Re: replace "memset(...,0,PAGE_SIZE)" calls with "clear_page()"?

[Paul Mundt <lethal@linux-sh.org>]

On Sat, Dec 30, 2006 at 06:04:14PM -0500, Robert P. J. Day wrote:
> fair enough.  *technically*, not every call of the form
> "memset(ptr,0,PAGE_SIZE)" necessarily represents an address that's on
> a page boundary.  but, *realistically*, i'm guessing most of them do.
> just grabbing a random example from some grep output:
> 
> arch/sh/mm/init.c:
>   ...
>   /* clear the zero-page */
>   memset(empty_zero_page, 0, PAGE_SIZE);
>   ...
> 
The problem with random grepping is that it doesn't give you any context.
clear_page() isn't available in this case since we have a couple of
different ways of implementing it, and the optimal approach is selected
later on. There are also additional assumptions regarding alignment that
don't allow clear_page() to be used directly as replacement for the
memset() callsites (as has already been pointed out for some of the other
architectures). While the empty_zero_page in this case sits on a full page
boundary, others do not.

You might find some places in drivers that do this where you might be
able to optimize things slightly with a clear_page() (or copy_page() in
the memcpy() case), but it's going to need a lot of manual auditing
rather than a find and replace. Any sort of wins you get out of this
would be marginal at best, anyways.

The more interesting case would be page clustering/bulk page clearing
with offload engines, and there's certainly room to build on the SGI
patches for this.