BUG: tracer_alloc_buffers returned with preemption imbalance

Hi Paul,

I got some boot hangs in the below branches and linux-next:

        rcu/dev.2012.06.21c-4f4a59a
        rcu/fixes.2012.06.20a-b43145c
        rcu/rcu.next-3f26dd3

They all start with an error

[    0.131005] initcall tracer_alloc_buffers+0x0/0x22e returned with preemption imbalance

Then followed by lots of scheduling while atomic bugs:

[    0.143206] Brought up 2 CPUs
[    0.144996] ----------------
[    0.145989] | NMI testsuite:
[    0.147988] --------------------
[    0.149238]   remote IPI:
[    0.149239] BUG: scheduling while atomic: swapper/0/1/0x00000002
[    0.151082] no locks held by swapper/0/1.
[    0.152991] Modules linked in:
[    0.153992] Pid: 1, comm: swapper/0 Not tainted 3.5.0-rc3+ #24
[    0.155990] Call Trace:
[    0.157056]  [<ffffffff81a21ee1>] __schedule_bug+0x67/0x75
[    0.158999]  [<ffffffff81a3717c>] __schedule+0xb0/0x8b0
[    0.160995]  [<ffffffff810b8fe2>] ? save_trace+0x3d/0xae
[    0.162993]  [<ffffffff8105ad19>] ? pvclock_clocksource_read+0x4d/0xc3
[    0.164992]  [<ffffffff810bbdff>] ? mark_lock+0x23c/0x502
[    0.166993]  [<ffffffff8105a10b>] ? kvm_clock_read+0x2e/0x36
[    0.168991]  [<ffffffff81a37c61>] schedule+0x65/0x67
[    0.170992]  [<ffffffff81a35542>] schedule_timeout+0x2e/0x22a
[    0.172991]  [<ffffffff810b8c1c>] ? trace_hardirqs_off+0xd/0xf
[    0.174992]  [<ffffffff8109c890>] ? local_clock+0x41/0x5a
[    0.176995]  [<ffffffff81a392f0>] ? _raw_spin_unlock_irq+0x2e/0x48
[    0.178990]  [<ffffffff810bd647>] ? trace_hardirqs_on_caller+0x125/0x181
[    0.180995]  [<ffffffff810bd6b0>] ? trace_hardirqs_on+0xd/0xf
[    0.182989]  [<ffffffff81a37b06>] wait_for_common+0x10a/0x140
[    0.184989]  [<ffffffff8109a2f2>] ? try_to_wake_up+0x2cb/0x2cb
[    0.186990]  [<ffffffff810ea4c6>] ? call_rcu_bh+0x19/0x19
[    0.188988]  [<ffffffff81a37bfa>] wait_for_completion+0x1d/0x1f
[    0.190989]  [<ffffffff8108752d>] wait_rcu_gp+0x4d/0x54
[    0.192988]  [<ffffffff81087534>] ? wait_rcu_gp+0x54/0x54
[    0.194987]  [<ffffffff81a37a40>] ? wait_for_common+0x44/0x140
[    0.196987]  [<ffffffff810e8713>] synchronize_sched+0x50/0x52
[    0.198987]  [<ffffffff8103b272>] unregister_nmi_handler+0xc6/0xd5
[    0.200989]  [<ffffffff8142fa79>] ? delay_tsc+0x49/0xa5
[    0.202988]  [<ffffffff82259db7>] test_nmi_ipi.constprop.2+0x72/0x85
[    0.204986]  [<ffffffff82259dca>] ? test_nmi_ipi.constprop.2+0x85/0x85
[    0.206986]  [<ffffffff82259e27>] remote_ipi+0x5d/0x62
[    0.208985]  [<ffffffff82259ce2>] dotest.constprop.1+0x6/0x69
[    0.210985]  [<ffffffff82259ede>] nmi_selftest+0x60/0x16d
[    0.212986]  [<ffffffff822558da>] native_smp_cpus_done+0x1c/0x111
[    0.214989]  [<ffffffff8226198b>] smp_init+0xe2/0xeb
[    0.216986]  [<ffffffff82248c0f>] kernel_init+0xa9/0x1c0
[    0.218984]  [<ffffffff810bd538>] ? trace_hardirqs_on_caller+0x16/0x181
[    0.220989]  [<ffffffff81a417b4>] kernel_thread_helper+0x4/0x10
[    0.222989]  [<ffffffff81a396f0>] ? retint_restore_args+0x13/0x13
[    0.224990]  [<ffffffff82248b66>] ? start_kernel+0x3a3/0x3a3
[    0.226983]  [<ffffffff81a417b0>] ? gs_change+0x13/0x13
[    0.231002]   ok  |
[    0.231998]    local IPI:
[    0.231998] BUG: scheduling while atomic: swapper/0/1/0x00000003

...

