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 int dw_spi_dma_caps_init(struct dw_spi *dws)
77 struct dma_slave_caps tx, rx;
80 ret = dma_get_slave_caps(dws->txchan, &tx);
84 ret = dma_get_slave_caps(dws->rxchan, &rx);
88 if (!(tx.directions & BIT(DMA_MEM_TO_DEV) &&
89 rx.directions & BIT(DMA_DEV_TO_MEM)))
92 if (tx.max_sg_burst > 0 && rx.max_sg_burst > 0)
93 dws->dma_sg_burst = min(tx.max_sg_burst, rx.max_sg_burst);
94 else if (tx.max_sg_burst > 0)
95 dws->dma_sg_burst = tx.max_sg_burst;
96 else if (rx.max_sg_burst > 0)
97 dws->dma_sg_burst = rx.max_sg_burst;
99 dws->dma_sg_burst = 0;
102 * Assuming both channels belong to the same DMA controller hence the
103 * peripheral side address width capabilities most likely would be
106 dws->dma_addr_widths = tx.dst_addr_widths & rx.src_addr_widths;
111 static int dw_spi_dma_init_mfld(struct device *dev, struct dw_spi *dws)
113 struct dw_dma_slave dma_tx = { .dst_id = 1 }, *tx = &dma_tx;
114 struct dw_dma_slave dma_rx = { .src_id = 0 }, *rx = &dma_rx;
115 struct pci_dev *dma_dev;
120 * Get pci device for DMA controller, currently it could only
121 * be the DMA controller of Medfield
123 dma_dev = pci_get_device(PCI_VENDOR_ID_INTEL, 0x0827, NULL);
128 dma_cap_set(DMA_SLAVE, mask);
130 /* 1. Init rx channel */
131 rx->dma_dev = &dma_dev->dev;
132 dws->rxchan = dma_request_channel(mask, dw_spi_dma_chan_filter, rx);
136 /* 2. Init tx channel */
137 tx->dma_dev = &dma_dev->dev;
138 dws->txchan = dma_request_channel(mask, dw_spi_dma_chan_filter, tx);
142 dws->host->dma_rx = dws->rxchan;
143 dws->host->dma_tx = dws->txchan;
145 init_completion(&dws->dma_completion);
147 ret = dw_spi_dma_caps_init(dws);
151 dw_spi_dma_maxburst_init(dws);
153 pci_dev_put(dma_dev);
158 dma_release_channel(dws->txchan);
161 dma_release_channel(dws->rxchan);
164 pci_dev_put(dma_dev);
168 static int dw_spi_dma_init_generic(struct device *dev, struct dw_spi *dws)
172 dws->rxchan = dma_request_chan(dev, "rx");
173 if (IS_ERR(dws->rxchan)) {
174 ret = PTR_ERR(dws->rxchan);
179 dws->txchan = dma_request_chan(dev, "tx");
180 if (IS_ERR(dws->txchan)) {
181 ret = PTR_ERR(dws->txchan);
186 dws->host->dma_rx = dws->rxchan;
187 dws->host->dma_tx = dws->txchan;
189 init_completion(&dws->dma_completion);
191 ret = dw_spi_dma_caps_init(dws);
195 dw_spi_dma_maxburst_init(dws);
200 dma_release_channel(dws->txchan);
203 dma_release_channel(dws->rxchan);
209 static void dw_spi_dma_exit(struct dw_spi *dws)
212 dmaengine_terminate_sync(dws->txchan);
213 dma_release_channel(dws->txchan);
217 dmaengine_terminate_sync(dws->rxchan);
218 dma_release_channel(dws->rxchan);
222 static irqreturn_t dw_spi_dma_transfer_handler(struct dw_spi *dws)
224 dw_spi_check_status(dws, false);
226 complete(&dws->dma_completion);
231 static enum dma_slave_buswidth dw_spi_dma_convert_width(u8 n_bytes)
235 return DMA_SLAVE_BUSWIDTH_1_BYTE;
237 return DMA_SLAVE_BUSWIDTH_2_BYTES;
239 return DMA_SLAVE_BUSWIDTH_4_BYTES;
241 return DMA_SLAVE_BUSWIDTH_UNDEFINED;
245 static bool dw_spi_can_dma(struct spi_controller *host,
