memcg: fix deadlock by inverting lrucare nesting
Hugh Dickins [Mon, 5 Mar 2012 22:59:16 +0000 (14:59 -0800)]
We have forgotten the rules of lock nesting: the irq-safe ones must be
taken inside the non-irq-safe ones, otherwise we are open to deadlock:

CPU0                          CPU1
----                          ----
lock(&(&pc->lock)->rlock);
                              local_irq_disable();
                              lock(&(&zone->lru_lock)->rlock);
                              lock(&(&pc->lock)->rlock);
<Interrupt>
lock(&(&zone->lru_lock)->rlock);

To check a different locking issue, I happened to add a spin_lock to
memcg's bit_spin_lock in lock_page_cgroup(), and lockdep very quickly
complained about __mem_cgroup_commit_charge_lrucare() (on CPU1 above).

So delete __mem_cgroup_commit_charge_lrucare(), passing a bool lrucare to
__mem_cgroup_commit_charge() instead, taking zone->lru_lock under
lock_page_cgroup() in the lrucare case.

The original was using spin_lock_irqsave, but we'd be in more trouble if
it were ever called at interrupt time: unconditional _irq is enough.  And
ClearPageLRU before del from lru, SetPageLRU before add to lru: no strong
reason, but that is the ordering used consistently elsewhere.

Fixes 36b62ad539498d00c2d280a151a ("memcg: simplify corner case handling
of LRU").

Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Konstantin Khlebnikov <khlebnikov@openvz.org>
Acked-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm/memcontrol.c

index 228d646..1097d80 100644 (file)
@@ -2408,8 +2408,12 @@ static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
                                       struct page *page,
                                       unsigned int nr_pages,
                                       struct page_cgroup *pc,
-                                      enum charge_type ctype)
+                                      enum charge_type ctype,
+                                      bool lrucare)
 {
+       struct zone *uninitialized_var(zone);
+       bool was_on_lru = false;
+
        lock_page_cgroup(pc);
        if (unlikely(PageCgroupUsed(pc))) {
                unlock_page_cgroup(pc);
@@ -2420,6 +2424,21 @@ static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
         * we don't need page_cgroup_lock about tail pages, becase they are not
         * accessed by any other context at this point.
         */
+
+       /*
+        * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
+        * may already be on some other mem_cgroup's LRU.  Take care of it.
+        */
+       if (lrucare) {
+               zone = page_zone(page);
+               spin_lock_irq(&zone->lru_lock);
+               if (PageLRU(page)) {
+                       ClearPageLRU(page);
+                       del_page_from_lru_list(zone, page, page_lru(page));
+                       was_on_lru = true;
+               }
+       }
+
        pc->mem_cgroup = memcg;
        /*
         * We access a page_cgroup asynchronously without lock_page_cgroup().
@@ -2443,9 +2462,18 @@ static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
                break;
        }
 
+       if (lrucare) {
+               if (was_on_lru) {
+                       VM_BUG_ON(PageLRU(page));
+                       SetPageLRU(page);
+                       add_page_to_lru_list(zone, page, page_lru(page));
+               }
+               spin_unlock_irq(&zone->lru_lock);
+       }
+
        mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), nr_pages);
        unlock_page_cgroup(pc);
-       WARN_ON_ONCE(PageLRU(page));
+
        /*
         * "charge_statistics" updated event counter. Then, check it.
         * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
@@ -2643,7 +2671,7 @@ static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
        ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
        if (ret == -ENOMEM)
                return ret;
-       __mem_cgroup_commit_charge(memcg, page, nr_pages, pc, ctype);
+       __mem_cgroup_commit_charge(memcg, page, nr_pages, pc, ctype, false);
        return 0;
 }
 
@@ -2663,35 +2691,6 @@ static void
 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
                                        enum charge_type ctype);
 
-static void
-__mem_cgroup_commit_charge_lrucare(struct page *page, struct mem_cgroup *memcg,
-                                       enum charge_type ctype)
-{
-       struct page_cgroup *pc = lookup_page_cgroup(page);
-       struct zone *zone = page_zone(page);
-       unsigned long flags;
-       bool removed = false;
-
-       /*
-        * In some case, SwapCache, FUSE(splice_buf->radixtree), the page
-        * is already on LRU. It means the page may on some other page_cgroup's
-        * LRU. Take care of it.
-        */
-       spin_lock_irqsave(&zone->lru_lock, flags);
-       if (PageLRU(page)) {
-               del_page_from_lru_list(zone, page, page_lru(page));
-               ClearPageLRU(page);
-               removed = true;
-       }
-       __mem_cgroup_commit_charge(memcg, page, 1, pc, ctype);
-       if (removed) {
-               add_page_to_lru_list(zone, page, page_lru(page));
-               SetPageLRU(page);
-       }
-       spin_unlock_irqrestore(&zone->lru_lock, flags);
-       return;
-}
-
 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
                                gfp_t gfp_mask)
 {
@@ -2769,13 +2768,16 @@ static void
 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
                                        enum charge_type ctype)
 {
+       struct page_cgroup *pc;
+
        if (mem_cgroup_disabled())
                return;
        if (!memcg)
                return;
        cgroup_exclude_rmdir(&memcg->css);
 
-       __mem_cgroup_commit_charge_lrucare(page, memcg, ctype);
+       pc = lookup_page_cgroup(page);
+       __mem_cgroup_commit_charge(memcg, page, 1, pc, ctype, true);
        /*
         * Now swap is on-memory. This means this page may be
         * counted both as mem and swap....double count.
@@ -3248,7 +3250,7 @@ int mem_cgroup_prepare_migration(struct page *page,
                ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
        else
                ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
-       __mem_cgroup_commit_charge(memcg, newpage, 1, pc, ctype);
+       __mem_cgroup_commit_charge(memcg, newpage, 1, pc, ctype, false);
        return ret;
 }
 
@@ -3332,7 +3334,7 @@ void mem_cgroup_replace_page_cache(struct page *oldpage,
         * the newpage may be on LRU(or pagevec for LRU) already. We lock
         * LRU while we overwrite pc->mem_cgroup.
         */
-       __mem_cgroup_commit_charge_lrucare(newpage, memcg, type);
+       __mem_cgroup_commit_charge(memcg, newpage, 1, pc, type, true);
 }
 
 #ifdef CONFIG_DEBUG_VM