Re: [PATCH] Remove 8 bytes from struct page on 64bit archs

[William Lee Irwin III <wli@holomorphy.com>]

On Mon, Jan 07, 2002 at 12:33:26AM +1100, Anton Blanchard wrote:
> For archs that need ->zone, merging it with ->flags sounds like a
> great idea. Id like to cram something into ->flags on 64 bit archs
> since its only a long due to bitop constraints. I had thought of
> stuffing ->count in the high word but now Im just getting silly
> since all non atomic accesses to ->flags would then have to be word
> ones.

At some point in the past, I wrote:
>> My i386 version, which makes ->virtual conditional on CONFIG_HIGHMEM as
>> well, is at:

On Mon, Jan 07, 2002 at 12:33:26AM +1100, Anton Blanchard wrote:
> Id like to do redo some profiling, on ppc64 we had page_address() doing
> pointer arithmetic (instead of page->virtual) and the compiler created
> an awful sequence of instructions in the acenic interrupt handler.
> A zero copy TCP benchmark made it rather obvious.

I'm not entirely surprised at this, CONFIG_HIGHMEM is probably just
not quite a strict enough criterion for eliminating ->virtual as it's
a time/space tradeoff that just happens to be bad on ppc64. Maybe we
should figure out some other criterion in addition to it.


Cheers,
Bill