Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Special handling for DW DMA core
4 *
5 * Copyright (c) 2009, 2014 Intel Corporation.
6 */
7
8#include <linux/completion.h>
9#include <linux/dma-mapping.h>
10#include <linux/dmaengine.h>
11#include <linux/irqreturn.h>
12#include <linux/jiffies.h>
13#include <linux/module.h>
14#include <linux/pci.h>
15#include <linux/platform_data/dma-dw.h>
16#include <linux/spi/spi.h>
17#include <linux/types.h>
18
19#include "spi-dw.h"
20
21#define DW_SPI_RX_BUSY 0
22#define DW_SPI_RX_BURST_LEVEL 16
23#define DW_SPI_TX_BUSY 1
24#define DW_SPI_TX_BURST_LEVEL 16
25
26static bool dw_spi_dma_chan_filter(struct dma_chan *chan, void *param)
27{
28 struct dw_dma_slave *s = param;
29
30 if (s->dma_dev != chan->device->dev)
31 return false;
32
33 chan->private = s;
34 return true;
35}
36
37static void dw_spi_dma_maxburst_init(struct dw_spi *dws)
38{
39 struct dma_slave_caps caps;
40 u32 max_burst, def_burst;
41 int ret;
42
43 def_burst = dws->fifo_len / 2;
44
45 ret = dma_get_slave_caps(dws->rxchan, &caps);
46 if (!ret && caps.max_burst)
47 max_burst = caps.max_burst;
48 else
49 max_burst = DW_SPI_RX_BURST_LEVEL;
50
51 dws->rxburst = min(max_burst, def_burst);
52 dw_writel(dws, DW_SPI_DMARDLR, dws->rxburst - 1);
53
54 ret = dma_get_slave_caps(dws->txchan, &caps);
55 if (!ret && caps.max_burst)
56 max_burst = caps.max_burst;
57 else
58 max_burst = DW_SPI_TX_BURST_LEVEL;
59
60 /*
61 * Having a Rx DMA channel serviced with higher priority than a Tx DMA
62 * channel might not be enough to provide a well balanced DMA-based
63 * SPI transfer interface. There might still be moments when the Tx DMA
64 * channel is occasionally handled faster than the Rx DMA channel.
65 * That in its turn will eventually cause the SPI Rx FIFO overflow if
66 * SPI bus speed is high enough to fill the SPI Rx FIFO in before it's
67 * cleared by the Rx DMA channel. In order to fix the problem the Tx
68 * DMA activity is intentionally slowed down by limiting the SPI Tx
69 * FIFO depth with a value twice bigger than the Tx burst length.
70 */
71 dws->txburst = min(max_burst, def_burst);
72 dw_writel(dws, DW_SPI_DMATDLR, dws->txburst);
73}
74
75static void dw_spi_dma_sg_burst_init(struct dw_spi *dws)
76{
77 struct dma_slave_caps tx = {0}, rx = {0};
78
79 dma_get_slave_caps(dws->txchan, &tx);
80 dma_get_slave_caps(dws->rxchan, &rx);
81
82 if (tx.max_sg_burst > 0 && rx.max_sg_burst > 0)
83 dws->dma_sg_burst = min(tx.max_sg_burst, rx.max_sg_burst);
84 else if (tx.max_sg_burst > 0)
85 dws->dma_sg_burst = tx.max_sg_burst;
86 else if (rx.max_sg_burst > 0)
87 dws->dma_sg_burst = rx.max_sg_burst;
88 else
89 dws->dma_sg_burst = 0;
90}
91
92static int dw_spi_dma_init_mfld(struct device *dev, struct dw_spi *dws)
93{
94 struct dw_dma_slave dma_tx = { .dst_id = 1 }, *tx = &dma_tx;
95 struct dw_dma_slave dma_rx = { .src_id = 0 }, *rx = &dma_rx;
96 struct pci_dev *dma_dev;
97 dma_cap_mask_t mask;
98
99 /*
100 * Get pci device for DMA controller, currently it could only
101 * be the DMA controller of Medfield
102 */
103 dma_dev = pci_get_device(PCI_VENDOR_ID_INTEL, 0x0827, NULL);
104 if (!dma_dev)
105 return -ENODEV;
106
107 dma_cap_zero(mask);
108 dma_cap_set(DMA_SLAVE, mask);
109
110 /* 1. Init rx channel */
111 rx->dma_dev = &dma_dev->dev;
112 dws->rxchan = dma_request_channel(mask, dw_spi_dma_chan_filter, rx);
113 if (!dws->rxchan)
114 goto err_exit;
115
116 /* 2. Init tx channel */
117 tx->dma_dev = &dma_dev->dev;
118 dws->txchan = dma_request_channel(mask, dw_spi_dma_chan_filter, tx);
119 if (!dws->txchan)
120 goto free_rxchan;
121
122 dws->master->dma_rx = dws->rxchan;
123 dws->master->dma_tx = dws->txchan;
124
125 init_completion(&dws->dma_completion);
126
127 dw_spi_dma_maxburst_init(dws);
128
129 dw_spi_dma_sg_burst_init(dws);
130
131 pci_dev_put(dma_dev);
132
133 return 0;
134
135free_rxchan:
136 dma_release_channel(dws->rxchan);
137 dws->rxchan = NULL;
138err_exit:
139 pci_dev_put(dma_dev);
140 return -EBUSY;
141}
142
143static int dw_spi_dma_init_generic(struct device *dev, struct dw_spi *dws)
144{
145 int ret;
146
147 dws->rxchan = dma_request_chan(dev, "rx");
148 if (IS_ERR(dws->rxchan)) {
149 ret = PTR_ERR(dws->rxchan);
150 dws->rxchan = NULL;
151 goto err_exit;
152 }
153
154 dws->txchan = dma_request_chan(dev, "tx");
155 if (IS_ERR(dws->txchan)) {
156 ret = PTR_ERR(dws->txchan);
157 dws->txchan = NULL;
158 goto free_rxchan;
159 }
160
161 dws->master->dma_rx = dws->rxchan;
162 dws->master->dma_tx = dws->txchan;
163
164 init_completion(&dws->dma_completion);
165
166 dw_spi_dma_maxburst_init(dws);
167
