nbpf_desc, async_tx);
675 	struct nbpf_channel *chan = desc->chan;
676 	unsigned long flags;
677 	dma_cookie_t cookie;
678 
679 	spin_lock_irqsave(&chan->lock, flags);
680 	cookie = dma_cookie_assign(tx);
681 	list_add_tail(&desc->node, &chan->queued);
682 	spin_unlock_irqrestore(&chan->lock, flags);
683 
684 	dev_dbg(chan->dma_chan.device->dev, "Entry %s(%d)\n", __func__, cookie);
685 
686 	return cookie;
687 }
688 
689 static int nbpf_desc_page_alloc(struct nbpf_channel *chan)
690 {
691 	struct dma_chan *dchan = &chan->dma_chan;
692 	struct nbpf_desc_page *dpage = (void *)get_zeroed_page(GFP_KERNEL | GFP_DMA);
693 	struct nbpf_link_desc *ldesc;
694 	struct nbpf_link_reg *hwdesc;
695 	struct nbpf_desc *desc;
696 	LIST_HEAD(head);
697 	LIST_HEAD(lhead);
698 	int i;
699 	struct device *dev = dchan->device->dev;
700 
701 	if (!dpage)
702 		return -ENOMEM;
703 
704 	dev_dbg(dev, "%s(): alloc %lu descriptors, %lu segments, total alloc %zu\n",
705 		__func__, NBPF_DESCS_PER_PAGE, NBPF_SEGMENTS_PER_PAGE, sizeof(*dpage));
706 
707 	for (i = 0, ldesc = dpage->ldesc, hwdesc = dpage->hwdesc;
708 	     i < ARRAY_SIZE(dpage->ldesc);
709 	     i++, ldesc++, hwdesc++) {
710 		ldesc->hwdesc = hwdesc;
711 		list_add_tail(&ldesc->node, &lhead);
712 		ldesc->hwdesc_dma_addr = dma_map_single(dchan->device->dev,
713 					hwdesc, sizeof(*hwdesc), DMA_TO_DEVICE);
714 
715 		dev_dbg(dev, "%s(): mapped 0x%p to %pad\n", __func__,
716 			hwdesc, &ldesc->hwdesc_dma_addr);
717 	}
718 
719 	for (i = 0, desc = dpage->desc;
720 	     i < ARRAY_SIZE(dpage->desc);
721 	     i++, desc++) {
722 		dma_async_tx_descriptor_init(&desc->async_tx, dchan);
723 		desc->async_tx.tx_submit = nbpf_tx_submit;
724 		desc->chan = chan;
725 		INIT_LIST_HEAD(&desc->sg);
726 		list_add_tail(&desc->node, &head);
727 	}
728 
729 	/*
730 	 * This function cannot be called from interrupt context, so, no need to
731 	 * save flags
732 	 */
733 	spin_lock_irq(&chan->lock);
734 	list_splice_tail(&lhead, &chan->free_links);
735 	list_splice_tail(&head, &chan->free);
736 	list_add(&dpage->node, &chan->desc_page);
737 	spin_unlock_irq(&chan->lock);
738 
739 	return ARRAY_SIZE(dpage->desc);
740 }
741 
742 static void nbpf_desc_put(struct nbpf_desc *desc)
743 {
744 	struct nbpf_channel *chan = desc->chan;
745 	struct nbpf_link_desc *ldesc, *tmp;
746 	unsigned long flags;
747 
748 	spin_lock_irqsave(&chan->lock, flags);
749 	list_for_each_entry_safe(ldesc, tmp, &desc->sg, node)
750 		list_move(&ldesc->node, &chan->free_links);
751 
752 	list_add(&desc->node, &chan->free);
753 	spin_unlock_irqrestore(&chan->lock, flags);
754 }
755 
756 static void nbpf_scan_acked(struct nbpf_channel *chan)
757 {
758 	struct nbpf_desc *desc, *tmp;
759 	unsigned long flags;
760 	LIST_HEAD(head);
761 
762 	spin_lock_irqsave(&chan->lock, flags);
763 	list_for_each_entry_safe(desc, tmp, &chan->done, node)
764 		if (async_tx_test_ack(&desc->async_tx) && desc->user_wait) {
765 			list_move(&desc->node, &head);
766 			desc->user_wait = false;
767 		}
768 	spin_unlock_irqrestore(&chan->lock, flags);
769 
770 	list_for_each_entry_safe(desc, tmp, &head, node) {
771 		list_del(&desc->node);
772 		nbpf_desc_put(desc);
773 	}
774 }
775 
776 /*
777  * We have to allocate descriptors with the channel lock dropped. This means,
778  * before we re-acquire the lock buffers can be taken already, so we have to
779  * re-check after re-acquiring the lock and possibly retry, if buffers are gone
780  * again.
