Re: a joint letter on low latency and Linux

[Richard Gooch]

David Schleef writes:
> On Thu, Jun 29, 2000 at 10:31:07PM -0600, Richard Gooch wrote:
> > 
> > I don't see why the MM stuff is an issue: just tap in at every place
> > where we do a user<->kernel transition. There's not that many, after
> > all.
> > 
> > So if a hard-RT user-space process suddenly takes a page fault, the
> > CPU switches to kernel mode (still running under the RTLinux
> > scheduler), and the first thing we do is call rtl_drop_rt(). Same for
> > all other kernel entry points.
> 
> This is, in fact, exactly what RTAI/LXRT is planning to do.
> (Actually, "does", if you count my current source tree.)  Except
> that you get to chose whether to drop priority and execute the trap,
> kill the process, or have the system call fail.  (If it is a system
> call that trapped -- eventually, we'll implement a few important
> system calls, like sys_gettimeofday().)

I don't like that at all. Adding a pile of extra "fast" syscalls is
just too hackish. Just say: "if you go into the kernel, you lose RT".
It's the thin edge of the wedge. Soon you'll be getting people to say
"we need read(2) to keep RT priority".

Now, if instead you can move between the RT run queue and Linux run
queue when you call schedule(), rather than at kernel entry/exit
points, then you'd get fast, simple syscalls for free. But that may be
harder than switching at kernel entry/exit points.

				Regards,

					Richard....
Permanent: [e-mail]
Current:   [e-mail]