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)
142 dws->rxchan = dma_request_slave_channel(dev, "rx");
146 dws->txchan = dma_request_slave_channel(dev, "tx");
148 dma_release_channel(dws->rxchan);
153 dws->master->dma_rx = dws->rxchan;
154 dws->master->dma_tx = dws->txchan;
156 init_completion(&dws->dma_completion);
158 dw_spi_dma_maxburst_init(dws);
160 dw_spi_dma_sg_burst_init(dws);
165 static void dw_spi_dma_exit(struct dw_spi *dws)
168 dmaengine_terminate_sync(dws->txchan);
169 dma_release_channel(dws->txchan);
173 dmaengine_terminate_sync(dws->rxchan);
174 dma_release_channel(dws->rxchan);
178 static irqreturn_t dw_spi_dma_transfer_handler(struct dw_spi *dws)
180 dw_spi_check_status(dws, false);
182 complete(&dws->dma_completion);
187 static bool dw_spi_can_dma(struct spi_controller *master,
188 struct spi_device *spi, struct spi_transfer *xfer)
190 struct dw_spi *dws = spi_controller_get_devdata(master);
192 return xfer->len > dws->fifo_len;
195 static enum dma_slave_buswidth dw_spi_dma_convert_width(u8 n_bytes)
198 return DMA_SLAVE_BUSWIDTH_1_BYTE;
199 else if (n_bytes == 2)
200 return DMA_SLAVE_BUSWIDTH_2_BYTES;
202 return DMA_SLAVE_BUSWIDTH_UNDEFINED;
205 static int dw_spi_dma_wait(struct dw_spi *dws, unsigned int len, u32 speed)
207 unsigned long long ms;
209 ms = len * MSEC_PER_SEC * BITS_PER_BYTE;
216 ms = wait_for_completion_timeout(&dws->dma_completion,
217 msecs_to_jiffies(ms));
220 dev_err(&dws->master->cur_msg->spi->dev,
221 "DMA transaction timed out\n");
228 static inline bool dw_spi_dma_tx_busy(struct dw_spi *dws)
230 return !(dw_readl(dws, DW_SPI_SR) & DW_SPI_SR_TF_EMPT);
233 static int dw_spi_dma_wait_tx_done(struct dw_spi *dws,
234 struct spi_transfer *xfer)
236 int retry = DW_SPI_WAIT_RETRIES;
237 struct spi_delay delay;
240 nents = dw_readl(dws, DW_SPI_TXFLR);
241 delay.unit = SPI_DELAY_UNIT_SCK;
242 delay.value = nents * dws->n_bytes * BITS_PER_BYTE;
244 while (dw_spi_dma_tx_busy(dws) && retry--)
245 spi_delay_exec(&delay, xfer);
248 dev_err(&dws->master->dev, "Tx hanged up\n");
256 * dws->dma_chan_busy is set before the dma transfer starts, callback for tx
257 * channel will clear a corresponding bit.
