xref: /openbmc/linux/drivers/dma/nbpfaxi.c (revision 0ad53fe3)
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  * @chan:	associated DMAC channel
148  * @sg:		list of hardware descriptors, represented by struct nbpf_link_desc
149  * @node:	member in channel descriptor lists
150  */
151 struct nbpf_desc {
152 	struct dma_async_tx_descriptor async_tx;
153 	bool user_wait;
154 	size_t length;
155 	struct nbpf_channel *chan;
156 	struct list_head sg;
157 	struct list_head node;
158 };
159 
160 /* Take a wild guess: allocate 4 segments per descriptor */
161 #define NBPF_SEGMENTS_PER_DESC 4
162 #define NBPF_DESCS_PER_PAGE ((PAGE_SIZE - sizeof(struct list_head)) /	\
163 	(sizeof(struct nbpf_desc) +					\
164 	 NBPF_SEGMENTS_PER_DESC *					\
165 	 (sizeof(struct nbpf_link_desc) + sizeof(struct nbpf_link_reg))))
166 #define NBPF_SEGMENTS_PER_PAGE (NBPF_SEGMENTS_PER_DESC * NBPF_DESCS_PER_PAGE)
167 
168 struct nbpf_desc_page {
169 	struct list_head node;
170 	struct nbpf_desc desc[NBPF_DESCS_PER_PAGE];
171 	struct nbpf_link_desc ldesc[NBPF_SEGMENTS_PER_PAGE];
172 	struct nbpf_link_reg hwdesc[NBPF_SEGMENTS_PER_PAGE];
173 };
174 
175 /**
176  * struct nbpf_channel - one DMAC channel
177  * @dma_chan:	standard dmaengine channel object
178  * @tasklet:	channel specific tasklet used for callbacks
179  * @base:	register address base
180  * @nbpf:	DMAC
181  * @name:	IRQ name
182  * @irq:	IRQ number
183  * @slave_src_addr:	source address for slave DMA
184  * @slave_src_width:	source slave data size in bytes
185  * @slave_src_burst:	maximum source slave burst size in bytes
186  * @slave_dst_addr:	destination address for slave DMA
187  * @slave_dst_width:	destination slave data size in bytes
188  * @slave_dst_burst:	maximum destination slave burst size in bytes
189  * @terminal:	DMA terminal, assigned to this channel
190  * @dmarq_cfg:	DMA request line configuration - high / low, edge / level for NBPF_CHAN_CFG
191  * @flags:	configuration flags from DT
192  * @lock:	protect descriptor lists
193  * @free_links:	list of free link descriptors
194  * @free:	list of free descriptors
195  * @queued:	list of queued descriptors
196  * @active:	list of descriptors, scheduled for processing
197  * @done:	list of completed descriptors, waiting post-processing
198  * @desc_page:	list of additionally allocated descriptor pages - if any
199  * @running:	linked descriptor of running transaction
200  * @paused:	are translations on this channel paused?
201  */
202 struct nbpf_channel {
203 	struct dma_chan dma_chan;
204 	struct tasklet_struct tasklet;
205 	void __iomem *base;
206 	struct nbpf_device *nbpf;
207 	char name[16];
208 	int irq;
209 	dma_addr_t slave_src_addr;
210 	size_t slave_src_width;
211 	size_t slave_src_burst;
212 	dma_addr_t slave_dst_addr;
213 	size_t slave_dst_width;
214 	size_t slave_dst_burst;
215 	unsigned int terminal;
216 	u32 dmarq_cfg;
217 	unsigned long flags;
218 	spinlock_t lock;
219 	struct list_head free_links;
220 	struct list_head free;
221 	struct list_head queued;
222 	struct list_head active;
223 	struct list_head done;
224 	struct list_head desc_page;
225 	struct nbpf_desc *running;
226 	bool paused;
227 };
228 
229 struct nbpf_device {
230 	struct dma_device dma_dev;
231 	void __iomem *base;
232 	u32 max_burst_mem_read;
233 	u32 max_burst_mem_write;
234 	struct clk *clk;
235 	const struct nbpf_config *config;
236 	unsigned int eirq;
237 	struct nbpf_channel chan[];
238 };
239 
240 enum nbpf_model {
241 	NBPF1B4,
242 	NBPF1B8,
243 	NBPF1B16,
244 	NBPF4B4,
245 	NBPF4B8,
246 	NBPF4B16,
247 	NBPF8B4,
248 	NBPF8B8,
249 	NBPF8B16,
250 };
251 
252 static struct nbpf_config nbpf_cfg[] = {
253 	[NBPF1B4] = {
254 		.num_channels = 1,
255 		.buffer_size = 4,
256 	},
257 	[NBPF1B8] = {
258 		.num_channels = 1,
259 		.buffer_size = 8,
260 	},
261 	[NBPF1B16] = {
262 		.num_channels = 1,
263 		.buffer_size = 16,
264 	},
265 	[NBPF4B4] = {
266 		.num_channels = 4,
267 		.buffer_size = 4,
268 	},
269 	[NBPF4B8] = {
270 		.num_channels = 4,
271 		.buffer_size = 8,
272 	},
273 	[NBPF4B16] = {
274 		.num_channels = 4,
275 		.buffer_size = 16,
276 	},
277 	[NBPF8B4] = {
278 		.num_channels = 8,
279 		.buffer_size = 4,
280 	},
281 	[NBPF8B8] = {
282 		.num_channels = 8,
283 		.buffer_size = 8,
284 	},
285 	[NBPF8B16] = {
286 		.num_channels = 8,
287 		.buffer_size = 16,
288 	},
289 };
290 
291 #define nbpf_to_chan(d) container_of(d, struct nbpf_channel, dma_chan)
292 
293 /*
294  * dmaengine drivers seem to have a lot in common and instead of sharing more
295  * code, they reimplement those common algorithms independently. In this driver
296  * we try to separate the hardware-specific part from the (largely) generic
297  * part. This improves code readability and makes it possible in the future to
298  * reuse the generic code in form of a helper library. That generic code should
299  * be suitable for various DMA controllers, using transfer descriptors in RAM
300  * and pushing one SG list at a time to the DMA controller.
