Re: [RFC 3/3] x86/ptrace: down with test_thread_flag(TIF_IA32)

[Andy Lutomirski]

On Mon, Apr 25, 2016 at 11:09 AM, Dmitry Safonov <dsafonov@xxxxxxxxxxxxx> wrote:
> On 04/25/2016 08:14 PM, Dmitry Safonov wrote:
>>
>> On 04/25/2016 07:53 PM, Andy Lutomirski wrote:
>>>
>>> On Mon, Apr 25, 2016 at 9:12 AM, Dmitry Safonov <dsafonov@xxxxxxxxxxxxx>
>>> wrote:
>>>>
>>>> As the task isn't executing at the moment of {GET,SET}REGS,
>>>> return regset that corresponds to code selector.
>>>> So, for i386 elf binary that changed it's CS to __USER_CS
>>>> it will return full x86_64 register set.
>>>>
>>>> That will change ABI: i.e, strace uses returned register size
>>>> to determine, in which mode the application is.
>>>> With the current ABI that way is buggy:
>>>
>>> Oleg, any comment here?
>>>
>>>> int main(int argc, char **argv, char **envp)
>>>> {
>>>>          printf("Here we exit\n");
>>>>          fflush(stdout);
>>>>          asm volatile ("int $0x80" : : "a" (__NR_exit), "D" (1));
>>>>          printf("After exit\n");
>>>>
>>>>          return 0;
>>>> }
>>>>
>>>> This program will confuse strace:
>>>>
>>>> [tst]$ strace ./confuse 2>&1 | tail
>>>> brk(0x1ca1000)                          = 0x1ca1000
>>>> write(1, "Here we exit\n", 13Here we exit
>>>> )          = 13
>>>> exit(1)                                 = ?
>>>> <... exit resumed> strace: _exit returned!
>>>> )                    = ?
>>>> write(1, "After exit\n", 11After exit
>>>> )            = 11
>>>> exit_group(0)                           = ?
>>>> +++ exited with 0 +++
>>>>
>>>> So this ABI change should make PTRACE_GETREGSET more reliable and
>>>> this will be another step to drop TIF_{IA32,X32} flags.
>>>
>>> Does strace start working again with this change?  I suspect that
>>> we'll eventually have to expose syscall_get_arch directly through
>>> ptrace, but that's a project for another day.
>>
>>
>> Oh, crap, not yet - seems like, I failed with my test.
>> I'll resend this patch as will get it fixed, sorry.
>
>
> I find out, what I have changed (and broke test):
>> if (!user_64bit_mode(task_pt_regs(task)))
> was
>> if (task_thread_info(task)->status & TS_COMPAT)
>
> That way the test runs now:
>>
>> brk(NULL)                               = 0x1145000
>> brk(0x1167000)                          = 0x1167000
>> write(1, "Here we exit\n", 13Here we exit
>> )          = 13
>> strace: [ Process PID=1608 runs in 32 bit mode. ]
>> umask(0)                                = 022
>> strace: [ Process PID=1608 runs in 64 bit mode. ]
>> write(1, "After exit\n", 11After exit
>> )            = 11
>> exit_group(0)                           = ?
>> +++ exited with 0 +++
>
>
> But I changed on signal patch rebase and now I'm
> thinking: should it be
>> if (task_thread_info(task)->status & TS_COMPAT ||
>> !user_64bit_mode(task_pt_regs(task)))
> or what?
> Should we count program that does compat syscall
> as compatible even if it's in 64-bit mode?

I think we should report 64-bit regs if the app is running in 64-bit
mode.  Then (not necessarily as part of your series), we should have a
way for ptrace users to query the *syscall* arch.

IOW, this is totally screwed up right now.  I think the goal of your
patch should be to stop using TIF_IA32 without breaking it any worse
than it already is.

--Andy