powerpc: syscall_dotrace() && retcode (Was: powerpc: fork && stepping)

[Oleg Nesterov <oleg@redhat.com>]

On 11/28, Ananth N Mavinakayanahalli wrote:
>
> syscall-reset is the only failure I see on
> powerpc:
>
> errno 14 (Bad address)
> syscall-reset: syscall-reset.c:95: main: Assertion `(*__errno_location
> ()) == 38' failed.
> unexpected child status 67f
> FAIL: syscall-reset

(to remind, it also fails without utrace)

Once again, I know nothing about powerc, perhaps I misread the code,
but I believe this test-case is just wrong on powerpc and should be
fixed.

On powerpc, syscall_get_nr() returns regs->gpr[0], this means this
register is used to pass the syscall number.

This matches do_syscall_trace_enter(), it returns regs->gpr[0] as a
(possibly changed by tracer) syscall nr.

arch/powerpc/kernel/entry_64.S does

	syscall_dotrace:

		 bl      .do_syscall_trace_enter
		 mr      r0,r3	// I guess, r3 = r0 ?
		 ...
		 b       syscall_dotrace_cont

	syscall_dotrace_cont:

		syscall_dotrace_cont:

			cmpldi  0,r0,NR_syscalls
			bge-    syscall_enosys

	syscall_enosys:

		li      r3,-ENOSYS
		b       syscall_exit


Now return to the test-case, syscall-reset.c. The tracee does
l = syscall (-23, 1, 2, 3) and stops.

The tracer does

	#define RETREG	offsetof(struct pt_regs, gpr[0])
	#define NEWVAL	((long) ENOTTY)

	l = ptrace(PTRACE_PEEKUSER, child, RETREG, 0l);

l == -23, this is correct, note syscall(-23) above.

	l = ptrace(PTRACE_POKEUSER, child, RETREG, NEWVAL);

And expects the tracee will see NEWVAL==ENOTTY after return from
the systame call.

Of course this can't happen. We changed the syscall number, the
new value is ENOTTY == 25 == __NR_stime, sys_stime() correctly
returns -EFAULT.

-----------------------------------------------------------------

If I change the test-case to use NEWVAL == 1000 (or any other value
greater than NR_syscalls), then the tracee sees ENOSYS and this is
correct too.

But I do not see how it is possible to change the retcode on powerpc.
Unlike x86, powepc doesn't set -ENOSYS "in advance", before doing
do_syscall_trace_enter() logic. This means that if the tracer "cancels"
syscall, r3 will be overwritten by syscall_enosys.

This probably means the kernel should be fixed too, but I am not
brave enough to change the asm which I can't understand ;)

Oleg.