spi: stm32: use STM32 DMA with STM32 MDMA to enhance DDR use

The STM32 DMA doesn't have the ability to generate convenient burst
transfer on the DDR, ensuring the best load of the AXI & DDR.
To avoid this bad load of the AXI & DDR, STM32 MDMA can be used to transfer
data to the DDR, being triggered by STM32 DMA channel transfer
completion.
An SRAM buffer is used between DMA and MDMA. So the MDMA always does
MEM_TO_MEM transfers (from/to SRAM to/from DDR), and the DMA uses SRAM
instead of DDR with DEV_TO_MEM transfers.
SPI RX DMA (DEV_TO_MEM) becomes:
SPI RX FIFO ==DMA==> SRAM ==MDMA==> DDR

In RX (DEV_TO_MEM), EOT interrupt is used to pause the DMA channel (which
will raise a transfer complete) to trigger the MDMA to flush the SRAM (when
transfer length is not aligned on SRAM period).
TX remains on the former implementation.

Signed-off-by: Clément Le Goffic <[e-mail]>
Link: https://patch.msgid.link/[e-mail]
Signed-off-by: Mark Brown <[e-mail]>

drivers/spi/spi-stm32.c

···
 #include <linux/debugfs.h>
 #include <linux/clk.h>
 #include <linux/delay.h>
+#include <linux/dma-mapping.h>
 #include <linux/dmaengine.h>
+#include <linux/genalloc.h>
 #include <linux/interrupt.h>
 #include <linux/iopoll.h>
 #include <linux/module.h>
···
  * @dma_rx: dma channel for RX transfer
  * @phys_addr: SPI registers physical base address
  * @device_mode: the controller is configured as SPI device
+ * @sram_pool: SRAM pool for DMA transfers
+ * @sram_rx_buf_size: size of SRAM buffer for RX transfer
+ * @sram_rx_buf: SRAM buffer for RX transfer
+ * @sram_dma_rx_buf: SRAM buffer physical address for RX transfer
+ * @mdma_rx: MDMA channel for RX transfer
  */
 struct stm32_spi {
 	struct device *dev;
···
 	dma_addr_t phys_addr;
 
 	bool device_mode;
+
+	struct gen_pool *sram_pool;
+	size_t sram_rx_buf_size;
+	void *sram_rx_buf;
+	dma_addr_t sram_dma_rx_buf;
+	struct dma_chan *mdma_rx;
 };
 
 static const struct stm32_spi_regspec stm32fx_spi_regspec = {
···
 
 	if (spi->cur_usedma && spi->dma_tx)
 		dmaengine_terminate_async(spi->dma_tx);
-	if (spi->cur_usedma && spi->dma_rx)
+	if (spi->cur_usedma && spi->dma_rx) {
 		dmaengine_terminate_async(spi->dma_rx);
+		if (spi->mdma_rx)
+			dmaengine_terminate_async(spi->mdma_rx);
+	}
 
 	stm32_spi_clr_bits(spi, STM32H7_SPI_CR1, STM32H7_SPI_CR1_SPE);
 
···
 	}
 
 	if (sr & STM32H7_SPI_SR_EOT) {
+		dev_dbg(spi->dev, "End of transfer\n");
 		if (!spi->cur_usedma && (spi->rx_buf && (spi->rx_len > 0)))
 			stm32h7_spi_read_rxfifo(spi);
 		if (!spi->cur_usedma ||
-		    (spi->cur_comm == SPI_SIMPLEX_TX || spi->cur_comm == SPI_3WIRE_TX))
+		    (spi->cur_comm == SPI_SIMPLEX_TX || spi->cur_comm == SPI_3WIRE_TX) ||
+		    (spi->mdma_rx && (spi->cur_comm == SPI_SIMPLEX_RX ||
+		     spi->cur_comm == SPI_FULL_DUPLEX)))
 			end = true;
 	}
 
···
 	spin_unlock_irqrestore(&spi->lock, flags);
 
 	if (end) {
+		if (spi->cur_usedma && spi->mdma_rx) {
+			dmaengine_pause(spi->dma_rx);
+			/* Wait for callback */
+			return IRQ_HANDLED;
+		}
 		stm32h7_spi_disable(spi);
 		spi_finalize_current_transfer(ctrl);
 	}
···
 	/* Enable the interrupts */
 	if (spi->cur_comm == SPI_SIMPLEX_TX || spi->cur_comm == SPI_3WIRE_TX)
 		ier |= STM32H7_SPI_IER_EOTIE | STM32H7_SPI_IER_TXTFIE;
+	if (spi->mdma_rx && (spi->cur_comm == SPI_SIMPLEX_RX || spi->cur_comm == SPI_FULL_DUPLEX))
+		ier |= STM32H7_SPI_IER_EOTIE;
 
 	stm32_spi_set_bits(spi, STM32H7_SPI_IER, ier);
 
