[PATCH for 4.12] Revert "pinctrl: rockchip: avoid hardirq-unsafe functions in irq_chip"

[Paul E. McKenney]

On Sat, Jun 24, 2017 at 12:12:49AM +0200, Thomas Gleixner wrote:
> On Fri, 23 Jun 2017, Brian Norris wrote:
> 
> > This reverts commit 88bb94216f59e10802aaf78c858a4146085faf18.
> > 
> > It introduced a new CONFIG_DEBUG_ATOMIC_SLEEP warning in v4.12-rc1:
> > 
> > [ 7226.716713] BUG: sleeping function called from invalid context at kernel/locking/mutex.c:238
> > [ 7226.716716] in_atomic(): 0, irqs_disabled(): 0, pid: 1708, name: bash
> > [ 7226.716722] CPU: 1 PID: 1708 Comm: bash Not tainted 4.12.0-rc6+ #1213
> > [ 7226.716724] Hardware name: Google Kevin (DT)
> > [ 7226.716726] Call trace:
> > [ 7226.716738] [<ffffff8008089928>] dump_backtrace+0x0/0x24c
> > [ 7226.716743] [<ffffff8008089b94>] show_stack+0x20/0x28
> > [ 7226.716749] [<ffffff8008371370>] dump_stack+0x90/0xb0
> > [ 7226.716755] [<ffffff80080cd2a0>] ___might_sleep+0x10c/0x124
> > [ 7226.716760] [<ffffff80080cd330>] __might_sleep+0x78/0x88
> > [ 7226.716765] [<ffffff800879e210>] mutex_lock+0x2c/0x64
> > [ 7226.716771] [<ffffff80083ad678>] rockchip_irq_bus_lock+0x30/0x3c
> > [ 7226.716777] [<ffffff80080f6d40>] __irq_get_desc_lock+0x78/0x98
> > [ 7226.716782] [<ffffff80080f7e6c>] irq_set_irq_wake+0x44/0x12c
> > [ 7226.716787] [<ffffff8008486e18>] dev_pm_arm_wake_irq+0x4c/0x58
> > [ 7226.716792] [<ffffff800848b80c>] device_wakeup_arm_wake_irqs+0x3c/0x58
> > [ 7226.716796] [<ffffff80084896fc>] dpm_suspend_noirq+0xf8/0x3a0
> > [ 7226.716800] [<ffffff80080f1384>] suspend_devices_and_enter+0x1a4/0x9a8
> > [ 7226.716803] [<ffffff80080f21ec>] pm_suspend+0x664/0x6a4
> > [ 7226.716807] [<ffffff80080f04d8>] state_store+0xd4/0xf8
> > ...
> > 
> > It was reported on -rc1, and it's still not fixed in -rc6, so it should
> > just be reverted.
> > 
> > + Thomas, in case he has thoughts
> 
> + Peter and Paul, Tony
> 
> > Subject was "[4.12 REGRESSION] pinctrl: rockchip: sleeping function
> > called from atomic context"
> > 
> > On Fri, May 26, 2017 at 07:19:00PM -0700, Brian Norris wrote:
> > > Any thoughts? Revert the offending patch? I can spend a little more time
> > > next week trying to debug what's actually going on if needed.
> > > 
> > > On Wed, May 17, 2017 at 03:56:34PM -0700, Brian Norris wrote:
> > 
> > > > The thing is, the documentation (and apparent design) suggest that
> > > > calling sleeping functions from ->irq_bus_lock() is perfectly valid. I'm
> > > > not 100% following the ___might_sleep() logic, but is this complaining
> > > > because of the RCU read locking in device_wakeup_arm_wake_irqs()? I have
> > > > CONFIG_PREEMPT_RCU and CONFIG_PREEMPT enabled, FWIW.
> 
> Sigh, The real wreckage happened in commit:
> 
> commit 4990d4fe327b9d9a7a3be7103a82699406fdde69
> Author: Tony Lindgren <tony at atomide.com>
> Date:   Mon May 18 15:40:29 2015 -0700
> 
>     PM / Wakeirq: Add automated device wake IRQ handling
> 
> which added that RCU locking stuff and thereby broke the long existing
> bus_lock() facility of the interrupt core.
> 
> irq_bus_lock/unlock was explicitely made to allow sleeping locks for
> interrupt chips which hang behind slow busses like i2c or spi. It took us
> quite some effort to get this done and that patch broke it permanently.
> 
> I'm not sure what to do here. This is an ever recurring issue simply
> because RT requires that sleeping locks can be taken inside rcu locked
> regions. So sooner than later we need a resoilution for that problem.

The usual advice would be for 4990d4fe327b ("PM / Wakeirq: Add automated
device wake IRQ handling") to use SRCU rather than RCU.  Is there some
reason that won't work?

And yes, my commit 5b72f9643b52a ("rcu: Complain if blocking in
preemptible RCU read-side critical section") that is now in -tip needs
adjustment for -rt.  I can easily add "&& IS_ENABLED(CONFIG_PREEMPT_RT)"
or whatever the current incantation is.  Just let me know!

							Thanx, Paul