In preload mode, we emulate sigaction, which prevents handlers from being installed natively. While we implement delivery of signals to those handlers for signals arriving via syscalls, we do not arrange to propagate signals that the host OS raises. e.g. if the managed thread executes an illegal instruction, the host Linux kernel will raise a native SIGILL, which will be delivered natively, killing the process. This is usually what would happen anyway, but a signal handler installed by the managed process to catch and "swallow" the SIGILL won't be invoked.

This is probably more common for SIGSEGV, where the signal handler might allocate memory or call mprotect to make the original access legal, and then return instead of crashing. This case is also slightly further complicated because in preload mode we do install our own SIGSEGV handler, but it only handles SIGSEGV's that were raised from executing RDTSC, and crashes on any other SIGSEGV without delegating to a signal handler that the managed process may have installed.

This issue appears to affect OpenJDK. #2084