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/pci.h>
14 #include <linux/platform_data/dma-dw.h>
15 #include <linux/spi/spi.h>
16 #include <linux/types.h>
21 #define RX_BURST_LEVEL 16
23 #define TX_BURST_LEVEL 16
25 static bool dw_spi_dma_chan_filter(struct dma_chan *chan, void *param)
27 struct dw_dma_slave *s = param;
29 if (s->dma_dev != chan->device->dev)
36 static void dw_spi_dma_maxburst_init(struct dw_spi *dws)
38 struct dma_slave_caps caps;
39 u32 max_burst, def_burst;
42 def_burst = dws->fifo_len / 2;
44 ret = dma_get_slave_caps(dws->rxchan, &caps);
45 if (!ret && caps.max_burst)
46 max_burst = caps.max_burst;
48 max_burst = RX_BURST_LEVEL;
50 dws->rxburst = min(max_burst, def_burst);
51 dw_writel(dws, DW_SPI_DMARDLR, dws->rxburst - 1);
53 ret = dma_get_slave_caps(dws->txchan, &caps);
54 if (!ret && caps.max_burst)
55 max_burst = caps.max_burst;
57 max_burst = TX_BURST_LEVEL;
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.
70 dws->txburst = min(max_burst, def_burst);
71 dw_writel(dws, DW_SPI_DMATDLR, dws->txburst);
74 static void dw_spi_dma_sg_burst_init(struct dw_spi *dws)
76 struct dma_slave_caps tx = {0}, rx = {0};
78 dma_get_slave_caps(dws->txchan, &tx);
79 dma_get_slave_caps(dws->rxchan, &rx);
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;
88 dws->dma_sg_burst = 0;
91 static int dw_spi_dma_init_mfld(struct device *dev, struct dw_spi *dws)
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;
99 * Get pci device for DMA controller, currently it could only
100 * be the DMA controller of Medfield
102 dma_dev = pci_get_device(PCI_VENDOR_ID_INTEL, 0x0827, NULL);
107 dma_cap_set(DMA_SLAVE, mask);
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);
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);
121 dws->master->dma_rx = dws->rxchan;
122 dws->master->dma_tx = dws->txchan;
124 init_completion(&dws->dma_completion);
126 dw_spi_dma_maxburst_init(dws);
128 dw_spi_dma_sg_burst_init(dws);
133 dma_release_channel(dws->rxchan);
139 static int dw_spi_dma_init_generic(struct device *dev, struct dw_spi *dws)
141 dws->rxchan = dma_request_slave_channel(dev, "rx");
145 dws->txchan = dma_request_slave_channel(dev, "tx");
147 dma_release_channel(dws->rxchan);
152 dws->master->dma_rx = dws->rxchan;
153 dws->master->dma_tx = dws->txchan;
155 init_completion(&dws->dma_completion);
157 dw_spi_dma_maxburst_init(dws);
159 dw_spi_dma_sg_burst_init(dws);
164 static void dw_spi_dma_exit(struct dw_spi *dws)
167 dmaengine_terminate_sync(dws->txchan);
168 dma_release_channel(dws->txchan);
172 dmaengine_terminate_sync(dws->rxchan);
173 dma_release_channel(dws->rxchan);
177 static irqreturn_t dw_spi_dma_transfer_handler(struct dw_spi *dws)
179 u16 irq_status = dw_readl(dws, DW_SPI_ISR);
184 dw_readl(dws, DW_SPI_ICR);
187 dev_err(&dws->master->dev, "%s: FIFO overrun/underrun\n", __func__);
188 dws->master->cur_msg->status = -EIO;
189 complete(&dws->dma_completion);
193 static bool dw_spi_can_dma(struct spi_controller *master,
194 struct spi_device *spi, struct spi_transfer *xfer)
196 struct dw_spi *dws = spi_controller_get_devdata(master);
198 return xfer->len > dws->fifo_len;
201 static enum dma_slave_buswidth dw_spi_dma_convert_width(u8 n_bytes)
204 return DMA_SLAVE_BUSWIDTH_1_BYTE;
205 else if (n_bytes == 2)
206 return DMA_SLAVE_BUSWIDTH_2_BYTES;
208 return DMA_SLAVE_BUSWIDTH_UNDEFINED;
211 static int dw_spi_dma_wait(struct dw_spi *dws, unsigned int len, u32 speed)
213 unsigned long long ms;
215 ms = len * MSEC_PER_SEC * BITS_PER_BYTE;
222 ms = wait_for_completion_timeout(&dws->dma_completion,
223 msecs_to_jiffies(ms));
226 dev_err(&dws->master->cur_msg->spi->dev,
227 "DMA transaction timed out\n");
234 static inline bool dw_spi_dma_tx_busy(struct dw_spi *dws)
236 return !(dw_readl(dws, DW_SPI_SR) & SR_TF_EMPT);
239 static int dw_spi_dma_wait_tx_done(struct dw_spi *dws,
240 struct spi_transfer *xfer)
242 int retry = SPI_WAIT_RETRIES;
243 struct spi_delay delay;
246 nents = dw_readl(dws, DW_SPI_TXFLR);
247 delay.unit = SPI_DELAY_UNIT_SCK;
248 delay.value = nents * dws->n_bytes * BITS_PER_BYTE;
250 while (dw_spi_dma_tx_busy(dws) && retry--)
251 spi_delay_exec(&delay, xfer);
254 dev_err(&dws->master->dev, "Tx hanged up\n");
262 * dws->dma_chan_busy is set before the dma transfer starts, callback for tx
263 * channel will clear a corresponding bit.
