cvs commit: src/sys/dev/bce if_bcereg.h
mj at feral.com
Wed Apr 26 06:30:06 UTC 2006
> Oh, I'm sure that you could contrive a silly device today to prove your
> point. But really, what are you going to do other than DMA completely
> randomly-associated memory pages that just happen to neighbor each
> other? It's not like the kernel or any user process can own more than
> 4GB of memory on i386+PAE.
There are a whole class of systems which are data movement engines that
have buttloads of RAM that needs to be moved in and out. The processor
doesn't touch hardly any of it. Most of the hoopla about VM and fast
processors are just *in the way* of moving this data at 200MB/s+ speeds.
The only reason such systems are in fact i386/amd64 is that you can
build them out of COTS- otherwise you might as well do a Sandpoint based
system for the cost.
Anyway- we're all pretty much in agreement that for now a bus_size_t on
i386 need not be > 32 bits. I was just observing that IMO, such as it
is, this is a pretty rapidly changing area that in a shorter time frame
than you envisioned might change.
More information about the cvs-all