[ Added Peter Update: Laurent noticed that sysrq 'n' (nice-all-RT-tasks) calls __sched_setscheduler() form interrupt context. At the start of that function, there's a BUG_ON(in_interrupt()). The reason for that was due to the rt mutex pi code calling wait_lock. Which was not irq safe. Now it is, but that's not good enough. ]
On Wed, 8 Mar 2017 18:03:55 +0100 Laurent Dufour <lduf...@linux.vnet.ibm.com> wrote: > On 08/03/2017 17:57, Steven Rostedt wrote: > > On Wed, 8 Mar 2017 11:51:14 -0500 > > Steven Rostedt <rost...@goodmis.org> wrote: > > > > > >> Hmm, that commit was added in 2.6.18, and you're right, a lot has > >> changed since then. Have you tried removing it and running it under > >> lockdep, and see if it triggers any warnings? > > > > I did a little digging, and it appears that its the rt mutex wait lock > > that the comment was referring to. Today that spin lock is irq safe. I > > believe its safe to remove the BUG_ON(). Want me to send a patch? > > Sure, go ahead ;) > Actually, it's still not safe :-/ I just noticed this in the call path: raw_spin_unlock_irq(&task->pi_lock); As well as other raw_spin_unlock_irq()s. Which would enable interrupts regardless of the previous state. One solution is to change all those to irqsave() but that seems to be a big step for something that is rarely done (how many years has it been since 2.6.18?). I wonder if we should just have a special flag sent by that sysrq trigger. Since it is causing all tasks to go "nice" there's no need to do the pi chain walk in __sched_setscheduler(). -- Steve