memcg: fix mis-accounting of file mapped racy with migration

FILE_MAPPED per memcg of migrated file cache is not properly updated,
because our hook in page_add_file_rmap() can't know to which memcg
FILE_MAPPED should be counted.

Basically, this patch is for fixing the bug but includes some big changes
to fix up other messes.

Now, at migrating mapped file, events happen in following sequence.

 1. allocate a new page.
 2. get memcg of an old page.
 3. charge ageinst a new page before migration. But at this point,
    no changes to new page's page_cgroup, no commit for the charge.
    (IOW, PCG_USED bit is not set.)
 4. page migration replaces radix-tree, old-page and new-page.
 5. page migration remaps the new page if the old page was mapped.
 6. Here, the new page is unlocked.
 7. memcg commits the charge for newpage, Mark the new page's page_cgroup
    as PCG_USED.

Because "commit" happens after page-remap, we can count FILE_MAPPED
at "5", because we should avoid to trust page_cgroup->mem_cgroup.
if PCG_USED bit is unset.
(Note: memcg's LRU removal code does that but LRU-isolation logic is used
 for helping it. When we overwrite page_cgroup->mem_cgroup, page_cgroup is
 not on LRU or page_cgroup->mem_cgroup is NULL.)

We can lose file_mapped accounting information at 5 because FILE_MAPPED
is updated only when mapcount changes 0->1. So we should catch it.

BTW, historically, above implemntation comes from migration-failure
of anonymous page. Because we charge both of old page and new page
with mapcount=0, we can't catch
  - the page is really freed before remap.
  - migration fails but it's freed before remap
or .....corner cases.

New migration sequence with memcg is:

 1. allocate a new page.
 2. mark PageCgroupMigration to the old page.
 3. charge against a new page onto the old page's memcg. (here, new page's pc
    is marked as PageCgroupUsed.)
 4. page migration replaces radix-tree, page table, etc...
 5. At remapping, new page's page_cgroup is now makrked as "USED"
    We can catch 0->1 event and FILE_MAPPED will be properly updated.

    And we can catch SWAPOUT event after unlock this and freeing this
    page by unmap() can be caught.

 7. Clear PageCgroupMigration of the old page.

So, FILE_MAPPED will be correctly updated.

Then, for what MIGRATION flag is ?
  Without it, at migration failure, we may have to charge old page again
  because it may be fully unmapped. "charge" means that we have to dive into
  memory reclaim or something complated. So, it's better to avoid
  charge it again. Before this patch, __commit_charge() was working for
  both of the old/new page and fixed up all. But this technique has some
  racy condtion around FILE_MAPPED and SWAPOUT etc...
  Now, the kernel use MIGRATION flag and don't uncharge old page until
  the end of migration.

I hope this change will make memcg's page migration much simpler.  This
page migration has caused several troubles.  Worth to add a flag for
simplification.

Reviewed-by: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>
Tested-by: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>
Reported-by: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>
Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Balbir Singh <balbir@in.ibm.com>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

2 files changed, 98 insertions, 39 deletions

diff --git a/mm/memcontrol.c b/mm/memcontrol.c
index 8c200e86da4c..df1234c0dac3 100644
--- a/mm/memcontrol.c
+++ b/mm/memcontrol.c
@@ -2258,7 +2258,8 @@ __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
         switch (ctype) {
         case MEM_CGROUP_CHARGE_TYPE_MAPPED:
         case MEM_CGROUP_CHARGE_TYPE_DROP:
-                if (page_mapped(page))
+                /* See mem_cgroup_prepare_migration() */
+                if (page_mapped(page) || PageCgroupMigration(pc))
                         goto unlock_out;
                 break;
         case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
@@ -2481,10 +2482,12 @@ static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
  * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
  * page belongs to.
  */
-int mem_cgroup_prepare_migration(struct page *page, struct mem_cgroup **ptr)
+int mem_cgroup_prepare_migration(struct page *page,
+        struct page *newpage, struct mem_cgroup **ptr)
 {
         struct page_cgroup *pc;
         struct mem_cgroup *mem = NULL;
+        enum charge_type ctype;
         int ret = 0;
 
         if (mem_cgroup_disabled())
@@ -2495,69 +2498,125 @@ int mem_cgroup_prepare_migration(struct page *page, struct mem_cgroup **ptr)
         if (PageCgroupUsed(pc)) {
                 mem = pc->mem_cgroup;
                 css_get(&mem->css);
+                /*
+                 * At migrating an anonymous page, its mapcount goes down
+                 * to 0 and uncharge() will be called. But, even if it's fully
+                 * unmapped, migration may fail and this page has to be
+                 * charged again. We set MIGRATION flag here and delay uncharge
+                 * until end_migration() is called
+                 *
+                 * Corner Case Thinking
+                 * A)
+                 * When the old page was mapped as Anon and it's unmap-and-freed
+                 * while migration was ongoing.
+                 * If unmap finds the old page, uncharge() of it will be delayed
+                 * until end_migration(). If unmap finds a new page, it's
+                 * uncharged when it make mapcount to be 1->0. If unmap code
+                 * finds swap_migration_entry, the new page will not be mapped
+                 * and end_migration() will find it(mapcount==0).
+                 *
+                 * B)
+                 * When the old page was mapped but migraion fails, the kernel
+                 * remaps it. A charge for it is kept by MIGRATION flag even
+                 * if mapcount goes down to 0. We can do remap successfully
+                 * without charging it again.
+                 *
+                 * C)
+                 * The "old" page is under lock_page() until the end of
+                 * migration, so, the old page itself will not be swapped-out.
+                 * If the new page is swapped out before end_migraton, our
+                 * hook to usual swap-out path will catch the event.
+                 */
+                if (PageAnon(page))
+                        SetPageCgroupMigration(pc);
         }
         unlock_page_cgroup(pc);
+        /*
+         * If the page is not charged at this point,
+         * we return here.
+         */
+        if (!mem)
+                return 0;
 