246 struct spi_device *spi, struct spi_transfer *xfer)
248 struct dw_spi *dws = spi_controller_get_devdata(host);
249 enum dma_slave_buswidth dma_bus_width;
251 if (xfer->len <= dws->fifo_len)
254 dma_bus_width = dw_spi_dma_convert_width(dws->n_bytes);
256 return dws->dma_addr_widths & BIT(dma_bus_width);
259 static int dw_spi_dma_wait(struct dw_spi *dws, unsigned int len, u32 speed)
261 unsigned long long ms;
263 ms = len * MSEC_PER_SEC * BITS_PER_BYTE;
270 ms = wait_for_completion_timeout(&dws->dma_completion,
271 msecs_to_jiffies(ms));
274 dev_err(&dws->host->cur_msg->spi->dev,
275 "DMA transaction timed out\n");
282 static inline bool dw_spi_dma_tx_busy(struct dw_spi *dws)
284 return !(dw_readl(dws, DW_SPI_SR) & DW_SPI_SR_TF_EMPT);
287 static int dw_spi_dma_wait_tx_done(struct dw_spi *dws,
288 struct spi_transfer *xfer)
290 int retry = DW_SPI_WAIT_RETRIES;
291 struct spi_delay delay;
294 nents = dw_readl(dws, DW_SPI_TXFLR);
295 delay.unit = SPI_DELAY_UNIT_SCK;
296 delay.value = nents * dws->n_bytes * BITS_PER_BYTE;
298 while (dw_spi_dma_tx_busy(dws) && retry--)
299 spi_delay_exec(&delay, xfer);
302 dev_err(&dws->host->dev, "Tx hanged up\n");
310 * dws->dma_chan_busy is set before the dma transfer starts, callback for tx
311 * channel will clear a corresponding bit.
313 static void dw_spi_dma_tx_done(void *arg)
315 struct dw_spi *dws = arg;
317 clear_bit(DW_SPI_TX_BUSY, &dws->dma_chan_busy);
318 if (test_bit(DW_SPI_RX_BUSY, &dws->dma_chan_busy)) {
319 dw_writel(dws, DW_SPI_DMARDLR, 0);
323 complete(&dws->dma_completion);
326 static int dw_spi_dma_config_tx(struct dw_spi *dws)
328 struct dma_slave_config txconf;
330 memset(&txconf, 0, sizeof(txconf));
331 txconf.direction = DMA_MEM_TO_DEV;
332 txconf.dst_addr = dws->dma_addr;
333 txconf.dst_maxburst = dws->txburst;
334 txconf.src_addr_width = DMA_SLAVE_BUSWIDTH_4_BYTES;
335 txconf.dst_addr_width = dw_spi_dma_convert_width(dws->n_bytes);
336 txconf.device_fc = false;
338 return dmaengine_slave_config(dws->txchan, &txconf);
341 static int dw_spi_dma_submit_tx(struct dw_spi *dws, struct scatterlist *sgl,
344 struct dma_async_tx_descriptor *txdesc;
348 txdesc = dmaengine_prep_slave_sg(dws->txchan, sgl, nents,
350 DMA_PREP_INTERRUPT | DMA_CTRL_ACK);
354 txdesc->callback = dw_spi_dma_tx_done;
355 txdesc->callback_param = dws;
357 cookie = dmaengine_submit(txdesc);
358 ret = dma_submit_error(cookie);
360 dmaengine_terminate_sync(dws->txchan);
364 set_bit(DW_SPI_TX_BUSY, &dws->dma_chan_busy);
369 static inline bool dw_spi_dma_rx_busy(struct dw_spi *dws)
371 return !!(dw_readl(dws, DW_SPI_SR) & DW_SPI_SR_RF_NOT_EMPT);
374 static int dw_spi_dma_wait_rx_done(struct dw_spi *dws)
376 int retry = DW_SPI_WAIT_RETRIES;
377 struct spi_delay delay;
378 unsigned long ns, us;
382 * It's unlikely that DMA engine is still doing the data fetching, but
383 * if it's let's give it some reasonable time. The timeout calculation
384 * is based on the synchronous APB/SSI reference clock rate, on a
