1 // SPDX-License-Identifier: GPL-2.0-only
3 * Special handling for DW DMA core
5 * Copyright (c) 2009, 2014 Intel Corporation.
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>
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
26 static bool dw_spi_dma_chan_filter(struct dma_chan *chan, void *param)
28 struct dw_dma_slave *s = param;
30 if (s->dma_dev != chan->device->dev)
37 static void dw_spi_dma_maxburst_init(struct dw_spi *dws)
39 struct dma_slave_caps caps;
40 u32 max_burst, def_burst;
43 def_burst = dws->fifo_len / 2;
45 ret = dma_get_slave_caps(dws->rxchan, &caps);
46 if (!ret && caps.max_burst)
47 max_burst = caps.max_burst;
49 max_burst = DW_SPI_RX_BURST_LEVEL;
51 dws->rxburst = min(max_burst, def_burst);
52 dw_writel(dws, DW_SPI_DMARDLR, dws->rxburst - 1);
54 ret = dma_get_slave_caps(dws->txchan, &caps);
55 if (!ret && caps.max_burst)
56 max_burst = caps.max_burst;
58 max_burst = DW_SPI_TX_BURST_LEVEL;
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.
71 dws->txburst = min(max_burst, def_burst);
72 dw_writel(dws, DW_SPI_DMATDLR, dws->txburst);
75 static void dw_spi_dma_sg_burst_init(struct dw_spi *dws)
77 struct dma_slave_caps tx = {0}, rx = {0};
79 dma_get_slave_caps(dws->txchan, &tx);
80 dma_get_slave_caps(dws->rxchan, &rx);
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;
89 dws->dma_sg_burst = 0;
92 static int dw_spi_dma_init_mfld(struct device *dev, struct dw_spi *dws)
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;
100 * Get pci device for DMA controller, currently it could only
101 * be the DMA controller of Medfield
103 dma_dev = pci_get_device(PCI_VENDOR_ID_INTEL, 0x0827, NULL);
108 dma_cap_set(DMA_SLAVE, mask);
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);
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);
122 dws->master->dma_rx = dws->rxchan;
123 dws->master->dma_tx = dws->txchan;
125 init_completion(&dws->dma_completion);
127 dw_spi_dma_maxburst_init(dws);
129 dw_spi_dma_sg_burst_init(dws);
134 dma_release_channel(dws->rxchan);
140 static int dw_spi_dma_init_generic(struct device *dev, struct dw_spi *dws)
144 dws->rxchan = dma_request_chan(dev, "rx");
145 if (IS_ERR(dws->rxchan)) {
146 ret = PTR_ERR(dws->rxchan);
151 dws->txchan = dma_request_chan(dev, "tx");
152 if (IS_ERR(dws->txchan)) {
153 ret = PTR_ERR(dws->txchan);
158 dws->master->dma_rx = dws->rxchan;
159 dws->master->dma_tx = dws->txchan;
161 init_completion(&dws->dma_completion);
163 dw_spi_dma_maxburst_init(dws);
165 dw_spi_dma_sg_burst_init(dws);
170 dma_release_channel(dws->rxchan);
176 static void dw_spi_dma_exit(struct dw_spi *dws)
179 dmaengine_terminate_sync(dws->txchan);
180 dma_release_channel(dws->txchan);
184 dmaengine_terminate_sync(dws->rxchan);
185 dma_release_channel(dws->rxchan);
189 static irqreturn_t dw_spi_dma_transfer_handler(struct dw_spi *dws)
191 dw_spi_check_status(dws, false);
193 complete(&dws->dma_completion);
198 static bool dw_spi_can_dma(struct spi_controller *master,
199 struct spi_device *spi, struct spi_transfer *xfer)
201 struct dw_spi *dws = spi_controller_get_devdata(master);
203 return xfer->len > dws->fifo_len;
206 static enum dma_slave_buswidth dw_spi_dma_convert_width(u8 n_bytes)
209 return DMA_SLAVE_BUSWIDTH_1_BYTE;
210 else if (n_bytes == 2)
211 return DMA_SLAVE_BUSWIDTH_2_BYTES;
213 return DMA_SLAVE_BUSWIDTH_UNDEFINED;
216 static int dw_spi_dma_wait(struct dw_spi *dws, unsigned int len, u32 speed)
218 unsigned long long ms;
220 ms = len * MSEC_PER_SEC * BITS_PER_BYTE;
227 ms = wait_for_completion_timeout(&dws->dma_completion,
228 msecs_to_jiffies(ms));
231 dev_err(&dws->master->cur_msg->spi->dev,
232 "DMA transaction timed out\n");
239 static inline bool dw_spi_dma_tx_busy(struct dw_spi *dws)
241 return !(dw_readl(dws, DW_SPI_SR) & DW_SPI_SR_TF_EMPT);
244 static int dw_spi_dma_wait_tx_done(struct dw_spi *dws,
245 struct spi_transfer *xfer)
247 int retry = DW_SPI_WAIT_RETRIES;
248 struct spi_delay delay;
251 nents = dw_readl(dws, DW_SPI_TXFLR);
252 delay.unit = SPI_DELAY_UNIT_SCK;
253 delay.value = nents * dws->n_bytes * BITS_PER_BYTE;
255 while (dw_spi_dma_tx_busy(dws) && retry--)
256 spi_delay_exec(&delay, xfer);
259 dev_err(&dws->master->dev, "Tx hanged up\n");
267 * dws->dma_chan_busy is set before the dma transfer starts, callback for tx
268 * channel will clear a corresponding bit.
