lkml.org 
[lkml]   [2017]   [Sep]   [1]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
Date
SubjectRe: tip -ENOBOOT - bisected to locking/refcounts, x86/asm: Implement fast refcount overflow protection
On Fri, Sep 1, 2017 at 11:58 AM, Kees Cook <keescook@chromium.org> wrote:
> On Fri, Sep 1, 2017 at 10:52 AM, Mike Galbraith <efault@gmx.de> wrote:
>> On Fri, 2017-09-01 at 10:12 -0700, Kees Cook wrote:
>>> On Fri, Sep 1, 2017 at 6:09 AM, Mike Galbraith <efault@gmx.de> wrote:
>>> > On Fri, 2017-09-01 at 08:57 +0200, Mike Galbraith wrote:
>>> >> On Thu, 2017-08-31 at 11:45 -0700, Kees Cook wrote:
>>> >> > On Thu, Aug 31, 2017 at 10:19 AM, Mike Galbraith <efault@gmx.de> wrote:
>>> >> > > On Thu, 2017-08-31 at 10:00 -0700, Kees Cook wrote:
>>> >> > >>
>>> >> > >> Oh! So it's gcc-version sensitive? That's alarming. Is this mapping correct:
>>> >> > >>
>>> >> > >> 4.8.5: WARN, eventual kernel hang
>>> >> > >> 6.3.1, 7.0.1: WARN, but continues working
>>> >> > >
>>> >> > > Yeah, that's correct. I find that troubling, simply because this gcc
>>> >> > > version has been through one hell of a lot of kernels with me. Yeah, I
>>> >> > > know, that doesn't exempt it from having bugs, but color me suspicious.
>>> >> >
>>> >> > I still can't hit this with a 4.8.5 build. :(
>>> >> >
>>> >> > With _RATELIMIT removed, this should, in theory, report whatever goes
>>> >> > negative first...
>>> >>
>>> >> I applied the other patch you posted, and built with gcc-6.3.1 to
>>> >> remove the gcc-4.8.5 aspect. Look below the resulting splat.
>>> >
>>> > Grr, that one has a in6_dev_getx() line missing for the first
>>> > increment, where things go pear shaped.
>>> >
>>> > With that added, looking at counter both before, and after incl, with a
>>> > trace_printk() in the exception handler showing it doing its saturate
>>> > thing, irqs disabled across the whole damn refcount_inc(), and even
>>> > booting box nr_cpus=1 for extra credit...
>>> >
>>> > HTH can that first refcount_inc() get there?
>>> >
>>> > # tracer: nop
>>> > #
>>> > # _-----=> irqs-off
>>> > # / _----=> need-resched
>>> > # | / _---=> hardirq/softirq
>>> > # || / _--=> preempt-depth
>>> > # ||| / delay
>>> > # TASK-PID CPU# |||| TIMESTAMP FUNCTION
>>> > # | | | |||| | |
>>> > systemd-1 [000] d..1 1.937284: in6_dev_getx: PRE refs.counter:3
>>> > systemd-1 [000] d..1 1.937295: ex_handler_refcount: *(int *)regs->cx = -1073741824
>>> > systemd-1 [000] d..1 1.937296: in6_dev_getx: POST refs.counter:-1073741824
>>>
>>> O_o
>>>
>>> Can you paste the disassembly of in6_dev_getx? I can't understand how
>>> we're landing in the exception handler.
>>
>> I was hoping you'd say that.
>>
>> 0xffffffff816b2f72 <+0>: push %rbp
>> 0xffffffff816b2f73 <+1>: mov %rsp,%rbp
>> 0xffffffff816b2f76 <+4>: push %r12
>> 0xffffffff816b2f78 <+6>: push %rbx
>> 0xffffffff816b2f79 <+7>: incl %gs:0x7e95a2d0(%rip) # 0xd250 <__preempt_count>
>> 0xffffffff816b2f80 <+14>: mov 0x308(%rdi),%rbx
