[PATCH 2/3] spi: nxp-xspi: Improve AHB read performance with DMA

[PATCH 2/3] spi: nxp-xspi: Improve AHB read performance with DMA

[han.xu]

From: Han Xu <han.xu@nxp.com>

Use eDMA for XSPI AHB read to improve the performance.

Read with CPU
root@imx943evk:~# dd if=/dev/mtd0 of=/dev/null bs=32M count=1
1+0 records in
1+0 records out
33554432 bytes (34 MB, 32 MiB) copied, 0.479719 s, 69.9 MB/s

Read with DMA
root@imx943evk:~# dd if=/dev/mtd0 of=/dev/null bs=32M count=1
1+0 records in
1+0 records out
33554432 bytes (34 MB, 32 MiB) copied, 0.115788 s, 290 MB/s

Signed-off-by: Han Xu <han.xu@nxp.com>
---
 drivers/spi/spi-nxp-xspi.c | 118 +++++++++++++++++++++++++++++++++++--
 1 file changed, 114 insertions(+), 4 deletions(-)

diff --git a/drivers/spi/spi-nxp-xspi.c b/drivers/spi/spi-nxp-xspi.c
index 037eac24e6fd2..0f614dc20799f 100644
--- a/drivers/spi/spi-nxp-xspi.c
+++ b/drivers/spi/spi-nxp-xspi.c
@@ -35,6 +35,8 @@
 #include <linux/clk.h>
 #include <linux/completion.h>
 #include <linux/delay.h>
+#include <linux/dmaengine.h>
+#include <linux/dma-mapping.h>
 #include <linux/err.h>
 #include <linux/errno.h>
 #include <linux/interrupt.h>
@@ -303,6 +305,8 @@
 	(opr)) << (((idx) % 2) * OPRND_SHIFT))
 
 #define NXP_XSPI_MIN_IOMAP	SZ_4M
+#define NXP_XSPI_DMA_TOUT	5000 /* ms */
+#define NXP_XSPI_DMA_ALIGN	32
 #define NXP_XSPI_MAX_CHIPSELECT		2
 #define POLL_TOUT_US		5000
 
@@ -336,6 +340,8 @@ struct nxp_xspi {
 	/* mutex lock for each operation */
 	struct mutex lock;
 	int selected;
+	struct dma_chan *dma_rx;
+	struct completion dma_rx_c;
 #define XSPI_DTR_PROTO		BIT(0)
 	int flags;
 	/* Save the previous operation clock rate */
@@ -796,6 +802,78 @@ static int nxp_xspi_ahb_read(struct nxp_xspi *xspi, const struct spi_mem_op *op)
 	return 0;
 }
 
