def test_CanUseImgdiff_ineligible(self):
# Disabled by caller.
block_image_diff = BlockImageDiff(EmptyImage(), EmptyImage(),
disable_imgdiff=True)
self.assertFalse(
block_image_diff.CanUseImgdiff(
"/system/app/app1.apk", RangeSet("10-15"), RangeSet("0-5")))
# Unsupported file type.
block_image_diff = BlockImageDiff(EmptyImage(), EmptyImage())
self.assertFalse(
block_image_diff.CanUseImgdiff(
"/system/bin/gzip", RangeSet("10-15"), RangeSet("0-5")))
# At least one of the ranges is in non-monotonic order.
self.assertFalse(
block_image_diff.CanUseImgdiff(
"/system/app/app2.apk", RangeSet("10-15"),
RangeSet("15-20 30 10-14")))
# At least one of the ranges is incomplete.
src_ranges = RangeSet("0-5")
src_ranges.extra['incomplete'] = True
self.assertFalse(
block_image_diff.CanUseImgdiff(
"/vendor/app/app4.apk", RangeSet("10-15"), src_ranges))
# The stats are correctly logged.
self.assertDictEqual(
{
ImgdiffStats.SKIPPED_NONMONOTONIC: {'/system/app/app2.apk'},
ImgdiffStats.SKIPPED_INCOMPLETE: {'/vendor/app/app4.apk'},
},
block_image_diff.imgdiff_stats.stats)
def AssertSequenceGood(self):
# Simulate the sequences of transfers we will output, and check that:
# - we never read a block after writing it, and
# - we write every block we care about exactly once.
# Start with no blocks having been touched yet.
touched = RangeSet()
# Imagine processing the transfers in order.
for xf in self.transfers:
# Check that the input blocks for this transfer haven't yet been touched.
x = xf.src_ranges
if self.version >= 2:
for _, sr in xf.use_stash:
x = x.subtract(sr)
assert not touched.overlaps(x)
# Check that the output blocks for this transfer haven't yet been touched.
assert not touched.overlaps(xf.tgt_ranges)
# Touch all the blocks written by this transfer.
touched = touched.union(xf.tgt_ranges)
# Check that we've written every target block.
assert touched == self.tgt.care_map
def AssertPartition(total, seq):
"""Assert that all the RangeSets in 'seq' form a partition of the
'total' RangeSet (ie, they are nonintersecting and their union
equals 'total')."""
so_far = RangeSet()
for i in seq:
assert not so_far.overlaps(i)
so_far = so_far.union(i)
assert so_far == total
def test_ValidateFileConsistency_incompleteRange(self):
input_tmp = common.MakeTempDir()
os.mkdir(os.path.join(input_tmp, 'IMAGES'))
system_image = os.path.join(input_tmp, 'IMAGES', 'system.img')
system_root = os.path.join(input_tmp, "SYSTEM")
os.mkdir(system_root)
# Write the test file that contain multiple blocks of zeros, and these
# zero blocks will be omitted by kernel. And the test files will occupy one
# block range each in the final system image.
with open(os.path.join(system_root, 'a'), 'w') as f:
f.write("aaa")
f.write('\0' * 4096 * 3)
with open(os.path.join(system_root, 'b'), 'w') as f:
f.write("bbb")
f.write('\0' * 4096 * 3)
raw_file_map = os.path.join(input_tmp, 'IMAGES', 'raw_system.map')
self._generate_system_image(system_image, system_root, raw_file_map)
# Parse the generated file map and update the block ranges for each file.
file_map_list = {}
image_ranges = RangeSet()
with open(raw_file_map, 'r') as f:
for line in f.readlines():
info = line.split()
self.assertEqual(2, len(info))
image_ranges = image_ranges.union(RangeSet(info[1]))
file_map_list[info[0]] = RangeSet(info[1])
# Add one unoccupied block as the shared block for all test files.
mock_shared_block = RangeSet("10-20").subtract(image_ranges).first(1)
with open(os.path.join(input_tmp, 'IMAGES', 'system.map'), 'w') as f:
for key in sorted(file_map_list.keys()):
line = "{} {}\n".format(
key, file_map_list[key].union(mock_shared_block))
f.write(line)
# Prepare for the target zip file
input_file = common.MakeTempFile()
all_entries = ['SYSTEM/', 'SYSTEM/b', 'SYSTEM/a', 'IMAGES/',
'IMAGES/system.map', 'IMAGES/system.img']
with zipfile.ZipFile(input_file, 'w') as input_zip:
for name in all_entries:
input_zip.write(os.path.join(input_tmp, name), arcname=name)
input_zip = zipfile.ZipFile(input_file, 'r')
info_dict = {'extfs_sparse_flag': '-s'}
# Expect the validation to pass and both files are skipped due to
# 'incomplete' block range.
ValidateFileConsistency(input_zip, input_tmp, info_dict)
def test_parse_raw(self):
self.assertEqual(
RangeSet.parse_raw(RangeSet("0-9").to_string_raw()),
RangeSet("0-9"))
self.assertEqual(RangeSet.parse_raw(
RangeSet("2-10 12").to_string_raw()),
RangeSet("2-10 12"))
self.assertEqual(
RangeSet.parse_raw(RangeSet("11 2-10 12 1 0").to_string_raw()),
RangeSet("11 2-10 12 1 0"))
with self.assertRaises(AssertionError):
RangeSet.parse_raw("4,0,10")
def __init__(self, tgt, src=None, threads=None, version=4,
disable_imgdiff=False):
if threads is None:
threads = multiprocessing.cpu_count() // 2
if threads == 0:
threads = 1
self.threads = threads
self.version = version
self.transfers = []
self.src_basenames = {}
self.src_numpatterns = {}
self._max_stashed_size = 0
self.touched_src_ranges = RangeSet()
self.touched_src_sha1 = None
self.disable_imgdiff = disable_imgdiff
assert version in (1, 2, 3, 4)
self.tgt = tgt
if src is None:
src = EmptyImage()
self.src = src
# The updater code that installs the patch always uses 4k blocks.
assert tgt.blocksize == 4096
assert src.blocksize == 4096
# The range sets in each filemap should comprise a partition of
# the care map.
self.AssertPartition(src.care_map, src.file_map.values())
self.AssertPartition(tgt.care_map, tgt.file_map.values())
class EmptyImage(Image):
"""A zero-length image."""
def __init__(self):
self.blocksize = 4096
self.care_map = RangeSet()
self.clobbered_blocks = RangeSet()
self.extended = RangeSet()
self.total_blocks = 0
self.file_map = {}
def RangeSha1(self, ranges):
return sha1().hexdigest()
def ReadRangeSet(self, ranges):
return ()
def TotalSha1(self, include_clobbered_blocks=False):
# EmptyImage always carries empty clobbered_blocks, so
# include_clobbered_blocks can be ignored.
assert self.clobbered_blocks.size() == 0
return sha1().hexdigest()
def WriteRangeDataToFd(self, ranges, fd):
raise ValueError("Can't write data from EmptyImage to file")
def __init__(self):
self.blocksize = 4096
self.care_map = RangeSet()
self.clobbered_blocks = RangeSet()
self.extended = RangeSet()
self.total_blocks = 0
self.file_map = {}
def __init__(self, data, trim=False, pad=False):
self.data = data
self.blocksize = 4096
assert not (trim and pad)
partial = len(self.data) % self.blocksize
padded = False
if partial > 0:
if trim:
self.data = self.data[:-partial]
elif pad:
self.data += '\0' * (self.blocksize - partial)
padded = True
else:
raise ValueError(("data for DataImage must be multiple of %d bytes "
"unless trim or pad is specified") %
(self.blocksize,))
assert len(self.data) % self.blocksize == 0
self.total_blocks = len(self.data) / self.blocksize
self.care_map = RangeSet(data=(0, self.total_blocks))
# When the last block is padded, we always write the whole block even for
# incremental OTAs. Because otherwise the last block may get skipped if
# unchanged for an incremental, but would fail the post-install
# verification if it has non-zero contents in the padding bytes.
# Bug: 23828506
if padded:
clobbered_blocks = [self.total_blocks-1, self.total_blocks]
else:
clobbered_blocks = []
self.clobbered_blocks = clobbered_blocks
self.extended = RangeSet()
zero_blocks = []
nonzero_blocks = []
reference = '\0' * self.blocksize
for i in range(self.total_blocks-1 if padded else self.total_blocks):
d = self.data[i*self.blocksize : (i+1)*self.blocksize]
if d == reference:
zero_blocks.append(i)
zero_blocks.append(i+1)
else:
nonzero_blocks.append(i)
nonzero_blocks.append(i+1)
assert zero_blocks or nonzero_blocks or clobbered_blocks
self.file_map = dict()
if zero_blocks:
self.file_map["__ZERO"] = RangeSet(data=zero_blocks)
if nonzero_blocks:
self.file_map["__NONZERO"] = RangeSet(data=nonzero_blocks)
if clobbered_blocks:
self.file_map["__COPY"] = RangeSet(data=clobbered_blocks)
class DataImage(Image):
"""An image wrapped around a single string of data."""
def __init__(self, data, trim=False, pad=False):
self.data = data
self.blocksize = 4096
assert not (trim and pad)
partial = len(self.data) % self.blocksize
if partial > 0:
if trim:
self.data = self.data[:-partial]
elif pad:
self.data += '\0' * (self.blocksize - partial)
else:
raise ValueError(("data for DataImage must be multiple of %d bytes "
"unless trim or pad is specified") %
(self.blocksize,))
assert len(self.data) % self.blocksize == 0
self.total_blocks = len(self.data) / self.blocksize
self.care_map = RangeSet(data=(0, self.total_blocks))
self.clobbered_blocks = RangeSet()
self.extended = RangeSet()
zero_blocks = []
nonzero_blocks = []
reference = '\0' * self.blocksize
for i in range(self.total_blocks):
d = self.data[i*self.blocksize : (i+1)*self.blocksize]
if d == reference:
zero_blocks.append(i)
zero_blocks.append(i+1)
else:
nonzero_blocks.append(i)
nonzero_blocks.append(i+1)
self.file_map = {"__ZERO": RangeSet(zero_blocks),
"__NONZERO": RangeSet(nonzero_blocks)}
def ReadRangeSet(self, ranges):
return [self.data[s*self.blocksize:e*self.blocksize] for (s, e) in ranges]
def TotalSha1(self, include_clobbered_blocks=False):
# DataImage always carries empty clobbered_blocks, so
# include_clobbered_blocks can be ignored.
assert self.clobbered_blocks.size() == 0
return sha1(self.data).hexdigest()
def __init__(self, data, trim=False, pad=False):
self.data = data
self.blocksize = 4096
assert not (trim and pad)
partial = len(self.data) % self.blocksize
if partial > 0:
if trim:
self.data = self.data[:-partial]
elif pad:
self.data += '\0' * (self.blocksize - partial)
else:
raise ValueError(("data for DataImage must be multiple of %d bytes "
"unless trim or pad is specified") %
(self.blocksize,))
assert len(self.data) % self.blocksize == 0
self.total_blocks = len(self.data) / self.blocksize
self.care_map = RangeSet(data=(0, self.total_blocks))
self.clobbered_blocks = RangeSet()
self.extended = RangeSet()
zero_blocks = []
nonzero_blocks = []
reference = '\0' * self.blocksize
for i in range(self.total_blocks):
d = self.data[i*self.blocksize : (i+1)*self.blocksize]
if d == reference:
zero_blocks.append(i)
zero_blocks.append(i+1)
else:
nonzero_blocks.append(i)
nonzero_blocks.append(i+1)
self.file_map = {"__ZERO": RangeSet(zero_blocks),
"__NONZERO": RangeSet(nonzero_blocks)}
def test_union(self):
self.assertEqual(RangeSet("10-19 30-34").union(RangeSet("18-29")),
RangeSet("10-34"))
self.assertEqual(RangeSet("10-19 30-34").union(RangeSet("22 32")),
RangeSet("10-19 22 30-34"))
def FindTransfers(self):
"""Parse the file_map to generate all the transfers."""
def AddTransfer(tgt_name, src_name, tgt_ranges, src_ranges, style, by_id,
split=False):
"""Wrapper function for adding a Transfer().
For BBOTA v3, we need to stash source blocks for resumable feature.
However, with the growth of file size and the shrink of the cache
partition source blocks are too large to be stashed. If a file occupies
too many blocks (greater than MAX_BLOCKS_PER_DIFF_TRANSFER), we split it
into smaller pieces by getting multiple Transfer()s.
The downside is that after splitting, we can no longer use imgdiff but
only bsdiff."""
