🦈 add bram tool i forgot about

This commit is contained in:
xenia 2020-09-17 13:47:41 -04:00
parent 84c3f71bb4
commit 430b5a9f13
3 changed files with 37144 additions and 0 deletions

View File

@ -12,3 +12,16 @@ WARNING: very poor code quality. expect that you'll probably have to make change
from working correctly. additionally, some low-level steps needed to be taken to ensure the from working correctly. additionally, some low-level steps needed to be taken to ensure the
readback was not corrupted. comments explain each step. you will need to adjust the parameters readback was not corrupted. comments explain each step. you will need to adjust the parameters
based on the fpga part you have, however. also it needs to be built with the xilinx toolchain based on the fpga part you have, however. also it needs to be built with the xilinx toolchain
- bram_decode_tool - decodes block RAM frames dumped from 7-series bitstreams by guessing the order
of bitlanes based on a provided "known file" using some hacky statistical analysis. uses a
project x-ray database to understand the content of the block RAM frames - included is the
zynq7 database file from https://github.com/SymbiFlow/prjxray-db/ (licensed CC0). however you
can provide a different database file for different parts. this tool only handles the BRAM
construction where 32 BRAMs compose each of the 32 bits in a word, because that's the specific
scenario i wrote it for. if you have a different setup, where BRAM elements hold more than one
bit of a word you'll need to make changes. additionally, it's based on pure guesswork and needs
you to know at least some of the contents of the RAM for comparison. perhaps a smarter version
could check where the BRAM elements are connected and guess the order based on that - but it
would require lifting the rest of the configuration space which is hard (though one chinese
paper very dubiously claims to have lifted xilinx AND intel bitstreams to RTL code -
https://doi.org/10.1109/ACCESS.2019.2901949 - to this i say code or it didn't happen tbh)

