Re: [syzbot] INFO: rcu detected stall in __hrtimer_run_queues

[Paul E. McKenney]

On Tue, Nov 16, 2021 at 08:42:39AM -0700, Jens Axboe wrote:
> On 11/16/21 8:41 AM, syzbot wrote:
> > syzbot suspects this issue was fixed by commit:
> > 
> > commit b60876296847e6cd7f1da4b8b7f0f31399d59aa1
> > Author: Jens Axboe <axboe@xxxxxxxxx>
> > Date:   Fri Oct 15 21:03:52 2021 +0000
> > 
> >     block: improve layout of struct request
> 
> No functional changes in that patch, so looks like a fluky bisection.

I am seeing an intermittent (and rather strange) stall warnings on
v5.16-rc1.  This is a self-detected stall from the idle loop.  The
reason that this is strange is that the usual reason that a CPU stalls
in the idle loop is due to a long-running interrupt, in which case
you would expect other CPUs to detect the stall.

Reproduce using RCU's TRE07 scenario, except that the MTBF looks to be
several hundred hours.  But I ran this scenario long enough on v5.15-rc*
to be confident that this stall warning is a regression introduced
recently.

And the reason is that the CPU, despite being in the idle loop, is not
marked as idle from an RCU perspective (see the "idle=d59/0/0x1"):

 rcu:     0-...!: (13 ticks this GP) idle=d59/0/0x1 softirq=281261
/281261 fqs=1
  (t=2199037 jiffies g=249449 q=5)
 NMI backtrace for cpu 0
 CPU: 0 PID: 0 Comm: swapper/0 Not tainted 5.16.0-rc1 #4571
 Hardware name: Red Hat KVM/RHEL-AV, BIOS 1.13.0-2.module_el8.5.0+
746+bbd5d70c 04/01/2014
 Call Trace:
  <IRQ>
  dump_stack_lvl+0x33/0x42
  nmi_cpu_backtrace.cold.6+0x30/0x70
  ? lapic_can_unplug_cpu+0x70/0x70
  nmi_trigger_cpumask_backtrace+0xbf/0xd0
  rcu_dump_cpu_stacks+0xc0/0x120
  rcu_sched_clock_irq.cold.110+0x15a/0x312
  ? get_nohz_timer_target+0x60/0x190
  ? lock_timer_base+0x62/0x80
  ? account_process_tick+0xd4/0x160
  ? tick_sched_handle.isra.24+0x40/0x40
  update_process_times+0x8e/0xc0
  tick_sched_handle.isra.24+0x30/0x40
  tick_sched_timer+0x6a/0x80
  __hrtimer_run_queues+0xfc/0x2a0
  hrtimer_interrupt+0x105/0x220
  ? resched_curr+0x1e/0xc0
  __sysvec_apic_timer_interrupt+0x7a/0x160
  sysvec_apic_timer_interrupt+0x85/0xb0
  </IRQ>
  <TASK>
  asm_sysvec_apic_timer_interrupt+0x12/0x20
 RIP: 0010:default_idle+0xb/0x10
 Code: ff 48 89 df e8 16 5c 90 ff eb d7 e8 bf 82 ff ff cc cc cc cc
 cc cc cc cc cc cc cc cc cc cc cc eb 07 0f 00 2d df 92 41 00 fb f4 <c3> 0f 1f 40
 00 65 48 8b 04 25 00 ad 01 00 f0 80 48 02 20 48 8b 10
 RSP: 0018:ffffffff9dc03e98 EFLAGS: 00000202
 RAX: ffffffff9d3ed200 RBX: 0000000000000000 RCX: 0000000000000000
 RDX: 0000000000000000 RSI: ffffffff9da3e0a1 RDI: ffffffff9da683ae
 RBP: ffffffff9de86050 R08: 00000000e141ad57 R09: ffffa03c5f229d40
 R10: 0000000000002400 R11: 0000000000002400 R12: 0000000000000000
 R13: 0000000000000000 R14: ffffffffffffffff R15: ffffffff9dc14940
  ? __cpuidle_text_start+0x8/0x8
  ? __cpuidle_text_start+0x8/0x8
  default_idle_call+0x28/0xd0
  do_idle+0x1fb/0x290
  cpu_startup_entry+0x14/0x20
  start_kernel+0x659/0x680
  secondary_startup_64_no_verify+0xc2/0xcb
  </TASK>

The usual reason for this odd situation is that someone forgot an
irq_enter() or added an extra irq_exit().  Or likewise for a number of
similar functions that tell RCU to start/stop ignoring the current CPU:
nmi_enter(), nmi_exit(), rcu_*_enter(), rcu_*_exit(), and so on.

Adding the x86 list on CC.

							Thanx, Paul