Re: [PATCH v2 2/2] mm/memory_hotplug: fix hwpoisoned large folio handling in do_migrate_range

[Zi Yan <ziy@nvidia.com>]

On 14 Jul 2025, at 11:33, David Hildenbrand wrote:

> On 14.07.25 17:28, Zi Yan wrote:
>> On 14 Jul 2025, at 11:25, Zi Yan wrote:
>>
>>> On 14 Jul 2025, at 11:14, David Hildenbrand wrote:
>>>
>>>> On 14.07.25 17:09, Pankaj Raghav (Samsung) wrote:
>>>>>>>>> So we will need to take care of madvise cold or pageout case?
>>>>>>>>>
>>>>>>>>> Hi Matthew, Pankaj, and Luis,
>>>>>>>>>
>>>>>>>>> Is it possible to partially map a min-order folio in a fs with LBS? Based on my
>>>>>>>>
>>>>>>>> Typically, FSs match the min order with the blocksize of the filesystem.
>>>>>>>> As a filesystem block is the smallest unit of data that the filesystem uses
>>>>>>>> to store file data on the disk, we cannot partially map them.
>>>>>>>>
>>>>>>>> So if I understand your question correctly, the answer is no.
>>>>>>
>>>>>> I'm confused. Shouldn't this be trivially possible?
>>>>>>
>>>>> Hmm, maybe I misunderstood the question?
>>>>>
>>>>>> E.g., just mmap() a single page of such a file? Who would make that fail?
>>>>>>
>>>>>
>>>>> My point was, even if you try to mmap a single page of a file, page
>>>>> cache will read the whole block (that corresponds to min order folio).
>>>>>
>>>>> Technically we can mmap a single page of file, but FS will always read
>>>>> and write **at least** in min folio order chunks.
>>>>
>>>> Okay, so it can be partially mapped into page tables :) What happens in the background (page cache management) is a different story
>>>
>>> David, thanks for getting to the bottom of this.
>>>
>>> OK. So we will see deadlock looping in madvise cold or pageout case.
>>> I wonder how to proceed with this. Since the folio is seen as a whole
>>> by fs, it should be marked cold/paged out as a whole. Maybe we should
>>> skip the partially mapped region?
>>
>> Actually, it is skipped, since split_folio() bumps new_order to the min
>> order, and if the folio order is already at min order, split code return
>> -EINVAL. This makes the madvise cold or pageout code move to the next
>> address.
>
> But what if the folio order is 2x min_order etc?

If folio order is greater than min_order, the code is suboptimal.
It will first split the original folio to min_order, then
loop through after-split folios. If the after-split ones are fully
mapped, madvise ops will be performed. Otherwise, the code will
try to split after-split ones again, fail, and end up with skipping
the partially mapped range.

An improvement can be done by skipping the folio if it is partially
mapped and its order is already min_order. Something like:

diff --git a/mm/madvise.c b/mm/madvise.c
index e61e32b2cd91..545ab7920a81 100644
--- a/mm/madvise.c
+++ b/mm/madvise.c
@@ -499,6 +499,13 @@ static int madvise_cold_or_pageout_pte_range(pmd_t *pmd,
 			if (nr < folio_nr_pages(folio)) {
 				int err;

+				/*
+				 * Skip partially mapped folios that are
+				 * already at their min order
+				 */
+				if (folio_order(folio) ==
+				    min_order_for_split(folio))
+					continue;
 				if (folio_maybe_mapped_shared(folio))
 					continue;
 				if (pageout_anon_only_filter && !folio_test_anon(folio))




Best Regards,
Yan, Zi