Re: [PATCH v5 14/14] mm: thp: always enable mTHP support

[Baolin Wang <baolin.wang@linux.alibaba.com>]

On 6/3/26 2:47 PM, Lance Yang wrote:
> Hi Luiz,
> 
> SHMEM_HUGE_FORCE still assumes PMD order in a few places.
> 
> Is that expected?
> 
> shmem_init() only sets the default inherit mask when PMD leaves are
> available.
> 
> 	if (!shmem_orders_configured && pgtable_has_pmd_leaves())
> 		huge_shmem_orders_inherit = BIT(HPAGE_PMD_ORDER);
> 
> But shmem_parse_huge() rejects "force" unless the mask is exactly PMD
> order.
> 
> 	if (huge == SHMEM_HUGE_FORCE &&
> 	    huge_shmem_orders_inherit != BIT(HPAGE_PMD_ORDER))
> 		return -EINVAL;
> 
> Even if "force" is selected, shmem_huge_global_enabled() still return
> only PMD order.
> 
> 	if (shmem_huge_force || shmem_huge == SHMEM_HUGE_FORCE)
> 		return maybe_pmd_order;
> 
> and shmem_allowable_huge_orders() mask it out.
> 
> 	if (!pgtable_has_pmd_leaves())
> 		disabled_orders = BIT(PMD_ORDER);
> 
> 	if (!vma || !vma_is_anon_shmem(vma))
> 		return global_orders & ~disabled_orders;
> 
> For anon shmem, it can also return 0 for same reason.
> 
> 	if (shmem_huge == SHMEM_HUGE_FORCE)
> 		return READ_ONCE(huge_shmem_orders_inherit) & ~disabled_orders;
> 
> Should SHMEM_HUGE_FORCE use the available mTHP orders below PMD when
> pgtable_has_pmd_leaves() is false?

No. That's exactly the strategy we agreed on after lots of discussion:
1. The options 'force' and 'deny' are dropped for mTHP interfaces, which 
are rather testing artifacts from the old ages.

2. If top level value is 'force', then only PMD-sized hugepages have 
enabled="inherit", otherwise configuration will be failed and vice versa.

You can check commit 4b98995530b7 ("mm: shmem: add multi-size THP sysfs 
interface for anonymous shmem") for more information.