Messages in this thread | | | Subject | Re: Considerations on sched APIs under RT patch | From | Peter Zijlstra <> | Date | Wed, 21 Apr 2010 22:58:16 +0200 |
| |
On Wed, 2010-04-21 at 22:38 +0200, Primiano Tucci wrote: > > No, any syscall can end up blocking/scheduling there are no exceptions. > > But blocking doesn't mean its non-deterministic, esp. when coupled with > > things like PI. > > > > But you do have to treat system resources as such, that is they can (and > > will) create cross-cpu dependencies, if you do not take that into > > account you will of course be surprised. > > > I actually don't understand why do you recall PI so frequently, it > seems to be the unique point of interest.
PI keeps preemptible locks working in a RT environment. Non-preemptible or preemptible+PI are both valid RT constructs that can be analyzed
> Actually I take care about not sharing cross-cpu resources, but I > cannot take care of what the kernel should do.
An SMP kernel must be treated as a cross-cpu resource. There's just no way around that. For instance, Unix allows two processes on different cpus to invoke sched_setscheduler/sched_setaffinity or any number of system calls on the same target process. Filesystems are shared etc..
> In my viewpoint is unacceptable that the scheduler apis can led into a > rescheduling.
They can even lead to pagefaults and disk IO if you're not careful.
I'm not sure if there are blocking locks left thereabout, but spinlocks or rt_mutex, both create cross-cpu dependencies that need to be analyzed, !preempt isn't magic in any way.
> It voids any form of process control. > If I lose the control while controlling other processes, Quis > custodiet ipsos custodes? > > P.S. It actually does not happen in other RTOSes, e.g., VxWorks SMP
I don't know any of those, but its impossible to migrate tasks from one cpu to another without creating cross-cpu dependencies.
Whether locks are preemptible or not doesn't make them any less analyzable, if you use system-calls in your RT program, their implementation needs to be considered.
| |