Linux lru_cache_add
2019-03-26
lru_cache_add Queue the page for addition to the LRU via pagevec. 是否加入[in]active [file|anon] 列表会被推迟到pagevec被drain. 这样就 可以给lru_cache_add的调用者一个机会通过mark_page_accessed来把page 重新加入active list.
436 void lru_cache_add(struct page *page)
437 {
438 VM_BUG_ON_PAGE(PageActive(page) && PageUnevictable(page), page);
439 VM_BUG_ON_PAGE(PageLRU(page), page);
440 __lru_cache_add(page);
441 }
438行，如果page是active和unevitable状态，表明出现了bug，unevictable表明不 可以被回收
439行，page已经在LRU list里面，重复加入也表明出现bug
__lru_cache_add是把page加入pagevec,如果pagevec满了，把pagevec加入LRU,
如果page是PageCompound则不判断pagevec是否已满直接drain pagevc
398 static void __lru_cache_add(struct page *page)
399 {
400 struct pagevec *pvec = &get_cpu_var(lru_add_pvec);
401
402 get_page(page);
403 if (!pagevec_add(pvec, page) || PageCompound(page))
404 __pagevec_lru_add(pvec);
405 put_cpu_var(lru_add_pvec);
406 }
400,405行是获取和释放lru_add_pvec percpu变量, get_cpu_var 和 put_cpu_var 会禁止当前cpu的调度，这是必须的，不然进程迁移会导致错误的percpu变量访问.
403行，把page加入到pagevec中，如果是compound page或者pagevec已满则直接加入lru
404行__pagevec_lru_add,遍历pvec中的page加入page所在的pgdata的lrurec list
914 /*
915 * Add the passed pages to the LRU, then drop the caller's refcount
916 * on them. Reinitialises the caller's pagevec.
917 */
918 void __pagevec_lru_add(struct pagevec *pvec)
919 {
920 pagevec_lru_move_fn(pvec, __pagevec_lru_add_fn, NULL);
921 }
922 EXPORT_SYMBOL(__pagevec_lru_add);
这里我们不考虑memcg的话,mem_cgroup_page_lruvec(page, pgdat);直接返回
当前pgdat的lrurec list/array,核心函数当然是__pagevec_lru_add_fn
859 static void __pagevec_lru_add_fn(struct page *page, struct lruvec *lruvec,
860 void *arg)
861 {
862 enum lru_list lru;
863 int was_unevictable = TestClearPageUnevictable(page);
864
865 VM_BUG_ON_PAGE(PageLRU(page), page);
866
867 SetPageLRU(page);
868 /*
869 * Page becomes evictable in two ways:
870 * 1) Within LRU lock [munlock_vma_pages() and __munlock_pagevec()].
871 * 2) Before acquiring LRU lock to put the page to correct LRU and then
872 * a) do PageLRU check with lock [check_move_unevictable_pages]
873 * b) do PageLRU check before lock [clear_page_mlock]
874 *
875 * (1) & (2a) are ok as LRU lock will serialize them. For (2b), we need
876 * following strict ordering:
877 *
878 * #0: __pagevec_lru_add_fn #1: clear_page_mlock
879 *
880 * SetPageLRU() TestClearPageMlocked()
881 * smp_mb() // explicit ordering // above provides strict
882 * // ordering
883 * PageMlocked() PageLRU()
884 *
885 *
886 * if '#1' does not observe setting of PG_lru by '#0' and fails
887 * isolation, the explicit barrier will make sure that page_evictable
888 * check will put the page in correct LRU. Without smp_mb(), SetPageLRU
889 * can be reordered after PageMlocked check and can make '#1' to fail
890 * the isolation of the page whose Mlocked bit is cleared (#0 is also
891 * looking at the same page) and the evictable page will be stranded
892 * in an unevictable LRU.
893 */
894 smp_mb();
895
896 if (page_evictable(page)) {
897 lru = page_lru(page);
898 update_page_reclaim_stat(lruvec, page_is_file_cache(page),
899 PageActive(page));
900 if (was_unevictable)
901 count_vm_event(UNEVICTABLE_PGRESCUED);
902 } else {
903 lru = LRU_UNEVICTABLE;
904 ClearPageActive(page);
905 SetPageUnevictable(page);
906 if (!was_unevictable)
907 count_vm_event(UNEVICTABLE_PGCULLED);
908 }
909
910 add_page_to_lru_list(page, lruvec, lru);
911 trace_mm_lru_insertion(page, lru);
912 }
863行判断当前page was_evictable, was表示之前,不表示现在, 并且清掉这个标志位.
865行如果page已经设置LRU状态,表明有bug,因为page已经在LRU链表里了,再次加入LRU 肯定有bug.
867行设置page的LRU标志
894行smp_mb,为什么要加呢？防止编译器和处理器乱序访问.根据注释,如果#1没有 观察到#0处的LRU置位,显示的barrier会确保page#evictable 检查可以把page放入 正确的LRU.如果没有smp#mb,SetPageLRU可能会被重排到PageMlocked之后,这样#1 处page ioslation会失败,这是#0处也同时在访问这个page的flags,那么就会把 evictable page放入unevictable LRU,所以需要显示的smp_mb.
896行如果page是evictable,page_lru是根据page来找到这个page应该被放在哪个LRU list
902行else代码段,表示page不能被evict,清active标志,置位Unevictable标志
910行把page加入对应的lru链表