HailoRT 4.24.0 — How can I verify async callback drain during HailoNet shutdown?

General Discussion
, ,
1

I am investigating shutdown behavior in a GStreamer/HailoNet application using:

  • Hailo-8 M.2 module, part HM218B1C2FAE

  • HailoRT and PCIe driver 4.24.0

  • Firmware 4.24.0

  • Orange Pi 5 Plus, AArch64

  • Armbian 26.5.1 / Ubuntu 26.04

  • Kernel 6.1.115-vendor-rk35xx

Two bounded 48-frame runs produced the same application-visible result:

  • 48 frames admitted

  • 47 frames acknowledged at the Hailo source boundary

  • 1 frame pending at EOS

  • Hailo sink EOS occurred before source EOS

  • Four external probe contexts were destroyed

  • Zero external probe callbacks remained in flight

  • Process exited normally without a watchdog timeout

The application observer localizes the unacknowledged frame to the Hailo sink-to-source boundary. This does not establish why it was not observed or attribute causality to HailoRT.

My main question is about internal asynchronous lifetime:

Does HailoRT 4.24.0 guarantee that all detached inference-completion callbacks or outstanding transfers have completed or been safely cancelled before HailoNet finishes its READY-to-NULL transition and releases dependent resources?

In particular:

  1. Does configured-inference-model shutdown wait for all detached completion callbacks?

  2. What happens if its internal wait reaches a timeout?

  3. Is completion of the HailoNet output thread separate from completion of inference callbacks?

  4. Is there a supported API, tracepoint, log category, counter, debug-symbol package, or instrumented build that can prove zero outstanding callbacks/transfers at shutdown?

  5. Can Hailo provide version-matched symbols or a supported diagnostic for HailoRT 4.24.0?

I have exact runtime/plugin hashes, ELF build IDs, and preserved comparison evidence available if required.