On Tue 12-01-21 19:43:21, Muchun Song wrote:

On Tue, Jan 12, 2021 at 7:17 PM Michal Hocko [off-list ref] wrote:

quoted

On Tue 12-01-21 18:13:02, Muchun Song wrote:

quoted

On Tue, Jan 12, 2021 at 6:02 PM Michal Hocko [off-list ref] wrote:

quoted

On Sun 10-01-21 20:40:14, Muchun Song wrote:
[...]

quoted

@@ -1770,6 +1788,14 @@ int dissolve_free_huge_page(struct page *page)
              int nid = page_to_nid(head);
              if (h->free_huge_pages - h->resv_huge_pages == 0)
                      goto out;
+
+             /*
+              * We should make sure that the page is already on the free list
+              * when it is dissolved.
+              */
+             if (unlikely(!PageHugeFreed(head)))
+                     goto out;
+

Do you really want to report EBUSY in this case? This doesn't make much
sense to me TBH. I believe you want to return 0 same as when you race
and the page is no longer PageHuge.

Return 0 is wrong. Because the page is not freed to the buddy allocator.
IIUC, dissolve_free_huge_page returns 0 when the page is already freed
to the buddy allocator. Right?

0 is return when the page is either dissolved or it doesn't need
dissolving. If there is a race with somebody else freeing the page then
there is nothing to dissolve. Under which condition it makes sense to
report the failure and/or retry dissolving?

If there is a race with somebody else freeing the page, the page
can be freed to the hugepage pool not the buddy allocator. Do
you think that this page is dissolved?

OK, I see what you mean. Effectively the page would be in a limbo, not
yet in the pool nor in the allocator but it can find its way to the
either of the two. But I still dislike returning a failure because that
would mean e.g. memory hotplug to fail. Can you simply retry inside this
code path (drop the lock, cond_resched and retry)?
-- 
Michal Hocko
SUSE Labs