Re: [RFC] page fault retry with NOPAGE_RETRY

[Andrew Morton <akpm@osdl.org>]

On Wed, 20 Sep 2006 09:50:35 +1000
Benjamin Herrenschmidt <benh@kernel.crashing.org> wrote:

> On Tue, 2006-09-19 at 16:35 -0700, Mike Waychison wrote:
> > Patch attached.
> > 
> > As Andrew points out, the logic is a bit hacky and using a flag in 
> > current->flags to determine whether we have done the retry or not already.
> > 
> > I too think the right approach to being able to handle these kinds of 
> > retries in a more general fashion is to introduce a struct 
> > pagefault_args along the page faulting path.  Within it, we could 
> > introduce a reason for the retry so the higher levels would be able to 
> > better understand what to do.
> 
>  .../...
> 
> I need to re-read your mail and Andrew as at this point, I don't quite
> see why we need that args and/or that current->flags bit instead of
> always returning all the way to userland and let the faulting
> instruction happen again (which means you don't block in the kernel, can
> take signals etc...

That would amount to a busy wait, waiting for the disk IO to complete.

So we need to go to sleep somewhere (in D state, because we _are_ waiting
for disk IO).  Returning all the way to userspace and immediately retaking
the fault is unneeded extra work.

> thus do you actually need to prevent multiple
> retries ?)

I expect there are livelock scenarios.  For example, process A could spin
on posix_fadvise(some libc text page, POSIX_FADV_DONTNEED), perhaps causing
other applications to get permanently stuck in the kernel.