···
 
 	if (STM32_SPI_HOST_MODE(spi))
 		stm32_spi_set_bits(spi, STM32H7_SPI_CR1, STM32H7_SPI_CR1_CSTART);
+}
+
+/**
+ * stm32_spi_prepare_rx_dma_mdma_chaining - Prepare RX DMA and MDMA chaining
+ * @spi: pointer to the spi controller data structure
+ * @xfer: pointer to the spi transfer
+ * @rx_dma_conf: pointer to the DMA configuration for RX channel
+ * @rx_dma_desc: pointer to the RX DMA descriptor
+ * @rx_mdma_desc: pointer to the RX MDMA descriptor
+ *
+ * It must return 0 if the chaining is possible or an error code if not.
+ */
+static int stm32_spi_prepare_rx_dma_mdma_chaining(struct stm32_spi *spi,
+						  struct spi_transfer *xfer,
+						  struct dma_slave_config *rx_dma_conf,
+						  struct dma_async_tx_descriptor **rx_dma_desc,
+						  struct dma_async_tx_descriptor **rx_mdma_desc)
+{
+	struct dma_slave_config rx_mdma_conf = {0};
+	u32 sram_period, nents = 0, spi_s_len;
+	struct sg_table dma_sgt, mdma_sgt;
+	struct scatterlist *spi_s, *s;
+	dma_addr_t dma_buf;
+	int i, ret;
+
+	sram_period = spi->sram_rx_buf_size / 2;
+
+	/* Configure MDMA RX channel */
+	rx_mdma_conf.direction = rx_dma_conf->direction;
+	rx_mdma_conf.src_addr = spi->sram_dma_rx_buf;
+	rx_mdma_conf.peripheral_config = rx_dma_conf->peripheral_config;
+	rx_mdma_conf.peripheral_size = rx_dma_conf->peripheral_size;
+	dmaengine_slave_config(spi->mdma_rx, &rx_mdma_conf);
+
+	/* Count the number of entries needed */
+	for_each_sg(xfer->rx_sg.sgl, spi_s, xfer->rx_sg.nents, i)
+		if (sg_dma_len(spi_s) > sram_period)
+			nents += DIV_ROUND_UP(sg_dma_len(spi_s), sram_period);
+		else
+			nents++;
+
+	/* Prepare DMA slave_sg DBM transfer DEV_TO_MEM (RX>MEM=SRAM) */
+	ret = sg_alloc_table(&dma_sgt, nents, GFP_ATOMIC);
+	if (ret)
+		return ret;
+
+	spi_s = xfer->rx_sg.sgl;
+	spi_s_len = sg_dma_len(spi_s);
+	dma_buf = spi->sram_dma_rx_buf;
+	for_each_sg(dma_sgt.sgl, s, dma_sgt.nents, i) {
+		size_t bytes = min_t(size_t, spi_s_len, sram_period);
+
+		sg_dma_len(s) = bytes;
+		sg_dma_address(s) = dma_buf;
+		spi_s_len -= bytes;
+
+		if (!spi_s_len && sg_next(spi_s)) {
+			spi_s = sg_next(spi_s);
+			spi_s_len = sg_dma_len(spi_s);
+			dma_buf = spi->sram_dma_rx_buf;
+		} else { /* DMA configured in DBM: it will swap between the SRAM periods */
+			if (i & 1)
+				dma_buf += sram_period;
+			else
+				dma_buf = spi->sram_dma_rx_buf;
+		}
+	}
+
+	*rx_dma_desc = dmaengine_prep_slave_sg(spi->dma_rx, dma_sgt.sgl,
+					       dma_sgt.nents, rx_dma_conf->direction,
+					       DMA_PREP_INTERRUPT);
+	sg_free_table(&dma_sgt);
+
+	if (!rx_dma_desc)
+		return -EINVAL;
+
+	/* Prepare MDMA slave_sg transfer MEM_TO_MEM (SRAM>DDR) */
+	ret = sg_alloc_table(&mdma_sgt, nents, GFP_ATOMIC);
+	if (ret) {
+		rx_dma_desc = NULL;
+		return ret;
+	}
+
+	spi_s = xfer->rx_sg.sgl;
+	spi_s_len = sg_dma_len(spi_s);
+	dma_buf = sg_dma_address(spi_s);
+	for_each_sg(mdma_sgt.sgl, s, mdma_sgt.nents, i) {
+		size_t bytes = min_t(size_t, spi_s_len, sram_period);
+
+		sg_dma_len(s) = bytes;
+		sg_dma_address(s) = dma_buf;
+		spi_s_len -= bytes;
+
+		if (!spi_s_len && sg_next(spi_s)) {
+			spi_s = sg_next(spi_s);
+			spi_s_len = sg_dma_len(spi_s);
+			dma_buf = sg_dma_address(spi_s);
+		} else {
+			dma_buf += bytes;
+		}
+	}
+
+	*rx_mdma_desc = dmaengine_prep_slave_sg(spi->mdma_rx, mdma_sgt.sgl,
+						mdma_sgt.nents, rx_mdma_conf.direction,
+						DMA_PREP_INTERRUPT);
+	sg_free_table(&mdma_sgt);
+
+	if (!rx_mdma_desc) {
+		rx_dma_desc = NULL;
+		return -EINVAL;
+	}
+
+	return 0;
 }
 
