Re: Feedback on preemptible kernel patch

[Robert Love <rml@tech9.net>]

On Fri, 2001-09-14 at 03:30, george anzinger wrote:
> Right, the same problem as using floating point in the kernel (mmx uses
> the FP regs and they are not saved).

Right, and I suspect we will find more problems of this type as we go
on.  In fact, the more general case "things that are SMP-safe but not
preempt safe" will be issues, too.  The highmem bug was one of these -
code that was SMP-safe but did not have lock points because it was
per-CPU code.  Preemption ruins all that.

> The question is: Just how long do these routines take?  If it is very long
> it may be best to just say no. One way would be to always pretend that
> the"in_interrupt" flag is set.  I think possibly some routines are
> short and the switch off/ switch on pair is right, but for the long ones,
> well the preemption patch is supposed to make the kernel more preemptable,
> not less.  Any one have execution times for these functions?

Well, its the routines in arch/i386/lib/mmx.c -- and just the ones that
call kernel_begin/end_fpu.  My patch pushes a ctx_sw_off/on pair into
those functions.  Anyhow, if you look, they aren't too long.

However, I agree that we may be destroying our purpose here.  A user of
the patch actually put together a patch that will disable the CONFIG to
use the fast MMX memcpy stuff if preemption was enabled.  He benchmarked
against the two and I can send you those results when I sort through
them.

-- 
Robert M. Love
rml at ufl.edu
rml at tech9.net