2 * Copyright (C) 2013-2014 Renesas Electronics Europe Ltd.
3 * Author: Guennadi Liakhovetski <g.liakhovetski@gmx.de>
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of version 2 of the GNU General Public License as
7 * published by the Free Software Foundation.
10 #include <linux/bitmap.h>
11 #include <linux/bitops.h>
12 #include <linux/clk.h>
13 #include <linux/dma-mapping.h>
14 #include <linux/dmaengine.h>
15 #include <linux/err.h>
16 #include <linux/interrupt.h>
18 #include <linux/log2.h>
19 #include <linux/module.h>
21 #include <linux/of_device.h>
22 #include <linux/of_dma.h>
23 #include <linux/platform_device.h>
24 #include <linux/slab.h>
26 #include <dt-bindings/dma/nbpfaxi.h>
28 #include "dmaengine.h"
30 #define NBPF_REG_CHAN_OFFSET 0
31 #define NBPF_REG_CHAN_SIZE 0x40
33 /* Channel Current Transaction Byte register */
34 #define NBPF_CHAN_CUR_TR_BYTE 0x20
36 /* Channel Status register */
37 #define NBPF_CHAN_STAT 0x24
38 #define NBPF_CHAN_STAT_EN 1
39 #define NBPF_CHAN_STAT_TACT 4
40 #define NBPF_CHAN_STAT_ERR 0x10
41 #define NBPF_CHAN_STAT_END 0x20
42 #define NBPF_CHAN_STAT_TC 0x40
43 #define NBPF_CHAN_STAT_DER 0x400
45 /* Channel Control register */
46 #define NBPF_CHAN_CTRL 0x28
47 #define NBPF_CHAN_CTRL_SETEN 1
48 #define NBPF_CHAN_CTRL_CLREN 2
49 #define NBPF_CHAN_CTRL_STG 4
50 #define NBPF_CHAN_CTRL_SWRST 8
51 #define NBPF_CHAN_CTRL_CLRRQ 0x10
52 #define NBPF_CHAN_CTRL_CLREND 0x20
53 #define NBPF_CHAN_CTRL_CLRTC 0x40
54 #define NBPF_CHAN_CTRL_SETSUS 0x100
55 #define NBPF_CHAN_CTRL_CLRSUS 0x200
57 /* Channel Configuration register */
58 #define NBPF_CHAN_CFG 0x2c
59 #define NBPF_CHAN_CFG_SEL 7 /* terminal SELect: 0..7 */
60 #define NBPF_CHAN_CFG_REQD 8 /* REQuest Direction: DMAREQ is 0: input, 1: output */
61 #define NBPF_CHAN_CFG_LOEN 0x10 /* LOw ENable: low DMA request line is: 0: inactive, 1: active */
62 #define NBPF_CHAN_CFG_HIEN 0x20 /* HIgh ENable: high DMA request line is: 0: inactive, 1: active */
63 #define NBPF_CHAN_CFG_LVL 0x40 /* LeVeL: DMA request line is sensed as 0: edge, 1: level */
64 #define NBPF_CHAN_CFG_AM 0x700 /* ACK Mode: 0: Pulse mode, 1: Level mode, b'1x: Bus Cycle */
65 #define NBPF_CHAN_CFG_SDS 0xf000 /* Source Data Size: 0: 8 bits,... , 7: 1024 bits */
66 #define NBPF_CHAN_CFG_DDS 0xf0000 /* Destination Data Size: as above */
67 #define NBPF_CHAN_CFG_SAD 0x100000 /* Source ADdress counting: 0: increment, 1: fixed */
68 #define NBPF_CHAN_CFG_DAD 0x200000 /* Destination ADdress counting: 0: increment, 1: fixed */
69 #define NBPF_CHAN_CFG_TM 0x400000 /* Transfer Mode: 0: single, 1: block TM */
70 #define NBPF_CHAN_CFG_DEM 0x1000000 /* DMAEND interrupt Mask */
71 #define NBPF_CHAN_CFG_TCM 0x2000000 /* DMATCO interrupt Mask */
72 #define NBPF_CHAN_CFG_SBE 0x8000000 /* Sweep Buffer Enable */
73 #define NBPF_CHAN_CFG_RSEL 0x10000000 /* RM: Register Set sELect */
74 #define NBPF_CHAN_CFG_RSW 0x20000000 /* RM: Register Select sWitch */
75 #define NBPF_CHAN_CFG_REN 0x40000000 /* RM: Register Set Enable */
76 #define NBPF_CHAN_CFG_DMS 0x80000000 /* 0: register mode (RM), 1: link mode (LM) */
78 #define NBPF_CHAN_NXLA 0x38
79 #define NBPF_CHAN_CRLA 0x3c
81 /* Link Header field */
82 #define NBPF_HEADER_LV 1
83 #define NBPF_HEADER_LE 2
84 #define NBPF_HEADER_WBD 4
85 #define NBPF_HEADER_DIM 8
87 #define NBPF_CTRL 0x300
88 #define NBPF_CTRL_PR 1 /* 0: fixed priority, 1: round robin */
89 #define NBPF_CTRL_LVINT 2 /* DMAEND and DMAERR signalling: 0: pulse, 1: level */
91 #define NBPF_DSTAT_ER 0x314
92 #define NBPF_DSTAT_END 0x318
94 #define NBPF_DMA_BUSWIDTHS \
95 (BIT(DMA_SLAVE_BUSWIDTH_UNDEFINED) | \
96 BIT(DMA_SLAVE_BUSWIDTH_1_BYTE) | \
97 BIT(DMA_SLAVE_BUSWIDTH_2_BYTES) | \
98 BIT(DMA_SLAVE_BUSWIDTH_4_BYTES) | \
99 BIT(DMA_SLAVE_BUSWIDTH_8_BYTES))
107 * We've got 3 types of objects, used to describe DMA transfers:
108 * 1. high-level descriptor, containing a struct dma_async_tx_descriptor object
109 * in it, used to communicate with the user
110 * 2. hardware DMA link descriptors, that we pass to DMAC for DMA transfer
111 * queuing, these must be DMAable, using either the streaming DMA API or
112 * allocated from coherent memory - one per SG segment
113 * 3. one per SG segment descriptors, used to manage HW link descriptors from
114 * (2). They do not have to be DMAable. They can either be (a) allocated
115 * together with link descriptors as mixed (DMA / CPU) objects, or (b)
116 * separately. Even if allocated separately it would be best to link them
117 * to link descriptors once during channel resource allocation and always
118 * use them as a single object.