 /**
···
 static int stm32_spi_transfer_one_dma(struct stm32_spi *spi,
 				      struct spi_transfer *xfer)
 {
+	struct dma_async_tx_descriptor *rx_mdma_desc = NULL, *rx_dma_desc = NULL;
+	struct dma_async_tx_descriptor *tx_dma_desc = NULL;
 	struct dma_slave_config tx_dma_conf, rx_dma_conf;
-	struct dma_async_tx_descriptor *tx_dma_desc, *rx_dma_desc;
 	unsigned long flags;
+	int ret = 0;
 
 	spin_lock_irqsave(&spi->lock, flags);
 
-	rx_dma_desc = NULL;
 	if (spi->rx_buf && spi->dma_rx) {
 		stm32_spi_dma_config(spi, spi->dma_rx, &rx_dma_conf, DMA_DEV_TO_MEM);
-		dmaengine_slave_config(spi->dma_rx, &rx_dma_conf);
+		if (spi->mdma_rx) {
+			rx_dma_conf.peripheral_size = 1;
+			dmaengine_slave_config(spi->dma_rx, &rx_dma_conf);
 
-		/* Enable Rx DMA request */
-		stm32_spi_set_bits(spi, spi->cfg->regs->dma_rx_en.reg,
-				   spi->cfg->regs->dma_rx_en.mask);
-
-		rx_dma_desc = dmaengine_prep_slave_sg(
-					spi->dma_rx, xfer->rx_sg.sgl,
-					xfer->rx_sg.nents,
-					rx_dma_conf.direction,
-					DMA_PREP_INTERRUPT);
+			ret = stm32_spi_prepare_rx_dma_mdma_chaining(spi, xfer, &rx_dma_conf,
+								     &rx_dma_desc, &rx_mdma_desc);
+			if (ret) { /* RX DMA MDMA chaining not possible, fallback to DMA only */
+				rx_dma_conf.peripheral_config = 0;
+				rx_dma_desc = NULL;
+			}
+		}
+		if (!rx_dma_desc) {
+			dmaengine_slave_config(spi->dma_rx, &rx_dma_conf);
+			rx_dma_desc = dmaengine_prep_slave_sg(spi->dma_rx, xfer->rx_sg.sgl,
+							      xfer->rx_sg.nents,
+							      rx_dma_conf.direction,
+							      DMA_PREP_INTERRUPT);
+		}
 	}
 
-	tx_dma_desc = NULL;
 	if (spi->tx_buf && spi->dma_tx) {
 		stm32_spi_dma_config(spi, spi->dma_tx, &tx_dma_conf, DMA_MEM_TO_DEV);
 		dmaengine_slave_config(spi->dma_tx, &tx_dma_conf);
-
-		tx_dma_desc = dmaengine_prep_slave_sg(
-					spi->dma_tx, xfer->tx_sg.sgl,
-					xfer->tx_sg.nents,
-					tx_dma_conf.direction,
-					DMA_PREP_INTERRUPT);
+		tx_dma_desc = dmaengine_prep_slave_sg(spi->dma_tx, xfer->tx_sg.sgl,
+						      xfer->tx_sg.nents,
+						      tx_dma_conf.direction,
+						      DMA_PREP_INTERRUPT);
 	}
 
 	if ((spi->tx_buf && spi->dma_tx && !tx_dma_desc) ||
···
 		goto dma_desc_error;
 
 	if (rx_dma_desc) {
-		rx_dma_desc->callback = spi->cfg->dma_rx_cb;
-		rx_dma_desc->callback_param = spi;
+		if (rx_mdma_desc) {
+			rx_mdma_desc->callback = spi->cfg->dma_rx_cb;
+			rx_mdma_desc->callback_param = spi;
+		} else {
+			rx_dma_desc->callback = spi->cfg->dma_rx_cb;
+			rx_dma_desc->callback_param = spi;
+		}
 
