Re: + x86-mm-handle-mm_fault_error-in-kernel-space.patch added to -mm tree

[Oleg Nesterov <oleg@redhat.com>]

On 03/11, Andrew Vagin wrote:
>
> On 03/11/2011 02:19 PM, Oleg Nesterov wrote:
>>
>> Btw, this may be true, but this is irrelevant. If we shouldn't call
>> out_of_memory() in this case, then we shouldn't call it at all, even
>> if PF_USER.
>
> Yes. We shouldn't call it at all, even if PF_USER.

Then why did you send this patch? If we should not call it, then we
should kill pagefault_out_of_memory() and update the callers instead
of adding the special 'if (PF_USER)' checks.

Yes, the current pagefault_out_of_memory() logic looks a bit suspicious,
but this needs another discussion. Once again, I am arguing against making
it depend on PF_USER, this was my point from the very beginning.

>>> Now pls think what is the
>>> difference between these page faults?
>> The difference is that oom-killer should free the memory in between.
>> _OR_ it can decide to kill us, and _this_ case should be fixed.
>
> We wait memory in __alloc_pages_may_oom(). I think now handle_mm_fault()
> returns VM_FAULT_OOM only if OOM-killer killed current task.

I don't think so, but this doesn't matter.

Once again, if OOM-killer killed current task we do not retry. That is
why my patch checks fatal_signal_pending() to fix the bug. That is all.

The point is, if current was _NOT_ killed we should follow the current
pagefault_out_of_memory() logic or remove pagefault_out_of_memory()
completely.

>> Why do you think the current task should be killed? In this case we
>> do not need oom-killer at all, we could always kill the caller of
>> alloc_page/etc.
>
> You don't understand. alloc_page calls oom-killer himself, then try
> allocate memory again. Pls look at __alloc_pages_slowpath().
> __alloc_pages_slowpat may fail if order > 3 || gfp_mask & __GFP_NOFAIL
> || test_thread_flag(TIF_MEMDIE)

Andrew, please, I know this.

> Probaly you think that oom-killer is called from mm_fault_error() only.
> It's incorrect.

And this too ;)

If nothing else. alloc_page doesn't call oom-killer if it is already
in progress. At least in this case we should retry after it completes.


Either way, I believe this patch should be dropped.

Oleg.