168 dw_spi_dma_sg_burst_init(dws);
169
170 return 0;
171
172free_rxchan:
173 dma_release_channel(dws->rxchan);
174 dws->rxchan = NULL;
175err_exit:
176 return ret;
177}
178
179static void dw_spi_dma_exit(struct dw_spi *dws)
180{
181 if (dws->txchan) {
182 dmaengine_terminate_sync(dws->txchan);
183 dma_release_channel(dws->txchan);
184 }
185
186 if (dws->rxchan) {
187 dmaengine_terminate_sync(dws->rxchan);
188 dma_release_channel(dws->rxchan);
189 }
190}
191
192static irqreturn_t dw_spi_dma_transfer_handler(struct dw_spi *dws)
193{
194 dw_spi_check_status(dws, false);
195
196 complete(&dws->dma_completion);
197
198 return IRQ_HANDLED;
199}
200
201static bool dw_spi_can_dma(struct spi_controller *master,
202 struct spi_device *spi, struct spi_transfer *xfer)
203{
204 struct dw_spi *dws = spi_controller_get_devdata(master);
205
206 return xfer->len > dws->fifo_len;
207}
208
209static enum dma_slave_buswidth dw_spi_dma_convert_width(u8 n_bytes)
210{
211 if (n_bytes == 1)
212 return DMA_SLAVE_BUSWIDTH_1_BYTE;
213 else if (n_bytes == 2)
214 return DMA_SLAVE_BUSWIDTH_2_BYTES;
215
216 return DMA_SLAVE_BUSWIDTH_UNDEFINED;
217}
218
219static int dw_spi_dma_wait(struct dw_spi *dws, unsigned int len, u32 speed)
220{
221 unsigned long long ms;
222
223 ms = len * MSEC_PER_SEC * BITS_PER_BYTE;
224 do_div(ms, speed);
225 ms += ms + 200;
226
227 if (ms > UINT_MAX)
228 ms = UINT_MAX;
229
230 ms = wait_for_completion_timeout(&dws->dma_completion,
231 msecs_to_jiffies(ms));
232
233 if (ms == 0) {
234 dev_err(&dws->master->cur_msg->spi->dev,
235 "DMA transaction timed out\n");
236 return -ETIMEDOUT;
237 }
238
239 return 0;
240}
241
242static inline bool dw_spi_dma_tx_busy(struct dw_spi *dws)
243{
244 return !(dw_readl(dws, DW_SPI_SR) & DW_SPI_SR_TF_EMPT);
245}
246
247static int dw_spi_dma_wait_tx_done(struct dw_spi *dws,
248 struct spi_transfer *xfer)
249{
250 int retry = DW_SPI_WAIT_RETRIES;
251 struct spi_delay delay;
252 u32 nents;
253
254 nents = dw_readl(dws, DW_SPI_TXFLR);
255 delay.unit = SPI_DELAY_UNIT_SCK;
256 delay.value = nents * dws->n_bytes * BITS_PER_BYTE;
257
258 while (dw_spi_dma_tx_busy(dws) && retry--)
259 spi_delay_exec(&delay, xfer);
260
261 if (retry < 0) {
262 dev_err(&dws->master->dev, "Tx hanged up\n");
263 return -EIO;
264 }
265
266 return 0;
267}
268
269/*
270 * dws->dma_chan_busy is set before the dma transfer starts, callback for tx
271 * channel will clear a corresponding bit.
272 */
273static void dw_spi_dma_tx_done(void *arg)
274{
275 struct dw_spi *dws = arg;
276
277 clear_bit(DW_SPI_TX_BUSY, &dws->dma_chan_busy);
278 if (test_bit(DW_SPI_RX_BUSY, &dws->dma_chan_busy))
279 return;
280
281 complete(&dws->dma_completion);
282}
283
284static int dw_spi_dma_config_tx(struct dw_spi *dws)
285{
286 struct dma_slave_config txconf;
287
288 memset(&txconf, 0, sizeof(txconf));
289 txconf.direction = DMA_MEM_TO_DEV;
290 txconf.dst_addr = dws->dma_addr;
291 txconf.dst_maxburst = dws->txburst;
292 txconf.src_addr_width = DMA_SLAVE_BUSWIDTH_4_BYTES;
293 txconf.dst_addr_width = dw_spi_dma_convert_width(dws->n_bytes);
294 txconf.device_fc = false;
295
296 return dmaengine_slave_config(dws->txchan, &txconf);
297}
298
299static int dw_spi_dma_submit_tx(struct dw_spi *dws, struct scatterlist *sgl,
300 unsigned int nents)
301{
302 struct dma_async_tx_descriptor *txdesc;
303 dma_cookie_t cookie;
304 int ret;
305
306 txdesc = dmaengine_prep_slave_sg(dws->txchan, sgl, nents,
307 DMA_MEM_TO_DEV,
308 DMA_PREP_INTERRUPT | DMA_CTRL_ACK);
309 if (!txdesc)
310 return -ENOMEM;
311
312 txdesc->callback = dw_spi_dma_tx_done;
313 txdesc->callback_param = dws;
314
315 cookie = dmaengine_submit(txdesc);
316 ret = dma_submit_error(cookie);
317 if (ret) {
318 dmaengine_terminate_sync(dws->txchan);
319 return ret;
320 }
321
322 set_bit(DW_SPI_TX_BUSY, &dws->dma_chan_busy);
323
324 return 0;
325}
326
327static inline bool dw_spi_dma_rx_busy(struct dw_spi *dws)
328{
329 return !!(dw_readl(dws, DW_SPI_SR) & DW_SPI_SR_RF_NOT_EMPT);
330}
331
332static int dw_spi_dma_wait_rx_done(struct dw_spi *dws)
333{
334 int retry = DW_SPI_WAIT_RETRIES;
335 struct spi_delay delay;
336 unsigned long ns, us;
337 u32 nents;
338
339 /*
340 * It's unlikely that DMA engine is still doing the data fetching, but
341 * if it's let's give it some reasonable time. The timeout calculation
342 * is based on the synchronous APB/SSI reference clock rate, on a
343 * number of data entries left in the Rx FIFO, times a number of clock
344 * periods normally needed for a single APB read/write transaction
345 * without PREADY signal utilized (which is true for the DW APB SSI
346 * controller).
