Re: new warning on sysrq kernel crash trigger

[Ani Sinha]

On Fri, Dec 11, 2015 at 2:27 PM, Paul E. McKenney
<paulmck@xxxxxxxxxxxxxxxxxx> wrote:
> On Fri, Dec 11, 2015 at 05:10:43PM -0500, Rik van Riel wrote:
>> On 12/11/2015 03:44 PM, Ani Sinha wrote:
>> >
>> >
>> > On Thu, 10 Dec 2015, Paul E. McKenney wrote:
>> >
>> >> On Thu, Dec 10, 2015 at 03:57:09PM -0800, Ani Sinha wrote:
>> >>> Hi guys
>> >>>
>> >>> I am noticing a new warning in linux 3.18 which we did not see before
>> >>> in linux 3.4 :
>> >>>
>> >>> bash-4.1# echo c > /proc/sysrq-trigger
>> >>> [  978.807185] BUG: sleeping function called from invalid context at
>> >>> ../arch/x86/mm/fault.c:1187
>> >>> [  978.909816] in_atomic(): 0, irqs_disabled(): 0, pid: 4706, name: bash
>> >>> [  978.987358] Preemption disabled at:[<ffffffff81484339>] printk+0x48/0x4a
>> >>>
>> >>>
>> >>> I have bisected this to the following change :
>> >>>
>> >>> commit 984d74a72076a12b400339973e8c98fd2fcd90e5
>> >>> Author: Rik van Riel <riel@xxxxxxxxxx>
>> >>> Date:   Fri Jun 6 14:38:13 2014 -0700
>> >>>
>> >>>     sysrq: rcu-ify __handle_sysrq
>> >>>
>> >>>
>> >>> the rcu_read_lock() in handle_sysrq() bumps up
>> >>> current->rcu_read_lock_nesting. Hence, in __do_page_fault() when it
>> >>> calls might_sleep() in x86/mm/fault.c line 1191,
>> >>> preempt_count_equals(0) returns false and hence the warning is
>> >>> printed.
>> >>>
>> >>> One way to handle this would be to do something like this:
>> >>>
>> >>> diff --git a/arch/x86/mm/fault.c b/arch/x86/mm/fault.c
>> >>> index eef44d9..d4dbe22 100644
>> >>> --- a/arch/x86/mm/fault.c
>> >>> +++ b/arch/x86/mm/fault.c
>> >>> @@ -1132,7 +1132,7 @@ __do_page_fault(struct pt_regs *regs, unsigned
>> >>> long error_code,
>> >>>   * If we're in an interrupt, have no user context or are running
>> >>>   * in a region with pagefaults disabled then we must not take the fault
>> >>>   */
>> >>> - if (unlikely(faulthandler_disabled() || !mm)) {
>> >>> + if (unlikely(faulthandler_disabled() || rcu_preempt_depth() || !mm)) {
>> >>
>> >> This works if CONFIG_PREEMPT=y, but if CONFIG_PREEMPT=n, then
>> >> rcu_preempt_depth() unconditionally returns zero.  And if
>> >> CONFIG_PREEMPT_COUNT=y && CONFIG_PREEMPT=n, you would still see
>> >> the might_sleep() splat.
>> >>
>> >> Maybe use SRCU instead of RCU for this purpose?
>> >>
>> >
>> > From ae232ce3fb167b2ad363bfac7aab69001bc55a50 Mon Sep 17 00:00:00 2001
>> > From: Ani Sinha <ani@xxxxxxxxxx>
>> > Date: Fri, 11 Dec 2015 12:07:42 -0800
>> > Subject: [PATCH 1/1] Fix 'sleeping function called from invalid context'
>> >  warning in sysrq generated crash.
>> >
>> > Commit 984d74a72076a1 ("sysrq: rcu-ify __handle_sysrq")
>> > replaced spin_lock_irqsave() calls with
>> > rcu_read_lock() calls in sysrq. Since rcu_read_lock() does not
>> > disable preemption, faulthandler_disabled() in
>> > __do_page_fault() in x86/fault.c returns false. When the code
>> > later calls might_sleep() in the pagefault handler, we get the
>> > following warning:
>> >
>> > BUG: sleeping function called from invalid context at ../arch/x86/mm/fault.c:1187
>> > in_atomic(): 0, irqs_disabled(): 0, pid: 4706, name: bash
>> > Preemption disabled at:[<ffffffff81484339>] printk+0x48/0x4a
>> >
>> > To fix this, replace RCU call in handle_sysrq() to use SRCU.
>>
>> The sysrq code can be called from irq context.
>>
>> Trying to use SRCU from an irq context sounds like it could
>> be a bad idea, though admittedly I do not know enough about
>> SRCU to know for sure :)
>
> Indeed, not the best idea!  ;-)
>
> I could imagine something like this:
>
>         if (in_irq())
>                 rcu_read_lock();
>         else
>                 idx = srcu_read_lock(&sysrq_rcu);
>
> And ditto for unlock.  Then, for the update:
>
>         synchronize_rcu_mult(call_rcu, call_sysrq_srcu);

This won't work on 3.18 as this api was introduced in linux 4.3.

>
> Where:
>
>         static void call_sysrq_srcu(struct rcu_head *head, rcu_callback_t func)
>         {
>                 call_srcu(&sysrq_rcu, head, func);
>         }
>
> Here I presume that the page-fault code avoids the might_sleep if invoked
> from irq context.

Quick look at the code seems to indicate that this is true.

>
> Thoughts?
>
>                                                         Thanx, Paul
>
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/