385 * number of data entries left in the Rx FIFO, times a number of clock
386 * periods normally needed for a single APB read/write transaction
387 * without PREADY signal utilized (which is true for the DW APB SSI
390 nents = dw_readl(dws, DW_SPI_RXFLR);
391 ns = 4U * NSEC_PER_SEC / dws->max_freq * nents;
392 if (ns <= NSEC_PER_USEC) {
393 delay.unit = SPI_DELAY_UNIT_NSECS;
396 us = DIV_ROUND_UP(ns, NSEC_PER_USEC);
397 delay.unit = SPI_DELAY_UNIT_USECS;
398 delay.value = clamp_val(us, 0, USHRT_MAX);
401 while (dw_spi_dma_rx_busy(dws) && retry--)
402 spi_delay_exec(&delay, NULL);
405 dev_err(&dws->host->dev, "Rx hanged up\n");
413 * dws->dma_chan_busy is set before the dma transfer starts, callback for rx
414 * channel will clear a corresponding bit.
416 static void dw_spi_dma_rx_done(void *arg)
418 struct dw_spi *dws = arg;
420 clear_bit(DW_SPI_RX_BUSY, &dws->dma_chan_busy);
421 if (test_bit(DW_SPI_TX_BUSY, &dws->dma_chan_busy))
424 complete(&dws->dma_completion);
427 static int dw_spi_dma_config_rx(struct dw_spi *dws)
429 struct dma_slave_config rxconf;
431 memset(&rxconf, 0, sizeof(rxconf));
432 rxconf.direction = DMA_DEV_TO_MEM;
433 rxconf.src_addr = dws->dma_addr;
434 rxconf.src_maxburst = dws->rxburst;
435 rxconf.dst_addr_width = DMA_SLAVE_BUSWIDTH_4_BYTES;
436 rxconf.src_addr_width = dw_spi_dma_convert_width(dws->n_bytes);
437 rxconf.device_fc = false;
439 return dmaengine_slave_config(dws->rxchan, &rxconf);
442 static int dw_spi_dma_submit_rx(struct dw_spi *dws, struct scatterlist *sgl,
445 struct dma_async_tx_descriptor *rxdesc;
449 rxdesc = dmaengine_prep_slave_sg(dws->rxchan, sgl, nents,
451 DMA_PREP_INTERRUPT | DMA_CTRL_ACK);
455 rxdesc->callback = dw_spi_dma_rx_done;
456 rxdesc->callback_param = dws;
458 cookie = dmaengine_submit(rxdesc);
459 ret = dma_submit_error(cookie);
461 dmaengine_terminate_sync(dws->rxchan);
465 set_bit(DW_SPI_RX_BUSY, &dws->dma_chan_busy);
470 static int dw_spi_dma_setup(struct dw_spi *dws, struct spi_transfer *xfer)
478 /* Setup DMA channels */
479 ret = dw_spi_dma_config_tx(dws);
484 ret = dw_spi_dma_config_rx(dws);
489 /* Set the DMA handshaking interface */
490 dma_ctrl = DW_SPI_DMACR_TDMAE;
492 dma_ctrl |= DW_SPI_DMACR_RDMAE;
493 dw_writel(dws, DW_SPI_DMACR, dma_ctrl);
495 /* Set the interrupt mask */
496 imr = DW_SPI_INT_TXOI;
498 imr |= DW_SPI_INT_RXUI | DW_SPI_INT_RXOI;
499 dw_spi_umask_intr(dws, imr);
501 reinit_completion(&dws->dma_completion);
503 dws->transfer_handler = dw_spi_dma_transfer_handler;
508 static int dw_spi_dma_transfer_all(struct dw_spi *dws,
509 struct spi_transfer *xfer)
513 /* Submit the DMA Tx transfer */
514 ret = dw_spi_dma_submit_tx(dws, xfer->tx_sg.sgl, xfer->tx_sg.nents);
518 /* Submit the DMA Rx transfer if required */
520 ret = dw_spi_dma_submit_rx(dws, xfer->rx_sg.sgl,
525 /* rx must be started before tx due to spi instinct */
526 dma_async_issue_pending(dws->rxchan);
529 dma_async_issue_pending(dws->txchan);
531 ret = dw_spi_dma_wait(dws, xfer->len, xfer->effective_speed_hz);
534 dw_writel(dws, DW_SPI_DMACR, 0);
