How to avoid or reduce GFP_ATOMIC allocation failed

[greg kh]

On Fri, Mar 08, 2019 at 03:34:18PM +0800, wuzhouhui wrote:
> > -----Original Messages-----
> > From: "Greg KH" <greg at kroah.com>
> > Sent Time: 2019-03-08 15:21:52 (Friday)
> > To: wuzhouhui <wuzhouhui14 at mails.ucas.ac.cn>
> > Cc: kernelnewbies at kernelnewbies.org
> > Subject: Re: How to avoid or reduce GFP_ATOMIC allocation failed
> > 
> > On Fri, Mar 08, 2019 at 01:37:26PM +0800, wuzhouhui wrote:
> > > Hi,
> > > 
> > > I check kernel code and found that GFP_ATOMIC allocation will
> > > use emergency pool and maybe failed if emergency pool is not
> > > enough. And GFP_ATOMIC doesn't trigger reclaim (because of
> > > ATOMIC) even if there are a lot of page caches. So my question
> > > is how to avoid or reduce GFP_ATOMIC allocation failed if there
> > > are enough reclaimable memory? Is there some kernel parameters
> > > can be configured?
> > 
> > Have you seen the ATOMIC pools be used up and not able to be reclaimed
> > in real-world usages?  If so, I'm sure the mm developers would love to
> 
> No, I haven't seen this scenario. But I encountered the similar issue
> with [1] (order is 5 in my scenario), and this issue is not resolved for
> now.

Please work with the company that created your out-of-tree kernel
networking code as it sounds like they do not know how to properly
handle this type of problem :)

good luck!

greg k-h