270 static void dw_spi_dma_tx_done(void *arg)
272 struct dw_spi *dws = arg;
274 clear_bit(DW_SPI_TX_BUSY, &dws->dma_chan_busy);
275 if (test_bit(DW_SPI_RX_BUSY, &dws->dma_chan_busy))
278 complete(&dws->dma_completion);
281 static int dw_spi_dma_config_tx(struct dw_spi *dws)
283 struct dma_slave_config txconf;
285 memset(&txconf, 0, sizeof(txconf));
286 txconf.direction = DMA_MEM_TO_DEV;
287 txconf.dst_addr = dws->dma_addr;
288 txconf.dst_maxburst = dws->txburst;
289 txconf.src_addr_width = DMA_SLAVE_BUSWIDTH_4_BYTES;
290 txconf.dst_addr_width = dw_spi_dma_convert_width(dws->n_bytes);
291 txconf.device_fc = false;
293 return dmaengine_slave_config(dws->txchan, &txconf);
296 static int dw_spi_dma_submit_tx(struct dw_spi *dws, struct scatterlist *sgl,
299 struct dma_async_tx_descriptor *txdesc;
303 txdesc = dmaengine_prep_slave_sg(dws->txchan, sgl, nents,
305 DMA_PREP_INTERRUPT | DMA_CTRL_ACK);
309 txdesc->callback = dw_spi_dma_tx_done;
310 txdesc->callback_param = dws;
312 cookie = dmaengine_submit(txdesc);
313 ret = dma_submit_error(cookie);
315 dmaengine_terminate_sync(dws->txchan);
319 set_bit(DW_SPI_TX_BUSY, &dws->dma_chan_busy);
324 static inline bool dw_spi_dma_rx_busy(struct dw_spi *dws)
326 return !!(dw_readl(dws, DW_SPI_SR) & DW_SPI_SR_RF_NOT_EMPT);
329 static int dw_spi_dma_wait_rx_done(struct dw_spi *dws)
331 int retry = DW_SPI_WAIT_RETRIES;
332 struct spi_delay delay;
333 unsigned long ns, us;
337 * It's unlikely that DMA engine is still doing the data fetching, but
338 * if it's let's give it some reasonable time. The timeout calculation
339 * is based on the synchronous APB/SSI reference clock rate, on a
340 * number of data entries left in the Rx FIFO, times a number of clock
341 * periods normally needed for a single APB read/write transaction
342 * without PREADY signal utilized (which is true for the DW APB SSI
345 nents = dw_readl(dws, DW_SPI_RXFLR);
346 ns = 4U * NSEC_PER_SEC / dws->max_freq * nents;
347 if (ns <= NSEC_PER_USEC) {
348 delay.unit = SPI_DELAY_UNIT_NSECS;
351 us = DIV_ROUND_UP(ns, NSEC_PER_USEC);
352 delay.unit = SPI_DELAY_UNIT_USECS;
353 delay.value = clamp_val(us, 0, USHRT_MAX);
356 while (dw_spi_dma_rx_busy(dws) && retry--)
357 spi_delay_exec(&delay, NULL);
360 dev_err(&dws->master->dev, "Rx hanged up\n");
368 * dws->dma_chan_busy is set before the dma transfer starts, callback for rx
369 * channel will clear a corresponding bit.
