On Sun, Dec 14, 2003 at 05:16:37PM +0000, Jamie Lokier wrote:
> Peter Horton wrote:
> > I've seen code written for X86 use MAP_FIXED to create self wrapping
> > ring buffers. Surely it's better to fail the mmap() on other archs
> > rather than for the code to fail in unexpected ways?
> Such code should test the buffers or just not create ring buffers on
> architectures it doesn't know about. (You can usually simulate them
> by copying data). On some architectures there is _no_ alignment which
> works, and even on x86 aligning aliases to 32k results in faster
> memory accesses on some chips (AMD ones).
> Also, sometimes a self wrapping ring buffer can work even when the
> separation isn't coherent, provided the code using it forces cache
> line flushes at the appropriate points.
Still I don't see why we shouldn't simply return EINVAL if a user is
trying to something obviously stupid - assuming full coherency in
application is a somewhat common thing and there's better things to waste
time on. And yes while we could support coherency for arbitrary mappings
I agree it's a bad idea - but there's a huge difference between just
checking arguments and adding the large extra complexity of supporting
arbitrary combinations of addresses for mappings.