119 * Therefore for both cases (a) and (b) at run-time objects (2) and (3) shall be
120 * treated as a single SG segment descriptor.
123 struct nbpf_link_reg {
127 u32 transaction_size;
138 struct nbpf_link_desc {
139 struct nbpf_link_reg *hwdesc;
140 dma_addr_t hwdesc_dma_addr;
141 struct nbpf_desc *desc;
142 struct list_head node;
146 * struct nbpf_desc - DMA transfer descriptor
147 * @async_tx: dmaengine object
148 * @user_wait: waiting for a user ack
149 * @length: total transfer length
150 * @sg: list of hardware descriptors, represented by struct nbpf_link_desc
151 * @node: member in channel descriptor lists
154 struct dma_async_tx_descriptor async_tx;
157 struct nbpf_channel *chan;
159 struct list_head node;
162 /* Take a wild guess: allocate 4 segments per descriptor */
163 #define NBPF_SEGMENTS_PER_DESC 4
164 #define NBPF_DESCS_PER_PAGE ((PAGE_SIZE - sizeof(struct list_head)) / \
165 (sizeof(struct nbpf_desc) + \
166 NBPF_SEGMENTS_PER_DESC * \
167 (sizeof(struct nbpf_link_desc) + sizeof(struct nbpf_link_reg))))
168 #define NBPF_SEGMENTS_PER_PAGE (NBPF_SEGMENTS_PER_DESC * NBPF_DESCS_PER_PAGE)
170 struct nbpf_desc_page {
171 struct list_head node;
172 struct nbpf_desc desc[NBPF_DESCS_PER_PAGE];
173 struct nbpf_link_desc ldesc[NBPF_SEGMENTS_PER_PAGE];
174 struct nbpf_link_reg hwdesc[NBPF_SEGMENTS_PER_PAGE];
178 * struct nbpf_channel - one DMAC channel
179 * @dma_chan: standard dmaengine channel object
180 * @base: register address base
184 * @slave_addr: address for slave DMA
185 * @slave_width:slave data size in bytes
186 * @slave_burst:maximum slave burst size in bytes
187 * @terminal: DMA terminal, assigned to this channel
188 * @dmarq_cfg: DMA request line configuration - high / low, edge / level for NBPF_CHAN_CFG
189 * @flags: configuration flags from DT
190 * @lock: protect descriptor lists
191 * @free_links: list of free link descriptors
192 * @free: list of free descriptors
193 * @queued: list of queued descriptors
194 * @active: list of descriptors, scheduled for processing
195 * @done: list of completed descriptors, waiting post-processing
196 * @desc_page: list of additionally allocated descriptor pages - if any
198 struct nbpf_channel {
199 struct dma_chan dma_chan;
201 struct nbpf_device *nbpf;
204 dma_addr_t slave_src_addr;
205 size_t slave_src_width;
206 size_t slave_src_burst;
207 dma_addr_t slave_dst_addr;
208 size_t slave_dst_width;
209 size_t slave_dst_burst;
210 unsigned int terminal;
214 struct list_head free_links;
215 struct list_head free;
216 struct list_head queued;
217 struct list_head active;
218 struct list_head done;
219 struct list_head desc_page;
220 struct nbpf_desc *running;
225 struct dma_device dma_dev;
228 const struct nbpf_config *config;
229 struct nbpf_channel chan[];
244 static struct nbpf_config nbpf_cfg[] = {
283 #define nbpf_to_chan(d) container_of(d, struct nbpf_channel, dma_chan)
286 * dmaengine drivers seem to have a lot in common and instead of sharing more
287 * code, they reimplement those common algorithms independently. In this driver
288 * we try to separate the hardware-specific part from the (largely) generic
289 * part. This improves code readability and makes it possible in the future to
290 * reuse the generic code in form of a helper library. That generic code should
291 * be suitable for various DMA controllers, using transfer descriptors in RAM
292 * and pushing one SG list at a time to the DMA controller.
295 /* Hardware-specific part */
297 static inline u32 nbpf_chan_read(struct nbpf_channel *chan,
300 u32 data = ioread32(chan->base + offset);
301 dev_dbg(chan->dma_chan.device->dev, "%s(0x%p + 0x%x) = 0x%x\n",
302 __func__, chan->base, offset, data);
306 static inline void nbpf_chan_write(struct nbpf_channel *chan,
307 unsigned int offset, u32 data)
309 iowrite32(data, chan->base + offset);
310 dev_dbg(chan->dma_chan.device->dev, "%s(0x%p + 0x%x) = 0x%x\n",
311 __func__, chan->base, offset, data);
314 static inline u32 nbpf_read(struct nbpf_device *nbpf,
317 u32 data = ioread32(nbpf->base + offset);
318 dev_dbg(nbpf->dma_dev.dev, "%s(0x%p + 0x%x) = 0x%x\n",
319 __func__, nbpf->base, offset, data);
323 static inline void nbpf_write(struct nbpf_device *nbpf,
324 unsigned int offset, u32 data)
326 iowrite32(data, nbpf->base + offset);
327 dev_dbg(nbpf->dma_dev.dev, "%s(0x%p + 0x%x) = 0x%x\n",
328 __func__, nbpf->base, offset, data);
331 static void nbpf_chan_halt(struct nbpf_channel *chan)
333 nbpf_chan_write(chan, NBPF_CHAN_CTRL, NBPF_CHAN_CTRL_CLREN);
336 static bool nbpf_status_get(struct nbpf_channel *chan)
338 u32 status = nbpf_read(chan->nbpf, NBPF_DSTAT_END);
340 return status & BIT(chan - chan->nbpf->chan);
343 static void nbpf_status_ack(struct nbpf_channel *chan)
345 nbpf_chan_write(chan, NBPF_CHAN_CTRL, NBPF_CHAN_CTRL_CLREND);
348 static u32 nbpf_error_get(struct nbpf_device *nbpf)
350 return nbpf_read(nbpf, NBPF_DSTAT_ER);
