Re: [PATCH] zram_drv: add __GFP_NOMEMALLOC not to use ALLOC_NO_WATERMARKS

From: Minchan Kim
Date: Tue Jun 07 2022 - 22:04:03 EST


Hi Jaewon,

On Tue, Jun 07, 2022 at 10:17:02AM +0900, Jaewon Kim wrote:
> >
> >
> >--------- Original Message ---------
> >Sender : Minchan Kim <minchan@xxxxxxxxxx>
> >Date : 2022-06-07 05:48 (GMT+9)
> >Title : Re: [PATCH] zram_drv: add __GFP_NOMEMALLOC not to use ALLOC_NO_WATERMARKS
> >
> >On Mon, Jun 06, 2022 at 12:59:39PM -0700, Andrew Morton wrote:
> >> On Mon, 6 Jun 2022 12:46:38 -0700 Minchan Kim <minchan@xxxxxxxxxx> wrote:
> >>
> >> > On Fri, Jun 03, 2022 at 02:57:47PM +0900, Jaewon Kim wrote:
> >> > > The atomic page allocation failure sometimes happened, and most of them
> >> > > seem to occur during boot time.
> >> > >
> >> > > <4>[ 59.707645] system_server: page allocation failure: order:0, mode:0xa20(GFP_ATOMIC), nodemask=(null),cpuset=foreground-boost,mems_allowed=0
> >> >
> >> > ...
> >> >
> >> > >
> >> > > The kswapd or other reclaim contexts may not prepare enough free pages
> >> > > for too many atomic allocations occurred in short time. But zram may not
> >> > > be helpful for this atomic allocation even though zram is used to
> >> > > reclaim.
> >> > >
> >> > > To get one zs object for a specific size, zram may allocate serveral
> >> > > pages. And this can be happened on different class sizes at the same
> >> > > time. It means zram may consume more pages to reclaim only one page.
> >> > > This inefficiency may consume all free pages below watmerk min by a
> >> > > process having PF_MEMALLOC like kswapd.
> >> >
> >> > However, that's how zram has worked for a long time(allocate memory
> >> > under memory pressure) and many folks already have raised min_free_kbytes
> >> > when they use zram as swap. If we don't allow the allocation, swap out
> >> > fails easier than old, which would break existing tunes.
>
>
> Hello.
>
> Yes correct. We may need to tune again to swap out as much as we did.
>
> But on my experiment, there were quite many zram allocations which might
> be failed unless it has the ALLOC_NO_WATERMARKS. I thought the zram
> allocations seem to be easy to affect atomic allocation failure.

I understand your concern but solution here would affect to existing common
users too much.

>
> >>
> >> So is there a better way of preventing this warning? Just suppress it
> >> with __GFP_NOWARN?
> >
> >For me, I usually tries to remove GFP_ATOMIC alllocation since the
> >atomic allocation can be failed easily(zram is not only source for
> >it). Otherwise, increase min_free_kbytes?
> >
>
> I also hope driver developers to handle this atomic allocation failure.
> However this selinux stuff, context_struct_to_string, is out of their domain.
> Do I need to report this to selinux community? Actualy I got several
> different callpaths to reach this context_struct_to_string.

I am not famliar with selinux stuff but if it's common to see the
GFP_ATOMIC failures in the path, I think it should have __GFP_NOWARN
or other solution to allocate memory in advance.
(BTW, I had similar problem before and fixed it with adding __GFP_NOWARN
648f2c6100cf, selinux: use __GFP_NOWARN with GFP_NOWAIT in the AVC)

>
> Yes we may need to increase min_free_kbytes. But I have an experience where
> changing wmark_min from 4MB to 8MB did not work last year. Could you share
> some advice about size?

I don't think we could have universal golden value for it since every
workload and configuration are different in their system. Maybe,
your zram size is rather big compared to system memory and swappiness
is rather high for boot.