Re: [PATCH v10 1/5] mm: introduce a common interface for balloon pages mobility

On Tue, Sep 25, 2012 at 03:05:49AM +0200, Michael S. Tsirkin wrote:

If these are all under page lock these barriers just confuse things,
because they are almost never enough by themselves.
So in that case it would be better to drop them and document
usage as you are going to.

Would the following make more sense (with the proprer comments, as well) ?

---8<---
+static inline void balloon_page_set(struct page *page,
+                                   struct address_space *mapping,
+                                   struct list_head *head)
+{
+       list_add(&page->lru, head);
+       smp_wmb();
+       page->mapping = mapping;
+}
+
+static inline void balloon_page_del(struct page *page)
+{
+       page->mapping = NULL;
+       smp_wmb();
+       list_del(&page->lru);
+}
+
+static inline bool __is_movable_balloon_page(struct page *page)
+{
+       struct address_space *mapping = ACCESS_ONCE(page->mapping);
+       smp_read_barrier_depends();
+       return mapping_balloon(mapping);
+}
+
---8<---

There's still a case where we have to test page->mapping->flags and we cannot
afford to wait for, or grab, the page lock @ isolate_migratepages_range().
The barriers won't avoid leak_ballon() racing against isolate_migratepages_range(),
but they surely will make tests for page->mapping more consistent. And for those
cases where leak_balloon() races against
isolate_migratepages_range->isolate_balloon_page(), we solve the conflict of
interest through page refcounting and page lock. I'm preparing a more extensive
doc to include at Documentation/ to explain the interfaces and how we cope with
these mentioned races, as well.