353 struct nbpf_channel *nbpf_error_get_channel(struct nbpf_device *nbpf, u32 error)
355 return nbpf->chan + __ffs(error);
358 static void nbpf_error_clear(struct nbpf_channel *chan)
363 /* Stop the channel, make sure DMA has been aborted */
364 nbpf_chan_halt(chan);
366 for (i = 1000; i; i--) {
367 status = nbpf_chan_read(chan, NBPF_CHAN_STAT);
368 if (!(status & NBPF_CHAN_STAT_TACT))
374 dev_err(chan->dma_chan.device->dev,
375 "%s(): abort timeout, channel status 0x%x\n", __func__, status);
377 nbpf_chan_write(chan, NBPF_CHAN_CTRL, NBPF_CHAN_CTRL_SWRST);
380 static int nbpf_start(struct nbpf_desc *desc)
382 struct nbpf_channel *chan = desc->chan;
383 struct nbpf_link_desc *ldesc = list_first_entry(&desc->sg, struct nbpf_link_desc, node);
385 nbpf_chan_write(chan, NBPF_CHAN_NXLA, (u32)ldesc->hwdesc_dma_addr);
386 nbpf_chan_write(chan, NBPF_CHAN_CTRL, NBPF_CHAN_CTRL_SETEN | NBPF_CHAN_CTRL_CLRSUS);
387 chan->paused = false;
389 /* Software trigger MEMCPY - only MEMCPY uses the block mode */
390 if (ldesc->hwdesc->config & NBPF_CHAN_CFG_TM)
391 nbpf_chan_write(chan, NBPF_CHAN_CTRL, NBPF_CHAN_CTRL_STG);
393 dev_dbg(chan->nbpf->dma_dev.dev, "%s(): next 0x%x, cur 0x%x\n", __func__,
394 nbpf_chan_read(chan, NBPF_CHAN_NXLA), nbpf_chan_read(chan, NBPF_CHAN_CRLA));
399 static void nbpf_chan_prepare(struct nbpf_channel *chan)
401 chan->dmarq_cfg = (chan->flags & NBPF_SLAVE_RQ_HIGH ? NBPF_CHAN_CFG_HIEN : 0) |
402 (chan->flags & NBPF_SLAVE_RQ_LOW ? NBPF_CHAN_CFG_LOEN : 0) |
403 (chan->flags & NBPF_SLAVE_RQ_LEVEL ?
404 NBPF_CHAN_CFG_LVL | (NBPF_CHAN_CFG_AM & 0x200) : 0) |
408 static void nbpf_chan_prepare_default(struct nbpf_channel *chan)
410 /* Don't output DMAACK */
411 chan->dmarq_cfg = NBPF_CHAN_CFG_AM & 0x400;
416 static void nbpf_chan_configure(struct nbpf_channel *chan)
419 * We assume, that only the link mode and DMA request line configuration
420 * have to be set in the configuration register manually. Dynamic
421 * per-transfer configuration will be loaded from transfer descriptors.
423 nbpf_chan_write(chan, NBPF_CHAN_CFG, NBPF_CHAN_CFG_DMS | chan->dmarq_cfg);
426 static u32 nbpf_xfer_ds(struct nbpf_device *nbpf, size_t size)
428 /* Maximum supported bursts depend on the buffer size */
429 return min_t(int, __ffs(size), ilog2(nbpf->config->buffer_size * 8));
432 static size_t nbpf_xfer_size(struct nbpf_device *nbpf,
433 enum dma_slave_buswidth width, u32 burst)
441 case DMA_SLAVE_BUSWIDTH_8_BYTES:
445 case DMA_SLAVE_BUSWIDTH_4_BYTES:
449 case DMA_SLAVE_BUSWIDTH_2_BYTES:
454 pr_warn("%s(): invalid bus width %u\n", __func__, width);
455 case DMA_SLAVE_BUSWIDTH_1_BYTE:
459 return nbpf_xfer_ds(nbpf, size);
463 * We need a way to recognise slaves, whose data is sent "raw" over the bus,
464 * i.e. it isn't known in advance how many bytes will be received. Therefore
465 * the slave driver has to provide a "large enough" buffer and either read the
466 * buffer, when it is full, or detect, that some data has arrived, then wait for
467 * a timeout, if no more data arrives - receive what's already there. We want to
468 * handle such slaves in a special way to allow an optimised mode for other
469 * users, for whom the amount of data is known in advance. So far there's no way
470 * to recognise such slaves. We use a data-width check to distinguish between
471 * the SD host and the PL011 UART.
474 static int nbpf_prep_one(struct nbpf_link_desc *ldesc,
475 enum dma_transfer_direction direction,
476 dma_addr_t src, dma_addr_t dst, size_t size, bool last)
478 struct nbpf_link_reg *hwdesc = ldesc->hwdesc;
479 struct nbpf_desc *desc = ldesc->desc;
480 struct nbpf_channel *chan = desc->chan;
481 struct device *dev = chan->dma_chan.device->dev;
482 size_t mem_xfer, slave_xfer;
485 hwdesc->header = NBPF_HEADER_WBD | NBPF_HEADER_LV |
486 (last ? NBPF_HEADER_LE : 0);
488 hwdesc->src_addr = src;
489 hwdesc->dst_addr = dst;
490 hwdesc->transaction_size = size;
493 * set config: SAD, DAD, DDS, SDS, etc.
494 * Note on transfer sizes: the DMAC can perform unaligned DMA transfers,
495 * but it is important to have transaction size a multiple of both
496 * receiver and transmitter transfer sizes. It is also possible to use
497 * different RAM and device transfer sizes, and it does work well with
498 * some devices, e.g. with V08R07S01E SD host controllers, which can use
499 * 128 byte transfers. But this doesn't work with other devices,
500 * especially when the transaction size is unknown. This is the case,
501 * e.g. with serial drivers like amba-pl011.c. For reception it sets up
502 * the transaction size of 4K and if fewer bytes are received, it
503 * pauses DMA and reads out data received via DMA as well as those left
504 * in the Rx FIFO. For this to work with the RAM side using burst
505 * transfers we enable the SBE bit and terminate the transfer in our
508 mem_xfer = nbpf_xfer_ds(chan->nbpf, size);
512 can_burst = chan->slave_src_width >= 3;
513 slave_xfer = min(mem_xfer, can_burst ?
