On Wed, Jan 07, 2015 at 07:48:48PM +0000, Arnd Bergmann wrote:
On Wednesday 07 January 2015 12:44:56 Jon Masters wrote:
I'm expecting to need new drivers for SoC IP blocks that are net new, but generational differences between iterations of the same SoC should be abstracted behind the firmware (and we are already seeing this with at least one platform). Platform wise, it's nice to already see e.g. mmconfig working to handle the specific ways a platform wires PCI.
Yes, the parts that are mandated by SBSA, like the way that PCI needs to be done are generally good. Unfortunately a lot of the hardware that I've seen has a rather lax interpretation of the spec, so just because something is mandated doesn't mean it's done that way ;-)
In other cases that's actually a good thing. One such example is the "Principles of ARM Memory Maps" document that tells hardware implementers to do a rather complex mapping "To support 36-bit x86 PAE compatible operating systems, such as Linux." but makes life much harder in the process than any of the random mappings we have seen in the wild.
Unfortunately, with any significant amount of RAM (say 16GB), this document becomes pretty useless. It basically forces you to have a very sparse physical address map from 0 to over 40-bit. I wouldn't apply the ARM memory maps doc to server systems.
- There is a general mindset about deprecating unwanted features early. ARMv8 aarch32 bit mode removes support for older instructions or makes them optional. Even the virtualization mode doesn't allow to trap on architecture version specific differences, so you can't completely emulate an older architecture level. This is nice for implementers but not so much for users that rely on old (mis-)features.
This mindset is (slowly) changing. There are, of course, instructions like SWP that just can't always be implemented at the SoC level (not necessarily CPU level; requiring bus locks) but others like CP15 barriers, I don't see why they should go away, it's just a decoder problem.
It's also not just the CPU core, other components also get easily replaced, like a GICv3 that is not a strict superset of GICv2.
That's not a problem for Linux, we can describe them in DT or ACPI and have drivers. GICv3 has an optional GICv2 compatible mode, though vendors may decide not to implement it.