Messages in this thread | | | From | Andy Lutomirski <> | Date | Mon, 16 Jun 2014 14:54:40 -0700 | Subject | Re: 3.15: kernel BUG at kernel/auditsc.c:1525! |
| |
On Mon, Jun 16, 2014 at 2:48 PM, H. Peter Anvin <hpa@zytor.com> wrote: > On 06/16/2014 02:35 PM, Andy Lutomirski wrote: >> >> To hpa, etc: It appears that entry_32.S is missing any call to the >> audit exit hook on the badsys path. If I'm diagnosing this bug report >> correctly, this causes OOPSes. >> >> The the world at large: it's increasingly apparent that no one (except >> maybe the blackhats) has ever scrutinized the syscall auditing code. >> This is two old severe bugs in the code that have probably been there >> for a long time. >> > > Yes, the audit code is a total mess. > >> The bad syscall nr paths are their own incomprehensible route >> through the entry control flow. Rearrange them to work just like >> syscalls that return -ENOSYS. > > I have to admit... it sort of lends itself to a solution like this: > > /* For the 64-bit case, analogous code for 32 bits */ > movl $__NR_syscall_max+1,%ecx # *Not* __NR_syscall_max > cmpq %rcx,%rax > cmovae %rcx,%rax > movq %r10,%rcx > call *sys_call_table(,%rax,8) > > ... and having an extra (invalid) system call slot in the syscall table > beyond the end instead of branching off separately. > > (Note: we could use either cmova or cmovae, and either the 32- or 64-bit > form... the reason why is left as an exercise to the reader.)
For 64-bit, I want to do this instead:
https://git.kernel.org/cgit/linux/kernel/git/luto/linux.git/commit/?h=x86/seccomp-fastpath&id=a5ec2d7af2c54b55fc7201fa662138b53fbbda39
I see no reason why the 64-bit badsys code needs its own code path at all. I haven't sent it yet because AFAICT it doesn't fix any bug, and the series it's a part of isn't ready.
I'm also contemplating rewriting the 64-bit syscall entry work path in C.
--Andy
| |