memcg: fix deadlock between cpuset and memcg
Daisuke Nishimura [Thu, 13 Jan 2011 23:47:41 +0000 (15:47 -0800)]
Commit b1dd693e ("memcg: avoid deadlock between move charge and
try_charge()") can cause another deadlock about mmap_sem on task migration
if cpuset and memcg are mounted onto the same mount point.

After the commit, cgroup_attach_task() has sequence like:

cgroup_attach_task()
  ss->can_attach()
    cpuset_can_attach()
    mem_cgroup_can_attach()
      down_read(&mmap_sem)        (1)
  ss->attach()
    cpuset_attach()
      mpol_rebind_mm()
        down_write(&mmap_sem)     (2)
        up_write(&mmap_sem)
      cpuset_migrate_mm()
        do_migrate_pages()
          down_read(&mmap_sem)
          up_read(&mmap_sem)
    mem_cgroup_move_task()
      mem_cgroup_clear_mc()
        up_read(&mmap_sem)

We can cause deadlock at (2) because we've already aquire the mmap_sem at (1).

But the commit itself is necessary to fix deadlocks which have existed
before the commit like:

Ex.1)
                move charge             |        try charge
  --------------------------------------+------------------------------
    mem_cgroup_can_attach()             |  down_write(&mmap_sem)
      mc.moving_task = current          |    ..
      mem_cgroup_precharge_mc()         |  __mem_cgroup_try_charge()
        mem_cgroup_count_precharge()    |    prepare_to_wait()
          down_read(&mmap_sem)          |    if (mc.moving_task)
          -> cannot aquire the lock     |    -> true
                                        |      schedule()
                                        |      -> move charge should wake it up

Ex.2)
                move charge             |        try charge
  --------------------------------------+------------------------------
    mem_cgroup_can_attach()             |
      mc.moving_task = current          |
      mem_cgroup_precharge_mc()         |
        mem_cgroup_count_precharge()    |
          down_read(&mmap_sem)          |
          ..                            |
          up_read(&mmap_sem)            |
                                        |  down_write(&mmap_sem)
    mem_cgroup_move_task()              |    ..
      mem_cgroup_move_charge()          |  __mem_cgroup_try_charge()
        down_read(&mmap_sem)            |    prepare_to_wait()
        -> cannot aquire the lock       |    if (mc.moving_task)
                                        |    -> true
                                        |      schedule()
                                        |      -> move charge should wake it up

This patch fixes all of these problems by:
1. revert the commit.
2. To fix the Ex.1, we set mc.moving_task after mem_cgroup_count_precharge()
   has released the mmap_sem.
3. To fix the Ex.2, we use down_read_trylock() instead of down_read() in
   mem_cgroup_move_charge() and, if it has failed to aquire the lock, cancel
   all extra charges, wake up all waiters, and retry trylock.

Signed-off-by: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>
Reported-by: Ben Blum <bblum@andrew.cmu.edu>
Cc: Miao Xie <miaox@cn.fujitsu.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Paul Menage <menage@google.com>
Cc: Hiroyuki Kamezawa <kamezawa.hiroyuki@gmail.com>
Cc: Balbir Singh <balbir@in.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm/memcontrol.c

index 1b44ad6..c339d74 100644 (file)
@@ -292,7 +292,6 @@ static struct move_charge_struct {
        unsigned long moved_charge;
        unsigned long moved_swap;
        struct task_struct *moving_task;        /* a task moving charges */
-       struct mm_struct *mm;
        wait_queue_head_t waitq;                /* a waitq for other context */
 } mc = {
        .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
@@ -4681,7 +4680,7 @@ static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
        unsigned long precharge;
        struct vm_area_struct *vma;
 
-       /* We've already held the mmap_sem */
+       down_read(&mm->mmap_sem);
        for (vma = mm->mmap; vma; vma = vma->vm_next) {
                struct mm_walk mem_cgroup_count_precharge_walk = {
                        .pmd_entry = mem_cgroup_count_precharge_pte_range,
@@ -4693,6 +4692,7 @@ static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
                walk_page_range(vma->vm_start, vma->vm_end,
                                        &mem_cgroup_count_precharge_walk);
        }
+       up_read(&mm->mmap_sem);
 
        precharge = mc.precharge;
        mc.precharge = 0;
@@ -4702,10 +4702,15 @@ static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
 
 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
 {
-       return mem_cgroup_do_precharge(mem_cgroup_count_precharge(mm));
+       unsigned long precharge = mem_cgroup_count_precharge(mm);
+
+       VM_BUG_ON(mc.moving_task);
+       mc.moving_task = current;
+       return mem_cgroup_do_precharge(precharge);
 }
 
