fix init image at creation

[rjodinchr]

We need to create the cvk_command_image_init with a valid context. Otherwise we can have a race condition between the destruction of queue and commands being deleted in the executor thread.

[rjodinchr]

switftshader artifact needs to be refreshed

[kpet]

Kicked off https://github.com/kpet/clvk/actions/runs/12951700691

[rjodinchr]

Can you re-trigger the failing bot? It's the first time I'm seeing that error. I wonder if it happens just very rarely. I'm not able to reproduce on my machine. That may be something we just want to filter in the test to avoid flakyness I guess

[rjodinchr]

In fact, the bug was much simpler than what I was thinking. It's just about not forcing the deletion of the queue.

[rjodinchr]

CI should be fixed with #771

[rjodinchr]

Using a cvk_command_queue_holder would be better looking, but we would go back to the headers issue: #735

[rjodinchr]

Fixing the bug also showed another deadlock between the executor and the main thread, for which I've added the fix.

The deadlock involves the lock of cvk_executor_thread.m_lock and cvk_executor_thread_pool.m_lock. The main thread has cvk_executor_thread_pool.m_lock running in cvk_executor_thread_pool::get_executor, and is trying to take cvk_executor_thread.m_lock in cvk_executor_thread::is_idle, while an executor is holding cvk_executor_thread.m_lock from cvk_executor_thread::executor and is trying to get cvk_executor_thread_pool.m_lock while deleting the queue that noone hold anymore in cvk_executor_thread_pool::return_executor.

main:
  cvk_executor_thread::is_idle -> tries to take cvk_executor_thread.m_lock
  cvk_executor_thread_pool::get_executor -> has cvk_executor_thread_pool.m_lock
executor:
  cvk_executor_thread_pool::return_executor -> tries to take cvk_executor_thread_pool.m_lock
  cvk_executor_thread::executor -> has cvk_executor_thread.m_lock

[kpet]

How did you bump into that deadlock? What test or application were you running?

[rjodinchr]

How did you bump into that deadlock? What test or application were you running?

Several tests in api_tests started to be flaky (with that deadlock) once I updated ChromeOS with that new version of the fix (with only the release on the command queue).

[rjodinchr]

TODO: move queue_holder into group?
Done