MAX_BLOCKS_PER_DIFF_TRANSFER = 1024
# We care about diff transfers only.
if style != "diff" or not split:
Transfer(tgt_name, src_name, tgt_ranges, src_ranges, style, by_id)
return
# Change nothing for small files.
if (tgt_ranges.size() <= MAX_BLOCKS_PER_DIFF_TRANSFER and
src_ranges.size() <= MAX_BLOCKS_PER_DIFF_TRANSFER):
Transfer(tgt_name, src_name, tgt_ranges, src_ranges, style, by_id)
return
pieces = 0
while (tgt_ranges.size() > MAX_BLOCKS_PER_DIFF_TRANSFER and
src_ranges.size() > MAX_BLOCKS_PER_DIFF_TRANSFER):
tgt_split_name = "%s-%d" % (tgt_name, pieces)
src_split_name = "%s-%d" % (src_name, pieces)
tgt_first = tgt_ranges.first(MAX_BLOCKS_PER_DIFF_TRANSFER)
src_first = src_ranges.first(MAX_BLOCKS_PER_DIFF_TRANSFER)
Transfer(tgt_split_name, src_split_name, tgt_first, src_first, style,
by_id)
tgt_ranges = tgt_ranges.subtract(tgt_first)
src_ranges = src_ranges.subtract(src_first)
pieces += 1
# Handle remaining blocks.
if tgt_ranges.size() or src_ranges.size():
# Must be both non-empty.
assert tgt_ranges.size() and src_ranges.size()
tgt_split_name = "%s-%d" % (tgt_name, pieces)
src_split_name = "%s-%d" % (src_name, pieces)
Transfer(tgt_split_name, src_split_name, tgt_ranges, src_ranges, style,
by_id)
empty = RangeSet()
for tgt_fn, tgt_ranges in self.tgt.file_map.items():
if tgt_fn == "__ZERO":
# the special "__ZERO" domain is all the blocks not contained
# in any file and that are filled with zeros. We have a
# special transfer style for zero blocks.
src_ranges = self.src.file_map.get("__ZERO", empty)
AddTransfer(tgt_fn, "__ZERO", tgt_ranges, src_ranges,
"zero", self.transfers)
continue
elif tgt_fn == "__COPY":
# "__COPY" domain includes all the blocks not contained in any
# file and that need to be copied unconditionally to the target.
AddTransfer(tgt_fn, None, tgt_ranges, empty, "new", self.transfers)
continue
elif tgt_fn in self.src.file_map:
# Look for an exact pathname match in the source.
AddTransfer(tgt_fn, tgt_fn, tgt_ranges, self.src.file_map[tgt_fn],
"diff", self.transfers, self.version >= 3)
continue
b = os.path.basename(tgt_fn)
if b in self.src_basenames:
# Look for an exact basename match in the source.
src_fn = self.src_basenames[b]
AddTransfer(tgt_fn, src_fn, tgt_ranges, self.src.file_map[src_fn],
"diff", self.transfers, self.version >= 3)
continue
b = re.sub("[0-9]+", "#", b)
if b in self.src_numpatterns:
# Look for a 'number pattern' match (a basename match after
# all runs of digits are replaced by "#"). (This is useful
# for .so files that contain version numbers in the filename
# that get bumped.)
src_fn = self.src_numpatterns[b]
AddTransfer(tgt_fn, src_fn, tgt_ranges, self.src.file_map[src_fn],
"diff", self.transfers, self.version >= 3)
continue
AddTransfer(tgt_fn, None, tgt_ranges, empty, "new", self.transfers)
def test_monotonic(self):
self.assertTrue(RangeSet("0-9").monotonic)
self.assertTrue(RangeSet("2-9").monotonic)
self.assertTrue(RangeSet("2-9 30 31 35").monotonic)
self.assertTrue(RangeSet("").monotonic)
self.assertTrue(RangeSet("0-4 5-9").monotonic)
self.assertFalse(RangeSet("5-9 0-4").monotonic)
self.assertFalse(RangeSet("258768-259211 196604").monotonic)
self.assertTrue(RangeSet(data=[0, 10]).monotonic)
self.assertTrue(RangeSet(data=[0, 10, 15, 20]).monotonic)
self.assertTrue(RangeSet(data=[2, 9, 30, 31, 31, 32, 35, 36]).monotonic)
self.assertTrue(RangeSet(data=[0, 5, 5, 10]).monotonic)
self.assertFalse(RangeSet(data=[5, 10, 0, 5]).monotonic)
class DataImage(Image):
"""An image wrapped around a single string of data."""
def __init__(self, data, trim=False, pad=False):
self.data = data
self.blocksize = 4096
assert not (trim and pad)
partial = len(self.data) % self.blocksize
padded = False
if partial > 0:
if trim:
self.data = self.data[:-partial]
elif pad:
self.data += '\0' * (self.blocksize - partial)
padded = True
else:
raise ValueError(("data for DataImage must be multiple of %d bytes "
"unless trim or pad is specified") %
(self.blocksize,))
assert len(self.data) % self.blocksize == 0
self.total_blocks = len(self.data) / self.blocksize
self.care_map = RangeSet(data=(0, self.total_blocks))
# When the last block is padded, we always write the whole block even for
# incremental OTAs. Because otherwise the last block may get skipped if
# unchanged for an incremental, but would fail the post-install
# verification if it has non-zero contents in the padding bytes.
# Bug: 23828506
if padded:
clobbered_blocks = [self.total_blocks-1, self.total_blocks]
else:
clobbered_blocks = []
self.clobbered_blocks = clobbered_blocks
self.extended = RangeSet()
zero_blocks = []
nonzero_blocks = []
reference = '\0' * self.blocksize
for i in range(self.total_blocks-1 if padded else self.total_blocks):
d = self.data[i*self.blocksize : (i+1)*self.blocksize]
if d == reference:
zero_blocks.append(i)
zero_blocks.append(i+1)
else:
nonzero_blocks.append(i)
nonzero_blocks.append(i+1)
assert zero_blocks or nonzero_blocks or clobbered_blocks
self.file_map = dict()
if zero_blocks:
self.file_map["__ZERO"] = RangeSet(data=zero_blocks)
if nonzero_blocks:
self.file_map["__NONZERO"] = RangeSet(data=nonzero_blocks)
if clobbered_blocks:
self.file_map["__COPY"] = RangeSet(data=clobbered_blocks)
def _GetRangeData(self, ranges):
for s, e in ranges:
yield self.data[s*self.blocksize:e*self.blocksize]
def RangeSha1(self, ranges):
h = sha1()
for data in self._GetRangeData(ranges):
h.update(data)
return h.hexdigest()
def ReadRangeSet(self, ranges):
return [self._GetRangeData(ranges)]
def TotalSha1(self, include_clobbered_blocks=False):
if not include_clobbered_blocks:
return self.RangeSha1(self.care_map.subtract(self.clobbered_blocks))
else:
return sha1(self.data).hexdigest()
def WriteRangeDataToFd(self, ranges, fd):
for data in self._GetRangeData(ranges):
fd.write(data)
class DataImage(Image):
"""An image wrapped around a single string of data."""
def __init__(self, data, trim=False, pad=False):
self.data = data
self.blocksize = 4096
assert not (trim and pad)
partial = len(self.data) % self.blocksize
padded = False
if partial > 0:
if trim:
self.data = self.data[:-partial]
elif pad:
self.data += '\0' * (self.blocksize - partial)
padded = True
else:
raise ValueError(("data for DataImage must be multiple of %d bytes "
"unless trim or pad is specified") %
(self.blocksize,))
assert len(self.data) % self.blocksize == 0
self.total_blocks = len(self.data) / self.blocksize
self.care_map = RangeSet(data=(0, self.total_blocks))
# When the last block is padded, we always write the whole block even for
# incremental OTAs. Because otherwise the last block may get skipped if
# unchanged for an incremental, but would fail the post-install
# verification if it has non-zero contents in the padding bytes.
# Bug: 23828506
if padded:
clobbered_blocks = [self.total_blocks - 1, self.total_blocks]
else:
clobbered_blocks = []
self.clobbered_blocks = clobbered_blocks
self.extended = RangeSet()
zero_blocks = []
nonzero_blocks = []
reference = '\0' * self.blocksize
for i in range(self.total_blocks - 1 if padded else self.total_blocks):
d = self.data[i * self.blocksize: (i + 1) * self.blocksize]
if d == reference:
zero_blocks.append(i)
zero_blocks.append(i + 1)
else:
nonzero_blocks.append(i)
nonzero_blocks.append(i + 1)
assert zero_blocks or nonzero_blocks or clobbered_blocks
self.file_map = dict()
if zero_blocks:
self.file_map["__ZERO"] = RangeSet(data=zero_blocks)
if nonzero_blocks:
self.file_map["__NONZERO"] = RangeSet(data=nonzero_blocks)
if clobbered_blocks:
self.file_map["__COPY"] = RangeSet(data=clobbered_blocks)
def ReadRangeSet(self, ranges):
return [self.data[s * self.blocksize:e * self.blocksize] for (s, e) in ranges]
def TotalSha1(self, include_clobbered_blocks=False):
if not include_clobbered_blocks:
ranges = self.care_map.subtract(self.clobbered_blocks)
return sha1(self.ReadRangeSet(ranges)).hexdigest()
else:
return sha1(self.data).hexdigest()
def ParseTransferList(self, name):
"""Simulate the transfer commands and calculate the amout of I/O."""
logging.info("\nSimulating commands in '{}':".format(name))
lines = self.package.read(name).strip().splitlines()
assert len(lines) >= 4, "{} is too short; Transfer list expects at least" \
"4 lines, it has {}".format(name, len(lines))
assert int(lines[0]) >= 3
logging.info("(version: {})".format(lines[0]))
blocks_written = 0
my_stash = Stash()
for line in lines[4:]:
cmd_list = line.strip().split(" ")
cmd_name = cmd_list[0]
try:
if cmd_name == "new" or cmd_name == "zero":
assert len(cmd_list) == 2, "command format error: {}".format(line)
target_range = RangeSet.parse_raw(cmd_list[1])
blocks_written += target_range.size()
elif cmd_name == "move":
# Example: move <onehash> <tgt_range> <src_blk_count> <src_range>
# [<loc_range> <stashed_blocks>]
assert len(cmd_list) >= 5, "command format error: {}".format(line)
target_range = RangeSet.parse_raw(cmd_list[2])
blocks_written += target_range.size()
if cmd_list[4] == '-':
continue
SHA1 = cmd_list[1]
source_range = RangeSet.parse_raw(cmd_list[4])
if target_range.overlaps(source_range):
my_stash.HandleOverlapBlocks(SHA1, source_range)
elif cmd_name == "bsdiff" or cmd_name == "imgdiff":
# Example: bsdiff <offset> <len> <src_hash> <tgt_hash> <tgt_range>
# <src_blk_count> <src_range> [<loc_range> <stashed_blocks>]
assert len(cmd_list) >= 8, "command format error: {}".format(line)
target_range = RangeSet.parse_raw(cmd_list[5])
blocks_written += target_range.size()
if cmd_list[7] == '-':
continue
source_SHA1 = cmd_list[3]
source_range = RangeSet.parse_raw(cmd_list[7])
if target_range.overlaps(source_range):
my_stash.HandleOverlapBlocks(source_SHA1, source_range)
elif cmd_name == "stash":
assert len(cmd_list) == 3, "command format error: {}".format(line)
SHA1 = cmd_list[1]
source_range = RangeSet.parse_raw(cmd_list[2])
my_stash.StashBlocks(SHA1, source_range)
elif cmd_name == "free":
assert len(cmd_list) == 2, "command format error: {}".format(line)
SHA1 = cmd_list[1]
my_stash.FreeBlocks(SHA1)
except:
logging.error("failed to parse command in: " + line)
raise
self.block_written += blocks_written
self.block_stashed += my_stash.blocks_stashed
logging.info("blocks written: {} (expected: {})".format(
blocks_written, lines[1]))
logging.info("max blocks stashed simultaneously: {} (expected: {})".
format(my_stash.max_stash_needed, lines[3]))
logging.info("total blocks stashed: {}".format(my_stash.blocks_stashed))
logging.info("blocks stashed implicitly: {}".format(
my_stash.overlap_blocks_stashed))
def FindTransfers(self):
"""Parse the file_map to generate all the transfers."""
def AddSplitTransfers(tgt_name, src_name, tgt_ranges, src_ranges,
style, by_id):
"""Add one or multiple Transfer()s by splitting large files.
For BBOTA v3, we need to stash source blocks for resumable feature.
However, with the growth of file size and the shrink of the cache
partition source blocks are too large to be stashed. If a file occupies
too many blocks, we split it into smaller pieces by getting multiple
Transfer()s.
The downside is that after splitting, we may increase the package size
since the split pieces don't align well. According to our experiments,
1/8 of the cache size as the per-piece limit appears to be optimal.
Compared to the fixed 1024-block limit, it reduces the overall package
size by 30% for volantis, and 20% for angler and bullhead."""
# Possibly split large files into smaller chunks.
pieces = 0
cache_size = common.OPTIONS.cache_size
split_threshold = 0.125
max_blocks_per_transfer = int(cache_size * split_threshold /
self.tgt.blocksize)
# Change nothing for small files.
if (tgt_ranges.size() <= max_blocks_per_transfer and
src_ranges.size() <= max_blocks_per_transfer):
Transfer(tgt_name, src_name, tgt_ranges, src_ranges, style, by_id)
return
while (tgt_ranges.size() > max_blocks_per_transfer and
src_ranges.size() > max_blocks_per_transfer):
tgt_split_name = "%s-%d" % (tgt_name, pieces)
src_split_name = "%s-%d" % (src_name, pieces)
tgt_first = tgt_ranges.first(max_blocks_per_transfer)
src_first = src_ranges.first(max_blocks_per_transfer)
Transfer(tgt_split_name, src_split_name, tgt_first, src_first, style,
by_id)
tgt_ranges = tgt_ranges.subtract(tgt_first)
src_ranges = src_ranges.subtract(src_first)
pieces += 1
# Handle remaining blocks.
if tgt_ranges.size() or src_ranges.size():
# Must be both non-empty.
assert tgt_ranges.size() and src_ranges.size()
tgt_split_name = "%s-%d" % (tgt_name, pieces)
src_split_name = "%s-%d" % (src_name, pieces)
Transfer(tgt_split_name, src_split_name, tgt_ranges, src_ranges, style,
by_id)
def AddTransfer(tgt_name, src_name, tgt_ranges, src_ranges, style, by_id,
split=False):
"""Wrapper function for adding a Transfer()."""
# We specialize diff transfers only (which covers bsdiff/imgdiff/move);
# otherwise add the Transfer() as is.
if style != "diff" or not split:
Transfer(tgt_name, src_name, tgt_ranges, src_ranges, style, by_id)
return
# Handle .odex files specially to analyze the block-wise difference. If
# most of the blocks are identical with only few changes (e.g. header),
# we will patch the changed blocks only. This avoids stashing unchanged
# blocks while patching. We limit the analysis to files without size
# changes only. This is to avoid sacrificing the OTA generation cost too
# much.
if (tgt_name.split(".")[-1].lower() == 'odex' and
tgt_ranges.size() == src_ranges.size()):
# 0.5 threshold can be further tuned. The tradeoff is: if only very
# few blocks remain identical, we lose the opportunity to use imgdiff
# that may have better compression ratio than bsdiff.
crop_threshold = 0.5
tgt_skipped = RangeSet()
src_skipped = RangeSet()
tgt_size = tgt_ranges.size()
tgt_changed = 0
for src_block, tgt_block in zip(src_ranges.next_item(),
tgt_ranges.next_item()):
src_rs = RangeSet(str(src_block))
tgt_rs = RangeSet(str(tgt_block))
if self.src.ReadRangeSet(src_rs) == self.tgt.ReadRangeSet(tgt_rs):
tgt_skipped = tgt_skipped.union(tgt_rs)
src_skipped = src_skipped.union(src_rs)
else:
tgt_changed += tgt_rs.size()
# Terminate early if no clear sign of benefits.
if tgt_changed > tgt_size * crop_threshold:
break
if tgt_changed < tgt_size * crop_threshold:
assert tgt_changed + tgt_skipped.size() == tgt_size
print('%10d %10d (%6.2f%%) %s' % (tgt_skipped.size(), tgt_size,
tgt_skipped.size() * 100.0 / tgt_size, tgt_name))
AddSplitTransfers(
"%s-skipped" % (tgt_name,),
"%s-skipped" % (src_name,),
tgt_skipped, src_skipped, style, by_id)
# Intentionally change the file extension to avoid being imgdiff'd as
# the files are no longer in their original format.
tgt_name = "%s-cropped" % (tgt_name,)
src_name = "%s-cropped" % (src_name,)
tgt_ranges = tgt_ranges.subtract(tgt_skipped)
src_ranges = src_ranges.subtract(src_skipped)
# Possibly having no changed blocks.
if not tgt_ranges:
return
# Add the transfer(s).
AddSplitTransfers(
tgt_name, src_name, tgt_ranges, src_ranges, style, by_id)
print("Finding transfers...")