301  */
302 
303 /*		Hardware-specific part		*/
304 
305 static inline u32 nbpf_chan_read(struct nbpf_channel *chan,
306 				 unsigned int offset)
307 {
308 	u32 data = ioread32(chan->base + offset);
309 	dev_dbg(chan->dma_chan.device->dev, "%s(0x%p + 0x%x) = 0x%x\n",
310 		__func__, chan->base, offset, data);
311 	return data;
312 }
313 
314 static inline void nbpf_chan_write(struct nbpf_channel *chan,
315 				   unsigned int offset, u32 data)
316 {
317 	iowrite32(data, chan->base + offset);
318 	dev_dbg(chan->dma_chan.device->dev, "%s(0x%p + 0x%x) = 0x%x\n",
319 		__func__, chan->base, offset, data);
320 }
321 
322 static inline u32 nbpf_read(struct nbpf_device *nbpf,
323 			    unsigned int offset)
324 {
325 	u32 data = ioread32(nbpf->base + offset);
326 	dev_dbg(nbpf->dma_dev.dev, "%s(0x%p + 0x%x) = 0x%x\n",
327 		__func__, nbpf->base, offset, data);
328 	return data;
329 }
330 
331 static inline void nbpf_write(struct nbpf_device *nbpf,
332 			      unsigned int offset, u32 data)
333 {
334 	iowrite32(data, nbpf->base + offset);
335 	dev_dbg(nbpf->dma_dev.dev, "%s(0x%p + 0x%x) = 0x%x\n",
336 		__func__, nbpf->base, offset, data);
337 }
338 
339 static void nbpf_chan_halt(struct nbpf_channel *chan)
340 {
341 	nbpf_chan_write(chan, NBPF_CHAN_CTRL, NBPF_CHAN_CTRL_CLREN);
342 }
343 
344 static bool nbpf_status_get(struct nbpf_channel *chan)
345 {
346 	u32 status = nbpf_read(chan->nbpf, NBPF_DSTAT_END);
347 
348 	return status & BIT(chan - chan->nbpf->chan);
349 }
350 
351 static void nbpf_status_ack(struct nbpf_channel *chan)
352 {
353 	nbpf_chan_write(chan, NBPF_CHAN_CTRL, NBPF_CHAN_CTRL_CLREND);
354 }
355 
356 static u32 nbpf_error_get(struct nbpf_device *nbpf)
357 {
358 	return nbpf_read(nbpf, NBPF_DSTAT_ER);
359 }
360 
361 static struct nbpf_channel *nbpf_error_get_channel(struct nbpf_device *nbpf, u32 error)
362 {
363 	return nbpf->chan + __ffs(error);
364 }
365 
366 static void nbpf_error_clear(struct nbpf_channel *chan)
367 {
368 	u32 status;
369 	int i;
370 
371 	/* Stop the channel, make sure DMA has been aborted */
372 	nbpf_chan_halt(chan);
373 
374 	for (i = 1000; i; i--) {
375 		status = nbpf_chan_read(chan, NBPF_CHAN_STAT);
376 		if (!(status & NBPF_CHAN_STAT_TACT))
377 			break;
378 		cpu_relax();
379 	}
380 
381 	if (!i)
382 		dev_err(chan->dma_chan.device->dev,
383 			"%s(): abort timeout, channel status 0x%x\n", __func__, status);
384 
385 	nbpf_chan_write(chan, NBPF_CHAN_CTRL, NBPF_CHAN_CTRL_SWRST);
386 }
387 
388 static int nbpf_start(struct nbpf_desc *desc)
389 {
390 	struct nbpf_channel *chan = desc->chan;
391 	struct nbpf_link_desc *ldesc = list_first_entry(&desc->sg, struct nbpf_link_desc, node);
392 
393 	nbpf_chan_write(chan, NBPF_CHAN_NXLA, (u32)ldesc->hwdesc_dma_addr);
394 	nbpf_chan_write(chan, NBPF_CHAN_CTRL, NBPF_CHAN_CTRL_SETEN | NBPF_CHAN_CTRL_CLRSUS);
395 	chan->paused = false;
396 
397 	/* Software trigger MEMCPY - only MEMCPY uses the block mode */
398 	if (ldesc->hwdesc->config & NBPF_CHAN_CFG_TM)
399 		nbpf_chan_write(chan, NBPF_CHAN_CTRL, NBPF_CHAN_CTRL_STG);
400 
401 	dev_dbg(chan->nbpf->dma_dev.dev, "%s(): next 0x%x, cur 0x%x\n", __func__,
402 		nbpf_chan_read(chan, NBPF_CHAN_NXLA), nbpf_chan_read(chan, NBPF_CHAN_CRLA));
403 
404 	return 0;
405 }
406 
407 static void nbpf_chan_prepare(struct nbpf_channel *chan)
408 {
409 	chan->dmarq_cfg = (chan->flags & NBPF_SLAVE_RQ_HIGH ? NBPF_CHAN_CFG_HIEN : 0) |
410 		(chan->flags & NBPF_SLAVE_RQ_LOW ? NBPF_CHAN_CFG_LOEN : 0) |
411 		(chan->flags & NBPF_SLAVE_RQ_LEVEL ?
412 		 NBPF_CHAN_CFG_LVL | (NBPF_CHAN_CFG_AM & 0x200) : 0) |
413 		chan->terminal;
414 }
415 
416 static void nbpf_chan_prepare_default(struct nbpf_channel *chan)
417 {
418 	/* Don't output DMAACK */
419 	chan->dmarq_cfg = NBPF_CHAN_CFG_AM & 0x400;
420 	chan->terminal = 0;
421 	chan->flags = 0;
422 }
423 
424 static void nbpf_chan_configure(struct nbpf_channel *chan)
425 {
426 	/*
427 	 * We assume, that only the link mode and DMA request line configuration
428 	 * have to be set in the configuration register manually. Dynamic
429 	 * per-transfer configuration will be loaded from transfer descriptors.
430 	 */
431 	nbpf_chan_write(chan, NBPF_CHAN_CFG, NBPF_CHAN_CFG_DMS | chan->dmarq_cfg);
432 }
433 
434 static u32 nbpf_xfer_ds(struct nbpf_device *nbpf, size_t size,
435 			enum dma_transfer_direction direction)
436 {
437 	int max_burst = nbpf->config->buffer_size * 8;
438 
439 	if (nbpf->max_burst_mem_read || nbpf->max_burst_mem_write) {
440 		switch (direction) {
441 		case DMA_MEM_TO_MEM:
442 			max_burst = min_not_zero(nbpf->max_burst_mem_read,
443 						 nbpf->max_burst_mem_write);
444 			break;
445 		case DMA_MEM_TO_DEV:
446 			if (nbpf->max_burst_mem_read)
447 				max_burst = nbpf->max_burst_mem_read;
448 			break;
449 		case DMA_DEV_TO_MEM:
450 			if (nbpf->max_burst_mem_write)
451 				max_burst = nbpf->max_burst_mem_write;
452 			break;
453 		case DMA_DEV_TO_DEV:
454 		default:
455 			break;
456 		}
457 	}
458 
459 	/* Maximum supported bursts depend on the buffer size */
460 	return min_t(int, __ffs(size), ilog2(max_burst));
461 }
462 
463 static size_t nbpf_xfer_size(struct nbpf_device *nbpf,
464 			     enum dma_slave_buswidth width, u32 burst)
465 {
466 	size_t size;
467 
468 	if (!burst)
469 		burst = 1;
470 
471 	switch (width) {
472 	case DMA_SLAVE_BUSWIDTH_8_BYTES:
473 		size = 8 * burst;
474 		break;
475 
476 	case DMA_SLAVE_BUSWIDTH_4_BYTES:
477 		size = 4 * burst;
478 		break;
479 
480 	case DMA_SLAVE_BUSWIDTH_2_BYTES:
481 		size = 2 * burst;
482 		break;
483 
484 	default:
485 		pr_warn("%s(): invalid bus width %u\n", __func__, width);
486 		fallthrough;
487 	case DMA_SLAVE_BUSWIDTH_1_BYTE:
488 		size = burst;
489 	}
490 
491 	return nbpf_xfer_ds(nbpf, size, DMA_TRANS_NONE);
492 }
493 
494 /*
495  * We need a way to recognise slaves, whose data is sent "raw" over the bus,
496  * i.e. it isn't known in advance how many bytes will be received. Therefore
497  * the slave driver has to provide a "large enough" buffer and either read the
498  * buffer, when it is full, or detect, that some data has arrived, then wait for
499  * a timeout, if no more data arrives - receive what's already there. We want to
500  * handle such slaves in a special way to allow an optimised mode for other
501  * users, for whom the amount of data is known in advance. So far there's no way
502  * to recognise such slaves. We use a data-width check to distinguish between
503  * the SD host and the PL011 UART.
