On Mon, 2008-08-11 at 08:09 +0200, Peter Zijlstra wrote: > On Mon, 2008-08-11 at 08:59 +0800, Huang Ying wrote: > > On Fri, 2008-08-08 at 12:13 +0200, Peter Zijlstra wrote: > > > On Fri, 2008-08-08 at 14:52 +0800, Huang Ying wrote: > > > > Replace local_irq_disable() with raw_local_irq_disable() to prevent > > > > lockdep complain. > > > Uhhm, please provide more information - just using raw_* to silence > > > lockdep is generally the wrong thing to do. > > > > In traditional kexec, the new kernel will replace current one, so the > > irq is simply disabled. But now jumping back from kexeced kernel is > > supported, so the irq should be enabled again. > > > > The code sequence of irq during kexec jump is as follow: > > > > local_irq_disable(); /* in kernel_kexec() */ > > local_irq_disable(); /* in machine_kexec() */ > > local_irq_enable(); /* in kernel_kexec() */ > > > > The disable and enable is not match. Maybe another method is to use > > local_irq_save(), local_irq_restore() pair in machine_kexec(), so the > > disable and enable is matched. > > And its the machine kernel's lockdep instance that goes complain? > > whichever annotation gets used - and I think I can agree that raw_* > might be approriate there, this should be accompanied with a rather > elaborate changelog and preferably a comment in the code too. Without > such we'll be wondering in the years to come WTH happens here. Sorry, I find there is no complain from lockdep. Un-paired irq disable/enable has no problem with lockdep, just increase something such as "redundant_hardirqs_off". Please ignore this thread. Best Regards, Huang Ying
