Re: emap

[yamt%mwd.biglobe.ne.jp@localhost (YAMAMOTO Takashi)]

hi,

> yamt%mwd.biglobe.ne.jp@localhost (YAMAMOTO Takashi) wrote:
>> hi,
>> 
>> what's the status of emap and pipe?
>> 
> 
> If I remember correctly, I could not get real performance improvement
> of pipe with UVM emap.  Higher chances are that it would improve socket
> performance.  Also, pipe has direct I/O disabled (PIPE_NODIRECT) due
> to some rare lock-ups when switching between direct/buffered.  AFAIK,
> we have this bug since netbsd-3 times or even earlier.

can you provide a recipe to reproduce the lock-up?

> 
> However, I tend to question whether UVM emap was a right approach to
> the problem.  For amd64, we should definitely have a direct mapping
> of kernel space (I was planning to revive the patch written by chs@,
> but ran out of time) and encourage our users to use amd64 instead
> of i386.  Most other architectures have a segment for direct mapping.
> I wonder if we could help i386 with large pages for kernel space..

the real problem for i386 is kva shortage.
direct-mapping only limited part of physical memory would work, tho.
(call the rest of memory highmem!)

YAMAMOTO Takshi

> 
> -- 
> Mindaugas