> -----Original Message-----
> From: Yao Zhao [mailto:yzhao@qnx.com] 
> Sent: September 18, 2007 11:36 AM
> To: ostech-core_os
> Subject: Re: kernel reading notes
> 
> Because of code here then you know only one cpu can enter 
> kernel state or get INKERNEL_NOW for system call. This makes 
> KERN_CPU sense. It means there is a big kernel lock and 
> kernel is not quite smp capable although interrupt, 
> exceptions may not need this. 
> 
> Current kernel code relies this too much, I guess it is not 
> easy to make it more smp capable but it is possible. Need a project?

[ Code comments snipped ]

Yao,

I'm not sure what your intention is with these posts?  

* Are you trying to fix a particular problem; if so then can you
create a post with a more accurate synopsis/title.

* Are you simply relaying your own personal notes from having read 
through the sources; if so, then perhaps contributing the commentary
to an appropriate wiki page would be better

* Are you doing something else?

This forum thread seems to be kind of random, and with your statement
above, I'm not sure where you are headed.

Thanks,
 Thomas
  

Yao Zhao wrote:
>
> Because of code here then you know only one cpu can enter kernel state 
> or get INKERNEL_NOW for system call. This makes KERN_CPU sense. It 
> means there is a big kernel lock and kernel is not quite smp capable 
> although interrupt, exceptions may not need this.
>
This is by design.  The amount of time spent in a microkernel kernel 
call is typically so short as to not be an issue.  Interrupts
acquire the kernel lock after the EOI, in order to process any event 
deliveries/scheduling decisions that were caused during interrupt
processing.

> Current kernel code relies this too much, I guess it is not easy to 
> make it more smp capable but it is possible. Need a project?
>
> it will be released in __ker_exit.
>  
> 1:
>     cmpl    $0,need_to_run      /*yzhao need_to_run thread == 0? */
>         jz              3f                      /*yzhao 
> need_to_run==0, that means no thread need to run and preempt me, just 
> go ahead, jump to 3 */
>
>         cmpl    %ebp,need_to_run_cpu /*yzhao need_to_run!=0, compare 
> current cpu == need_to_run_cpu? */
>         je              3f                                      
> /*yzhao need_to_run_cpu is current cpu, jump to 3*/
>         pause                                           /*yzhao pause, 
> it was suggested by Intel for spin-loop wait, in Intel Xeon, P4 and 
> dual cores
>
>                                                                 
>                 we have to */
>         jmp             1b
>
>         // See if anybody else is in the kernel
> 3:
>         mov             inkernel,%eax
>         test    $INKERNEL_NOW,%eax /*yzhao and immed32, eax, set 
> SF,ZF,PF according to result*/
>         jnz             1b                              /*yzhao if 
> (eax & INKERNAL_NOW) jmp 1b*/
>
>
>         cli                                             /*yzhao 
> disable irq*/
> end_aquire_kernel_attempt:
>
>         mov             %eax,%edi               /*yzhao mov inkernel 
> to edi */
>         andl    $0x00ffffff,%edi                /*yzhao get edi lower 
> 24 bits*/
>         mov     %ebp,%ecx              
>         shl             $24,%ecx                /*yzhao get cpunum, 
> ebp lower 8 bits represent cpunum */
>         orl             %edi,%ecx               / Set cpunum
>
>         orl             $INKERNEL_NOW,%ecx      /*yzhao 
> cpunum|inkernel|INKERNEL_NOW*/
>         lock; cmpxchg   %ecx,inkernel   /*yzhao lock bus and xchg ecx 
> and inkernel*/
>         jnz             aquire_kernel_attempt   /*yzhao if ecx is not 
> same with inkernel, again */
>         // We are the kernel
>         mov             %esi,%eax       // restore kernel call number
>
> _______________________________________________
> OSTech
> http://community.qnx.com/sf/go/post1353
>

-- 
cburgess@qnx.com