Re: Rising io_load results Re: 2.5.63-mm1

[Andrew Morton]

Dave McCracken <[e-mail]> wrote:
>
> 
> --On Thursday, February 27, 2003 13:44:03 -0800 Andrew Morton
> <[e-mail]> wrote:
> 
> >> ...
> >> Mapped:       4294923652 kB
> > 
> > Well that's gotta hurt.  This metric is used in making writeback
> > decisions.  Probably the objrmap patch.
> 
> Oops.  You're right.  Here's a patch to fix it.
> 

Thanks.

I'm just looking at page_mapped().  It is now implicitly assuming that the
architecture's representation of a zero-count atomic_t is all-bits-zero.

This is not true on sparc32 if some other CPU is in the middle of an
atomic_foo() against that counter.  Maybe the assumption is false on other
architectures too.

So page_mapped() really should be performing an atomic_read() if that is
appropriate to the particular page.  I guess this involves testing
page->mapping.  Which is stable only when the page is locked or
mapping->page_lock is held.

It appears that all page_mapped() callers are inside lock_page() at present,
so a quick audit and addition of a comment would be appropriate there please.