540 * In case if at least one of the requested DMA channels doesn't support the
541 * hardware accelerated SG list entries traverse, the DMA driver will most
542 * likely work that around by performing the IRQ-based SG list entries
543 * resubmission. That might and will cause a problem if the DMA Tx channel is
544 * recharged and re-executed before the Rx DMA channel. Due to
545 * non-deterministic IRQ-handler execution latency the DMA Tx channel will
546 * start pushing data to the SPI bus before the Rx DMA channel is even
547 * reinitialized with the next inbound SG list entry. By doing so the DMA Tx
548 * channel will implicitly start filling the DW APB SSI Rx FIFO up, which while
549 * the DMA Rx channel being recharged and re-executed will eventually be
552 * In order to solve the problem we have to feed the DMA engine with SG list
553 * entries one-by-one. It shall keep the DW APB SSI Tx and Rx FIFOs
554 * synchronized and prevent the Rx FIFO overflow. Since in general the tx_sg
555 * and rx_sg lists may have different number of entries of different lengths
556 * (though total length should match) let's virtually split the SG-lists to the
557 * set of DMA transfers, which length is a minimum of the ordered SG-entries
558 * lengths. An ASCII-sketch of the implemented algo is following:
561 * tx_sg list: |___|____|__|
562 * rx_sg list: |_|____|____|
563 * DMA transfers: |_|_|__|_|__|
565 * Note in order to have this workaround solving the denoted problem the DMA
566 * engine driver should properly initialize the max_sg_burst capability and set
567 * the DMA device max segment size parameter with maximum data block size the
568 * DMA engine supports.
571 static int dw_spi_dma_transfer_one(struct dw_spi *dws,
572 struct spi_transfer *xfer)
574 struct scatterlist *tx_sg = NULL, *rx_sg = NULL, tx_tmp, rx_tmp;
575 unsigned int tx_len = 0, rx_len = 0;
576 unsigned int base, len;
579 sg_init_table(&tx_tmp, 1);
580 sg_init_table(&rx_tmp, 1);
582 for (base = 0, len = 0; base < xfer->len; base += len) {
583 /* Fetch next Tx DMA data chunk */
585 tx_sg = !tx_sg ? &xfer->tx_sg.sgl[0] : sg_next(tx_sg);
586 sg_dma_address(&tx_tmp) = sg_dma_address(tx_sg);
587 tx_len = sg_dma_len(tx_sg);
590 /* Fetch next Rx DMA data chunk */
592 rx_sg = !rx_sg ? &xfer->rx_sg.sgl[0] : sg_next(rx_sg);
593 sg_dma_address(&rx_tmp) = sg_dma_address(rx_sg);
594 rx_len = sg_dma_len(rx_sg);
597 len = min(tx_len, rx_len);
599 sg_dma_len(&tx_tmp) = len;
600 sg_dma_len(&rx_tmp) = len;
602 /* Submit DMA Tx transfer */
603 ret = dw_spi_dma_submit_tx(dws, &tx_tmp, 1);
607 /* Submit DMA Rx transfer */
608 ret = dw_spi_dma_submit_rx(dws, &rx_tmp, 1);
612 /* Rx must be started before Tx due to SPI instinct */
613 dma_async_issue_pending(dws->rxchan);
615 dma_async_issue_pending(dws->txchan);
618 * Here we only need to wait for the DMA transfer to be
619 * finished since SPI controller is kept enabled during the
620 * procedure this loop implements and there is no risk to lose
621 * data left in the Tx/Rx FIFOs.
