Re: 7.0-rc4 dcn31_program_compbuf_size trace

[+a few people]

On 5/18/26 10:08, Minxi Hou wrote:
> On Mon, 16 Mar 2026 09:42:32 +0100, Julian Wollrath wrote:
>> when turning the display off via 'xset dpms force off' I observe the
>> following warning and trace under 7.0-rc4 with a AMD Ryzen 7 PRO 8840U
> 
> I can reproduce the same warning on different hardware with a different
> trigger.
> 
> Machine: HP Z66 G6 14 (ProBook 4 G1ah), AMD Ryzen 7 H 255 (HawkPoint1),
> kernel 7.0.8-200.fc44.x86_64.
> 
> Trigger: HDMI hotplug with an external 4K monitor. The Z66's HDMI port
> uses a DP-to-HDMI protocol converter with FRL support (dmesg shows
> "DP-HDMI FRL PCON supported"). Connecting the monitor at boot causes
> the first timeout during initialization; subsequent hotplug events
> repeat it at runtime.
> 
> The call trace matches yours exactly:
>   dcn31_program_compbuf_size line:141
>   dcn20_optimize_bandwidth
>   dc_commit_state_no_check
>   dc_commit_streams
>   amdgpu_dm_atomic_commit_tail
> 
> Not seen with 6.19.x. CachyOS community bisected this to:
> 
>   commit 592c5b80110d5e9e50873b5364818cb6f401e26d
>   "drm/amd/display: Migrate HUBBUB register access from hwseq to hubbub component"

Antonio posted a RFC fix for that comment here with a reference to a
warning that on a *very quick* look looks similar:
https://lore.kernel.org/all/[email protected]/

Wondering if we need a ticket here to get this resolved:
https://gitlab.freedesktop.org/drm/amd/-/work_items/

But from the reply to Antonio's patch is sounds like it's not needed.

Ciao, Thorsten

> One thing I noticed on my setup: the amdgpu driver defaults to
> max_bpc=16 for the HDMI connector. With a HDMI 2.0 monitor (BenQ
> EL2870U, Max TMDS 340 MHz) at 4K@60Hz, the higher color depth
> increases bandwidth demand during compbuf reprogramming and makes the
> timeout happen more often. Adding amdgpu.deep_color=0 helps but does
> not fix the root cause.
> 
> Partial workaround:
>   amdgpu.runpm=0 amdgpu.gfx_off=0 amdgpu.deep_color=0
> 
> A patch proposed by Giuseppe Ranieri in the CachyOS issue tracker [1]
> changes the REG_WAIT parameters from 1us*100 to 1000us*30 and resolves
> the warning for several users with DCN 3.1 hardware. Discussion there
> is ongoing about whether 2-10us*1000 would be a better fit.
> 
> [1] https://github.com/CachyOS/linux-cachyos/issues/810
> 
> Minxi Hou