347 */
348 nents = dw_readl(dws, DW_SPI_RXFLR);
349 ns = 4U * NSEC_PER_SEC / dws->max_freq * nents;
350 if (ns <= NSEC_PER_USEC) {
351 delay.unit = SPI_DELAY_UNIT_NSECS;
352 delay.value = ns;
353 } else {
354 us = DIV_ROUND_UP(ns, NSEC_PER_USEC);
355 delay.unit = SPI_DELAY_UNIT_USECS;
356 delay.value = clamp_val(us, 0, USHRT_MAX);
357 }
358
359 while (dw_spi_dma_rx_busy(dws) && retry--)
360 spi_delay_exec(&delay, NULL);
361
362 if (retry < 0) {
363 dev_err(&dws->master->dev, "Rx hanged up\n");
364 return -EIO;
365 }
366
367 return 0;
368}
369
370/*
371 * dws->dma_chan_busy is set before the dma transfer starts, callback for rx
372 * channel will clear a corresponding bit.
373 */
374static void dw_spi_dma_rx_done(void *arg)
375{
376 struct dw_spi *dws = arg;
377
378 clear_bit(DW_SPI_RX_BUSY, &dws->dma_chan_busy);
379 if (test_bit(DW_SPI_TX_BUSY, &dws->dma_chan_busy))
380 return;
381
382 complete(&dws->dma_completion);
383}
384
385static int dw_spi_dma_config_rx(struct dw_spi *dws)
386{
387 struct dma_slave_config rxconf;
388
389 memset(&rxconf, 0, sizeof(rxconf));
390 rxconf.direction = DMA_DEV_TO_MEM;
391 rxconf.src_addr = dws->dma_addr;
392 rxconf.src_maxburst = dws->rxburst;
393 rxconf.dst_addr_width = DMA_SLAVE_BUSWIDTH_4_BYTES;
394 rxconf.src_addr_width = dw_spi_dma_convert_width(dws->n_bytes);
395 rxconf.device_fc = false;
396
397 return dmaengine_slave_config(dws->rxchan, &rxconf);
398}
399
400static int dw_spi_dma_submit_rx(struct dw_spi *dws, struct scatterlist *sgl,
401 unsigned int nents)
402{
403 struct dma_async_tx_descriptor *rxdesc;
404 dma_cookie_t cookie;
405 int ret;
406
407 rxdesc = dmaengine_prep_slave_sg(dws->rxchan, sgl, nents,
408 DMA_DEV_TO_MEM,
409 DMA_PREP_INTERRUPT | DMA_CTRL_ACK);
410 if (!rxdesc)
411 return -ENOMEM;
412
413 rxdesc->callback = dw_spi_dma_rx_done;
414 rxdesc->callback_param = dws;
415
416 cookie = dmaengine_submit(rxdesc);
417 ret = dma_submit_error(cookie);
418 if (ret) {
419 dmaengine_terminate_sync(dws->rxchan);
420 return ret;
421 }
422
423 set_bit(DW_SPI_RX_BUSY, &dws->dma_chan_busy);
424
425 return 0;
426}
427
428static int dw_spi_dma_setup(struct dw_spi *dws, struct spi_transfer *xfer)
429{
430 u16 imr, dma_ctrl;
431 int ret;
432
433 if (!xfer->tx_buf)
434 return -EINVAL;
435
436 /* Setup DMA channels */
437 ret = dw_spi_dma_config_tx(dws);
438 if (ret)
439 return ret;
440
441 if (xfer->rx_buf) {
442 ret = dw_spi_dma_config_rx(dws);
443 if (ret)
444 return ret;
445 }
446
447 /* Set the DMA handshaking interface */
448 dma_ctrl = DW_SPI_DMACR_TDMAE;
449 if (xfer->rx_buf)
450 dma_ctrl |= DW_SPI_DMACR_RDMAE;
451 dw_writel(dws, DW_SPI_DMACR, dma_ctrl);
452
453 /* Set the interrupt mask */
454 imr = DW_SPI_INT_TXOI;
455 if (xfer->rx_buf)
456 imr |= DW_SPI_INT_RXUI | DW_SPI_INT_RXOI;
457 dw_spi_umask_intr(dws, imr);
458
459 reinit_completion(&dws->dma_completion);
460
461 dws->transfer_handler = dw_spi_dma_transfer_handler;
462
463 return 0;
464}
465
466static int dw_spi_dma_transfer_all(struct dw_spi *dws,
467 struct spi_transfer *xfer)
468{
469 int ret;
470
471 /* Submit the DMA Tx transfer */
472 ret = dw_spi_dma_submit_tx(dws, xfer->tx_sg.sgl, xfer->tx_sg.nents);
473 if (ret)
474 goto err_clear_dmac;
475
476 /* Submit the DMA Rx transfer if required */
477 if (xfer->rx_buf) {
478 ret = dw_spi_dma_submit_rx(dws, xfer->rx_sg.sgl,
479 xfer->rx_sg.nents);
480 if (ret)
481 goto err_clear_dmac;
482
483 /* rx must be started before tx due to spi instinct */
484 dma_async_issue_pending(dws->rxchan);
485 }
486
487 dma_async_issue_pending(dws->txchan);
488
489 ret = dw_spi_dma_wait(dws, xfer->len, xfer->effective_speed_hz);
490
491err_clear_dmac:
492 dw_writel(dws, DW_SPI_DMACR, 0);
493
494 return ret;
495}
496
497/*
498 * In case if at least one of the requested DMA channels doesn't support the
499 * hardware accelerated SG list entries traverse, the DMA driver will most
500 * likely work that around by performing the IRQ-based SG list entries
501 * resubmission. That might and will cause a problem if the DMA Tx channel is
502 * recharged and re-executed before the Rx DMA channel. Due to
503 * non-deterministic IRQ-handler execution latency the DMA Tx channel will
504 * start pushing data to the SPI bus before the Rx DMA channel is even
505 * reinitialized with the next inbound SG list entry. By doing so the DMA Tx
506 * channel will implicitly start filling the DW APB SSI Rx FIFO up, which while
507 * the DMA Rx channel being recharged and re-executed will eventually be
508 * overflown.
