copy_mutex duplicated the pthread_mutex_t of vk.queue.mutex into
main_queue->mutex by value, which creates an independent mutex: monado
submissions on IPC client threads (comp_swapchain image transitions,
vkDeviceWaitIdle in image_cleanup) did not exclude wivrn's layer_commit
and encoder submissions on the same VkQueue. This intermittently
corrupts driver state when a client creates or destroys swapchains
while streaming, observed as SIGSEGV inside vkQueueSubmit2 and as
uncaught vk::SystemError from layer_commit on NVIDIA.
Initialize monado's queue mutexes and redirect vk.queue.mutex to the
main queue os_mutex so every submitter takes the same lock.
In the layer squasher, world pose is retrieved from a projection layer
if possible, but the eye pose is incorrect in that case.
Get the data from hmd if it is needed by any of the layers.
Optional server profiling built on the Perfetto SDK. Adds CPU spans
(compositor, encoder, network), GPU encode/copy slices from Vulkan timestamp
queries (Vulkan/NVENC/VAAPI), and headset feedback markers from dump_time.
WIVRN_TRACING enables tracing and selects the backend (system, inprocess, or
both); WIVRN_USE_PERFETTO gates it at build time and compiles it out otherwise.
WIVRN_TRACE_MONADO additionally drives Monado's percetto tracing from the same
switch so WiVRn and Monado land on one timeline.
Tooling: server-tracing[-monado] presets and a CI build, two Perfetto configs
(WiVRn-only and full-system), a wivrn_capture.py capture wrapper, a pftrace
summary script, and docs/profiling.md.
Compositor release left layouts as eUndefined while encoder acquire
transitioned to eVideoEncodeSrcKHR (or eTransferSrcOptimal on the NVIDIA
tmp_image path); the spec forbids that layout mismatch for exclusive-mode
QFOT. NVIDIA honoured it by treating encode-queue contents as undefined
(solid green frame).
Encoders now publish target_layout. The compositor's release and the
encoder's acquire both transition eGeneral -> target_layout in a single
barrier; per VkImageMemoryBarrier2 the layout transition is performed
exactly once between the queues, so no second non-QFOT barrier is needed.
target_layout is eVideoEncodeSrcKHR on the direct vulkan encode path and
eGeneral elsewhere.
Also fix optimal_transfer(): the inner `auto to = ...` shadowed the
parameter, so the maintenance9 check was effectively always false.