Re: [PATCH]: Option to run cache reap in thread mode

[Manfred Spraul <manfred@colorfullife.com>]

Andrew Morton wrote:

>Dimitri Sivanich <sivanich@sgi.com> wrote:
>  
>
>>At the time of the holdoff (the point where we've spent a total of 30 usec in
>>the timer_interrupt), we've looped through more than 100 of the 131 caches,
>>usually closer to 120.
>>    
>>
>
>ahh, ooh, ow, of course.
>
>Manfred, we need a separate list of "slabs which might need reaping".
>  
>
A cache that might need reaping is a cache that has seen at least one 
kmem_cache_free(). The list would trade less time in the timer context 
at the expense of slower kmem_cache_free calls. I'm fairly certain that 
this would be end up as a big net loss.

>That'll help the average case.  To help the worst case we should change
>cache_reap() to only reap (say) ten caches from the head of the new list
>and to then return.
>
I'll write something:
- allow to disable the DMA kmalloc caches for archs that do not need them.
- increase the timer frequency and scan only a few caches in each timer.
- perhaps a quicker test for cache_reap to notice that nothing needs to 
be done. Right now four tests are done (!flags & _NO_REAP, 
ac->touched==0, ac->avail != 0, global timer not yet expired). It's 
possible to skip some tests. e.g. move the _NO_REAP caches on a separate 
list, replace the time_after(.next_reap,jiffies) with a separate timer.

--
    Manfred