781  */
782 static struct nbpf_desc *nbpf_desc_get(struct nbpf_channel *chan, size_t len)
783 {
784 	struct nbpf_desc *desc = NULL;
785 	struct nbpf_link_desc *ldesc, *prev = NULL;
786 
787 	nbpf_scan_acked(chan);
788 
789 	spin_lock_irq(&chan->lock);
790 
791 	do {
792 		int i = 0, ret;
793 
794 		if (list_empty(&chan->free)) {
795 			/* No more free descriptors */
796 			spin_unlock_irq(&chan->lock);
797 			ret = nbpf_desc_page_alloc(chan);
798 			if (ret < 0)
799 				return NULL;
800 			spin_lock_irq(&chan->lock);
801 			continue;
802 		}
803 		desc = list_first_entry(&chan->free, struct nbpf_desc, node);
804 		list_del(&desc->node);
805 
806 		do {
807 			if (list_empty(&chan->free_links)) {
808 				/* No more free link descriptors */
809 				spin_unlock_irq(&chan->lock);
810 				ret = nbpf_desc_page_alloc(chan);
811 				if (ret < 0) {
812 					nbpf_desc_put(desc);
813 					return NULL;
814 				}
815 				spin_lock_irq(&chan->lock);
816 				continue;
817 			}
818 
819 			ldesc = list_first_entry(&chan->free_links,
820 						 struct nbpf_link_desc, node);
821 			ldesc->desc = desc;
822 			if (prev)
823 				prev->hwdesc->next = (u32)ldesc->hwdesc_dma_addr;
824 
825 			prev = ldesc;
826 			list_move_tail(&ldesc->node, &desc->sg);
827 
828 			i++;
829 		} while (i < len);
830 	} while (!desc);
831 
832 	prev->hwdesc->next = 0;
833 
834 	spin_unlock_irq(&chan->lock);
835 
836 	return desc;
837 }
838 
839 static void nbpf_chan_idle(struct nbpf_channel *chan)
840 {
841 	struct nbpf_desc *desc, *tmp;
842 	unsigned long flags;
843 	LIST_HEAD(head);
844 
845 	spin_lock_irqsave(&chan->lock, flags);
846 
847 	list_splice_init(&chan->done, &head);
848 	list_splice_init(&chan->active, &head);
849 	list_splice_init(&chan->queued, &head);
850 
851 	chan->running = NULL;
852 
853 	spin_unlock_irqrestore(&chan->lock, flags);
854 
855 	list_for_each_entry_safe(desc, tmp, &head, node) {
856 		dev_dbg(chan->nbpf->dma_dev.dev, "%s(): force-free desc %p cookie %d\n",
857 			__func__, desc, desc->async_tx.cookie);
858 		list_del(&desc->node);
859 		nbpf_desc_put(desc);
860 	}
861 }
862 
863 static int nbpf_pause(struct dma_chan *dchan)
864 {
865 	struct nbpf_channel *chan = nbpf_to_chan(dchan);
866 
867 	dev_dbg(dchan->device->dev, "Entry %s\n", __func__);
868 
869 	chan->paused = true;
870 	nbpf_chan_write(chan, NBPF_CHAN_CTRL, NBPF_CHAN_CTRL_SETSUS);
871 	/* See comment in nbpf_prep_one() */
872 	nbpf_chan_write(chan, NBPF_CHAN_CTRL, NBPF_CHAN_CTRL_CLREN);