265 static void dw_spi_dma_tx_done(void *arg)
267 struct dw_spi *dws = arg;
269 clear_bit(TX_BUSY, &dws->dma_chan_busy);
270 if (test_bit(RX_BUSY, &dws->dma_chan_busy))
273 complete(&dws->dma_completion);
276 static int dw_spi_dma_config_tx(struct dw_spi *dws)
278 struct dma_slave_config txconf;
280 memset(&txconf, 0, sizeof(txconf));
281 txconf.direction = DMA_MEM_TO_DEV;
282 txconf.dst_addr = dws->dma_addr;
283 txconf.dst_maxburst = dws->txburst;
284 txconf.src_addr_width = DMA_SLAVE_BUSWIDTH_4_BYTES;
285 txconf.dst_addr_width = dw_spi_dma_convert_width(dws->n_bytes);
286 txconf.device_fc = false;
288 return dmaengine_slave_config(dws->txchan, &txconf);
291 static int dw_spi_dma_submit_tx(struct dw_spi *dws, struct scatterlist *sgl,
294 struct dma_async_tx_descriptor *txdesc;
298 txdesc = dmaengine_prep_slave_sg(dws->txchan, sgl, nents,
300 DMA_PREP_INTERRUPT | DMA_CTRL_ACK);
304 txdesc->callback = dw_spi_dma_tx_done;
305 txdesc->callback_param = dws;
307 cookie = dmaengine_submit(txdesc);
308 ret = dma_submit_error(cookie);
310 dmaengine_terminate_sync(dws->txchan);
314 set_bit(TX_BUSY, &dws->dma_chan_busy);
319 static inline bool dw_spi_dma_rx_busy(struct dw_spi *dws)
321 return !!(dw_readl(dws, DW_SPI_SR) & SR_RF_NOT_EMPT);
324 static int dw_spi_dma_wait_rx_done(struct dw_spi *dws)
326 int retry = SPI_WAIT_RETRIES;
327 struct spi_delay delay;
328 unsigned long ns, us;
332 * It's unlikely that DMA engine is still doing the data fetching, but
333 * if it's let's give it some reasonable time. The timeout calculation
334 * is based on the synchronous APB/SSI reference clock rate, on a
335 * number of data entries left in the Rx FIFO, times a number of clock
336 * periods normally needed for a single APB read/write transaction
337 * without PREADY signal utilized (which is true for the DW APB SSI
340 nents = dw_readl(dws, DW_SPI_RXFLR);
341 ns = 4U * NSEC_PER_SEC / dws->max_freq * nents;
342 if (ns <= NSEC_PER_USEC) {
343 delay.unit = SPI_DELAY_UNIT_NSECS;
346 us = DIV_ROUND_UP(ns, NSEC_PER_USEC);
347 delay.unit = SPI_DELAY_UNIT_USECS;
348 delay.value = clamp_val(us, 0, USHRT_MAX);
351 while (dw_spi_dma_rx_busy(dws) && retry--)
352 spi_delay_exec(&delay, NULL);
355 dev_err(&dws->master->dev, "Rx hanged up\n");
363 * dws->dma_chan_busy is set before the dma transfer starts, callback for rx
364 * channel will clear a corresponding bit.