+static int nxp_xspi_dma_init(struct device *dev, struct nxp_xspi *xspi)
+{
+	xspi->dma_rx = dma_request_chan(dev, "rx");
+	if (IS_ERR(xspi->dma_rx)) {
+		int ret = PTR_ERR(xspi->dma_rx);
+
+		xspi->dma_rx = NULL;
+		if (ret == -EPROBE_DEFER)
+			return ret;
+		dev_dbg(dev, "NO DMA RX channel, falling back to CPU read\n");
+		return 0;
+	}
+	init_completion(&xspi->dma_rx_c);
+
+	return 0;
+}
+
+static void nxp_xspi_dma_rx_callback(void *data)
+{
+	struct nxp_xspi *xspi = data;
+
+	complete(&xspi->dma_rx_c);
+}
+
+static int nxp_xspi_ahb_dma_read(struct nxp_xspi *xspi,
+				 const struct spi_mem_op *op)
+{
+	struct dma_async_tx_descriptor *desc;
+	struct dma_chan *chan = xspi->dma_rx;
+	unsigned long timeout;
+	dma_addr_t dma_addr;
+	int ret = 0;
+
+	dma_addr = dma_map_single(chan->device->dev,
+				  op->data.buf.in, op->data.nbytes,
+				  DMA_FROM_DEVICE);
+
+	if (dma_mapping_error(chan->device->dev, dma_addr)) {
+		dev_err(xspi->dev, "failed to map DMA buffer for AHB read\n");
+		return -ENOMEM;
+	}
+
+	desc = dmaengine_prep_dma_memcpy(chan, dma_addr,
+					 xspi->memmap_phy + op->addr.val,
+					 op->data.nbytes,
+					 DMA_PREP_INTERRUPT | DMA_CTRL_ACK);
+	if (!desc) {
+		dev_err(xspi->dev, "failed to prepare AHB RX DMA descriptor\n");
+		ret = -EIO;
+		goto err_unmap;
+	}
+	desc->callback = nxp_xspi_dma_rx_callback;
+	desc->callback_param = xspi;
+	reinit_completion(&xspi->dma_rx_c);
+	dmaengine_submit(desc);
+	dma_async_issue_pending(chan);
+
+	timeout = wait_for_completion_timeout(&xspi->dma_rx_c,
+					      msecs_to_jiffies(NXP_XSPI_DMA_TOUT));
+	if (!timeout) {
+		dev_err(xspi->dev, "AHB RX DMA timeout\n");
+		dmaengine_terminate_sync(chan);
+		ret = -ETIMEDOUT;
+	}
+
+err_unmap:
+	dma_unmap_single(chan->device->dev, dma_addr,
+			 op->data.nbytes, DMA_FROM_DEVICE);
+
+	return ret;
+}
+
 static int nxp_xspi_fill_txfifo(struct nxp_xspi *xspi,
 				 const struct spi_mem_op *op)
 {
@@ -1016,10 +1094,15 @@ static int nxp_xspi_exec_op(struct spi_mem *mem, const struct spi_mem_op *op)
 	 *     all use IP write.
 	 */
 	if ((op->data.dir == SPI_MEM_DATA_IN) && !needs_ip_only(xspi)
-		&& ((op->addr.val + op->data.nbytes) <= xspi->memmap_phy_size))
-		err = nxp_xspi_ahb_read(xspi, op);
-	else
+		&& ((op->addr.val + op->data.nbytes) <= xspi->memmap_phy_size)) {
+	/* use DMA for transfers no less than ahb_buf_size, when DMA is available */
+		if (xspi->dma_rx && op->data.nbytes >= xspi->devtype_data->ahb_buf_size)
+			err = nxp_xspi_ahb_dma_read(xspi, op);
+		else
+			err = nxp_xspi_ahb_read(xspi, op);
+	} else {
 		err = nxp_xspi_do_op(xspi, op);
+	}
 
 	nxp_xspi_sw_reset(xspi);
 
@@ -1042,6 +1125,24 @@ static int nxp_xspi_adjust_op_size(struct spi_mem *mem, struct spi_mem_op *op)
 		if (!needs_ip_only(xspi) && (op->addr.val < xspi->memmap_phy_size)
 			&& ((op->addr.val + op->data.nbytes) > xspi->memmap_phy_size))
 			op->data.nbytes = xspi->memmap_phy_size - op->addr.val;
+
+		/*
+		 * For AHB DMA read, align the transfer to NXP_XSPI_DMA_ALIGN
+		 * boundaries. If the start address is unaligned, shorten this
+		 * transfer so the next one starts on an aligned boundary.
+		 * Otherwise, if the length is unaligned, round it down.
+		 */
+		if (xspi->dma_rx && !needs_ip_only(xspi) &&
+		    (op->data.nbytes >= xspi->devtype_data->ahb_buf_size)) {
+			if (op->addr.val % NXP_XSPI_DMA_ALIGN)
+				op->data.nbytes =
+					ALIGN(op->addr.val, NXP_XSPI_DMA_ALIGN) -
+					op->addr.val;
+			else if (op->data.nbytes % NXP_XSPI_DMA_ALIGN)
+				op->data.nbytes =
+					ALIGN_DOWN(op->data.nbytes,
+					NXP_XSPI_DMA_ALIGN);
+		}
 	}
 
 	return 0;
@@ -1203,6 +1304,10 @@ static void nxp_xspi_cleanup(void *data)
 
 	if (xspi->ahb_addr)
 		iounmap(xspi->ahb_addr);
+	if (xspi->dma_rx) {
+		dmaengine_terminate_sync(xspi->dma_rx);
+		dma_release_channel(xspi->dma_rx);
+	}
 }
 
