Using DMA with native CS for a SPI peripheral

[mehul_dabhi]

Hi everyone,

I had faced an issue with the delay between CS assertion and SCLK trigger for SPI for which I received a suggestion to use native CS with DMA disabled. This the link of the entire discussion.

The problem with disabling the DMA is that the CPU will be burdened a lot, so I am still looking for the solutions of below two queries from my previous post :

1. How does enabling DMA causes the native CS pin to go high after each byte transfer?
2. The project that I am working on also has ADC interfaced to SPI (not the same bus as DAC). So communicating continuously to both ADC and DAC for high sample rates will burden the CPU a lot. For this reason I was thinking of forcing DMA for every transaction over SPI. With your current solution I will not be able to do that. So, is there any way to get this issue resolved with DMA enabled? If not, what can I do to lower the burden from CPU?

Thanks,
Mehul

IMX8MPLUS 

[Manuel_Salas]

Hello @mehul_dabhi 

I hope you are doing very well.

Could you please share steps to replicate the issue by my side?

Also, are you using Cortex M or Cortex A?

SPI as Master or Slave? (I guess is master).

Also, please share your Device tree related to SPI.

A captured Image of signals would be helpful.

Best regards,

Salas.

[mehul_dabhi]

Hi @Manuel_Salas,

Apologies for my delayed response.

Could you please share steps to replicate the issue by my side?

>> 1. Configure the SPI node as per shared snap of dts.

    2. Disable dma from spi_imx.c as below snap

        mehul_dabhi_0-1738142973196.png

     3. Transmit some data over SPI.

Also, are you using Cortex M or Cortex A?

>> I am using Cortex A.

SPI as Master or Slave? (I guess is master).

>> Yes, the SPI is configured as Master.

Also, please share your Device tree related to SPI.

>>

       mehul_dabhi_1-1738143415671.png

     mehul_dabhi_2-1738143522758.png

A captured Image of signals would be helpful.
>> When I switched to native CS I had observed below waveform when I was trying to transmit 24bits with 8 bits-per-word (Yellow - CS & Blue - SCLK). So the CS was going high after each byte transfer.

mehul_dabhi_3-1738143671852.png

To resolve the above issue I was suggested to disable DMA as I mentioned in above steps and got it resolved (below snap yellow - clock and blue - clock).

Screenshot from 2025-02-04 12-49-35.png

If anything is yet required to recreate the issue, I am happy to help with that.

Thanks,
Mehul

[Manuel_Salas]

Hello @mehul_dabhi 

Actually, this behavior is expected. When you are using the SPI as Master, the CS signal should be an IO pad, then the driver will handle when CS should be asserted.

When you are using the SPI as Slave, you should select the CS signal as the dedicated CS.

You can see this examples on device tree (i.MX93 just for reference):

imx93-11x11-evk-lpspi-slave.dts

and 

imx93-11x11-evk-lpspi.dts

Please try changing the native ECSPI2_SS0 for any GPIO without disable DMA.

Best regards,

Salas.

[mehul_dabhi]

Hi @Manuel_Salas,

Thanks for your prompt reply.

Please try changing the native ECSPI2_SS0 for any GPIO without disable DMA.
>> Initially this is what I had configured and had faced an issue of the delay between CS assertion and first rising edge of SCLK and then switched to dedicated CS.

The original issue link is here. Also below is the snap for your reference.

Screenshot from 2025-02-05 12-35-33.png

So if I use the GPIO pad for CS won't I face this issue again and go back in the same loop?

Thanks,
Mehul 

[Manuel_Salas]

Hello @mehul_dabhi 

It appears that delay is from the driver.

Best regards,

Salas.

[mehul_dabhi]

Hello @Manuel_Salas,

I tried tweaking the driver source and I could only find hard delay being used in the section shown in the below snap.
Screenshot from 2025-02-06 12-27-37.png

Here, the delay to be generated was found to be 0. Apart from this, is there any hard delays that might affect? If I intend to reduce these delays, where would you suggest I make changes to improve the throughput?

Thanks,

Mehul

[mehul_dabhi]

Hi @Manuel_Salas ,

I would really appreciate some insights or suggestions on my recent querry.

Thanks,

Mehul

[mehul_dabhi]

Hi @Manuel_Salas 

We are facing a roadblock in development due to an issue with SPI.

Without DMA, the CS delay is minimal, but CPU utilisation becomes extremely high. Using DMA helps offload the CPU, but we are encountering significant CS delays as described in above issue, making it infeasible.

We are currently stuck and would appreciate any insights on reducing these delays while using DMA. If optimisation isn't possible, could you shed some light on the underlying reasons? Understanding this will help us reconsider our hardware choices.

Looking forward to your guidance.

Thanks,
Mehul