873 
874 	return 0;
875 }
876 
877 static int nbpf_terminate_all(struct dma_chan *dchan)
878 {
879 	struct nbpf_channel *chan = nbpf_to_chan(dchan);
880 
881 	dev_dbg(dchan->device->dev, "Entry %s\n", __func__);
882 	dev_dbg(dchan->device->dev, "Terminating\n");
883 
884 	nbpf_chan_halt(chan);
885 	nbpf_chan_idle(chan);
886 
887 	return 0;
888 }
889 
890 static int nbpf_config(struct dma_chan *dchan,
891 		       struct dma_slave_config *config)
892 {
893 	struct nbpf_channel *chan = nbpf_to_chan(dchan);
894 
895 	dev_dbg(dchan->device->dev, "Entry %s\n", __func__);
896 
897 	/*
898 	 * We could check config->slave_id to match chan->terminal here,
899 	 * but with DT they would be coming from the same source, so
900 	 * such a check would be superflous
901 	 */
902 
903 	chan->slave_dst_addr = config->dst_addr;
904 	chan->slave_dst_width = nbpf_xfer_size(chan->nbpf,
905 					       config->dst_addr_width, 1);
906 	chan->slave_dst_burst = nbpf_xfer_size(chan->nbpf,
907 					       config->dst_addr_width,
908 					       config->dst_maxburst);
909 	chan->slave_src_addr = config->src_addr;
910 	chan->slave_src_width = nbpf_xfer_size(chan->nbpf,
911 					       config->src_addr_width, 1);
912 	chan->slave_src_burst = nbpf_xfer_size(chan->nbpf,
913 					       config->src_addr_width,
914 					       config->src_maxburst);
915 
916 	return 0;
917 }
918 
919 static struct dma_async_tx_descriptor *nbpf_prep_sg(struct nbpf_channel *chan,
920 		struct scatterlist *src_sg, struct scatterlist *dst_sg,
921 		size_t len, enum dma_transfer_direction direction,
922 		unsigned long flags)
923 {
924 	struct nbpf_link_desc *ldesc;
925 	struct scatterlist *mem_sg;
926 	struct nbpf_desc *desc;
927 	bool inc_src, inc_dst;
928 	size_t data_len = 0;
929 	int i = 0;
930 
931 	switch (direction) {
932 	case DMA_DEV_TO_MEM:
933 		mem_sg = dst_sg;
934 		inc_src = false;
935 		inc_dst = true;
936 		break;
937 
938 	case DMA_MEM_TO_DEV:
939 		mem_sg = src_sg;
940 		inc_src = true;
941 		inc_dst = false;
942 		break;
943 
944 	default:
945 	case DMA_MEM_TO_MEM:
946 		mem_sg = src_sg;
947 		inc_src = true;
948 		inc_dst = true;
949 	}
950 
951 	desc = nbpf_desc_get(chan, len);
952 	if (!desc)
953 		return NULL;
954 
955 	desc->async_tx.flags = flags;
956 	desc->async_tx.cookie = -EBUSY;
957 	desc->user_wait = false;
958 
959 	/*
960 	 * This is a private descriptor list, and we own the descriptor. No need
961 	 * to lock.
