On 07/23/2014 05:25 PM, Will Deacon wrote:

On Wed, Jul 23, 2014 at 08:03:47AM +0100, AKASHI Takahiro wrote:

quoted

On 07/23/2014 05:15 AM, Kees Cook wrote:

quoted

On Tue, Jul 22, 2014 at 2:14 AM, AKASHI Takahiro
[off-list ref] wrote:

quoted

   asmlinkage int syscall_trace_enter(struct pt_regs *regs)
   {
+       unsigned long saved_x0, saved_x8;
+
+       saved_x0 = regs->regs[0];
+       saved_x8 = regs->regs[8];
+
          if (test_thread_flag(TIF_SYSCALL_TRACE))
                  tracehook_report_syscall(regs, PTRACE_SYSCALL_ENTER);

+       regs->syscallno = regs->regs[8];
+       if ((long)regs->syscallno == ~0UL) { /* skip this syscall */
+               regs->regs[8] = saved_x8;
+               if (regs->regs[0] == saved_x0) /* not changed by user */
+                       regs->regs[0] = -ENOSYS;

I'm not sure this is right compared to other architectures. Generally
when a tracer performs a syscall skip, it's up to them to also adjust
the return value. They may want to be faking a syscall, and what if
the value they want to return happens to be what x0 was going into the
tracer? It would have no way to avoid this -ENOSYS case. I think
things are fine without this test.

Yeah, I know this issue, but was not sure that setting a return value
is mandatory. (x86 seems to return -ENOSYS by default if not explicitly
specified.)
Is "fake a system call" a more appropriate word than "skip"?

I will defer to Will.

I agree with Kees -- iirc, I only suggested restoring x8.

OK.

-Takahiro AKASHI

Will