Re: [PATCH] Fix get ERESTARTSYS with m32 in x86_64 when debug by GDB

[Andrew Pinski <pinskia@gmail.com>]

On Tue, Apr 29, 2014 at 9:50 PM, H. Peter Anvin <hpa@zytor.com> wrote:
> On 04/29/2014 08:44 PM, Hui Zhu wrote:
>>
>> I am sorry that the root cause of issue has something wrong.
>> The right root cause is:
>> When inferior call 32 bits syscall "read", Linux kernel function
>> "ia32_cstar_target" will set TS_COMPAT to current_thread_info->status.
>>
>> syscall read is interrupt by ctrl-c.   Then the $rax will be set to
>> errno -512 in 64 bits.
>> And the inferior will be stopped by Linux kernel function ptrace_stop,
>> the call trace is:
>> #0  freezable_schedule () at include/linux/freezer.h:172
>> #1  ptrace_stop (exit_code=exit_code@entry=5, why=why@entry=262148,
>>     clear_code=clear_code@entry=0, info=info@entry=0xffff88001d833e78)
>>     at kernel/signal.c:1920
>> #2  0xffffffff8107ec33 in ptrace_signal (info=0xffff88001d833e78, signr=5)
>>     at kernel/signal.c:2157
>> #3  get_signal_to_deliver (info=info@entry=0xffff88001d833e78,
>>     return_ka=return_ka@entry=0xffff88001d833e58, regs=<optimized out>,
>>     cookie=cookie@entry=0x0 <irq_stack_union>) at kernel/signal.c:2269
>> #4  0xffffffff81013438 in do_signal (regs=regs@entry=0xffff88001d833f58)
>>     at arch/x86/kernel/signal.c:696
>> #5  0xffffffff81013a40 in do_notify_resume (regs=0xffff88001d833f58,
>>     unused=<optimized out>, thread_info_flags=4) at arch/x86/kernel/signal.c:747
>> #6  <signal handler called>
>> #7  0x0000000000000000 in irq_stack_union ()
>>
>> After that, GDB can control the stopped inferior.
>> To call function "func1()" of inferior, GDB need:
>> Step 1, save current values of registers ($rax 0xfffffffffffffe00(64 bits -512)
>> is cut to 0xfffffe00(32 bits -512) because inferior is a 32 bits program).
>
> So gdb just corrupted the system state.

Except GDB in 32bit mode does not know the registers are full 64bits so ...

>
>> Step 2, change the values of registers.
>> Step 3, Push a dummy frame to stack.
>> Step 4, set a breakpint in the return address.
>>
>> When GDB resume the inferior, it will keep execut from ptrace_stop
>> with new values of registers that set by GDB.
>
>> And TS_COMPAT inside current_thread_info->status will be cleared when
>> inferior switch back to user space.
>
> As it should, because TS_COMPAT *only is meaningful while a system call
> is executing*.
>
>> When function "func1()" return, inferior will be stoped by breakpoint
>> inferior will be stopped by Linux kernel function "ptrace_stop" again.
>> current_thread_info->status will not set TS_COMPAT when inferior swith
>> from user space to kernel space because breakpoint handler "int3" doesn't
>> has code for that.
>
> As it shouldn't, because there is no system call entry involved.
>
>> GDB begin to set saved values of registers back to inferior that use
>> function "amd64_collect_native_gregset".  Because this function just
>> zero-extend each 32 bits value to 64 bits value before put them to inferior.
>> $rax's value is set to 0xfffffe00(32 bits -512) but not
>> 0xfffffffffffffe00(64 bits -512).
>>
>> When GDB continue syscall "read" that is interrupted by "ctrl-c", it will
>> keep execute from ptrace_stop without "TS_COMPAT".
>
> gdb has corrupted the state, and it fails to execute.
>
> I'm wondering if we need to add additional state here, to carry the
> TS_COMPAT bit.  We have talked about this kind of issues in the past.
>
>> Then in Linux kernel function "syscall_get_error", current_thread_info->status
>> doesn't have TS_COMPAT and $rax is 0xfffffe00(32 bits -512).  Then in
>> function do_signal will not handle this -ERESTARTSYS.
>>
>> -ERESTARTSYS will be return back to inferior, that is why inferior got a
>> errno -ERESTARTSYS.
>>
>> I made a new patch that before call do_notify_resume(will call do_signal)
>> in the int3 handler, set TS_COMPAT to status if this task is TIF_IA32.
>> Then after GDB call a function of inferior, it will still has TS_COMPAT.
>
> I'm not sure if I want to label this a gdb bug or not (my main feeling
> is that gdb should save and restore the register set presented to it,
> and that truncating values to 32 bits is the root of the problem), but
> the above is definitely a hack which doesn't really address the real
> problem.

restoring the values is hard since even the ptrace interface does not
allow for that.

Thanks,
Andrew Pinski

>
>         -hpa
>