discovery: identify SDR by VID:PID with sysfs fallback; document three observer failure modes

This commit is contained in:
Lars Nolden
2026-09-01 10:07:50 +02:00
parent 3a2e23e204
commit 773343c178
2 changed files with 61 additions and 12 deletions
+31 -5
View File
@@ -73,8 +73,34 @@ Nooelec NESDR SMArt v5
RTL2832U + R820T, USB ID 0bda:2838, serial 33999948
```
Quirk (EXPERIMENTALLY VERIFIED): the R820T PLL can wedge ("PLL not
locked"), after which captures contain gain-independent white noise on any
frequency. A USB reset (`usbreset 0bda:2838`) recovers it. The HIL runner
resets the dongle before each session. Suspected trigger: `rtl_test -t`'s
direct-sampling probe. Avoid `rtl_test -t` in automation.
### Observer reliability (EXPERIMENTALLY VERIFIED)
This dongle is not a stable instrument and every measurement must record
its state. Three distinct failure modes were observed and characterised:
1. **Sensitivity drift.** Within minutes of continuous streaming the tuner
loses roughly 17 dB. Measured with `hil/sdr_health.py`: fresh after
replug ~ +22 dB mean power at 42.1 dB gain and ~51 dB spectrum span;
after a few minutes ~ +5 dB and ~19 dB span. Captures at "the same
gain" taken minutes apart are therefore NOT comparable.
2. **Tuner death mid-stream.** `r82xx_write: i2c wr failed` appears and the
receiver goes deaf, but `rtl_sdr` does NOT exit -- it keeps producing
nothing. A naive long capture silently returns "no traffic". Observed at
~192 s into a 1800 s capture. Mitigation: bounded capture segments with
`-n` plus `usbreset` between them (`hil/capture_both.sh`).
3. **Descriptor / enumeration faults.** The device can enumerate with empty
descriptor strings (`rtl_test` prints `0: , , SN: `) or disappear from
the bus entirely, in which case `usbreset` reports "No such device
found" and only a physical replug recovers it. Discovery therefore
identifies the dongle by USB VID:PID 0bda:2838, not descriptor text.
`usbreset 0bda:2838` recovers modes 1-3 as long as the device is still
enumerated. Kernel `over_current_count` stayed 0 throughout, so the port
never tripped protection; the dongle's own controller hangs.
Note: "[R82XX] PLL not locked!" printed while tuning is BENIGN -- it is
emitted while `rtl_sdr` probes and disables direct-sampling mode, and
appears on both rtl-sdr-blog and osmocom builds. An earlier revision of
this document blamed `rtl_test -t` for wedging the tuner; that was
DISPROVEN. The real symptom of a deaf receiver is railed, gain-independent
noise, which is what `hil/sdr_health.py` tests for.