366 static void dw_spi_dma_rx_done(void *arg)
368 struct dw_spi *dws = arg;
370 clear_bit(RX_BUSY, &dws->dma_chan_busy);
371 if (test_bit(TX_BUSY, &dws->dma_chan_busy))
374 complete(&dws->dma_completion);
377 static int dw_spi_dma_config_rx(struct dw_spi *dws)
379 struct dma_slave_config rxconf;
381 memset(&rxconf, 0, sizeof(rxconf));
382 rxconf.direction = DMA_DEV_TO_MEM;
383 rxconf.src_addr = dws->dma_addr;
384 rxconf.src_maxburst = dws->rxburst;
385 rxconf.dst_addr_width = DMA_SLAVE_BUSWIDTH_4_BYTES;
386 rxconf.src_addr_width = dw_spi_dma_convert_width(dws->n_bytes);
387 rxconf.device_fc = false;
389 return dmaengine_slave_config(dws->rxchan, &rxconf);
392 static int dw_spi_dma_submit_rx(struct dw_spi *dws, struct scatterlist *sgl,
395 struct dma_async_tx_descriptor *rxdesc;
399 rxdesc = dmaengine_prep_slave_sg(dws->rxchan, sgl, nents,
401 DMA_PREP_INTERRUPT | DMA_CTRL_ACK);
405 rxdesc->callback = dw_spi_dma_rx_done;
406 rxdesc->callback_param = dws;
408 cookie = dmaengine_submit(rxdesc);
409 ret = dma_submit_error(cookie);
411 dmaengine_terminate_sync(dws->rxchan);
415 set_bit(RX_BUSY, &dws->dma_chan_busy);
420 static int dw_spi_dma_setup(struct dw_spi *dws, struct spi_transfer *xfer)
428 /* Setup DMA channels */
429 ret = dw_spi_dma_config_tx(dws);
434 ret = dw_spi_dma_config_rx(dws);
439 /* Set the DMA handshaking interface */
440 dma_ctrl = SPI_DMA_TDMAE;
442 dma_ctrl |= SPI_DMA_RDMAE;
443 dw_writel(dws, DW_SPI_DMACR, dma_ctrl);
445 /* Set the interrupt mask */
448 imr |= SPI_INT_RXUI | SPI_INT_RXOI;
449 spi_umask_intr(dws, imr);
451 reinit_completion(&dws->dma_completion);
453 dws->transfer_handler = dw_spi_dma_transfer_handler;
458 static int dw_spi_dma_transfer_all(struct dw_spi *dws,
459 struct spi_transfer *xfer)
463 /* Submit the DMA Tx transfer */
464 ret = dw_spi_dma_submit_tx(dws, xfer->tx_sg.sgl, xfer->tx_sg.nents);
468 /* Submit the DMA Rx transfer if required */
470 ret = dw_spi_dma_submit_rx(dws, xfer->rx_sg.sgl,
475 /* rx must be started before tx due to spi instinct */
476 dma_async_issue_pending(dws->rxchan);
479 dma_async_issue_pending(dws->txchan);
481 ret = dw_spi_dma_wait(dws, xfer->len, xfer->effective_speed_hz);
484 dw_writel(dws, DW_SPI_DMACR, 0);
490 * In case if at least one of the requested DMA channels doesn't support the
491 * hardware accelerated SG list entries traverse, the DMA driver will most
492 * likely work that around by performing the IRQ-based SG list entries
493 * resubmission. That might and will cause a problem if the DMA Tx channel is
494 * recharged and re-executed before the Rx DMA channel. Due to
495 * non-deterministic IRQ-handler execution latency the DMA Tx channel will
496 * start pushing data to the SPI bus before the Rx DMA channel is even
497 * reinitialized with the next inbound SG list entry. By doing so the DMA Tx
498 * channel will implicitly start filling the DW APB SSI Rx FIFO up, which while
499 * the DMA Rx channel being recharged and re-executed will eventually be
502 * In order to solve the problem we have to feed the DMA engine with SG list
503 * entries one-by-one. It shall keep the DW APB SSI Tx and Rx FIFOs
504 * synchronized and prevent the Rx FIFO overflow. Since in general the tx_sg
505 * and rx_sg lists may have different number of entries of different lengths
506 * (though total length should match) let's virtually split the SG-lists to the
507 * set of DMA transfers, which length is a minimum of the ordered SG-entries
508 * lengths. An ASCII-sketch of the implemented algo is following:
511 * tx_sg list: |___|____|__|
512 * rx_sg list: |_|____|____|
513 * DMA transfers: |_|_|__|_|__|
515 * Note in order to have this workaround solving the denoted problem the DMA
516 * engine driver should properly initialize the max_sg_burst capability and set
517 * the DMA device max segment size parameter with maximum data block size the
518 * DMA engine supports.