504  */
505 
506 static int nbpf_prep_one(struct nbpf_link_desc *ldesc,
507 			 enum dma_transfer_direction direction,
508 			 dma_addr_t src, dma_addr_t dst, size_t size, bool last)
509 {
510 	struct nbpf_link_reg *hwdesc = ldesc->hwdesc;
511 	struct nbpf_desc *desc = ldesc->desc;
512 	struct nbpf_channel *chan = desc->chan;
513 	struct device *dev = chan->dma_chan.device->dev;
514 	size_t mem_xfer, slave_xfer;
515 	bool can_burst;
516 
517 	hwdesc->header = NBPF_HEADER_WBD | NBPF_HEADER_LV |
518 		(last ? NBPF_HEADER_LE : 0);
519 
520 	hwdesc->src_addr = src;
521 	hwdesc->dst_addr = dst;
522 	hwdesc->transaction_size = size;
523 
524 	/*
525 	 * set config: SAD, DAD, DDS, SDS, etc.
526 	 * Note on transfer sizes: the DMAC can perform unaligned DMA transfers,
527 	 * but it is important to have transaction size a multiple of both
528 	 * receiver and transmitter transfer sizes. It is also possible to use
529 	 * different RAM and device transfer sizes, and it does work well with
530 	 * some devices, e.g. with V08R07S01E SD host controllers, which can use
531 	 * 128 byte transfers. But this doesn't work with other devices,
532 	 * especially when the transaction size is unknown. This is the case,
533 	 * e.g. with serial drivers like amba-pl011.c. For reception it sets up
534 	 * the transaction size of 4K and if fewer bytes are received, it
535 	 * pauses DMA and reads out data received via DMA as well as those left
536 	 * in the Rx FIFO. For this to work with the RAM side using burst
537 	 * transfers we enable the SBE bit and terminate the transfer in our
538 	 * .device_pause handler.
539 	 */
540 	mem_xfer = nbpf_xfer_ds(chan->nbpf, size, direction);
541 
542 	switch (direction) {
543 	case DMA_DEV_TO_MEM:
544 		can_burst = chan->slave_src_width >= 3;
545 		slave_xfer = min(mem_xfer, can_burst ?
546 				 chan->slave_src_burst : chan->slave_src_width);
547 		/*
548 		 * Is the slave narrower than 64 bits, i.e. isn't using the full
549 		 * bus width and cannot use bursts?
550 		 */
551 		if (mem_xfer > chan->slave_src_burst && !can_burst)
552 			mem_xfer = chan->slave_src_burst;
553 		/* Device-to-RAM DMA is unreliable without REQD set */
554 		hwdesc->config = NBPF_CHAN_CFG_SAD | (NBPF_CHAN_CFG_DDS & (mem_xfer << 16)) |
555 			(NBPF_CHAN_CFG_SDS & (slave_xfer << 12)) | NBPF_CHAN_CFG_REQD |
556 			NBPF_CHAN_CFG_SBE;
557 		break;
558 
559 	case DMA_MEM_TO_DEV:
560 		slave_xfer = min(mem_xfer, chan->slave_dst_width >= 3 ?
561 				 chan->slave_dst_burst : chan->slave_dst_width);
562 		hwdesc->config = NBPF_CHAN_CFG_DAD | (NBPF_CHAN_CFG_SDS & (mem_xfer << 12)) |
563 			(NBPF_CHAN_CFG_DDS & (slave_xfer << 16)) | NBPF_CHAN_CFG_REQD;
564 		break;
565 
566 	case DMA_MEM_TO_MEM:
567 		hwdesc->config = NBPF_CHAN_CFG_TCM | NBPF_CHAN_CFG_TM |
568 			(NBPF_CHAN_CFG_SDS & (mem_xfer << 12)) |
569 			(NBPF_CHAN_CFG_DDS & (mem_xfer << 16));
570 		break;
571 
572 	default:
573 		return -EINVAL;
574 	}
575 
576 	hwdesc->config |= chan->dmarq_cfg | (last ? 0 : NBPF_CHAN_CFG_DEM) |
577 		NBPF_CHAN_CFG_DMS;
578 
579 	dev_dbg(dev, "%s(): desc @ %pad: hdr 0x%x, cfg 0x%x, %zu @ %pad -> %pad\n",
580 		__func__, &ldesc->hwdesc_dma_addr, hwdesc->header,
581 		hwdesc->config, size, &src, &dst);
582 
583 	dma_sync_single_for_device(dev, ldesc->hwdesc_dma_addr, sizeof(*hwdesc),
584 				   DMA_TO_DEVICE);
585 
586 	return 0;
587 }
588 
589 static size_t nbpf_bytes_left(struct nbpf_channel *chan)
590 {
591 	return nbpf_chan_read(chan, NBPF_CHAN_CUR_TR_BYTE);
592 }
593 
594 static void nbpf_configure(struct nbpf_device *nbpf)
595 {
596 	nbpf_write(nbpf, NBPF_CTRL, NBPF_CTRL_LVINT);
597 }
598 
599 /*		Generic part			*/
600 
601 /* DMA ENGINE functions */
602 static void nbpf_issue_pending(struct dma_chan *dchan)
603 {
604 	struct nbpf_channel *chan = nbpf_to_chan(dchan);
605 	unsigned long flags;
606 
607 	dev_dbg(dchan->device->dev, "Entry %s()\n", __func__);
608 
609 	spin_lock_irqsave(&chan->lock, flags);
610 	if (list_empty(&chan->queued))
611 		goto unlock;
612 
613 	list_splice_tail_init(&chan->queued, &chan->active);
614 
615 	if (!chan->running) {
616 		struct nbpf_desc *desc = list_first_entry(&chan->active,
617 						struct nbpf_desc, node);
618 		if (!nbpf_start(desc))
619 			chan->running = desc;
620 	}
621 
622 unlock:
623 	spin_unlock_irqrestore(&chan->lock, flags);
624 }
625 
626 static enum dma_status nbpf_tx_status(struct dma_chan *dchan,
627 		dma_cookie_t cookie, struct dma_tx_state *state)
628 {
629 	struct nbpf_channel *chan = nbpf_to_chan(dchan);
630 	enum dma_status status = dma_cookie_status(dchan, cookie, state);
631 
632 	if (state) {