514 chan->slave_src_burst : chan->slave_src_width);
516 * Is the slave narrower than 64 bits, i.e. isn't using the full
517 * bus width and cannot use bursts?
519 if (mem_xfer > chan->slave_src_burst && !can_burst)
520 mem_xfer = chan->slave_src_burst;
521 /* Device-to-RAM DMA is unreliable without REQD set */
522 hwdesc->config = NBPF_CHAN_CFG_SAD | (NBPF_CHAN_CFG_DDS & (mem_xfer << 16)) |
523 (NBPF_CHAN_CFG_SDS & (slave_xfer << 12)) | NBPF_CHAN_CFG_REQD |
528 slave_xfer = min(mem_xfer, chan->slave_dst_width >= 3 ?
529 chan->slave_dst_burst : chan->slave_dst_width);
530 hwdesc->config = NBPF_CHAN_CFG_DAD | (NBPF_CHAN_CFG_SDS & (mem_xfer << 12)) |
531 (NBPF_CHAN_CFG_DDS & (slave_xfer << 16)) | NBPF_CHAN_CFG_REQD;
535 hwdesc->config = NBPF_CHAN_CFG_TCM | NBPF_CHAN_CFG_TM |
536 (NBPF_CHAN_CFG_SDS & (mem_xfer << 12)) |
537 (NBPF_CHAN_CFG_DDS & (mem_xfer << 16));
544 hwdesc->config |= chan->dmarq_cfg | (last ? 0 : NBPF_CHAN_CFG_DEM) |
547 dev_dbg(dev, "%s(): desc @ %pad: hdr 0x%x, cfg 0x%x, %zu @ %pad -> %pad\n",
548 __func__, &ldesc->hwdesc_dma_addr, hwdesc->header,
549 hwdesc->config, size, &src, &dst);
551 dma_sync_single_for_device(dev, ldesc->hwdesc_dma_addr, sizeof(*hwdesc),
557 static size_t nbpf_bytes_left(struct nbpf_channel *chan)
559 return nbpf_chan_read(chan, NBPF_CHAN_CUR_TR_BYTE);
562 static void nbpf_configure(struct nbpf_device *nbpf)
564 nbpf_write(nbpf, NBPF_CTRL, NBPF_CTRL_LVINT);
567 static void nbpf_pause(struct nbpf_channel *chan)
569 nbpf_chan_write(chan, NBPF_CHAN_CTRL, NBPF_CHAN_CTRL_SETSUS);
570 /* See comment in nbpf_prep_one() */
571 nbpf_chan_write(chan, NBPF_CHAN_CTRL, NBPF_CHAN_CTRL_CLREN);
576 /* DMA ENGINE functions */
577 static void nbpf_issue_pending(struct dma_chan *dchan)
579 struct nbpf_channel *chan = nbpf_to_chan(dchan);
582 dev_dbg(dchan->device->dev, "Entry %s()\n", __func__);
584 spin_lock_irqsave(&chan->lock, flags);
585 if (list_empty(&chan->queued))
588 list_splice_tail_init(&chan->queued, &chan->active);
590 if (!chan->running) {
591 struct nbpf_desc *desc = list_first_entry(&chan->active,
592 struct nbpf_desc, node);
593 if (!nbpf_start(desc))
594 chan->running = desc;
598 spin_unlock_irqrestore(&chan->lock, flags);
601 static enum dma_status nbpf_tx_status(struct dma_chan *dchan,
602 dma_cookie_t cookie, struct dma_tx_state *state)
604 struct nbpf_channel *chan = nbpf_to_chan(dchan);
605 enum dma_status status = dma_cookie_status(dchan, cookie, state);
608 dma_cookie_t running;
611 spin_lock_irqsave(&chan->lock, flags);
612 running = chan->running ? chan->running->async_tx.cookie : -EINVAL;
614 if (cookie == running) {
615 state->residue = nbpf_bytes_left(chan);
616 dev_dbg(dchan->device->dev, "%s(): residue %u\n", __func__,
618 } else if (status == DMA_IN_PROGRESS) {
619 struct nbpf_desc *desc;
622 list_for_each_entry(desc, &chan->active, node)
623 if (desc->async_tx.cookie == cookie) {
629 list_for_each_entry(desc, &chan->queued, node)
630 if (desc->async_tx.cookie == cookie) {
636 state->residue = found ? desc->length : 0;
639 spin_unlock_irqrestore(&chan->lock, flags);
648 static dma_cookie_t nbpf_tx_submit(struct dma_async_tx_descriptor *tx)
650 struct nbpf_desc *desc = container_of(tx, struct nbpf_desc, async_tx);
651 struct nbpf_channel *chan = desc->chan;
655 spin_lock_irqsave(&chan->lock, flags);
656 cookie = dma_cookie_assign(tx);
657 list_add_tail(&desc->node, &chan->queued);
658 spin_unlock_irqrestore(&chan->lock, flags);
660 dev_dbg(chan->dma_chan.device->dev, "Entry %s(%d)\n", __func__, cookie);
665 static int nbpf_desc_page_alloc(struct nbpf_channel *chan)
667 struct dma_chan *dchan = &chan->dma_chan;
668 struct nbpf_desc_page *dpage = (void *)get_zeroed_page(GFP_KERNEL | GFP_DMA);
669 struct nbpf_link_desc *ldesc;
670 struct nbpf_link_reg *hwdesc;
671 struct nbpf_desc *desc;
675 struct device *dev = dchan->device->dev;
680 dev_dbg(dev, "%s(): alloc %lu descriptors, %lu segments, total alloc %zu\n",
681 __func__, NBPF_DESCS_PER_PAGE, NBPF_SEGMENTS_PER_PAGE, sizeof(*dpage));
683 for (i = 0, ldesc = dpage->ldesc, hwdesc = dpage->hwdesc;
684 i < ARRAY_SIZE(dpage->ldesc);
685 i++, ldesc++, hwdesc++) {
686 ldesc->hwdesc = hwdesc;
687 list_add_tail(&ldesc->node, &lhead);
688 ldesc->hwdesc_dma_addr = dma_map_single(dchan->device->dev,
689 hwdesc, sizeof(*hwdesc), DMA_TO_DEVICE);
691 dev_dbg(dev, "%s(): mapped 0x%p to %pad\n", __func__,
692 hwdesc, &ldesc->hwdesc_dma_addr);
695 for (i = 0, desc = dpage->desc;
696 i < ARRAY_SIZE(dpage->desc);
698 dma_async_tx_descriptor_init(&desc->async_tx, dchan);
699 desc->async_tx.tx_submit = nbpf_tx_submit;
701 INIT_LIST_HEAD(&desc->sg);
702 list_add_tail(&desc->node, &head);