371 static void dw_spi_dma_rx_done(void *arg)
373 struct dw_spi *dws = arg;
375 clear_bit(DW_SPI_RX_BUSY, &dws->dma_chan_busy);
376 if (test_bit(DW_SPI_TX_BUSY, &dws->dma_chan_busy))
379 complete(&dws->dma_completion);
382 static int dw_spi_dma_config_rx(struct dw_spi *dws)
384 struct dma_slave_config rxconf;
386 memset(&rxconf, 0, sizeof(rxconf));
387 rxconf.direction = DMA_DEV_TO_MEM;
388 rxconf.src_addr = dws->dma_addr;
389 rxconf.src_maxburst = dws->rxburst;
390 rxconf.dst_addr_width = DMA_SLAVE_BUSWIDTH_4_BYTES;
391 rxconf.src_addr_width = dw_spi_dma_convert_width(dws->n_bytes);
392 rxconf.device_fc = false;
394 return dmaengine_slave_config(dws->rxchan, &rxconf);
397 static int dw_spi_dma_submit_rx(struct dw_spi *dws, struct scatterlist *sgl,
400 struct dma_async_tx_descriptor *rxdesc;
404 rxdesc = dmaengine_prep_slave_sg(dws->rxchan, sgl, nents,
406 DMA_PREP_INTERRUPT | DMA_CTRL_ACK);
410 rxdesc->callback = dw_spi_dma_rx_done;
411 rxdesc->callback_param = dws;
413 cookie = dmaengine_submit(rxdesc);
414 ret = dma_submit_error(cookie);
416 dmaengine_terminate_sync(dws->rxchan);
420 set_bit(DW_SPI_RX_BUSY, &dws->dma_chan_busy);
425 static int dw_spi_dma_setup(struct dw_spi *dws, struct spi_transfer *xfer)
433 /* Setup DMA channels */
434 ret = dw_spi_dma_config_tx(dws);
439 ret = dw_spi_dma_config_rx(dws);
444 /* Set the DMA handshaking interface */
445 dma_ctrl = DW_SPI_DMACR_TDMAE;
447 dma_ctrl |= DW_SPI_DMACR_RDMAE;
448 dw_writel(dws, DW_SPI_DMACR, dma_ctrl);
450 /* Set the interrupt mask */
451 imr = DW_SPI_INT_TXOI;
453 imr |= DW_SPI_INT_RXUI | DW_SPI_INT_RXOI;
454 dw_spi_umask_intr(dws, imr);
456 reinit_completion(&dws->dma_completion);
458 dws->transfer_handler = dw_spi_dma_transfer_handler;
463 static int dw_spi_dma_transfer_all(struct dw_spi *dws,
464 struct spi_transfer *xfer)
468 /* Submit the DMA Tx transfer */
469 ret = dw_spi_dma_submit_tx(dws, xfer->tx_sg.sgl, xfer->tx_sg.nents);
473 /* Submit the DMA Rx transfer if required */
475 ret = dw_spi_dma_submit_rx(dws, xfer->rx_sg.sgl,
480 /* rx must be started before tx due to spi instinct */
481 dma_async_issue_pending(dws->rxchan);
484 dma_async_issue_pending(dws->txchan);
486 ret = dw_spi_dma_wait(dws, xfer->len, xfer->effective_speed_hz);
489 dw_writel(dws, DW_SPI_DMACR, 0);
495 * In case if at least one of the requested DMA channels doesn't support the
496 * hardware accelerated SG list entries traverse, the DMA driver will most
497 * likely work that around by performing the IRQ-based SG list entries
498 * resubmission. That might and will cause a problem if the DMA Tx channel is
499 * recharged and re-executed before the Rx DMA channel. Due to
500 * non-deterministic IRQ-handler execution latency the DMA Tx channel will
501 * start pushing data to the SPI bus before the Rx DMA channel is even
502 * reinitialized with the next inbound SG list entry. By doing so the DMA Tx
503 * channel will implicitly start filling the DW APB SSI Rx FIFO up, which while
504 * the DMA Rx channel being recharged and re-executed will eventually be
507 * In order to solve the problem we have to feed the DMA engine with SG list
508 * entries one-by-one. It shall keep the DW APB SSI Tx and Rx FIFOs
509 * synchronized and prevent the Rx FIFO overflow. Since in general the tx_sg
510 * and rx_sg lists may have different number of entries of different lengths
511 * (though total length should match) let's virtually split the SG-lists to the
512 * set of DMA transfers, which length is a minimum of the ordered SG-entries
513 * lengths. An ASCII-sketch of the implemented algo is following:
516 * tx_sg list: |___|____|__|
517 * rx_sg list: |_|____|____|
518 * DMA transfers: |_|_|__|_|__|
520 * Note in order to have this workaround solving the denoted problem the DMA
521 * engine driver should properly initialize the max_sg_burst capability and set
522 * the DMA device max segment size parameter with maximum data block size the
523 * DMA engine supports.