 static int nxp_xspi_probe(struct platform_device *pdev)
@@ -1250,7 +1355,12 @@ static int nxp_xspi_probe(struct platform_device *pdev)
 	/* Find the irq */
 	irq = platform_get_irq(pdev, 0);
 	if (irq < 0)
-		return dev_err_probe(dev, irq,  "Failed to get irq source");
+		return dev_err_probe(dev, irq, "Failed to get irq source");
+
+	/* DMA is optional, failure(other than -EPROBE_DEFER) falls back to CPU */
+	ret = nxp_xspi_dma_init(dev, xspi);
+	if (ret == -EPROBE_DEFER)
+		return ret;
 
 	pm_runtime_set_autosuspend_delay(dev, XSPI_RPM_TIMEOUT_MS);
 	pm_runtime_use_autosuspend(dev);
-- 
2.34.1

[PATCH 3/3] dt-bindings: spi: nxp,imx94-xspi: add DMA properties

[han.xu]

From: Han Xu <han.xu@nxp.com>

Add dmas and dma-names to describe TX and RX DMA channels for the i.MX94
XSPI controller.

Signed-off-by: Han Xu <han.xu@nxp.com>
---
 .../devicetree/bindings/spi/nxp,imx94-xspi.yaml        | 10 ++++++++++
 1 file changed, 10 insertions(+)

diff --git a/Documentation/devicetree/bindings/spi/nxp,imx94-xspi.yaml b/Documentation/devicetree/bindings/spi/nxp,imx94-xspi.yaml
index 16a0598c6d033..ccf841f194c06 100644
--- a/Documentation/devicetree/bindings/spi/nxp,imx94-xspi.yaml
+++ b/Documentation/devicetree/bindings/spi/nxp,imx94-xspi.yaml
@@ -30,6 +30,16 @@ properties:
       - const: base
       - const: mmap
 
+  dmas:
+    items:
+      - description: Transmit DMA
+      - description: Receive DMA
+
+  dma-names:
+    items:
+      - const: tx
+      - const: rx
+
   interrupts:
     items:
       - description: interrupt for EENV0
-- 
2.34.1

Re: [PATCH 2/3] spi: nxp-xspi: Improve AHB read performance with DMA

[Mark Brown]

[-- Attachment #1: Type: text/plain, Size: 725 bytes --]

On Wed, Jun 17, 2026 at 04:55:18PM -0500, han.xu@oss.nxp.com wrote:
> From: Han Xu <han.xu@nxp.com>
> 
> Use eDMA for XSPI AHB read to improve the performance.

You've not copied me on the rest of the series so I don't know what's
going on with dependencies.  When sending a patch series it is important
to ensure that all the various maintainers understand what the
relationship between the patches as the expecation is that there will be
interdependencies.  Either copy everyone on the whole series or at least
copy them on the cover letter and explain what's going on.  If there are
no strong interdependencies then it's generally simplest to just send
the patches separately to avoid any possible confusion.

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 488 bytes --]

Re: [PATCH 3/3] dt-bindings: spi: nxp,imx94-xspi: add DMA properties

[Frank Li]

On Wed, Jun 17, 2026 at 04:55:19PM -0500, han.xu@oss.nxp.com wrote:
> From: Han Xu <han.xu@nxp.com>
>
> Add dmas and dma-names to describe TX and RX DMA channels for the i.MX94
> XSPI controller.
>
> Signed-off-by: Han Xu <han.xu@nxp.com>
> ---

This patch should be fix one.

