Conversation

Notices

1. Sergey Bugaev (bugaevc@floss.social)'s status on Monday, 13-Mar-2023 06:09:37 JST Sergey Bugaev

   🎉 x86_64-gnu glibc works 🎉

   ...kind of.

   I was unable to make x86_64 GNU Mach run it (needs further investigation), and then even if I would, gnumach still doesn't implement i386_fsgs_base_state, so glibc would crash & burn very early at TLS setup.

   So instead, I'm running it... on Linux. Yes, really 🙂

   I'm of course no stranger to running Mach software under Linux, if you know what I mean — but no, I have not implemented the (in?)famous HELL (the Hurd Emulation Layer for Linux).

   I'm using GDB!

   • Adrian Cochrane repeated this.
   • Sergey Bugaev (bugaevc@floss.social)'s status on Monday, 13-Mar-2023 06:09:36 JST Sergey Bugaev

     in reply to

     Linux happily agreed to exec my statically linked ELF, even though the ELF says it's for the Hurd. I had to gaslight GDB into thinking it the executable is for Linux, but that was easy ('set os default').

     Then I ran it, instruction by instruction (si, si, si, si...). When it attempts to make a syscall, I make it skip the syscall and model its effect in my head (you want __mach_task_self? here, say that returned a port name of 1).

     Adrian Cochrane repeated this.
   • Sergey Bugaev (bugaevc@floss.social)'s status on Monday, 13-Mar-2023 06:13:52 JST Sergey Bugaev

     in reply to

     The two interesting calls are setting fs_base (that very i386_fsgs_base_state thing) and allocating memory with vm_map.

     The former is easy to do, recent GDB versions let me just read and write $fs_base like any other register. So there I go, emulating a syscall (technically an RPC) that does not exist in the real Mach yet!

     Adrian Cochrane repeated this.
   • Sergey Bugaev (bugaevc@floss.social)'s status on Monday, 13-Mar-2023 06:14:20 JST Sergey Bugaev

     in reply to

     vm_map is a bit harder, because we have to allocate actual memory from Linux. I break'd before the very 'syscall' instruction, set registers as for the Linux's mmap syscall, held my breath, and let it execute 'syscall'. It worked! — so now I just had to save the newly allocated address as Mach would, and make vm_map return KERN_SUCCESS.

     Adrian Cochrane repeated this.
   • Sergey Bugaev (bugaevc@floss.social)'s status on Monday, 13-Mar-2023 06:20:31 JST Sergey Bugaev

     in reply to

     Anyways, it does all the stack magic and Hurd data block and TLS setup and malloc and tunables and symbols and whatnot, and finally calls user's main!

     So I can be reasonably sure that once GNU Mach 'gets there', glibc will not crash *right away* — maybe a little bit later, when you try to do something more complex, such as (oh horror) actually start the signal thread.