Message-ID: <CAO6TR8XtRF0Jc_EKT8tUm72dd9cyPqBk+pjtyTfiKBOMksF=Fg@mail.gmail.com>
Date:	Mon, 11 Jan 2016 20:09:12 -0700
From:	Jeff Merkey <linux.mdb@...il.com>
To:	Andy Lutomirski <luto@...capital.net>
Cc:	Linus Torvalds <torvalds@...ux-foundation.org>,
	LKML <linux-kernel@...r.kernel.org>,
	Thomas Gleixner <tglx@...utronix.de>,
	Ingo Molnar <mingo@...hat.com>,
	"H. Peter Anvin" <hpa@...or.com>, X86 ML <x86@...nel.org>,
	Peter Zijlstra <peterz@...radead.org>,
	Andy Lutomirski <luto@...nel.org>,
	Masami Hiramatsu <masami.hiramatsu.pt@...achi.com>,
	Steven Rostedt <rostedt@...dmis.org>,
	Borislav Petkov <bp@...en8.de>, Jiri Olsa <jolsa@...nel.org>
Subject: Re: [GIT PULL v4.5] Fix INT1 recursion with unregistered breakpoints

On 1/11/16, Andy Lutomirski <luto@...capital.net> wrote:
> On Mon, Jan 11, 2016 at 6:50 PM, Jeff Merkey <linux.mdb@...il.com> wrote:
>> On 1/11/16, Andy Lutomirski <luto@...capital.net> wrote:
>>> On Mon, Jan 11, 2016 at 6:26 PM, Jeff Merkey <linux.mdb@...il.com>
>>> wrote:
>>>> On 1/11/16, Andy Lutomirski <luto@...capital.net> wrote:
>>>>> On Mon, Jan 11, 2016 at 6:07 PM, Jeff Merkey <linux.mdb@...il.com>
>>>>> wrote:
>>>>>> On 1/11/16, Andy Lutomirski <luto@...capital.net> wrote:
>>>>>>> On Mon, Jan 11, 2016 at 5:30 PM, Jeff Merkey <linux.mdb@...il.com>
>>>>>>> wrote:
>>>>>>>> On 1/11/16, Andy Lutomirski <luto@...capital.net> wrote:
>>>>>>>>> On Mon, Jan 11, 2016 at 4:44 PM, Jeff Merkey <linux.mdb@...il.com>
>>>>>>>>> wrote:
>>>>>>>>>> Hi Thomas,
>>>>>>>>>>
>>>>>>>>>> I agree with #2, we should clear the breakpoint.  As for #1, if
>>>>>>>>>> there's an execute breakpoint it MUST be cleared or it will just
>>>>>>>>>> fire
>>>>>>>>>> off again when it sees the iretd from the int1 exception handler.
>>>>>>>>>> I
>>>>>>>>>> do use the breakpoint API Thomas, this showed up while debugging
>>>>>>>>>> and
>>>>>>>>>> testing the API with "lazy debug register switching".
>>>>>>>>>>
>>>>>>>>>> So do you want me to expand the patch and clear the breakpoint?
>>>>>>>>>> Just
>>>>>>>>>> give the word and I'll get busy and GIT -R- DONE.
>>>>>>>>>
>>>>>>>>> It seems to me that you're papering over some issue instead of
>>>>>>>>> fixing
>>>>>>>>> the root cause.  If you're using the API, then either you're doing
>>>>>>>>> it
>>>>>>>>> wrong or the API is broken.  Can you figure out which and fix it?
>>>>>>>>>
>>>>>>>>> --Andy
>>>>>>>>>
>>>>>>>>
>>>>>>>> Andy,
>>>>>>>>
>>>>>>>> Linux should not crash because someone triggered a breakpoint or
>>>>>>>> one
>>>>>>>> got triggered due to a program leaving some bits lying in a read
>>>>>>>> only
>>>>>>>> register (DR6) which for some strange reason someone in the linux
>>>>>>>> world decided could be used as local storage and to pass arguments
>>>>>>>> between subsystems - a register intel designed to be read from for
>>>>>>>> status.    I did not design what's in that API, I have to live with
>>>>>>>> it.
>>>>>>>
>>>>>>> The API appears to work, though.  Are you *sure* you're using it
>>>>>>> correctly?  Are you telling the code in kernel/hw_breakpoint.c about
>>>>>>> your breakpoint?
>>>>>>>
>>>>>>>> So all I am asking is that we fix this issue.  It does not matter
>>>>>>>> to my debugger is this is fixed or not in Linux, since I carry the
>>>>>>>> fix
>>>>>>>> in my patch, but it does matter to the overall robustness of Linux.
>>>>>>>
>>>>>>> Robust against what, exactly?  What's the bug?
>>>>>>>
>>>>>>> I will grant that the comments about lazy dr7 switching are
>>>>>>> mystifying, and cleaning them up might be nice.  But there's no
>>>>>>> adequate explanation of what the failure mode is, how to trigger it,