509 *
510 * In order to solve the problem we have to feed the DMA engine with SG list
511 * entries one-by-one. It shall keep the DW APB SSI Tx and Rx FIFOs
512 * synchronized and prevent the Rx FIFO overflow. Since in general the tx_sg
513 * and rx_sg lists may have different number of entries of different lengths
514 * (though total length should match) let's virtually split the SG-lists to the
515 * set of DMA transfers, which length is a minimum of the ordered SG-entries
516 * lengths. An ASCII-sketch of the implemented algo is following:
517 * xfer->len
518 * |___________|
519 * tx_sg list: |___|____|__|
520 * rx_sg list: |_|____|____|
521 * DMA transfers: |_|_|__|_|__|
522 *
523 * Note in order to have this workaround solving the denoted problem the DMA
524 * engine driver should properly initialize the max_sg_burst capability and set
525 * the DMA device max segment size parameter with maximum data block size the
526 * DMA engine supports.
527 */
528
529static int dw_spi_dma_transfer_one(struct dw_spi *dws,
530 struct spi_transfer *xfer)
531{
532 struct scatterlist *tx_sg = NULL, *rx_sg = NULL, tx_tmp, rx_tmp;
533 unsigned int tx_len = 0, rx_len = 0;
534 unsigned int base, len;
535 int ret;
536
537 sg_init_table(&tx_tmp, 1);
538 sg_init_table(&rx_tmp, 1);
539
540 for (base = 0, len = 0; base < xfer->len; base += len) {
541 /* Fetch next Tx DMA data chunk */
542 if (!tx_len) {
543 tx_sg = !tx_sg ? &xfer->tx_sg.sgl[0] : sg_next(tx_sg);
544 sg_dma_address(&tx_tmp) = sg_dma_address(tx_sg);
545 tx_len = sg_dma_len(tx_sg);
546 }
547
548 /* Fetch next Rx DMA data chunk */
549 if (!rx_len) {
550 rx_sg = !rx_sg ? &xfer->rx_sg.sgl[0] : sg_next(rx_sg);
551 sg_dma_address(&rx_tmp) = sg_dma_address(rx_sg);
552 rx_len = sg_dma_len(rx_sg);
553 }
554
555 len = min(tx_len, rx_len);
556
557 sg_dma_len(&tx_tmp) = len;
558 sg_dma_len(&rx_tmp) = len;
559
560 /* Submit DMA Tx transfer */
561 ret = dw_spi_dma_submit_tx(dws, &tx_tmp, 1);
562 if (ret)
563 break;
564
565 /* Submit DMA Rx transfer */
566 ret = dw_spi_dma_submit_rx(dws, &rx_tmp, 1);
567 if (ret)
568 break;
569
570 /* Rx must be started before Tx due to SPI instinct */
571 dma_async_issue_pending(dws->rxchan);
572
573 dma_async_issue_pending(dws->txchan);
574
575 /*
576 * Here we only need to wait for the DMA transfer to be
577 * finished since SPI controller is kept enabled during the
578 * procedure this loop implements and there is no risk to lose
579 * data left in the Tx/Rx FIFOs.
580 */
581 ret = dw_spi_dma_wait(dws, len, xfer->effective_speed_hz);
582 if (ret)
583 break;
584
585 reinit_completion(&dws->dma_completion);
586
587 sg_dma_address(&tx_tmp) += len;
588 sg_dma_address(&rx_tmp) += len;
589 tx_len -= len;
590 rx_len -= len;
591 }
592
593 dw_writel(dws, DW_SPI_DMACR, 0);
594
595 return ret;
596}
597
598static int dw_spi_dma_transfer(struct dw_spi *dws, struct spi_transfer *xfer)
599{
600 unsigned int nents;
601 int ret;
602
603 nents = max(xfer->tx_sg.nents, xfer->rx_sg.nents);
604
605 /*
606 * Execute normal DMA-based transfer (which submits the Rx and Tx SG
607 * lists directly to the DMA engine at once) if either full hardware
608 * accelerated SG list traverse is supported by both channels, or the
609 * Tx-only SPI transfer is requested, or the DMA engine is capable to
610 * handle both SG lists on hardware accelerated basis.
611 */
612 if (!dws->dma_sg_burst || !xfer->rx_buf || nents <= dws->dma_sg_burst)
613 ret = dw_spi_dma_transfer_all(dws, xfer);
614 else
615 ret = dw_spi_dma_transfer_one(dws, xfer);
616 if (ret)
617 return ret;
618
619 if (dws->master->cur_msg->status == -EINPROGRESS) {
620 ret = dw_spi_dma_wait_tx_done(dws, xfer);
621 if (ret)
622 return ret;
623 }
624
625 if (xfer->rx_buf && dws->master->cur_msg->status == -EINPROGRESS)
626 ret = dw_spi_dma_wait_rx_done(dws);
627
628 return ret;
629}
630
631static void dw_spi_dma_stop(struct dw_spi *dws)
632{
633 if (test_bit(DW_SPI_TX_BUSY, &dws->dma_chan_busy)) {
634 dmaengine_terminate_sync(dws->txchan);
635 clear_bit(DW_SPI_TX_BUSY, &dws->dma_chan_busy);
636 }
637 if (test_bit(DW_SPI_RX_BUSY, &dws->dma_chan_busy)) {
638 dmaengine_terminate_sync(dws->rxchan);
639 clear_bit(DW_SPI_RX_BUSY, &dws->dma_chan_busy);
640 }
641}
642
643static const struct dw_spi_dma_ops dw_spi_dma_mfld_ops = {
644 .dma_init = dw_spi_dma_init_mfld,
645 .dma_exit = dw_spi_dma_exit,
646 .dma_setup = dw_spi_dma_setup,
647 .can_dma = dw_spi_can_dma,
648 .dma_transfer = dw_spi_dma_transfer,
649 .dma_stop = dw_spi_dma_stop,
650};
651
652void dw_spi_dma_setup_mfld(struct dw_spi *dws)
653{
654 dws->dma_ops = &dw_spi_dma_mfld_ops;
655}
656EXPORT_SYMBOL_NS_GPL(dw_spi_dma_setup_mfld, SPI_DW_CORE);
657
658static const struct dw_spi_dma_ops dw_spi_dma_generic_ops = {
659 .dma_init = dw_spi_dma_init_generic,
660 .dma_exit = dw_spi_dma_exit,
661 .dma_setup = dw_spi_dma_setup,
662 .can_dma = dw_spi_can_dma,
663 .dma_transfer = dw_spi_dma_transfer,
664 .dma_stop = dw_spi_dma_stop,
665};
666
667void dw_spi_dma_setup_generic(struct dw_spi *dws)
668{
669 dws->dma_ops = &dw_spi_dma_generic_ops;
670}
671EXPORT_SYMBOL_NS_GPL(dw_spi_dma_setup_generic, SPI_DW_CORE);
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Special handling for DW DMA core
4 *
5 * Copyright (c) 2009, 2014 Intel Corporation.