706 * This function cannot be called from interrupt context, so, no need to
709 spin_lock_irq(&chan->lock);
710 list_splice_tail(&lhead, &chan->free_links);
711 list_splice_tail(&head, &chan->free);
712 list_add(&dpage->node, &chan->desc_page);
713 spin_unlock_irq(&chan->lock);
715 return ARRAY_SIZE(dpage->desc);
718 static void nbpf_desc_put(struct nbpf_desc *desc)
720 struct nbpf_channel *chan = desc->chan;
721 struct nbpf_link_desc *ldesc, *tmp;
724 spin_lock_irqsave(&chan->lock, flags);
725 list_for_each_entry_safe(ldesc, tmp, &desc->sg, node)
726 list_move(&ldesc->node, &chan->free_links);
728 list_add(&desc->node, &chan->free);
729 spin_unlock_irqrestore(&chan->lock, flags);
732 static void nbpf_scan_acked(struct nbpf_channel *chan)
734 struct nbpf_desc *desc, *tmp;
738 spin_lock_irqsave(&chan->lock, flags);
739 list_for_each_entry_safe(desc, tmp, &chan->done, node)
740 if (async_tx_test_ack(&desc->async_tx) && desc->user_wait) {
741 list_move(&desc->node, &head);
742 desc->user_wait = false;
744 spin_unlock_irqrestore(&chan->lock, flags);
746 list_for_each_entry_safe(desc, tmp, &head, node) {
747 list_del(&desc->node);
753 * We have to allocate descriptors with the channel lock dropped. This means,
754 * before we re-acquire the lock buffers can be taken already, so we have to
755 * re-check after re-acquiring the lock and possibly retry, if buffers are gone
758 static struct nbpf_desc *nbpf_desc_get(struct nbpf_channel *chan, size_t len)
760 struct nbpf_desc *desc = NULL;
761 struct nbpf_link_desc *ldesc, *prev = NULL;
763 nbpf_scan_acked(chan);
765 spin_lock_irq(&chan->lock);
770 if (list_empty(&chan->free)) {
771 /* No more free descriptors */
772 spin_unlock_irq(&chan->lock);
773 ret = nbpf_desc_page_alloc(chan);
776 spin_lock_irq(&chan->lock);
779 desc = list_first_entry(&chan->free, struct nbpf_desc, node);
780 list_del(&desc->node);
783 if (list_empty(&chan->free_links)) {
784 /* No more free link descriptors */
785 spin_unlock_irq(&chan->lock);
786 ret = nbpf_desc_page_alloc(chan);
791 spin_lock_irq(&chan->lock);
795 ldesc = list_first_entry(&chan->free_links,
796 struct nbpf_link_desc, node);
799 prev->hwdesc->next = (u32)ldesc->hwdesc_dma_addr;
802 list_move_tail(&ldesc->node, &desc->sg);
808 prev->hwdesc->next = 0;
810 spin_unlock_irq(&chan->lock);
815 static void nbpf_chan_idle(struct nbpf_channel *chan)
817 struct nbpf_desc *desc, *tmp;
821 spin_lock_irqsave(&chan->lock, flags);
823 list_splice_init(&chan->done, &head);
824 list_splice_init(&chan->active, &head);
825 list_splice_init(&chan->queued, &head);
827 chan->running = NULL;
829 spin_unlock_irqrestore(&chan->lock, flags);
831 list_for_each_entry_safe(desc, tmp, &head, node) {
832 dev_dbg(chan->nbpf->dma_dev.dev, "%s(): force-free desc %p cookie %d\n",
833 __func__, desc, desc->async_tx.cookie);
834 list_del(&desc->node);
839 static int nbpf_control(struct dma_chan *dchan, enum dma_ctrl_cmd cmd,
842 struct nbpf_channel *chan = nbpf_to_chan(dchan);
843 struct dma_slave_config *config;
845 dev_dbg(dchan->device->dev, "Entry %s(%d)\n", __func__, cmd);
848 case DMA_TERMINATE_ALL:
849 dev_dbg(dchan->device->dev, "Terminating\n");
850 nbpf_chan_halt(chan);
851 nbpf_chan_idle(chan);
854 case DMA_SLAVE_CONFIG:
857 config = (struct dma_slave_config *)arg;
860 * We could check config->slave_id to match chan->terminal here,
861 * but with DT they would be coming from the same source, so
862 * such a check would be superflous
865 chan->slave_dst_addr = config->dst_addr;
866 chan->slave_dst_width = nbpf_xfer_size(chan->nbpf,
867 config->dst_addr_width, 1);
868 chan->slave_dst_burst = nbpf_xfer_size(chan->nbpf,
869 config->dst_addr_width,
870 config->dst_maxburst);
871 chan->slave_src_addr = config->src_addr;
872 chan->slave_src_width = nbpf_xfer_size(chan->nbpf,
873 config->src_addr_width, 1);
874 chan->slave_src_burst = nbpf_xfer_size(chan->nbpf,
875 config->src_addr_width,
876 config->src_maxburst);
891 static struct dma_async_tx_descriptor *nbpf_prep_sg(struct nbpf_channel *chan,
892 struct scatterlist *src_sg, struct scatterlist *dst_sg,
893 size_t len, enum dma_transfer_direction direction,
896 struct nbpf_link_desc *ldesc;
897 struct scatterlist *mem_sg;
898 struct nbpf_desc *desc;
899 bool inc_src, inc_dst;
923 desc = nbpf_desc_get(chan, len);
927 desc->async_tx.flags = flags;
928 desc->async_tx.cookie = -EBUSY;
929 desc->user_wait = false;
932 * This is a private descriptor list, and we own the descriptor. No need
935 list_for_each_entry(ldesc, &desc->sg, node) {
936 int ret = nbpf_prep_one(ldesc, direction,
937 sg_dma_address(src_sg),
938 sg_dma_address(dst_sg),
