| 2 |
- |
1 |
This is a preliminary Plan 9 port to the Compulab Trimslice,
|
|
|
2 |
containing a Tegra 2 SoC: a dual-core, (truly) dual-issue 1GHz
|
|
|
3 |
Cortex-A9 v7a-architecture ARM system, *and* it comes in a case. VFP
|
|
|
4 |
3 floating-point hardware is present, but 5l doesn't yet generate
|
|
|
5 |
those instructions. This is the first multiprocessor ARM port we've
|
|
|
6 |
done, and much of the code should be reusable in future ports. There
|
|
|
7 |
are still things to be done but it can run both processors and is
|
|
|
8 |
believed to have adequate kernel support for VFP 3 floating-point.
|
|
|
9 |
|
|
|
10 |
|
|
|
11 |
What's implemented.
|
|
|
12 |
|
|
|
13 |
Two cpus running concurrently with level 1 and 2 caches enabled.
|
|
|
14 |
|
|
|
15 |
Realtek 8168 Ethernet. A slightly dimmer 8169. Has to be jabbed with
|
|
|
16 |
an electric cattle prod by software about once per day when it wedges.
|
|
|
17 |
|
|
|
18 |
Profiling. Charles Forsyth fixed various bugs to make user-mode
|
|
|
19 |
profiling on ARMs work for the first time ever.
|
|
|
20 |
|
|
|
21 |
|
|
|
22 |
What's not (yet) implemented.
|
|
|
23 |
|
|
|
24 |
USB. It probably just needs initialisation.
|
|
|
25 |
|
|
|
26 |
NOR flash.
|
|
|
27 |
|
|
|
28 |
Video.
|
|
|
29 |
|
|
|
30 |
VFP3 floating point. The go 5l generates VFP 3 floating-point
|
|
|
31 |
instructions (among other changes). Attempts to transplant just that
|
|
|
32 |
code into our 5l failed to generate correct code. Eventually someone
|
|
|
33 |
will get this to work, and then we'll be able to use the hardware
|
|
|
34 |
floating-point. [Eventually someone did, thanks.] Even with only
|
|
|
35 |
software emulation of floating-point, astro runs in under 3 seconds.
|
|
|
36 |
|
|
|
37 |
In-line 64-bit arithmetic in 5[cl].
|
|
|
38 |
|
|
|
39 |
And the really horrid peripherals: NAND flash and MMC.
|
|
|
40 |
|
|
|
41 |
|
|
|
42 |
Known problems.
|
|
|
43 |
|
|
|
44 |
kprof. kprof profiling doesn't work correctly, charging all CPU time
|
|
|
45 |
to _start.
|
|
|
46 |
|
|
|
47 |
Reboot. After an fshalt -r reboot (or two) with cpu1 enabled,
|
|
|
48 |
accesses to pci registers (notably 0x80015000) in the newly-loaded
|
|
|
49 |
kernel often hang. One of three watchdogs' reset should jolt the
|
|
|
50 |
system back to life and force a reboot through u-boot when this
|
|
|
51 |
happens. Sometimes the ethernet goes dead instead ("waiting for
|
|
|
52 |
dhcp..." forever); this could be a different symptom of pci illness.
|
|
|
53 |
|
|
|
54 |
Also following a reboot, cpu1's local (not tegra SoC shared) timers
|
|
|
55 |
don't interrupt. Since the local watchdogs don't seem to actually
|
|
|
56 |
interrupt nor generate resets when used in anger (as opposed to
|
|
|
57 |
boot-time check-out), their loss is merely a mystery. The local timer
|
|
|
58 |
not interrupting is more worrying.
|