Re: 2.2.18Pre Lan Performance Rocks!

["Jeff V. Merkey" <jmerkey@timpanogas.org>]

One more optimization it has.  NetWare never "calls" functions in the
kernel.  There's a template of register assignments in between kernel
modules that's very strict (esi contains a WTD head, edi has the target
thread, etc.) and all function calls are jumps in a linear space. 
layout of all functions are 16 bytes aligned for speed, and all
arguments in kernel are passed via registers.  It's a level of
optimization no C compiler does -- all of it was done by hand, and most
function s in fast paths are layed out in 512 byte chunks to increases
speed.  Stack memory activity in the NetWare kernel is almost
non-existent in almost all of the "fast paths"

Jeff

"Jeff V. Merkey" wrote:
> 
> A "context" is usually assued to be a "stack".  The simplest of all
> context switches
> is:
> 
>    mov    x, esp
>    mov    esp, y
> 
> A context switch can be as short as two instructions, or as big as a TSS
> with CR3 hardware switching,
> 
> i.e.
> 
>    ltr    ax
>    jmp    task_gate
> 
> (500 clocks later)
> 
>    ts->eip gets exec'd
> 
> you can also have a context switch that does an int X where X is a gate
> or TSS.
> 
> you can also have a context switch (like linux) that does
> 
>     mov    x, esp
>     mov    esp, y
>     mov    z, CR3
>     mov    CR3, w
> 
> etc.
> 
> In NetWare, a context switch is an in-line assembly language macro that
> is 2 instructions long for a stack switch and 4 instructions for a CR3
> reload -- this is a lot shorter than Linux.
> Only EBX, EBP, ESI, and EDI are saved and this is never done in the
> kernel, but is a natural
> affect of the Watcom C compiler.  There's also strict rules about
> register assignments that re enforced between assembler modules in
> NetWare to reduce the overhead of a context switch.  The code path is
> very complex in NetWare, and priorities and all this stuff exists, but
> these code paths are segragated so these types of checks only happen
> once in a while and check a pre-calc'd "scoreboard" that is read only
> across processors and updated and recal'd by a timer every 18 ticks.
> 
> Jeff
> 
> 
> 
> Andi Kleen wrote:
> >
> > On Tue, Oct 31, 2000 at 02:52:11PM -0700, Jeff V. Merkey wrote:
> > > The numbers don't lie.  You know where the code is.  You notice that
> > > there is a version of
> > > the kernel hand coded in assembly language.  You'l also noticed that
> > > it's SMP and takes ZERO LOCKS during context switching, in fact, most of
> > > the design is completely lockless.
> >
> > I suspect most of the confusion in this thread comes because you seem to
> > use a different definition of context switch than Ingo and others. Could
> > you explain what you exactly mean with a context switch ?
> >
> > -Andi
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
Please read the FAQ at http://www.tux.org/lkml/