259 static void dw_spi_dma_tx_done(void *arg)
261 struct dw_spi *dws = arg;
263 clear_bit(DW_SPI_TX_BUSY, &dws->dma_chan_busy);
264 if (test_bit(DW_SPI_RX_BUSY, &dws->dma_chan_busy))
267 complete(&dws->dma_completion);
270 static int dw_spi_dma_config_tx(struct dw_spi *dws)
272 struct dma_slave_config txconf;
274 memset(&txconf, 0, sizeof(txconf));
275 txconf.direction = DMA_MEM_TO_DEV;
276 txconf.dst_addr = dws->dma_addr;
277 txconf.dst_maxburst = dws->txburst;
278 txconf.src_addr_width = DMA_SLAVE_BUSWIDTH_4_BYTES;
279 txconf.dst_addr_width = dw_spi_dma_convert_width(dws->n_bytes);
280 txconf.device_fc = false;
282 return dmaengine_slave_config(dws->txchan, &txconf);
285 static int dw_spi_dma_submit_tx(struct dw_spi *dws, struct scatterlist *sgl,
288 struct dma_async_tx_descriptor *txdesc;
292 txdesc = dmaengine_prep_slave_sg(dws->txchan, sgl, nents,
294 DMA_PREP_INTERRUPT | DMA_CTRL_ACK);
298 txdesc->callback = dw_spi_dma_tx_done;
299 txdesc->callback_param = dws;
301 cookie = dmaengine_submit(txdesc);
302 ret = dma_submit_error(cookie);
304 dmaengine_terminate_sync(dws->txchan);
308 set_bit(DW_SPI_TX_BUSY, &dws->dma_chan_busy);
313 static inline bool dw_spi_dma_rx_busy(struct dw_spi *dws)
315 return !!(dw_readl(dws, DW_SPI_SR) & DW_SPI_SR_RF_NOT_EMPT);
318 static int dw_spi_dma_wait_rx_done(struct dw_spi *dws)
320 int retry = DW_SPI_WAIT_RETRIES;
321 struct spi_delay delay;
322 unsigned long ns, us;
326 * It's unlikely that DMA engine is still doing the data fetching, but
327 * if it's let's give it some reasonable time. The timeout calculation
328 * is based on the synchronous APB/SSI reference clock rate, on a
329 * number of data entries left in the Rx FIFO, times a number of clock
330 * periods normally needed for a single APB read/write transaction
331 * without PREADY signal utilized (which is true for the DW APB SSI
334 nents = dw_readl(dws, DW_SPI_RXFLR);
335 ns = 4U * NSEC_PER_SEC / dws->max_freq * nents;
336 if (ns <= NSEC_PER_USEC) {
337 delay.unit = SPI_DELAY_UNIT_NSECS;
340 us = DIV_ROUND_UP(ns, NSEC_PER_USEC);
341 delay.unit = SPI_DELAY_UNIT_USECS;
342 delay.value = clamp_val(us, 0, USHRT_MAX);
345 while (dw_spi_dma_rx_busy(dws) && retry--)
346 spi_delay_exec(&delay, NULL);
349 dev_err(&dws->master->dev, "Rx hanged up\n");
357 * dws->dma_chan_busy is set before the dma transfer starts, callback for rx
358 * channel will clear a corresponding bit.
360 static void dw_spi_dma_rx_done(void *arg)
362 struct dw_spi *dws = arg;
364 clear_bit(DW_SPI_RX_BUSY, &dws->dma_chan_busy);
365 if (test_bit(DW_SPI_TX_BUSY, &dws->dma_chan_busy))
368 complete(&dws->dma_completion);
371 static int dw_spi_dma_config_rx(struct dw_spi *dws)
373 struct dma_slave_config rxconf;
375 memset(&rxconf, 0, sizeof(rxconf));
376 rxconf.direction = DMA_DEV_TO_MEM;
377 rxconf.src_addr = dws->dma_addr;
378 rxconf.src_maxburst = dws->rxburst;
379 rxconf.dst_addr_width = DMA_SLAVE_BUSWIDTH_4_BYTES;
380 rxconf.src_addr_width = dw_spi_dma_convert_width(dws->n_bytes);
381 rxconf.device_fc = false;
383 return dmaengine_slave_config(dws->rxchan, &rxconf);
386 static int dw_spi_dma_submit_rx(struct dw_spi *dws, struct scatterlist *sgl,
389 struct dma_async_tx_descriptor *rxdesc;
393 rxdesc = dmaengine_prep_slave_sg(dws->rxchan, sgl, nents,
395 DMA_PREP_INTERRUPT | DMA_CTRL_ACK);
399 rxdesc->callback = dw_spi_dma_rx_done;
400 rxdesc->callback_param = dws;
402 cookie = dmaengine_submit(rxdesc);
403 ret = dma_submit_error(cookie);
405 dmaengine_terminate_sync(dws->rxchan);
409 set_bit(DW_SPI_RX_BUSY, &dws->dma_chan_busy);
414 static int dw_spi_dma_setup(struct dw_spi *dws, struct spi_transfer *xfer)
422 /* Setup DMA channels */
423 ret = dw_spi_dma_config_tx(dws);
428 ret = dw_spi_dma_config_rx(dws);
433 /* Set the DMA handshaking interface */
434 dma_ctrl = DW_SPI_DMACR_TDMAE;
436 dma_ctrl |= DW_SPI_DMACR_RDMAE;
