On Linux targets with high-rate external sensor data (>1000Hz), all
sensor calibrations (gyro, accel, mag) can freeze PX4 by starving
other threads of CPU. Normal flight is unaffected — only calibration
triggers the problem.
Two compounding issues in the calibration worker threads:
1. calibrate_cancel_check() creates a new uORB::Subscription on every
call, which triggers getDeviceNodeLocked() — an O(n) linear strcmp
scan through all uORB nodes. In gyro/mag calibration this was called
on every sensor sample, consuming the majority of CPU in strcmp alone.
2. SubscriptionBlocking::updatedBlocking() returns immediately when data
is already available (it only blocks when no data is pending). With
continuous high-rate sensor data, the calibration loops never yield,
spinning at 100% CPU.
These problems are addressed with this patch as follows:
- Throttle calibrate_cancel_check() to once per 200ms in gyro
and mag calibration loops.
- Add 1ms px4_usleep() yield before updatedBlocking()/updateBlocking()
in all calibration loops (gyro, accel, mag, orientation detection).
This caps the effective loop rate at ~1000Hz — still far above what
calibration needs (250-750 samples).
- Force Commander main loop to sleep during calibration so it does not
compete with calibration worker threads for CPU.
Tested under Linux (x64, aarch64) both with RT and non-RT scheduling,
with sensor data arriving at ~3600Hz. Calibration completes normally
and no longer results in a deadlocked process.
