lkml.org 
[lkml]   [2021]   [Jun]   [21]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: Kernel stack read with PTRACE_EVENT_EXIT and io_uring threads
From
Date
Hi Linus,

On 22/06/21 11:14 am, Linus Torvalds wrote:
> On Mon, Jun 21, 2021 at 12:45 PM Al Viro <viro@zeniv.linux.org.uk> wrote:
>>> Looks like sys_exit() and do_group_exit() would be the two places to
>>> do it (do_group_exit() would handle the signal case and
>>> sys_group_exit()).
>> Maybe... I'm digging through that pile right now, will follow up when
>> I get a reasonably complete picture
> We might have another possible way to solve this:
>
> (a) make it the rule that everybody always saves the full (integer)
> register set in pt_regs
>
> (b) make m68k just always create that switch-stack for all system
> calls (it's really not that big, I think it's like six words or
> something)

Correct - six words for registers, one for the return address. Probably
still a win compared to setting and clearing flag bits all over the
place in an attempt to catch any as yet undetected unsafe cases of
ptrace_stop.

I'll have to see how much of a performance impact I can see (not that I
can even remotely measure that accurately - it's more of a 'does it now
feel real sluggish' thing).

Cheers,

    Michael

>
> (c) admit that alpha is broken, but nobody really cares
>
>> In the meanwhile, do kernel/kthread.c uses look even remotely sane?
>> Intentional - sure, but it really looks wrong to use thread exit code
>> as communication channel there...
> I really doubt that it is even "intentional".
>
> I think it's "use some errno as a random exit code" and nobody ever
> really thought about it, or thought about how that doesn't really
> work. People are used to the error numbers, not thinking about how
> do_exit() doesn't take an error number, but a signal number (and an
> 8-bit positive error code in bits 8-15).
>
> Because no, it's not even remotely sane.
>
> I think the do_exit(-EINTR) could be do_exit(SIGINT) and it would make
> more sense. And the -ENOMEM might be SIGBUS, perhaps.
>
> It does look like the usermode-helper code does save the exit code
> with things like
>
> kernel_wait(pid, &sub_info->retval);
>
> and I see call_usermodehelper_exec() doing
>
> retval = sub_info->retval;
>
> and treating it as an error code. But I think those have never been
> tested with that (bogus) exit code thing from kernel_wait(), because
> it wouldn't have worked. It has only ever been tested with the (real)
> exit code things like
>
> if (pid < 0) {
> sub_info->retval = pid;
>
> which does actually assign a negative error code to it.
>
> So I think that
>
> kernel_wait(pid, &sub_info->retval);
>
> line is buggy, and should be something like
>
> int wstatus;
> kernel_wait(pid, &wstatus);
> sub_info->retval = WEXITSTATUS(wstatus) ? -EINVAL : 0;
>
> or something.
>
> Linus

\
 
 \ /
  Last update: 2021-06-22 02:01    [W:0.298 / U:1.016 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site