Re: [PATCH v1] ringbuffer: Don't choose the process with adj equal OOM_SCORE_ADJ_MIN

From: Zhaoyang Huang
Date: Mon Apr 09 2018 - 22:32:42 EST


On Tue, Apr 10, 2018 at 8:32 AM, Zhaoyang Huang <huangzhaoyang@xxxxxxxxx> wrote:
> On Mon, Apr 9, 2018 at 9:49 PM, Steven Rostedt <rostedt@xxxxxxxxxxx> wrote:
>> On Mon, 9 Apr 2018 08:56:01 +0800
>> Zhaoyang Huang <huangzhaoyang@xxxxxxxxx> wrote:
>>
>>> >>
>>> >> if (oom_task_origin(task)) {
>>> >> points = ULONG_MAX;
>>> >> goto select;
>>> >> }
>>> >>
>>> >> points = oom_badness(task, NULL, oc->nodemask, oc->totalpages);
>>> >> if (!points || points < oc->chosen_points)
>>> >> goto next;
>>> >
>>> > And what's wrong with that?
>>> >
>>> > -- Steve
>>> I think the original thought of OOM is the flag 'OOM_SCORE_ADJ_MIN' is
>>> most likely to be set by process himself via accessing the proc file,
>>> if it does so, OOM can select it as the victim. except, it is
>>> reluctant to choose the critical process to be killed, so I suggest
>>> not to set such heavy flag as OOM_SCORE_ADJ_MIN on behalf of -1000
>>> process.
>>
>> Really, I don't think tasks that are setting OOM_CORE_ADJ_MIN should be
>> allocating a lot of memory in the kernel (via ring buffer). It sounds
>> like a good way to wreck havoc on the system.
>>
>> It's basically saying, "I'm going to take up all memory, but don't kill
>> me, just kill some random user on the system".
>>
>> -- Steve
> Sure, but the memory status is dynamic, the process could also exceed the limit
> at the moment even it check the available memory before. We have to
> add protection
> for such kind of risk. It could also happen that the critical process
> be preempted by
> another huge memory allocating process, which may cause insufficient memory when
> it schedule back.

For bellowing scenario, process A have no intension to exhaust the
memory, but will be likely to be selected by OOM for we set
OOM_CORE_ADJ_MIN for it.
process A(-1000) process B

i = si_mem_available();
if (i < nr_pages)
return -ENOMEM;
schedule
--------------->
allocate huge memory
<-------------
if (user_thread)
set_current_oom_origin();

for (i = 0; i < nr_pages; i++) {
bpage = kzalloc_node