empty = RangeSet()
for tgt_fn, tgt_ranges in self.tgt.file_map.items():
if tgt_fn == "__ZERO":
# the special "__ZERO" domain is all the blocks not contained
# in any file and that are filled with zeros. We have a
# special transfer style for zero blocks.
src_ranges = self.src.file_map.get("__ZERO", empty)
AddTransfer(tgt_fn, "__ZERO", tgt_ranges, src_ranges,
"zero", self.transfers)
continue
elif tgt_fn == "__COPY":
# "__COPY" domain includes all the blocks not contained in any
# file and that need to be copied unconditionally to the target.
AddTransfer(tgt_fn, None, tgt_ranges, empty, "new", self.transfers)
continue
elif tgt_fn in self.src.file_map:
# Look for an exact pathname match in the source.
AddTransfer(tgt_fn, tgt_fn, tgt_ranges, self.src.file_map[tgt_fn],
"diff", self.transfers, self.version >= 3)
continue
b = os.path.basename(tgt_fn)
if b in self.src_basenames:
# Look for an exact basename match in the source.
src_fn = self.src_basenames[b]
AddTransfer(tgt_fn, src_fn, tgt_ranges, self.src.file_map[src_fn],
"diff", self.transfers, self.version >= 3)
continue
b = re.sub("[0-9]+", "#", b)
if b in self.src_numpatterns:
# Look for a 'number pattern' match (a basename match after
# all runs of digits are replaced by "#"). (This is useful
# for .so files that contain version numbers in the filename
# that get bumped.)
src_fn = self.src_numpatterns[b]
AddTransfer(tgt_fn, src_fn, tgt_ranges, self.src.file_map[src_fn],
"diff", self.transfers, self.version >= 3)
continue
AddTransfer(tgt_fn, None, tgt_ranges, empty, "new", self.transfers)
def test_overlaps(self):
self.assertTrue(RangeSet("10-19 30-34").overlaps(RangeSet("18-32")))
self.assertFalse(RangeSet("10-19 30-34").overlaps(RangeSet("22-28")))
def WriteTransfers(self, prefix):
out = []
total = 0
stashes = {}
stashed_blocks = 0
max_stashed_blocks = 0
free_stash_ids = []
next_stash_id = 0
for xf in self.transfers:
if self.version < 2:
assert not xf.stash_before
assert not xf.use_stash
for s, sr in xf.stash_before:
assert s not in stashes
if free_stash_ids:
sid = heapq.heappop(free_stash_ids)
else:
sid = next_stash_id
next_stash_id += 1
stashes[s] = sid
stashed_blocks += sr.size()
if self.version == 2:
out.append("stash %d %s\n" % (sid, sr.to_string_raw()))
else:
sh = self.HashBlocks(self.src, sr)
if sh in stashes:
stashes[sh] += 1
else:
stashes[sh] = 1
out.append("stash %s %s\n" % (sh, sr.to_string_raw()))
if stashed_blocks > max_stashed_blocks:
max_stashed_blocks = stashed_blocks
free_string = []
if self.version == 1:
src_str = xf.src_ranges.to_string_raw()
elif self.version >= 2:
# <# blocks> <src ranges>
# OR
# <# blocks> <src ranges> <src locs> <stash refs...>
# OR
# <# blocks> - <stash refs...>
size = xf.src_ranges.size()
src_str = [str(size)]
unstashed_src_ranges = xf.src_ranges
mapped_stashes = []
for s, sr in xf.use_stash:
sid = stashes.pop(s)
stashed_blocks -= sr.size()
unstashed_src_ranges = unstashed_src_ranges.subtract(sr)
sh = self.HashBlocks(self.src, sr)
sr = xf.src_ranges.map_within(sr)
mapped_stashes.append(sr)
if self.version == 2:
src_str.append("%d:%s" % (sid, sr.to_string_raw()))
else:
assert sh in stashes
src_str.append("%s:%s" % (sh, sr.to_string_raw()))
stashes[sh] -= 1
if stashes[sh] == 0:
free_string.append("free %s\n" % sh)
stashes.pop(sh)
heapq.heappush(free_stash_ids, sid)
if unstashed_src_ranges:
src_str.insert(1, unstashed_src_ranges.to_string_raw())
if xf.use_stash:
mapped_unstashed = xf.src_ranges.map_within(unstashed_src_ranges)
src_str.insert(2, mapped_unstashed.to_string_raw())
mapped_stashes.append(mapped_unstashed)
self.AssertPartition(RangeSet(data=(0, size)), mapped_stashes)
else:
src_str.insert(1, "-")
self.AssertPartition(RangeSet(data=(0, size)), mapped_stashes)
src_str = " ".join(src_str)
# all versions:
# zero <rangeset>
# new <rangeset>
# erase <rangeset>
#
# version 1:
# bsdiff patchstart patchlen <src rangeset> <tgt rangeset>
# imgdiff patchstart patchlen <src rangeset> <tgt rangeset>
# move <src rangeset> <tgt rangeset>
#
# version 2:
# bsdiff patchstart patchlen <tgt rangeset> <src_str>
# imgdiff patchstart patchlen <tgt rangeset> <src_str>
# move <tgt rangeset> <src_str>
#
# version 3:
# bsdiff patchstart patchlen srchash tgthash <tgt rangeset> <src_str>
# imgdiff patchstart patchlen srchash tgthash <tgt rangeset> <src_str>
# move hash <tgt rangeset> <src_str>
tgt_size = xf.tgt_ranges.size()
if xf.style == "new":
assert xf.tgt_ranges
out.append("%s %s\n" % (xf.style, xf.tgt_ranges.to_string_raw()))
total += tgt_size
elif xf.style == "move":
assert xf.tgt_ranges
assert xf.src_ranges.size() == tgt_size
if xf.src_ranges != xf.tgt_ranges:
if self.version == 1:
out.append("%s %s %s\n" % (
xf.style,
xf.src_ranges.to_string_raw(), xf.tgt_ranges.to_string_raw()))
elif self.version == 2:
out.append("%s %s %s\n" % (
xf.style,
xf.tgt_ranges.to_string_raw(), src_str))
elif self.version >= 3:
# take into account automatic stashing of overlapping blocks
if xf.src_ranges.overlaps(xf.tgt_ranges):
temp_stash_usage = stashed_blocks + xf.src_ranges.size()
if temp_stash_usage > max_stashed_blocks:
max_stashed_blocks = temp_stash_usage
out.append("%s %s %s %s\n" % (
xf.style,
self.HashBlocks(self.tgt, xf.tgt_ranges),
xf.tgt_ranges.to_string_raw(), src_str))
total += tgt_size
elif xf.style in ("bsdiff", "imgdiff"):
assert xf.tgt_ranges
assert xf.src_ranges
if self.version == 1:
out.append("%s %d %d %s %s\n" % (
xf.style, xf.patch_start, xf.patch_len,
xf.src_ranges.to_string_raw(), xf.tgt_ranges.to_string_raw()))
elif self.version == 2:
out.append("%s %d %d %s %s\n" % (
xf.style, xf.patch_start, xf.patch_len,
xf.tgt_ranges.to_string_raw(), src_str))
elif self.version >= 3:
# take into account automatic stashing of overlapping blocks
if xf.src_ranges.overlaps(xf.tgt_ranges):
temp_stash_usage = stashed_blocks + xf.src_ranges.size()
if temp_stash_usage > max_stashed_blocks:
max_stashed_blocks = temp_stash_usage
out.append("%s %d %d %s %s %s %s\n" % (
xf.style,
xf.patch_start, xf.patch_len,
self.HashBlocks(self.src, xf.src_ranges),
self.HashBlocks(self.tgt, xf.tgt_ranges),
xf.tgt_ranges.to_string_raw(), src_str))
total += tgt_size
elif xf.style == "zero":
assert xf.tgt_ranges
to_zero = xf.tgt_ranges.subtract(xf.src_ranges)
if to_zero:
out.append("%s %s\n" % (xf.style, to_zero.to_string_raw()))
total += to_zero.size()
else:
raise ValueError("unknown transfer style '%s'\n" % xf.style)
if free_string:
out.append("".join(free_string))
if self.version >= 2 and common.OPTIONS.cache_size is not None:
# Sanity check: abort if we're going to need more stash space than
# the allowed size (cache_size * threshold). There are two purposes
# of having a threshold here. a) Part of the cache may have been
# occupied by some recovery logs. b) It will buy us some time to deal
# with the oversize issue.
cache_size = common.OPTIONS.cache_size
stash_threshold = common.OPTIONS.stash_threshold
max_allowed = cache_size * stash_threshold
assert max_stashed_blocks * self.tgt.blocksize < max_allowed, \
'Stash size %d (%d * %d) exceeds the limit %d (%d * %.2f)' % (
max_stashed_blocks * self.tgt.blocksize, max_stashed_blocks,
self.tgt.blocksize, max_allowed, cache_size,
stash_threshold)
# Zero out extended blocks as a workaround for bug 20881595.
if self.tgt.extended:
out.append("zero %s\n" % (self.tgt.extended.to_string_raw(),))
# We erase all the blocks on the partition that a) don't contain useful
# data in the new image and b) will not be touched by dm-verity.
all_tgt = RangeSet(data=(0, self.tgt.total_blocks))
all_tgt_minus_extended = all_tgt.subtract(self.tgt.extended)
new_dontcare = all_tgt_minus_extended.subtract(self.tgt.care_map)
if new_dontcare:
out.append("erase %s\n" % (new_dontcare.to_string_raw(),))
out.insert(0, "%d\n" % (self.version,)) # format version number
out.insert(1, str(total) + "\n")
if self.version >= 2:
# version 2 only: after the total block count, we give the number
# of stash slots needed, and the maximum size needed (in blocks)
out.insert(2, str(next_stash_id) + "\n")
out.insert(3, str(max_stashed_blocks) + "\n")
with open(prefix + ".transfer.list", "wb") as f:
for i in out:
f.write(i)
if self.version >= 2:
max_stashed_size = max_stashed_blocks * self.tgt.blocksize
OPTIONS = common.OPTIONS
if OPTIONS.cache_size is not None:
max_allowed = OPTIONS.cache_size * OPTIONS.stash_threshold
print("max stashed blocks: %d (%d bytes), "
"limit: %d bytes (%.2f%%)\n" % (
max_stashed_blocks, max_stashed_size, max_allowed,
max_stashed_size * 100.0 / max_allowed))
else:
print("max stashed blocks: %d (%d bytes), limit: <unknown>\n" % (
max_stashed_blocks, max_stashed_size))
def WriteTransfers(self, prefix):
def WriteTransfersZero(out, to_zero):
"""Limit the number of blocks in command zero to 1024 blocks.
This prevents the target size of one command from being too large; and
might help to avoid fsync errors on some devices."""
zero_blocks_limit = 1024
total = 0
while to_zero.size() > 0:
zero_blocks = to_zero.first(zero_blocks_limit)
out.append("zero %s\n" % (zero_blocks.to_string_raw(),))
total += zero_blocks.size()
to_zero = to_zero.subtract(zero_blocks)
return total
out = []
total = 0
stashes = {}
stashed_blocks = 0
max_stashed_blocks = 0
free_stash_ids = []
next_stash_id = 0
for xf in self.transfers:
if self.version < 2:
assert not xf.stash_before
assert not xf.use_stash
for s, sr in xf.stash_before:
assert s not in stashes
if free_stash_ids:
sid = heapq.heappop(free_stash_ids)
else:
sid = next_stash_id
next_stash_id += 1
stashes[s] = sid
if self.version == 2:
stashed_blocks += sr.size()
out.append("stash %d %s\n" % (sid, sr.to_string_raw()))
else:
sh = self.HashBlocks(self.src, sr)
if sh in stashes:
stashes[sh] += 1
else:
stashes[sh] = 1
stashed_blocks += sr.size()
self.touched_src_ranges = self.touched_src_ranges.union(sr)
out.append("stash %s %s\n" % (sh, sr.to_string_raw()))
if stashed_blocks > max_stashed_blocks:
max_stashed_blocks = stashed_blocks
free_string = []
free_size = 0
if self.version == 1:
src_str = xf.src_ranges.to_string_raw() if xf.src_ranges else ""
elif self.version >= 2:
# <# blocks> <src ranges>
# OR
# <# blocks> <src ranges> <src locs> <stash refs...>
# OR
# <# blocks> - <stash refs...>
size = xf.src_ranges.size()
src_str = [str(size)]
unstashed_src_ranges = xf.src_ranges
mapped_stashes = []
for s, sr in xf.use_stash:
sid = stashes.pop(s)
unstashed_src_ranges = unstashed_src_ranges.subtract(sr)
sh = self.HashBlocks(self.src, sr)
sr = xf.src_ranges.map_within(sr)
mapped_stashes.append(sr)
if self.version == 2:
src_str.append("%d:%s" % (sid, sr.to_string_raw()))
