Re: [RFC PATCH v1 09/37] KVM: guest_memfd: Skip LRU for guest_memfd folios

[Ackerley Tng <ackerleytng@google.com>]

Vlastimil Babka <vbabka@suse.cz> writes:

> On 10/17/25 22:11, Ackerley Tng wrote:
>> filemap_add_folio(), called from filemap_grab_folio(), adds folios to
>> an LRU list. This is unnecessary for guest_memfd, which does not
>> participate in swapping.
>
> IIRC guest_memfd mappings are unevictable. That should mean they are not
> ultimately added to a list (see lruvec_add_folio()).
>
>> In addition, the LRU list takes a reference count on the folio. With
>
> IIUC the refcount is temporary while being on the percpu
> &cpu_fbatches.lru_add, added by __folio_batch_add_and_move().

Thanks for pointing this out. You're right about this, I misunderstood
this refcounting earlier.

> When flushed
> via folio_batch_move_lru(), the refcount is removed and there's only the LRU
> folio flag that remains. The fbatch flushing can be triggered if you see an
> unexpected refcount increase.

The new plan is, to update kvm_gmem_is_safe_for_conversion() to drain
the fbatch if it some elevated refcount is found:

static bool kvm_gmem_is_safe_for_conversion(struct inode *inode,
					    pgoff_t start, size_t nr_pages,
					    pgoff_t *err_index)
{
	struct address_space *mapping = inode->i_mapping;
	const int filemap_get_folios_refcount = 1;
	pgoff_t last = start + nr_pages - 1;
	struct folio_batch fbatch;
	bool lru_drained = false;
	bool safe = true;
	int i;

	folio_batch_init(&fbatch);
	while (safe && filemap_get_folios(mapping, &start, last, &fbatch)) {

		for (i = 0; i < folio_batch_count(&fbatch);) {
			struct folio *folio = fbatch.folios[i];

			safe = (folio_ref_count(folio) ==
				folio_nr_pages(folio) +
				filemap_get_folios_refcount);

			if (safe) {
				++i;
			} else if (!lru_drained) {
				lru_add_drain_all();
				lru_drained = true;
			} else {
				*err_index = folio->index;
				break;
			}
		}

		folio_batch_release(&fbatch);
	}

	return safe;
}

I hope this is what you meant!

> So it might be feasible to do without this
> patch (maybe it was already tried and there were substantial issues, in
> which case should be mentioned).
>

The patch "KVM: guest_memfd: Skip LRU for guest_memfd folios" will be
dropped from the next revision, and "KVM: guest_memfd: Don't set
FGP_ACCESSED when getting folios" is no longer a requirement for this
patch series.

>> shared-to-private memory conversions for KVM guests dependent on folio
>> refcounts, this extra reference can cause conversions to fail due to
>> unexpected refcounts.
>>
>>
>> [...snip...]
>>