cgroup: fix race condition around termination check in css_task_iter_next()
authorTejun Heo <tj@kernel.org>
Thu, 29 Oct 2015 02:43:05 +0000 (11:43 +0900)
committerTejun Heo <tj@kernel.org>
Thu, 29 Oct 2015 02:43:05 +0000 (11:43 +0900)
css_task_iter_next() checked @it->cur_task before grabbing
css_set_lock and assumed that the result won't change afterwards;
however, tasks could leave the cgroup being iterated terminating the
iterator before css_task_lock is acquired.  If this happens,
css_task_iter_next() tries to calculate the current task from NULL
cg_list pointer leading to the following oops.

 BUG: unable to handle kernel paging request at fffffffffffff7d0
 IP: [<ffffffff810d5f22>] css_task_iter_next+0x42/0x80
 ...
 CPU: 4 PID: 6391 Comm: JobQDisp2 Not tainted 4.0.9-22_fbk4_rc3_81616_ge8d9cb6 #1
 Hardware name: Quanta Freedom/Winterfell, BIOS F03_3B08 03/04/2014
 task: ffff880868e46400 ti: ffff88083404c000 task.ti: ffff88083404c000
 RIP: 0010:[<ffffffff810d5f22>]  [<ffffffff810d5f22>] css_task_iter_next+0x42/0x80
 RSP: 0018:ffff88083404fd28  EFLAGS: 00010246
 RAX: 0000000000000000 RBX: ffff88083404fd68 RCX: ffff8804697fb8b0
 RDX: fffffffffffff7c0 RSI: ffff8803b7dff800 RDI: ffffffff822c0278
 RBP: ffff88083404fd38 R08: 0000000000017160 R09: ffff88046f4070c0
 R10: ffffffff810d61f7 R11: 0000000000000293 R12: ffff880863bf8400
 R13: ffff88046b87fd80 R14: 0000000000000000 R15: ffff88083404fe58
 FS:  00007fa0567e2700(0000) GS:ffff88046f900000(0000) knlGS:0000000000000000
 CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
 CR2: fffffffffffff7d0 CR3: 0000000469568000 CR4: 00000000001406e0
 Stack:
  0000000000000246 0000000000000000 ffff88083404fde8 ffffffff810d6248
  ffff88083404fd68 0000000000000000 ffff8803b7dff800 000001ef000001ee
  0000000000000000 0000000000000000 ffff880863bf8568 0000000000000000
 Call Trace:
  [<ffffffff810d6248>] cgroup_pidlist_start+0x258/0x550
  [<ffffffff810cf66d>] cgroup_seqfile_start+0x1d/0x20
  [<ffffffff8121f8ef>] kernfs_seq_start+0x5f/0xa0
  [<ffffffff811cab76>] seq_read+0x166/0x380
  [<ffffffff812200fd>] kernfs_fop_read+0x11d/0x180
  [<ffffffff811a7398>] __vfs_read+0x18/0x50
  [<ffffffff811a745d>] vfs_read+0x8d/0x150
  [<ffffffff811a756f>] SyS_read+0x4f/0xb0
  [<ffffffff818d4772>] system_call_fastpath+0x12/0x17

Fix it by moving the termination condition check inside css_set_lock.
@it->cur_task is now cleared after being put and @it->task_pos is
tested for termination instead of @it->cset_pos as they indicate the
same condition and @it->task_pos is what's being dereferenced.

Signed-off-by: Tejun Heo <tj@kernel.org>
Reported-by: Calvin Owens <calvinowens@fb.com>
Fixes: ed27b9f7a17d ("cgroup: don't hold css_set_rwsem across css task iteration")
Acked-by: Zefan Li <lizefan@huawei.com>
kernel/cgroup.c

index 4f4fc538cb1b605da8f36f7faf5b9bde0581c240..b9d0cce3f9ce54937fea988b531d0cc7bf52f692 100644 (file)
@@ -3930,18 +3930,19 @@ void css_task_iter_start(struct cgroup_subsys_state *css,
  */
 struct task_struct *css_task_iter_next(struct css_task_iter *it)
 {
-       if (!it->cset_pos)
-               return NULL;
-
-       if (it->cur_task)
+       if (it->cur_task) {
                put_task_struct(it->cur_task);
+               it->cur_task = NULL;
+       }
 
        spin_lock_bh(&css_set_lock);
 
-       it->cur_task = list_entry(it->task_pos, struct task_struct, cg_list);
-       get_task_struct(it->cur_task);
-
-       css_task_iter_advance(it);
+       if (it->task_pos) {
+               it->cur_task = list_entry(it->task_pos, struct task_struct,
+                                         cg_list);
+               get_task_struct(it->cur_task);
+               css_task_iter_advance(it);
+       }
 
        spin_unlock_bh(&css_set_lock);