Re: Issues with i.MX SPI DMA transfers

From: Igor Plyatov
Date: Wed Apr 03 2019 - 11:51:35 EST


This is a multi-part message in MIME format. Dear Robin,


Please apply the attached patch which is based on linux-next commit 05d08e2995cbe6efdb993482ee0d38a77040861a.
Please notice it has already contained two sdma patches revert - "ad0d92d7ba6a" and "25aaa75df1e6"

1) Yours source code is same as mine with exception of SDMA description for eCSPI in Device Tree.

I have changed Device Tree for i.MX6Q/DL as in attached patch and finally SPI interfaces operate more or less.

My patch revert back patches df07101e1c4a29e820df02f9989a066988b160e6 and dd4b487b32a3571fdcc66062e661e3a3e360e35b. It is strange, because they are merged into mainline while ago. Maybe they are valid for some specific variant of i.MX SOC?

"More or less" means I have come to state described in first e-mail of this e-mail thread. Byte duplication (ERR009165) happens very often for eCSPI5 interface operating through DMA.

Test Conditions:
1.1. Interface under test is eCSPI1 or eCSPI5;
1.2. Four exemplars of "burn-neon" (CPU burn) executes in background to have 100% load for all 4 CPU cores (source code taken from https://raw.githubusercontent.com/ssvb/cpuburn-arm/dd5c5ba58d2b0b23cfab4a286f9d3f5510000f20/cpuburn-a8.S and https://hardwarebug.org/files/burn.S);
1.3. PC has running "ping -f embedded_device" to have network activity around 1 MiB/s Rx and Tx;
1.4. One exemplar of "spidev_test -D /dev/spidevX.0 -s FREQUENCY -b 8 -S 512 -I 1000000 -l" executes at different frequencies;

Test Results for eCSPI1:
21 MHzÂÂÂ - success;
20 MHzÂÂÂ - success;
...
16 MHzÂÂÂ - success;
...
Â8 MHzÂÂÂ - success;

Test Results for eCSPI5:
21 MHzÂÂÂ - failed;
20 MHzÂÂÂ - failed;
19 MHzÂÂÂ - failed;
...
12 MHzÂÂÂ - failed;
11 MHzÂÂÂ - failed;
10 MHzÂÂÂ - failed;
Â9 MHzÂÂÂ - failed;
Â8 MHzÂÂÂ - failed;
Â7 MHzÂÂÂ - failed;
Â6 MHzÂÂÂ - failed;
Â5 MHzÂÂÂ - failed;
Â4 MHzÂÂÂ - success.

Maybe it is worth to dump content of registers for eCSPI1 and eCSPI5 to compare them? I can do this if you will describe how to make it.

Maybe I'm wrong, but I suppose incorrect clock source for eCSPI5. Looks like it is worth to check correctness of driver "clk-imx6q.c" or something else responsible for eCSPI5 clock.


2) I want to improve description and replace magic numbers by constants in Device Tree for SDMA. I mean strings like "dmas = <&sdma 11 7 1>, <&sdma 12 7 2>;"?

So, finally Device Tree will have strings like

dmas = <&sdma SOME_DEF_A IMX_DMATYPE_.. DMA_PRIO_..>, <&sdma SOME_DEF_B IMX_DMATYPE_.. DMA_PRIO_..>;

I think, this will stop black magic manipulations for SDMA in Device Tree, by various developers.

Does first digit means "DMA request/event ID"? Where to find more info about this?

Does second digit means "enum sdma_peripheral_type"?

Does third digit means "enum imx_dma_prio"?

Where can I find description of difference between IMX_DMATYPE_CSPI (MCU domain CSPI) and IMX_DMATYPE_CSPI_SP (Shared CSPI)? Googling does not help too much.

Best wishes.
--
Igor Plyatov