Re: [PATCH v7 9/9] iio: ssp_sensors: reuse embedded RX buffer for SPI transfers

[Jonathan Cameron <jic23@kernel.org>]

On Sun, 26 Apr 2026 14:47:10 +0530
Sanjay Chitroda <sanjayembeddedse@gmail.com> wrote:

> From: Sanjay Chitroda <sanjayembeddedse@gmail.com>
> 
> Avoid allocating a temporary DMA buffer in the interrupt context when
> handling hub-to-AP and AP-to-hub SPI write messages.
> 
> Replace the dynamically allocated RX buffer with a fixed-size,
> preallocated buffer embedded in the driver structure and reused for all
> SPI receive operations. This removes memory allocation from the
> IRQ path, simplifies lifetime management.
> 
> Signed-off-by: Sanjay Chitroda <sanjayembeddedse@gmail.com>
> ---
> Changes in v6:
> - Replace dynamically allocated RX buffer with embedded fixed-size buffer
> - Fix struct layout to satisfy DMA alignment constraints comment from David
> - v5 change: https://lore.kernel.org/all/20260406080852.2727453-6-sanjayembedded@gmail.com/
> Changes in v5:
> - Rebase change on top of latest v5 patch series.
> Changes in v4:
> - Use preallocated buffer and stash a buffer that gets reused each time instead of a fresh allocation.
> - v3 change: https://lore.kernel.org/all/20260315125509.857195-3-sanjayembedded@gmail.com/
> Changes in v3:
> - prepare series to have all respective cleanup API support for the ssp_sensors following input from Andy Shevchenko
> - v2 change: https://lore.kernel.org/all/20260311174151.3441429-1-sanjayembedded@gmail.com/
> Changes in v2:
> - split series to individual patch
> - address review comment from Andy Shevchenko
> - v1 change: https://lore.kernel.org/all/20260310200513.2162018-3-sanjayembedded@gmail.com/
> ---
>  drivers/iio/common/ssp_sensors/ssp.h     |  3 +++
>  drivers/iio/common/ssp_sensors/ssp_spi.c | 20 ++------------------
>  2 files changed, 5 insertions(+), 18 deletions(-)
> 
> diff --git a/drivers/iio/common/ssp_sensors/ssp.h b/drivers/iio/common/ssp_sensors/ssp.h
> index f649cdecc277..8295bb7062a3 100644
> --- a/drivers/iio/common/ssp_sensors/ssp.h
> +++ b/drivers/iio/common/ssp_sensors/ssp.h
> @@ -174,6 +174,7 @@ struct ssp_sensorhub_info {
>   * @pending_list:	pending list for messages queued to be sent/read
>   * @sensor_devs:	registered IIO devices table
>   * @enable_refcount:	enable reference count for wdt (watchdog timer)
> + * @rx_buf:		buffer to receive SPI data
>   * @header_buffer:	cache aligned buffer for packet header
>   */
>  struct ssp_data {
> @@ -221,6 +222,8 @@ struct ssp_data {
>  	struct iio_dev *sensor_devs[SSP_SENSOR_MAX];
>  	atomic_t enable_refcount;
>  
> +	u8 rx_buf[SSP_DATA_PACKET_SIZE];

Whilst the define exists for this size, do we have anything to confirm it is right?
It wasn't previously used for anything so is a bit high risk.

Also, why doesn't it need to be DMA safe?  I'd expect it either after the following buffer
or with a __aligned(IIO_DMA_MINALIGN) marking in which case the one below isn't used.


> +
>  	__le16 header_buffer[SSP_HEADER_BUFFER_SIZE / sizeof(__le16)] __aligned(IIO_DMA_MINALIGN);
>  };
>  
> diff --git a/drivers/iio/common/ssp_sensors/ssp_spi.c b/drivers/iio/common/ssp_sensors/ssp_spi.c
> index 870214551f0b..e41da88bf96d 100644
> --- a/drivers/iio/common/ssp_sensors/ssp_spi.c
> +++ b/drivers/iio/common/ssp_sensors/ssp_spi.c
> @@ -339,7 +339,7 @@ static int ssp_parse_dataframe(struct ssp_data *data, char *dataframe, int len)
>  /* threaded irq */
>  int ssp_irq_msg(struct ssp_data *data)
>  {
> -	char *buffer;
> +	char *buffer = data->rx_buf;

Put this in inline. The local variable may reduce code churn but
it also hides the important info that this is not locally allocated.

>  	u8 msg_type;
>  	int ret;
>  	u16 length, msg_options;
> @@ -383,19 +383,12 @@ int ssp_irq_msg(struct ssp_data *data)
>  			 * but the slave should not send such ones - it is to
>  			 * check but let's handle this
>  			 */
> -			buffer = kmalloc(length, GFP_KERNEL | GFP_DMA);
> -			if (!buffer) {
> -				ret = -ENOMEM;
> -				goto _unlock;
> -			}
>  
>  			/* got dead packet so it is always an error */
>  			ret = spi_read(data->spi, buffer, length);
>  			if (ret >= 0)
>  				ret = -EPROTO;
>  
> -			kfree(buffer);
> -
>  			dev_err(SSP_DEV, "No match error %x\n",
>  				msg_options);
>  
> @@ -428,22 +421,13 @@ int ssp_irq_msg(struct ssp_data *data)
>  		mutex_unlock(&data->pending_lock);
>  		break;
>  	case SSP_HUB2AP_WRITE:
> -		buffer = kzalloc(length, GFP_KERNEL | GFP_DMA);
> -		if (!buffer)
> -			return -ENOMEM;
> -
>  		ret = spi_read(data->spi, buffer, length);
>  		if (ret < 0) {
>  			dev_err(SSP_DEV, "spi read fail\n");
> -			kfree(buffer);
>  			break;
>  		}
>  
> -		ret = ssp_parse_dataframe(data, buffer, length);
> -
> -		kfree(buffer);
> -		break;
> -
> +		return ssp_parse_dataframe(data, buffer, length);

This now makes this path different from the other switch legs.
I would precede this patch with one doing direct returns from
all the legs in here.  Then this just becomes an obvious update
in this patch.

>  	default:
>  		dev_err(SSP_DEV, "unknown msg type\n");
>  		return -EPROTO;