-static void mem_cgroup_clear_mc(void)
+/* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
+static void __mem_cgroup_clear_mc(void)
 {
        struct mem_cgroup *from = mc.from;
        struct mem_cgroup *to = mc.to;
@@ -4740,23 +4745,28 @@ static void mem_cgroup_clear_mc(void)
                                                PAGE_SIZE * mc.moved_swap);
                }
                /* we've already done mem_cgroup_get(mc.to) */
-
                mc.moved_swap = 0;
        }
-       if (mc.mm) {
-               up_read(&mc.mm->mmap_sem);
-               mmput(mc.mm);
-       }
+       memcg_oom_recover(from);
+       memcg_oom_recover(to);
+       wake_up_all(&mc.waitq);
+}
+
+static void mem_cgroup_clear_mc(void)
+{
+       struct mem_cgroup *from = mc.from;
+
+       /*
+        * we must clear moving_task before waking up waiters at the end of
+        * task migration.
+        */
+       mc.moving_task = NULL;
+       __mem_cgroup_clear_mc();
        spin_lock(&mc.lock);
        mc.from = NULL;
        mc.to = NULL;
        spin_unlock(&mc.lock);
-       mc.moving_task = NULL;
-       mc.mm = NULL;
        mem_cgroup_end_move(from);
-       memcg_oom_recover(from);
-       memcg_oom_recover(to);
-       wake_up_all(&mc.waitq);
 }
 
 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
@@ -4778,38 +4788,23 @@ static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
                        return 0;
                /* We move charges only when we move a owner of the mm */
                if (mm->owner == p) {
-                       /*
-                        * We do all the move charge works under one mmap_sem to
-                        * avoid deadlock with down_write(&mmap_sem)
-                        * -> try_charge() -> if (mc.moving_task) -> sleep.
-                        */
-                       down_read(&mm->mmap_sem);
-
                        VM_BUG_ON(mc.from);
                        VM_BUG_ON(mc.to);
                        VM_BUG_ON(mc.precharge);
                        VM_BUG_ON(mc.moved_charge);
                        VM_BUG_ON(mc.moved_swap);
-                       VM_BUG_ON(mc.moving_task);
-                       VM_BUG_ON(mc.mm);
-
                        mem_cgroup_start_move(from);
                        spin_lock(&mc.lock);
                        mc.from = from;
                        mc.to = mem;
-                       mc.precharge = 0;
-                       mc.moved_charge = 0;
-                       mc.moved_swap = 0;
                        spin_unlock(&mc.lock);
-                       mc.moving_task = current;
-                       mc.mm = mm;
+                       /* We set mc.moving_task later */
 
                        ret = mem_cgroup_precharge_mc(mm);
                        if (ret)
                                mem_cgroup_clear_mc();
-                       /* We call up_read() and mmput() in clear_mc(). */
-               } else
-                       mmput(mm);
+               }
+               mmput(mm);
        }
        return ret;
 }
@@ -4898,7 +4893,19 @@ static void mem_cgroup_move_charge(struct mm_struct *mm)
        struct vm_area_struct *vma;
 
        lru_add_drain_all();
-       /* We've already held the mmap_sem */
+retry:
+       if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
+               /*
+                * Someone who are holding the mmap_sem might be waiting in
+                * waitq. So we cancel all extra charges, wake up all waiters,
+                * and retry. Because we cancel precharges, we might not be able
+                * to move enough charges, but moving charge is a best-effort
+                * feature anyway, so it wouldn't be a big problem.
+                */
+               __mem_cgroup_clear_mc();
+               cond_resched();
+               goto retry;
+       }
        for (vma = mm->mmap; vma; vma = vma->vm_next) {
                int ret;
                struct mm_walk mem_cgroup_move_charge_walk = {
@@ -4917,6 +4924,7 @@ static void mem_cgroup_move_charge(struct mm_struct *mm)
                         */
                        break;
        }
+       up_read(&mm->mmap_sem);
 }
 
 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
@@ -4925,11 +4933,17 @@ static void mem_cgroup_move_task(struct cgroup_subsys *ss,
                                struct task_struct *p,
                                bool threadgroup)
 {
-       if (!mc.mm)
+       struct mm_struct *mm;
+
+       if (!mc.to)
                /* no need to move charge */
                return;
 
-       mem_cgroup_move_charge(mc.mm);
+       mm = get_task_mm(p);
+       if (mm) {
+               mem_cgroup_move_charge(mm);
+               mmput(mm);
+       }
        mem_cgroup_clear_mc();
 }
 #else  /* !CONFIG_MMU */