521 static int dw_spi_dma_transfer_one(struct dw_spi *dws,
522 struct spi_transfer *xfer)
524 struct scatterlist *tx_sg = NULL, *rx_sg = NULL, tx_tmp, rx_tmp;
525 unsigned int tx_len = 0, rx_len = 0;
526 unsigned int base, len;
529 sg_init_table(&tx_tmp, 1);
530 sg_init_table(&rx_tmp, 1);
532 for (base = 0, len = 0; base < xfer->len; base += len) {
533 /* Fetch next Tx DMA data chunk */
535 tx_sg = !tx_sg ? &xfer->tx_sg.sgl[0] : sg_next(tx_sg);
536 sg_dma_address(&tx_tmp) = sg_dma_address(tx_sg);
537 tx_len = sg_dma_len(tx_sg);
540 /* Fetch next Rx DMA data chunk */
542 rx_sg = !rx_sg ? &xfer->rx_sg.sgl[0] : sg_next(rx_sg);
543 sg_dma_address(&rx_tmp) = sg_dma_address(rx_sg);
544 rx_len = sg_dma_len(rx_sg);
547 len = min(tx_len, rx_len);
549 sg_dma_len(&tx_tmp) = len;
550 sg_dma_len(&rx_tmp) = len;
552 /* Submit DMA Tx transfer */
553 ret = dw_spi_dma_submit_tx(dws, &tx_tmp, 1);
557 /* Submit DMA Rx transfer */
558 ret = dw_spi_dma_submit_rx(dws, &rx_tmp, 1);
562 /* Rx must be started before Tx due to SPI instinct */
563 dma_async_issue_pending(dws->rxchan);
565 dma_async_issue_pending(dws->txchan);
568 * Here we only need to wait for the DMA transfer to be
569 * finished since SPI controller is kept enabled during the
570 * procedure this loop implements and there is no risk to lose
571 * data left in the Tx/Rx FIFOs.
573 ret = dw_spi_dma_wait(dws, len, xfer->effective_speed_hz);
577 reinit_completion(&dws->dma_completion);
579 sg_dma_address(&tx_tmp) += len;
580 sg_dma_address(&rx_tmp) += len;
585 dw_writel(dws, DW_SPI_DMACR, 0);
590 static int dw_spi_dma_transfer(struct dw_spi *dws, struct spi_transfer *xfer)
595 nents = max(xfer->tx_sg.nents, xfer->rx_sg.nents);
598 * Execute normal DMA-based transfer (which submits the Rx and Tx SG
599 * lists directly to the DMA engine at once) if either full hardware
600 * accelerated SG list traverse is supported by both channels, or the
601 * Tx-only SPI transfer is requested, or the DMA engine is capable to
602 * handle both SG lists on hardware accelerated basis.
604 if (!dws->dma_sg_burst || !xfer->rx_buf || nents <= dws->dma_sg_burst)
605 ret = dw_spi_dma_transfer_all(dws, xfer);
607 ret = dw_spi_dma_transfer_one(dws, xfer);
611 if (dws->master->cur_msg->status == -EINPROGRESS) {
612 ret = dw_spi_dma_wait_tx_done(dws, xfer);
617 if (xfer->rx_buf && dws->master->cur_msg->status == -EINPROGRESS)
618 ret = dw_spi_dma_wait_rx_done(dws);
623 static void dw_spi_dma_stop(struct dw_spi *dws)
625 if (test_bit(TX_BUSY, &dws->dma_chan_busy)) {
626 dmaengine_terminate_sync(dws->txchan);
627 clear_bit(TX_BUSY, &dws->dma_chan_busy);
629 if (test_bit(RX_BUSY, &dws->dma_chan_busy)) {
630 dmaengine_terminate_sync(dws->rxchan);
631 clear_bit(RX_BUSY, &dws->dma_chan_busy);
635 static const struct dw_spi_dma_ops dw_spi_dma_mfld_ops = {
636 .dma_init = dw_spi_dma_init_mfld,
637 .dma_exit = dw_spi_dma_exit,
638 .dma_setup = dw_spi_dma_setup,
639 .can_dma = dw_spi_can_dma,
640 .dma_transfer = dw_spi_dma_transfer,
641 .dma_stop = dw_spi_dma_stop,
644 void dw_spi_dma_setup_mfld(struct dw_spi *dws)
646 dws->dma_ops = &dw_spi_dma_mfld_ops;
648 EXPORT_SYMBOL_GPL(dw_spi_dma_setup_mfld);
650 static const struct dw_spi_dma_ops dw_spi_dma_generic_ops = {
651 .dma_init = dw_spi_dma_init_generic,
652 .dma_exit = dw_spi_dma_exit,
653 .dma_setup = dw_spi_dma_setup,
654 .can_dma = dw_spi_can_dma,
655 .dma_transfer = dw_spi_dma_transfer,
656 .dma_stop = dw_spi_dma_stop,
659 void dw_spi_dma_setup_generic(struct dw_spi *dws)
661 dws->dma_ops = &dw_spi_dma_generic_ops;
663 EXPORT_SYMBOL_GPL(dw_spi_dma_setup_generic);