Flicker Investigation Continued

1. Visual evidence — the artifact is "screen shift"

Captured slow-motion phone video of the screen during a flicker event and pulled individual frames out. Frame 362 is a clean baseline. Frames 363, 370, 376, 381, 382 are flickers. The consistent signature:

• Image content is intact (no garbage colours, no noise, no black frames)
• Image is vertically displaced or split into horizontal bands
• Frame 381 is the clearest — the image appears stacked, multiple partial frames at different vertical positions within one displayed frame

To my eye this is the signature of a frame-tearing / VSYNC-alignment slip. Not a data integrity problem — a timing-alignment problem.

Reference frames (slow-motion phone capture, resized for inclusion; I have the original full-size PNGs if you want them):

2. Test-pattern bisection rules out the LVDS output stage and panel

While a burst of flicker was actively occurring on the MIPI input path, I enabled the SN65DSI83's internal LVDS test pattern (write 0x10 to CSR 0x3C, i.e. bit 4 of LVDS_FORMAT — confirmed via the known i2cset sequence). With no other change:

The internal test pattern is generated inside the SN65 at the LVDS output stage, downstream of the MIPI input and the MIPI-to-LVDS conversion. The fact that it is clean while the MIPI-fed image flickers means everything from the SN65 LVDS PLL through the panel is physically healthy. The fault must therefore lie upstream of that point — somewhere in the i.MX DSIM controller output, the MIPI bus, the SN65's MIPI receiver, or the SN65's internal MIPI-to-LVDS conversion logic.

3. On-target instrumentation I built today

To complement the visual capture I wrote a small Flask HTTP service that runs on the Hawk and polls register state at high rate, plus a host-side Python tool that drives it and records events. This is not production-quality software — it's a measurement harness, and Claude helped me write the Python.

4. Captured a burst of 93 flickers with no register-level signal

In one session, with the monitor running and an operator (me) pressing f on each visible flicker, I recorded a single burst of 93 visible flickers in ~29 s (~3 Hz, randomly spaced from roughly 100 ms to 700 ms apart). All 93 marks fell within a single ~30-second window between two scheduled video switches.

Throughout the burst, the on-target poller reported:

• SN65 PLL stayed locked: csr_0a = 0x85 the entire time
• No SN65 error bits set: csr_e5 = 0x00
• DSIM INTSRC stayed at 0x02000018 (the PllStable bit plus some status bits the kernel does not clear; no transient error bits caught)
• No DSIM configuration registers changed during the burst

So from the register-polling perspective the system looks healthy at the sampling rates I used. I have the full session log if useful.

How confident am I that nothing happened?

Less confident than the above bullet list might suggest, and I want to be honest about that. 500 ms is a long time compared to a single video frame (40 ms at 25 Hz), so the wide DSIM-register snapshot could easily miss a transient that completes within one poll interval. Specifically:

• What I can say with confidence. The SN65 PLL did not lose lock during the burst — that was polled at 10 ms, fast enough to catch the ~50 ms unlocks we see around video switches, and it caught none. The SN65 bridge did not undergo a driver-driven teardown either (those leave a clear ~500 ms imprint, and there was nothing in the log).
• What I might have missed. A DSIM configuration register flipping briefly and flipping back within 500 ms would not show up in the wide snapshot diff. Likewise DSIM_INTSRC error bits are latched but the kernel's interrupt handler clears them, typically much faster than 500 ms, so a transient error bit could be set and cleared between my polls and I would never see it. Pure hardware-level events that do not change any register I am polling (e.g. a brief PHY FIFO timing slip with no associated interrupt) are also outside what this instrumentation can detect.
• Why I did not push the wide poll faster. The memtool subprocess calls have non-trivial overhead and pushing the wide poll much below ~100 ms started loading the device CPU and impacting the I²C poll loop.

Follow-up trial: high-rate /dev/mem mmap poll

To try to see below the 500 ms ceiling I added a second poller that bypasses memtool by mmap-ing /dev/mem directly in Python and reading the DSIM registers via load instructions, aiming for ~1 ms resolution. A short captured window showed:

• DSIM_CLKCTRL (offset 0x008) is active at sub- millisecond timescales — it flips from its steady-state value of 0x02 to 0x10, 0x100 or 0x01 in 1–2 ms wide pulses, occurring roughly every 200 ms during normal operation.
• The pulses appear at a very regular cadence and are not visibly aligned with the video switches I was driving. The pattern looks consistent with routine per-frame clock-enable management by the kernel driver rather than fault-correlated events — but I am not claiming to have shown that, only that the activity I saw was regular.
• The mmap poll at that rate destabilised the device — the full unit (not just the userspace service) crashed/rebooted within seconds of starting, on multiple attempts at 1 ms target rate. I therefore could not run it long enough to mark visible flicker events and compare timestamps against the CLKCTRL transitions.

