[PATCH] dma: dw-axi-dmac: ensure segment lengths obey reg_width limits

[hverkuil]

Describe the bug

While experimenting with the rp1-pio driver of the Raspberry Pi 5 I got 'length X not aligned to device width 4" messages, preventing the DMA to work.

Debugging showed that the dw_axi_dma_chan_prep_slave_sg() did not take the reg_width into account.

Patch fixing this is attached.

0001-dma-dw-axi-dmac-ensure-segment-lengths-obey-reg_widt.patch

Steps to reproduce the behaviour

Test utility is found here: https://git.linuxtv.org/hverkuil/hdmi-dbg-ctl.git/tree/utils/hdmi-pio?h=hdmi-dbg-ctl-v2

gcc -o hdmi-pio hdmi-pio.c
./hdmi-pio 20 2 10 # sample GPIO 20+21 at 10 MHz

Device (s)

Raspberry Pi 5

System

$ cat /etc/rpi-issue
Raspberry Pi reference 2024-11-19
Generated using pi-gen, https://github.com/RPi-Distro/pi-gen, 891df1e21ed2b6099a2e6a13e26c91dea44b34d4, stage4
$ vcgencmd version
2026/06/17 00:06:53
Copyright (c) 2012 Broadcom
version 6219bf00 (release) (embedded)
$ uname -a
Linux rpi5 6.18.42-rpi5 #20 SMP PREEMPT Sun Aug 9 10:17:09 CEST 2026 aarch64 GNU/Linux

Logs

No response

Additional context

No response

[pelwell]

You are going to have to convince me that there is a correct way to handle lengths which are not a multiple of the register width in a generic way in a DMA driver.

[hverkuil]

It's this code in the axi driver:

   for_each_sg(sgl, sg, sg_len, i) {
            mem = sg_dma_address(sg);
            len = sg_dma_len(sg);
            num_segments = DIV_ROUND_UP(sg_dma_len(sg), axi_block_len);
            if (!num_segments)
                    continue;

            segment_len = DIV_ROUND_UP(sg_dma_len(sg), num_segments);

len = 2500000 (multiple of 4), axi_block_len = 1048576 (multiple of 4), num_segments = 3, and with those numbers segment_len is 833334, which is NOT a multiple of 4.

This code is broken in several ways: 833334 is not a multiple of 4 (causing the error), and even if it was, then it would create 2 hw_desc entries, but skip the third because 2500000 - 2 * 833334 = 833332, so it is skipped by the 'while (len >= segment_len).

With this patch it creates two dma_desc entries with size 833336 and one with 833328, all dividable by 4.

Obviously, sg_dma_len(sg) must be a multiple of 4, but the problem here is how it is split up in smaller chunks, each <= axi_block_len.

[pelwell]

Oh, I see. I think the motivation behind the original code is faulty - what benefit is there in dividing the block evenly between the segments? All it does is delay more bytes than is necessary. I think all segments should be as full as possible, even if that means the last one contains a single word.

[pelwell]

#7547 is an alternative fix. You can install a trial kernel build using sudo rpi-update pulls/7547, after reading the caveats.