         *ptr = mem;
-        if (mem) {
-                ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, ptr, false);
-                css_put(&mem->css);
+        ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, ptr, false);
+        css_put(&mem->css);/* drop extra refcnt */
+        if (ret || *ptr == NULL) {
+                if (PageAnon(page)) {
+                        lock_page_cgroup(pc);
+                        ClearPageCgroupMigration(pc);
+                        unlock_page_cgroup(pc);
+                        /*
+                         * The old page may be fully unmapped while we kept it.
+                         */
+                        mem_cgroup_uncharge_page(page);
+                }
+                return -ENOMEM;
         }
+        /*
+         * We charge new page before it's used/mapped. So, even if unlock_page()
+         * is called before end_migration, we can catch all events on this new
+         * page. In the case new page is migrated but not remapped, new page's
+         * mapcount will be finally 0 and we call uncharge in end_migration().
+         */
+        pc = lookup_page_cgroup(newpage);
+        if (PageAnon(page))
+                ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
+        else if (page_is_file_cache(page))
+                ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
+        else
+                ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
+        __mem_cgroup_commit_charge(mem, pc, ctype);
         return ret;
 }
 
 /* remove redundant charge if migration failed*/
 void mem_cgroup_end_migration(struct mem_cgroup *mem,
-                struct page *oldpage, struct page *newpage)
+        struct page *oldpage, struct page *newpage)
 {
-        struct page *target, *unused;
+        struct page *used, *unused;
         struct page_cgroup *pc;
-        enum charge_type ctype;
 
         if (!mem)
                 return;
+        /* blocks rmdir() */
         cgroup_exclude_rmdir(&mem->css);
         /* at migration success, oldpage->mapping is NULL. */
         if (oldpage->mapping) {
-                target = oldpage;
-                unused = NULL;
+                used = oldpage;
+                unused = newpage;
         } else {
-                target = newpage;
+                used = newpage;
                 unused = oldpage;
         }
-
-        if (PageAnon(target))
-                ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
-        else if (page_is_file_cache(target))
-                ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
-        else
-                ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
-
-        /* unused page is not on radix-tree now. */
-        if (unused)
-                __mem_cgroup_uncharge_common(unused, ctype);
-
-        pc = lookup_page_cgroup(target);
         /*
-         * __mem_cgroup_commit_charge() check PCG_USED bit of page_cgroup.
-         * So, double-counting is effectively avoided.
+         * We disallowed uncharge of pages under migration because mapcount
+         * of the page goes down to zero, temporarly.
+         * Clear the flag and check the page should be charged.
          */
-        __mem_cgroup_commit_charge(mem, pc, ctype);
+        pc = lookup_page_cgroup(oldpage);
+        lock_page_cgroup(pc);
+        ClearPageCgroupMigration(pc);
+        unlock_page_cgroup(pc);
 
+        if (unused != oldpage)
+                pc = lookup_page_cgroup(unused);
+        __mem_cgroup_uncharge_common(unused, MEM_CGROUP_CHARGE_TYPE_FORCE);
+
+        pc = lookup_page_cgroup(used);
         /*
-         * Both of oldpage and newpage are still under lock_page().
-         * Then, we don't have to care about race in radix-tree.
-         * But we have to be careful that this page is unmapped or not.
-         *
-         * There is a case for !page_mapped(). At the start of
-         * migration, oldpage was mapped. But now, it's zapped.
-         * But we know *target* page is not freed/reused under us.
-         * mem_cgroup_uncharge_page() does all necessary checks.
+         * If a page is a file cache, radix-tree replacement is very atomic
+         * and we can skip this check. When it was an Anon page, its mapcount
+         * goes down to 0. But because we added MIGRATION flage, it's not
+         * uncharged yet. There are several case but page->mapcount check
+         * and USED bit check in mem_cgroup_uncharge_page() will do enough
+         * check. (see prepare_charge() also)
          */
-        if (ctype == MEM_CGROUP_CHARGE_TYPE_MAPPED)
-                mem_cgroup_uncharge_page(target);
+        if (PageAnon(used))
+                mem_cgroup_uncharge_page(used);
         /*
-         * At migration, we may charge account against cgroup which has no tasks
+         * At migration, we may charge account against cgroup which has no
+         * tasks.
          * So, rmdir()->pre_destroy() can be called while we do this charge.
          * In that case, we need to call pre_destroy() again. check it here.
          */
diff --git a/mm/migrate.c b/mm/migrate.c
index 09e2471afa0f..4205b1d6049e 100644
--- a/mm/migrate.c
+++ b/mm/migrate.c
@@ -590,7 +590,7 @@ static int unmap_and_move(new_page_t get_new_page, unsigned long private,
         }
 
         /* charge against new page */
-        charge = mem_cgroup_prepare_migration(page, &mem);
+        charge = mem_cgroup_prepare_migration(page, newpage, &mem);
         if (charge == -ENOMEM) {
                 rc = -ENOMEM;
                 goto unlock;