bugfix for memory cgroup controller: avoid !PageLRU page in mem_cgroup_isolate_pages
[linux-2.6.git] / mm / memcontrol.c
index 3270ce7..e8493fb 100644 (file)
@@ -250,7 +250,7 @@ unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
        unsigned long scan;
        LIST_HEAD(pc_list);
        struct list_head *src;
-       struct page_cgroup *pc;
+       struct page_cgroup *pc, *tmp;
 
        if (active)
                src = &mem_cont->active_list;
@@ -258,11 +258,18 @@ unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
                src = &mem_cont->inactive_list;
 
        spin_lock(&mem_cont->lru_lock);
-       for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
-               pc = list_entry(src->prev, struct page_cgroup, lru);
+       scan = 0;
+       list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
+               if (scan++ > nr_to_scan)
+                       break;
                page = pc->page;
                VM_BUG_ON(!pc);
 
+               if (unlikely(!PageLRU(page))) {
+                       scan--;
+                       continue;
+               }
+
                if (PageActive(page) && !active) {
                        __mem_cgroup_move_lists(pc, true);
                        scan--;
@@ -492,6 +499,49 @@ void mem_cgroup_uncharge(struct page_cgroup *pc)
                }
        }
 }
+/*
+ * Returns non-zero if a page (under migration) has valid page_cgroup member.
+ * Refcnt of page_cgroup is incremented.
+ */
+
+int mem_cgroup_prepare_migration(struct page *page)
+{
+       struct page_cgroup *pc;
+       int ret = 0;
+       lock_page_cgroup(page);
+       pc = page_get_page_cgroup(page);
+       if (pc && atomic_inc_not_zero(&pc->ref_cnt))
+               ret = 1;
+       unlock_page_cgroup(page);
+       return ret;
+}
+
+void mem_cgroup_end_migration(struct page *page)
+{
+       struct page_cgroup *pc = page_get_page_cgroup(page);
+       mem_cgroup_uncharge(pc);
+}
+/*
+ * We know both *page* and *newpage* are now not-on-LRU and Pg_locked.
+ * And no race with uncharge() routines because page_cgroup for *page*
+ * has extra one reference by mem_cgroup_prepare_migration.
+ */
+
+void mem_cgroup_page_migration(struct page *page, struct page *newpage)
+{
+       struct page_cgroup *pc;
+retry:
+       pc = page_get_page_cgroup(page);
+       if (!pc)
+               return;
+       if (clear_page_cgroup(page, pc) != pc)
+               goto retry;
+       pc->page = newpage;
+       lock_page_cgroup(newpage);
+       page_assign_page_cgroup(newpage, pc);
+       unlock_page_cgroup(newpage);
+       return;
+}
 
 int mem_cgroup_write_strategy(char *buf, unsigned long long *tmp)
 {