962 	 */
963 	list_for_each_entry(ldesc, &desc->sg, node) {
964 		int ret = nbpf_prep_one(ldesc, direction,
965 					sg_dma_address(src_sg),
966 					sg_dma_address(dst_sg),
967 					sg_dma_len(mem_sg),
968 					i == len - 1);
969 		if (ret < 0) {
970 			nbpf_desc_put(desc);
971 			return NULL;
972 		}
973 		data_len += sg_dma_len(mem_sg);
974 		if (inc_src)
975 			src_sg = sg_next(src_sg);
976 		if (inc_dst)
977 			dst_sg = sg_next(dst_sg);
978 		mem_sg = direction == DMA_DEV_TO_MEM ? dst_sg : src_sg;
979 		i++;
980 	}
981 
982 	desc->length = data_len;
983 
984 	/* The user has to return the descriptor to us ASAP via .tx_submit() */
985 	return &desc->async_tx;
986 }
987 
988 static struct dma_async_tx_descriptor *nbpf_prep_memcpy(
989 	struct dma_chan *dchan, dma_addr_t dst, dma_addr_t src,
990 	size_t len, unsigned long flags)
991 {
992 	struct nbpf_channel *chan = nbpf_to_chan(dchan);
993 	struct scatterlist dst_sg;
994 	struct scatterlist src_sg;
995 
996 	sg_init_table(&dst_sg, 1);
997 	sg_init_table(&src_sg, 1);
998 
999 	sg_dma_address(&dst_sg) = dst;
1000 	sg_dma_address(&src_sg) = src;
1001 
1002 	sg_dma_len(&dst_sg) = len;
1003 	sg_dma_len(&src_sg) = len;
1004 
1005 	dev_dbg(dchan->device->dev, "%s(): %zu @ %pad -> %pad\n",
1006 		__func__, len, &src, &dst);
1007 
1008 	return nbpf_prep_sg(chan, &src_sg, &dst_sg, 1,
1009 			    DMA_MEM_TO_MEM, flags);
1010 }
1011 
1012 static struct dma_async_tx_descriptor *nbpf_prep_slave_sg(
1013 	struct dma_chan *dchan, struct scatterlist *sgl, unsigned int sg_len,
1014 	enum dma_transfer_direction direction, unsigned long flags, void *context)
1015 {
1016 	struct nbpf_channel *chan = nbpf_to_chan(dchan);
1017 	struct scatterlist slave_sg;
1018 
1019 	dev_dbg(dchan->device->dev, "Entry %s()\n", __func__);
1020 
1021 	sg_init_table(&slave_sg, 1);
1022 
1023 	switch (direction) {
1024 	case DMA_MEM_TO_DEV:
1025 		sg_dma_address(&slave_sg) = chan->slave_dst_addr;
1026 		return nbpf_prep_sg(chan, sgl, &slave_sg, sg_len,
1027 				    direction, flags);
1028 
1029 	case DMA_DEV_TO_MEM:
1030 		sg_dma_address(&slave_sg) = chan->slave_src_addr;
1031 		return nbpf_prep_sg(chan, &slave_sg, sgl, sg_len,
1032 				    direction, flags);
1033 
1034 	default:
1035 		return NULL;
1036 	}
1037 }
1038 
1039 static int nbpf_alloc_chan_resources(struct dma_chan *dchan)
1040 {
1041 	struct nbpf_channel *chan = nbpf_to_chan(dchan);
1042 	int ret;
1043 
1044 	INIT_LIST_HEAD(&chan->free);
1045 	INIT_LIST_HEAD(&chan->free_links);
1046 	INIT_LIST_HEAD(&chan->queued);
1047 	INIT_LIST_HEAD(&chan->active);
1048 	INIT_LIST_HEAD(&chan->done);
1049 
1050 	ret = nbpf_desc_page_alloc(chan);
1051 	if (ret < 0)
1052 		return ret;
1053 
1054 	dev_dbg(dchan->device->dev, "Entry %s(): terminal %u\n", __func__,
1055 		chan->terminal);
1056 
1057 	nbpf_chan_configure(chan);
1058 
1059 	return ret;
1060 }
1061 
1062 static void nbpf_free_chan_resources(struct dma_chan *dchan)
1063 {
1064 	struct nbpf_channel *chan = nbpf_to_chan(dchan);
1065 	struct nbpf_desc_page *dpage, *tmp;
1066 
1067 	dev_dbg(dchan->device->dev, "Entry %s()\n", __func__);