>> 0xffffffff816b2f87 <+21>: test %rbx,%rbx
>> 0xffffffff816b2f8a <+24>: je 0xffffffff816b2feb <in6_dev_getx+121>
>> 0xffffffff816b2f8c <+26>: callq *0xffffffff81c35a00
>> 0xffffffff816b2f93 <+33>: mov %rax,%r12
>> 0xffffffff816b2f96 <+36>: callq *0xffffffff81c35a10
>> 0xffffffff816b2f9d <+43>: mov 0x769ad4(%rip),%rsi # 0xffffffff81e1ca78 <trace_printk_fmt.21733>
>> 0xffffffff816b2fa4 <+50>: mov 0xf0(%rbx),%edx
>> 0xffffffff816b2faa <+56>: mov $0xffffffff816b2f8c,%rdi
>> 0xffffffff816b2fb1 <+63>: callq 0xffffffff81171fc0 <__trace_bprintk>
>> 0xffffffff816b2fb6 <+68>: lock incl 0xf0(%rbx)
>> 0xffffffff816b2fbd <+75>: js 0xffffffff816b2fbf <in6_dev_getx+77>
>> 0xffffffff816b2fbf <+77>: lea 0xf0(%rbx),%rcx
>> 0xffffffff816b2fc6 <+84>: (bad)
>> 0xffffffff816b2fc8 <+86>: mov 0x769a99(%rip),%rsi # 0xffffffff81e1ca68 <trace_printk_fmt.21744>
>> 0xffffffff816b2fcf <+93>: mov 0xf0(%rbx),%edx
>> 0xffffffff816b2fd5 <+99>: mov $0xffffffff816b2f8c,%rdi
>> 0xffffffff816b2fdc <+106>: callq 0xffffffff81171fc0 <__trace_bprintk>
>> 0xffffffff816b2fe1 <+111>: mov %r12,%rdi
>> 0xffffffff816b2fe4 <+114>: callq *0xffffffff81c35a08
>> 0xffffffff816b2feb <+121>: decl %gs:0x7e95a25e(%rip) # 0xd250 <__preempt_count>
>> 0xffffffff816b2ff2 <+128>: mov %rbx,%rax
>> 0xffffffff816b2ff5 <+131>: pop %rbx
>> 0xffffffff816b2ff6 <+132>: pop %r12
>> 0xffffffff816b2ff8 <+134>: pop %rbp
>> 0xffffffff816b2ff9 <+135>: retq
>>
>> I don't get the section business at all, +75 looks to me like we're
>> gonna trap no matter what.. as we appear to be doing.
>
> The section stuff is supposed to be a trick to push the error case off
> into the .text.unlikely area to avoid needing a jmp over the handler
> and with possibly some redundancy removal done by the compiler (though
> this appears to be rather limited) if it notices a bunch of error
> paths are the same. However, in your disassembly, it's inline (!!) in
> the code, as if "pushsection" and "popsection" were entirely ignored.
>
> And when I make my own in6_dev_getx(), I see the same disassembly:
>
> 0xffffffff818a757b <+181>: lock incl 0x1e0(%rbx)
> 0xffffffff818a7582 <+188>: js 0xffffffff818a7584 <in6_dev_getx+190>
> 0xffffffff818a7584 <+190>: lea 0x1e0(%rbx),%rcx
> 0xffffffff818a758b <+197>: (bad)
>
> Which is VERY different from how it looks in other places!

Found it.

If the compiler already pushed the entire function into
.text.unlikely, x86-refcount's .pushsection doesn't do any good
(obviously). Durrr.

.section .text.unlikely,"ax",@progbits
.type in6_dev_getx, @function
in6_dev_getx:
.LFB4673:
.loc 2 4128 0
.cfi_startproc
...
lock; incl 480(%rbx)
js 111f
.pushsection .text.unlikely
111: lea 480(%rbx), %rcx
112: .byte 0x0f, 0xff
.popsection
113:

I will get this fixed. Thank you again for helping track this down!

-Kees

--
Kees Cook
Pixel Security

\
 
 \ /
  Last update: 2017-09-01 21:40    [W:0.542 / U:1.008 seconds]
©2003-2017 Jasper Spaans. hosted at Digital OceanAdvertise on this site