633 		dma_cookie_t running;
634 		unsigned long flags;
635 
636 		spin_lock_irqsave(&chan->lock, flags);
637 		running = chan->running ? chan->running->async_tx.cookie : -EINVAL;
638 
639 		if (cookie == running) {
640 			state->residue = nbpf_bytes_left(chan);
641 			dev_dbg(dchan->device->dev, "%s(): residue %u\n", __func__,
642 				state->residue);
643 		} else if (status == DMA_IN_PROGRESS) {
644 			struct nbpf_desc *desc;
645 			bool found = false;
646 
647 			list_for_each_entry(desc, &chan->active, node)
648 				if (desc->async_tx.cookie == cookie) {
649 					found = true;
650 					break;
651 				}
652 
653 			if (!found)
654 				list_for_each_entry(desc, &chan->queued, node)
655 					if (desc->async_tx.cookie == cookie) {
656 						found = true;
657 						break;
658 
659 					}
660 
661 			state->residue = found ? desc->length : 0;
662 		}
663 
664 		spin_unlock_irqrestore(&chan->lock, flags);
665 	}
666 
667 	if (chan->paused)
668 		status = DMA_PAUSED;
669 
670 	return status;
671 }
672 
673 static dma_cookie_t nbpf_tx_submit(struct dma_async_tx_descriptor *tx)
674 {
675 	struct nbpf_desc *desc = container_of(tx, struct nbpf_desc, async_tx);
676 	struct nbpf_channel *chan = desc->chan;
677 	unsigned long flags;
678 	dma_cookie_t cookie;
679 
680 	spin_lock_irqsave(&chan->lock, flags);
681 	cookie = dma_cookie_assign(tx);
682 	list_add_tail(&desc->node, &chan->queued);
683 	spin_unlock_irqrestore(&chan->lock, flags);
684 
685 	dev_dbg(chan->dma_chan.device->dev, "Entry %s(%d)\n", __func__, cookie);
686 
687 	return cookie;
688 }
689 
690 static int nbpf_desc_page_alloc(struct nbpf_channel *chan)
691 {
692 	struct dma_chan *dchan = &chan->dma_chan;
693 	struct nbpf_desc_page *dpage = (void *)get_zeroed_page(GFP_KERNEL | GFP_DMA);
694 	struct nbpf_link_desc *ldesc;
695 	struct nbpf_link_reg *hwdesc;
696 	struct nbpf_desc *desc;
697 	LIST_HEAD(head);
698 	LIST_HEAD(lhead);
699 	int i;
700 	struct device *dev = dchan->device->dev;
701 
702 	if (!dpage)
703 		return -ENOMEM;
704 
705 	dev_dbg(dev, "%s(): alloc %lu descriptors, %lu segments, total alloc %zu\n",
706 		__func__, NBPF_DESCS_PER_PAGE, NBPF_SEGMENTS_PER_PAGE, sizeof(*dpage));
707 
708 	for (i = 0, ldesc = dpage->ldesc, hwdesc = dpage->hwdesc;
709 	     i < ARRAY_SIZE(dpage->ldesc);
710 	     i++, ldesc++, hwdesc++) {
711 		ldesc->hwdesc = hwdesc;
712 		list_add_tail(&ldesc->node, &lhead);
713 		ldesc->hwdesc_dma_addr = dma_map_single(dchan->device->dev,
714 					hwdesc, sizeof(*hwdesc), DMA_TO_DEVICE);
715 
716 		dev_dbg(dev, "%s(): mapped 0x%p to %pad\n", __func__,
717 			hwdesc, &ldesc->hwdesc_dma_addr);
718 	}
719 
720 	for (i = 0, desc = dpage->desc;
721 	     i < ARRAY_SIZE(dpage->desc);
722 	     i++, desc++) {
723 		dma_async_tx_descriptor_init(&desc->async_tx, dchan);
724 		desc->async_tx.tx_submit = nbpf_tx_submit;
725 		desc->chan = chan;
726 		INIT_LIST_HEAD(&desc->sg);
727 		list_add_tail(&desc->node, &head);
728 	}
729 
730 	/*
731 	 * This function cannot be called from interrupt context, so, no need to
732 	 * save flags
733 	 */
734 	spin_lock_irq(&chan->lock);
735 	list_splice_tail(&lhead, &chan->free_links);
736 	list_splice_tail(&head, &chan->free);
737 	list_add(&dpage->node, &chan->desc_page);
738 	spin_unlock_irq(&chan->lock);
739 
740 	return ARRAY_SIZE(dpage->desc);
741 }
742 
743 static void nbpf_desc_put(struct nbpf_desc *desc)
744 {
745 	struct nbpf_channel *chan = desc->chan;
746 	struct nbpf_link_desc *ldesc, *tmp;
747 	unsigned long flags;
748 
749 	spin_lock_irqsave(&chan->lock, flags);
750 	list_for_each_entry_safe(ldesc, tmp, &desc->sg, node)
751 		list_move(&ldesc->node, &chan->free_links);
752 
753 	list_add(&desc->node, &chan->free);
754 	spin_unlock_irqrestore(&chan->lock, flags);
755 }
756 
757 static void nbpf_scan_acked(struct nbpf_channel *chan)
758 {
759 	struct nbpf_desc *desc, *tmp;
760 	unsigned long flags;
761 	LIST_HEAD(head);
762 
763 	spin_lock_irqsave(&chan->lock, flags);
764 	list_for_each_entry_safe(desc, tmp, &chan->done, node)
765 		if (async_tx_test_ack(&desc->async_tx) && desc->user_wait) {
766 			list_move(&desc->node, &head);
767 			desc->user_wait = false;
768 		}
769 	spin_unlock_irqrestore(&chan->lock, flags);
770 
771 	list_for_each_entry_safe(desc, tmp, &head, node) {
772 		list_del(&desc->node);
773 		nbpf_desc_put(desc);
774 	}
775 }
776 
777 /*
778  * We have to allocate descriptors with the channel lock dropped. This means,
779  * before we re-acquire the lock buffers can be taken already, so we have to
780  * re-check after re-acquiring the lock and possibly retry, if buffers are gone
781  * again.
