Re: [PATCH RT 3.18] irq_work: Provide a soft-irq based queue

[Jan Kiszka <jan.kiszka@siemens.com>]

On 2015-04-24 08:54, Mike Galbraith wrote:
> On Thu, 2015-04-23 at 17:00 -0400, Steven Rostedt wrote:
>> On Thu, 23 Apr 2015 09:19:26 +0200
>> Mike Galbraith <umgwanakikbuti@gmail.com> wrote:
>>
>>>>   CC      kernel/irq_work.o
>>>> In file included from ../include/asm-generic/percpu.h:6:0,
>>>>                  from ../arch/x86/include/asm/percpu.h:522,
>>>>                  from ../arch/x86/include/asm/current.h:5,
>>>>                  from ../arch/x86/include/asm/processor.h:15,
>>>>                  from ../arch/x86/include/asm/irq_work.h:4,
>>>>                  from ../include/linux/irq_work.h:47,
>>>>                  from ../kernel/irq_work.c:11:
>>>> ../kernel/irq_work.c: In function ‘irq_work_queue_on’:
>>>> ../kernel/irq_work.c:85:17: error: ‘hirq_work_list’ undeclared 
>>>> (first use in this function)
>>>>         &per_cpu(hirq_work_list, cpu));
>>>
>>> Aw poo, so that's just what I _thought_ it was for.
>>
>> It helps optimization but does nothing for undefined symbols.
>>
>> That said, why don't we clean up that irq_work code and at least
>> declare both lists, and get rid of all the #ifdefs. I wonder if gcc is
>> smart enough to not allocate a static variable if it happens to be
>> optimized out?
> 
> Nope, it didn't notice a thing.
> 
> This is a stab at that cleanup.  Usable as is with Jan's ok, or as
> fodder for your bitmaster-9000 patch shredder, or whatever.  Box works
> and it makes line count shrink...
> 
> I downgraded evolution v3.16->v3.12 to restore its ability to read it's
> own fscking "Preformatted" switch, so whitespace should be fine. 
> 
> Oh, btw, if anyone (else) makes a 4.1-rt, your rt push work will want
> one of those nifty hirq tags lest box make boom due to trying to do that
> not only way late, but with irqs enabled which pisses sched all off.
> 
> Subject: [PATCH RT 3.18] irq_work: Provide a soft-irq based queue
> Date:	Thu, 16 Apr 2015 18:28:16 +0200
> From:	Jan Kiszka <jan.kiszka@siemens.com>
> 
> Instead of turning all irq_work requests into lazy ones on -rt, just
> move their execution from hard into soft-irq context.
> 
> This resolves deadlocks of ftrace which will queue work from arbitrary
> contexts, including those that have locks held that are needed for
> raising a soft-irq.
> 
> Mike: cleanup ifdef mess and kill hirq_work_list.  We need two lists,
> and already have them, merely need to select according to work type.
> In -rt all work not tagged for hirq execution is queued to the lazy
> list and runs via irq_work_tick().  Raising SOFTIRQ_TIMER is always
> done via IPI for deadlock safety, if the work item is not a lazy work
> or the tick is stopped, fire IPI immediately, otherwise let it wait.
> IOW, lazy work is lazy in -rt only until someone queues immediate work.

The approach looks good to me, but the commit log deserves a rework now.

Jan

-- 
Siemens AG, Corporate Technology, CT RTC ITP SES-DE
Corporate Competence Center Embedded Linux