Skip to content

debug-log.md

Latest commit

 

History

History
113 lines (74 loc) · 4.72 KB

debug-log.md

File metadata and controls

113 lines (74 loc) · 4.72 KB

Debug Log

Changes made to the Pi outside of the crust binary itself. AKA: Everything that went wrong.

2026-02-20 — I2C speed

Change: /boot/firmware/config.txt

dtparam=i2c_arm=on
dtparam=i2c_arm_baudrate=400000

Raised I2C to 400kHz fast-mode for lower display latency.

2026-02-21 — Fix MPD ALSA device (no audio from MPD)

Problem: speaker-test produced audio but MPD playback was silent.

Root cause: hw:vc4hdmi,0 only accepts IEC958_SUBFRAME_LE (raw S/PDIF frames). MPD was sending standard PCM and ALSA rejected it silently. The speaker-test tool uses ALSA's plughw layer by default, which transparently converts PCM → IEC958, so it worked while MPD didn't.

Fix: /etc/mpd.conf — changed ALSA output device:

# before
device      "hw:vc4hdmi,0"

# after
device      "plughw:vc4hdmi,0"

plughw enables ALSA's plugin layer, which handles the PCM → IEC958 format conversion required by the vc4-hdmi driver.

Applied with:

sudo sed -i 's/device.*"hw:vc4hdmi,0"/device      "plughw:vc4hdmi,0"/' /etc/mpd.conf
sudo systemctl restart mpd

2026-02-21 — Bar layout tuning (X_OFFSET, BAR_STEP)

Change: src/main.rs constants.

The SSD1305 has 132 column drivers for a 128px panel; hardware columns 0–3 are off-screen on the left. A full X_OFFSET=4 pushes bars right but clips bar 15 by 3px. X_OFFSET=2 with BAR_STEP=8 (7px bar + 1px gap) distributes the deficit: bar 0 loses 2px (shows 5px, flush with panel left edge), bar 15 loses 1px (shows 6px). All 14 inner bars render at full 7px width.

Note: avoid bars=15 in CAVA — empirically, CAVA outputs 32 bytes per frame regardless (not 30), causing frame-boundary drift and scrolling. 16 bars is stable.

2026-02-21 — CAVA sensitivity tuning

Change: ~/.config/cava/config — added sensitivity = 50 (default is 100).

Halving sensitivity prevents bars from maxing out on moderately loud passages, giving more visible dynamic range.

2026-02-22 — Systemd auto-start services

Problem: After a power cycle, MPD gets "Unknown error 524" (~75s post-boot) because the vc4-hdmi ALSA driver requires the HDMI link to be fully negotiated first (~3m post-boot for the terrible HDMI extractor I started with). cava and crust had no auto-start at all.

Root cause of error 524: The vc4-hdmi audio driver is coupled to the DRM display driver. Until the HDMI extractor completes link negotiation, the PCM device cannot be opened. speaker-test worked as a workaround because it ran after the link was up.

Fix: three files deployed to /etc/systemd/system/

mpd.service.d/10-hdmi-wait.conf — drop-in that blocks MPD start until the HDMI audio device successfully accepts a 1-second silent probe (establishes the link, then exits):

[Service]
ExecStartPre=/bin/bash -c 'until aplay -D plughw:vc4hdmi,0 -f S16_LE -r 44100 -c 2 -d 1 /dev/zero -q 2>/dev/null; do sleep 5; done'
TimeoutStartSec=600

cava.service — starts CAVA after MPD, restarts on failure.

crust.service — starts crust after CAVA; ExecStartPre waits for /tmp/cava.fifo to exist as a named pipe before opening it.

Applied with:

sudo systemctl enable mpd cava crust
sudo systemctl start cava crust

Service files are version-controlled in the repo under systemd/.

2026-02-22 — Audio xrun fix (crust threading + MPD buffer)

Problem: Intermittent audio dropouts every few seconds. MPD logged: alsa_output: Decoder is too slow; playing silence to avoid xrun

Root cause: crust's main loop was strictly sequential: read_exact(cava.fifo) → render → display.flush() (I2C, ~15ms, uninterruptible D-state). While the I2C write blocked, nothing drained cava.fifo. Backpressure propagated: cava.fifo → CAVA → mpd.fifo → MPD output thread stall → xrun. Confirmed by stopping crust: zero xruns for 60s; restarting crust: xruns resumed immediately.

The ~25% wa in vmstat was caused by crust's frequent I2C D-state blocks, not SD card reads (diskstats showed near-zero block I/O during the same window).

Fix 1 — src/main.rs: moved FIFO read onto a dedicated thread so it drains continuously regardless of I2C write duration. The render loop reads the latest frame from a shared Arc<Mutex<[u8; 32]>> and calls display.flush() independently.

Fix 2 — /etc/mpd.conf: increased audio_buffer_size from default 4 MB to 32 MB as a secondary safeguard against any remaining pipeline latency spikes.

audio_buffer_size    "32768"

Applied with:

sudo sed -i 's/^filesystem_charset.*"UTF-8"/filesystem_charset\t\t"UTF-8"\naudio_buffer_size\t\t"32768"/' /etc/mpd.conf
sudo systemctl restart mpd
# deploy new crust binary (built with cargo zigbuild --release)
sudo systemctl restart crust