On Mon, Sep 21, 2015 at 03:01:42PM -0400, Steven Rostedt wrote: > On Mon, 21 Sep 2015 12:04:16 -0500 > Felipe Balbi <balbi@xxxxxx> wrote: > > > > no, no. you're missing the point here. The problem is that when RT > > is applied, spinlocks get reimplemented as RT-aware mutexes which > > works pretty well as long as you don't install your own top and bottom > > halves. If you do, RT patch can't force your handler to run as a thread > > and you're left with, essentially, a mutex for synchronization in > > hardirq context. > > > > So the question is really for Steve, what should we do here ? Use a > > raw_spinlock ? I'd like to avoid that if possible. If we have another > > option, I'm all ears. > > > > I may be missing something here. Yes, in RT spinlocks become rtmutexes > and can sleep. But in RT all interrupts become threads. With the slight > exception that if you declare your interrupt handle as a thread, then > you do have a top half handler. This top half should not be grabbing > any spin locks, and should simply disable any more interrupts from > happening on the device until the thread handler can run. and that's basically we're doing. $SUBJECT is removing the spin lock for that reason. Register accesses should be satomic anyway, right?? -- balbi
Attachment:
signature.asc
Description: PGP signature