526 static int dw_spi_dma_transfer_one(struct dw_spi *dws,
527 struct spi_transfer *xfer)
529 struct scatterlist *tx_sg = NULL, *rx_sg = NULL, tx_tmp, rx_tmp;
530 unsigned int tx_len = 0, rx_len = 0;
531 unsigned int base, len;
534 sg_init_table(&tx_tmp, 1);
535 sg_init_table(&rx_tmp, 1);
537 for (base = 0, len = 0; base < xfer->len; base += len) {
538 /* Fetch next Tx DMA data chunk */
540 tx_sg = !tx_sg ? &xfer->tx_sg.sgl[0] : sg_next(tx_sg);
541 sg_dma_address(&tx_tmp) = sg_dma_address(tx_sg);
542 tx_len = sg_dma_len(tx_sg);
545 /* Fetch next Rx DMA data chunk */
547 rx_sg = !rx_sg ? &xfer->rx_sg.sgl[0] : sg_next(rx_sg);
548 sg_dma_address(&rx_tmp) = sg_dma_address(rx_sg);
549 rx_len = sg_dma_len(rx_sg);
552 len = min(tx_len, rx_len);
554 sg_dma_len(&tx_tmp) = len;
555 sg_dma_len(&rx_tmp) = len;
557 /* Submit DMA Tx transfer */
558 ret = dw_spi_dma_submit_tx(dws, &tx_tmp, 1);
562 /* Submit DMA Rx transfer */
563 ret = dw_spi_dma_submit_rx(dws, &rx_tmp, 1);
567 /* Rx must be started before Tx due to SPI instinct */
568 dma_async_issue_pending(dws->rxchan);
570 dma_async_issue_pending(dws->txchan);
573 * Here we only need to wait for the DMA transfer to be
574 * finished since SPI controller is kept enabled during the
575 * procedure this loop implements and there is no risk to lose
576 * data left in the Tx/Rx FIFOs.
578 ret = dw_spi_dma_wait(dws, len, xfer->effective_speed_hz);
582 reinit_completion(&dws->dma_completion);
584 sg_dma_address(&tx_tmp) += len;
585 sg_dma_address(&rx_tmp) += len;
590 dw_writel(dws, DW_SPI_DMACR, 0);
595 static int dw_spi_dma_transfer(struct dw_spi *dws, struct spi_transfer *xfer)
600 nents = max(xfer->tx_sg.nents, xfer->rx_sg.nents);
603 * Execute normal DMA-based transfer (which submits the Rx and Tx SG
604 * lists directly to the DMA engine at once) if either full hardware
605 * accelerated SG list traverse is supported by both channels, or the
606 * Tx-only SPI transfer is requested, or the DMA engine is capable to
607 * handle both SG lists on hardware accelerated basis.
609 if (!dws->dma_sg_burst || !xfer->rx_buf || nents <= dws->dma_sg_burst)
610 ret = dw_spi_dma_transfer_all(dws, xfer);
612 ret = dw_spi_dma_transfer_one(dws, xfer);
616 if (dws->master->cur_msg->status == -EINPROGRESS) {
617 ret = dw_spi_dma_wait_tx_done(dws, xfer);
622 if (xfer->rx_buf && dws->master->cur_msg->status == -EINPROGRESS)
623 ret = dw_spi_dma_wait_rx_done(dws);
628 static void dw_spi_dma_stop(struct dw_spi *dws)
630 if (test_bit(DW_SPI_TX_BUSY, &dws->dma_chan_busy)) {
631 dmaengine_terminate_sync(dws->txchan);
632 clear_bit(DW_SPI_TX_BUSY, &dws->dma_chan_busy);
634 if (test_bit(DW_SPI_RX_BUSY, &dws->dma_chan_busy)) {
635 dmaengine_terminate_sync(dws->rxchan);
636 clear_bit(DW_SPI_RX_BUSY, &dws->dma_chan_busy);
640 static const struct dw_spi_dma_ops dw_spi_dma_mfld_ops = {
641 .dma_init = dw_spi_dma_init_mfld,
642 .dma_exit = dw_spi_dma_exit,
643 .dma_setup = dw_spi_dma_setup,
644 .can_dma = dw_spi_can_dma,
645 .dma_transfer = dw_spi_dma_transfer,
646 .dma_stop = dw_spi_dma_stop,
649 void dw_spi_dma_setup_mfld(struct dw_spi *dws)
651 dws->dma_ops = &dw_spi_dma_mfld_ops;
653 EXPORT_SYMBOL_NS_GPL(dw_spi_dma_setup_mfld, SPI_DW_CORE);
655 static const struct dw_spi_dma_ops dw_spi_dma_generic_ops = {
656 .dma_init = dw_spi_dma_init_generic,
657 .dma_exit = dw_spi_dma_exit,
658 .dma_setup = dw_spi_dma_setup,
659 .can_dma = dw_spi_can_dma,
660 .dma_transfer = dw_spi_dma_transfer,
661 .dma_stop = dw_spi_dma_stop,
664 void dw_spi_dma_setup_generic(struct dw_spi *dws)
666 dws->dma_ops = &dw_spi_dma_generic_ops;
668 EXPORT_SYMBOL_NS_GPL(dw_spi_dma_setup_generic, SPI_DW_CORE);