# A stash will be used only once. We need to free the stash
# immediately after the use, instead of waiting for the automatic
# clean-up at the end. Because otherwise it may take up extra space
# and lead to OTA failures.
# Bug: 23119955
free_string.append("free %d\n" % (sid,))
free_size += sr.size()
else:
assert sh in stashes
src_str.append("%s:%s" % (sh, sr.to_string_raw()))
stashes[sh] -= 1
if stashes[sh] == 0:
free_size += sr.size()
free_string.append("free %s\n" % (sh))
stashes.pop(sh)
heapq.heappush(free_stash_ids, sid)
if unstashed_src_ranges is None:
src_str.insert(1, unstashed_src_ranges.to_string_raw())
if xf.use_stash:
mapped_unstashed = xf.src_ranges.map_within(unstashed_src_ranges)
src_str.insert(2, mapped_unstashed.to_string_raw())
mapped_stashes.append(mapped_unstashed)
self.AssertPartition(RangeSet(data=(0, size)), mapped_stashes)
else:
src_str.insert(1, "-")
self.AssertPartition(RangeSet(data=(0, size)), mapped_stashes)
src_str = " ".join(src_str)
# all versions:
# zero <rangeset>
# new <rangeset>
# erase <rangeset>
#
# version 1:
# bsdiff patchstart patchlen <src rangeset> <tgt rangeset>
# imgdiff patchstart patchlen <src rangeset> <tgt rangeset>
# move <src rangeset> <tgt rangeset>
#
# version 2:
# bsdiff patchstart patchlen <tgt rangeset> <src_str>
# imgdiff patchstart patchlen <tgt rangeset> <src_str>
# move <tgt rangeset> <src_str>
#
# version 3:
# bsdiff patchstart patchlen srchash tgthash <tgt rangeset> <src_str>
# imgdiff patchstart patchlen srchash tgthash <tgt rangeset> <src_str>
# move hash <tgt rangeset> <src_str>
tgt_size = xf.tgt_ranges.size()
if xf.style == "new":
assert xf.tgt_ranges
out.append("%s %s\n" % (xf.style, xf.tgt_ranges.to_string_raw()))
total += tgt_size
elif xf.style == "move":
assert xf.tgt_ranges
assert xf.src_ranges.size() == tgt_size
if xf.src_ranges != xf.tgt_ranges:
if self.version == 1:
out.append("%s %s %s\n" % (
xf.style,
xf.src_ranges.to_string_raw(), xf.tgt_ranges.to_string_raw()))
elif self.version == 2:
out.append("%s %s %s\n" % (
xf.style,
xf.tgt_ranges.to_string_raw(), src_str))
elif self.version >= 3:
# take into account automatic stashing of overlapping blocks
if xf.src_ranges.overlaps(xf.tgt_ranges):
temp_stash_usage = stashed_blocks + xf.src_ranges.size()
if temp_stash_usage > max_stashed_blocks:
max_stashed_blocks = temp_stash_usage
self.touched_src_ranges = self.touched_src_ranges.union(
xf.src_ranges)
out.append("%s %s %s %s\n" % (
xf.style,
self.HashBlocks(self.tgt, xf.tgt_ranges),
xf.tgt_ranges.to_string_raw(), src_str))
total += tgt_size
elif xf.style in ("bsdiff", "imgdiff"):
assert xf.tgt_ranges
assert xf.src_ranges
if self.version == 1:
out.append("%s %d %d %s %s\n" % (
xf.style, xf.patch_start, xf.patch_len,
xf.src_ranges.to_string_raw(), xf.tgt_ranges.to_string_raw()))
elif self.version == 2:
out.append("%s %d %d %s %s\n" % (
xf.style, xf.patch_start, xf.patch_len,
xf.tgt_ranges.to_string_raw(), src_str))
elif self.version >= 3:
# take into account automatic stashing of overlapping blocks
if xf.src_ranges.overlaps(xf.tgt_ranges):
temp_stash_usage = stashed_blocks + xf.src_ranges.size()
if temp_stash_usage > max_stashed_blocks:
max_stashed_blocks = temp_stash_usage
self.touched_src_ranges = self.touched_src_ranges.union(
xf.src_ranges)
out.append("%s %d %d %s %s %s %s\n" % (
xf.style,
xf.patch_start, xf.patch_len,
self.HashBlocks(self.src, xf.src_ranges),
self.HashBlocks(self.tgt, xf.tgt_ranges),
xf.tgt_ranges.to_string_raw(), src_str))
total += tgt_size
elif xf.style == "zero":
assert xf.tgt_ranges
to_zero = xf.tgt_ranges.subtract(xf.src_ranges)
assert WriteTransfersZero(out, to_zero) == to_zero.size()
total += to_zero.size()
else:
raise ValueError("unknown transfer style '%s'\n" % xf.style)
if free_string:
out.append("".join(free_string))
stashed_blocks -= free_size
if self.version >= 2 and common.OPTIONS.cache_size is not None:
# Sanity check: abort if we're going to need more stash space than
# the allowed size (cache_size * threshold). There are two purposes
# of having a threshold here. a) Part of the cache may have been
# occupied by some recovery logs. b) It will buy us some time to deal
# with the oversize issue.
cache_size = common.OPTIONS.cache_size
stash_threshold = common.OPTIONS.stash_threshold
max_allowed = cache_size * stash_threshold
assert max_stashed_blocks * self.tgt.blocksize < max_allowed, \
'Stash size %d (%d * %d) exceeds the limit %d (%d * %.2f)' % (
max_stashed_blocks * self.tgt.blocksize, max_stashed_blocks,
self.tgt.blocksize, max_allowed, cache_size,
stash_threshold)
if self.version >= 3:
self.touched_src_sha1 = self.HashBlocks(
self.src, self.touched_src_ranges)
# Zero out extended blocks as a workaround for bug 20881595.
if self.tgt.extended.size() > 0:
assert (WriteTransfersZero(out, self.tgt.extended) ==
self.tgt.extended.size())
total += self.tgt.extended.size()
# We erase all the blocks on the partition that a) don't contain useful
# data in the new image; b) will not be touched by dm-verity. Out of those
# blocks, we erase the ones that won't be used in this update at the
# beginning of an update. The rest would be erased at the end. This is to
# work around the eMMC issue observed on some devices, which may otherwise
# get starving for clean blocks and thus fail the update. (b/28347095)
all_tgt = RangeSet(data=(0, self.tgt.total_blocks))
all_tgt_minus_extended = all_tgt.subtract(self.tgt.extended)
new_dontcare = all_tgt_minus_extended.subtract(self.tgt.care_map)
erase_first = new_dontcare.subtract(self.touched_src_ranges)
if erase_first.size() > 0:
out.insert(0, "erase %s\n" % (erase_first.to_string_raw(),))
erase_last = new_dontcare.subtract(erase_first)
if erase_last.size() > 0:
out.append("erase %s\n" % (erase_last.to_string_raw(),))
out.insert(0, "%d\n" % (self.version,)) # format version number
out.insert(1, "%d\n" % (total,))
if self.version >= 2:
# version 2 only: after the total block count, we give the number
# of stash slots needed, and the maximum size needed (in blocks)
out.insert(2, str(next_stash_id) + "\n")
out.insert(3, str(max_stashed_blocks) + "\n")
with open(prefix + ".transfer.list", "wb") as f:
for i in out:
f.write(i.encode('UTF-8'))
if self.version >= 2:
self._max_stashed_size = max_stashed_blocks * self.tgt.blocksize
OPTIONS = common.OPTIONS
if OPTIONS.cache_size is not None:
max_allowed = OPTIONS.cache_size * OPTIONS.stash_threshold
print("max stashed blocks: %d (%d bytes), "
"limit: %d bytes (%.2f%%)\n" % (
max_stashed_blocks, self._max_stashed_size, max_allowed,
self._max_stashed_size * 100.0 / max_allowed))
else:
print("max stashed blocks: %d (%d bytes), limit: <unknown>\n" % (
max_stashed_blocks, self._max_stashed_size))
def test_str(self):
self.assertEqual(str(RangeSet("0-9")), "0-9")
self.assertEqual(str(RangeSet("2-10 12")), "2-10 12")
self.assertEqual(str(RangeSet("11 2-10 12 1 0")), "0-12")
self.assertEqual(str(RangeSet("")), "empty")
class BlockImageDiff(object):
def __init__(self, tgt, src=None, version=4, threads=None,
disable_imgdiff=False):
if threads is None:
threads = multiprocessing.cpu_count() // 2
if threads == 0:
threads = 1
self.threads = threads
self.version = version
self.transfers = []
self.src_basenames = {}
self.src_numpatterns = {}
self._max_stashed_size = 0
self.touched_src_ranges = RangeSet()
self.touched_src_sha1 = None
self.disable_imgdiff = disable_imgdiff
assert version in (1, 2, 3, 4)
self.tgt = tgt
if src is None:
src = EmptyImage()
self.src = src
# The updater code that installs the patch always uses 4k blocks.
assert tgt.blocksize == 4096
assert src.blocksize == 4096
# The range sets in each filemap should comprise a partition of
# the care map.
self.AssertPartition(src.care_map, src.file_map.values())
self.AssertPartition(tgt.care_map, tgt.file_map.values())
@property
def max_stashed_size(self):
return self._max_stashed_size
def Compute(self, prefix):
# When looking for a source file to use as the diff input for a
# target file, we try:
# 1) an exact path match if available, otherwise
# 2) a exact basename match if available, otherwise
# 3) a basename match after all runs of digits are replaced by
# "#" if available, otherwise
# 4) we have no source for this target.
self.AbbreviateSourceNames()
self.FindTransfers()
# Find the ordering dependencies among transfers (this is O(n^2)
# in the number of transfers).
self.GenerateDigraph()
# Find a sequence of transfers that satisfies as many ordering
# dependencies as possible (heuristically).
self.FindVertexSequence()
# Fix up the ordering dependencies that the sequence didn't
# satisfy.
if self.version == 1:
self.RemoveBackwardEdges()
else:
self.ReverseBackwardEdges()
self.ImproveVertexSequence()
# Ensure the runtime stash size is under the limit.
if self.version >= 2 and common.OPTIONS.cache_size is not None:
self.ReviseStashSize()
# Double-check our work.
self.AssertSequenceGood()
self.ComputePatches(prefix)
self.WriteTransfers(prefix)
def HashBlocks(self, source, ranges): # pylint: disable=no-self-use
data = source.ReadRangeSet(ranges)
ctx = sha1()
for p in data:
ctx.update(p)
return ctx.hexdigest()
def WriteTransfers(self, prefix):
def WriteTransfersZero(out, to_zero):
"""Limit the number of blocks in command zero to 1024 blocks.
This prevents the target size of one command from being too large; and
might help to avoid fsync errors on some devices."""
zero_blocks_limit = 1024
total = 0
while to_zero.size() > 0:
zero_blocks = to_zero.first(zero_blocks_limit)
out.append("zero %s\n" % (zero_blocks.to_string_raw(),))
total += zero_blocks.size()
to_zero = to_zero.subtract(zero_blocks)
return total
out = []
total = 0
stashes = {}
stashed_blocks = 0
max_stashed_blocks = 0
free_stash_ids = []
next_stash_id = 0
for xf in self.transfers:
if self.version < 2:
assert not xf.stash_before
assert not xf.use_stash
for s, sr in xf.stash_before:
assert s not in stashes
if free_stash_ids:
sid = heapq.heappop(free_stash_ids)
else:
sid = next_stash_id
next_stash_id += 1
stashes[s] = sid
if self.version == 2:
stashed_blocks += sr.size()
out.append("stash %d %s\n" % (sid, sr.to_string_raw()))
else:
sh = self.HashBlocks(self.src, sr)
if sh in stashes:
stashes[sh] += 1
else:
stashes[sh] = 1
stashed_blocks += sr.size()
self.touched_src_ranges = self.touched_src_ranges.union(sr)
out.append("stash %s %s\n" % (sh, sr.to_string_raw()))
if stashed_blocks > max_stashed_blocks:
max_stashed_blocks = stashed_blocks
free_string = []
free_size = 0
if self.version == 1:
src_str = xf.src_ranges.to_string_raw() if xf.src_ranges else ""
elif self.version >= 2:
# <# blocks> <src ranges>
# OR
# <# blocks> <src ranges> <src locs> <stash refs...>
# OR
# <# blocks> - <stash refs...>
size = xf.src_ranges.size()
src_str = [str(size)]
unstashed_src_ranges = xf.src_ranges
mapped_stashes = []
for s, sr in xf.use_stash:
sid = stashes.pop(s)
unstashed_src_ranges = unstashed_src_ranges.subtract(sr)
sh = self.HashBlocks(self.src, sr)
sr = xf.src_ranges.map_within(sr)
mapped_stashes.append(sr)
if self.version == 2:
src_str.append("%d:%s" % (sid, sr.to_string_raw()))
# A stash will be used only once. We need to free the stash
# immediately after the use, instead of waiting for the automatic
# clean-up at the end. Because otherwise it may take up extra space
# and lead to OTA failures.
# Bug: 23119955
free_string.append("free %d\n" % (sid,))
free_size += sr.size()
else:
assert sh in stashes
src_str.append("%s:%s" % (sh, sr.to_string_raw()))
stashes[sh] -= 1
if stashes[sh] == 0:
free_size += sr.size()
free_string.append("free %s\n" % (sh))
stashes.pop(sh)
heapq.heappush(free_stash_ids, sid)
if unstashed_src_ranges is None:
src_str.insert(1, unstashed_src_ranges.to_string_raw())
if xf.use_stash:
mapped_unstashed = xf.src_ranges.map_within(unstashed_src_ranges)
src_str.insert(2, mapped_unstashed.to_string_raw())
mapped_stashes.append(mapped_unstashed)
self.AssertPartition(RangeSet(data=(0, size)), mapped_stashes)
else:
src_str.insert(1, "-")
self.AssertPartition(RangeSet(data=(0, size)), mapped_stashes)
src_str = " ".join(src_str)
# all versions:
# zero <rangeset>
# new <rangeset>
# erase <rangeset>
#
# version 1:
# bsdiff patchstart patchlen <src rangeset> <tgt rangeset>
# imgdiff patchstart patchlen <src rangeset> <tgt rangeset>
# move <src rangeset> <tgt rangeset>
#
# version 2:
# bsdiff patchstart patchlen <tgt rangeset> <src_str>
# imgdiff patchstart patchlen <tgt rangeset> <src_str>
# move <tgt rangeset> <src_str>
#
# version 3:
# bsdiff patchstart patchlen srchash tgthash <tgt rangeset> <src_str>
# imgdiff patchstart patchlen srchash tgthash <tgt rangeset> <src_str>
# move hash <tgt rangeset> <src_str>
tgt_size = xf.tgt_ranges.size()
if xf.style == "new":
assert xf.tgt_ranges
out.append("%s %s\n" % (xf.style, xf.tgt_ranges.to_string_raw()))
total += tgt_size
elif xf.style == "move":
assert xf.tgt_ranges
assert xf.src_ranges.size() == tgt_size
if xf.src_ranges != xf.tgt_ranges:
if self.version == 1:
out.append("%s %s %s\n" % (
xf.style,
xf.src_ranges.to_string_raw(), xf.tgt_ranges.to_string_raw()))
elif self.version == 2:
out.append("%s %s %s\n" % (
xf.style,
xf.tgt_ranges.to_string_raw(), src_str))
elif self.version >= 3:
# take into account automatic stashing of overlapping blocks
if xf.src_ranges.overlaps(xf.tgt_ranges):
temp_stash_usage = stashed_blocks + xf.src_ranges.size()
if temp_stash_usage > max_stashed_blocks:
max_stashed_blocks = temp_stash_usage
self.touched_src_ranges = self.touched_src_ranges.union(
xf.src_ranges)
out.append("%s %s %s %s\n" % (
xf.style,
self.HashBlocks(self.tgt, xf.tgt_ranges),
xf.tgt_ranges.to_string_raw(), src_str))
total += tgt_size
elif xf.style in ("bsdiff", "imgdiff"):
assert xf.tgt_ranges
assert xf.src_ranges
if self.version == 1:
out.append("%s %d %d %s %s\n" % (
xf.style, xf.patch_start, xf.patch_len,
xf.src_ranges.to_string_raw(), xf.tgt_ranges.to_string_raw()))
elif self.version == 2:
out.append("%s %d %d %s %s\n" % (
xf.style, xf.patch_start, xf.patch_len,
xf.tgt_ranges.to_string_raw(), src_str))
elif self.version >= 3:
# take into account automatic stashing of overlapping blocks
if xf.src_ranges.overlaps(xf.tgt_ranges):
temp_stash_usage = stashed_blocks + xf.src_ranges.size()
if temp_stash_usage > max_stashed_blocks:
max_stashed_blocks = temp_stash_usage
self.touched_src_ranges = self.touched_src_ranges.union(
xf.src_ranges)
out.append("%s %d %d %s %s %s %s\n" % (
xf.style,
xf.patch_start, xf.patch_len,
self.HashBlocks(self.src, xf.src_ranges),
self.HashBlocks(self.tgt, xf.tgt_ranges),
xf.tgt_ranges.to_string_raw(), src_str))
total += tgt_size
elif xf.style == "zero":
assert xf.tgt_ranges
to_zero = xf.tgt_ranges.subtract(xf.src_ranges)
assert WriteTransfersZero(out, to_zero) == to_zero.size()
total += to_zero.size()
else:
raise ValueError("unknown transfer style '%s'\n" % xf.style)
if free_string:
out.append("".join(free_string))
stashed_blocks -= free_size
if self.version >= 2 and common.OPTIONS.cache_size is not None:
# Sanity check: abort if we're going to need more stash space than
# the allowed size (cache_size * threshold). There are two purposes
# of having a threshold here. a) Part of the cache may have been
# occupied by some recovery logs. b) It will buy us some time to deal
# with the oversize issue.
cache_size = common.OPTIONS.cache_size
stash_threshold = common.OPTIONS.stash_threshold
max_allowed = cache_size * stash_threshold
assert max_stashed_blocks * self.tgt.blocksize < max_allowed, \
'Stash size %d (%d * %d) exceeds the limit %d (%d * %.2f)' % (
max_stashed_blocks * self.tgt.blocksize, max_stashed_blocks,
self.tgt.blocksize, max_allowed, cache_size,
stash_threshold)
if self.version >= 3:
self.touched_src_sha1 = self.HashBlocks(
self.src, self.touched_src_ranges)
# Zero out extended blocks as a workaround for bug 20881595.
if self.tgt.extended.size() > 0:
assert (WriteTransfersZero(out, self.tgt.extended) ==
self.tgt.extended.size())
total += self.tgt.extended.size()
# We erase all the blocks on the partition that a) don't contain useful
# data in the new image; b) will not be touched by dm-verity. Out of those
# blocks, we erase the ones that won't be used in this update at the
# beginning of an update. The rest would be erased at the end. This is to
# work around the eMMC issue observed on some devices, which may otherwise
# get starving for clean blocks and thus fail the update. (b/28347095)
all_tgt = RangeSet(data=(0, self.tgt.total_blocks))
all_tgt_minus_extended = all_tgt.subtract(self.tgt.extended)
new_dontcare = all_tgt_minus_extended.subtract(self.tgt.care_map)
erase_first = new_dontcare.subtract(self.touched_src_ranges)
if erase_first.size() > 0:
out.insert(0, "erase %s\n" % (erase_first.to_string_raw(),))
erase_last = new_dontcare.subtract(erase_first)
if erase_last.size() > 0:
out.append("erase %s\n" % (erase_last.to_string_raw(),))
out.insert(0, "%d\n" % (self.version,)) # format version number
out.insert(1, "%d\n" % (total,))
if self.version >= 2:
# version 2 only: after the total block count, we give the number
# of stash slots needed, and the maximum size needed (in blocks)
out.insert(2, str(next_stash_id) + "\n")
out.insert(3, str(max_stashed_blocks) + "\n")
with open(prefix + ".transfer.list", "wb") as f:
for i in out:
f.write(i.encode('UTF-8'))
if self.version >= 2:
self._max_stashed_size = max_stashed_blocks * self.tgt.blocksize
OPTIONS = common.OPTIONS
if OPTIONS.cache_size is not None:
max_allowed = OPTIONS.cache_size * OPTIONS.stash_threshold
print("max stashed blocks: %d (%d bytes), "
"limit: %d bytes (%.2f%%)\n" % (
max_stashed_blocks, self._max_stashed_size, max_allowed,
self._max_stashed_size * 100.0 / max_allowed))
else:
print("max stashed blocks: %d (%d bytes), limit: <unknown>\n" % (
max_stashed_blocks, self._max_stashed_size))
def ReviseStashSize(self):
print("Revising stash size...")
stashes = {}
# Create the map between a stash and its def/use points. For example, for a
# given stash of (idx, sr), stashes[idx] = (sr, def_cmd, use_cmd).
for xf in self.transfers:
# Command xf defines (stores) all the stashes in stash_before.
for idx, sr in xf.stash_before:
stashes[idx] = (sr, xf)
# Record all the stashes command xf uses.
for idx, _ in xf.use_stash:
stashes[idx] += (xf,)
# Compute the maximum blocks available for stash based on /cache size and
# the threshold.
cache_size = common.OPTIONS.cache_size
stash_threshold = common.OPTIONS.stash_threshold
max_allowed = cache_size * stash_threshold / self.tgt.blocksize
stashed_blocks = 0
new_blocks = 0
# Now go through all the commands. Compute the required stash size on the
# fly. If a command requires excess stash than available, it deletes the
# stash by replacing the command that uses the stash with a "new" command
# instead.
for xf in self.transfers:
replaced_cmds = []
# xf.stash_before generates explicit stash commands.
for idx, sr in xf.stash_before:
if stashed_blocks + sr.size() > max_allowed:
# We cannot stash this one for a later command. Find out the command
# that will use this stash and replace the command with "new".
use_cmd = stashes[idx][2]
replaced_cmds.append(use_cmd)
print("%10d %9s %s" % (sr.size(), "explicit", use_cmd))
else:
stashed_blocks += sr.size()
# xf.use_stash generates free commands.
for _, sr in xf.use_stash:
stashed_blocks -= sr.size()
# "move" and "diff" may introduce implicit stashes in BBOTA v3. Prior to
# ComputePatches(), they both have the style of "diff".
if xf.style == "diff" and self.version >= 3:
assert xf.tgt_ranges and xf.src_ranges
if xf.src_ranges.overlaps(xf.tgt_ranges):
if stashed_blocks + xf.src_ranges.size() > max_allowed:
replaced_cmds.append(xf)
print("%10d %9s %s" % (xf.src_ranges.size(), "implicit", xf))
# Replace the commands in replaced_cmds with "new"s.
for cmd in replaced_cmds:
# It no longer uses any commands in "use_stash". Remove the def points
# for all those stashes.
for idx, sr in cmd.use_stash:
def_cmd = stashes[idx][1]
assert (idx, sr) in def_cmd.stash_before
def_cmd.stash_before.remove((idx, sr))
# Add up blocks that violates space limit and print total number to
# screen later.
new_blocks += cmd.tgt_ranges.size()
cmd.ConvertToNew()
num_of_bytes = new_blocks * self.tgt.blocksize
print(" Total %d blocks (%d bytes) are packed as new blocks due to "
"insufficient cache size." % (new_blocks, num_of_bytes))
def ComputePatches(self, prefix):
print("Reticulating splines...")
diff_q = []
patch_num = 0
with open(prefix + ".new.dat", "wb") as new_f:
for xf in self.transfers:
if xf.style == "zero":
pass
elif xf.style == "new":
for piece in self.tgt.ReadRangeSet(xf.tgt_ranges):
new_f.write(piece)
elif xf.style == "diff":
src = self.src.ReadRangeSet(xf.src_ranges)
tgt = self.tgt.ReadRangeSet(xf.tgt_ranges)
# We can't compare src and tgt directly because they may have
# the same content but be broken up into blocks differently, eg:
#
# ["he", "llo"] vs ["h", "ello"]
#
# We want those to compare equal, ideally without having to
# actually concatenate the strings (these may be tens of
# megabytes).
src_sha1 = sha1()
for p in src:
src_sha1.update(p)
tgt_sha1 = sha1()
tgt_size = 0
for p in tgt:
tgt_sha1.update(p)
tgt_size += len(p)
if src_sha1.digest() == tgt_sha1.digest():
# These are identical; we don't need to generate a patch,
# just issue copy commands on the device.
xf.style = "move"
else:
# For files in zip format (eg, APKs, JARs, etc.) we would
# like to use imgdiff -z if possible (because it usually
# produces significantly smaller patches than bsdiff).
# This is permissible if:
#
# - imgdiff is not disabled, and
# - the source and target files are monotonic (ie, the
# data is stored with blocks in increasing order), and
# - we haven't removed any blocks from the source set.
#
# If these conditions are satisfied then appending all the
# blocks in the set together in order will produce a valid
# zip file (plus possibly extra zeros in the last block),
# which is what imgdiff needs to operate. (imgdiff is
# fine with extra zeros at the end of the file.)
imgdiff = (not self.disable_imgdiff and xf.intact and
xf.tgt_name.split(".")[-1].lower()
in ("apk", "jar", "zip"))
xf.style = "imgdiff" if imgdiff else "bsdiff"
diff_q.append((tgt_size, src, tgt, xf, patch_num))
patch_num += 1
else:
assert False, "unknown style " + xf.style
if diff_q:
if self.threads > 1:
print("Computing patches (using %d threads)..." % (self.threads,))
else:
print("Computing patches...")
diff_q.sort()
patches = [None] * patch_num
# TODO: Rewrite with multiprocessing.ThreadPool?
lock = threading.Lock()
def diff_worker():
while True:
with lock:
if not diff_q:
return
tgt_size, src, tgt, xf, patchnum = diff_q.pop()
patch = compute_patch(src, tgt, imgdiff=(xf.style == "imgdiff"))
size = len(patch)
with lock:
patches[patchnum] = (patch, xf)
print("%10d %10d (%6.2f%%) %7s %s" % (
size, tgt_size, size * 100.0 / tgt_size, xf.style,
xf.tgt_name if xf.tgt_name == xf.src_name else (
xf.tgt_name + " (from " + xf.src_name + ")")))
threads = [threading.Thread(target=diff_worker)
for _ in range(self.threads)]
for th in threads:
th.start()
while threads:
threads.pop().join()
else:
patches = []
p = 0
with open(prefix + ".patch.dat", "wb") as patch_f:
for patch, xf in patches:
xf.patch_start = p
xf.patch_len = len(patch)
patch_f.write(patch)
p += len(patch)
def AssertSequenceGood(self):
# Simulate the sequences of transfers we will output, and check that:
# - we never read a block after writing it, and
# - we write every block we care about exactly once.
# Start with no blocks having been touched yet.
touched = array.array("B", (0,) * self.tgt.total_blocks)
# Imagine processing the transfers in order.
for xf in self.transfers:
# Check that the input blocks for this transfer haven't yet been touched.
x = xf.src_ranges
if self.version >= 2:
for _, sr in xf.use_stash:
x = x.subtract(sr)
for s, e in x:
# Source image could be larger. Don't check the blocks that are in the
# source image only. Since they are not in 'touched', and won't ever
# be touched.
for i in range(s, min(e, self.tgt.total_blocks)):
assert touched[i] == 0
# Check that the output blocks for this transfer haven't yet
# been touched, and touch all the blocks written by this
# transfer.
for s, e in xf.tgt_ranges:
for i in range(s, e):
assert touched[i] == 0
touched[i] = 1
# Check that we've written every target block.
for s, e in self.tgt.care_map:
for i in range(s, e):
assert touched[i] == 1
def ImproveVertexSequence(self):
print("Improving vertex order...")
# At this point our digraph is acyclic; we reversed any edges that
# were backwards in the heuristically-generated sequence. The
# previously-generated order is still acceptable, but we hope to
# find a better order that needs less memory for stashed data.
# Now we do a topological sort to generate a new vertex order,
# using a greedy algorithm to choose which vertex goes next
# whenever we have a choice.
# Make a copy of the edge set; this copy will get destroyed by the
# algorithm.
for xf in self.transfers:
xf.incoming = xf.goes_after.copy()
xf.outgoing = xf.goes_before.copy()
L = [] # the new vertex order
# S is the set of sources in the remaining graph; we always choose
# the one that leaves the least amount of stashed data after it's
# executed.
S = [(u.NetStashChange(), u.order, u) for u in self.transfers
if not u.incoming]
heapq.heapify(S)
while S:
_, _, xf = heapq.heappop(S)
L.append(xf)
for u in xf.outgoing:
del u.incoming[xf]
if not u.incoming:
heapq.heappush(S, (u.NetStashChange(), u.order, u))
# if this fails then our graph had a cycle.
assert len(L) == len(self.transfers)
self.transfers = L
for i, xf in enumerate(L):
xf.order = i
def RemoveBackwardEdges(self):
print("Removing backward edges...")
in_order = 0
out_of_order = 0
lost_source = 0
for xf in self.transfers:
lost = 0
size = xf.src_ranges.size()
for u in xf.goes_before:
# xf should go before u
if xf.order < u.order:
# it does, hurray!
in_order += 1
else:
# it doesn't, boo. trim the blocks that u writes from xf's
# source, so that xf can go after u.
out_of_order += 1
assert xf.src_ranges.overlaps(u.tgt_ranges)
xf.src_ranges = xf.src_ranges.subtract(u.tgt_ranges)
xf.intact = False
if xf.style == "diff" and not xf.src_ranges:
# nothing left to diff from; treat as new data
xf.style = "new"
lost = size - xf.src_ranges.size()
lost_source += lost
print((" %d/%d dependencies (%.2f%%) were violated; "
"%d source blocks removed.") %
(out_of_order, in_order + out_of_order,
(out_of_order * 100.0 / (in_order + out_of_order))
if (in_order + out_of_order) else 0.0,
lost_source))
def ReverseBackwardEdges(self):
print("Reversing backward edges...")
in_order = 0
out_of_order = 0
stashes = 0
stash_size = 0
for xf in self.transfers:
for u in xf.goes_before.copy():
# xf should go before u
if xf.order < u.order:
# it does, hurray!
in_order += 1
else:
# it doesn't, boo. modify u to stash the blocks that it
# writes that xf wants to read, and then require u to go
# before xf.
out_of_order += 1
overlap = xf.src_ranges.intersect(u.tgt_ranges)
assert overlap
u.stash_before.append((stashes, overlap))
xf.use_stash.append((stashes, overlap))
stashes += 1
stash_size += overlap.size()
# reverse the edge direction; now xf must go after u
del xf.goes_before[u]
del u.goes_after[xf]
xf.goes_after[u] = None # value doesn't matter
u.goes_before[xf] = None
print((" %d/%d dependencies (%.2f%%) were violated; "
"%d source blocks stashed.") %
(out_of_order, in_order + out_of_order,
(out_of_order * 100.0 / (in_order + out_of_order))
if (in_order + out_of_order) else 0.0,
stash_size))
def FindVertexSequence(self):
print("Finding vertex sequence...")
