Subject: What's the relationship between Linux driver running priority and FIFO policy?

Hi
I have a question disturb me very much.
You know linux FIFO policy is common,and also there are many drivers
running in system,so
which level are these drivers running on? If I have a progrom which
priority is FIFO99, does it disturb the driver running?
I want know which level does drivers running on? For example, driver
running on FIFOXX? and how does linux schedule on this
case?

Thanks!
Best Regards!

---------------------------
ZhangYongChao
2010.07.01
---------------------------


2010-07-01 07:36:18

by Peter Zijlstra

[permalink] [raw]
Subject: Re: What's the relationship between Linux driver running priority and FIFO policy?

On Thu, 2010-07-01 at 13:21 +0800, Zhang, Yongchao (GE Healthcare)
wrote:
> Hi
> I have a question disturb me very much.
> You know linux FIFO policy is common,and also there are many drivers
> running in system,so
> which level are these drivers running on? If I have a progrom which
> priority is FIFO99, does it disturb the driver running?
> I want know which level does drivers running on? For example, driver
> running on FIFOXX? and how does linux schedule on this
> case?
>

Fully depends on what a driver is and what you expect it to. If part of
the driver consists of a (or multiple) kernel thread, and that thread
has lower priority than FIFO99 (very typical) then of course.

If the driver is a pure IRQ driven thingy, then no, since hardirq
context will still preempt any task context.

Of course, if whatever proglet is talking to the driver is running at a
lower priority, that will get starved and your driver might again not
make any progress because all its queues are filled (incomming) or empty
(outgoing).

In general having a FIFO99 task around that consumes unbounded amounts
of cpu time is not something that is healthy for the system and it not
something I can recommend.

The Real-Time classes (FIFO/RR) are designed for deterministic programs
with bounded runtime. Things like spinning in userspace are very bad for
a number of reasons (which nevertheless doesn't seem to stop people from
doing just that).

The only answer that I can give you is: don't do that.

As to how Linux schedules, the scheduler is concerned only with tasks;
it will run the highest priority FIFO/RR task available, if there are no
FIFO/RR tasks present it will run SCHED_OTHER tasks, for those it will
divide the remaining time fairly based on their nice value (which is
mapped to a weight).

Interrupts -- at a hardware level -- interrupt task context and are thus
always ran whenever they happen.

That's all, there's really nothing more to it.