88W8388

This is Marvell's wireless controller chip. Outside of this page, it is undocumented.

Gross memory map

00000000 0000ffff code, part 1
04000000 04001fff stack
80000000 8000ffff IO memory
90000000 9000ffff IO memory
c0000000 c000ffff code, part 2
c0010000 c0027fff heap and bss
ffff0000 ffffffff boot code

CPU

This is an little-endian ARM chip. "xscale" appears to be a good match. Mostly it is used to run thumb code, but some regular ARM code is used as well. ARM code can be spotted in hex dumps because most instructions will start (4th byte in little-endian) with the hex digit "e".

Note that objdump is completely unsuitable for disassembling this code. You need an interactive disassembler and/or one that traces code flow. Something like objdump will be unaware of the transitions between plain ARM code, thumb code, and data. It is normal for functions to be followed by constants; a good disassembler will match these up with the code instead of stupidly disassembling them.

Main firmware format

Firmware is a stream of packets like this.

le32      0x00000001 if data follows, else 0x00000004
le32      destination address
le32      data length including final CRC, normally 0x200 bytes
be32      header CRC (1st 12 bytes and 4 zero bytes)
char[508] may be smaller for the last block
be32      data CRC (the 508 bytes and 4 zero bytes)

Data length goes short at the end of a segment. After the very last block of the whole file, there is one special header. It has code 0x00000004, address zero, and length zero.

Use bin2elf.c and elf2bin.c to convert firmware files to/from ELF..

CRC

polynomial 0x04c11db7 (common CRC32)
initial remainder 0
no reflections
no inversion at the end
stored in big-endian format!
on creation, include 4 trailing zero bytes
nice property: the CRC of the data with following CRC will be zero

Boot firmware format

It's just 10240 bytes of byte-swapped code. This code, known as Boot2, is flashed into the Marvell chipset. It implements the USB interface required for loading the main firmware.

IO memory in the 9000xxxx area

90000000 w8 
90000004 w8 

At 90000200 and 90000204, we find a pair of 16-bit registers full of 2-bit fields. At 90000208 we find a 16-bit register full of 1-bit fields that match up with the 2-bit fields. This is two arrays, with the first being split because electrical engineers love to torture the software guys. Each array element appears to control a 256 MiB memory region, with a 1 bit in the 90000208 register being used to enable the cache. When such a bit is set, the corresponding 2-bit field in the 90000200 or 90000204 register should be set to 1. (values 0,2,3 are likely to be various cache policies like write-through)

90000200 r32 w16 w32 
90000204 r32 w16 w32 
90000208 r32 w16 w32 
9000020c r32 w16 w32 
90000214 r32 w32   
90000218 ?         
90000220 r32 w32 
90000228 r32 w16 
9000022c r32 w16 
90000230 r32 w16 
90000234 r32 w16 
90000238 r32 w16 
9000023c r32 w16 
90000240 r32 w16 
90000244 r32 w16 
9000024c r32 w16 
90000254 r32 w32 

90000800 r16 
90000804 r32 w32 r8 w8 
90000806 r8 
90000808 r32 r16 w16 
9000080c r8 
9000081c r8 
90000824 r32 w32 r8 w8 
90000828 w32 w8 

9000b010 w32 
9000b014 w32 
9000b018 w32 
9000b024 w32 
9000b028 w32 
9000b02c w32 
9000b030 w32 
9000b034 w32 
9000b038 w32 
9000b03c w32 
9000b040 w32 

9000b100 r32 w32 
9000b118 w32 
9000b11c w32 
9000b120 w32 
9000b124 w32 
9000b128 r32 
9000b12c r32 
9000b130 r32 
9000b134 r32