1.5 KiB
Device loss during native presentation
The replay already treats VK_ERROR_DEVICE_LOST as terminal: device_lost()
writes its checkpoint report once, then replay_fatal() flushes the host log
before exiting. Continuing cannot complete GPU work and can retain captures and
staging allocations indefinitely.
Native presentation uses the same policy for the frame-fence wait/reset, command-pool reset, command-buffer begin/end, image acquisition, blit submission and presentation. The check precedes dependent work: a failed fence wait cannot lead to readback, a failed submission cannot enqueue a present waiting for its unsignaled semaphore, and a failed present cannot be counted as completed by Reflex. Both the shared and dedicated presentation queues are covered.
This does not recover a lost device, change successful presentation, or replace the existing out-of-date/suboptimal/Wayland-timeout handling. Other error classes retain their existing behavior, including allocation errors; this is not a general submission-recovery mechanism.
tests/test_present_device_loss.py executes the actual call/guard sites with
injected return values. It covers nine sites (two presentation queues), checks
that loss reports before the site continues, and supplies non-loss controls.
It tests routing rather than simulating an entire Vulkan implementation. The
existing test_device_lost_exit.py separately exercises the real concurrent
report-once and host-log-flush/exit contract. Hardware-reset recovery is not
claimed, and no physical driver reset is needed to run these tests.