782  */
783 static struct nbpf_desc *nbpf_desc_get(struct nbpf_channel *chan, size_t len)
784 {
785 	struct nbpf_desc *desc = NULL;
786 	struct nbpf_link_desc *ldesc, *prev = NULL;
787 
788 	nbpf_scan_acked(chan);
789 
790 	spin_lock_irq(&chan->lock);
791 
792 	do {
793 		int i = 0, ret;
794 
795 		if (list_empty(&chan->free)) {
796 			/* No more free descriptors */
797 			spin_unlock_irq(&chan->lock);
798 			ret = nbpf_desc_page_alloc(chan);
799 			if (ret < 0)
800 				return NULL;
801 			spin_lock_irq(&chan->lock);
802 			continue;
803 		}
804 		desc = list_first_entry(&chan->free, struct nbpf_desc, node);
805 		list_del(&desc->node);
806 
807 		do {
808 			if (list_empty(&chan->free_links)) {
809 				/* No more free link descriptors */
810 				spin_unlock_irq(&chan->lock);
811 				ret = nbpf_desc_page_alloc(chan);
812 				if (ret < 0) {
813 					nbpf_desc_put(desc);
814 					return NULL;
815 				}
816 				spin_lock_irq(&chan->lock);
817 				continue;
818 			}
819 
820 			ldesc = list_first_entry(&chan->free_links,
821 						 struct nbpf_link_desc, node);
822 			ldesc->desc = desc;
823 			if (prev)
824 				prev->hwdesc->next = (u32)ldesc->hwdesc_dma_addr;
825 
826 			prev = ldesc;
827 			list_move_tail(&ldesc->node, &desc->sg);
828 
829 			i++;
830 		} while (i < len);
831 	} while (!desc);
832 
833 	prev->hwdesc->next = 0;
834 
835 	spin_unlock_irq(&chan->lock);
836 
837 	return desc;
838 }
839 
840 static void nbpf_chan_idle(struct nbpf_channel *chan)
841 {
842 	struct nbpf_desc *desc, *tmp;
843 	unsigned long flags;
844 	LIST_HEAD(head);
845 
846 	spin_lock_irqsave(&chan->lock, flags);
847 
848 	list_splice_init(&chan->done, &head);
849 	list_splice_init(&chan->active, &head);
850 	list_splice_init(&chan->queued, &head);
851 
852 	chan->running = NULL;
853 
854 	spin_unlock_irqrestore(&chan->lock, flags);
855 
856 	list_for_each_entry_safe(desc, tmp, &head, node) {
857 		dev_dbg(chan->nbpf->dma_dev.dev, "%s(): force-free desc %p cookie %d\n",
858 			__func__, desc, desc->async_tx.cookie);
859 		list_del(&desc->node);
860 		nbpf_desc_put(desc);
861 	}
862 }
863 
864 static int nbpf_pause(struct dma_chan *dchan)
865 {
866 	struct nbpf_channel *chan = nbpf_to_chan(dchan);
867 
868 	dev_dbg(dchan->device->dev, "Entry %s\n", __func__);
869 
870 	chan->paused = true;
871 	nbpf_chan_write(chan, NBPF_CHAN_CTRL, NBPF_CHAN_CTRL_SETSUS);
872 	/* See comment in nbpf_prep_one() */
873 	nbpf_chan_write(chan, NBPF_CHAN_CTRL, NBPF_CHAN_CTRL_CLREN);
874 
875 	return 0;
876 }
877 
878 static int nbpf_terminate_all(struct dma_chan *dchan)
879 {
880 	struct nbpf_channel *chan = nbpf_to_chan(dchan);
881 
882 	dev_dbg(dchan->device->dev, "Entry %s\n", __func__);
883 	dev_dbg(dchan->device->dev, "Terminating\n");
884 
885 	nbpf_chan_halt(chan);
886 	nbpf_chan_idle(chan);
887 
888 	return 0;
889 }
890 
891 static int nbpf_config(struct dma_chan *dchan,
892 		       struct dma_slave_config *config)
893 {
894 	struct nbpf_channel *chan = nbpf_to_chan(dchan);
895 
896 	dev_dbg(dchan->device->dev, "Entry %s\n", __func__);
897 
898 	/*
899 	 * We could check config->slave_id to match chan->terminal here,
900 	 * but with DT they would be coming from the same source, so
901 	 * such a check would be superflous
902 	 */
903 
904 	chan->slave_dst_addr = config->dst_addr;
905 	chan->slave_dst_width = nbpf_xfer_size(chan->nbpf,
906 					       config->dst_addr_width, 1);
907 	chan->slave_dst_burst = nbpf_xfer_size(chan->nbpf,
908 					       config->dst_addr_width,
909 					       config->dst_maxburst);
910 	chan->slave_src_addr = config->src_addr;
911 	chan->slave_src_width = nbpf_xfer_size(chan->nbpf,
912 					       config->src_addr_width, 1);
913 	chan->slave_src_burst = nbpf_xfer_size(chan->nbpf,
914 					       config->src_addr_width,
915 					       config->src_maxburst);
916 
917 	return 0;
918 }
919 
920 static struct dma_async_tx_descriptor *nbpf_prep_sg(struct nbpf_channel *chan,
921 		struct scatterlist *src_sg, struct scatterlist *dst_sg,
922 		size_t len, enum dma_transfer_direction direction,
923 		unsigned long flags)
924 {
925 	struct nbpf_link_desc *ldesc;
926 	struct scatterlist *mem_sg;
927 	struct nbpf_desc *desc;
928 	bool inc_src, inc_dst;
929 	size_t data_len = 0;
930 	int i = 0;
931 
932 	switch (direction) {
933 	case DMA_DEV_TO_MEM:
934 		mem_sg = dst_sg;
935 		inc_src = false;
936 		inc_dst = true;
937 		break;
938 
939 	case DMA_MEM_TO_DEV:
940 		mem_sg = src_sg;
941 		inc_src = true;
942 		inc_dst = false;
943 		break;
944 
945 	default:
946 	case DMA_MEM_TO_MEM:
947 		mem_sg = src_sg;
948 		inc_src = true;
949 		inc_dst = true;
950 	}
951 
952 	desc = nbpf_desc_get(chan, len);
953 	if (!desc)
954 		return NULL;
955 
956 	desc->async_tx.flags = flags;
957 	desc->async_tx.cookie = -EBUSY;
958 	desc->user_wait = false;
959 
960 	/*
961 	 * This is a private descriptor list, and we own the descriptor. No need
962 	 * to lock.
963 	 */
964 	list_for_each_entry(ldesc, &desc->sg, node) {
965 		int ret = nbpf_prep_one(ldesc, direction,
966 					sg_dma_address(src_sg),
967 					sg_dma_address(dst_sg),
968 					sg_dma_len(mem_sg),
969 					i == len - 1);
970 		if (ret < 0) {
971 			nbpf_desc_put(desc);
972 			return NULL;
973 		}
974 		data_len += sg_dma_len(mem_sg);
975 		if (inc_src)
976 			src_sg = sg_next(src_sg);
977 		if (inc_dst)
978 			dst_sg = sg_next(dst_sg);
979 		mem_sg = direction == DMA_DEV_TO_MEM ? dst_sg : src_sg;
980 		i++;
981 	}
982 
983 	desc->length = data_len;
984 