6 */
7
8#include <linux/completion.h>
9#include <linux/dma-mapping.h>
10#include <linux/dmaengine.h>
11#include <linux/irqreturn.h>
12#include <linux/jiffies.h>
13#include <linux/pci.h>
14#include <linux/platform_data/dma-dw.h>
15#include <linux/spi/spi.h>
16#include <linux/types.h>
17
18#include "spi-dw.h"
19
20#define RX_BUSY 0
21#define RX_BURST_LEVEL 16
22#define TX_BUSY 1
23#define TX_BURST_LEVEL 16
24
25static bool dw_spi_dma_chan_filter(struct dma_chan *chan, void *param)
26{
27 struct dw_dma_slave *s = param;
28
29 if (s->dma_dev != chan->device->dev)
30 return false;
31
32 chan->private = s;
33 return true;
34}
35
36static void dw_spi_dma_maxburst_init(struct dw_spi *dws)
37{
38 struct dma_slave_caps caps;
39 u32 max_burst, def_burst;
40 int ret;
41
42 def_burst = dws->fifo_len / 2;
43
44 ret = dma_get_slave_caps(dws->rxchan, &caps);
45 if (!ret && caps.max_burst)
46 max_burst = caps.max_burst;
47 else
48 max_burst = RX_BURST_LEVEL;
49
50 dws->rxburst = min(max_burst, def_burst);
51 dw_writel(dws, DW_SPI_DMARDLR, dws->rxburst - 1);
52
53 ret = dma_get_slave_caps(dws->txchan, &caps);
54 if (!ret && caps.max_burst)
55 max_burst = caps.max_burst;
56 else
57 max_burst = TX_BURST_LEVEL;
58
59 /*
60 * Having a Rx DMA channel serviced with higher priority than a Tx DMA
61 * channel might not be enough to provide a well balanced DMA-based
62 * SPI transfer interface. There might still be moments when the Tx DMA
63 * channel is occasionally handled faster than the Rx DMA channel.
64 * That in its turn will eventually cause the SPI Rx FIFO overflow if
65 * SPI bus speed is high enough to fill the SPI Rx FIFO in before it's
66 * cleared by the Rx DMA channel. In order to fix the problem the Tx
67 * DMA activity is intentionally slowed down by limiting the SPI Tx
68 * FIFO depth with a value twice bigger than the Tx burst length.
69 */
70 dws->txburst = min(max_burst, def_burst);
71 dw_writel(dws, DW_SPI_DMATDLR, dws->txburst);
72}
73
74static void dw_spi_dma_sg_burst_init(struct dw_spi *dws)
75{
76 struct dma_slave_caps tx = {0}, rx = {0};
77
78 dma_get_slave_caps(dws->txchan, &tx);
79 dma_get_slave_caps(dws->rxchan, &rx);
80
81 if (tx.max_sg_burst > 0 && rx.max_sg_burst > 0)
82 dws->dma_sg_burst = min(tx.max_sg_burst, rx.max_sg_burst);
83 else if (tx.max_sg_burst > 0)
84 dws->dma_sg_burst = tx.max_sg_burst;
85 else if (rx.max_sg_burst > 0)
86 dws->dma_sg_burst = rx.max_sg_burst;
87 else
88 dws->dma_sg_burst = 0;
89}
90
91static int dw_spi_dma_init_mfld(struct device *dev, struct dw_spi *dws)
92{
93 struct dw_dma_slave dma_tx = { .dst_id = 1 }, *tx = &dma_tx;
94 struct dw_dma_slave dma_rx = { .src_id = 0 }, *rx = &dma_rx;
95 struct pci_dev *dma_dev;
96 dma_cap_mask_t mask;
97
98 /*
99 * Get pci device for DMA controller, currently it could only
100 * be the DMA controller of Medfield
101 */
102 dma_dev = pci_get_device(PCI_VENDOR_ID_INTEL, 0x0827, NULL);
103 if (!dma_dev)
104 return -ENODEV;
105
106 dma_cap_zero(mask);
107 dma_cap_set(DMA_SLAVE, mask);
108
109 /* 1. Init rx channel */
110 rx->dma_dev = &dma_dev->dev;
111 dws->rxchan = dma_request_channel(mask, dw_spi_dma_chan_filter, rx);
112 if (!dws->rxchan)
113 goto err_exit;
114
115 /* 2. Init tx channel */
116 tx->dma_dev = &dma_dev->dev;
117 dws->txchan = dma_request_channel(mask, dw_spi_dma_chan_filter, tx);
118 if (!dws->txchan)
119 goto free_rxchan;
120
121 dws->master->dma_rx = dws->rxchan;
122 dws->master->dma_tx = dws->txchan;
123
124 init_completion(&dws->dma_completion);
125
126 dw_spi_dma_maxburst_init(dws);
127
128 dw_spi_dma_sg_burst_init(dws);
129
130 return 0;
131
132free_rxchan:
133 dma_release_channel(dws->rxchan);
134 dws->rxchan = NULL;
135err_exit:
136 return -EBUSY;
137}
138
139static int dw_spi_dma_init_generic(struct device *dev, struct dw_spi *dws)
140{
141 dws->rxchan = dma_request_slave_channel(dev, "rx");
142 if (!dws->rxchan)
143 return -ENODEV;
144
145 dws->txchan = dma_request_slave_channel(dev, "tx");
146 if (!dws->txchan) {
147 dma_release_channel(dws->rxchan);
148 dws->rxchan = NULL;
149 return -ENODEV;
150 }
151
152 dws->master->dma_rx = dws->rxchan;
153 dws->master->dma_tx = dws->txchan;
154
155 init_completion(&dws->dma_completion);
156
157 dw_spi_dma_maxburst_init(dws);
158
159 dw_spi_dma_sg_burst_init(dws);
160
161 return 0;
162}
163
164static void dw_spi_dma_exit(struct dw_spi *dws)