# This is based on "A Fast & Effective Heuristic for the Feedback
# Arc Set Problem" by P. Eades, X. Lin, and W.F. Smyth. Think of
# it as starting with the digraph G and moving all the vertices to
# be on a horizontal line in some order, trying to minimize the
# number of edges that end up pointing to the left. Left-pointing
# edges will get removed to turn the digraph into a DAG. In this
# case each edge has a weight which is the number of source blocks
# we'll lose if that edge is removed; we try to minimize the total
# weight rather than just the number of edges.
# Make a copy of the edge set; this copy will get destroyed by the
# algorithm.
for xf in self.transfers:
xf.incoming = xf.goes_after.copy()
xf.outgoing = xf.goes_before.copy()
xf.score = sum(xf.outgoing.values()) - sum(xf.incoming.values())
# We use an OrderedDict instead of just a set so that the output
# is repeatable; otherwise it would depend on the hash values of
# the transfer objects.
G = OrderedDict()
for xf in self.transfers:
G[xf] = None
s1 = deque() # the left side of the sequence, built from left to right
s2 = deque() # the right side of the sequence, built from right to left
heap = []
for xf in self.transfers:
xf.heap_item = HeapItem(xf)
heap.append(xf.heap_item)
heapq.heapify(heap)
sinks = set(u for u in G if not u.outgoing)
sources = set(u for u in G if not u.incoming)
def adjust_score(iu, delta):
iu.score += delta
iu.heap_item.clear()
iu.heap_item = HeapItem(iu)
heapq.heappush(heap, iu.heap_item)
while G:
# Put all sinks at the end of the sequence.
while sinks:
new_sinks = set()
for u in sinks:
if u not in G: continue
s2.appendleft(u)
del G[u]
for iu in u.incoming:
adjust_score(iu, -iu.outgoing.pop(u))
if not iu.outgoing: new_sinks.add(iu)
sinks = new_sinks
# Put all the sources at the beginning of the sequence.
while sources:
new_sources = set()
for u in sources:
if u not in G: continue
s1.append(u)
del G[u]
for iu in u.outgoing:
adjust_score(iu, +iu.incoming.pop(u))
if not iu.incoming: new_sources.add(iu)
sources = new_sources
if not G: break
# Find the "best" vertex to put next. "Best" is the one that
# maximizes the net difference in source blocks saved we get by
# pretending it's a source rather than a sink.
while True:
u = heapq.heappop(heap)
if u and u.item in G:
u = u.item
break
s1.append(u)
del G[u]
for iu in u.outgoing:
adjust_score(iu, +iu.incoming.pop(u))
if not iu.incoming: sources.add(iu)
for iu in u.incoming:
adjust_score(iu, -iu.outgoing.pop(u))
if not iu.outgoing: sinks.add(iu)
# Now record the sequence in the 'order' field of each transfer,
# and by rearranging self.transfers to be in the chosen sequence.
new_transfers = []
for x in itertools.chain(s1, s2):
x.order = len(new_transfers)
new_transfers.append(x)
del x.incoming
del x.outgoing
self.transfers = new_transfers
def GenerateDigraph(self):
print("Generating digraph...")
# Each item of source_ranges will be:
# - None, if that block is not used as a source,
# - a transfer, if one transfer uses it as a source, or
# - a set of transfers.
source_ranges = []
for b in self.transfers:
for s, e in b.src_ranges:
if e > len(source_ranges):
source_ranges.extend([None] * (e - len(source_ranges)))
for i in range(s, e):
if source_ranges[i] is None:
source_ranges[i] = b
else:
if not isinstance(source_ranges[i], set):
source_ranges[i] = set([source_ranges[i]])
source_ranges[i].add(b)
for a in self.transfers:
intersections = set()
for s, e in a.tgt_ranges:
for i in range(s, e):
if i >= len(source_ranges): break
b = source_ranges[i]
if b is not None:
if isinstance(b, set):
intersections.update(b)
else:
intersections.add(b)
for b in intersections:
if a is b: continue
# If the blocks written by A are read by B, then B needs to go before A.
i = a.tgt_ranges.intersect(b.src_ranges)
if i:
if b.src_name == "__ZERO":
# the cost of removing source blocks for the __ZERO domain
# is (nearly) zero.
size = 0
else:
size = i.size()
b.goes_before[a] = size
a.goes_after[b] = size
def FindTransfers(self):
"""Parse the file_map to generate all the transfers."""
def AddTransfer(tgt_name, src_name, tgt_ranges, src_ranges, style, by_id,
split=False):
"""Wrapper function for adding a Transfer().
For BBOTA v3, we need to stash source blocks for resumable feature.
However, with the growth of file size and the shrink of the cache
partition source blocks are too large to be stashed. If a file occupies
too many blocks (greater than MAX_BLOCKS_PER_DIFF_TRANSFER), we split it
into smaller pieces by getting multiple Transfer()s.
The downside is that after splitting, we may increase the package size
since the split pieces don't align well. According to our experiments,
1/8 of the cache size as the per-piece limit appears to be optimal.
Compared to the fixed 1024-block limit, it reduces the overall package
size by 30% volantis, and 20% for angler and bullhead."""
# We care about diff transfers only.
if style != "diff" or not split:
Transfer(tgt_name, src_name, tgt_ranges, src_ranges, style, by_id)
return
pieces = 0
cache_size = common.OPTIONS.cache_size
split_threshold = 0.125
max_blocks_per_transfer = int(cache_size * split_threshold /
self.tgt.blocksize)
# Change nothing for small files.
if (tgt_ranges.size() <= max_blocks_per_transfer and
src_ranges.size() <= max_blocks_per_transfer):
Transfer(tgt_name, src_name, tgt_ranges, src_ranges, style, by_id)
return
while (tgt_ranges.size() > max_blocks_per_transfer and
src_ranges.size() > max_blocks_per_transfer):
tgt_split_name = "%s-%d" % (tgt_name, pieces)
src_split_name = "%s-%d" % (src_name, pieces)
tgt_first = tgt_ranges.first(max_blocks_per_transfer)
src_first = src_ranges.first(max_blocks_per_transfer)
Transfer(tgt_split_name, src_split_name, tgt_first, src_first, style,
by_id)
tgt_ranges = tgt_ranges.subtract(tgt_first)
src_ranges = src_ranges.subtract(src_first)
pieces += 1
# Handle remaining blocks.
if tgt_ranges.size() or src_ranges.size():
# Must be both non-empty.
assert tgt_ranges.size() and src_ranges.size()
tgt_split_name = "%s-%d" % (tgt_name, pieces)
src_split_name = "%s-%d" % (src_name, pieces)
Transfer(tgt_split_name, src_split_name, tgt_ranges, src_ranges, style,
by_id)
empty = RangeSet()
for tgt_fn, tgt_ranges in self.tgt.file_map.items():
if tgt_fn == "__ZERO":
# the special "__ZERO" domain is all the blocks not contained
# in any file and that are filled with zeros. We have a
# special transfer style for zero blocks.
src_ranges = self.src.file_map.get("__ZERO", empty)
AddTransfer(tgt_fn, "__ZERO", tgt_ranges, src_ranges,
"zero", self.transfers)
continue
elif tgt_fn == "__COPY":
# "__COPY" domain includes all the blocks not contained in any
# file and that need to be copied unconditionally to the target.
AddTransfer(tgt_fn, None, tgt_ranges, empty, "new", self.transfers)
continue
elif tgt_fn in self.src.file_map:
# Look for an exact pathname match in the source.
AddTransfer(tgt_fn, tgt_fn, tgt_ranges, self.src.file_map[tgt_fn],
"diff", self.transfers, self.version >= 3)
continue
b = os.path.basename(tgt_fn)
if b in self.src_basenames:
# Look for an exact basename match in the source.
src_fn = self.src_basenames[b]
AddTransfer(tgt_fn, src_fn, tgt_ranges, self.src.file_map[src_fn],
"diff", self.transfers, self.version >= 3)
continue
b = re.sub("[0-9]+", "#", b)
if b in self.src_numpatterns:
# Look for a 'number pattern' match (a basename match after
# all runs of digits are replaced by "#"). (This is useful
# for .so files that contain version numbers in the filename
# that get bumped.)
src_fn = self.src_numpatterns[b]
AddTransfer(tgt_fn, src_fn, tgt_ranges, self.src.file_map[src_fn],
"diff", self.transfers, self.version >= 3)
continue
AddTransfer(tgt_fn, None, tgt_ranges, empty, "new", self.transfers)
def AbbreviateSourceNames(self):
for k in self.src.file_map.keys():
b = os.path.basename(k)
self.src_basenames[b] = k
b = re.sub("[0-9]+", "#", b)
self.src_numpatterns[b] = k
@staticmethod
def AssertPartition(total, seq):
"""Assert that all the RangeSets in 'seq' form a partition of the
'total' RangeSet (ie, they are nonintersecting and their union
equals 'total')."""
so_far = RangeSet()
for i in seq:
assert not so_far.overlaps(i)
so_far = so_far.union(i)
assert so_far == total
def ConvertToNew(self):
assert self.style != "new"
self.use_stash = []
self.style = "new"
self.src_ranges = RangeSet()
def FindTransfers(self):
"""Parse the file_map to generate all the transfers."""
def AddTransfer(tgt_name, src_name, tgt_ranges, src_ranges, style, by_id,
split=False):
"""Wrapper function for adding a Transfer().
For BBOTA v3, we need to stash source blocks for resumable feature.
However, with the growth of file size and the shrink of the cache
partition source blocks are too large to be stashed. If a file occupies
too many blocks (greater than MAX_BLOCKS_PER_DIFF_TRANSFER), we split it
into smaller pieces by getting multiple Transfer()s.
The downside is that after splitting, we may increase the package size
since the split pieces don't align well. According to our experiments,
1/8 of the cache size as the per-piece limit appears to be optimal.
Compared to the fixed 1024-block limit, it reduces the overall package
size by 30% volantis, and 20% for angler and bullhead."""
# We care about diff transfers only.
if style != "diff" or not split:
Transfer(tgt_name, src_name, tgt_ranges, src_ranges, style, by_id)
return
pieces = 0
cache_size = common.OPTIONS.cache_size
split_threshold = 0.125
max_blocks_per_transfer = int(cache_size * split_threshold /
self.tgt.blocksize)
# Change nothing for small files.
if (tgt_ranges.size() <= max_blocks_per_transfer and
src_ranges.size() <= max_blocks_per_transfer):
Transfer(tgt_name, src_name, tgt_ranges, src_ranges, style, by_id)
return
while (tgt_ranges.size() > max_blocks_per_transfer and
src_ranges.size() > max_blocks_per_transfer):
tgt_split_name = "%s-%d" % (tgt_name, pieces)
src_split_name = "%s-%d" % (src_name, pieces)
tgt_first = tgt_ranges.first(max_blocks_per_transfer)
src_first = src_ranges.first(max_blocks_per_transfer)
Transfer(tgt_split_name, src_split_name, tgt_first, src_first, style,
by_id)
tgt_ranges = tgt_ranges.subtract(tgt_first)
src_ranges = src_ranges.subtract(src_first)
pieces += 1
# Handle remaining blocks.
if tgt_ranges.size() or src_ranges.size():
# Must be both non-empty.
assert tgt_ranges.size() and src_ranges.size()
tgt_split_name = "%s-%d" % (tgt_name, pieces)
src_split_name = "%s-%d" % (src_name, pieces)
Transfer(tgt_split_name, src_split_name, tgt_ranges, src_ranges, style,
by_id)
empty = RangeSet()
for tgt_fn, tgt_ranges in self.tgt.file_map.items():
if tgt_fn == "__ZERO":
# the special "__ZERO" domain is all the blocks not contained
# in any file and that are filled with zeros. We have a
# special transfer style for zero blocks.
src_ranges = self.src.file_map.get("__ZERO", empty)
AddTransfer(tgt_fn, "__ZERO", tgt_ranges, src_ranges,
"zero", self.transfers)
continue
elif tgt_fn == "__COPY":
# "__COPY" domain includes all the blocks not contained in any
# file and that need to be copied unconditionally to the target.
AddTransfer(tgt_fn, None, tgt_ranges, empty, "new", self.transfers)
continue
elif tgt_fn in self.src.file_map:
# Look for an exact pathname match in the source.
AddTransfer(tgt_fn, tgt_fn, tgt_ranges, self.src.file_map[tgt_fn],
"diff", self.transfers, self.version >= 3)
continue
b = os.path.basename(tgt_fn)
if b in self.src_basenames:
# Look for an exact basename match in the source.
src_fn = self.src_basenames[b]
AddTransfer(tgt_fn, src_fn, tgt_ranges, self.src.file_map[src_fn],
"diff", self.transfers, self.version >= 3)
continue
b = re.sub("[0-9]+", "#", b)
if b in self.src_numpatterns:
# Look for a 'number pattern' match (a basename match after
# all runs of digits are replaced by "#"). (This is useful
# for .so files that contain version numbers in the filename
# that get bumped.)
src_fn = self.src_numpatterns[b]
AddTransfer(tgt_fn, src_fn, tgt_ranges, self.src.file_map[src_fn],
"diff", self.transfers, self.version >= 3)
continue
AddTransfer(tgt_fn, None, tgt_ranges, empty, "new", self.transfers)
def test_intersect(self):
self.assertEqual(RangeSet("10-19 30-34").intersect(RangeSet("18-32")),
RangeSet("18-19 30-32"))
self.assertEqual(RangeSet("10-19 30-34").intersect(RangeSet("22-28")),
RangeSet(""))
def AddTransfer(tgt_name, src_name, tgt_ranges, src_ranges, style, by_id,
split=False):
"""Wrapper function for adding a Transfer()."""
