workqueue: always clear WORKER_REBIND in busy_worker_rebind_fn()
authorLai Jiangshan <laijs@cn.fujitsu.com>
Mon, 17 Sep 2012 22:42:31 +0000 (15:42 -0700)
committerTejun Heo <tj@kernel.org>
Mon, 17 Sep 2012 22:42:31 +0000 (15:42 -0700)
busy_worker_rebind_fn() didn't clear WORKER_REBIND if rebinding failed
(CPU is down again).  This used to be okay because the flag wasn't
used for anything else.

However, after 25511a477 "workqueue: reimplement CPU online rebinding
to handle idle workers", WORKER_REBIND is also used to command idle
workers to rebind.  If not cleared, the worker may confuse the next
CPU_UP cycle by having REBIND spuriously set or oops / get stuck by
prematurely calling idle_worker_rebind().

  WARNING: at /work/os/wq/kernel/workqueue.c:1323 worker_thread+0x4cd/0x5
 00()
  Hardware name: Bochs
  Modules linked in: test_wq(O-)
  Pid: 33, comm: kworker/1:1 Tainted: G           O 3.6.0-rc1-work+ #3
  Call Trace:
   [<ffffffff8109039f>] warn_slowpath_common+0x7f/0xc0
   [<ffffffff810903fa>] warn_slowpath_null+0x1a/0x20
   [<ffffffff810b3f1d>] worker_thread+0x4cd/0x500
   [<ffffffff810bc16e>] kthread+0xbe/0xd0
   [<ffffffff81bd2664>] kernel_thread_helper+0x4/0x10
  ---[ end trace e977cf20f4661968 ]---
  BUG: unable to handle kernel NULL pointer dereference at           (null)
  IP: [<ffffffff810b3db0>] worker_thread+0x360/0x500
  PGD 0
  Oops: 0000 [#1] PREEMPT SMP DEBUG_PAGEALLOC
  Modules linked in: test_wq(O-)
  CPU 0
  Pid: 33, comm: kworker/1:1 Tainted: G        W  O 3.6.0-rc1-work+ #3 Bochs Bochs
  RIP: 0010:[<ffffffff810b3db0>]  [<ffffffff810b3db0>] worker_thread+0x360/0x500
  RSP: 0018:ffff88001e1c9de0  EFLAGS: 00010086
  RAX: 0000000000000000 RBX: ffff88001e633e00 RCX: 0000000000004140
  RDX: 0000000000000000 RSI: 0000000000000000 RDI: 0000000000000009
  RBP: ffff88001e1c9ea0 R08: 0000000000000000 R09: 0000000000000001
  R10: 0000000000000002 R11: 0000000000000000 R12: ffff88001fc8d580
  R13: ffff88001fc8d590 R14: ffff88001e633e20 R15: ffff88001e1c6900
  FS:  0000000000000000(0000) GS:ffff88001fc00000(0000) knlGS:0000000000000000
  CS:  0010 DS: 0000 ES: 0000 CR0: 000000008005003b
  CR2: 0000000000000000 CR3: 00000000130e8000 CR4: 00000000000006f0
  DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
  DR3: 0000000000000000 DR6: 00000000ffff0ff0 DR7: 0000000000000400
  Process kworker/1:1 (pid: 33, threadinfo ffff88001e1c8000, task ffff88001e1c6900)
  Stack:
   ffff880000000000 ffff88001e1c9e40 0000000000000001 ffff88001e1c8010
   ffff88001e519c78 ffff88001e1c9e58 ffff88001e1c6900 ffff88001e1c6900
   ffff88001e1c6900 ffff88001e1c6900 ffff88001fc8d340 ffff88001fc8d340
  Call Trace:
   [<ffffffff810bc16e>] kthread+0xbe/0xd0
   [<ffffffff81bd2664>] kernel_thread_helper+0x4/0x10
  Code: b1 00 f6 43 48 02 0f 85 91 01 00 00 48 8b 43 38 48 89 df 48 8b 00 48 89 45 90 e8 ac f0 ff ff 3c 01 0f 85 60 01 00 00 48 8b 53 50 <8b> 02 83 e8 01 85 c0 89 02 0f 84 3b 01 00 00 48 8b 43 38 48 8b
  RIP  [<ffffffff810b3db0>] worker_thread+0x360/0x500
   RSP <ffff88001e1c9de0>
  CR2: 0000000000000000

There was no reason to keep WORKER_REBIND on failure in the first
place - WORKER_UNBOUND is guaranteed to be set in such cases
preventing incorrectly activating concurrency management.  Always
clear WORKER_REBIND.

tj: Updated comment and description.

Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
kernel/workqueue.c

index 1e1373bcb3e3125f72baf77cf396689d18de9bdd..b80065a2450ad082d941b757bdcf49f4ebacd5d7 100644 (file)
@@ -1349,8 +1349,16 @@ static void busy_worker_rebind_fn(struct work_struct *work)
        struct worker *worker = container_of(work, struct worker, rebind_work);
        struct global_cwq *gcwq = worker->pool->gcwq;
 
-       if (worker_maybe_bind_and_lock(worker))
-               worker_clr_flags(worker, WORKER_REBIND);
+       worker_maybe_bind_and_lock(worker);
+
+       /*
+        * %WORKER_REBIND must be cleared even if the above binding failed;
+        * otherwise, we may confuse the next CPU_UP cycle or oops / get
+        * stuck by calling idle_worker_rebind() prematurely.  If CPU went
+        * down again inbetween, %WORKER_UNBOUND would be set, so clearing
+        * %WORKER_REBIND is always safe.
+        */
+       worker_clr_flags(worker, WORKER_REBIND);
 
        spin_unlock_irq(&gcwq->lock);
 }