437 dw_writel(dws, DW_SPI_DMACR, dma_ctrl);
439 /* Set the interrupt mask */
440 imr = DW_SPI_INT_TXOI;
442 imr |= DW_SPI_INT_RXUI | DW_SPI_INT_RXOI;
443 dw_spi_umask_intr(dws, imr);
445 reinit_completion(&dws->dma_completion);
447 dws->transfer_handler = dw_spi_dma_transfer_handler;
452 static int dw_spi_dma_transfer_all(struct dw_spi *dws,
453 struct spi_transfer *xfer)
457 /* Submit the DMA Tx transfer */
458 ret = dw_spi_dma_submit_tx(dws, xfer->tx_sg.sgl, xfer->tx_sg.nents);
462 /* Submit the DMA Rx transfer if required */
464 ret = dw_spi_dma_submit_rx(dws, xfer->rx_sg.sgl,
469 /* rx must be started before tx due to spi instinct */
470 dma_async_issue_pending(dws->rxchan);
473 dma_async_issue_pending(dws->txchan);
475 ret = dw_spi_dma_wait(dws, xfer->len, xfer->effective_speed_hz);
478 dw_writel(dws, DW_SPI_DMACR, 0);
484 * In case if at least one of the requested DMA channels doesn't support the
485 * hardware accelerated SG list entries traverse, the DMA driver will most
486 * likely work that around by performing the IRQ-based SG list entries
487 * resubmission. That might and will cause a problem if the DMA Tx channel is
488 * recharged and re-executed before the Rx DMA channel. Due to
489 * non-deterministic IRQ-handler execution latency the DMA Tx channel will
490 * start pushing data to the SPI bus before the Rx DMA channel is even
491 * reinitialized with the next inbound SG list entry. By doing so the DMA Tx
492 * channel will implicitly start filling the DW APB SSI Rx FIFO up, which while
493 * the DMA Rx channel being recharged and re-executed will eventually be
496 * In order to solve the problem we have to feed the DMA engine with SG list
497 * entries one-by-one. It shall keep the DW APB SSI Tx and Rx FIFOs
498 * synchronized and prevent the Rx FIFO overflow. Since in general the tx_sg
499 * and rx_sg lists may have different number of entries of different lengths
500 * (though total length should match) let's virtually split the SG-lists to the
501 * set of DMA transfers, which length is a minimum of the ordered SG-entries
502 * lengths. An ASCII-sketch of the implemented algo is following:
505 * tx_sg list: |___|____|__|
506 * rx_sg list: |_|____|____|
507 * DMA transfers: |_|_|__|_|__|
509 * Note in order to have this workaround solving the denoted problem the DMA
510 * engine driver should properly initialize the max_sg_burst capability and set
511 * the DMA device max segment size parameter with maximum data block size the
512 * DMA engine supports.
515 static int dw_spi_dma_transfer_one(struct dw_spi *dws,
516 struct spi_transfer *xfer)
518 struct scatterlist *tx_sg = NULL, *rx_sg = NULL, tx_tmp, rx_tmp;
519 unsigned int tx_len = 0, rx_len = 0;
520 unsigned int base, len;
523 sg_init_table(&tx_tmp, 1);
524 sg_init_table(&rx_tmp, 1);
526 for (base = 0, len = 0; base < xfer->len; base += len) {
527 /* Fetch next Tx DMA data chunk */
529 tx_sg = !tx_sg ? &xfer->tx_sg.sgl[0] : sg_next(tx_sg);
530 sg_dma_address(&tx_tmp) = sg_dma_address(tx_sg);
531 tx_len = sg_dma_len(tx_sg);
534 /* Fetch next Rx DMA data chunk */
536 rx_sg = !rx_sg ? &xfer->rx_sg.sgl[0] : sg_next(rx_sg);
537 sg_dma_address(&rx_tmp) = sg_dma_address(rx_sg);
538 rx_len = sg_dma_len(rx_sg);
541 len = min(tx_len, rx_len);
543 sg_dma_len(&tx_tmp) = len;
544 sg_dma_len(&rx_tmp) = len;
546 /* Submit DMA Tx transfer */
547 ret = dw_spi_dma_submit_tx(dws, &tx_tmp, 1);
551 /* Submit DMA Rx transfer */
552 ret = dw_spi_dma_submit_rx(dws, &rx_tmp, 1);
556 /* Rx must be started before Tx due to SPI instinct */
557 dma_async_issue_pending(dws->rxchan);
559 dma_async_issue_pending(dws->txchan);
562 * Here we only need to wait for the DMA transfer to be
563 * finished since SPI controller is kept enabled during the
564 * procedure this loop implements and there is no risk to lose
565 * data left in the Tx/Rx FIFOs.
