On 21/12/2015 11:13, Richard Weinberger wrote:
> Am 21.12.2015 um 10:23 schrieb Mickaël Salaün:
>>>>> Doesn't this break the support for changing syscall numbers using PTRACE_SETREGS?
>>>>
>>>> The logic is unchanged except updating the UPT_SYSCALL_NR before syscall_trace_enter(). I did my last tests with the x86_32 subarchitecture and all tests (from selftest/seccomp), including PTRACE_SETREGS for syscall numbers tests, passed. However, 2 of this tests still fail for x86_64 (only).
>>>
>>> No, the logic is different.
>>> syscall_trace_enter(regs) enters the ptrace() path and here registers can be changed.
>>> Hence "syscall = UPT_SYSCALL_NR(r);" will see the old syscall number.
>>> UPT_SYSCALL_NR() returns the syscall number before the ptrace() path...
>>
>> The thing is, PTRACE_SETREGS give access to *orig_ax* in the user_regs_struct from arch/x86/include/asm/user_*.h and selftest/seccomp only update this (virtual) register, not the EAX/RAX. Am I missing something?
> 
> Sorry, meant orig...
> 
> Please see the attached program. It proves that your patch is breaking stuff.
> The test is extracted from UML's selftests.

OK, I found the origin of this misunderstanding. On x86_32, PTRACE_SETREGS set regs->syscall when updating orig_eax, which is not the case on x86_64, hence the difference of behavior. I fixed this bug in the v2 series. The ptsc test and all the seccomp tests pass for 32 and 64 bits!

Where can we find the UML selftests?

Thanks,
 Mickaël