[PATCH v6 4/9] mm: thp: Introduce anon_orders and anon_always_mask sysfs files

[Michael Ellerman]

Andrew Morton <akpm at linux-foundation.org> writes:
> On Fri, 29 Sep 2023 12:44:15 +0100 Ryan Roberts <ryan.roberts at arm.com> wrote:
>
>> In preparation for adding support for anonymous large folios that are
>> smaller than the PMD-size, introduce 2 new sysfs files that will be used
>> to control the new behaviours via the transparent_hugepage interface.
>> For now, the kernel still only supports PMD-order anonymous THP, so when
>> reading back anon_orders, it will reflect that. Therefore there are no
>> behavioural changes intended here.
>
> powerpc strikes again.  ARCH=powerpc allmodconfig:
>
>
> In file included from ./include/linux/bits.h:6,
>                  from ./include/linux/ratelimit_types.h:5,
>                  from ./include/linux/printk.h:9,
>                  from ./include/asm-generic/bug.h:22,
>                  from ./arch/powerpc/include/asm/bug.h:116,
>                  from ./include/linux/bug.h:5,
>                  from ./include/linux/mmdebug.h:5,
>                  from ./include/linux/mm.h:6,
>                  from mm/huge_memory.c:8:
> ./include/vdso/bits.h:7:33: error: initializer element is not constant
>     7 | #define BIT(nr)                 (UL(1) << (nr))
>       |                                 ^
> mm/huge_memory.c:77:47: note: in expansion of macro 'BIT'
>    77 | unsigned int huge_anon_orders __read_mostly = BIT(PMD_ORDER);
>       |                                               ^~~
>
> We keep tripping over this.  I wish there was a way to fix it.

I can't think of any solution, other than ripping the code out.

To catch it earlier we'd need a generic compile-time test that all values
derived from the page table geometry are only used in places that don't
require a constant. I can't think of a way to write a test for that.

Or submitters could compile-test for powerpc - one can dream :D

> Style whine: an all-caps identifier is supposed to be a constant,
> dammit.
>
> 	#define PTE_INDEX_SIZE  __pte_index_size
>
> Nope.

I agree it's ugly. It was done that way because PTE_INDEX_SIZE used to
be constant, and still is for 32-bit PPC and 64-bit Book3E PPC.

We could rename PTE_INDEX_SIZE itself, but we'd still have eg.
PTE_TABLE_SIZE which is used in generic code, and which would be
sometimes constant and sometimes not for different powerpc subarches.

> I did this:
>
> --- a/mm/huge_memory.c~mm-thp-introduce-anon_orders-and-anon_always_mask-sysfs-files-fix
> +++ a/mm/huge_memory.c
> @@ -74,7 +74,7 @@ static unsigned long deferred_split_scan
>  static atomic_t huge_zero_refcount;
>  struct page *huge_zero_page __read_mostly;
>  unsigned long huge_zero_pfn __read_mostly = ~0UL;
> -unsigned int huge_anon_orders __read_mostly = BIT(PMD_ORDER);
> +unsigned int huge_anon_orders __read_mostly;
>  static unsigned int huge_anon_always_mask __read_mostly;
>  
>  /**
> @@ -528,6 +528,9 @@ static int __init hugepage_init_sysfs(st
>  {
>  	int err;
>  
> +	/* powerpc's PMD_ORDER isn't a compile-time constant */
> +	huge_anon_orders = BIT(PMD_ORDER);
> +
>  	*hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
>  	if (unlikely(!*hugepage_kobj)) {
>  		pr_err("failed to create transparent hugepage kobject\n");
> _
>
>
> I assume this is set up early enough.

Yes it should be.

> I don't know why powerpc's PTE_INDEX_SIZE is variable.

To allow a single vmlinux to boot using either the Hashed Page Table
MMU, or Radix Tree MMU, which have different page table geometry.

That's a pretty crucial feature for distros, so that they can build a
single kernel to boot on Power8/9/10.

cheers