165{
166 if (dws->txchan) {
167 dmaengine_terminate_sync(dws->txchan);
168 dma_release_channel(dws->txchan);
169 }
170
171 if (dws->rxchan) {
172 dmaengine_terminate_sync(dws->rxchan);
173 dma_release_channel(dws->rxchan);
174 }
175}
176
177static irqreturn_t dw_spi_dma_transfer_handler(struct dw_spi *dws)
178{
179 dw_spi_check_status(dws, false);
180
181 complete(&dws->dma_completion);
182
183 return IRQ_HANDLED;
184}
185
186static bool dw_spi_can_dma(struct spi_controller *master,
187 struct spi_device *spi, struct spi_transfer *xfer)
188{
189 struct dw_spi *dws = spi_controller_get_devdata(master);
190
191 return xfer->len > dws->fifo_len;
192}
193
194static enum dma_slave_buswidth dw_spi_dma_convert_width(u8 n_bytes)
195{
196 if (n_bytes == 1)
197 return DMA_SLAVE_BUSWIDTH_1_BYTE;
198 else if (n_bytes == 2)
199 return DMA_SLAVE_BUSWIDTH_2_BYTES;
200
201 return DMA_SLAVE_BUSWIDTH_UNDEFINED;
202}
203
204static int dw_spi_dma_wait(struct dw_spi *dws, unsigned int len, u32 speed)
205{
206 unsigned long long ms;
207
208 ms = len * MSEC_PER_SEC * BITS_PER_BYTE;
209 do_div(ms, speed);
210 ms += ms + 200;
211
212 if (ms > UINT_MAX)
213 ms = UINT_MAX;
214
215 ms = wait_for_completion_timeout(&dws->dma_completion,
216 msecs_to_jiffies(ms));
217
218 if (ms == 0) {
219 dev_err(&dws->master->cur_msg->spi->dev,
220 "DMA transaction timed out\n");
221 return -ETIMEDOUT;
222 }
223
224 return 0;
225}
226
227static inline bool dw_spi_dma_tx_busy(struct dw_spi *dws)
228{
229 return !(dw_readl(dws, DW_SPI_SR) & SR_TF_EMPT);
230}
231
232static int dw_spi_dma_wait_tx_done(struct dw_spi *dws,
233 struct spi_transfer *xfer)
234{
235 int retry = SPI_WAIT_RETRIES;
236 struct spi_delay delay;
237 u32 nents;
238
239 nents = dw_readl(dws, DW_SPI_TXFLR);
240 delay.unit = SPI_DELAY_UNIT_SCK;
241 delay.value = nents * dws->n_bytes * BITS_PER_BYTE;
242
243 while (dw_spi_dma_tx_busy(dws) && retry--)
244 spi_delay_exec(&delay, xfer);
245
246 if (retry < 0) {
247 dev_err(&dws->master->dev, "Tx hanged up\n");
248 return -EIO;
249 }
250
251 return 0;
252}
253
254/*
255 * dws->dma_chan_busy is set before the dma transfer starts, callback for tx
256 * channel will clear a corresponding bit.
257 */
258static void dw_spi_dma_tx_done(void *arg)
259{
260 struct dw_spi *dws = arg;
261
262 clear_bit(TX_BUSY, &dws->dma_chan_busy);
263 if (test_bit(RX_BUSY, &dws->dma_chan_busy))
264 return;
265
266 complete(&dws->dma_completion);
267}
268
269static int dw_spi_dma_config_tx(struct dw_spi *dws)
270{
271 struct dma_slave_config txconf;
272
273 memset(&txconf, 0, sizeof(txconf));
274 txconf.direction = DMA_MEM_TO_DEV;
275 txconf.dst_addr = dws->dma_addr;
276 txconf.dst_maxburst = dws->txburst;
277 txconf.src_addr_width = DMA_SLAVE_BUSWIDTH_4_BYTES;
278 txconf.dst_addr_width = dw_spi_dma_convert_width(dws->n_bytes);
279 txconf.device_fc = false;
280
281 return dmaengine_slave_config(dws->txchan, &txconf);
282}
283
284static int dw_spi_dma_submit_tx(struct dw_spi *dws, struct scatterlist *sgl,
285 unsigned int nents)
286{
287 struct dma_async_tx_descriptor *txdesc;
288 dma_cookie_t cookie;
289 int ret;
290
291 txdesc = dmaengine_prep_slave_sg(dws->txchan, sgl, nents,
292 DMA_MEM_TO_DEV,
293 DMA_PREP_INTERRUPT | DMA_CTRL_ACK);
294 if (!txdesc)
295 return -ENOMEM;
296
297 txdesc->callback = dw_spi_dma_tx_done;
298 txdesc->callback_param = dws;
299
300 cookie = dmaengine_submit(txdesc);
301 ret = dma_submit_error(cookie);
302 if (ret) {
303 dmaengine_terminate_sync(dws->txchan);
304 return ret;
305 }
306
307 set_bit(TX_BUSY, &dws->dma_chan_busy);
308
309 return 0;
310}
311
312static inline bool dw_spi_dma_rx_busy(struct dw_spi *dws)
313{
314 return !!(dw_readl(dws, DW_SPI_SR) & SR_RF_NOT_EMPT);
315}
316
317static int dw_spi_dma_wait_rx_done(struct dw_spi *dws)
318{
319 int retry = SPI_WAIT_RETRIES;
320 struct spi_delay delay;
321 unsigned long ns, us;
322 u32 nents;
323
324 /*
325 * It's unlikely that DMA engine is still doing the data fetching, but
326 * if it's let's give it some reasonable time. The timeout calculation
327 * is based on the synchronous APB/SSI reference clock rate, on a
328 * number of data entries left in the Rx FIFO, times a number of clock
329 * periods normally needed for a single APB read/write transaction
330 * without PREADY signal utilized (which is true for the DW APB SSI
331 * controller).
