Problem

libvgpu.so registers signal handlers for SIGUSR1 and SIGUSR2 using signal(), which overwrites any previously installed handlers without saving them. This causes JVM processes to crash with SIGSEGV in Monitor::wait() because the JVM uses SIGUSR1/SIGUSR2 internally for GC safepoints and thread management.

Observed on HAMi volcano-vgpu nodes (hami-core mode) when running PyTorch jobs with a JVM component — the crash occurs at startup before CUDA initializes.

Additional risks from the current implementation:

• libvgpu.so intercepts dlsym(), which the JVM also uses for native library loading
• The ENSURE_RUNNING() spin loop can cause a deadlock if a Java thread holding a JVM monitor gets suspended