>>>>>>> or why your patch is a reasonable fix.  As it stands, you're
>>>>>>> duplicating code.
>>>>>>>
>>>>>>> --Andy
>>>>>>
>>>>>> Andy,
>>>>>>
>>>>>> Couple of things:
>>>>>>
>>>>>> Would you like a copy of the test harness that creates this bug to
>>>>>> test for yourself?  I previously posted it on the list.  If you don't
>>>>>> have it, I'll provide it.
>>>>>
>>>>> If you can send a short, buildable thing that triggers it, I'll read
>>>>> it.
>>>>>
>>>>>>
>>>>>> Since the dr6 bits get shifted around, it doesn't matter if the
>>>>>> breakpoint was registered or not in the API because the broken
>>>>>> handler
>>>>>> will call NULL bp structures and crash whether its registered or not.
>>>>>>
>>>>>
>>>>> And what exactly does this have to do with anything?  Your patch is
>>>>> all about spurious breakpoints triggered by dr7 and should have
>>>>> nothing much to do with the value in dr6.  Unless dr6 is missing a bit
>>>>> due to some issue, but you never suggested any problem like that.
>>>>>
>>>>
>>>> It's about setting the resume flag when an execute breakpoint occurs,
>>>> no
>>>> matter
>>>> what caused the breakpoint.  If is not set, the system will hang with
>>>> that processor
>>>> hung on the same execution address.  You cannot have an int1 exception
>>>> path
>>>> that does not set the resume flag which is the case here -- there
>>>> should be no path
>>>> where this flag does not get set on an execute breakpoint.
>>>
>>> There are many, many ways that one can corrupt kernel state to break
>>> things.  You could screw up IST state basically anywhere and crash.
>>> You could screw up GSBASE.  You could poke bad values into pt_regs in
>>> a fast syscall and hit the infamous SYSRET failure.  You can write a
>>> buggy .fault handler that returns success and doesn't actually do
>>> anything.  And yes, you can set a bit in dr7 without telling the
>>> hw_breakpoint code about it and thus infinite loop.
>>>
>>> Meanwhile, you keep claiming that kernel has a bug and that the bug
>>> can't be triggered without out-of-tree code.  In my book, that's not a
>>> bug.
>>>
>>
>> The handler that fails to set the resume flag is in tree code.
>
> It's an unreachable code path.
>
>>
>>> If you want to submit a nice clean patch to hw_breakpoint_handler to
>>> change the behavior on an unmatched breakpoint, then submit such a
>>> patch and justify why (a) the new behavior is better and (b) why it
>>> doesn't break any actual in-tree code.
>>>
>>
>> At last, a compromise -- accepted.  In the meantime, put this patch in
>> to get rid of the crash.  I'll code up another series and you can help me
>> by
>> reviewing it and keeping me on my toes.
>>
>
> No, because it still doesn't fix a bug *and* it's not a cleanup.
>

I'll work on the patch series you asked me to do.  This has been in
the kernel for a long time and only affects kernel debuggers for the
most part, and I have coded around it.  It's not unreachable though --
trigger a breakpoint without registering it -- or register another
int1 exception handler (its the two handler case when it shows up).
So if someone other than hw_breakpoint.c registers a second handler,
it shows up there too -- that's how I found it in the first place.
It can be triggered by simply calling in tree code and setting a
breakpoint with a second int1 handler registered as well, which I did
not report in the patch documentation because I wanted to keep it
short.    So consumers of the in-tree code can trigger it as well --
just like you said, in the notify handler, which is where the real
problem lies.

Jeff

Powered by blists - more mailing lists