# We specialize diff transfers only (which covers bsdiff/imgdiff/move);
# otherwise add the Transfer() as is.
if style != "diff" or not split:
Transfer(tgt_name, src_name, tgt_ranges, src_ranges, style, by_id)
return
# Handle .odex files specially to analyze the block-wise difference. If
# most of the blocks are identical with only few changes (e.g. header),
# we will patch the changed blocks only. This avoids stashing unchanged
# blocks while patching. We limit the analysis to files without size
# changes only. This is to avoid sacrificing the OTA generation cost too
# much.
if (tgt_name.split(".")[-1].lower() == 'odex' and
tgt_ranges.size() == src_ranges.size()):
# 0.5 threshold can be further tuned. The tradeoff is: if only very
# few blocks remain identical, we lose the opportunity to use imgdiff
# that may have better compression ratio than bsdiff.
crop_threshold = 0.5
tgt_skipped = RangeSet()
src_skipped = RangeSet()
tgt_size = tgt_ranges.size()
tgt_changed = 0
for src_block, tgt_block in zip(src_ranges.next_item(),
tgt_ranges.next_item()):
src_rs = RangeSet(str(src_block))
tgt_rs = RangeSet(str(tgt_block))
if self.src.ReadRangeSet(src_rs) == self.tgt.ReadRangeSet(tgt_rs):
tgt_skipped = tgt_skipped.union(tgt_rs)
src_skipped = src_skipped.union(src_rs)
else:
tgt_changed += tgt_rs.size()
# Terminate early if no clear sign of benefits.
if tgt_changed > tgt_size * crop_threshold:
break
if tgt_changed < tgt_size * crop_threshold:
assert tgt_changed + tgt_skipped.size() == tgt_size
print('%10d %10d (%6.2f%%) %s' % (tgt_skipped.size(), tgt_size,
tgt_skipped.size() * 100.0 / tgt_size, tgt_name))
AddSplitTransfers(
"%s-skipped" % (tgt_name,),
"%s-skipped" % (src_name,),
tgt_skipped, src_skipped, style, by_id)
# Intentionally change the file extension to avoid being imgdiff'd as
# the files are no longer in their original format.
tgt_name = "%s-cropped" % (tgt_name,)
src_name = "%s-cropped" % (src_name,)
tgt_ranges = tgt_ranges.subtract(tgt_skipped)
src_ranges = src_ranges.subtract(src_skipped)
# Possibly having no changed blocks.
if not tgt_ranges:
return
# Add the transfer(s).
AddSplitTransfers(
tgt_name, src_name, tgt_ranges, src_ranges, style, by_id)
def test_subtract(self):
self.assertEqual(RangeSet("10-19 30-34").subtract(RangeSet("18-32")),
RangeSet("10-17 33-34"))
self.assertEqual(RangeSet("10-19 30-34").subtract(RangeSet("22-28")),
RangeSet("10-19 30-34"))
def WriteTransfers(self, prefix):
out = []
total = 0
performs_read = False
stashes = {}
stashed_blocks = 0
max_stashed_blocks = 0
free_stash_ids = []
next_stash_id = 0
for xf in self.transfers:
if self.version < 2:
assert not xf.stash_before
assert not xf.use_stash
for s, sr in xf.stash_before:
assert s not in stashes
if free_stash_ids:
sid = heapq.heappop(free_stash_ids)
else:
sid = next_stash_id
next_stash_id += 1
stashes[s] = sid
stashed_blocks += sr.size()
if self.version == 2:
out.append("stash %d %s\n" % (sid, sr.to_string_raw()))
else:
sh = self.HashBlocks(self.src, sr)
if sh in stashes:
stashes[sh] += 1
else:
stashes[sh] = 1
out.append("stash %s %s\n" % (sh, sr.to_string_raw()))
if stashed_blocks > max_stashed_blocks:
max_stashed_blocks = stashed_blocks
free_string = []
if self.version == 1:
src_str = xf.src_ranges.to_string_raw()
elif self.version >= 2:
# <# blocks> <src ranges>
# OR
# <# blocks> <src ranges> <src locs> <stash refs...>
# OR
# <# blocks> - <stash refs...>
size = xf.src_ranges.size()
src_str = [str(size)]
unstashed_src_ranges = xf.src_ranges
mapped_stashes = []
for s, sr in xf.use_stash:
sid = stashes.pop(s)
stashed_blocks -= sr.size()
unstashed_src_ranges = unstashed_src_ranges.subtract(sr)
sh = self.HashBlocks(self.src, sr)
sr = xf.src_ranges.map_within(sr)
mapped_stashes.append(sr)
if self.version == 2:
src_str.append("%d:%s" % (sid, sr.to_string_raw()))
else:
assert sh in stashes
src_str.append("%s:%s" % (sh, sr.to_string_raw()))
stashes[sh] -= 1
if stashes[sh] == 0:
free_string.append("free %s\n" % (sh))
stashes.pop(sh)
heapq.heappush(free_stash_ids, sid)
if unstashed_src_ranges:
src_str.insert(1, unstashed_src_ranges.to_string_raw())
if xf.use_stash:
mapped_unstashed = xf.src_ranges.map_within(
unstashed_src_ranges)
src_str.insert(2, mapped_unstashed.to_string_raw())
mapped_stashes.append(mapped_unstashed)
self.AssertPartition(RangeSet(data=(0, size)),
mapped_stashes)
else:
src_str.insert(1, "-")
self.AssertPartition(RangeSet(data=(0, size)),
mapped_stashes)
src_str = " ".join(src_str)
# all versions:
# zero <rangeset>
# new <rangeset>
# erase <rangeset>
#
# version 1:
# bsdiff patchstart patchlen <src rangeset> <tgt rangeset>
# imgdiff patchstart patchlen <src rangeset> <tgt rangeset>
# move <src rangeset> <tgt rangeset>
#
# version 2:
# bsdiff patchstart patchlen <tgt rangeset> <src_str>
# imgdiff patchstart patchlen <tgt rangeset> <src_str>
# move <tgt rangeset> <src_str>
#
# version 3:
# bsdiff patchstart patchlen srchash tgthash <tgt rangeset> <src_str>
# imgdiff patchstart patchlen srchash tgthash <tgt rangeset> <src_str>
# move hash <tgt rangeset> <src_str>
tgt_size = xf.tgt_ranges.size()
if xf.style == "new":
assert xf.tgt_ranges
out.append("%s %s\n" %
(xf.style, xf.tgt_ranges.to_string_raw()))
total += tgt_size
elif xf.style == "move":
performs_read = True
assert xf.tgt_ranges
assert xf.src_ranges.size() == tgt_size
if xf.src_ranges != xf.tgt_ranges:
if self.version == 1:
out.append("%s %s %s\n" %
(xf.style, xf.src_ranges.to_string_raw(),
xf.tgt_ranges.to_string_raw()))
elif self.version == 2:
out.append(
"%s %s %s\n" %
(xf.style, xf.tgt_ranges.to_string_raw(), src_str))
elif self.version >= 3:
# take into account automatic stashing of overlapping blocks
if xf.src_ranges.overlaps(xf.tgt_ranges):
temp_stash_usage = stashed_blocks + xf.src_ranges.size(
)
if temp_stash_usage > max_stashed_blocks:
max_stashed_blocks = temp_stash_usage
out.append("%s %s %s %s\n" %
(xf.style,
self.HashBlocks(self.tgt, xf.tgt_ranges),
xf.tgt_ranges.to_string_raw(), src_str))
total += tgt_size
elif xf.style in ("bsdiff", "imgdiff"):
performs_read = True
assert xf.tgt_ranges
assert xf.src_ranges
if self.version == 1:
out.append("%s %d %d %s %s\n" %
(xf.style, xf.patch_start, xf.patch_len,
xf.src_ranges.to_string_raw(),
xf.tgt_ranges.to_string_raw()))
elif self.version == 2:
out.append("%s %d %d %s %s\n" %
(xf.style, xf.patch_start, xf.patch_len,
xf.tgt_ranges.to_string_raw(), src_str))
elif self.version >= 3:
# take into account automatic stashing of overlapping blocks
if xf.src_ranges.overlaps(xf.tgt_ranges):
temp_stash_usage = stashed_blocks + xf.src_ranges.size(
)
if temp_stash_usage > max_stashed_blocks:
max_stashed_blocks = temp_stash_usage
out.append("%s %d %d %s %s %s %s\n" %
(xf.style, xf.patch_start, xf.patch_len,
self.HashBlocks(self.src, xf.src_ranges),
self.HashBlocks(self.tgt, xf.tgt_ranges),
xf.tgt_ranges.to_string_raw(), src_str))
total += tgt_size
elif xf.style == "zero":
assert xf.tgt_ranges
to_zero = xf.tgt_ranges.subtract(xf.src_ranges)
if to_zero:
out.append("%s %s\n" % (xf.style, to_zero.to_string_raw()))
total += to_zero.size()
else:
raise ValueError("unknown transfer style '%s'\n" % xf.style)
if free_string:
out.append("".join(free_string))
# sanity check: abort if we're going to need more than 512 MB if
# stash space
assert max_stashed_blocks * self.tgt.blocksize < (512 << 20)
# Zero out extended blocks as a workaround for bug 20881595.
if self.tgt.extended:
out.append("zero %s\n" % (self.tgt.extended.to_string_raw(), ))
# We erase all the blocks on the partition that a) don't contain useful
# data in the new image and b) will not be touched by dm-verity.
all_tgt = RangeSet(data=(0, self.tgt.total_blocks))
all_tgt_minus_extended = all_tgt.subtract(self.tgt.extended)
new_dontcare = all_tgt_minus_extended.subtract(self.tgt.care_map)
if new_dontcare:
out.append("erase %s\n" % (new_dontcare.to_string_raw(), ))
out.insert(0, "%d\n" % (self.version, )) # format version number
out.insert(1, str(total) + "\n")
if self.version >= 2:
# version 2 only: after the total block count, we give the number
# of stash slots needed, and the maximum size needed (in blocks)
out.insert(2, str(next_stash_id) + "\n")
out.insert(3, str(max_stashed_blocks) + "\n")
with open(prefix + ".transfer.list", "wb") as f:
for i in out:
f.write(i)
if self.version >= 2:
print(
"max stashed blocks: %d (%d bytes)\n" %
(max_stashed_blocks, max_stashed_blocks * self.tgt.blocksize))
def AddTransfer(tgt_name, src_name, tgt_ranges, src_ranges, style, by_id,
split=False):
"""Wrapper function for adding a Transfer()."""
# We specialize diff transfers only (which covers bsdiff/imgdiff/move);
# otherwise add the Transfer() as is.
if style != "diff" or not split:
Transfer(tgt_name, src_name, tgt_ranges, src_ranges, style, by_id)
return
# Handle .odex files specially to analyze the block-wise difference. If
# most of the blocks are identical with only few changes (e.g. header),
# we will patch the changed blocks only. This avoids stashing unchanged
# blocks while patching. We limit the analysis to files without size
# changes only. This is to avoid sacrificing the OTA generation cost too
# much.
if (tgt_name.split(".")[-1].lower() == 'odex' and
tgt_ranges.size() == src_ranges.size()):
# 0.5 threshold can be further tuned. The tradeoff is: if only very
# few blocks remain identical, we lose the opportunity to use imgdiff
# that may have better compression ratio than bsdiff.
crop_threshold = 0.5
tgt_skipped = RangeSet()
src_skipped = RangeSet()
tgt_size = tgt_ranges.size()
tgt_changed = 0
for src_block, tgt_block in zip(src_ranges.next_item(),
tgt_ranges.next_item()):
src_rs = RangeSet(str(src_block))
tgt_rs = RangeSet(str(tgt_block))
if self.src.ReadRangeSet(src_rs) == self.tgt.ReadRangeSet(tgt_rs):
tgt_skipped = tgt_skipped.union(tgt_rs)
src_skipped = src_skipped.union(src_rs)
else:
tgt_changed += tgt_rs.size()
# Terminate early if no clear sign of benefits.
if tgt_changed > tgt_size * crop_threshold:
break
if tgt_changed < tgt_size * crop_threshold:
assert tgt_changed + tgt_skipped.size() == tgt_size
print('%10d %10d (%6.2f%%) %s' % (tgt_skipped.size(), tgt_size,
tgt_skipped.size() * 100.0 / tgt_size, tgt_name))
AddSplitTransfers(
"%s-skipped" % (tgt_name,),
"%s-skipped" % (src_name,),
tgt_skipped, src_skipped, style, by_id)
# Intentionally change the file extension to avoid being imgdiff'd as
# the files are no longer in their original format.
tgt_name = "%s-cropped" % (tgt_name,)
src_name = "%s-cropped" % (src_name,)
tgt_ranges = tgt_ranges.subtract(tgt_skipped)
src_ranges = src_ranges.subtract(src_skipped)
# Possibly having no changed blocks.
if not tgt_ranges:
return
# Add the transfer(s).
AddSplitTransfers(
tgt_name, src_name, tgt_ranges, src_ranges, style, by_id)
def WriteTransfers(self, prefix):
out = []
total = 0
performs_read = False
stashes = {}
stashed_blocks = 0
max_stashed_blocks = 0
free_stash_ids = []
next_stash_id = 0
for xf in self.transfers:
if self.version < 2:
assert not xf.stash_before
assert not xf.use_stash
for s, sr in xf.stash_before:
assert s not in stashes
if free_stash_ids:
sid = heapq.heappop(free_stash_ids)
else:
sid = next_stash_id
next_stash_id += 1
stashes[s] = sid
stashed_blocks += sr.size()
if self.version == 2:
out.append("stash %d %s\n" % (sid, sr.to_string_raw()))
else:
sh = self.HashBlocks(self.src, sr)
if sh in stashes:
stashes[sh] += 1
else:
stashes[sh] = 1
out.append("stash %s %s\n" % (sh, sr.to_string_raw()))
if stashed_blocks > max_stashed_blocks:
max_stashed_blocks = stashed_blocks
free_string = []
if self.version == 1:
src_str = xf.src_ranges.to_string_raw()
elif self.version >= 2:
# <# blocks> <src ranges>
# OR
# <# blocks> <src ranges> <src locs> <stash refs...>
# OR
# <# blocks> - <stash refs...>
size = xf.src_ranges.size()
src_str = [str(size)]
unstashed_src_ranges = xf.src_ranges
mapped_stashes = []
for s, sr in xf.use_stash:
sid = stashes.pop(s)
stashed_blocks -= sr.size()
unstashed_src_ranges = unstashed_src_ranges.subtract(sr)
sh = self.HashBlocks(self.src, sr)
sr = xf.src_ranges.map_within(sr)
mapped_stashes.append(sr)
if self.version == 2:
src_str.append("%d:%s" % (sid, sr.to_string_raw()))
