tree a4ebd66301bb2dd95ba07214d62b6684555e92e7
parent 76e8476bfe8097698d788f01f52ccbe4725eea70
author Barret Rhoden <brho@cs.berkeley.edu> 1571942260 -0400
committer Barret Rhoden <brho@cs.berkeley.edu> 1583266408 -0500

pci: map all memory BARs into the kernel

All device memory BARs get mapped, and drivers can just ask for a
specific, preexisting mapping.  They will stay mapped for the life of
the system.

This makes it easier to reset and initialize a device after the kernel
boots.  We don't need to map and unmap the KVA at runtime, which would
have changed the kernel mapping, requiring changes to all existing
address spaces / pgdirs.

It's also just easier for the devices - just ask for the BAR number, and
we'll provide the address.

Note the 82563 wasn't using 'nocache' for its register space.  Probably
a bad idea.

Also, I'm still up-in-the-air with respects to uinptr_t vs void*.  Plan
9 and Linux drivers tend to use void* for their iomem, even though you
aren't supposed to dereference it with normal operations.  In Linux, you
use writel(void*) (write32() here).  Our older code uses e.g.
write_mmreg32(uintptr_t).

Signed-off-by: Barret Rhoden <brho@cs.berkeley.edu>