985 	/* The user has to return the descriptor to us ASAP via .tx_submit() */
986 	return &desc->async_tx;
987 }
988 
989 static struct dma_async_tx_descriptor *nbpf_prep_memcpy(
990 	struct dma_chan *dchan, dma_addr_t dst, dma_addr_t src,
991 	size_t len, unsigned long flags)
992 {
993 	struct nbpf_channel *chan = nbpf_to_chan(dchan);
994 	struct scatterlist dst_sg;
995 	struct scatterlist src_sg;
996 
997 	sg_init_table(&dst_sg, 1);
998 	sg_init_table(&src_sg, 1);
999 
1000 	sg_dma_address(&dst_sg) = dst;
1001 	sg_dma_address(&src_sg) = src;
1002 
1003 	sg_dma_len(&dst_sg) = len;
1004 	sg_dma_len(&src_sg) = len;
1005 
1006 	dev_dbg(dchan->device->dev, "%s(): %zu @ %pad -> %pad\n",
1007 		__func__, len, &src, &dst);
1008 
1009 	return nbpf_prep_sg(chan, &src_sg, &dst_sg, 1,
1010 			    DMA_MEM_TO_MEM, flags);
1011 }
1012 
1013 static struct dma_async_tx_descriptor *nbpf_prep_slave_sg(
1014 	struct dma_chan *dchan, struct scatterlist *sgl, unsigned int sg_len,
1015 	enum dma_transfer_direction direction, unsigned long flags, void *context)
1016 {
1017 	struct nbpf_channel *chan = nbpf_to_chan(dchan);
1018 	struct scatterlist slave_sg;
1019 
1020 	dev_dbg(dchan->device->dev, "Entry %s()\n", __func__);
1021 
1022 	sg_init_table(&slave_sg, 1);
1023 
1024 	switch (direction) {
1025 	case DMA_MEM_TO_DEV:
1026 		sg_dma_address(&slave_sg) = chan->slave_dst_addr;
1027 		return nbpf_prep_sg(chan, sgl, &slave_sg, sg_len,
1028 				    direction, flags);
1029 
1030 	case DMA_DEV_TO_MEM:
1031 		sg_dma_address(&slave_sg) = chan->slave_src_addr;
1032 		return nbpf_prep_sg(chan, &slave_sg, sgl, sg_len,
1033 				    direction, flags);
1034 
1035 	default:
1036 		return NULL;
1037 	}
1038 }
1039 
1040 static int nbpf_alloc_chan_resources(struct dma_chan *dchan)
1041 {
1042 	struct nbpf_channel *chan = nbpf_to_chan(dchan);
1043 	int ret;
1044 
1045 	INIT_LIST_HEAD(&chan->free);
1046 	INIT_LIST_HEAD(&chan->free_links);
1047 	INIT_LIST_HEAD(&chan->queued);
1048 	INIT_LIST_HEAD(&chan->active);
1049 	INIT_LIST_HEAD(&chan->done);
1050 
1051 	ret = nbpf_desc_page_alloc(chan);
1052 	if (ret < 0)
1053 		return ret;
1054 
1055 	dev_dbg(dchan->device->dev, "Entry %s(): terminal %u\n", __func__,
1056 		chan->terminal);
1057 
1058 	nbpf_chan_configure(chan);
1059 
1060 	return ret;
1061 }
1062 
1063 static void nbpf_free_chan_resources(struct dma_chan *dchan)
1064 {
1065 	struct nbpf_channel *chan = nbpf_to_chan(dchan);
1066 	struct nbpf_desc_page *dpage, *tmp;
1067 
1068 	dev_dbg(dchan->device->dev, "Entry %s()\n", __func__);
1069 
1070 	nbpf_chan_halt(chan);
1071 	nbpf_chan_idle(chan);
1072 	/* Clean up for if a channel is re-used for MEMCPY after slave DMA */
1073 	nbpf_chan_prepare_default(chan);
1074 
1075 	list_for_each_entry_safe(dpage, tmp, &chan->desc_page, node) {
1076 		struct nbpf_link_desc *ldesc;
1077 		int i;
1078 		list_del(&dpage->node);
1079 		for (i = 0, ldesc = dpage->ldesc;
1080 		     i < ARRAY_SIZE(dpage->ldesc);
1081 		     i++, ldesc++)
1082 			dma_unmap_single(dchan->device->dev, ldesc->hwdesc_dma_addr,
1083 					 sizeof(*ldesc->hwdesc), DMA_TO_DEVICE);
1084 		free_page((unsigned long)dpage);
1085 	}
1086 }
1087 
1088 static struct dma_chan *nbpf_of_xlate(struct of_phandle_args *dma_spec,
1089 				      struct of_dma *ofdma)
1090 {
1091 	struct nbpf_device *nbpf = ofdma->of_dma_data;
1092 	struct dma_chan *dchan;
1093 	struct nbpf_channel *chan;
1094 
1095 	if (dma_spec->args_count != 2)
1096 		return NULL;
1097 
1098 	dchan = dma_get_any_slave_channel(&nbpf->dma_dev);
1099 	if (!dchan)
1100 		return NULL;
1101 
1102 	dev_dbg(dchan->device->dev, "Entry %s(%pOFn)\n", __func__,
1103 		dma_spec->np);
1104 
1105 	chan = nbpf_to_chan(dchan);
1106 
1107 	chan->terminal = dma_spec->args[0];
1108 	chan->flags = dma_spec->args[1];
1109 
1110 	nbpf_chan_prepare(chan);
1111 	nbpf_chan_configure(chan);
1112 
1113 	return dchan;
1114 }
1115 
1116 static void nbpf_chan_tasklet(struct tasklet_struct *t)
1117 {
1118 	struct nbpf_channel *chan = from_tasklet(chan, t, tasklet);
1119 	struct nbpf_desc *desc, *tmp;
1120 	struct dmaengine_desc_callback cb;
1121 
1122 	while (!list_empty(&chan->done)) {
1123 		bool found = false, must_put, recycling = false;
1124 
1125 		spin_lock_irq(&chan->lock);
1126 
1127 		list_for_each_entry_safe(desc, tmp, &chan->done, node) {
1128 			if (!desc->user_wait) {
1129 				/* Newly completed descriptor, have to process */
1130 				found = true;
1131 				break;
1132 			} else if (async_tx_test_ack(&desc->async_tx)) {
1133 				/*
1134 				 * This descriptor was waiting for a user ACK,
1135 				 * it can be recycled now.
1136 				 */
1137 				list_del(&desc->node);
1138 				spin_unlock_irq(&chan->lock);
1139 				nbpf_desc_put(desc);
1140 				recycling = true;
1141 				break;
1142 			}
1143 		}
1144 
1145 		if (recycling)
1146 			continue;
1147 
1148 		if (!found) {
1149 			/* This can happen if TERMINATE_ALL has been called */
1150 			spin_unlock_irq(&chan->lock);
1151 			break;
1152 		}
1153 
1154 		dma_cookie_complete(&desc->async_tx);
1155 
1156 		/*
1157 		 * With released lock we cannot dereference desc, maybe it's
1158 		 * still on the "done" list
1159 		 */
1160 		if (async_tx_test_ack(&desc->async_tx)) {
1161 			list_del(&desc->node);
1162 			must_put = true;
1163 		} else {
1164 			desc->user_wait = true;
1165 			must_put = false;
1166 		}
1167 
1168 		dmaengine_desc_get_callback(&desc->async_tx, &cb);
1169 
1170 		/* ack and callback completed descriptor */
1171 		spin_unlock_irq(&chan->lock);