So the most accurate position on register-level diagnostics for this investigation is:

• At the resolutions I could safely sustain (500 ms wide poll, 10 ms PLL poll), nothing showed up during a confirmed flicker burst.
• At higher rate (1 ms) the registers are demonstrably busy, but the activity I captured looks like normal driver behaviour, and I could not keep the device alive long enough to test correlation with flicker timing.
• The combined evidence is consistent with — but not proof of — the fault being below register-level visibility on this rig, e.g. a sub-frame transient timing event that does not change any register I am able to read safely.

5. Reproducer without GStreamer or video

To check whether GStreamer / userspace was contributing, I stopped both the kiosk and the device-server service, leaving just the Linux text console on screen. Then ran the cycle script:

echo 4 > /sys/class/graphics/fb0/blank
echo 0 > /sys/class/graphics/fb0/blank
sleep ...
(repeat)

A random subset of the cycles produced the same vertical-shift artifact as during video playback, at roughly the same hit rate. The flicker is also visible on the U-Boot splash screen during boot, before Linux is up.

As I read this, the fault is reachable from both the U-Boot DSI bring-up path and the Linux DRM/KMS pipeline-restart path on every cycle. GStreamer is not required and not the cause.

6. dmesg signature on every pipeline restart

Every blank/unblank cycle produces this exact 3-line sequence in dmesg, deterministically across many cycles:

mxsfb 32e00000.lcdif: [drm] magic pixel crtc=0 offset 0x3e7ffc
sn65dsi83 4-002c: Using DSI clock for LVDS
@MF@ sn65dsi83_get_dsi_range dsi_rate=216000000 mode_clock=72000000 min_rate=216000000 => range=0x2b

So the SN65 bridge driver is re-initialising on every cycle (recomputing its DSI range and re-applying LVDS configuration). The visible flicker variability appears to be in how this re-init lands relative to the panel's own scan cycle — sometimes the timing aligns and the panel locks cleanly; sometimes it does not. But that is software-side interpretation and I am not confident in it.

At boot the dmesg also shows the line:

samsung-dsim 32e10000.dsi: @MF@ flb-flags=0x0

which led me to poke around in the source code, see section 7.

7. Something I noticed in the BSP source — may well be a red herring

This may turn out to be unrelated, but while poking around in the kernel patches looking for anything that might be connected to what I am seeing, I came across flb-hack-dsi-autoflush in parkeon_linux_bsp/trunk/kernel/patches-6.6/ (dated 2025-05-16, authored by Martin Fuzzey). The commit message reads:

The phrase "screen shift" jumped out because it is exactly the words I would use to describe the artifact I am seeing. From my limited reading of the patch source, it looks like it adds two device-tree flags on the Samsung DSIM driver:

The patch also adds dev_info(dev, "@MF@ flb-flags=0x%x\n", dsi->flb_flags); in samsung_dsim_parse_dt(), which I believe is where the @MF@ flb-flags=0x0 line in my boot dmesg comes from. If that mapping is right, then the patch is in the running kernel, but flb_flags == 0 at runtime, and the behavioural changes the patch adds are wrapped in two if blocks that both evaluate to false:

if (dsi->flb_flags & FLB_FLAG_AUTO_FLUSH)
    reg &= ~DSIM_MFLUSH_VS;
if (dsi->flb_flags & FLB_FLAG_NO_BURST)
    reg &= ~DSIM_BURST_MODE;

So, again from my limited understanding: it looks like this patch may have been written as a fix associated with the same kind of artifact I am investigating, and on the unit I have it is compiled in but not actually doing anything because no flb-flags property has been set in the device tree. If that interpretation is correct, the fix would be to add flb-flags = <1>; to the DSIM device-tree node (which I think lives in the flb-dt-hawk patch, but I have not opened it to confirm).

I am writing this section mainly to check on its validity — or at the very least to try to understand it. I am not making a claim that this is the fix. Specifically I do not know:

• Whether matching on the commit-message phrase "screen shift" really means the patch addresses the same issue I am seeing.
• Whether my reading of what the two bits actually do is correct.
• Whether the device tree is even the right place to enable it, or whether there is a reason it has been left off on this Hawk unit.

If you can tell me whether this is a real lead or a red herring, that alone would be very useful. And if it is a real lead, a one-line device-tree change would let me test it against the blank/unblank reproducer immediately — I would expect dmesg to then show flb-flags=0x1 and the visible flicker rate to drop sharply if the hypothesis is right.

8. What I have on my side, if any of it is useful

All of this lives on my workstation; happy to share any of it on request:

• Slow-mo phone captures of the flicker events (frames 362–382 are the labelled examples used above)
• Register-polling session logs from the on-target instrumentation, with operator flicker marks
• The on-target Flask measurement service (runs as device-server.service on the Hawk)
• The host-side CLI driver / monitor that talks to it
• The modified GStreamer kiosk that allows remote video switching
• The blank/unblank cycle reproducer script