945 data_len += sg_dma_len(mem_sg);
947 src_sg = sg_next(src_sg);
949 dst_sg = sg_next(dst_sg);
950 mem_sg = direction == DMA_DEV_TO_MEM ? dst_sg : src_sg;
954 desc->length = data_len;
956 /* The user has to return the descriptor to us ASAP via .tx_submit() */
957 return &desc->async_tx;
960 static struct dma_async_tx_descriptor *nbpf_prep_memcpy(
961 struct dma_chan *dchan, dma_addr_t dst, dma_addr_t src,
962 size_t len, unsigned long flags)
964 struct nbpf_channel *chan = nbpf_to_chan(dchan);
965 struct scatterlist dst_sg;
966 struct scatterlist src_sg;
968 sg_init_table(&dst_sg, 1);
969 sg_init_table(&src_sg, 1);
971 sg_dma_address(&dst_sg) = dst;
972 sg_dma_address(&src_sg) = src;
974 sg_dma_len(&dst_sg) = len;
975 sg_dma_len(&src_sg) = len;
977 dev_dbg(dchan->device->dev, "%s(): %zu @ %pad -> %pad\n",
978 __func__, len, &src, &dst);
980 return nbpf_prep_sg(chan, &src_sg, &dst_sg, 1,
981 DMA_MEM_TO_MEM, flags);
984 static struct dma_async_tx_descriptor *nbpf_prep_memcpy_sg(
985 struct dma_chan *dchan,
986 struct scatterlist *dst_sg, unsigned int dst_nents,
987 struct scatterlist *src_sg, unsigned int src_nents,
990 struct nbpf_channel *chan = nbpf_to_chan(dchan);
992 if (dst_nents != src_nents)
995 return nbpf_prep_sg(chan, src_sg, dst_sg, src_nents,
996 DMA_MEM_TO_MEM, flags);
999 static struct dma_async_tx_descriptor *nbpf_prep_slave_sg(
1000 struct dma_chan *dchan, struct scatterlist *sgl, unsigned int sg_len,
1001 enum dma_transfer_direction direction, unsigned long flags, void *context)
1003 struct nbpf_channel *chan = nbpf_to_chan(dchan);
1004 struct scatterlist slave_sg;
1006 dev_dbg(dchan->device->dev, "Entry %s()\n", __func__);
1008 sg_init_table(&slave_sg, 1);
1010 switch (direction) {
1011 case DMA_MEM_TO_DEV:
1012 sg_dma_address(&slave_sg) = chan->slave_dst_addr;
1013 return nbpf_prep_sg(chan, sgl, &slave_sg, sg_len,
1016 case DMA_DEV_TO_MEM:
1017 sg_dma_address(&slave_sg) = chan->slave_src_addr;
1018 return nbpf_prep_sg(chan, &slave_sg, sgl, sg_len,
1026 static int nbpf_alloc_chan_resources(struct dma_chan *dchan)
1028 struct nbpf_channel *chan = nbpf_to_chan(dchan);
1031 INIT_LIST_HEAD(&chan->free);
1032 INIT_LIST_HEAD(&chan->free_links);
1033 INIT_LIST_HEAD(&chan->queued);
1034 INIT_LIST_HEAD(&chan->active);
1035 INIT_LIST_HEAD(&chan->done);
1037 ret = nbpf_desc_page_alloc(chan);
1041 dev_dbg(dchan->device->dev, "Entry %s(): terminal %u\n", __func__,
1044 nbpf_chan_configure(chan);
1049 static void nbpf_free_chan_resources(struct dma_chan *dchan)
1051 struct nbpf_channel *chan = nbpf_to_chan(dchan);
1052 struct nbpf_desc_page *dpage, *tmp;
1054 dev_dbg(dchan->device->dev, "Entry %s()\n", __func__);
1056 nbpf_chan_halt(chan);
1057 nbpf_chan_idle(chan);
1058 /* Clean up for if a channel is re-used for MEMCPY after slave DMA */
1059 nbpf_chan_prepare_default(chan);
1061 list_for_each_entry_safe(dpage, tmp, &chan->desc_page, node) {
1062 struct nbpf_link_desc *ldesc;
1064 list_del(&dpage->node);
1065 for (i = 0, ldesc = dpage->ldesc;
1066 i < ARRAY_SIZE(dpage->ldesc);
1068 dma_unmap_single(dchan->device->dev, ldesc->hwdesc_dma_addr,
1069 sizeof(*ldesc->hwdesc), DMA_TO_DEVICE);
1070 free_page((unsigned long)dpage);
1074 static int nbpf_slave_caps(struct dma_chan *dchan,
1075 struct dma_slave_caps *caps)
1077 caps->src_addr_widths = NBPF_DMA_BUSWIDTHS;
1078 caps->dstn_addr_widths = NBPF_DMA_BUSWIDTHS;
1079 caps->directions = BIT(DMA_DEV_TO_MEM) | BIT(DMA_MEM_TO_DEV);
1080 caps->cmd_pause = false;
1081 caps->cmd_terminate = true;
1086 static struct dma_chan *nbpf_of_xlate(struct of_phandle_args *dma_spec,
1087 struct of_dma *ofdma)
1089 struct nbpf_device *nbpf = ofdma->of_dma_data;
1090 struct dma_chan *dchan;
1091 struct nbpf_channel *chan;
1093 if (dma_spec->args_count != 2)
1096 dchan = dma_get_any_slave_channel(&nbpf->dma_dev);
1100 dev_dbg(dchan->device->dev, "Entry %s(%s)\n", __func__,
1101 dma_spec->np->name);
1103 chan = nbpf_to_chan(dchan);
1105 chan->terminal = dma_spec->args[0];
1106 chan->flags = dma_spec->args[1];
1108 nbpf_chan_prepare(chan);
1109 nbpf_chan_configure(chan);
1114 static irqreturn_t nbpf_chan_irqt(int irq, void *dev)
1116 struct nbpf_channel *chan = dev;
1117 struct nbpf_desc *desc, *tmp;
1118 dma_async_tx_callback callback;
1121 while (!list_empty(&chan->done)) {
1122 bool found = false, must_put, recycling = false;
1124 spin_lock_irq(&chan->lock);
1126 list_for_each_entry_safe(desc, tmp, &chan->done, node) {
1127 if (!desc->user_wait) {
1128 /* Newly completed descriptor, have to process */
1131 } else if (async_tx_test_ack(&desc->async_tx)) {
1133 * This descriptor was waiting for a user ACK,
1134 * it can be recycled now.
