Re: Linux 2.6.x on 8xx status

[Guillaume Autran <gautran@mrv.com>]

[-- Attachment #1: Type: text/plain, Size: 3299 bytes --]



Marcelo Tosatti wrote:

>On Tue, Mar 22, 2005 at 03:57:08PM -0500, Dan Malek wrote:
>  
>
>>On Mar 22, 2005, at 8:04 AM, Marcelo Tosatti wrote:
>>
>>    
>>
>>>I'm quite puzzled. Why v2.6 calls the "tlbie" instruction 100-or-so
>>>less times than v2.4 ?
>>>      
>>>
>
>That was rather a _factor_ of "100-or-so" less.
>
>  
>
>>Oh my ...  I'm more worried about the high number of TLB misses
>>in 2.6 compared to 2.4.  That's really bad.  
>>    
>>
>
>Newbie question: What prevents the initial kernel map (tuple of 8Mbyte I/D-TLB entries) 
>and the IMMR 8Mbyte D-TLB entry from getting unmapped by translation pressure,
>in case CONFIG_PIN_TLB is disabled ? 
>
>  
>
>>How did you instrument the tlbie measurement? 
>>    
>>
>
>By a counter at the end of _tlbie function, similar to other counters which 
>you suggested.
>
>  
>
>>It could be that 2.4 used lots more 'tlbia' which were replaced by tlbie in 2.6.
>>    
>>
>
>Dont think thats the case given that v2.4 calls tlbia through flush_tlb_mm() at exit_mmap() 
>only. And at vmalloc_free which shouldnt be called at all.
>
>I just noticed this conditional at switch_mm() (v2.6), which _can_ partly 
>explain the reduced tlbie's  (its just a guess for now, though):
>
>static inline void switch_mm(struct mm_struct *prev, struct mm_struct *next,
>                             struct task_struct *tsk)
>{
>#ifdef CONFIG_ALTIVEC
>        asm volatile (
> BEGIN_FTR_SECTION
>        "dssall;\n"
>#ifndef CONFIG_POWER4
>         "sync;\n" /* G4 needs a sync here, G5 apparently not */
>#endif
> END_FTR_SECTION_IFSET(CPU_FTR_ALTIVEC)
>         : : );
>#endif /* CONFIG_ALTIVEC */
>                                                                                            
>        tsk->thread.pgdir = next->pgd;
>                                                                                            
>        /* No need to flush userspace segments if the mm doesnt change */
>	if (prev == next) 				<--------------
>		return;   				<--------------
>                                                                                            
>        /* Setup new userspace context */
>        get_mmu_context(next);
>        set_context(next->context, next->pgd);
>}
>
>I'm about to disable it and retry.
>
>Spent part of the day reading the MMU section of 860 manual, I think I have kind 
>of a clue how things are supposed to work at the lowlevel now. 
>
>I'll continue tracking it down - any help is appreciated.
>
>PS: I can't reproduce the invalid TLB crash anymore. i.e. even by removing
>the _tlbie() at update_mmu_cache() everything is working as expected.
>
>How can I reproduce it again? Guillaume, what kernel version are you using?
>
>  
>
It is very timing dependant. Running 2.6.11, ldconfig crashes in 
__flush_dcache_icache(...) just after boot time (first time called). 
Unfortunatly, it only happens every now and then. And of course, never 
when my BDI2000 is plugged in. :(

I also noticed that with kernel preemption disable, oops are less 
frequent. Probably does not mean anything anyway...

Guillaume.


-- 
=======================================
Guillaume Autran
Senior Software Engineer
MRV Communications, Inc.
Tel: (978) 952-4932 office
E-mail: gautran@mrv.com
======================================= 


[-- Attachment #2: Type: text/html, Size: 4130 bytes --]