1172 
1173 		dmaengine_desc_callback_invoke(&cb, NULL);
1174 
1175 		if (must_put)
1176 			nbpf_desc_put(desc);
1177 	}
1178 }
1179 
1180 static irqreturn_t nbpf_chan_irq(int irq, void *dev)
1181 {
1182 	struct nbpf_channel *chan = dev;
1183 	bool done = nbpf_status_get(chan);
1184 	struct nbpf_desc *desc;
1185 	irqreturn_t ret;
1186 	bool bh = false;
1187 
1188 	if (!done)
1189 		return IRQ_NONE;
1190 
1191 	nbpf_status_ack(chan);
1192 
1193 	dev_dbg(&chan->dma_chan.dev->device, "%s()\n", __func__);
1194 
1195 	spin_lock(&chan->lock);
1196 	desc = chan->running;
1197 	if (WARN_ON(!desc)) {
1198 		ret = IRQ_NONE;
1199 		goto unlock;
1200 	} else {
1201 		ret = IRQ_HANDLED;
1202 		bh = true;
1203 	}
1204 
1205 	list_move_tail(&desc->node, &chan->done);
1206 	chan->running = NULL;
1207 
1208 	if (!list_empty(&chan->active)) {
1209 		desc = list_first_entry(&chan->active,
1210 					struct nbpf_desc, node);
1211 		if (!nbpf_start(desc))
1212 			chan->running = desc;
1213 	}
1214 
1215 unlock:
1216 	spin_unlock(&chan->lock);
1217 
1218 	if (bh)
1219 		tasklet_schedule(&chan->tasklet);
1220 
1221 	return ret;
1222 }
1223 
1224 static irqreturn_t nbpf_err_irq(int irq, void *dev)
1225 {
1226 	struct nbpf_device *nbpf = dev;
1227 	u32 error = nbpf_error_get(nbpf);
1228 
1229 	dev_warn(nbpf->dma_dev.dev, "DMA error IRQ %u\n", irq);
1230 
1231 	if (!error)
1232 		return IRQ_NONE;
1233 
1234 	do {
1235 		struct nbpf_channel *chan = nbpf_error_get_channel(nbpf, error);
1236 		/* On error: abort all queued transfers, no callback */
1237 		nbpf_error_clear(chan);
1238 		nbpf_chan_idle(chan);
1239 		error = nbpf_error_get(nbpf);
1240 	} while (error);
1241 
1242 	return IRQ_HANDLED;
1243 }
1244 
1245 static int nbpf_chan_probe(struct nbpf_device *nbpf, int n)
1246 {
1247 	struct dma_device *dma_dev = &nbpf->dma_dev;
1248 	struct nbpf_channel *chan = nbpf->chan + n;
1249 	int ret;
1250 
1251 	chan->nbpf = nbpf;
1252 	chan->base = nbpf->base + NBPF_REG_CHAN_OFFSET + NBPF_REG_CHAN_SIZE * n;
1253 	INIT_LIST_HEAD(&chan->desc_page);
1254 	spin_lock_init(&chan->lock);
1255 	chan->dma_chan.device = dma_dev;
1256 	dma_cookie_init(&chan->dma_chan);
1257 	nbpf_chan_prepare_default(chan);
1258 
1259 	dev_dbg(dma_dev->dev, "%s(): channel %d: -> %p\n", __func__, n, chan->base);
1260 
1261 	snprintf(chan->name, sizeof(chan->name), "nbpf %d", n);
1262 
1263 	tasklet_setup(&chan->tasklet, nbpf_chan_tasklet);
1264 	ret = devm_request_irq(dma_dev->dev, chan->irq,
1265 			nbpf_chan_irq, IRQF_SHARED,
1266 			chan->name, chan);
1267 	if (ret < 0)
1268 		return ret;
1269 
1270 	/* Add the channel to DMA device channel list */
1271 	list_add_tail(&chan->dma_chan.device_node,
1272 		      &dma_dev->channels);
1273 
1274 	return 0;
1275 }
1276 
1277 static const struct of_device_id nbpf_match[] = {
1278 	{.compatible = "renesas,nbpfaxi64dmac1b4",	.data = &nbpf_cfg[NBPF1B4]},
1279 	{.compatible = "renesas,nbpfaxi64dmac1b8",	.data = &nbpf_cfg[NBPF1B8]},
1280 	{.compatible = "renesas,nbpfaxi64dmac1b16",	.data = &nbpf_cfg[NBPF1B16]},
1281 	{.compatible = "renesas,nbpfaxi64dmac4b4",	.data = &nbpf_cfg[NBPF4B4]},
1282 	{.compatible = "renesas,nbpfaxi64dmac4b8",	.data = &nbpf_cfg[NBPF4B8]},
1283 	{.compatible = "renesas,nbpfaxi64dmac4b16",	.data = &nbpf_cfg[NBPF4B16]},
1284 	{.compatible = "renesas,nbpfaxi64dmac8b4",	.data = &nbpf_cfg[NBPF8B4]},
1285 	{.compatible = "renesas,nbpfaxi64dmac8b8",	.data = &nbpf_cfg[NBPF8B8]},
1286 	{.compatible = "renesas,nbpfaxi64dmac8b16",	.data = &nbpf_cfg[NBPF8B16]},
1287 	{}
1288 };
1289 MODULE_DEVICE_TABLE(of, nbpf_match);
1290 
1291 static int nbpf_probe(struct platform_device *pdev)
1292 {
1293 	struct device *dev = &pdev->dev;
1294 	struct device_node *np = dev->of_node;
1295 	struct nbpf_device *nbpf;
1296 	struct dma_device *dma_dev;
1297 	struct resource *iomem, *irq_res;
1298 	const struct nbpf_config *cfg;
1299 	int num_channels;
1300 	int ret, irq, eirq, i;
1301 	int irqbuf[9] /* maximum 8 channels + error IRQ */;
1302 	unsigned int irqs = 0;
1303 
1304 	BUILD_BUG_ON(sizeof(struct nbpf_desc_page) > PAGE_SIZE);
1305 
1306 	/* DT only */
1307 	if (!np)
1308 		return -ENODEV;
1309 
1310 	cfg = of_device_get_match_data(dev);
1311 	num_channels = cfg->num_channels;
1312 
1313 	nbpf = devm_kzalloc(dev, struct_size(nbpf, chan, num_channels),
1314 			    GFP_KERNEL);
1315 	if (!nbpf)
1316 		return -ENOMEM;
1317 
1318 	dma_dev = &nbpf->dma_dev;
1319 	dma_dev->dev = dev;
1320 
1321 	iomem = platform_get_resource(pdev, IORESOURCE_MEM, 0);
1322 	nbpf->base = devm_ioremap_resource(dev, iomem);
1323 	if (IS_ERR(nbpf->base))
1324 		return PTR_ERR(nbpf->base);
1325 
1326 	nbpf->clk = devm_clk_get(dev, NULL);
1327 	if (IS_ERR(nbpf->clk))
1328 		return PTR_ERR(nbpf->clk);
1329 
1330 	of_property_read_u32(np, "max-burst-mem-read",
1331 			     &nbpf->max_burst_mem_read);
1332 	of_property_read_u32(np, "max-burst-mem-write",
1333 			     &nbpf->max_burst_mem_write);
1334 
1335 	nbpf->config = cfg;
1336 
1337 	for (i = 0; irqs < ARRAY_SIZE(irqbuf); i++) {
1338 		irq_res = platform_get_resource(pdev, IORESOURCE_IRQ, i);
1339 		if (!irq_res)
1340 			break;
1341 
1342 		for (irq = irq_res->start; irq <= irq_res->end;
1343 		     irq++, irqs++)
1344 			irqbuf[irqs] = irq;
1345 	}
1346 
1347 	/*
1348 	 * 3 IRQ resource schemes are supported:
1349 	 * 1. 1 shared IRQ for error and all channels
1350 	 * 2. 2 IRQs: one for error and one shared for all channels
1351 	 * 3. 1 IRQ for error and an own IRQ for each channel