1068 
1069 	nbpf_chan_halt(chan);
1070 	nbpf_chan_idle(chan);
1071 	/* Clean up for if a channel is re-used for MEMCPY after slave DMA */
1072 	nbpf_chan_prepare_default(chan);
1073 
1074 	list_for_each_entry_safe(dpage, tmp, &chan->desc_page, node) {
1075 		struct nbpf_link_desc *ldesc;
1076 		int i;
1077 		list_del(&dpage->node);
1078 		for (i = 0, ldesc = dpage->ldesc;
1079 		     i < ARRAY_SIZE(dpage->ldesc);
1080 		     i++, ldesc++)
1081 			dma_unmap_single(dchan->device->dev, ldesc->hwdesc_dma_addr,
1082 					 sizeof(*ldesc->hwdesc), DMA_TO_DEVICE);
1083 		free_page((unsigned long)dpage);
1084 	}
1085 }
1086 
1087 static struct dma_chan *nbpf_of_xlate(struct of_phandle_args *dma_spec,
1088 				      struct of_dma *ofdma)
1089 {
1090 	struct nbpf_device *nbpf = ofdma->of_dma_data;
1091 	struct dma_chan *dchan;
1092 	struct nbpf_channel *chan;
1093 
1094 	if (dma_spec->args_count != 2)
1095 		return NULL;
1096 
1097 	dchan = dma_get_any_slave_channel(&nbpf->dma_dev);
1098 	if (!dchan)
1099 		return NULL;
1100 
1101 	dev_dbg(dchan->device->dev, "Entry %s(%pOFn)\n", __func__,
1102 		dma_spec->np);
1103 
1104 	chan = nbpf_to_chan(dchan);
1105 
1106 	chan->terminal = dma_spec->args[0];
1107 	chan->flags = dma_spec->args[1];
1108 
1109 	nbpf_chan_prepare(chan);
1110 	nbpf_chan_configure(chan);
1111 
1112 	return dchan;
1113 }
1114 
1115 static void nbpf_chan_tasklet(struct tasklet_struct *t)
1116 {
1117 	struct nbpf_channel *chan = from_tasklet(chan, t, tasklet);
1118 	struct nbpf_desc *desc, *tmp;
1119 	struct dmaengine_desc_callback cb;
1120 
1121 	while (!list_empty(&chan->done)) {
1122 		bool found = false, must_put, recycling = false;
1123 
1124 		spin_lock_irq(&chan->lock);
1125 
1126 		list_for_each_entry_safe(desc, tmp, &chan->done, node) {
1127 			if (!desc->user_wait) {
1128 				/* Newly completed descriptor, have to process */
1129 				found = true;
1130 				break;
1131 			} else if (async_tx_test_ack(&desc->async_tx)) {
1132 				/*
1133 				 * This descriptor was waiting for a user ACK,
1134 				 * it can be recycled now.
1135 				 */
1136 				list_del(&desc->node);
1137 				spin_unlock_irq(&chan->lock);
1138 				nbpf_desc_put(desc);
1139 				recycling = true;
1140 				break;
1141 			}
1142 		}
1143 
1144 		if (recycling)
1145 			continue;
1146 
1147 		if (!found) {
1148 			/* This can happen if TERMINATE_ALL has been called */
1149 			spin_unlock_irq(&chan->lock);
1150 			break;
1151 		}
1152 
1153 		dma_cookie_complete(&desc->async_tx);
1154 
1155 		/*
1156 		 * With released lock we cannot dereference desc, maybe it's
1157 		 * still on the "done" list
1158 		 */
1159 		if (async_tx_test_ack(&desc->async_tx)) {
1160 			list_del(&desc->node);
1161 			must_put = true;
1162 		} else {
1163 			desc->user_wait = true;
1164 			must_put = false;
1165 		}
1166 
1167 		dmaengine_desc_get_callback(&desc->async_tx, &cb);
1168 
1169 		/* ack and callback completed descriptor */
1170 		spin_unlock_irq(&chan->lock);
1171 
1172 		dmaengine_desc_callback_invoke(&cb, NULL);
1173 
1174 		if (must_put)