The first bad commit is

commit b43145c912b7dabd190ec45cc31c2b96d64715ee

Author: Paul E. McKenney <[email protected]>
Date:   Tue Jun 19 11:58:27 2012 -0700

rcu: Disable preemption in rcu_blocking_is_gp()

It is time to optimize CONFIG_TREE_PREEMPT_RCU's synchronize_rcu() for uniprocessor optimization, which means that rcu_blocking_is_gp() can no longer rely on RCU read-side critical sections having disabled preemption. This commit therefore disables preemption across rcu_blocking_is_gp()'s scan of the cpu_online_mask.

Signed-off-by Paul E. McKenney <[email protected]>

kernel/rcutree.c +0 -0

index ce175b6..c0cc41f 100644

Review

Quoted 7 lines

static inline int rcu_blocking_is_gp(void)

 {
        might_sleep();  /* Check for RCU read-side critical section. */

(quoted diff) +2 -0

+       preempt_disable();
        return num_online_cpus() <= 1;
+       preempt_enable();
 }

Thank you! I have no idea how a preempt_disable() causes that badness to happen, but this commit is not yet critically important, so I will drop it.

Quoted 8 lines

Quoted 7 lines

static inline int rcu_blocking_is_gp(void)

 {
        might_sleep();  /* Check for RCU read-side critical section. */

(quoted diff) +2 -0

+       preempt_disable();
        return num_online_cpus() <= 1;
+       preempt_enable();
 }

Quoted 3 lines

Thank you! I have no idea how a preempt_disable() causes that badness to happen, but this commit is not yet critically important, so I will drop it.

preempt_enable() becomes dead code because of the return statement? I wonder why gcc didn't issue a warning (or I failed to catch it)...

Quoted 8 lines

Quoted 7 lines

 static inline int rcu_blocking_is_gp(void)
 {
        might_sleep();  /* Check for RCU read-side critical section. */
+       preempt_disable();
        return num_online_cpus() <= 1;
+       preempt_enable();
 }

Quoted 3 lines

Thank you!  I have no idea how a preempt_disable() causes that badness to happen, but this commit is not yet critically important, so I will drop it.

preempt_enable() becomes dead code because of the return statement? I wonder why gcc didn't issue a warning (or I failed to catch it)...

Quoted 8 lines · Fengguang Wu

Quoted 8 lines

Quoted 7 lines

 static inline int rcu_blocking_is_gp(void)
 {
        might_sleep();  /* Check for RCU read-side critical section. */
+       preempt_disable();
        return num_online_cpus() <= 1;
+       preempt_enable();
 }

Quoted 7 lines

Quoted 3 lines

Thank you!  I have no idea how a preempt_disable() causes that badness to happen, but this commit is not yet critically important, so I will drop it.

preempt_enable() becomes dead code because of the return statement? I wonder why gcc didn't issue a warning (or I failed to catch it)...

gcc has an option -Wunreachable-code, but we don't enable it when building kernel, nor it will be enabled with -Wall. If we enable it, we will have many false-positives as we have lots of debugging code which is not reachable unless we enable some debugging option.

However, when I test it manually with the following code:

~% cat /tmp/unreachable.c
int main(void)

{
	int a = 0;
	a++;
	return ++a;
	a++;
}

~% gcc -Wunreachable-code -O0 -c /tmp/unreachable.c

gcc still doesn't give me any warning for the last line of the code, gcc optimizes it out silently, I am wondering if this is a gcc bug.

But in my case, the trailing preempt_enable() should not have been optimized away, right? Wouldn't it be more like the following?

int a = 0;
int main(void)

{
	a++;
	return ++a;
	a++;
}

Hmmm... But this -still- doesn't emit any warnings.

Quoted 11 lines · Paul E. McKenney

But in my case, the trailing preempt_enable() should not have been optimized away, right? Wouldn't it be more like the following?

int a = 0;
int main(void)

{
	a++;
	return ++a;
	a++;
}

Quoted 2 lines

Hmmm... But this -still- doesn't emit any warnings.

gcc removed -Wunreachable-code a couple years ago. It doesn't complain when you pass it for backwards compatibility with old Makefiles.