Re: [PATCH] mm/memory-failure: allow memory allocation from emergency reserves

[Andrew Morton]

On Sat, 29 Jun 2024 10:09:46 +0800 Miaohe Lin <linmiaohe@xxxxxxxxxx> wrote:

> On 2024/6/25 10:23, Rui Qi wrote:
> > From: Rui Qi <qirui.001@xxxxxxxxxxxxx>
> > 
> > we hope that memory errors can be successfully handled quickly, using
> > __GFP_MEMALLOC can help us improve the success rate of processing
> 
> Comments of __GFP_MEMALLOC says:
> 
>  * Users of this flag have to be extremely careful to not deplete the reserve
>  * completely and implement a throttling mechanism which controls the
>  * consumption of the reserve based on the amount of freed memory.
> 
> It seems there's no such throttling mechanism in memory_failure.
> 
> > under memory pressure, because to_kill struct is freed very quickly,
> > so using __GFP_MEMALLOC will not exacerbate memory pressure for a long time,
> > and  more memory will be freed after killed task exiting, which will also
> 
> Tasks might not be killed even to_kill struct is allocated.
> 
> ...
>
> > -	raw_hwp = kmalloc(sizeof(struct raw_hwp_page), GFP_ATOMIC);
> > +	raw_hwp = kmalloc(sizeof(struct raw_hwp_page), GFP_ATOMIC | __GFP_MEMALLOC);
> 
> In already hardware poisoned code path, raw_hwp can be allocated to store raw page info
> without killing anything. So __GFP_MEMALLOC might not be suitable to use.
> Or am I miss something?

Yes, I'm doubtful about this patch.  I think that rather than poking at a
particular implementation, it would be helpful for us to see a complete
description of the issues which were observed, please.  Let's see the
bug report and we can discuss fixes later.