1175 			nbpf_desc_put(desc);
1176 	}
1177 }
1178 
1179 static irqreturn_t nbpf_chan_irq(int irq, void *dev)
1180 {
1181 	struct nbpf_channel *chan = dev;
1182 	bool done = nbpf_status_get(chan);
1183 	struct nbpf_desc *desc;
1184 	irqreturn_t ret;
1185 	bool bh = false;
1186 
1187 	if (!done)
1188 		return IRQ_NONE;
1189 
1190 	nbpf_status_ack(chan);
1191 
1192 	dev_dbg(&chan->dma_chan.dev->device, "%s()\n", __func__);
1193 
1194 	spin_lock(&chan->lock);
1195 	desc = chan->running;
1196 	if (WARN_ON(!desc)) {
1197 		ret = IRQ_NONE;
1198 		goto unlock;
1199 	} else {
1200 		ret = IRQ_HANDLED;
1201 		bh = true;
1202 	}
1203 
1204 	list_move_tail(&desc->node, &chan->done);
1205 	chan->running = NULL;
1206 
1207 	if (!list_empty(&chan->active)) {
1208 		desc = list_first_entry(&chan->active,
1209 					struct nbpf_desc, node);
1210 		if (!nbpf_start(desc))
1211 			chan->running = desc;
1212 	}
1213 
1214 unlock:
1215 	spin_unlock(&chan->lock);
1216 
1217 	if (bh)
1218 		tasklet_schedule(&chan->tasklet);
1219 
1220 	return ret;
1221 }
1222 
1223 static irqreturn_t nbpf_err_irq(int irq, void *dev)
1224 {
1225 	struct nbpf_device *nbpf = dev;
1226 	u32 error = nbpf_error_get(nbpf);
1227 
1228 	dev_warn(nbpf->dma_dev.dev, "DMA error IRQ %u\n", irq);
1229 
1230 	if (!error)
1231 		return IRQ_NONE;
1232 
1233 	do {
1234 		struct nbpf_channel *chan = nbpf_error_get_channel(nbpf, error);
1235 		/* On error: abort all queued transfers, no callback */
1236 		nbpf_error_clear(chan);
1237 		nbpf_chan_idle(chan);
1238 		error = nbpf_error_get(nbpf);
1239 	} while (error);
1240 
1241 	return IRQ_HANDLED;
1242 }
1243 
1244 static int nbpf_chan_probe(struct nbpf_device *nbpf, int n)
1245 {
1246 	struct dma_device *dma_dev = &nbpf->dma_dev;
1247 	struct nbpf_channel *chan = nbpf->chan + n;
1248 	int ret;
1249 
1250 	chan->nbpf = nbpf;
1251 	chan->base = nbpf->base + NBPF_REG_CHAN_OFFSET + NBPF_REG_CHAN_SIZE * n;
1252 	INIT_LIST_HEAD(&chan->desc_page);
1253 	spin_lock_init(&chan->lock);
1254 	chan->dma_chan.device = dma_dev;
1255 	dma_cookie_init(&chan->dma_chan);
1256 	nbpf_chan_prepare_default(chan);
1257 
1258 	dev_dbg(dma_dev->dev, "%s(): channel %d: -> %p\n", __func__, n, chan->base);
1259 
1260 	snprintf(chan->name, sizeof(chan->name), "nbpf %d", n);
1261 
1262 	tasklet_setup(&chan->tasklet, nbpf_chan_tasklet);
1263 	ret = devm_request_irq(dma_dev->dev, chan->irq,
1264 			nbpf_chan_irq, IRQF_SHARED,
1265 			chan->name, chan);
1266 	if (ret < 0)
1267 		return ret;
1268 
1269 	/* Add the channel to DMA device channel list */
1270 	list_add_tail(&chan->dma_chan.device_node,
1271 		      &dma_dev->channels);
1272 
1273 	return 0;
1274 }
1275 
1276 static const struct of_device_id nbpf_match[] = {
1277 	{.compatible = "renesas,nbpfaxi64dmac1b4",	.data = &nbpf_cfg[NBPF1B4]},
1278 	{.compatible = "renesas,nbpfaxi64dmac1b8",	.data = &nbpf_cfg[NBPF1B8]},
1279 	{.compatible = "renesas,nbpfaxi64dmac1b16",	.data = &nbpf_cfg[NBPF1B16]},
1280 	{.compatible = "renesas,nbpfaxi64dmac4b4",	.data = &nbpf_cfg[NBPF4B4]},