Reviewed-by: Frank Li <Frank.Li@nxp.com>

>  .../devicetree/bindings/spi/nxp,imx94-xspi.yaml        | 10 ++++++++++
>  1 file changed, 10 insertions(+)
>
> diff --git a/Documentation/devicetree/bindings/spi/nxp,imx94-xspi.yaml b/Documentation/devicetree/bindings/spi/nxp,imx94-xspi.yaml
> index 16a0598c6d033..ccf841f194c06 100644
> --- a/Documentation/devicetree/bindings/spi/nxp,imx94-xspi.yaml
> +++ b/Documentation/devicetree/bindings/spi/nxp,imx94-xspi.yaml
> @@ -30,6 +30,16 @@ properties:
>        - const: base
>        - const: mmap
>
> +  dmas:
> +    items:
> +      - description: Transmit DMA
> +      - description: Receive DMA
> +
> +  dma-names:
> +    items:
> +      - const: tx
> +      - const: rx
> +
>    interrupts:
>      items:
>        - description: interrupt for EENV0
> --
> 2.34.1
>
>

Re: [PATCH 3/3] dt-bindings: spi: nxp,imx94-xspi: add DMA properties

[Frank Li]

On Wed, Jun 17, 2026 at 04:55:19PM -0500, han.xu@oss.nxp.com wrote:
>

tags is reversed in spi.

spi: dt-bindings: imx94-xspi ....

you can use git log --oneline Document/devicetree/bindings/spi

Frank

> Add dmas and dma-names to describe TX and RX DMA channels for the i.MX94
> XSPI controller.
>
> Signed-off-by: Han Xu <han.xu@nxp.com>
> ---
>  .../devicetree/bindings/spi/nxp,imx94-xspi.yaml        | 10 ++++++++++
>  1 file changed, 10 insertions(+)
>
> diff --git a/Documentation/devicetree/bindings/spi/nxp,imx94-xspi.yaml b/Documentation/devicetree/bindings/spi/nxp,imx94-xspi.yaml
> index 16a0598c6d033..ccf841f194c06 100644
> --- a/Documentation/devicetree/bindings/spi/nxp,imx94-xspi.yaml
> +++ b/Documentation/devicetree/bindings/spi/nxp,imx94-xspi.yaml
> @@ -30,6 +30,16 @@ properties:
>        - const: base
>        - const: mmap
>
> +  dmas:
> +    items:
> +      - description: Transmit DMA
> +      - description: Receive DMA
> +
> +  dma-names:
> +    items:
> +      - const: tx
> +      - const: rx
> +
>    interrupts:
>      items:
>        - description: interrupt for EENV0
> --
> 2.34.1
>
>

Re: [PATCH 2/3] spi: nxp-xspi: Improve AHB read performance with DMA

[Frank Li]

On Wed, Jun 17, 2026 at 04:55:18PM -0500, han.xu@oss.nxp.com wrote:
> Use eDMA for XSPI AHB read to improve the performance.

It'd better mention 4x performance improve here.

>
> Read with CPU
> root@imx943evk:~# dd if=/dev/mtd0 of=/dev/null bs=32M count=1
> 1+0 records in
> 1+0 records out
> 33554432 bytes (34 MB, 32 MiB) copied, 0.479719 s, 69.9 MB/s
>
> Read with DMA
> root@imx943evk:~# dd if=/dev/mtd0 of=/dev/null bs=32M count=1
> 1+0 records in
> 1+0 records out
> 33554432 bytes (34 MB, 32 MiB) copied, 0.115788 s, 290 MB/s
>
> Signed-off-by: Han Xu <han.xu@nxp.com>
> ---
>  drivers/spi/spi-nxp-xspi.c | 118 +++++++++++++++++++++++++++++++++++--
>  1 file changed, 114 insertions(+), 4 deletions(-)
>
> diff --git a/drivers/spi/spi-nxp-xspi.c b/drivers/spi/spi-nxp-xspi.c
> index 037eac24e6fd2..0f614dc20799f 100644
> --- a/drivers/spi/spi-nxp-xspi.c
> +++ b/drivers/spi/spi-nxp-xspi.c
> @@ -35,6 +35,8 @@
>  #include <linux/clk.h>
>  #include <linux/completion.h>
>  #include <linux/delay.h>
> +#include <linux/dmaengine.h>
> +#include <linux/dma-mapping.h>
>  #include <linux/err.h>
>  #include <linux/errno.h>
>  #include <linux/interrupt.h>
> @@ -303,6 +305,8 @@
>         (opr)) << (((idx) % 2) * OPRND_SHIFT))
>
>  #define NXP_XSPI_MIN_IOMAP     SZ_4M
> +#define NXP_XSPI_DMA_TOUT      5000 /* ms */

Need unit, NXP_XPI_DMA_TOUT_MS,

Add comments why choose 5ms here, in case someone need change it in future.

