Rev 2 | Blame | Compare with Previous | Last modification | View Log | RSS feed
global scale sheevaplug & guruplug
marvell 88f6281 (feroceon kirkwood) SoC
arm926ej-s rev 1 [56251311] (armv5tejl) 1.2GHz cpu
l1 I & D VIVT caches 16K each: 4-way, 128 sets, 32-byte lines
l1 D is write-through, l1 I is write-back
unified l2 PIPT cache 256K: 4-way, 2048 sets, 32-byte lines
potentially 512K: 8-way
apparently the mmu walks the page tables in dram and won't look in the
l2 cache. there is no hardware cache coherence, thus the l1 caches
need to be flushed or invalidated when mmu mappings change, but l2
only needs to be flushed or invalidated around dma operations and page
table changes, and only the affected dma buffers and descriptors or
page table entries need to be flushed or invalidated in l2.
we arrange that device registers are uncached.
be aware that cache operations act on cache lines (of CACHELINESZ
bytes) as atomic units, so if you invalidate one word of a cache line,
you invalidate the entire cache line, whether it's been written back
(is clean) or not (is dirty). mixed data structures with parts
maintained by hardware and other parts by software are especially
tricky. we try to pad the initial hardware parts so that the software
parts start in a new cache line.
there are no video controllers so far, so this port is a cpu
kernel only.
512MB of dram at physical address 0
512MB of nand flash
16550 uart for console
see http://www.marvell.com/files/products/embedded_processors/kirkwood/\
FS_88F6180_9x_6281_OpenSource.pdf, stored locally as
/public/doc/marvell/88f61xx.kirkwood.pdf
If you plan to use flash, it would be wise to avoid touching the first
megabyte, which contains u-boot, right up to 0x100000. There's a
linux kernel from there to 0x400000, if you care. You'll also likely
want to use paqfs rather than fossil or kfs for file systems in flash
since there is no wear-levelling.
The code is fairly heavy-handed with the use of barrier instructions
(BARRIERS in assembler, coherence in C), partly in reaction to bad
experience doing Power PC ports, but also just as precautions against
modern processors, which may feel free to execute instructions out of
order or some time later, store to memory out of order or some time
later, otherwise break the model of traditional sequential processors,
or any combination of the above.
this plan 9 port is based on the port of native inferno to the
sheevaplug by Salva Peiró (saoret.one@gmail.com) and Mechiel Lukkien
(mechiel@ueber.net).
___
# type this once at u-boot, replacing 00504301c49e with your plug's
# mac address; thereafter the plug will pxe boot:
setenv bootdelay 2
setenv bootcmd 'bootp; bootp; tftp 0x1000 /cfg/pxe/00504301c49e; bootp; tftp 0x800000; go 0x800000'
saveenv
# see /cfg/pxe/example-kw
physical mem map
hex addr size what
----
0 512MB sdram
c8010000 2K cesa sram
virtual mem map (from cpu address map & mmu mappings)
hex addr size what
----
0 512MB user process address space
60000000 512MB kzero, mapped to 0
90000000 256MB pcie mem # identity mapped by u-boot
d0000000 1MB internal regs default address at reset
d8000000 128MB nand flash # 512MB addressed through this (guru)
f0000000 16MB pcie i/o # mapped to 0xc0000000 by u-boot
f1000000 1MB internal regs (on-chip devices) as mapped by u-boot
f1000000 64K dram regs
f1010000 64K uart, flashes, rtc, gpio, etc.
f1030000 64K crypto accelerator (cesa)
f1040000 64K pcie regs
f1050000 64K usb otg regs (ehci-like)
f1070000 64K gbe regs
f1080000 64K non-ahci sata regs
f1090000 64K sdio regs
f8000000 16MB spi flash # mapped to 0 by u-boot
f9000000 8MB nand flash # identity mapped by u-boot (sheeva/openrd)
fb000000 64KB crypto engine # identity mapped by u-boot
ff000000 16MB boot rom # identity mapped by u-boot
... as per physical map