Chris Friesen wrote:
Douglas Niehaus wrote:
(1.1) Will the use of system services (system calls) by RT threads
need to be explicitly supported by, perhaps, explicitly making the
schedulers of *other* threads also using those system calls aware of
that and take it into account to make those other tasks non-preemptible
while holding system call related locks.
(1.2) Can Linux SMP ever support this in an acceptable manner since
locks associated with systems services are shared across CPU boundaries.
THus, I can isolate a set of RT tasks on a CPU easily, and they are well
isolated and can run under strict predictability, *until they use a
system call that uses a lock*. Then, the system call is an interaction
channel with every other thread on the system using the same system call.
This may be a terminology issue, but I think it would be more correct to
say that the lock is the interaction channel, not the system call, and
it is an interaction channel with every other entity on the system using
the same lock. Depending on the specific lock, this could be other
userspace tasks, kernel threads, or even hardware interrupt handlers.
Sorry, sloppiness on my part while typing quickly, resulting in the
terminolgy problem of --- my using the wrong terminology.
Yes, the lock, is the interaction channel.
Admittedly, which locks are the interaction channels is correlated with
which system services are used by threads, but sometimes more and
sometimes less strongly correlated.
When I explain it to less expert audiences than this, I tend to talk in
terms of the system services because they at least know what some of
them are, while many have no idea what the concurrency control in the
kernel is. They can fairly easily see that if RT tasks use a range of
services used by generic Linux tasks, then some interaction between RT
and non-RT tasks is a result.
Still, no excuses, only explanation. Sorry to have over-simplified to
this audience.
Thanks for clarifying.
This goes back to your first question of which system services an RT
task can use while maintaining schedulability analysis. Unfortunately
this may be a moving target, since the exact answer depends on what
locks are taken in the underlying kernel implementation.
Chris
Yes. This is true and is also the point I was trying to make. When
talking to people about RT over the years I have observed that it is
often hard to communicate the full cost of the predictability required
for RT tasks
Extremely detailed information is required, and getting it can be
expensive. This is one reason why supporting RT in Linux proper is even
harder than it first appears.
However, I think that while completely arbitrary use of Linux system
services by RT tasks is extremely complicated, many RT applications can
be happy using only a small subset of services, and so various classes
of applications can be supported successfully with merely extreme
effort, as opposed to completely insane effort.
It is a really hard problem, no doubt, though.
Doug
--
To unsubscribe from this list: send the line "unsubscribe linux-rt-users" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