> +#define NXP_XSPI_DMA_ALIGN     32
>  #define NXP_XSPI_MAX_CHIPSELECT                2
>  #define POLL_TOUT_US           5000
>
> @@ -336,6 +340,8 @@ struct nxp_xspi {
>         /* mutex lock for each operation */
>         struct mutex lock;
>         int selected;
> +       struct dma_chan *dma_rx;
> +       struct completion dma_rx_c;
>  #define XSPI_DTR_PROTO         BIT(0)
>         int flags;
>         /* Save the previous operation clock rate */
> @@ -796,6 +802,78 @@ static int nxp_xspi_ahb_read(struct nxp_xspi *xspi, const struct spi_mem_op *op)
>         return 0;
>  }
>
> +static int nxp_xspi_dma_init(struct device *dev, struct nxp_xspi *xspi)
> +{
> +       xspi->dma_rx = dma_request_chan(dev, "rx");
> +       if (IS_ERR(xspi->dma_rx)) {
> +               int ret = PTR_ERR(xspi->dma_rx);
> +
> +               xspi->dma_rx = NULL;
> +               if (ret == -EPROBE_DEFER)
> +                       return ret;
> +               dev_dbg(dev, "NO DMA RX channel, falling back to CPU read\n");
> +               return 0;
> +       }

use local variable will a little bit simpile

	dma_chan *rx = dma_request_chan()
	if (IS_ERR(rx) ){
		if (PTR_ERR(rx) == -EPROBE_DEFER)
			return ret;

		return 0;
	}

	xspi->dma_rx = rx;

> +       init_completion(&xspi->dma_rx_c);
> +
> +       return 0;
> +}
> +
> +static void nxp_xspi_dma_rx_callback(void *data)
> +{
> +       struct nxp_xspi *xspi = data;
> +
> +       complete(&xspi->dma_rx_c);
> +}
> +
> +static int nxp_xspi_ahb_dma_read(struct nxp_xspi *xspi,
> +                                const struct spi_mem_op *op)
> +{
> +       struct dma_async_tx_descriptor *desc;
> +       struct dma_chan *chan = xspi->dma_rx;
> +       unsigned long timeout;
> +       dma_addr_t dma_addr;
> +       int ret = 0;
> +
> +       dma_addr = dma_map_single(chan->device->dev,
> +                                 op->data.buf.in, op->data.nbytes,
> +                                 DMA_FROM_DEVICE);
> +
> +       if (dma_mapping_error(chan->device->dev, dma_addr)) {
> +               dev_err(xspi->dev, "failed to map DMA buffer for AHB read\n");
> +               return -ENOMEM;
> +       }
> +
> +       desc = dmaengine_prep_dma_memcpy(chan, dma_addr,
> +                                        xspi->memmap_phy + op->addr.val,
> +                                        op->data.nbytes,
> +                                        DMA_PREP_INTERRUPT | DMA_CTRL_ACK);
> +       if (!desc) {
> +               dev_err(xspi->dev, "failed to prepare AHB RX DMA descriptor\n");
> +               ret = -EIO;
> +               goto err_unmap;
> +       }
> +       desc->callback = nxp_xspi_dma_rx_callback;
> +       desc->callback_param = xspi;
> +       reinit_completion(&xspi->dma_rx_c);
> +       dmaengine_submit(desc);
> +       dma_async_issue_pending(chan);
> +
> +       timeout = wait_for_completion_timeout(&xspi->dma_rx_c,
> +                                             msecs_to_jiffies(NXP_XSPI_DMA_TOUT));
> +       if (!timeout) {
> +               dev_err(xspi->dev, "AHB RX DMA timeout\n");
> +               dmaengine_terminate_sync(chan);
> +               ret = -ETIMEDOUT;
> +       }
> +
> +err_unmap:
> +       dma_unmap_single(chan->device->dev, dma_addr,
> +                        op->data.nbytes, DMA_FROM_DEVICE);
> +
> +       return ret;
> +}
> +
>  static int nxp_xspi_fill_txfifo(struct nxp_xspi *xspi,
>                                  const struct spi_mem_op *op)
>  {
> @@ -1016,10 +1094,15 @@ static int nxp_xspi_exec_op(struct spi_mem *mem, const struct spi_mem_op *op)
>          *     all use IP write.
>          */
>         if ((op->data.dir == SPI_MEM_DATA_IN) && !needs_ip_only(xspi)
> -               && ((op->addr.val + op->data.nbytes) <= xspi->memmap_phy_size))
> -               err = nxp_xspi_ahb_read(xspi, op);
> -       else
> +               && ((op->addr.val + op->data.nbytes) <= xspi->memmap_phy_size)) {
> +       /* use DMA for transfers no less than ahb_buf_size, when DMA is available */
> +               if (xspi->dma_rx && op->data.nbytes >= xspi->devtype_data->ahb_buf_size)
> +                       err = nxp_xspi_ahb_dma_read(xspi, op);

