sched: make printk safe when rq lock is held

The dead lock can happen if we try to use printk(), such as a call of
SCHED_WARN_ON(), during the rq->__lock is held. The printk() will try to
print the message to the console, and the console driver can call
queue_work_on(), which will try to obtain rq->__lock again.

This means that any WARN during the kernel function that hold the
rq->__lock, such as schedule(), sched_ttwu_pending(), etc, can cause dead
lock.

Following is the call trace of the deadlock case that I encounter:

  PID: 0      TASK: ff36bfda010c8000  CPU: 156  COMMAND: "swapper/156"
   #0 crash_nmi_callback+30
   #1 nmi_handle+85
   #2 default_do_nmi+66
   #3 exc_nmi+291
   #4 end_repeat_nmi+22
      [exception RIP: native_queued_spin_lock_slowpath+96]
   #5 native_queued_spin_lock_slowpath+96
   torvalds#6 _raw_spin_lock+30
   torvalds#7 ttwu_queue+111
   torvalds#8 try_to_wake_up+375
   torvalds#9 __queue_work+462
  torvalds#10 queue_work_on+32
  torvalds#11 soft_cursor+420
  torvalds#12 bit_cursor+898
  torvalds#13 hide_cursor+39
  torvalds#14 vt_console_print+995
  torvalds#15 call_console_drivers.constprop.0+204
  torvalds#16 console_unlock+374
  torvalds#17 vprintk_emit+280
  torvalds#18 printk+88
  torvalds#19 __warn_printk+71
  torvalds#20 enqueue_task_fair+1779
  torvalds#21 activate_task+102
  torvalds#22 ttwu_do_activate+155
  torvalds#23 sched_ttwu_pending+177
  torvalds#24 flush_smp_call_function_from_idle+42
  torvalds#25 do_idle+161
  torvalds#26 cpu_startup_entry+25
  torvalds#27 secondary_startup_64_no_verify+194

Fix this by using __printk_safe_enter()/__printk_safe_exit() in
rq_pin_lock()/rq_unpin_lock(). Then, printk will defer to print out the
buffers to the console.

Signed-off-by: Menglong Dong <dongml2@chinatelecom.cn>
Signed-off-by: Bin Lai <laib2@chinatelecom.cn>

kernel/sched/sched.h

@@ -1702,6 +1702,7 @@ static inline void rq_pin_lock(struct rq *rq, struct rq_flags *rf)
 {
 	rf->cookie = lockdep_pin_lock(__rq_lockp(rq));
 
+	__printk_safe_enter();
 #ifdef CONFIG_SCHED_DEBUG
 	rq->clock_update_flags &= (RQCF_REQ_SKIP|RQCF_ACT_SKIP);
 	rf->clock_update_flags = 0;
@@ -1718,6 +1719,7 @@ static inline void rq_unpin_lock(struct rq *rq, struct rq_flags *rf)
 		rf->clock_update_flags = RQCF_UPDATED;
 #endif
 
+	__printk_safe_exit();
 	lockdep_unpin_lock(__rq_lockp(rq), rf->cookie);
 }