Re: [PATCH mm] vmalloc: back off when the current task is OOM-killed

[Vasily Averin <vvs@virtuozzo.com>]

On 9/21/21 9:55 PM, Andrew Morton wrote:
> On Mon, 20 Sep 2021 13:59:35 +0300 Vasily Averin <vvs@virtuozzo.com> wrote:
> 
>> On 9/20/21 4:22 AM, Tetsuo Handa wrote:
>>> On 2021/09/20 8:31, Andrew Morton wrote:
>>>> On Fri, 17 Sep 2021 11:06:49 +0300 Vasily Averin <vvs@virtuozzo.com> wrote:
>>>>
>>>>> Huge vmalloc allocation on heavy loaded node can lead to a global
>>>>> memory shortage. A task called vmalloc can have the worst badness
>>>>> and be chosen by OOM-killer, however received fatal signal and
>>>>> oom victim mark does not interrupt allocation cycle. Vmalloc will
>>>>> continue allocating pages over and over again, exacerbating the crisis
>>>>> and consuming the memory freed up by another killed tasks.
>>>>>
>>>>> This patch allows OOM-killer to break vmalloc cycle, makes OOM more
>>>>> effective and avoid host panic.
>>>>>
>>>>> Unfortunately it is not 100% safe. Previous attempt to break vmalloc
>>>>> cycle was reverted by commit b8c8a338f75e ("Revert "vmalloc: back off when
>>>>> the current task is killed"") due to some vmalloc callers did not handled
>>>>> failures properly. Found issues was resolved, however, there may
>>>>> be other similar places.
>>>>
>>>> Well that was lame of us.
>>>>
>>>> I believe that at least one of the kernel testbots can utilize fault
>>>> injection.  If we were to wire up vmalloc (as we have done with slab
>>>> and pagealloc) then this will help to locate such buggy vmalloc callers.
>>
>> Andrew, could you please clarify how we can do it?
>> Do you mean we can use exsiting allocation fault injection infrastructure to trigger
>> such kind of issues? Unfortunately I found no ways to reach this goal.
>> It  allows to emulate single faults with small probability, however it is not enough,
>> we need to completely disable all vmalloc allocations. 
> 
> I don't see why there's a problem?  You're saying "there might still be
> vmalloc() callers which don't correctly handle allocation failures",
> yes?
> 
> I'm suggesting that we use fault injection to cause a small proportion
> of vmalloc() calls to artificially fail, so such buggy callers will
> eventually be found and fixed.  Why does such a scheme require that
> *all* vmalloc() calls fail?

Let me explain.
1) it is not trivial to use current allocation fault injection to cause
a small proportion of vmalloc() calls to artificially fail.

vmalloc
 __vmalloc_node
  __vmalloc_node_range
   __vmalloc_area_node
    vm_area_alloc_pages
 
vm_area_alloc_pages uses new __alloc_pages_bulk subsystem, requesting up to 100 pages in cycle.
__alloc_pages_bulk() can be interrupted by allocation fault injection, however in this case
vm_area_alloc_pages() will failback to old-style page allocation cycle.
In general case it successfully finishes allocation and vmalloc itself will not fail.

To fail vmalloc we need to fail both alloc_pages_bulk_array_node() and alloc_pages_node() together.

2) if we failed single vmalloc it is not enough.
I would remind, we want to emulate fatal signal reaction.
However I afraid dying task can execute a quite complex rollback procedure.
This rollback can call another vmalloc and last one will be failed
again on fatal_signal_pending check.

To emulate this behavior in fault injection we need to disable all following
vmalloc calls of our victim, pseudo-"dying" task.

I doubt both these goals can be reached by current allocation fault injection subsystem,
I do not understand how to configure it accordingly.

Thank you,
	Vasily Averin