The GPU-map observer imported a new mapping into Vulkan on the thread
that made the mapping, on whatever stack it ran. Guest fibers run on
contexts the title allocates in guest memory, with no guard page;
Demon's Souls packs its 16 KiB job fiber contexts back to back. A GPU
mapping made on a fiber used 10,616 bytes of the fiber's stack on
NVIDIA (Add, NoteGpuMapping, the registry walk in GuestAllocationsEnd,
the import, the fiber's entry and return), about 7.5 KiB of it for the
import, so a fiber that mapped GPU memory deep in a job could overrun
its context into the next one.
The observer now imports inline when the stack pointer lies inside the
thread's stack bounds with at least 256 KiB below it, and otherwise
starts a host thread for the import and joins it. On Windows the bounds
come from GetCurrentThreadStackLimits, which reads the TIB bounds that
libSceFiber sets to the fiber context while a fiber runs; on Linux from
pthread_getattr_np, whose bounds a fiber's stack lies outside of. A
mapping made while the thread holds the GPU mutex still imports inline:
a host thread could wait for that mutex forever. A host thread that
cannot be started is reported like any failed import, which is still
printed to stderr while the mapping carries on.
agc_driver_gpu_map_on_fiber maps GPU memory from a fiber with a 64 KiB
context and requires that the mapping use at most 8 KiB of it (10,616
bytes inline, 2,960 on a host thread) and start one host thread, then
maps the same memory from the main thread and requires that it start
none.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
ef022d53b8