Re: [PATCH v4 9/9] mm: thp: always enable mTHP support

[Baolin Wang]

On 5/13/26 11:58 PM, David Hildenbrand (Arm) wrote:
> On 5/1/26 21:18, Luiz Capitulino wrote:
> > If PMD-sized pages are not supported on an architecture (ie. the
> > arch implements arch_has_pmd_leaves() and it returns false) then the
> > current code disables all THP, including mTHP.
> >
> > This commit fixes this by allowing mTHP to be always enabled for all
> > archs. When PMD-sized pages are not supported, its sysfs entry won't be
> > created and their mapping will be disallowed at page-fault time.
> >
> > Similarly, this commit implements the following changes for shmem in
> > shmem_allowable_huge_orders():
> >
> >   - Drop the pgtable_has_pmd_leaves() check so that mTHP sizes are
> >     considered
> >   - Filter out PMD and PUD orders from allowable orders when
> >     PMD-sized pages are not supported by the CPU
> >
> > Signed-off-by: Luiz Capitulino <luizcap@xxxxxxxxxx>
> > ---
> >   mm/huge_memory.c | 23 ++++++++++++++++++-----
> >   mm/shmem.c       | 14 +++++++++-----
> >   2 files changed, 27 insertions(+), 10 deletions(-)
> >
> > diff --git a/mm/huge_memory.c b/mm/huge_memory.c
> > index 32254febe097..c1765c8e3dc6 100644

[snip]

> > diff --git a/mm/shmem.c b/mm/shmem.c
> > index a48f034830cd..23893c2bc2dd 100644
> > --- a/mm/shmem.c
> > +++ b/mm/shmem.c
> > @@ -1840,16 +1840,19 @@ unsigned long shmem_allowable_huge_orders(struct inode *inode,
> >   	unsigned long mask = READ_ONCE(huge_shmem_orders_always);
> >   	unsigned long within_size_orders = READ_ONCE(huge_shmem_orders_within_size);
> >   	vm_flags_t vm_flags = vma ? vma->vm_flags : 0;
> > -	unsigned int global_orders;
> > +	unsigned int global_orders, filter_orders = 0;
> >   
> > -	if (!pgtable_has_pmd_leaves() || (vma && vma_thp_disabled(vma, vm_flags, shmem_huge_force)))
> > +	if (vma && vma_thp_disabled(vma, vm_flags, shmem_huge_force))
> >   		return 0;
> >   
> > +	if (!pgtable_has_pmd_leaves())
> > +		filter_orders = BIT(PMD_ORDER) | BIT(PUD_ORDER);
>
> Would "disabled_orders" or "unavailable_orders" be more appropriate?

Thanks. Better than my naming:)

> There is no need to disable PUD-orders, as shmem does not support PUD-orders
> (unlike DAX). So you can keep it simpler here.

Agree.

> >   	global_orders = shmem_huge_global_enabled(inode, index, write_end,
> >   						  shmem_huge_force, vma, vm_flags);
> >   	/* Tmpfs huge pages allocation */
> >   	if (!vma || !vma_is_anon_shmem(vma))
> > -		return global_orders;
> > +		return global_orders & ~filter_orders;
> >   
> >   	/*
> >   	 * Following the 'deny' semantics of the top level, force the huge
> > @@ -1863,7 +1866,7 @@ unsigned long shmem_allowable_huge_orders(struct inode *inode,
> >   	 * means non-PMD sized THP can not override 'huge' mount option now.
> >   	 */
> >   	if (shmem_huge == SHMEM_HUGE_FORCE)
> > -		return READ_ONCE(huge_shmem_orders_inherit);
> > +		return READ_ONCE(huge_shmem_orders_inherit) & ~filter_orders;
> >   
> >   	/* Allow mTHP that will be fully within i_size. */
> >   	mask |= shmem_get_orders_within_size(inode, within_size_orders, index, 0);
> > @@ -1874,6 +1877,7 @@ unsigned long shmem_allowable_huge_orders(struct inode *inode,
> >   	if (global_orders > 0)
> >   		mask |= READ_ONCE(huge_shmem_orders_inherit);
> >   
> > +	mask &= ~filter_orders;
> >   	return THP_ORDERS_ALL_FILE_DEFAULT & mask;
> >   }
> >   
> > @@ -5457,7 +5461,7 @@ void __init shmem_init(void)
> >   	 * Default to setting PMD-sized THP to inherit the global setting and
> >   	 * disable all other multi-size THPs.
> >   	 */
> > -	if (!shmem_orders_configured)
> > +	if (!shmem_orders_configured && pgtable_has_pmd_leaves())
> >   		huge_shmem_orders_inherit = BIT(HPAGE_PMD_ORDER);
>
> Do we really want to change that? We can just leave the defaults as is, no?

Yes, this needs to be handled the same way as anonymous memory. If 
pgtable_has_pmd_leaves() returns false, but huge_shmem_orders_inherit 
still has the PMD order set by default, then shmem would still be able 
to allocate PMD-sized large folios if the global shmem interface is enabled.