1136 list_del(&desc->node);
1137 spin_unlock_irq(&chan->lock);
1138 nbpf_desc_put(desc);
1148 /* This can happen if TERMINATE_ALL has been called */
1149 spin_unlock_irq(&chan->lock);
1153 dma_cookie_complete(&desc->async_tx);
1156 * With released lock we cannot dereference desc, maybe it's
1157 * still on the "done" list
1159 if (async_tx_test_ack(&desc->async_tx)) {
1160 list_del(&desc->node);
1163 desc->user_wait = true;
1167 callback = desc->async_tx.callback;
1168 param = desc->async_tx.callback_param;
1170 /* ack and callback completed descriptor */
1171 spin_unlock_irq(&chan->lock);
1177 nbpf_desc_put(desc);
1183 static irqreturn_t nbpf_chan_irq(int irq, void *dev)
1185 struct nbpf_channel *chan = dev;
1186 bool done = nbpf_status_get(chan);
1187 struct nbpf_desc *desc;
1193 nbpf_status_ack(chan);
1195 dev_dbg(&chan->dma_chan.dev->device, "%s()\n", __func__);
1197 spin_lock(&chan->lock);
1198 desc = chan->running;
1199 if (WARN_ON(!desc)) {
1203 ret = IRQ_WAKE_THREAD;
1206 list_move_tail(&desc->node, &chan->done);
1207 chan->running = NULL;
1209 if (!list_empty(&chan->active)) {
1210 desc = list_first_entry(&chan->active,
1211 struct nbpf_desc, node);
1212 if (!nbpf_start(desc))
1213 chan->running = desc;
1217 spin_unlock(&chan->lock);
1222 static irqreturn_t nbpf_err_irq(int irq, void *dev)
1224 struct nbpf_device *nbpf = dev;
1225 u32 error = nbpf_error_get(nbpf);
1227 dev_warn(nbpf->dma_dev.dev, "DMA error IRQ %u\n", irq);
1233 struct nbpf_channel *chan = nbpf_error_get_channel(nbpf, error);
1234 /* On error: abort all queued transfers, no callback */
1235 nbpf_error_clear(chan);
1236 nbpf_chan_idle(chan);
1237 error = nbpf_error_get(nbpf);
1243 static int nbpf_chan_probe(struct nbpf_device *nbpf, int n)
1245 struct dma_device *dma_dev = &nbpf->dma_dev;
1246 struct nbpf_channel *chan = nbpf->chan + n;
1250 chan->base = nbpf->base + NBPF_REG_CHAN_OFFSET + NBPF_REG_CHAN_SIZE * n;
1251 INIT_LIST_HEAD(&chan->desc_page);
1252 spin_lock_init(&chan->lock);
1253 chan->dma_chan.device = dma_dev;
1254 dma_cookie_init(&chan->dma_chan);
1255 nbpf_chan_prepare_default(chan);
1257 dev_dbg(dma_dev->dev, "%s(): channel %d: -> %p\n", __func__, n, chan->base);
1259 snprintf(chan->name, sizeof(chan->name), "nbpf %d", n);
1261 ret = devm_request_threaded_irq(dma_dev->dev, chan->irq,
1262 nbpf_chan_irq, nbpf_chan_irqt, IRQF_SHARED,
1267 /* Add the channel to DMA device channel list */
1268 list_add_tail(&chan->dma_chan.device_node,
1269 &dma_dev->channels);
1274 static const struct of_device_id nbpf_match[] = {
1275 {.compatible = "renesas,nbpfaxi64dmac1b4", .data = &nbpf_cfg[NBPF1B4]},
1276 {.compatible = "renesas,nbpfaxi64dmac1b8", .data = &nbpf_cfg[NBPF1B8]},
1277 {.compatible = "renesas,nbpfaxi64dmac1b16", .data = &nbpf_cfg[NBPF1B16]},
1278 {.compatible = "renesas,nbpfaxi64dmac4b4", .data = &nbpf_cfg[NBPF4B4]},
1279 {.compatible = "renesas,nbpfaxi64dmac4b8", .data = &nbpf_cfg[NBPF4B8]},
1280 {.compatible = "renesas,nbpfaxi64dmac4b16", .data = &nbpf_cfg[NBPF4B16]},
1281 {.compatible = "renesas,nbpfaxi64dmac8b4", .data = &nbpf_cfg[NBPF8B4]},
1282 {.compatible = "renesas,nbpfaxi64dmac8b8", .data = &nbpf_cfg[NBPF8B8]},
1283 {.compatible = "renesas,nbpfaxi64dmac8b16", .data = &nbpf_cfg[NBPF8B16]},
1286 MODULE_DEVICE_TABLE(of, nbpf_match);
1288 static int nbpf_probe(struct platform_device *pdev)
1290 struct device *dev = &pdev->dev;
1291 const struct of_device_id *of_id = of_match_device(nbpf_match, dev);
1292 struct device_node *np = dev->of_node;
1293 struct nbpf_device *nbpf;
1294 struct dma_device *dma_dev;
1295 struct resource *iomem, *irq_res;
1296 const struct nbpf_config *cfg;
1298 int ret, irq, eirq, i;
1299 int irqbuf[9] /* maximum 8 channels + error IRQ */;
1300 unsigned int irqs = 0;
1302 BUILD_BUG_ON(sizeof(struct nbpf_desc_page) > PAGE_SIZE);
1305 if (!np || !of_id || !of_id->data)
1309 num_channels = cfg->num_channels;
1311 nbpf = devm_kzalloc(dev, sizeof(*nbpf) + num_channels *
1312 sizeof(nbpf->chan[0]), GFP_KERNEL);
1314 dev_err(dev, "Memory allocation failed\n");
1317 dma_dev = &nbpf->dma_dev;
1320 iomem = platform_get_resource(pdev, IORESOURCE_MEM, 0);
1321 nbpf->base = devm_ioremap_resource(dev, iomem);
1322 if (IS_ERR(nbpf->base))
1323 return PTR_ERR(nbpf->base);
1325 nbpf->clk = devm_clk_get(dev, NULL);
