Re: [PATCH] irq: add a new function irq_in_progress to check pending IRQ handlers

[Yanmin Zhang <yanmin_zhang@linux.intel.com>]

On Thu, 2013-06-06 at 18:02 -0700, Greg KH wrote:
> On Fri, Jun 07, 2013 at 08:53:29AM +0800, Yanmin Zhang wrote:
> > On Thu, 2013-06-06 at 15:18 +0200, Thomas Gleixner wrote:
> > > On Thu, 6 Jun 2013, shuox.liu@intel.com wrote:
> > > > From: Zhang Yanmin <yanmin.zhang@intel.com>
> > > > 
> > > > synchronize_irq waits pending IRQ handlers to be finished. If using this
> > > > function while holding a resource, the IRQ handler may cause deadlock.
> > > > 
> > > > Here we add a new function irq_in_progress which doesn't wait for the handlers
> > > > to be finished.
> > > > 
> > > > A typical use case at suspend-to-ram:
> > > > 
> > > > device driver's irq handler is complicated and might hold a mutex at rare cases.
> > > > Its suspend function is called and a suspended flag is set.
> > > > In case its IRQ handler is running, suspend function calls irq_in_progress. if
> > > > handler is running, abort suspend.
> > > > The irq handler checks the suspended flag. If the device is suspended, irq handler
> > > > either ignores the interrupt, or wakes up the whole system, and the driver's
> > > > resume function could deal with the delayed interrupt handling.
> > > 
> > > This is as wrong as it can be. Fix the driver instead of hacking racy
> > > functions into the core code.
> > > 
> > > So your problem looks like this:
> > > 
> > > CPU 0				CPU 1
> > > irq_handler_thread()		suspend()
> > >    .....			mutex_lock(&m);
> > >    mutex_lock(&m);		synchronize_irq();
> > > 
> > > So why needs the mutex to be taken before synchronize_irq()? Why not
> > > doing the obvious?
> > > 
> > > suspend()
> > >   disable_irq(); (Implies synchronize_irq)
> > >   mutex_lock(&m);
> > >   ....
> > >   mutex_unlock(&m);
> > >   enable_irq();
> > Thanks for the kind comment.
> > 
> > We do consider your solution before and it works well indeed with some specific
> > simple drivers. For example, some drives use GPIO pin as interrupt source.
> > 
> > On one specific platform, disable_irq would really disable the irq at RTE entry,
> > which means we lose the wakeup capability of this device.
> > synchronize_irq can be another solution. But we did hit 'DPM device timeout' issue
> > reported by dpm_wd_handler at suspend-to-ram.
> > 
> > The driver is complicated. We couldn't change too many functions to optimize it.
> > In addition, we have to use the driver instead of throwing it away.
> 
> What is preventing you from rewriting it to work properly?
It's complicated.

> 
> > With irq_in_progress, we can resolve this issue and it does work, although it
> > looks like ugly.
> 
> Don't paper over driver bugs in the core kernel, fix the driver.
It's hard to say it's a driver bug. Could generic kernel provide some flexibility
for such complicated drivers?

For example, any driver's suspend can return error and the whole suspend-to-ram
aborts. Can we say the driver has a bug? If so, why not to change all suspend/resume
callbacks to return VOID?
Current kernel already allows drivers to abort suspend-to-ram at some rare corner cases.
Of course, if the abort happens frequently, it's a bug and we need fix it in driver.
If it happens rarely, can we provide some flexibility for the driver?

Thanks,
Yanmin