1281 	{.compatible = "renesas,nbpfaxi64dmac4b8",	.data = &nbpf_cfg[NBPF4B8]},
1282 	{.compatible = "renesas,nbpfaxi64dmac4b16",	.data = &nbpf_cfg[NBPF4B16]},
1283 	{.compatible = "renesas,nbpfaxi64dmac8b4",	.data = &nbpf_cfg[NBPF8B4]},
1284 	{.compatible = "renesas,nbpfaxi64dmac8b8",	.data = &nbpf_cfg[NBPF8B8]},
1285 	{.compatible = "renesas,nbpfaxi64dmac8b16",	.data = &nbpf_cfg[NBPF8B16]},
1286 	{}
1287 };
1288 MODULE_DEVICE_TABLE(of, nbpf_match);
1289 
1290 static int nbpf_probe(struct platform_device *pdev)
1291 {
1292 	struct device *dev = &pdev->dev;
1293 	struct device_node *np = dev->of_node;
1294 	struct nbpf_device *nbpf;
1295 	struct dma_device *dma_dev;
1296 	const struct nbpf_config *cfg;
1297 	int num_channels;
1298 	int ret, irq, eirq, i;
1299 	int irqbuf[9] /* maximum 8 channels + error IRQ */;
1300 	unsigned int irqs = 0;
1301 
1302 	BUILD_BUG_ON(sizeof(struct nbpf_desc_page) > PAGE_SIZE);
1303 
1304 	/* DT only */
1305 	if (!np)
1306 		return -ENODEV;
1307 
1308 	cfg = of_device_get_match_data(dev);
1309 	num_channels = cfg->num_channels;
1310 
1311 	nbpf = devm_kzalloc(dev, struct_size(nbpf, chan, num_channels),
1312 			    GFP_KERNEL);
1313 	if (!nbpf)
1314 		return -ENOMEM;
1315 
1316 	dma_dev = &nbpf->dma_dev;
1317 	dma_dev->dev = dev;
1318 
1319 	nbpf->base = devm_platform_ioremap_resource(pdev, 0);
1320 	if (IS_ERR(nbpf->base))
1321 		return PTR_ERR(nbpf->base);
1322 
1323 	nbpf->clk = devm_clk_get(dev, NULL);
1324 	if (IS_ERR(nbpf->clk))
1325 		return PTR_ERR(nbpf->clk);
1326 
1327 	of_property_read_u32(np, "max-burst-mem-read",
1328 			     &nbpf->max_burst_mem_read);
1329 	of_property_read_u32(np, "max-burst-mem-write",
1330 			     &nbpf->max_burst_mem_write);
1331 
1332 	nbpf->config = cfg;
1333 
1334 	for (i = 0; irqs < ARRAY_SIZE(irqbuf); i++) {
1335 		irq = platform_get_irq_optional(pdev, i);
1336 		if (irq < 0 && irq != -ENXIO)
1337 			return irq;
1338 		if (irq > 0)
1339 			irqbuf[irqs++] = irq;
1340 	}
1341 
1342 	/*
1343 	 * 3 IRQ resource schemes are supported:
1344 	 * 1. 1 shared IRQ for error and all channels
1345 	 * 2. 2 IRQs: one for error and one shared for all channels
1346 	 * 3. 1 IRQ for error and an own IRQ for each channel
1347 	 */
1348 	if (irqs != 1 && irqs != 2 && irqs != num_channels + 1)
1349 		return -ENXIO;
1350 
1351 	if (irqs == 1) {
1352 		eirq = irqbuf[0];
1353 
1354 		for (i = 0; i <= num_channels; i++)
1355 			nbpf->chan[i].irq = irqbuf[0];
1356 	} else {
1357 		eirq = platform_get_irq_byname(pdev, "error");
1358 		if (eirq < 0)
1359 			return eirq;
1360 
1361 		if (irqs == num_channels + 1) {
1362 			struct nbpf_channel *chan;
1363 
1364 			for (i = 0, chan = nbpf->chan; i <= num_channels;
1365 			     i++, chan++) {
1366 				/* Skip the error IRQ */
1367 				if (irqbuf[i] == eirq)
1368 					i++;
1369 				chan->irq = irqbuf[i];
1370 			}
1371 
1372 			if (chan != nbpf->chan + num_channels)
1373 				return -EINVAL;
1374 		} else {
1375 			/* 2 IRQs and more than one channel */
1376 			if (irqbuf[0] == eirq)
1377 				irq = irqbuf[1];
1378 			else
1379 				irq = irqbuf[0];
1380 
