On Fri, May 26, 2023 at 7:22 PM Samuel Thibault <samuel.thiba...@gnu.org> wrote: > I.e. gnumach without optimizations gets things wrong. We do want to fix > that :)
Dumping what I've been able to debug (so far): The crash indeed happens when loading the executable image, inside copyout (). The addresses/sizes are all fine, that's a false lead. What's apparently important here is that it's copying to the memory just vm_allocated inside the task's own map, and as always the page gets mapped lazily, on first access. The -O2 version attempts to access the page and lands in t_page_fault, which calls through kernel_trap to the VM subsystem, finds a free physical page, maps it in, irets back, and everything works out fine. The -O0 version for some reason lands in a_dbl_fault, and then proceeds to crash. This is probably the first time a page fault happens / is supposed to happen during execution. So it's probably unrelated to ELF loading and related to the VM subsystem and/or interrupt handling. The relevant parts of the code (copyout, interrupt/exception/trap handlers) are all written in assembly and so cannot be impacted by C optimization level. I don't know/understand enough about how interrupts work and x86 hardware details to make anything out of this. Sergey
