Re: [Xenomai-core] x86: Endless minor faults

[Gilles Chanteperdrix <gilles.chanteperdrix@xenomai.org>]

Philippe Gerum wrote:
> On Tue, 2009-06-30 at 11:21 +0200, Gilles Chanteperdrix wrote:
>> Philippe Gerum wrote:
>>> On Tue, 2009-06-30 at 10:42 +0200, Gilles Chanteperdrix wrote:
>>>> Jan Kiszka wrote:
>>>>> Jan Kiszka wrote:
>>>>>> Hi all,
>>>>>>
>>>>>> seen such loops before? This particular trace is from a 2.6.29.3 kernel
>>>>>> with ipipe-2.3-01 (SMP/PREEMPT_VOLUNTARY), but the same happens with
>>>>>> 2.6.29.5/2.3-03:
>>>>>>
>>>>>> :|   +func                -653    0.084  __ipipe_handle_exception+0x11 (page_fault+0x26)
>>>>>> :|   +func                -653    0.096  ipipe_check_context+0xd (__ipipe_handle_exception+0x71)
>>>>>> :|   #end     0x80000000  -653    0.069  do_page_fault+0x33 (__ipipe_handle_exception+0x1ff)
>>>>>> :    #func                -653    0.078  __ipipe_unstall_root+0x9 (do_page_fault+0x3cb)
>>>>>> :|   #begin   0x80000000  -653    0.068  __ipipe_unstall_root+0x34 (do_page_fault+0x3cb)
>>>>>> :|   +end     0x80000000  -653    0.069  __ipipe_unstall_root+0x59 (do_page_fault+0x3cb)
>>>>>> :    +func                -653    0.060  down_read_trylock+0x4 (do_page_fault+0x424)
>>>>>> :    +func                -653    0.068  _spin_lock_irqsave+0x9 (__down_read_trylock+0x16)
>>>>>> :    +func                -653    0.108  ipipe_check_context+0xd (_spin_lock_irqsave+0x1d)
>>>>>> :    #func                -652    0.066  _spin_unlock_irqrestore+0x4 (__down_read_trylock+0x3f)
>>>>>> :    #func                -652    0.069  __ipipe_restore_root+0x4 (_spin_unlock_irqrestore+0x21)
>>>>>> :    #func                -652    0.074  __ipipe_unstall_root+0x9 (__ipipe_restore_root+0x2c)
>>>>>> :|   #begin   0x80000000  -652    0.066  __ipipe_unstall_root+0x34 (__ipipe_restore_root+0x2c)
>>>>>> :|   +end     0x80000000  -652    0.069  __ipipe_unstall_root+0x59 (__ipipe_restore_root+0x2c)
>>>>>> :    +func                -652    0.096  find_vma+0x4 (do_page_fault+0x465)
>>>>>> :    +func                -652    0.150  ltt_run_filter_default+0x4 (_ltt_specialized_trace+0xc1)
>>>>>> :    +func                -652    0.098  handle_mm_fault+0x11 (do_page_fault+0x537)
>>>>>> :    +func                -652    0.090  _spin_lock+0x4 (handle_mm_fault+0x680)
>>>>>> :    +func                -652    0.063  ptep_set_access_flags+0x9 (handle_mm_fault+0x6d1)
>>>>>> :    +func                -652    0.282  flush_tlb_page+0xd (handle_mm_fault+0x6e7)
>>>>>> :    +func                -651    0.162  ltt_run_filter_default+0x4 (_ltt_specialized_trace+0xc1)
>>>>>> :    +func                -651    0.062  up_read+0x4 (do_page_fault+0x5a9)
>>>>>> :    +func                -651    0.072  _spin_lock_irqsave+0x9 (__up_read+0x1c)
>>>>>> :    +func                -651    0.117  ipipe_check_context+0xd (_spin_lock_irqsave+0x1d)
>>>>>> :    #func                -651    0.074  _spin_unlock_irqrestore+0x4 (__up_read+0x92)
>>>>>> :    #func                -651    0.069  __ipipe_restore_root+0x4 (_spin_unlock_irqrestore+0x21)
>>>>>> :    #func                -651    0.060  __ipipe_unstall_root+0x9 (__ipipe_restore_root+0x2c)
>>>>>> :|   #begin   0x80000000  -651    0.056  __ipipe_unstall_root+0x34 (__ipipe_restore_root+0x2c)
>>>>>> :|   +end     0x80000000  -651    0.420  __ipipe_unstall_root+0x59 (__ipipe_restore_root+0x2c)
>>>>>> :|   +func                -650    0.084  __ipipe_handle_exception+0x11 (page_fault+0x26)
>>>>>>
>>>>>> and again and again...
>>>>>>
>>>>>> We are looping over a minor fault here (according to /proc/PID/stat),
>>>>>> the context is a Xenomai task in secondary mode. As the task no longer
>>>>>> processes signals in this state, the whole system is more or less
>>>>>> broken. Tomorrow I will try to find out the faulting address with an
>>>>>> instrumented kernel, but maybe you already have some ideas.
>>>>> The fault is apparently triggered by __xn_put_user(XNRELAX,
>>>>> thread->u_mode) in xnshadow_relax. thread->u_mode is pointing to an
>>>>> invalid region ATM. The questions are now: Who corrupted this, user
>>>>> space on init (not that likely) or kernel space later on (unpleasant
>>>>> thought)? Moreover: Why can't we recover from a fault on u_mode?
>>>> I already investigated such an issue, and my conclusion was that there
>>>> are some places in the code where we can not cope with a fault.
>>>> xnshadow_relax being such a place, because, if relax faults, then what
>>>> will the fault handler do? Call relax again. Fortunately, mlockall and
>>>> the nocow stuff fixes this.
>>>
>>> xnshadow_relax() faulting before the current thread bears the XNRELAX
>>> bit would mean that a creepy issue involving ondemand PTEs in _kernel_
>>> space must have caused this. Having the init_mm mappings known from all
>>> processes seems more relevant to this issue than anything nocow and/or
>>> mlockall could ever do to fix it.
>> u_mode is a user-space address.
>>
> 
> Why do you think xnshadow_relax() would be called for an already relaxed
> thread?

Because the fault happens before it has finished relaxing ?

-- 
                                          Gilles