Re: [PATCH v5 0/3] drm/msm/mdss: reset all block resets described in DT

From: Qianhui Xu

Date: Mon Nov 16 2026 - 23:17:28 EST


Hi all,

3/3 says "5 of 9 cold PORs became 15 out of 15" and leaves it at that. That
is the only place the Q28 evidence has ever appeared upstream, and it does not
let anyone judge it. Here is the underlying data and, more importantly, the
boundary of what it does and does not show. No code change in this message;
take it as material for the commit message if you think it belongs there.

Board: Lenovo Y700 Gen4 (TB322FC), SM8750P, bonded dual-DSI panel
"csot,pp8807hb1-1" (driver novatek,nt36536), 1904x3040, DSC, two DSI hosts.

1. What actually fails, and how it is classified

The failure is a cold boot (real power-on reset) that brings up no image. It
is ~50% of cold PORs, it is random per boot, and it clusters in time; it is
independent of the kernel image - the same binary is good and then bad in
neighbouring boots, and seven different kernel variants were tested against
this. Three distinguishable forms:

A. cardN exists, panel init fails. Five dmesg lines, in order:
command DMA timeout: status=0x2200200
peer command DMA timeout: status=0x2200200
bonded command DMA failed: master=-110 peer=-110
sending dcs data ff 20 failed: -110
Failed to initialize panel: -5
followed by intf1/ctl 1 reset failure -22 and a vblank/frame-done storm.
B. no DRM card at all, with "DSI PLL(0) lock failed, status=0x00000000" at
about 0.86 s and a failed device link for panel@0.
C. dark panel with everything in the DRM state looking healthy: connector
connected, DPMS On, frame counter advancing, backlight non-zero, and the
DSI pclk held. This one is the reason "no DMA timeout in dmesg" must not
be read as "display works".

The only physical discriminator we found for C is the DSI interrupt counter:
41 on a lit panel, 0 on a dark one, with identical DRM state.

2. Why "reset the MDSS blocks" was picked

Because in a bad boot the failure is not inside the DSI controller and a DSI
host reset does not clear it. Registers read from a live bad boot:
STATUS0 = 0x20003 (engine and DMA busy), DSI interrupt count 0, PLL unlocked
bit not set (the PLL is locked), and CLK_STATUS sitting at 0x1743/0x3743 with
DYN_PCLK_ACTIVE never released, where a healthy boot shows 0x343 -> 0x743 ->
0x343 around each command. Repeated retries after that point fail with
identical values, which is why a single bad boot costs the whole session.

Against that, the change under review makes the display driver de-assert the
MDSS block resets described in DT before anything is programmed, which puts
the whole MDSS block into a known state before the first DCS command.

3. The measurements

Campaign of 23 real power-on resets, arms sampled alternately in the same
session, identical kernel/config/ramdisk/dtb/cmdline except for the change:

base (no reset) 9 boots: 4 good, 5 storm
base + MDSS block reset 5 boots: 5 good
base + block reset + DIGTOP 10 boots: 10 good
base + DIGTOP only 4 boots: 1 good, 3 storm

Fisher exact, two-sided: block-reset+DIGTOP vs base p = 0.011; DIGTOP alone
vs base p = 1.0 (no effect); the two arms containing the block reset merged,
15/15 vs 4/9, p = 0.003. So the block reset is the thing that does something;
the DIGTOP/PLL-lock tweak does not, on its own, at all.

Moving the same reset out of the clock driver's probe into the display driver
(the shape of this series) then measured 5/5 good cold PORs.

4. Two boundaries I do not want to paper over

- The 15/15 figure is a merged group and it predates this series' shape. It
measures "MDSS block resets are de-asserted by the display side early", not
literally patch 2/3 of v5.
- The reset set that was measured on hardware is the three-entry set (core,
core_int2, rscc). After sashiko's remark about 0xa000 v4/v5 dropped to two
entries, and the two-entry set has not been through a cold-boot sample yet,
nor has 0xa000 been varied on its own. A paired cold-boot run of two-entry
vs three-entry is queued; until it completes, 3/3 stays worded as "not
repeated with this exact set" rather than as equivalent.

5. What I would like to know

Is this evidence useful for merging the series, or would you rather see it
condensed into 2/3's commit message? And separately: given that the failure
is a cold-boot race that upstream cannot reproduce on any other board we know
of, is a quirk-style "reset everything described in DT" acceptable here on its
own merits, or does it need to be tied to a named hardware erratum to be
mergeable?

AI-assistance disclosure: this message was drafted with an LLM coding
assistant; the author verified every number and register value above against
his own logs and test records and takes responsibility for it.

Thanks,
Qianhui