ioremap() vs. ioremap_nocache()

[David Mosberger <davidm@hpl.hp.com>]

Jes Sorensen was kind enough to point out a longstanding
mis-conception that I had about ioremap(): I thought that ioremap()
was being deprecated in favor of ioremap_nocache() because the former
does not clearly define what kind of memory attribute will be used to
access the mapped memory (cached, write-through cached,
write-coalescing, etc).  But I seem to have been wrong about that.
Now, as far as I know, on x86, ioremap() will give write-through
cached mappings (in the absence of mtrr games).  If this is true, how
can this work?  There are many drivers out there that use ioremap() on
memory mapped I/O regions that do NOT have the "Prefetchable" bit set
in the PCI BAR.  For example, the eepro100 driver does this and it has
a routine called wait_for_cmd_done(), which spins on an ioremapped
read.  On an x86, what prevents these reads from being cached?

Thanks,

	--david