alternative armv6 PMAP strictly derived from i386 pmap

[Svatopluk Kraus]

Hi, arm community!

About two years ago, when I and Michal Melound have started porting FreeBSD
on newly developed board, we had to make decision which was not too easy.
The new board was based on arm11 mpcore and there was not real SMP support
for arm in FreeBSD source tree. We grew up on i386 architecture. This arch
is quite familiar for us. When we was porting FreeBSD on Coldfire, we was
using i386 tree as starting point. So, we had to choose: First, to start to
port arm11 mpcore using i386 tree or second, try to learn existing arm tree
and change it to support SMP. The decision was not so easy. However, we had
no coding experience with SMP till then and i386 arch was familiar and
already SMP. Current arm port looked influenced too many by prior arm
platforms. Futher, we already had made a successful port on Coldfire this
way.

Unfortunatelly, due to other work, we could not present our armv6 pmap
implementation until now. So now, when one pmap-v6.c is in FreeBSD tree
already, this is, at least, inteded for everybody who could be interested
in.

Now, some words about our implementation.

(1) PIPT data cache was taken as prerequisite. This way the code is not
influenced by VIVT data cache impact.

(2) The code is SMP and L2 cache compliant. Fully tested on arm11 mpcore
platform.

(3) We do not use domains. Instead, global and not global mappings are
implemented.

(4) As we had experience with port for Colfire, where page table entries
are two 32 bit words, and because it has many good implications, we
strictly are using dedicated functions for manipulation with page tables
entries. For example, this way we easily can seek for every mapping made in
so large pmap.c file.

(5) We use similar aproach for L2 page tables as in i386 arch. So each
process has mapped all its L2 page tables in dedicated address space. The
address space is shared by all processes. The cost is only one page per a
process, but this page keeps track of all L2 page tables in a process
simultaneously.

(6) Section mapping is implemented. Super section not, but it could be
easily coded too.

(7) We joined data and prefetch aborts into one routine.

(8) We are using inlines instead of macros because of type checking.

(9) The code is fully commented. Mainly arm specific parts. Plus, our
internal comments are leaved in code too.

(10) We still keep track of i386 implementation. For large parts, which are
almost same, we extended testing community a lot.

Our work is in progress, so some files are not cleaned yet. Our
implementation as it is not prepared for compilation in FreeBSD source
tree. We plan to try to make it alive for pandaboard we already have got.
Unfortunatelly, we have not too much time we can spend on it right now.
However, if somebody would like to take a try, we can support him/her with
everything needed.
Some more files are provided for better insight to our port.

https://drive.google.com/folderview?id=0Bx9OFLzdl65mNVRiQUs1UXhPZmc&usp=sharing


Svatopluk Kraus