1352 	 */
1353 	if (irqs != 1 && irqs != 2 && irqs != num_channels + 1)
1354 		return -ENXIO;
1355 
1356 	if (irqs == 1) {
1357 		eirq = irqbuf[0];
1358 
1359 		for (i = 0; i <= num_channels; i++)
1360 			nbpf->chan[i].irq = irqbuf[0];
1361 	} else {
1362 		eirq = platform_get_irq_byname(pdev, "error");
1363 		if (eirq < 0)
1364 			return eirq;
1365 
1366 		if (irqs == num_channels + 1) {
1367 			struct nbpf_channel *chan;
1368 
1369 			for (i = 0, chan = nbpf->chan; i <= num_channels;
1370 			     i++, chan++) {
1371 				/* Skip the error IRQ */
1372 				if (irqbuf[i] == eirq)
1373 					i++;
1374 				chan->irq = irqbuf[i];
1375 			}
1376 
1377 			if (chan != nbpf->chan + num_channels)
1378 				return -EINVAL;
1379 		} else {
1380 			/* 2 IRQs and more than one channel */
1381 			if (irqbuf[0] == eirq)
1382 				irq = irqbuf[1];
1383 			else
1384 				irq = irqbuf[0];
1385 
1386 			for (i = 0; i <= num_channels; i++)
1387 				nbpf->chan[i].irq = irq;
1388 		}
1389 	}
1390 
1391 	ret = devm_request_irq(dev, eirq, nbpf_err_irq,
1392 			       IRQF_SHARED, "dma error", nbpf);
1393 	if (ret < 0)
1394 		return ret;
1395 	nbpf->eirq = eirq;
1396 
1397 	INIT_LIST_HEAD(&dma_dev->channels);
1398 
1399 	/* Create DMA Channel */
1400 	for (i = 0; i < num_channels; i++) {
1401 		ret = nbpf_chan_probe(nbpf, i);
1402 		if (ret < 0)
1403 			return ret;
1404 	}
1405 
1406 	dma_cap_set(DMA_MEMCPY, dma_dev->cap_mask);
1407 	dma_cap_set(DMA_SLAVE, dma_dev->cap_mask);
1408 	dma_cap_set(DMA_PRIVATE, dma_dev->cap_mask);
1409 
1410 	/* Common and MEMCPY operations */
1411 	dma_dev->device_alloc_chan_resources
1412 		= nbpf_alloc_chan_resources;
1413 	dma_dev->device_free_chan_resources = nbpf_free_chan_resources;
1414 	dma_dev->device_prep_dma_memcpy = nbpf_prep_memcpy;
1415 	dma_dev->device_tx_status = nbpf_tx_status;
1416 	dma_dev->device_issue_pending = nbpf_issue_pending;
1417 
1418 	/*
1419 	 * If we drop support for unaligned MEMCPY buffer addresses and / or
1420 	 * lengths by setting
1421 	 * dma_dev->copy_align = 4;
1422 	 * then we can set transfer length to 4 bytes in nbpf_prep_one() for
1423 	 * DMA_MEM_TO_MEM
1424 	 */
1425 
1426 	/* Compulsory for DMA_SLAVE fields */
1427 	dma_dev->device_prep_slave_sg = nbpf_prep_slave_sg;
1428 	dma_dev->device_config = nbpf_config;
1429 	dma_dev->device_pause = nbpf_pause;
1430 	dma_dev->device_terminate_all = nbpf_terminate_all;
1431 
1432 	dma_dev->src_addr_widths = NBPF_DMA_BUSWIDTHS;
1433 	dma_dev->dst_addr_widths = NBPF_DMA_BUSWIDTHS;
1434 	dma_dev->directions = BIT(DMA_DEV_TO_MEM) | BIT(DMA_MEM_TO_DEV);
1435 
1436 	platform_set_drvdata(pdev, nbpf);
1437 
1438 	ret = clk_prepare_enable(nbpf->clk);
1439 	if (ret < 0)
1440 		return ret;
1441 
1442 	nbpf_configure(nbpf);
1443 
1444 	ret = dma_async_device_register(dma_dev);
1445 	if (ret < 0)
1446 		goto e_clk_off;
1447 
1448 	ret = of_dma_controller_register(np, nbpf_of_xlate, nbpf);
1449 	if (ret < 0)
1450 		goto e_dma_dev_unreg;
1451 
1452 	return 0;
1453 
1454 e_dma_dev_unreg:
1455 	dma_async_device_unregister(dma_dev);
1456 e_clk_off:
1457 	clk_disable_unprepare(nbpf->clk);
1458 
1459 	return ret;
1460 }
1461 
1462 static int nbpf_remove(struct platform_device *pdev)
1463 {
1464 	struct nbpf_device *nbpf = platform_get_drvdata(pdev);
1465 	int i;
1466 
1467 	devm_free_irq(&pdev->dev, nbpf->eirq, nbpf);
1468 
1469 	for (i = 0; i < nbpf->config->num_channels; i++) {
1470 		struct nbpf_channel *chan = nbpf->chan + i;
1471 
1472 		devm_free_irq(&pdev->dev, chan->irq, chan);
1473 
1474 		tasklet_kill(&chan->tasklet);
1475 	}
1476 
1477 	of_dma_controller_free(pdev->dev.of_node);
1478 	dma_async_device_unregister(&nbpf->dma_dev);
1479 	clk_disable_unprepare(nbpf->clk);
1480 
1481 	return 0;
1482 }
1483 
1484 static const struct platform_device_id nbpf_ids[] = {
1485 	{"nbpfaxi64dmac1b4",	(kernel_ulong_t)&nbpf_cfg[NBPF1B4]},
1486 	{"nbpfaxi64dmac1b8",	(kernel_ulong_t)&nbpf_cfg[NBPF1B8]},
1487 	{"nbpfaxi64dmac1b16",	(kernel_ulong_t)&nbpf_cfg[NBPF1B16]},
1488 	{"nbpfaxi64dmac4b4",	(kernel_ulong_t)&nbpf_cfg[NBPF4B4]},
1489 	{"nbpfaxi64dmac4b8",	(kernel_ulong_t)&nbpf_cfg[NBPF4B8]},
1490 	{"nbpfaxi64dmac4b16",	(kernel_ulong_t)&nbpf_cfg[NBPF4B16]},
1491 	{"nbpfaxi64dmac8b4",	(kernel_ulong_t)&nbpf_cfg[NBPF8B4]},
1492 	{"nbpfaxi64dmac8b8",	(kernel_ulong_t)&nbpf_cfg[NBPF8B8]},
1493 	{"nbpfaxi64dmac8b16",	(kernel_ulong_t)&nbpf_cfg[NBPF8B16]},
1494 	{},
1495 };
1496 MODULE_DEVICE_TABLE(platform, nbpf_ids);
1497 
1498 #ifdef CONFIG_PM
1499 static int nbpf_runtime_suspend(struct device *dev)
1500 {
1501 	struct nbpf_device *nbpf = dev_get_drvdata(dev);
1502 	clk_disable_unprepare(nbpf->clk);
1503 	return 0;
1504 }
1505 
1506 static int nbpf_runtime_resume(struct device *dev)
1507 {
1508 	struct nbpf_device *nbpf = dev_get_drvdata(dev);
1509 	return clk_prepare_enable(nbpf->clk);
1510 }
1511 #endif
1512 
1513 static const struct dev_pm_ops nbpf_pm_ops = {
1514 	SET_RUNTIME_PM_OPS(nbpf_runtime_suspend, nbpf_runtime_resume, NULL)
1515 };
1516 
1517 static struct platform_driver nbpf_driver = {
1518 	.driver = {
1519 		.name = "dma-nbpf",
1520 		.of_match_table = nbpf_match,
1521 		.pm = &nbpf_pm_ops,
1522 	},
1523 	.id_table = nbpf_ids,
1524 	.probe = nbpf_probe,
1525 	.remove = nbpf_remove,
1526 };
1527 
1528 module_platform_driver(nbpf_driver);
1529 
1530 MODULE_AUTHOR("Guennadi Liakhovetski <g.liakhovetski@gmx.de>");
1531 MODULE_DESCRIPTION("dmaengine driver for NBPFAXI64* DMACs");
1532 MODULE_LICENSE("GPL v2");
1533