Re: raid5d, page_launder and scheduling latency

[Andrew Morton <andrewm@uow.edu.au>]

Alan Cox wrote:
> 
> > Happily, we've just fixed the four most gross sources of poor
> > interactivity in the kernel, so let's knock over some of the others as
> > well - a few /proc functions.  That mainly leaves zap_page_range() and
> > exit() with a lot of open files.
> 
> Nowhere near it. We have to fix copy_*_user and strlen_user (there are reasons
> 2.2 uses strnlen_user). Map the same page into 2Gig of address space filled with
> non zero bytes. Map a zero terminator on the end of it. Pass pointers to this
> for all your args and do an exec().

By "interactivity" I mean "things which make it feel jerky".
The commonly occurring things, not the oddball corner cases.

- huge reads from /dev/mem
- exit with 1,000 files open
- exit with half a million pages to be zapped

And "fixing" copy_*_user is outright dumb.  Just fix the four
or five places where it matters.

-