Re: [PATCH v2 1/2] mm: huge_memory: disallow hugepages if the system-wide THP sysfs settings are disabled

[Baolin Wang]

On 2025/6/7 20:21, Lorenzo Stoakes wrote:
> A couple follow up points that occurred to me:
>
> On Sat, Jun 07, 2025 at 12:55:19PM +0100, Lorenzo Stoakes wrote:
> > Not related to your patch at all, but man this whole thing (thp allowed orders)
> > needs significant improvement, it seems always perversely complicated for a
> > relatively simple operation.
> >
> > Overall I LOVE what you're doing here, but I feel we can clarify things a
> > little while we're at it to make it clear exactly what we're doing.
> >
> > This is a very important change so forgive my fiddling about here but I'm
> > hoping we can take the opportunity to make things a little simpler!
> >
> > On Thu, Jun 05, 2025 at 04:00:58PM +0800, Baolin Wang wrote:
> > > The MADV_COLLAPSE will ignore the system-wide Anon THP sysfs settings, which
> > > means that even though we have disabled the Anon THP configuration, MADV_COLLAPSE
> > > will still attempt to collapse into a Anon THP. This violates the rule we have
> > > agreed upon: never means never.
> > >
> > > Another rule for madvise, referring to David's suggestion: “allowing for collapsing
> > > in a VM without VM_HUGEPAGE in the "madvise" mode would be fine".
> >
> > I'm generally not sure it's worth talking only about MADV_COLLAPSE here when
> > you're changing what THP is permitted across the board, I may have missed some
> > discussion and forgive me if so, but what is special about MADV_COLLAPSE's use
> > of thp_vma_allowable_orders() that makes it ignore 'never's moreso than other
> > users?
>
> I'd mention that MADV_COLLAPSE is special because of not specifying
> TVA_ENFORCE_SYSFS but you are making this change across the board for all
> callers who do not specify this.

Currently, besides MADV_COLLAPSE not setting TVA_ENFORCE_SYSFS, there is 
only one other instance where TVA_ENFORCE_SYSFS is not set, which is in 
the collapse_pte_mapped_thp() function, but I believe this is reasonable 
from the comments:

        /*
         * If we are here, we've succeeded in replacing all the native 
pages
         * in the page cache with a single hugepage. If a mm were to 
fault-in
         * this memory (mapped by a suitably aligned VMA), we'd get the 
hugepage
         * and map it by a PMD, regardless of sysfs THP settings. As 
such, let's
         * analogously elide sysfs THP settings here.
         */
        if (!thp_vma_allowable_order(vma, vma->vm_flags, 0, PMD_ORDER))
                return SCAN_VMA_CHECK;

> I'd also CLEARLY mention that you handle David's request re: madvise by
> restricting yourself to checking only for NEVER and retaining the existing logic
> of not enforcing sysfs settings when TVA_ENFORCE_SYSFS, which includes not
> checking the VMA for VM_HUGEPAGE if the madvise mode is enabled.

Sure.

> (i.e. addressing David's request).
>
> [snip]
>
> > I feel this is compressing a lot of logic in a way that took me several
> > readings to understand (hey I might not be the smartest cookie in the jar,
> > but we need to account for all levels of kernel developer ;)
> >
> > I feel like we can make things a lot clearer here by separating out with a
> > helper function (means we can drop some indentation too), and also take
> > advantage of the fact that, if orders == 0, __thp_vma_allowable_orders()
> > exits with 0 early so no need for us to do so ourselves:
> >
> > /* Strictly mask requested anonymous orders according to sysfs settings. */
> > static inline unsigned long __thp_mask_anon_orders(unsigned long vm_flags,
> > 		unsigned long tva_flags, unsigned long orders)
> > {
> > 	unsigned long always = READ_ONCE(huge_anon_orders_always);
> > 	unsigned long madvise = READ_ONCE(huge_anon_orders_madvise);
> > 	unsigned long inherit = READ_ONCE(huge_anon_orders_inherit);;
> > 	bool inherit_enabled = hugepage_global_enabled();
> > 	bool has_madvise =  vm_flags & VM_HUGEPAGE;
> > 	unsigned long mask = always | madvise;
> >
> > 	mask = always | madvise;
> > 	if (inherit_enabled)
> > 		mask |= inherit;
> >
> > 	/* All set to/inherit NEVER - never means never globally, abort. */
> > 	if (!(mask & orders))
> > 		return 0;
> >
> > 	/* Otherwise, we only enforce sysfs settings if asked. */
>
> Perhaps worth adding a comment here noting that, if the user sets a sysfs mode
> of madvise and if TVA_ENFORCE_SYSFS is not set, we don't bother checking whether
> the VMA has VM_HUGEPAGE set.

Sure, will do. Thanks for reviewing.