Re: syscall_get_error() && TS_ checks

[Andy Lutomirski]

On Wed, Mar 29, 2017 at 9:45 AM, Linus Torvalds
<torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
> On Wed, Mar 29, 2017 at 9:33 AM, Oleg Nesterov <oleg@xxxxxxxxxx> wrote:
>>
>> Firstly, why do we need the IS_ERR_VALUE() check? This is only used by
>> do_signal/handle_signal, we do not care if it returns non-zero as long
>> as the value can't be confused with -ERESTART.* codes.
>
> There are system calls that can return "negative" values that aren't errors.
>
> Notably mmap() can return a valid pointer with the high bit set.
>
> So syscall_get_error() should return 0 for not just positive return
> values, but for those kinds of negative non-error values.
>
>> And why do we need the TS_ checks?
>
> Those may be bogus.
>
>> So why we can't simply change putreg32() to always sign-extend regs->ax
>> regs->orig_ax and just do
>>
>>         static inline long syscall_get_error(struct task_struct *task,
>>                                              struct pt_regs *regs)
>>         {
>>                 return regs-ax;
>>         }
>
> That would be *complete* garbage. Lots of system calls return positive
> values that sure as hell aren't errors.

Does this cause an observable problem?  The only things that care are:

a) 32-bit debugger pokes some value with the high bit and a 64-bit
debugger reads it back.  I seriously doubt we care.

b) 32-bit debugger pokes some value with the high bit set and the user
code switches to 64-bit mode and reads RAX.  This case is so
terminally broken anyway that we definitely don't care.

c) 32-bit debugger pokes some value with the high bit set and
syscall_get_error happens.  Oleg's proposed change won't change what
we do, but it will dramatically simplify the code.