On Aug 2, 11:17pm, manu@netbsd.org (Emmanuel Dreyfus) wrote:
-- Subject: Topdown UVM breaking COMPAT_IRIX?

Do you have TOPDOWN enabled?

christos

| Hi
| 
| COMPAT_IRIX has been broken for a few months, I'm not 100% sure, but I
| suspect it's related to this commit:
| 
| http://cvsweb.netbsd.org/bsdweb.cgi/
| src/sys/kern/exec_elf32.c.diff?r1=3D1.82&r2=3D1.83&f=3Dh
| 
| When I run an IRIX program, the kernel loads the interpreter at a
| default address it chooses itself, and it prepare the user register so
| that control is transfered to the interpreter entry point, relocated to
| its actual position. That works, the interpreter starts up.
| 
| But IRIX's dynamic linker does not seems to be relocatable. So quickly,
| the interpreter code does a jump at an absolute address, where it
| expects to find another dynamic linker routine. As we relocated it, it
| jumps in the middle of nowhere and crashes (memory fault: there is
| nothing mapped there).
| 
| Program startup:
| 0xe0774 in ?? ()
| (gdb) x/4i $pc
| 0xe0774:        bal     0xe077c
| 0xe0778:        nop
| 0xe077c:        lui     gp,0x7
| 0xe0780:        addiu   gp,gp,31604
| [ok, this is the interpreter start code, as observed with objdump]
| (gdb) c
| Continuing.
| 
| Program received signal SIGSEGV, Segmentation fault.
| 0xfa85850 in ?? ()
| 
| In the ELF headers, interpreter's startup is at 0xfae0774. This is where
| IRIX starts it, and this is also what objdump tells me.
| 
| First question: does that analysis seem right?
| Second question: if it's right, how to fix that?
| 
| I see two ways of fising that. First way is to modify elf_load_file() so
| that it searches the ELF headers for the interpreter load address and
| use that address.
| 
| Second way is to use the ability of elf_load_file') to get the
| interpreter load address from the probe function. I could look for the
| intepreter load address in irix_probe_{n,o}32(), and hand it to
| elf_load_file().
| 
| What is the right way? Looking for the right load address in
| elf_load_file seems a bad choice on the performance front, but on the
| other hand, this change might have broken more emulations that just
| IRIX. What about creating a function in sys/kern/exec_elf32.c that would
| look for the load address, and use that function in emulation's probe
| function when it is apporpriate?
| 
| Last question: what other emulation might have been broken? As I
| understand, this is a mips sp=E9cific problem? COMPAT_LINUX on mips is
| okay, I checked it. No idea for COMPAT_ULTRIX. Who uses that?
| 
| --=20
| Emmanuel Dreyfus
| Il y a 10 sortes de personnes dans le monde: ceux qui comprennent=20
| le binaire et ceux qui ne le comprennent pas.
| manu@netbsd.org
-- End of excerpt from Emmanuel Dreyfus