1381 			for (i = 0; i <= num_channels; i++)
1382 				nbpf->chan[i].irq = irq;
1383 		}
1384 	}
1385 
1386 	ret = devm_request_irq(dev, eirq, nbpf_err_irq,
1387 			       IRQF_SHARED, "dma error", nbpf);
1388 	if (ret < 0)
1389 		return ret;
1390 	nbpf->eirq = eirq;
1391 
1392 	INIT_LIST_HEAD(&dma_dev->channels);
1393 
1394 	/* Create DMA Channel */
1395 	for (i = 0; i < num_channels; i++) {
1396 		ret = nbpf_chan_probe(nbpf, i);
1397 		if (ret < 0)
1398 			return ret;
1399 	}
1400 
1401 	dma_cap_set(DMA_MEMCPY, dma_dev->cap_mask);
1402 	dma_cap_set(DMA_SLAVE, dma_dev->cap_mask);
1403 	dma_cap_set(DMA_PRIVATE, dma_dev->cap_mask);
1404 
1405 	/* Common and MEMCPY operations */
1406 	dma_dev->device_alloc_chan_resources
1407 		= nbpf_alloc_chan_resources;
1408 	dma_dev->device_free_chan_resources = nbpf_free_chan_resources;
1409 	dma_dev->device_prep_dma_memcpy = nbpf_prep_memcpy;
1410 	dma_dev->device_tx_status = nbpf_tx_status;
1411 	dma_dev->device_issue_pending = nbpf_issue_pending;
1412 
1413 	/*
1414 	 * If we drop support for unaligned MEMCPY buffer addresses and / or
1415 	 * lengths by setting
1416 	 * dma_dev->copy_align = 4;
1417 	 * then we can set transfer length to 4 bytes in nbpf_prep_one() for
1418 	 * DMA_MEM_TO_MEM
1419 	 */
1420 
1421 	/* Compulsory for DMA_SLAVE fields */
1422 	dma_dev->device_prep_slave_sg = nbpf_prep_slave_sg;
1423 	dma_dev->device_config = nbpf_config;
1424 	dma_dev->device_pause = nbpf_pause;
1425 	dma_dev->device_terminate_all = nbpf_terminate_all;
1426 
1427 	dma_dev->src_addr_widths = NBPF_DMA_BUSWIDTHS;
1428 	dma_dev->dst_addr_widths = NBPF_DMA_BUSWIDTHS;
1429 	dma_dev->directions = BIT(DMA_DEV_TO_MEM) | BIT(DMA_MEM_TO_DEV);
1430 
1431 	platform_set_drvdata(pdev, nbpf);
1432 
1433 	ret = clk_prepare_enable(nbpf->clk);
1434 	if (ret < 0)
1435 		return ret;
1436 
1437 	nbpf_configure(nbpf);
1438 
1439 	ret = dma_async_device_register(dma_dev);
1440 	if (ret < 0)
1441 		goto e_clk_off;
1442 
1443 	ret = of_dma_controller_register(np, nbpf_of_xlate, nbpf);
1444 	if (ret < 0)
1445 		goto e_dma_dev_unreg;
1446 
1447 	return 0;
1448 
1449 e_dma_dev_unreg:
1450 	dma_async_device_unregister(dma_dev);
1451 e_clk_off:
1452 	clk_disable_unprepare(nbpf->clk);
1453 
1454 	return ret;
1455 }
1456 
1457 static void nbpf_remove(struct platform_device *pdev)
1458 {
1459 	struct nbpf_device *nbpf = platform_get_drvdata(pdev);
1460 	int i;
1461 
1462 	devm_free_irq(&pdev->dev, nbpf->eirq, nbpf);
1463 
1464 	for (i = 0; i < nbpf->config->num_channels; i++) {
1465 		struct nbpf_channel *chan = nbpf->chan + i;
1466 
1467 		devm_free_irq(&pdev->dev, chan->irq, chan);
1468 
1469 		tasklet_kill(&chan->tasklet);
1470 	}
1471 
1472 	of_dma_controller_free(pdev->dev.of_node);
1473 	dma_async_device_unregister(&nbpf->dma_dev);
1474 	clk_disable_unprepare(nbpf->clk);
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 = dev_get_drvdata(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 = dev_get_drvdata(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_new = 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