Qcom lava log

Accessing LAVA logs

The LAVA web UI (/scheduler/job/<id>) is behind Anubis anti-bot protection, so WebFetch / plain browser fetches return HTTP 403. Use the REST API v0.2 endpoints instead — they serve public jobs without authentication.

Default host: lava.infra.foundries.io. A URL like https://lava.infra.foundries.io/scheduler/job/256592#L3656 means job ID 256592 (the #L3656 is a UI line anchor, not part of the API).

Endpoints (no auth for public jobs)

HOST=lava.infra.foundries.io
JOB=256592

# Metadata: status, device_type, actual_device, submitter, times, and the full job definition
curl -s "https://$HOST/api/v0.2/jobs/$JOB/"

# Full logs (YAML list of JSON log records) — this is the main artifact
curl -s "https://$HOST/api/v0.2/jobs/$JOB/logs/"

# Same log content via the scheduler plain endpoint (fallback if the API path changes)
curl -s "https://$HOST/scheduler/job/$JOB/log_file/plain"

# Per-test-suite results (if needed)
curl -s "https://$HOST/api/v0.2/jobs/$JOB/tests/"

A helper is provided: ./scripts/lava.sh <job_id_or_url> [host] downloads metadata + logs to /tmp/lava_<job>.{json,logs} and prints a quick result summary. It transparently handles XZ-compressed logs (see below).

⚠️ Compressed (archived) logs

The /logs/ (and /log_file/plain) endpoint serves plain text for recent/live jobs but raw XZ-compressed bytes for older / archived jobs. This is not HTTP Content-Encoding — it's the stored .xz file served as-is, so curl does not decompress it, and your greps silently return nothing (or UnicodeDecodeError in Python). Detect it by the XZ magic at byte 0: FD 37 7A 58 5A 00 (\xFD7zXZ).

Always decompress-or-passthrough after fetching:

# robust fetch that works for BOTH plain and xz logs
curl -s "https://$HOST/api/v0.2/jobs/$JOB/logs/" -o /tmp/raw.bin
{ xz -dc /tmp/raw.bin 2>/dev/null || cat /tmp/raw.bin; } > /tmp/lava_$JOB.logs

# or streaming, in one line:
curl -s "https://$HOST/api/v0.2/jobs/$JOB/logs/" | { xz -dc 2>/dev/null || cat; } > /tmp/lava_$JOB.logs

# quick check whether a downloaded file is xz:
xxd /tmp/raw.bin | head -1   # leading "fd37 7a58 5a00" => xz

When sweeping many jobs (e.g. extracting per-job kernel versions), this matters a lot: a naive curl | grep will mark every archived job as "blank/missing" if you skip the xz -dc step.

Log format

Each log line is a YAML list item wrapping one JSON record:

- {"dt": "2026-06-10T14:11:19.154147", "lvl": "target", "msg": "..."}

• lvl values: info, debug, warning, error, results, and target.
• target = serial console output from the device under test (DUT) — this is where the kernel boot log / dmesg and test-script stdout live. Grep these first for hardware/driver root causes.
• results records carry test-case outcomes; msg is itself a JSON object with definition, case, result (pass/fail/skip), duration, commit_id, etc.
• Test scripts print [PASS]/[FAIL]/[INFO]/[WARN] lines and emit <<<LAVA_SIGNAL_TESTCASE TEST_CASE_ID=<name> RESULT=<PASS|FAIL>>>>.

Useful queries

LOGS=/tmp/lava_$JOB.logs   # downloaded via the helper or curl above

# All test-case results (pass/fail/skip) in order
grep -oE '"definition": "[0-9][^"]*", "case": "[^"]*", "result": "[^"]*"' "$LOGS"

# Find a specific failing test and its [FAIL] reason
grep -n "RESULT=FAIL\|\[FAIL\]" "$LOGS"

# Read only the DUT serial console / kernel log (lvl == target), de-noised
grep '"lvl": "target"' "$LOGS" | sed 's/.*"msg": //; s/}}\?$//'

# Pretty-print a line range (UI #L<n> anchor maps roughly to file line n)
sed -n '3636,3760p' "$LOGS" | sed 's/.*"lvl": "\([a-z]*\)", "msg": /[\1] /; s/}}\?$//'

# Kernel hardware probe failures (common root causes)
grep '"lvl": "target"' "$LOGS" | grep -iE "never came up|failed|error|cut here|WARNING|Call trace|timed out"

# Job status + device from metadata
python3 -c "import json,sys; d=json.load(open('/tmp/lava_$JOB.json')); print(d['state'],d['health'],d['requested_device_type'],d['actual_device'])"

Investigation workflow

1. Resolve the job ID from the URL (digits after /job/).
2. Fetch metadata → note device_type, actual_device, state, health, and read the job definition (lists deploy image URL and the test definitions/expected cases).
3. Fetch logs → list all test-case results to see which case failed.
4. Jump to the failing case's target lines for the [FAIL] reason and surrounding script output.
5. Correlate with the kernel boot log (lvl: target) for the underlying hardware/ driver cause (e.g. Phy link never came up, probe -EPROBE_DEFER, firmware load errors, missing DT node). A driver that prints zero kernel messages usually means it never bound to any device.
6. Check whether the triggering change (e.g. the GitHub Actions build/commit) could plausibly cause the failure, or whether it's a pre-existing board/lab/DT issue.