AM62L-EVSE-DEV-EVM: dmaengine: ti: k3-udma: 1 second polling delay

Olle Blomgren

Hi,

I run ti-linux-kernel version 6.18.15 on an AM62P5-SK-EVM. I run k3-udma and spi-omap2-mcspi. I run 20MHz SPI to an off-board STM32. I occassionally see 1 second delay in the transfer of an SPI frame. I have tracked that delay to be caused by this function:

void udma_check_tx_completion(struct work_struct *work);

in file: k3-udma-common.c.

specifically in this retry part of the code which is called if the mcspi peer does not read data to its fifo fast enough:

/* No progress, check again in 1 second */
schedule_delayed_work(&uc->tx_drain.work, HZ);

I see discussions online and my guess is that TI know about this but doesn't want to change this wait time to a shorter time or detect DMA completion in another way which these online discussions suggest.:
https://lkml.org/lkml/2022/8/22/299
https://lkml.org/lkml/2025/7/31/950

I'm in the situation where I need to address this issue with a kernel update for our products and seeks your advice.

I'd like to know:
1. Why hasn't TI acted on this?
2. Right now I'm looking into simply shorten the 1 second polling dalay to a few millisecond. Do you see any problems with that?

4 months ago

0 Bin Liu 4 months ago

TI__Guru**** 181796 points

Hi Olle,

In general, the ti-linux-kernel follows the open source development model, any upstream kernrel update (the v6.18 repo in this case) is periodically pulled in, and any patch implemented by TI development team will be sent to kernel upstream first whenever applicable, then applied to ti-linux-kernel as well.

Regarding the two patch sets in lkml.org you linked above, I see the patch 1/2 in the first link is already in ti-linux-kernel.

[PATCH 1/2] dmaengine: ti: k3-udma: Respond TX done if DMA_PREP_INTERRUPT is not requested — Linux SPI

but patch 2/2 in this set is not in, not even in the upstream kernel.

[PATCH 2/2] spi: spi-omap2-mcspi: Use EOW interrupt for completion when DMA enabled — Linux SPI

I am wondering if this is because its comments from the kernel test bot were not addressed.

Re: [PATCH 2/2] spi: spi-omap2-mcspi: Use EOW interrupt for completion when DMA enabled — Linux SPI

The patch in your second link is not in upstream kernel, so won't be ti-linux-kernel either.

However, the patch 2/2 in the first set is in ti-linux-kernel v6.1 branch, but it is not in any of newer kernel branches. I am wondering that it is dropped only because the upstream kernel didn't take this patch. I am checking with our sw dev team for any insight.

0 Olle Blomgren 4 months ago

Prodigy 40 points

Hi Bin!
Thanks for looking into this. It seems that patch 2/2 negates the peer accounting which I think has been put there for a reason. That's why I suggest only shortening the timeout here:

/* No progress, check again in 1 second */
schedule_delayed_work(&uc->tx_drain.work, HERE!!!);

and keep the rest as is. I'd appreciate if you could ask the dev team that as well. In my case I could even consider looking at DMA channel to identify McSPI use and only shorten the time for SPI traffic if the dev team sees problem with a generic solution used by all DMA users.

Thanks agan.

+1 Bin Liu 4 months ago in reply to Olle Blomgren

TI__Guru**** 181796 points

Hi Olle,

The patch 2/2 (to the spi-omap2-mcspi driver) mentioned above is actually in ti-linux-kernel 6.1.y branch and SDK9.x releases too. But it is not in any new kernel (branch 6.{6,12,18}.y). I have discussed this with our sw dev team and they decided to take this patch to ti-linux-kernel 6.18.y. Therefor I think you can take patch 2/2 into your project now to solve the mcspi latency issue.

0 Olle Blomgren 4 months ago in reply to Bin Liu

Prodigy 40 points

The links you post are dead. Where can I find the patch?

+1 Bin Liu 4 months ago in reply to Olle Blomgren

TI__Guru**** 181796 points

spi: spi-omap2-mcspi: Use EOW interrupt for completion when DMA enabled - ti-linux-kernel/ti-linux-kernel - This repo contains a Linux kernel that has been integrated with outstanding

Here is it from ti-linux-kernel 6.1.y branch.

0 Tony Tang 3 months ago in reply to Bin Liu

TI__Mastermind 30767 points

Replace with this file to SDK12.0, can't build successfully.

0 Olle Blomgren 3 months ago

Prodigy 40 points

What do you mean?

0 Tony Tang 3 months ago in reply to Olle Blomgren

TI__Mastermind 30767 points

Olle,

I would like to try this patch also, as can't apply it to SDK12.0, so replace the file directly, but build return error.

0 Olle Blomgren 3 months ago in reply to Tony Tang

Prodigy 40 points

Hi,

ok I see. It looks like TI dma subsystem has been refactored. That's why I only applied the patch to 6.12. TI has promised to take this patch into 6.18. I don't know when though.

0 Bin Liu 3 months ago in reply to Olle Blomgren

TI__Guru**** 181796 points

SDK 12.1 development cycle is already done. The patch 2/2 should be in SDK 12.2.

Processors

Processors forum

AM62L-EVSE-DEV-EVM: dmaengine: ti: k3-udma: 1 second polling delay