332 */
333 nents = dw_readl(dws, DW_SPI_RXFLR);
334 ns = 4U * NSEC_PER_SEC / dws->max_freq * nents;
335 if (ns <= NSEC_PER_USEC) {
336 delay.unit = SPI_DELAY_UNIT_NSECS;
337 delay.value = ns;
338 } else {
339 us = DIV_ROUND_UP(ns, NSEC_PER_USEC);
340 delay.unit = SPI_DELAY_UNIT_USECS;
341 delay.value = clamp_val(us, 0, USHRT_MAX);
342 }
343
344 while (dw_spi_dma_rx_busy(dws) && retry--)
345 spi_delay_exec(&delay, NULL);
346
347 if (retry < 0) {
348 dev_err(&dws->master->dev, "Rx hanged up\n");
349 return -EIO;
350 }
351
352 return 0;
353}
354
355/*
356 * dws->dma_chan_busy is set before the dma transfer starts, callback for rx
357 * channel will clear a corresponding bit.
358 */
359static void dw_spi_dma_rx_done(void *arg)
360{
361 struct dw_spi *dws = arg;
362
363 clear_bit(RX_BUSY, &dws->dma_chan_busy);
364 if (test_bit(TX_BUSY, &dws->dma_chan_busy))
365 return;
366
367 complete(&dws->dma_completion);
368}
369
370static int dw_spi_dma_config_rx(struct dw_spi *dws)
371{
372 struct dma_slave_config rxconf;
373
374 memset(&rxconf, 0, sizeof(rxconf));
375 rxconf.direction = DMA_DEV_TO_MEM;
376 rxconf.src_addr = dws->dma_addr;
377 rxconf.src_maxburst = dws->rxburst;
378 rxconf.dst_addr_width = DMA_SLAVE_BUSWIDTH_4_BYTES;
379 rxconf.src_addr_width = dw_spi_dma_convert_width(dws->n_bytes);
380 rxconf.device_fc = false;
381
382 return dmaengine_slave_config(dws->rxchan, &rxconf);
383}
384
385static int dw_spi_dma_submit_rx(struct dw_spi *dws, struct scatterlist *sgl,
386 unsigned int nents)
387{
388 struct dma_async_tx_descriptor *rxdesc;
389 dma_cookie_t cookie;
390 int ret;
391
392 rxdesc = dmaengine_prep_slave_sg(dws->rxchan, sgl, nents,
393 DMA_DEV_TO_MEM,
394 DMA_PREP_INTERRUPT | DMA_CTRL_ACK);
395 if (!rxdesc)
396 return -ENOMEM;
397
398 rxdesc->callback = dw_spi_dma_rx_done;
399 rxdesc->callback_param = dws;
400
401 cookie = dmaengine_submit(rxdesc);
402 ret = dma_submit_error(cookie);
403 if (ret) {
404 dmaengine_terminate_sync(dws->rxchan);
405 return ret;
406 }
407
408 set_bit(RX_BUSY, &dws->dma_chan_busy);
409
410 return 0;
411}
412
413static int dw_spi_dma_setup(struct dw_spi *dws, struct spi_transfer *xfer)
414{
415 u16 imr, dma_ctrl;
416 int ret;
417
418 if (!xfer->tx_buf)
419 return -EINVAL;
420
421 /* Setup DMA channels */
422 ret = dw_spi_dma_config_tx(dws);
423 if (ret)
424 return ret;
425
426 if (xfer->rx_buf) {
427 ret = dw_spi_dma_config_rx(dws);
428 if (ret)
429 return ret;
430 }
431
432 /* Set the DMA handshaking interface */
433 dma_ctrl = SPI_DMA_TDMAE;
434 if (xfer->rx_buf)
435 dma_ctrl |= SPI_DMA_RDMAE;
436 dw_writel(dws, DW_SPI_DMACR, dma_ctrl);
437
438 /* Set the interrupt mask */
439 imr = SPI_INT_TXOI;
440 if (xfer->rx_buf)
441 imr |= SPI_INT_RXUI | SPI_INT_RXOI;
442 spi_umask_intr(dws, imr);
443
444 reinit_completion(&dws->dma_completion);
445
446 dws->transfer_handler = dw_spi_dma_transfer_handler;
447
448 return 0;
449}
450
451static int dw_spi_dma_transfer_all(struct dw_spi *dws,
452 struct spi_transfer *xfer)
453{
454 int ret;
455
456 /* Submit the DMA Tx transfer */
457 ret = dw_spi_dma_submit_tx(dws, xfer->tx_sg.sgl, xfer->tx_sg.nents);
458 if (ret)
459 goto err_clear_dmac;
460
461 /* Submit the DMA Rx transfer if required */
462 if (xfer->rx_buf) {
463 ret = dw_spi_dma_submit_rx(dws, xfer->rx_sg.sgl,
464 xfer->rx_sg.nents);
465 if (ret)
466 goto err_clear_dmac;
467
468 /* rx must be started before tx due to spi instinct */
469 dma_async_issue_pending(dws->rxchan);
470 }
471
472 dma_async_issue_pending(dws->txchan);
473
474 ret = dw_spi_dma_wait(dws, xfer->len, xfer->effective_speed_hz);
475
476err_clear_dmac:
477 dw_writel(dws, DW_SPI_DMACR, 0);
478
479 return ret;
480}
481
482/*
483 * In case if at least one of the requested DMA channels doesn't support the
484 * hardware accelerated SG list entries traverse, the DMA driver will most
485 * likely work that around by performing the IRQ-based SG list entries
486 * resubmission. That might and will cause a problem if the DMA Tx channel is
487 * recharged and re-executed before the Rx DMA channel. Due to
488 * non-deterministic IRQ-handler execution latency the DMA Tx channel will
489 * start pushing data to the SPI bus before the Rx DMA channel is even
490 * reinitialized with the next inbound SG list entry. By doing so the DMA Tx
491 * channel will implicitly start filling the DW APB SSI Rx FIFO up, which while
492 * the DMA Rx channel being recharged and re-executed will eventually be
493 * overflown.
494 *
495 * In order to solve the problem we have to feed the DMA engine with SG list
496 * entries one-by-one. It shall keep the DW APB SSI Tx and Rx FIFOs
497 * synchronized and prevent the Rx FIFO overflow. Since in general the tx_sg
498 * and rx_sg lists may have different number of entries of different lengths
499 * (though total length should match) let's virtually split the SG-lists to the
500 * set of DMA transfers, which length is a minimum of the ordered SG-entries
501 * lengths. An ASCII-sketch of the implemented algo is following:
502 * xfer->len
503 * |___________|
504 * tx_sg list: |___|____|__|
505 * rx_sg list: |_|____|____|
506 * DMA transfers: |_|_|__|_|__|
507 *
508 * Note in order to have this workaround solving the denoted problem the DMA
509 * engine driver should properly initialize the max_sg_burst capability and set
510 * the DMA device max segment size parameter with maximum data block size the
511 * DMA engine supports.
