Re: [PATCH 5/8] rcu: eliminate deadlock for rcu read site

[Lai Jiangshan <laijs@cn.fujitsu.com>]

On 08/21/2013 11:17 AM, Paul E. McKenney wrote:
> On Sat, Aug 10, 2013 at 08:07:15AM -0700, Paul E. McKenney wrote:
>> On Sat, Aug 10, 2013 at 11:43:59AM +0800, Lai Jiangshan wrote:
> 
> [ . . . ]
> 
>>> So I have to narrow the range of suspect locks. Two choices:
>>> A) don't call rt_mutex_unlock() from rcu_read_unlock(), only call it
>>>    from rcu_preempt_not_context_switch(). we need to rework these
>>>    two functions and it will add complexity to RCU, and it also still
>>>    adds some probability of deferring.
>>
>> One advantage of bh-disable locks is that enabling bh checks
>> TIF_NEED_RESCHED, so that there is no deferring beyond that
>> needed by bh disable.  The same of course applies to preempt_disable().
>>
>> So one approach is to defer when rcu_read_unlock_special() is entered
>> with either preemption or bh disabled.  Your current set_need_resched()
>> trick would work fine in this case.  Unfortunately, re-enabling interrupts
>> does -not- check TIF_NEED_RESCHED, which is why we have latency problems
>> in that case.  (Hence my earlier question about making self-IPI safe
>> on all arches, which would result in an interrupt as soon as interrupts
>> were re-enabled.)
>>
>> Another possibility is to defer only when preemption or bh are disabled
>> on entry ro rcu_read_unlock_special(), but to retain the current
>> (admittedly ugly) nesting rules for the scheduler locks.
> 
> Would you be willing to do a patch that deferred rt_mutex_unlock() in
> the preempt/bh cases?  This of course does not solve the irq-disable
> case, but it should at least narrow the problem to the scheduler locks.
> 
> Not a big hurry, given the testing required, this is 3.13 or 3.14 material,
> I think.
> 
> If you are busy, no problem, I can do it, just figured you have priority
> if you want it.
> 
> 	


I'm writing a special rt_mutex_unlock() for rcu deboost only.
I hope Steven accept it.

Thanks,
Lai