Review finalization of LibUSB (1.x) devices and context

libusb_unref_device() is not supposed to be called after the context (for that device) has been destroyed with libsub_exit().

However, with

# file: repro.py

import usb

devices = list(usb.core.find(find_all=True))
print(f"Found {len(devices)} USB devices")

del(devices)

and current PyUSB HEAD:

$ LIBUSB_DEBUG=4 vmaster/bin/python repro.py
[timestamp] [threadID] facility level [function call] <message>
--------------------------------------------------------------------------------
[ 0.000008] [0000554e] libusb: debug [libusb_init] libusb v1.0.25.11692
[ 0.000015] [0000554e] libusb: debug [usbi_add_event_source] add fd 3 events 1
[ 0.000017] [0000554e] libusb: debug [usbi_io_init] using timer for timeouts
[ 0.000018] [0000554e] libusb: debug [usbi_add_event_source] add fd 4 events 1
[ 0.000021] [0000554e] libusb: debug [get_kernel_version] reported kernel version is 5.16.5-arch1-1
[ 0.000028] [0000554e] libusb: debug [op_init] found usbfs at /dev/bus/usb
[ 0.000029] [0000554e] libusb: debug [op_init] max iso packet length is (likely) 98304 bytes
[ 0.000032] [0000554e] libusb: debug [op_init] sysfs is available
[ 0.000229] [00005551] libusb: debug [linux_udev_event_thread_main] udev event thread entering
[ 0.002767] [0000554e] libusb: debug [linux_get_device_address] getting address for device: usb1 detached: 0
[ 0.002771] [0000554e] libusb: debug [linux_get_device_address] scan usb1
[ 0.002782] [0000554e] libusb: debug [linux_get_device_address] bus=1 dev=1
[ 0.002783] [0000554e] libusb: debug [linux_enumerate_device] busnum 1 devaddr 1 session_id 257
[ 0.002785] [0000554e] libusb: debug [linux_enumerate_device] allocating new device for 1/1 (session 257)
[... corresponding calls to 12 other devices ...]
[ 0.003536] [0000554e] libusb: debug [linux_get_device_address] getting address for device: usb4 detached: 0
[ 0.003537] [0000554e] libusb: debug [linux_get_device_address] scan usb4
[ 0.003547] [0000554e] libusb: debug [linux_get_device_address] bus=4 dev=1
[ 0.003549] [0000554e] libusb: debug [linux_enumerate_device] busnum 4 devaddr 1 session_id 1025
[ 0.003550] [0000554e] libusb: debug [linux_enumerate_device] allocating new device for 4/1 (session 1025)
[ 0.003624] [0000554e] libusb: debug [libusb_get_device_list]  
[ 0.003649] [0000554e] libusb: debug [libusb_get_device_descriptor]  
[... 12 other calls to libusb_get_device_descriptor ...]
[ 0.003939] [0000554e] libusb: debug [libusb_get_device_descriptor]  
Found 14 USB devices
[ 0.004073] [0000554e] libusb: debug [libusb_exit]  
[ 0.004104] [00005551] libusb: debug [linux_udev_event_thread_main] udev event thread exiting
[ 0.004127] [0000554e] libusb: debug [libusb_unref_device] destroy device 4.1
[... 12 other calls to libusb_unref_device ...]
[ 0.004144] [0000554e] libusb: debug [libusb_unref_device] destroy device 1.1
[ 0.004146] [0000554e] libusb: debug [usbi_remove_event_source] remove fd 4
[ 0.004149] [0000554e] libusb: debug [usbi_remove_event_source] remove fd 3

Related: libusb/libusb#1058

[jonasmalacofilho]

Hah, I fell for exactly that myself!

Good to know I’m not the only one : )

---

I think what allows the code to work currently (under both CPython and PyPy, as I’ve just confirmed) is:

• _Device instances can’t outlive the _Device class;
• the _Device class can’t outlive the libusb1 module;
• and the _LibUSB singleton instance lives as long as the libusb module.

I also think that if _Device._finalize_object gets to run (finalizers aren’t guaranteed to run, especially in some cases when the interpreter is exiting), then that counts to the _Device class still being there, and thus the _LibUSB instance as well.

So I’m going to close this as it seems everything is working properly, both practically and in theory.

Thanks for all the help!

[benzea]

Nevermind, module is a string, obviously…

Hah, I fell for exactly that myself!

I fear I am a bit out of knowledge about python garbage collection and object destruction at this point 😕

My naive approach would be to create a chain like Device -> _ResourceManager -> _Device -> _LibUSB, and then hopefully python cannot release the objects in the wrong order. So, maybe it is enough to attach a reference to _LibUSB to the _Device instance?