In article <1992Aug20.164928.24749@ramsey.cs.laurentian.ca> ron@ramsey.cs.laurentian.ca (Ron Prediger [Velociraptor]) writes:
quoted text
>Here is another technical question about the Linux kernel.  Context
>switching in particular.

Ok, more answers...

quoted text
>        #define switch_to(n) {\
>        struct {long a,b;} __tmp;\

This just fools gcc to get 8 bytes of space on the stack, so that we
don't have to get them anywhere else..

quoted text
>        __asm__("cmpl %%ecx,_current\n\t"\
>                "je 1f\n\t

%ecx is set up with the task we want to switch to when we enter the
inline assembly statement, so this is obviosuly testing whether we are
trying to switch to the current task, in which case nothing is done. 
Besides being an obvious optimization, I don't think the 386 allows a
tss-jump to the current task (but not having my books around me, I can't
check). 

quoted text
>                "movw %%dx, %1\n\t"\

This just moves %dx (which contains the tss-selector) into the word at
__tmp+4. So ___tmp looks like this:

__tmp+0: undefined long
__tmp+4: tss-selector word
__tmp+6: undefined word

quoted text
>                "xchgl %%ecx,_current\n\t"\

This exchanges %ecx and 'current' - %ecx will contain the task we are
now executing in, and 'current' will contain the task we are just about
to jump to.

quoted text
>                "ljmp %0\n\t"\

This does the actual task switch: we do an indirect long-jump to __tmp.
This is also the reason we updated just the word at __tmp+4: a long jump
to a new tss-selector will ignore all other values. It will
automatically jump to the place indicated by the tss segment.

quoted text
>                "cmpl %%ecx,_last_task_used_math\n\t"\

Yes, this looks illogical: we just did an unconditional jump, and this
should never be reached.  Right? Wrong.  The ljmp will return once some
other task does it's own ljmp to this task.  Now %ecx contains the
current task, so we check if this is the task that used the math
coprocessor last... 

quoted text
>                "jne 1f\n\t"\
>                "clts\n"\

...in which case we can clear the TS flag, giving better
math-performance. Due to these three lines, linux never saves math state
unnecessarily.

quoted text
>                "1:"\
>                "::"m" (*&__tmp.a),"m" (*&__tmp.b), \

This just sets up %0 - pointing to '__tmp', and %1, pointing to
'__tmp+4'.

quoted text
>                "d" (_TSS(n)), "c" ((long) task[n]) \

%dx contains the tss-descriptor for the task we want to switch to, and
%ecx contains the task-pointer of the task we want to switch to.

quoted text
>                :"cx");\

And this just tells gcc that %ecx is changed by the inline-asm.

That's it: 8 lines of assembly does the linux context switch with
math-state optimizations.  Note that newer sources should have a cli-sti
pair around the "xchg %%ecx,_current ; ljmp %0", so that 'current' is
always up-to-date, even when running under interrupts.

                Linus