Follow @Openwall on Twitter for new release announcements and other news
[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Date: Fri, 6 Jul 2018 14:46:10 +0300
From: Alexander Popov <>
To: Ingo Molnar <>
Cc:, Kees Cook <>,
 PaX Team <>, Brad Spengler <>,
 Andy Lutomirski <>, Tycho Andersen <>,
 Laura Abbott <>, Mark Rutland <>,
 Ard Biesheuvel <>, Borislav Petkov <>,
 Richard Sandiford <>,
 Thomas Gleixner <>, "H . Peter Anvin" <>,
 Peter Zijlstra <>, "Dmitry V . Levin"
 <>, Emese Revfy <>,
 Jonathan Corbet <>, Andrey Ryabinin <>,
 "Kirill A . Shutemov" <>,
 Thomas Garnier <>,
 Andrew Morton <>,
 Alexei Starovoitov <>, Josef Bacik <>,
 Masami Hiramatsu <>, Nicholas Piggin <>,
 Al Viro <>, "David S . Miller" <>,
 Ding Tianhong <>, David Woodhouse
 <>, Josh Poimboeuf <>,
 Steven Rostedt <>,
 Dominik Brodowski <>,
 Juergen Gross <>,
 Linus Torvalds <>,
 Greg Kroah-Hartman <>,
 Dan Williams <>,
 Dave Hansen <>,
 Mathias Krause <>,
 Vikas Shivappa <>, Kyle Huey
 <>, Dmitry Safonov <>,
 Will Deacon <>, Arnd Bergmann <>,
 Florian Weimer <>,
 Boris Lukashev <>,
 Andrey Konovalov <>,,
Subject: Re: [PATCH v13 (resend) 2/6] x86/entry: Add STACKLEAK erasing the
 kernel stack at the end of syscalls

On 06.07.2018 01:20, Ingo Molnar wrote:
> * Alexander Popov <> wrote:
>> Hello Ingo,
>> Thanks for your review! I'll fix the style issues you pointed at.
>> Please also see my answers below.
>> On 05.07.2018 11:12, Ingo Molnar wrote:
>>>> +	  The tradeoff is the performance impact: on a single CPU system kernel
>>>> +	  compilation sees a 1% slowdown, other systems and workloads may vary
>>>> +	  and you are advised to test this feature on your expected workload
>>>> +	  before deploying it.
>>> Is there a way to patch this out runtime? I.e. if a distro enabled it, is there an 
>>> easy way to disable much of the overhead without rebooting the kernel?
>> Hm. We can't completely disable STACKLEAK in runtime, since STACKLEAK gcc plugin
>> performs compile-time instrumentation of the kernel code. So we can only chop
>> off a part of functionality, for example, by introducing some variable and
>> checking it before every stack erasing (additional performance impact), but the
>> kernel will stay uselessly instrumented. It doesn't look reasonable to me.
> Or we could use what every other performance critical instrumentation feature uses 
> to reduce overhead (ftrace, perf): kernel patching.

I see. It would be a big separate research - how to combine those different
kinds of instrumentation. I would propose to postpone it until we have a request
for STACKLEAK runtime disabling.

>>> If so then please make this:
>>> 	if (WARN_ON(boundary - kstack_ptr >= THREAD_SIZE))
>>> 		return;
>>> or so, to make it non-fatal and to allow users to report it, should it trigger 
>>> against all expectations.
>> I've made an experiment. The results:
>>  1. BUG_ON() here doesn't freeze the kernel output - I see a full 'PANIC: double
>> fault' report;
> Only in text mode - very few users are using text mode.
>>  2. WARN_ON() here gives absolutely same 'PANIC: double fault' here.
> that should only happen if the kernel is otherwise already fatally corrupted, 
> right?

No, I mean WARN_ON() in stackleak_erase_kstack() gives the double fault just
like BUG_ON() (without any corruption). In my experiment I've made the following

-       BUG_ON(boundary - kstack_ptr >= THREAD_SIZE);
+//     BUG_ON(boundary - kstack_ptr >= THREAD_SIZE);
+       WARN_ON(1);

It might be caused by the fact, that stackleak_erase_kstack() is called from the
trampoline stack just before returning to the userspace.

So I mean 'WARN_ON() + return' here wouldn't give any profit over a single
BUG_ON() check.

Best regards,

Powered by blists - more mailing lists

Confused about mailing lists and their use? Read about mailing lists on Wikipedia and check out these guidelines on proper formatting of your messages.