36864
bram_decode_tool/bram.db Normal file

File diff suppressed because it is too large Load Diff

267
bram_decode_tool/decoder.py Executable file
View File

@ -0,0 +1,267 @@
#!/usr/bin/env python3
import re
import struct
import os
import argparse
from difflib import SequenceMatcher
WORD_BE = struct.Struct(">I")
def _unpack_word(word: bytes) -> int:
return WORD_BE.unpack(word)[0]
WORD_LE = struct.Struct("<I")
def _unpack_word_le(word: bytes) -> int:
return WORD_LE.unpack(word)[0]
# a section of the microblaze code that's always the same (libc init stuff)
# KNOWN_CODE_BOUNDS = (0x50, 0x2f4)
# KNOWN_CODE_BOUNDS = (0x50, 0x120)
# KNOWN_CODE_BOUNDS = (0x50, 0x100)
KNOWN_CODE_BOUNDS = (0, 0x900)
BRAM_DB_FNAME = "bram.db"
XRAY_DB_FMT = re.compile(r"(.*?) ([0-9]+)_([0-9]+)")
XRAY_BRAM_WORD_OFFSETS = [0, 10, 20, 30, 40, 51, 61, 71, 81, 91]
def load_db(name):
db = {}
with open(name, "r") as f:
for line in f:
m = XRAY_DB_FMT.search(line)
if m is not None:
initstring, framenum, bitnum = [m.group(x) for x in range(1,4)]
db[(int(framenum), int(bitnum))] = line.strip().split(" ")[0]
return db
def make_far_addr(bottom_top, row, col, minor):
addr = 0x00800000 | (bottom_top << 22)
addr |= (row << 17)
addr |= (col << 7)
addr |= minor
return addr
def all_bram_cols():
for bottom_top in [0, 1]:
for row in [0]:
for col in range(5):
yield (bottom_top, row, col)
def load_init_bits(inputfile):
with open(inputfile, "rb") as f:
buf = f.read()
# old dump
bottom = b"\x00" * 404 + buf[0:101*4*128*5 - 404]
# skip 2 pad frames in between
# top = buf[101*4*(128*5 + 2):]
top = buf[101*4*(128*5 + 1):101*4*(128*5 + 1)+404*128*5]
# end old dump
# adjusted for new dump
# offs = 404
# bottom = buf[offs:offs+404*128*5]
# offs = offs + 404*128*5 + 808
# top = buf[offs:offs+404*128*5]
# end new dump
buf = bottom + top
assert len(buf) == 1280 * 4 * 101
init_bits = set()
addrs = set()
frame_idx = 0
for (bottom_top, row, col) in all_bram_cols():
for minor in range(128):
frame = buf[frame_idx*101*4:(frame_idx+1)*101*4]
assert len(frame) == 404
frame_idx += 1
words = [_unpack_word(frame[j*4:j*4+4]) for j in range(101)]
assert len(words) == 101
addr = make_far_addr(bottom_top, row, col, minor)
num_bits = 0
for i, word in enumerate(words):
for bit in range(32):
if (word >> bit) & 1 == 1:
num_bits += 1
bit_repr = (addr, i, bit)
init_bits.add(bit_repr)
if num_bits > 101*32/30:
addrs.add(addr)
addrs = sorted(addrs)
start = None
last = None
in_range = False
print("addrs:")
for i in range(len(addrs)):
if not in_range:
start = addrs[i]
last = addrs[i]
in_range = True
elif addrs[i] - 1 != last:
print("Addrs: ", hex(start), "-", hex(last))
start = addrs[i]
last = addrs[i]
else:
last = addrs[i]
if in_range:
print("Addrs: ", hex(start), "-", hex(last))
# init_bits = set()
# for (bottom_top, row, col) in all_bram_cols():
# for minor in range(128):
# s = "bram_orig/initmem_initialized.bit-frames-_bram_"
# # s = "bram_test/test_out.bit-frames-_bram_"
# # s = "bram_sweep/test_out_sweep.bit-frames-_bram_"
# # s = "bram_onelane/test_out_onelane.bit-frames-_bram_"
# # s = "bram_incr/test_out_incr.bit-frames-_bram_"
# s += f"{bottom_top}_{row}_{col}_{minor}.dat"
# with open(s, "rb") as f:
# frame = f.read()
# assert len(frame) == 404
# words = [_unpack_word(frame[j*4:j*4+4]) for j in range(101)]
# assert len(words) == 101
#
# addr = make_far_addr(bottom_top, row, col, minor)
# for i, word in enumerate(words):
# for bit in range(32):
# if (word >> bit) & 1 == 1:
# # in the test data, we appear to have dumped it wrong oops
# # offset addr by 1 to fix
# bit_repr = (addr+1, i, bit)
# init_bits.add(bit_repr)
return init_bits
def preview_bitlane(bitlane, fmt_hex=True):
s = ""
for i in range(0, 256//2, 4):
chars = "".join([str(x) for x in bitlane[i:i+4]])
s += hex(int(chars, 2))[2:] if fmt_hex else chars
return s
def decode_bitlanes(init_bits, db):
bitlanes = []
for (bottom_top, row, col) in all_bram_cols():
for word_offs in XRAY_BRAM_WORD_OFFSETS:
# print("processing", bottom_top, row, col, word_offs)
initstrings = set()
for minor in range(128):
for word in range(10):
for bit in range(32):
addr = make_far_addr(bottom_top, row, col, minor)
if (addr, word_offs + word, bit) in init_bits:
# lookup the init string, it's the frame in 0-127, word number
# (within the 10-word segment)
initline = db.get((minor, word*32 + bit), None)
# if initline is None:
# print("[!] warn: unexpected bit", addr, word_offs, minor,
# word*32 + bit)
initstrings.add(initline)
if len(initstrings) == 0:
continue
# print(len(initstrings))
bitlane = bytearray(b"\x00" * 32768)
# cnt = 0
for i in range(len(bitlane)):
position = i//2
y_index = i%2
initstring = f"{position//256:02x}".upper()
initbit = f"{position%256:03d}"
if f"BRAM_L.RAMB18_Y{y_index}.INIT_{initstring}[{initbit}]" in initstrings:
# cnt += 1
# print(i, position, y_index, initstring, initbit)
bitlane[i] = 1
# print(cnt)
# print(preview_bitlane(bitlane))
# raise SystemExit()
bitlanes.append(bitlane)
return bitlanes
def get_similarity(a, b):
l = len(a)
assert len(a) == len(b)
d = 0
for i in range(l):
if a[i] != b[i]:
d += 1
return d/l
# matcher = SequenceMatcher(None, a, b)
# return (l - sum([b.size for b in matcher.get_matching_blocks()])) / l
def main(known_code, inputfile, outputfile):
print("[+] initializing database")
db = load_db(BRAM_DB_FNAME)
print("[+] loading file")
init_bits = load_init_bits(inputfile)
print("[+] decoding bitlanes")
bitlanes = decode_bitlanes(init_bits, db)
print("[+] brute forcing order")
# for bn, bitlane in enumerate(bitlanes):
# print(bn, preview_bitlane(bitlane))
mapping = {}
with open(known_code, "rb") as f:
bounds = KNOWN_CODE_BOUNDS
bounds = (0, 0x100)
data = f.read()[0:bounds[1]*4]
words = [_unpack_word(data[j*4:j*4+4]) for j in range(0, bounds[1])]
# hax
# bounds = (0xfad0//4, 0xfb60//4)
# data = b"\x00" * bounds[0]*4 + f.read()[0x11ba0 + 4:]
# words = [_unpack_word(data[j*4:j*4+4]) for j in range(0, len(data)//4)]
# mapping[7] = (15, -1)
# mapping[15] = (59, -1)
# mapping[23] = (4, -1)
# mapping[31] = (0, -1)
# end hax
for i in range(32):
bitlane = bytearray(b"\x00"*bounds[1])
for j in range(*bounds):
bitlane[j] = (words[j] >> i) & 1
for bn, rb_lane in enumerate(bitlanes):
a = rb_lane[bounds[0]:bounds[1]]
b = bitlane[bounds[0]:bounds[1]]
similarity = get_similarity(a, b)
if similarity < 0.2:
if mapping.get(i, None) is not None:
if similarity < mapping[i][1]:
mapping[i] = (bn, similarity)
else:
mapping[i] = (bn, similarity)
if mapping.get(i, None) is None:
print("[!] ERROR: no candidate bitlane found for bit", i)
for k,v in mapping.items():
print(f"{k:02d} -> {v[0]:02d} (similarity {v[1]})")
for bn in set(range(len(bitlanes))) - set([v[0] for v in mapping.values()]):
if bitlanes[bn].count(1) > 200:
print("candidate left", bn)
if len(mapping) < 32:
print("[!] brute force failure!")
print("[*] bitlanes are shown below")
return
print("[+] brute force success, writing bin")
words = [0] * 32768
for bit in range(32):
bitlane = bitlanes[mapping[bit][0]]
for word_num in range(len(words)):
words[word_num] |= (bitlane[word_num]&1) << bit
with open(outputfile, "wb") as f:
for word in words:
f.write(WORD_BE.pack(word))
if __name__ == "__main__":
parse = argparse.ArgumentParser()
parse.add_argument("knownfile")
parse.add_argument("inputfile")
parse.add_argument("outputfile")
args = parse.parse_args()
main(args.knownfile, args.inputfile, args.outputfile)