Will,

When I was looking into syscall_trace_exit() more closely, I found
another (big) problem.
There are two system calls, execve() and rt_sigreturn(), which change
'syscallno' in pt_regs to -1 in start_thread() and restore_sigframe(),
respectively.

Since syscallno is not valid anymore in syscall_trace_exit() for these
system calls, we cannot create a correct syscall exit record for tracepoint
in trace_sys_exit() (=> ftrace_syscall_exit()) and for audit in audit_syscall_exit().

This does not happen on arm because syscall numbers are kept in
thread_info on arm.

How can we deal with this issue?

-Takahiro AKASHI


On 08/27/2014 02:51 AM, Will Deacon wrote:

On Fri, Aug 22, 2014 at 01:35:17AM +0100, AKASHI Takahiro wrote:

quoted

On 08/22/2014 02:08 AM, Kees Cook wrote:

quoted

On Thu, Aug 21, 2014 at 3:56 AM, AKASHI Takahiro
[off-list ref] wrote:

quoted

diff --git a/arch/arm64/kernel/ptrace.c b/arch/arm64/kernel/ptrace.c
index 8876049..c54dbcc 100644
--- a/arch/arm64/kernel/ptrace.c
+++ b/arch/arm64/kernel/ptrace.c

@@ -1121,9 +1121,29 @@ static void tracehook_report_syscall(struct pt_regs *regs,

   asmlinkage int syscall_trace_enter(struct pt_regs *regs)
   {
+       unsigned int saved_syscallno = regs->syscallno;
+
          if (test_thread_flag(TIF_SYSCALL_TRACE))
                  tracehook_report_syscall(regs, PTRACE_SYSCALL_ENTER);

+       if (IS_SKIP_SYSCALL(regs->syscallno)) {
+               /*
+                * RESTRICTION: we can't modify a return value of user
+                * issued syscall(-1) here. In order to ease this flavor,
+                * we need to treat whatever value in x0 as a return value,
+                * but this might result in a bogus value being returned.
+                */
+               /*
+                * NOTE: syscallno may also be set to -1 if fatal signal is
+                * detected in tracehook_report_syscall_entry(), but since
+                * a value set to x0 here is not used in this case, we may
+                * neglect the case.
+                */
+               if (!test_thread_flag(TIF_SYSCALL_TRACE) ||
+                               (IS_SKIP_SYSCALL(saved_syscallno)))
+                       regs->regs[0] = -ENOSYS;
+       }
+

I don't have a runtime environment yet for arm64, so I can't test this
directly myself, so I'm just trying to eyeball this. :)

Once the seccomp logic is added here, I don't think using -2 as a
special value will work. Doesn't this mean the Oops is possible by the
user issuing a "-2" syscall? As in, if TIF_SYSCALL_WORK is set, and
the user passed -2 as the syscall, audit will be called only on entry,
and then skipped on exit?

Oops, you're absolutely right. I didn't think of this case.
syscall_trace_enter() should not return a syscallno directly, but always
return -1 if syscallno < 0. (except when secure_computing() returns with -1)
This also implies that tracehook_report_syscall() should also have a return value.

Will, is this fine with you?

Well, the first thing that jumps out at me is why this is being done
completely differently for arm64 and arm. I thought adding the new ptrace
requests would reconcile the differences?

Will