Re: schedule_delayed_work with a 0 jiffy delay

[Joe Lawrence]

On Sun, 28 Sep 2014, Tejun Heo wrote:

> Hello, Joe.
> 
> On Fri, Sep 26, 2014 at 10:52:27AM -0400, Joe Lawrence wrote:
> ...
> > I was wondering if such behavior was expected on !PREEMPT kernels,
> > especially after b22ce2785d97 "workqueue: cond_resched() after
> > processing each work item".  In the ftraces I've observed from the RCU
> > stall, cond_resched() never seems to schedule.  Under what conditions
> > would it be expected to do so?
> 
> Hmmm... it should yield if there's a higher priority task scheduled.
> Maybe the workqueue in question is a highpri one?

I don't believe any high priority tasks or workqueues are involved here.  
Consider a synthetic repro module like:

#include <linux/module.h>
#include <linux/sched.h>
#include <linux/workqueue.h>

struct delayed_work dwork;
struct delayed_work dwork2;

static void warn_thread(struct work_struct *work)
{
        schedule_delayed_work_on(4, &dwork, 0);
}

static void warn_thread2(struct work_struct *work)
{
        schedule_delayed_work_on(4, &dwork2, 0);
}

int __init init_module(void)
{
        INIT_DELAYED_WORK(&dwork, warn_thread);
        schedule_delayed_work_on(4, &dwork, 1000);

        INIT_DELAYED_WORK(&dwork2, warn_thread2);
        schedule_delayed_work_on(4, &dwork2, 1000);

        pr_info("%s loaded.\n", THIS_MODULE->name);

        return 0;
}

void cleanup_module(void)
{
        cancel_delayed_work_sync(&dwork);
        cancel_delayed_work_sync(&dwork2);

        pr_info("%s unloaded.\n", THIS_MODULE->name);
}

When I load this, "kworker/4:1" goes to 100% CPU and the RCU grace 
periods reported by rcusched freeze on CPU4 for minutes at a time.
 
> > Changing the worker function to reschedule with at least 1 jiffy avoids
> > this problem -- I was just curious if a fix at that end was appropriate.
> 
> The fact that 0 delay means immediate execution is depended upon in
> multiple places and changing it will break other things.  The
> workqueue user itself sounds buggy to me.  Can't that be fixed?

Certainly modifying the workqueue user should be sufficient here.  I was 
only wondering if the cond_resched in process_one_work was expected to 
give the CPU some breathing room in this scenario.

Thanks,

-- Joe
--
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/