mm owner: fix race between swapoff and exit
authorBalbir Singh <balbir@linux.vnet.ibm.com>
Sun, 28 Sep 2008 22:09:31 +0000 (23:09 +0100)
committerLinus Torvalds <torvalds@linux-foundation.org>
Mon, 29 Sep 2008 15:41:47 +0000 (08:41 -0700)
There's a race between mm->owner assignment and swapoff, more easily
seen when task slab poisoning is turned on.  The condition occurs when
try_to_unuse() runs in parallel with an exiting task.  A similar race
can occur with callers of get_task_mm(), such as /proc/<pid>/<mmstats>
or ptrace or page migration.

CPU0                                    CPU1
                                        try_to_unuse
                                        looks at mm = task0->mm
                                        increments mm->mm_users
task 0 exits
mm->owner needs to be updated, but no
new owner is found (mm_users > 1, but
no other task has task->mm = task0->mm)
mm_update_next_owner() leaves
                                        mmput(mm) decrements mm->mm_users
task0 freed
                                        dereferencing mm->owner fails

The fix is to notify the subsystem via mm_owner_changed callback(),
if no new owner is found, by specifying the new task as NULL.

Jiri Slaby:
mm->owner was set to NULL prior to calling cgroup_mm_owner_callbacks(), but
must be set after that, so as not to pass NULL as old owner causing oops.

Daisuke Nishimura:
mm_update_next_owner() may set mm->owner to NULL, but mem_cgroup_from_task()
and its callers need to take account of this situation to avoid oops.

Hugh Dickins:
Lockdep warning and hang below exec_mmap() when testing these patches.
exit_mm() up_reads mmap_sem before calling mm_update_next_owner(),
so exec_mmap() now needs to do the same.  And with that repositioning,
there's now no point in mm_need_new_owner() allowing for NULL mm.

Reported-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Balbir Singh <balbir@linux.vnet.ibm.com>
Signed-off-by: Jiri Slaby <jirislaby@gmail.com>
Signed-off-by: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>
Signed-off-by: Hugh Dickins <hugh@veritas.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Paul Menage <menage@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
fs/exec.c
kernel/cgroup.c
kernel/exit.c
mm/memcontrol.c

index 32993beecbe91a10670cc4f5a9a72a5faf813c56..cecee501ce78805c32c9d007c658fce1e87b0a97 100644 (file)
--- a/fs/exec.c
+++ b/fs/exec.c
@@ -752,11 +752,11 @@ static int exec_mmap(struct mm_struct *mm)
        tsk->active_mm = mm;
        activate_mm(active_mm, mm);
        task_unlock(tsk);
-       mm_update_next_owner(old_mm);
        arch_pick_mmap_layout(mm);
        if (old_mm) {
                up_read(&old_mm->mmap_sem);
                BUG_ON(active_mm != old_mm);
+               mm_update_next_owner(old_mm);
                mmput(old_mm);
                return 0;
        }
index 13932abde15914e36a0d4f33e10b4a9a653f8c31..a0123d75ec9a7624f897bc2984ca4009b61abfbc 100644 (file)
@@ -2738,14 +2738,15 @@ void cgroup_fork_callbacks(struct task_struct *child)
  */
 void cgroup_mm_owner_callbacks(struct task_struct *old, struct task_struct *new)
 {
-       struct cgroup *oldcgrp, *newcgrp;
+       struct cgroup *oldcgrp, *newcgrp = NULL;
 
        if (need_mm_owner_callback) {
                int i;
                for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
                        struct cgroup_subsys *ss = subsys[i];
                        oldcgrp = task_cgroup(old, ss->subsys_id);
-                       newcgrp = task_cgroup(new, ss->subsys_id);
+                       if (new)
+                               newcgrp = task_cgroup(new, ss->subsys_id);
                        if (oldcgrp == newcgrp)
                                continue;
                        if (ss->mm_owner_changed)
index 16395644a98ff8c060b2f5fd776fe7abadd61c6a..85a83c831856c193570e40a3b7d3e03ef8862d1c 100644 (file)
@@ -583,8 +583,6 @@ mm_need_new_owner(struct mm_struct *mm, struct task_struct *p)
         * If there are other users of the mm and the owner (us) is exiting
         * we need to find a new owner to take on the responsibility.
         */
-       if (!mm)
-               return 0;
        if (atomic_read(&mm->mm_users) <= 1)
                return 0;
        if (mm->owner != p)
@@ -627,6 +625,16 @@ retry:
        } while_each_thread(g, c);
 
        read_unlock(&tasklist_lock);
+       /*
+        * We found no owner yet mm_users > 1: this implies that we are
+        * most likely racing with swapoff (try_to_unuse()) or /proc or
+        * ptrace or page migration (get_task_mm()).  Mark owner as NULL,
+        * so that subsystems can understand the callback and take action.
+        */
+       down_write(&mm->mmap_sem);
+       cgroup_mm_owner_callbacks(mm->owner, NULL);
+       mm->owner = NULL;
+       up_write(&mm->mmap_sem);
        return;
 
 assign_new_owner:
index c0500e4d3a2f9d3c5d392888326780379d064a4f..36896f3eb7f5e5c2e4c3803cc8125fe461a71669 100644 (file)
@@ -250,6 +250,14 @@ static struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
 
 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
 {
+       /*
+        * mm_update_next_owner() may clear mm->owner to NULL
+        * if it races with swapoff, page migration, etc.
+        * So this can be called with p == NULL.
+        */
+       if (unlikely(!p))
+               return NULL;
+
        return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
                                struct mem_cgroup, css);
 }
@@ -549,6 +557,11 @@ static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
        if (likely(!memcg)) {
                rcu_read_lock();
                mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
+               if (unlikely(!mem)) {
+                       rcu_read_unlock();
+                       kmem_cache_free(page_cgroup_cache, pc);
+                       return 0;
+               }
                /*
                 * For every charge from the cgroup, increment reference count
                 */
@@ -801,6 +814,10 @@ int mem_cgroup_shrink_usage(struct mm_struct *mm, gfp_t gfp_mask)
 
        rcu_read_lock();
        mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
+       if (unlikely(!mem)) {
+               rcu_read_unlock();
+               return 0;
+       }
        css_get(&mem->css);
        rcu_read_unlock();