+		/* Enable Rx DMA request */
+		stm32_spi_set_bits(spi, spi->cfg->regs->dma_rx_en.reg,
+				   spi->cfg->regs->dma_rx_en.mask);
+		if (rx_mdma_desc) {
+			if (dma_submit_error(dmaengine_submit(rx_mdma_desc))) {
+				dev_err(spi->dev, "Rx MDMA submit failed\n");
+				goto dma_desc_error;
+			}
+			/* Enable Rx MDMA channel */
+			dma_async_issue_pending(spi->mdma_rx);
+		}
 		if (dma_submit_error(dmaengine_submit(rx_dma_desc))) {
 			dev_err(spi->dev, "Rx DMA submit failed\n");
 			goto dma_desc_error;
···
 	return 1;
 
 dma_submit_error:
+	if (spi->mdma_rx)
+		dmaengine_terminate_sync(spi->mdma_rx);
 	if (spi->dma_rx)
 		dmaengine_terminate_sync(spi->dma_rx);
 
···
 	spin_unlock_irqrestore(&spi->lock, flags);
 
 	dev_info(spi->dev, "DMA issue: fall back to irq transfer\n");
+
+	if (spi->sram_rx_buf)
+		memset(spi->sram_rx_buf, 0, spi->sram_rx_buf_size);
 
 	spi->cur_usedma = false;
 	return spi->cfg->transfer_one_irq(spi);
···
 
 	spi->cfg->disable(spi);
 
+	if (spi->sram_rx_buf)
+		memset(spi->sram_rx_buf, 0, spi->sram_rx_buf_size);
+
 	return 0;
 }
 
···
 	if (spi->dma_tx || spi->dma_rx)
 		ctrl->can_dma = stm32_spi_can_dma;
 
+	spi->sram_pool = of_gen_pool_get(pdev->dev.of_node, "sram", 0);
+	if (spi->sram_pool) {
+		spi->sram_rx_buf_size = gen_pool_size(spi->sram_pool);
+		dev_info(&pdev->dev, "SRAM pool: %zu KiB for RX DMA/MDMA chaining\n",
+			 spi->sram_rx_buf_size / 1024);
+		spi->sram_rx_buf = gen_pool_dma_zalloc(spi->sram_pool, spi->sram_rx_buf_size,
+						       &spi->sram_dma_rx_buf);
+		if (!spi->sram_rx_buf) {
+			dev_err(&pdev->dev, "failed to allocate SRAM buffer\n");
+		} else {
+			spi->mdma_rx = dma_request_chan(spi->dev, "rxm2m");
+			if (IS_ERR(spi->mdma_rx)) {
+				ret = PTR_ERR(spi->mdma_rx);
+				spi->mdma_rx = NULL;
+				if (ret == -EPROBE_DEFER) {
+					goto err_pool_free;
+				} else {
+					gen_pool_free(spi->sram_pool,
+						      (unsigned long)spi->sram_rx_buf,
+						      spi->sram_rx_buf_size);
+					dev_warn(&pdev->dev,
+						 "failed to request rx mdma channel, DMA only\n");
+				}
+			}
+		}
+	}
+
 	pm_runtime_set_autosuspend_delay(&pdev->dev,
 					 STM32_SPI_AUTOSUSPEND_DELAY);
 	pm_runtime_use_autosuspend(&pdev->dev);
···
 	pm_runtime_put_noidle(&pdev->dev);
 	pm_runtime_set_suspended(&pdev->dev);
 	pm_runtime_dont_use_autosuspend(&pdev->dev);
+
+	if (spi->mdma_rx)
+		dma_release_channel(spi->mdma_rx);
+err_pool_free:
+	gen_pool_free(spi->sram_pool, (unsigned long)spi->sram_rx_buf, spi->sram_rx_buf_size);
 err_dma_release:
 	if (spi->dma_tx)
 		dma_release_channel(spi->dma_tx);
···
 		dma_release_channel(ctrl->dma_tx);
 	if (ctrl->dma_rx)
 		dma_release_channel(ctrl->dma_rx);
+	if (spi->mdma_rx)
+		dma_release_channel(spi->mdma_rx);
+	if (spi->sram_rx_buf)
+		gen_pool_free(spi->sram_pool, (unsigned long)spi->sram_rx_buf,
+			      spi->sram_rx_buf_size);
 
 	clk_disable_unprepare(spi->clk);