Re: [PATCH] io_thread/x86: don't reset 'cs', 'ss', 'ds' and 'es' registers for io_threads

From: Andy Lutomirski
Date: Mon May 03 2021 - 19:15:13 EST




> On May 3, 2021, at 3:56 PM, Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote:
>
> On Mon, May 03 2021 at 15:08, Linus Torvalds wrote:
>>> On Mon, May 3, 2021 at 2:49 PM Andy Lutomirski <luto@xxxxxxxxxx> wrote:
>>>
>>> To be clear, I'm suggesting that we -EINVAL the PTRACE_GETREGS calls
>>> and such, not the ATTACH. I have no idea what gdb will do if this
>>> happens, though.
>>
>> I feel like the likelihood that it will make gdb work any better is
>> basically zero.
>>
>> I think we should just do Stefan's patch - I assume it generates
>> something like four instructions (two loads, two stores) on x86-64,
>> and it "just works".
>>
>> Yeah, yeah, it presumably generates 8 instructions on 32-bit x86, and
>> we could fix that by just using the constant __USER_CS/DS instead (no
>> loads necessary) since 32-bit doesn't have any compat issues.
>>
>> But is it worth complicating the patch for a couple of instructions in
>> a non-critical path?
>>
>> And I don't see anybody stepping up to say "yes, I will do the patch
>> for gdb", so I really think the least pain is to just take the very
>> straightforward and tested kernel patch.
>>
>> Yes, yes, that also means admitting to ourselves that the gdb
>> situation isn't likely going to improve, but hey, if nobody in this
>> thread is willing to work on the gdb side to fix the known issues
>> there, isn't that the honest thing to do anyway?
>
> GDB is one thing. But is this setup actually correct under all
> circumstances?
>
> It's all fine that we have lots of blurb about GDB, but there is no
> reasoning why this does not affect regular kernel threads which take the
> same code path.
>
> Neither is there an answer what happens in case of a signal delivered to
> this thread and what any other GDB/ptraced induced poking might cause.
>
> This is a half setup user space thread which is assumed to behave like a
> regular kernel thread, but is this assumption actually true?
>
>

I’m personally concerned about FPU state. No one ever imagined when writing and reviewing the FPU state code that we were going to let ptrace poke the state on a kernel thread.

Now admittedly kernel_execve() magically turns kernel threads into user threads, but, again, I see no evidence that anyone has thought through all the implications of letting ptrace go to town before doing so.

(Is the io_uring thread a kthread style kernel thread? kthread does horrible, horrible things with the thread stack.)