567 ret = dw_spi_dma_wait(dws, len, xfer->effective_speed_hz);
571 reinit_completion(&dws->dma_completion);
573 sg_dma_address(&tx_tmp) += len;
574 sg_dma_address(&rx_tmp) += len;
579 dw_writel(dws, DW_SPI_DMACR, 0);
584 static int dw_spi_dma_transfer(struct dw_spi *dws, struct spi_transfer *xfer)
589 nents = max(xfer->tx_sg.nents, xfer->rx_sg.nents);
592 * Execute normal DMA-based transfer (which submits the Rx and Tx SG
593 * lists directly to the DMA engine at once) if either full hardware
594 * accelerated SG list traverse is supported by both channels, or the
595 * Tx-only SPI transfer is requested, or the DMA engine is capable to
596 * handle both SG lists on hardware accelerated basis.
598 if (!dws->dma_sg_burst || !xfer->rx_buf || nents <= dws->dma_sg_burst)
599 ret = dw_spi_dma_transfer_all(dws, xfer);
601 ret = dw_spi_dma_transfer_one(dws, xfer);
605 if (dws->master->cur_msg->status == -EINPROGRESS) {
606 ret = dw_spi_dma_wait_tx_done(dws, xfer);
611 if (xfer->rx_buf && dws->master->cur_msg->status == -EINPROGRESS)
612 ret = dw_spi_dma_wait_rx_done(dws);
617 static void dw_spi_dma_stop(struct dw_spi *dws)
619 if (test_bit(DW_SPI_TX_BUSY, &dws->dma_chan_busy)) {
620 dmaengine_terminate_sync(dws->txchan);
621 clear_bit(DW_SPI_TX_BUSY, &dws->dma_chan_busy);
623 if (test_bit(DW_SPI_RX_BUSY, &dws->dma_chan_busy)) {
624 dmaengine_terminate_sync(dws->rxchan);
625 clear_bit(DW_SPI_RX_BUSY, &dws->dma_chan_busy);
629 static const struct dw_spi_dma_ops dw_spi_dma_mfld_ops = {
630 .dma_init = dw_spi_dma_init_mfld,
631 .dma_exit = dw_spi_dma_exit,
632 .dma_setup = dw_spi_dma_setup,
633 .can_dma = dw_spi_can_dma,
634 .dma_transfer = dw_spi_dma_transfer,
635 .dma_stop = dw_spi_dma_stop,
638 void dw_spi_dma_setup_mfld(struct dw_spi *dws)
640 dws->dma_ops = &dw_spi_dma_mfld_ops;
642 EXPORT_SYMBOL_NS_GPL(dw_spi_dma_setup_mfld, SPI_DW_CORE);
644 static const struct dw_spi_dma_ops dw_spi_dma_generic_ops = {
645 .dma_init = dw_spi_dma_init_generic,
646 .dma_exit = dw_spi_dma_exit,
647 .dma_setup = dw_spi_dma_setup,
648 .can_dma = dw_spi_can_dma,
649 .dma_transfer = dw_spi_dma_transfer,
650 .dma_stop = dw_spi_dma_stop,
653 void dw_spi_dma_setup_generic(struct dw_spi *dws)
655 dws->dma_ops = &dw_spi_dma_generic_ops;
657 EXPORT_SYMBOL_NS_GPL(dw_spi_dma_setup_generic, SPI_DW_CORE);