Re: [BISECTED][REGRESSION] 5.10.56 longterm kernel breakage on m68k/aranym

[Mikael Pettersson]

On Tue, Aug 10, 2021 at 5:59 PM Mike Rapoport <rppt@xxxxxxxxxx> wrote:
>
> On Mon, Aug 09, 2021 at 03:40:04PM +0200, Geert Uytterhoeven wrote:
> > CC Mike
> >
> > On Mon, Aug 9, 2021 at 3:32 PM Mikael Pettersson <mikpelinux@xxxxxxxxx> wrote:
> > > On Mon, Aug 9, 2021 at 3:59 AM Finn Thain <fthain@xxxxxxxxxxxxxx> wrote:
> > > > On Sun, 8 Aug 2021, Mikael Pettersson wrote:
> > > > > On Sun, Aug 8, 2021 at 1:20 AM Finn Thain <fthain@xxxxxxxxxxxxxx> wrote:
> > > > > > On Sat, 7 Aug 2021, Mikael Pettersson wrote:
> > > > > > > I updated the 5.10 longterm kernel on one of my m68k/aranym VMs from
> > > > > > > 5.10.47 to 5.10.56, and the new kernel failed to boot:
> > > > > > >
> > > > > > > ARAnyM 1.1.0
> > > > > > > Using config file: 'aranym1.headless.config'
> > > > > > > Could not open joystick 0
> > > > > > > ARAnyM RTC Timer: /dev/rtc: Permission denied
> > > > > > > ARAnyM LILO: Error loading ramdisk 'root.bin'
> > > > > > > Blitter tried to read byte from register ff8a00 at 0077ee
> > > > > > >
> > > > > > > At this point it kept running, but produced no output to the console,
> > > > > > > and would never get to the point of starting user-space. Attaching gdb
> > > > > > > to aranym showed nothing interesting, i.e. it seemed to be executing
> > > > > > > normally.
> > >
> > > My initial bisect was wrong. I tried reverting 8f34f1eac382 from
> > > 5.10.57 but that made no difference, so I re-ran the git bisect with
> > > all known good points pre-marked. This landed on:
> > > # first bad commit: [ce6ee46e0f39ed97e23ebf7b5a565e0266a8a1a3]
> > > mm/page_alloc: fix memory map initialization for descending nodes
> > >
> > > Reverting _that_ from 5.10.57 does unbreak that kernel.
>
> Indeed there is a problem with that commit in 5.10. The memmap
> initialization relies on availability of zone_to_nid() to link struct page
> to a node. But in 5.10 zone_to_nid() is only defined for NUMA, but not for
> DISCONTIGMEM.
>
> Mikael, can you please try the patch below:
>
> diff --git a/include/linux/mmzone.h b/include/linux/mmzone.h
> index 9d0c454d23cd..63b550403317 100644
> --- a/include/linux/mmzone.h
> +++ b/include/linux/mmzone.h
> @@ -445,7 +445,7 @@ struct zone {
>          */
>         long lowmem_reserve[MAX_NR_ZONES];
>
> -#ifdef CONFIG_NUMA
> +#ifdef CONFIG_NEED_MULTIPLE_NODES
>         int node;
>  #endif
>         struct pglist_data      *zone_pgdat;
> @@ -896,7 +896,7 @@ static inline bool populated_zone(struct zone *zone)
>         return zone->present_pages;
>  }
>
> -#ifdef CONFIG_NUMA
> +#ifdef CONFIG_NEED_MULTIPLE_NODES
>  static inline int zone_to_nid(struct zone *zone)
>  {
>         return zone->node;
>
> --
> Sincerely yours,
> Mike.

Applying this on top of 5.10.57 fixes the problem, thanks.

Tested-by: Mikael Pettersson <mikpelinux@xxxxxxxxx>