623 ret = dw_spi_dma_wait(dws, len, xfer->effective_speed_hz);
627 reinit_completion(&dws->dma_completion);
629 sg_dma_address(&tx_tmp) += len;
630 sg_dma_address(&rx_tmp) += len;
635 dw_writel(dws, DW_SPI_DMACR, 0);
640 static int dw_spi_dma_transfer(struct dw_spi *dws, struct spi_transfer *xfer)
645 nents = max(xfer->tx_sg.nents, xfer->rx_sg.nents);
647 dw_writel(dws, DW_SPI_DMARDLR, xfer->tx_buf ? (dws->rxburst - 1) : 0);
650 * Execute normal DMA-based transfer (which submits the Rx and Tx SG
651 * lists directly to the DMA engine at once) if either full hardware
652 * accelerated SG list traverse is supported by both channels, or the
653 * Tx-only SPI transfer is requested, or the DMA engine is capable to
654 * handle both SG lists on hardware accelerated basis.
656 if (!dws->dma_sg_burst || !xfer->rx_buf || nents <= dws->dma_sg_burst)
657 ret = dw_spi_dma_transfer_all(dws, xfer);
659 ret = dw_spi_dma_transfer_one(dws, xfer);
663 if (dws->host->cur_msg->status == -EINPROGRESS) {
664 ret = dw_spi_dma_wait_tx_done(dws, xfer);
669 if (xfer->rx_buf && dws->host->cur_msg->status == -EINPROGRESS)
670 ret = dw_spi_dma_wait_rx_done(dws);
675 static void dw_spi_dma_stop(struct dw_spi *dws)
677 if (test_bit(DW_SPI_TX_BUSY, &dws->dma_chan_busy)) {
678 dmaengine_terminate_sync(dws->txchan);
679 clear_bit(DW_SPI_TX_BUSY, &dws->dma_chan_busy);
681 if (test_bit(DW_SPI_RX_BUSY, &dws->dma_chan_busy)) {
682 dmaengine_terminate_sync(dws->rxchan);
683 clear_bit(DW_SPI_RX_BUSY, &dws->dma_chan_busy);
687 static const struct dw_spi_dma_ops dw_spi_dma_mfld_ops = {
688 .dma_init = dw_spi_dma_init_mfld,
689 .dma_exit = dw_spi_dma_exit,
690 .dma_setup = dw_spi_dma_setup,
691 .can_dma = dw_spi_can_dma,
692 .dma_transfer = dw_spi_dma_transfer,
693 .dma_stop = dw_spi_dma_stop,
696 void dw_spi_dma_setup_mfld(struct dw_spi *dws)
698 dws->dma_ops = &dw_spi_dma_mfld_ops;
700 EXPORT_SYMBOL_NS_GPL(dw_spi_dma_setup_mfld, SPI_DW_CORE);
702 static const struct dw_spi_dma_ops dw_spi_dma_generic_ops = {
703 .dma_init = dw_spi_dma_init_generic,
704 .dma_exit = dw_spi_dma_exit,
705 .dma_setup = dw_spi_dma_setup,
706 .can_dma = dw_spi_can_dma,
707 .dma_transfer = dw_spi_dma_transfer,
708 .dma_stop = dw_spi_dma_stop,
711 void dw_spi_dma_setup_generic(struct dw_spi *dws)
713 dws->dma_ops = &dw_spi_dma_generic_ops;
715 EXPORT_SYMBOL_NS_GPL(dw_spi_dma_setup_generic, SPI_DW_CORE);