512 */
513
514static int dw_spi_dma_transfer_one(struct dw_spi *dws,
515 struct spi_transfer *xfer)
516{
517 struct scatterlist *tx_sg = NULL, *rx_sg = NULL, tx_tmp, rx_tmp;
518 unsigned int tx_len = 0, rx_len = 0;
519 unsigned int base, len;
520 int ret;
521
522 sg_init_table(&tx_tmp, 1);
523 sg_init_table(&rx_tmp, 1);
524
525 for (base = 0, len = 0; base < xfer->len; base += len) {
526 /* Fetch next Tx DMA data chunk */
527 if (!tx_len) {
528 tx_sg = !tx_sg ? &xfer->tx_sg.sgl[0] : sg_next(tx_sg);
529 sg_dma_address(&tx_tmp) = sg_dma_address(tx_sg);
530 tx_len = sg_dma_len(tx_sg);
531 }
532
533 /* Fetch next Rx DMA data chunk */
534 if (!rx_len) {
535 rx_sg = !rx_sg ? &xfer->rx_sg.sgl[0] : sg_next(rx_sg);
536 sg_dma_address(&rx_tmp) = sg_dma_address(rx_sg);
537 rx_len = sg_dma_len(rx_sg);
538 }
539
540 len = min(tx_len, rx_len);
541
542 sg_dma_len(&tx_tmp) = len;
543 sg_dma_len(&rx_tmp) = len;
544
545 /* Submit DMA Tx transfer */
546 ret = dw_spi_dma_submit_tx(dws, &tx_tmp, 1);
547 if (ret)
548 break;
549
550 /* Submit DMA Rx transfer */
551 ret = dw_spi_dma_submit_rx(dws, &rx_tmp, 1);
552 if (ret)
553 break;
554
555 /* Rx must be started before Tx due to SPI instinct */
556 dma_async_issue_pending(dws->rxchan);
557
558 dma_async_issue_pending(dws->txchan);
559
560 /*
561 * Here we only need to wait for the DMA transfer to be
562 * finished since SPI controller is kept enabled during the
563 * procedure this loop implements and there is no risk to lose
564 * data left in the Tx/Rx FIFOs.
565 */
566 ret = dw_spi_dma_wait(dws, len, xfer->effective_speed_hz);
567 if (ret)
568 break;
569
570 reinit_completion(&dws->dma_completion);
571
572 sg_dma_address(&tx_tmp) += len;
573 sg_dma_address(&rx_tmp) += len;
574 tx_len -= len;
575 rx_len -= len;
576 }
577
578 dw_writel(dws, DW_SPI_DMACR, 0);
579
580 return ret;
581}
582
583static int dw_spi_dma_transfer(struct dw_spi *dws, struct spi_transfer *xfer)
584{
585 unsigned int nents;
586 int ret;
587
588 nents = max(xfer->tx_sg.nents, xfer->rx_sg.nents);
589
590 /*
591 * Execute normal DMA-based transfer (which submits the Rx and Tx SG
592 * lists directly to the DMA engine at once) if either full hardware
593 * accelerated SG list traverse is supported by both channels, or the
594 * Tx-only SPI transfer is requested, or the DMA engine is capable to
595 * handle both SG lists on hardware accelerated basis.
596 */
597 if (!dws->dma_sg_burst || !xfer->rx_buf || nents <= dws->dma_sg_burst)
598 ret = dw_spi_dma_transfer_all(dws, xfer);
599 else
600 ret = dw_spi_dma_transfer_one(dws, xfer);
601 if (ret)
602 return ret;
603
604 if (dws->master->cur_msg->status == -EINPROGRESS) {
605 ret = dw_spi_dma_wait_tx_done(dws, xfer);
606 if (ret)
607 return ret;
608 }
609
610 if (xfer->rx_buf && dws->master->cur_msg->status == -EINPROGRESS)
611 ret = dw_spi_dma_wait_rx_done(dws);
612
613 return ret;
614}
615
616static void dw_spi_dma_stop(struct dw_spi *dws)
617{
618 if (test_bit(TX_BUSY, &dws->dma_chan_busy)) {
619 dmaengine_terminate_sync(dws->txchan);
620 clear_bit(TX_BUSY, &dws->dma_chan_busy);
621 }
622 if (test_bit(RX_BUSY, &dws->dma_chan_busy)) {
623 dmaengine_terminate_sync(dws->rxchan);
624 clear_bit(RX_BUSY, &dws->dma_chan_busy);
625 }
626}
627
628static const struct dw_spi_dma_ops dw_spi_dma_mfld_ops = {
629 .dma_init = dw_spi_dma_init_mfld,
630 .dma_exit = dw_spi_dma_exit,
631 .dma_setup = dw_spi_dma_setup,
632 .can_dma = dw_spi_can_dma,
633 .dma_transfer = dw_spi_dma_transfer,
634 .dma_stop = dw_spi_dma_stop,
635};
636
637void dw_spi_dma_setup_mfld(struct dw_spi *dws)
638{
639 dws->dma_ops = &dw_spi_dma_mfld_ops;
640}
641EXPORT_SYMBOL_GPL(dw_spi_dma_setup_mfld);
642
643static const struct dw_spi_dma_ops dw_spi_dma_generic_ops = {
644 .dma_init = dw_spi_dma_init_generic,
645 .dma_exit = dw_spi_dma_exit,
646 .dma_setup = dw_spi_dma_setup,
647 .can_dma = dw_spi_can_dma,
648 .dma_transfer = dw_spi_dma_transfer,
649 .dma_stop = dw_spi_dma_stop,
650};
651
652void dw_spi_dma_setup_generic(struct dw_spi *dws)
653{
654 dws->dma_ops = &dw_spi_dma_generic_ops;
655}
656EXPORT_SYMBOL_GPL(dw_spi_dma_setup_generic);