The story continues. The last fix of the syscall tracing code was wrong,
unfortunately. (The bug was a user could invoke any function in the
kernel. The fix was not to use t2 as pointer to a syscall, s0 was chosen
for it.) The problem we discovered is a few syscalls do SAVE_STATIC
(those declared as save_static_function), so s0 (which holds pointer to
the syscall at the time the syscall is invoked) is saved on the stack
overwriting a value saved from the process being traced. No wonder, s0
that restored on syscall exit differs from s0 saved on syscall enter.
See, arch/mips/kernel/scall32-o32.S, syscall_trace_entry, for example.
The rest of ABIs are the same.
There are several ways to fix this:
1. Make syscall handling code to be close to other arches. I mean, check
for the trace flag first, then parse arguments and invoke a syscall.
2. Remove save_static_functions and do SAVE_STATIC early for several
syscalls (yes, one big switch or its asm equivalent).
3. Store t2 in pt_regs (it means we have to expand this structure).
4. I know there should be yet another way.
Any ideas ?