Re: Widespread crashes in -next, bisected to 'mm: drop HASH_ADAPT'

[Guenter Roeck]

On 05/22/2017 01:45 AM, Michal Hocko wrote:
> On Sat 20-05-17 09:26:34, Michal Hocko wrote:
> > On Fri 19-05-17 09:46:23, Guenter Roeck wrote:
> > > Hi,
> > >
> > > my qemu tests of next-20170519 show the following results:
> > > 	total: 122 pass: 30 fail: 92
> > >
> > > I won't bother listing all of the failures; they are available at
> > > http://kerneltests.org/builders. I bisected one (openrisc, because
> > > it gives me some console output before dying). It points to
> > > 'mm: drop HASH_ADAPT' as the culprit. Bisect log is attached.
> > >
> > > A quick glance suggests that 64 bit kernels pass and 32 bit kernels fail.
> > > 32-bit x86 images fail and should provide an easy test case.
> >
> > Hmm, this is quite unexpected as the patch is not supposed to change
> > things much. It just removes the flag and perform the new hash scaling
> > automatically for all requeusts which do not have any high limit.
> > Some of those didn't have HASH_ADAPT before but that shouldn't change
> > the picture much. The only thing that I can imagine is that what
> > formerly failed for early memblock allocations is now suceeding and that
> > depletes the early memory. Do you have any serial console from the boot?
>
> OK, I guess I know what it going on here. Adaptive has scaling is not
> really suited for 32b. ADAPT_SCALE_BASE is just too large for the word
> size and so we end up in the endless loop. So the issue has been
> introduced by the original "mm: adaptive hash table scaling" but my
> patch made it more visible because [di]cache has tables most probably
> suceeded in the early initialization which didn't have HASH_ADAPT.
> The following should fix the hang. I am not yet sure about the maximum
> size for the scaling and something even smaller would make sense to me
> because kernel address space is just too small for such a large has
> tables.
> ---
> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> index a26e19c3e1ff..70c5fc1fb89a 100644
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -7174,11 +7174,15 @@ static unsigned long __init arch_reserved_kernel_pages(void)
>   /*
>    * Adaptive scale is meant to reduce sizes of hash tables on large memory
>    * machines. As memory size is increased the scale is also increased but at
> - * slower pace.  Starting from ADAPT_SCALE_BASE (64G), every time memory
> - * quadruples the scale is increased by one, which means the size of hash table
> - * only doubles, instead of quadrupling as well.
> + * slower pace.  Starting from ADAPT_SCALE_BASE (64G on 64b systems and 32M
> + * on 32b), every time memory quadruples the scale is increased by one, which
> + * means the size of hash table only doubles, instead of quadrupling as well.
>    */
> +#if __BITS_PER_LONG == 64
>   #define ADAPT_SCALE_BASE	(64ul << 30)
> +#else
> +#define ADAPT_SCALE_BASE	(32ul << 20)
> +#endif
>   #define ADAPT_SCALE_SHIFT	2
>   #define ADAPT_SCALE_NPAGES	(ADAPT_SCALE_BASE >> PAGE_SHIFT)
>   
I have seen another patch making it 64ull. Not sure if adaptive scaling
on 32 bit systems really makes sense; unless there is a clear need I'd rather
leave it alone.

Guenter