If err happen, do we need failback to use CPU to read()?

> +               else
> +                       err = nxp_xspi_ahb_read(xspi, op);
> +       } else {
>                 err = nxp_xspi_do_op(xspi, op);
> +       }
>
>         nxp_xspi_sw_reset(xspi);
>
> @@ -1042,6 +1125,24 @@ static int nxp_xspi_adjust_op_size(struct spi_mem *mem, struct spi_mem_op *op)
>                 if (!needs_ip_only(xspi) && (op->addr.val < xspi->memmap_phy_size)
>                         && ((op->addr.val + op->data.nbytes) > xspi->memmap_phy_size))
>                         op->data.nbytes = xspi->memmap_phy_size - op->addr.val;
> +
> +               /*
> +                * For AHB DMA read, align the transfer to NXP_XSPI_DMA_ALIGN
> +                * boundaries. If the start address is unaligned, shorten this
> +                * transfer so the next one starts on an aligned boundary.
> +                * Otherwise, if the length is unaligned, round it down.
> +                */

DMA_ALGIN restriction come from DMA or XSPI? If come from DMA, need use
DMA API to get such align requirement.

> +               if (xspi->dma_rx && !needs_ip_only(xspi) &&
> +                   (op->data.nbytes >= xspi->devtype_data->ahb_buf_size)) {
> +                       if (op->addr.val % NXP_XSPI_DMA_ALIGN)
> +                               op->data.nbytes =
> +                                       ALIGN(op->addr.val, NXP_XSPI_DMA_ALIGN) -
> +                                       op->addr.val;
> +                       else if (op->data.nbytes % NXP_XSPI_DMA_ALIGN)
> +                               op->data.nbytes =
> +                                       ALIGN_DOWN(op->data.nbytes,
> +                                       NXP_XSPI_DMA_ALIGN);
> +               }
>         }
>
>         return 0;
> @@ -1203,6 +1304,10 @@ static void nxp_xspi_cleanup(void *data)
>
>         if (xspi->ahb_addr)
>                 iounmap(xspi->ahb_addr);
> +       if (xspi->dma_rx) {
> +               dmaengine_terminate_sync(xspi->dma_rx);
> +               dma_release_channel(xspi->dma_rx);
> +       }
>  }
>
>  static int nxp_xspi_probe(struct platform_device *pdev)
> @@ -1250,7 +1355,12 @@ static int nxp_xspi_probe(struct platform_device *pdev)
>         /* Find the irq */
>         irq = platform_get_irq(pdev, 0);
>         if (irq < 0)
> -               return dev_err_probe(dev, irq,  "Failed to get irq source");
> +               return dev_err_probe(dev, irq, "Failed to get irq source");

This change is not related this patch.

Frank
> +
> +       /* DMA is optional, failure(other than -EPROBE_DEFER) falls back to CPU */
> +       ret = nxp_xspi_dma_init(dev, xspi);
> +       if (ret == -EPROBE_DEFER)
> +               return ret;
>
>         pm_runtime_set_autosuspend_delay(dev, XSPI_RPM_TIMEOUT_MS);
>         pm_runtime_use_autosuspend(dev);
> --
> 2.34.1
>
>