1326 if (IS_ERR(nbpf->clk))
1327 return PTR_ERR(nbpf->clk);
1331 for (i = 0; irqs < ARRAY_SIZE(irqbuf); i++) {
1332 irq_res = platform_get_resource(pdev, IORESOURCE_IRQ, i);
1336 for (irq = irq_res->start; irq <= irq_res->end;
1342 * 3 IRQ resource schemes are supported:
1343 * 1. 1 shared IRQ for error and all channels
1344 * 2. 2 IRQs: one for error and one shared for all channels
1345 * 3. 1 IRQ for error and an own IRQ for each channel
1347 if (irqs != 1 && irqs != 2 && irqs != num_channels + 1)
1353 for (i = 0; i <= num_channels; i++)
1354 nbpf->chan[i].irq = irqbuf[0];
1356 eirq = platform_get_irq_byname(pdev, "error");
1360 if (irqs == num_channels + 1) {
1361 struct nbpf_channel *chan;
1363 for (i = 0, chan = nbpf->chan; i <= num_channels;
1365 /* Skip the error IRQ */
1366 if (irqbuf[i] == eirq)
1368 chan->irq = irqbuf[i];
1371 if (chan != nbpf->chan + num_channels)
1374 /* 2 IRQs and more than one channel */
1375 if (irqbuf[0] == eirq)
1380 for (i = 0; i <= num_channels; i++)
1381 nbpf->chan[i].irq = irq;
1385 ret = devm_request_irq(dev, eirq, nbpf_err_irq,
1386 IRQF_SHARED, "dma error", nbpf);
1390 INIT_LIST_HEAD(&dma_dev->channels);
1392 /* Create DMA Channel */
1393 for (i = 0; i < num_channels; i++) {
1394 ret = nbpf_chan_probe(nbpf, i);
1399 dma_cap_set(DMA_MEMCPY, dma_dev->cap_mask);
1400 dma_cap_set(DMA_SLAVE, dma_dev->cap_mask);
1401 dma_cap_set(DMA_PRIVATE, dma_dev->cap_mask);
1402 dma_cap_set(DMA_SG, dma_dev->cap_mask);
1404 /* Common and MEMCPY operations */
1405 dma_dev->device_alloc_chan_resources
1406 = nbpf_alloc_chan_resources;
1407 dma_dev->device_free_chan_resources = nbpf_free_chan_resources;
1408 dma_dev->device_prep_dma_sg = nbpf_prep_memcpy_sg;
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 dma_dev->device_slave_caps = nbpf_slave_caps;
1415 * If we drop support for unaligned MEMCPY buffer addresses and / or
1416 * lengths by setting
1417 * dma_dev->copy_align = 4;
1418 * then we can set transfer length to 4 bytes in nbpf_prep_one() for
1422 /* Compulsory for DMA_SLAVE fields */
1423 dma_dev->device_prep_slave_sg = nbpf_prep_slave_sg;
1424 dma_dev->device_control = nbpf_control;
1426 platform_set_drvdata(pdev, nbpf);
1428 ret = clk_prepare_enable(nbpf->clk);
1432 nbpf_configure(nbpf);
1434 ret = dma_async_device_register(dma_dev);
1438 ret = of_dma_controller_register(np, nbpf_of_xlate, nbpf);
1440 goto e_dma_dev_unreg;
1445 dma_async_device_unregister(dma_dev);
1447 clk_disable_unprepare(nbpf->clk);
1452 static int nbpf_remove(struct platform_device *pdev)
1454 struct nbpf_device *nbpf = platform_get_drvdata(pdev);
1456 of_dma_controller_free(pdev->dev.of_node);
1457 dma_async_device_unregister(&nbpf->dma_dev);
1458 clk_disable_unprepare(nbpf->clk);
1463 static struct platform_device_id nbpf_ids[] = {
1464 {"nbpfaxi64dmac1b4", (kernel_ulong_t)&nbpf_cfg[NBPF1B4]},
1465 {"nbpfaxi64dmac1b8", (kernel_ulong_t)&nbpf_cfg[NBPF1B8]},
1466 {"nbpfaxi64dmac1b16", (kernel_ulong_t)&nbpf_cfg[NBPF1B16]},
1467 {"nbpfaxi64dmac4b4", (kernel_ulong_t)&nbpf_cfg[NBPF4B4]},
1468 {"nbpfaxi64dmac4b8", (kernel_ulong_t)&nbpf_cfg[NBPF4B8]},
1469 {"nbpfaxi64dmac4b16", (kernel_ulong_t)&nbpf_cfg[NBPF4B16]},
1470 {"nbpfaxi64dmac8b4", (kernel_ulong_t)&nbpf_cfg[NBPF8B4]},
1471 {"nbpfaxi64dmac8b8", (kernel_ulong_t)&nbpf_cfg[NBPF8B8]},
1472 {"nbpfaxi64dmac8b16", (kernel_ulong_t)&nbpf_cfg[NBPF8B16]},
1475 MODULE_DEVICE_TABLE(platform, nbpf_ids);
1477 #ifdef CONFIG_PM_RUNTIME
1478 static int nbpf_runtime_suspend(struct device *dev)
1480 struct nbpf_device *nbpf = platform_get_drvdata(to_platform_device(dev));
1481 clk_disable_unprepare(nbpf->clk);
1485 static int nbpf_runtime_resume(struct device *dev)
1487 struct nbpf_device *nbpf = platform_get_drvdata(to_platform_device(dev));
1488 return clk_prepare_enable(nbpf->clk);
1492 static const struct dev_pm_ops nbpf_pm_ops = {
1493 SET_RUNTIME_PM_OPS(nbpf_runtime_suspend, nbpf_runtime_resume, NULL)
1496 static struct platform_driver nbpf_driver = {
1498 .owner = THIS_MODULE,
1500 .of_match_table = nbpf_match,
1503 .id_table = nbpf_ids,
1504 .probe = nbpf_probe,
1505 .remove = nbpf_remove,
1508 module_platform_driver(nbpf_driver);
1510 MODULE_AUTHOR("Guennadi Liakhovetski <g.liakhovetski@gmx.de>");
1511 MODULE_DESCRIPTION("dmaengine driver for NBPFAXI64* DMACs");
1512 MODULE_LICENSE("GPL v2");