# A stash will be used only once. We need to free the stash
# immediately after the use, instead of waiting for the automatic
# clean-up at the end. Because otherwise it may take up extra space
# and lead to OTA failures.
# Bug: 23119955
free_string.append("free %d\n" % (sid,))
else:
assert sh in stashes
src_str.append("%s:%s" % (sh, sr.to_string_raw()))
stashes[sh] -= 1
if stashes[sh] == 0:
free_string.append("free %s\n" % (sh))
stashes.pop(sh)
heapq.heappush(free_stash_ids, sid)
if unstashed_src_ranges:
src_str.insert(1, unstashed_src_ranges.to_string_raw())
if xf.use_stash:
mapped_unstashed = xf.src_ranges.map_within(unstashed_src_ranges)
src_str.insert(2, mapped_unstashed.to_string_raw())
mapped_stashes.append(mapped_unstashed)
self.AssertPartition(RangeSet(data=(0, size)), mapped_stashes)
else:
src_str.insert(1, "-")
self.AssertPartition(RangeSet(data=(0, size)), mapped_stashes)
src_str = " ".join(src_str)
# all versions:
# zero <rangeset>
# new <rangeset>
# erase <rangeset>
#
# version 1:
# bsdiff patchstart patchlen <src rangeset> <tgt rangeset>
# imgdiff patchstart patchlen <src rangeset> <tgt rangeset>
# move <src rangeset> <tgt rangeset>
#
# version 2:
# bsdiff patchstart patchlen <tgt rangeset> <src_str>
# imgdiff patchstart patchlen <tgt rangeset> <src_str>
# move <tgt rangeset> <src_str>
#
# version 3:
# bsdiff patchstart patchlen srchash tgthash <tgt rangeset> <src_str>
# imgdiff patchstart patchlen srchash tgthash <tgt rangeset> <src_str>
# move hash <tgt rangeset> <src_str>
tgt_size = xf.tgt_ranges.size()
if xf.style == "new":
assert xf.tgt_ranges
out.append("%s %s\n" % (xf.style, xf.tgt_ranges.to_string_raw()))
total += tgt_size
elif xf.style == "move":
performs_read = True
assert xf.tgt_ranges
assert xf.src_ranges.size() == tgt_size
if xf.src_ranges != xf.tgt_ranges:
if self.version == 1:
out.append("%s %s %s\n" % (
xf.style,
xf.src_ranges.to_string_raw(), xf.tgt_ranges.to_string_raw()))
elif self.version == 2:
out.append("%s %s %s\n" % (
xf.style,
xf.tgt_ranges.to_string_raw(), src_str))
elif self.version >= 3:
# take into account automatic stashing of overlapping blocks
if xf.src_ranges.overlaps(xf.tgt_ranges):
temp_stash_usage = stashed_blocks + xf.src_ranges.size()
if temp_stash_usage > max_stashed_blocks:
max_stashed_blocks = temp_stash_usage
out.append("%s %s %s %s\n" % (
xf.style,
self.HashBlocks(self.tgt, xf.tgt_ranges),
xf.tgt_ranges.to_string_raw(), src_str))
total += tgt_size
elif xf.style in ("bsdiff", "imgdiff"):
performs_read = True
assert xf.tgt_ranges
assert xf.src_ranges
if self.version == 1:
out.append("%s %d %d %s %s\n" % (
xf.style, xf.patch_start, xf.patch_len,
xf.src_ranges.to_string_raw(), xf.tgt_ranges.to_string_raw()))
elif self.version == 2:
out.append("%s %d %d %s %s\n" % (
xf.style, xf.patch_start, xf.patch_len,
xf.tgt_ranges.to_string_raw(), src_str))
elif self.version >= 3:
# take into account automatic stashing of overlapping blocks
if xf.src_ranges.overlaps(xf.tgt_ranges):
temp_stash_usage = stashed_blocks + xf.src_ranges.size()
if temp_stash_usage > max_stashed_blocks:
max_stashed_blocks = temp_stash_usage
out.append("%s %d %d %s %s %s %s\n" % (
xf.style,
xf.patch_start, xf.patch_len,
self.HashBlocks(self.src, xf.src_ranges),
self.HashBlocks(self.tgt, xf.tgt_ranges),
xf.tgt_ranges.to_string_raw(), src_str))
total += tgt_size
elif xf.style == "zero":
assert xf.tgt_ranges
to_zero = xf.tgt_ranges.subtract(xf.src_ranges)
if to_zero:
out.append("%s %s\n" % (xf.style, to_zero.to_string_raw()))
total += to_zero.size()
else:
raise ValueError("unknown transfer style '%s'\n" % xf.style)
if free_string:
out.append("".join(free_string))
if self.version >= 2 and common.OPTIONS.cache_size is not None:
# Sanity check: abort if we're going to need more stash space than
# the allowed size (cache_size * threshold). There are two purposes
# of having a threshold here. a) Part of the cache may have been
# occupied by some recovery logs. b) It will buy us some time to deal
# with the oversize issue.
cache_size = common.OPTIONS.cache_size
stash_threshold = common.OPTIONS.stash_threshold
max_allowed = cache_size * stash_threshold
assert max_stashed_blocks * self.tgt.blocksize < max_allowed, \
'Stash size %d (%d * %d) exceeds the limit %d (%d * %.2f)' % (
max_stashed_blocks * self.tgt.blocksize, max_stashed_blocks,
self.tgt.blocksize, max_allowed, cache_size,
stash_threshold)
# Zero out extended blocks as a workaround for bug 20881595.
if self.tgt.extended:
out.append("zero %s\n" % (self.tgt.extended.to_string_raw(),))
# We erase all the blocks on the partition that a) don't contain useful
# data in the new image and b) will not be touched by dm-verity.
all_tgt = RangeSet(data=(0, self.tgt.total_blocks))
all_tgt_minus_extended = all_tgt.subtract(self.tgt.extended)
new_dontcare = all_tgt_minus_extended.subtract(self.tgt.care_map)
if new_dontcare:
out.append("erase %s\n" % (new_dontcare.to_string_raw(),))
out.insert(0, "%d\n" % (self.version,)) # format version number
out.insert(1, str(total) + "\n")
if self.version >= 2:
# version 2 only: after the total block count, we give the number
# of stash slots needed, and the maximum size needed (in blocks)
out.insert(2, str(next_stash_id) + "\n")
out.insert(3, str(max_stashed_blocks) + "\n")
with open(prefix + ".transfer.list", "wb") as f:
for i in out:
f.write(i)
if self.version >= 2:
max_stashed_size = max_stashed_blocks * self.tgt.blocksize
OPTIONS = common.OPTIONS
if OPTIONS.cache_size is not None:
max_allowed = OPTIONS.cache_size * OPTIONS.stash_threshold
print("max stashed blocks: %d (%d bytes), "
"limit: %d bytes (%.2f%%)\n" % (
max_stashed_blocks, max_stashed_size, max_allowed,
max_stashed_size * 100.0 / max_allowed))
else:
print("max stashed blocks: %d (%d bytes), limit: <unknown>\n" % (
max_stashed_blocks, max_stashed_size))
def test_CanUseImgdiff(self):
block_image_diff = BlockImageDiff(EmptyImage(), EmptyImage())
self.assertTrue(
block_image_diff.CanUseImgdiff("/system/app/app1.apk",
RangeSet("10-15"), RangeSet("0-5")))
def WriteTransfers(self, prefix):
out = []
total = 0
performs_read = False
stashes = {}
stashed_blocks = 0
max_stashed_blocks = 0
free_stash_ids = []
next_stash_id = 0
for xf in self.transfers:
if self.version < 2:
assert not xf.stash_before
assert not xf.use_stash
for s, sr in xf.stash_before:
assert s not in stashes
if free_stash_ids:
sid = heapq.heappop(free_stash_ids)
else:
sid = next_stash_id
next_stash_id += 1
stashes[s] = sid
stashed_blocks += sr.size()
if self.version == 2:
out.append("stash %d %s\n" % (sid, sr.to_string_raw()))
else:
sh = self.HashBlocks(self.src, sr)
if sh in stashes:
stashes[sh] += 1
else:
stashes[sh] = 1
out.append("stash %s %s\n" % (sh, sr.to_string_raw()))
if stashed_blocks > max_stashed_blocks:
max_stashed_blocks = stashed_blocks
free_string = []
if self.version == 1:
src_str = xf.src_ranges.to_string_raw()
elif self.version >= 2:
# <# blocks> <src ranges>
# OR
# <# blocks> <src ranges> <src locs> <stash refs...>
# OR
# <# blocks> - <stash refs...>
size = xf.src_ranges.size()
src_str = [str(size)]
unstashed_src_ranges = xf.src_ranges
mapped_stashes = []
for s, sr in xf.use_stash:
sid = stashes.pop(s)
stashed_blocks -= sr.size()
unstashed_src_ranges = unstashed_src_ranges.subtract(sr)
sh = self.HashBlocks(self.src, sr)
sr = xf.src_ranges.map_within(sr)
mapped_stashes.append(sr)
if self.version == 2:
src_str.append("%d:%s" % (sid, sr.to_string_raw()))
else:
assert sh in stashes
src_str.append("%s:%s" % (sh, sr.to_string_raw()))
stashes[sh] -= 1
if stashes[sh] == 0:
free_string.append("free %s\n" % (sh))
stashes.pop(sh)
heapq.heappush(free_stash_ids, sid)
if unstashed_src_ranges:
src_str.insert(1, unstashed_src_ranges.to_string_raw())
if xf.use_stash:
mapped_unstashed = xf.src_ranges.map_within(unstashed_src_ranges)
src_str.insert(2, mapped_unstashed.to_string_raw())
mapped_stashes.append(mapped_unstashed)
self.AssertPartition(RangeSet(data=(0, size)), mapped_stashes)
else:
src_str.insert(1, "-")
self.AssertPartition(RangeSet(data=(0, size)), mapped_stashes)
src_str = " ".join(src_str)
# all versions:
# zero <rangeset>
# new <rangeset>
# erase <rangeset>
#
# version 1:
# bsdiff patchstart patchlen <src rangeset> <tgt rangeset>
# imgdiff patchstart patchlen <src rangeset> <tgt rangeset>
# move <src rangeset> <tgt rangeset>
#
# version 2:
# bsdiff patchstart patchlen <tgt rangeset> <src_str>
# imgdiff patchstart patchlen <tgt rangeset> <src_str>
# move <tgt rangeset> <src_str>
#
# version 3:
# bsdiff patchstart patchlen srchash tgthash <tgt rangeset> <src_str>
# imgdiff patchstart patchlen srchash tgthash <tgt rangeset> <src_str>
# move hash <tgt rangeset> <src_str>
tgt_size = xf.tgt_ranges.size()
if xf.style == "new":
assert xf.tgt_ranges
out.append("%s %s\n" % (xf.style, xf.tgt_ranges.to_string_raw()))
total += tgt_size
elif xf.style == "move":
performs_read = True
assert xf.tgt_ranges
assert xf.src_ranges.size() == tgt_size
if xf.src_ranges != xf.tgt_ranges:
if self.version == 1:
out.append("%s %s %s\n" % (
xf.style,
xf.src_ranges.to_string_raw(), xf.tgt_ranges.to_string_raw()))
elif self.version == 2:
out.append("%s %s %s\n" % (
xf.style,
xf.tgt_ranges.to_string_raw(), src_str))
elif self.version >= 3:
# take into account automatic stashing of overlapping blocks
if xf.src_ranges.overlaps(xf.tgt_ranges):
temp_stash_usage = stashed_blocks + xf.src_ranges.size();
if temp_stash_usage > max_stashed_blocks:
max_stashed_blocks = temp_stash_usage
out.append("%s %s %s %s\n" % (
xf.style,
self.HashBlocks(self.tgt, xf.tgt_ranges),
xf.tgt_ranges.to_string_raw(), src_str))
total += tgt_size
elif xf.style in ("bsdiff", "imgdiff"):
performs_read = True
assert xf.tgt_ranges
assert xf.src_ranges
if self.version == 1:
out.append("%s %d %d %s %s\n" % (
xf.style, xf.patch_start, xf.patch_len,
xf.src_ranges.to_string_raw(), xf.tgt_ranges.to_string_raw()))
elif self.version == 2:
out.append("%s %d %d %s %s\n" % (
xf.style, xf.patch_start, xf.patch_len,
xf.tgt_ranges.to_string_raw(), src_str))
elif self.version >= 3:
# take into account automatic stashing of overlapping blocks
if xf.src_ranges.overlaps(xf.tgt_ranges):
temp_stash_usage = stashed_blocks + xf.src_ranges.size();
if temp_stash_usage > max_stashed_blocks:
max_stashed_blocks = temp_stash_usage
out.append("%s %d %d %s %s %s %s\n" % (
xf.style,
xf.patch_start, xf.patch_len,
self.HashBlocks(self.src, xf.src_ranges),
self.HashBlocks(self.tgt, xf.tgt_ranges),
xf.tgt_ranges.to_string_raw(), src_str))
total += tgt_size
elif xf.style == "zero":
assert xf.tgt_ranges
to_zero = xf.tgt_ranges.subtract(xf.src_ranges)
if to_zero:
out.append("%s %s\n" % (xf.style, to_zero.to_string_raw()))
total += to_zero.size()
else:
raise ValueError("unknown transfer style '%s'\n" % xf.style)
if free_string:
out.append("".join(free_string))
# sanity check: abort if we're going to need more than 512 MB if
# stash space
assert max_stashed_blocks * self.tgt.blocksize < (512 << 20)
all_tgt = RangeSet(data=(0, self.tgt.total_blocks))
if performs_read:
# if some of the original data is used, then at the end we'll
# erase all the blocks on the partition that don't contain data
# in the new image.
new_dontcare = all_tgt.subtract(self.tgt.care_map)
if new_dontcare:
out.append("erase %s\n" % (new_dontcare.to_string_raw(),))
else:
# if nothing is read (ie, this is a full OTA), then we can start
# by erasing the entire partition.
out.insert(0, "erase %s\n" % (all_tgt.to_string_raw(),))
out.insert(0, "%d\n" % (self.version,)) # format version number
out.insert(1, str(total) + "\n")
if self.version >= 2:
# version 2 only: after the total block count, we give the number
# of stash slots needed, and the maximum size needed (in blocks)
out.insert(2, str(next_stash_id) + "\n")
out.insert(3, str(max_stashed_blocks) + "\n")
with open(prefix + ".transfer.list", "wb") as f:
for i in out:
f.write(i)
if self.version >= 2:
print("max stashed blocks: %d (%d bytes)\n" % (
max_stashed_blocks, max_stashed_blocks * self.tgt.blocksize))
def WriteTransfers(self, prefix):
def WriteSplitTransfers(out, style, target_blocks):
"""Limit the size of operand in command 'new' and 'zero' to 1024 blocks.
This prevents the target size of one command from being too large; and
might help to avoid fsync errors on some devices."""
assert style == "new" or style == "zero"
blocks_limit = 1024
total = 0
while target_blocks:
blocks_to_write = target_blocks.first(blocks_limit)
out.append("%s %s\n" % (style, blocks_to_write.to_string_raw()))
total += blocks_to_write.size()
target_blocks = target_blocks.subtract(blocks_to_write)
return total
out = []
total = 0
# In BBOTA v3+, it uses the hash of the stashed blocks as the stash slot
# id. 'stashes' records the map from 'hash' to the ref count. The stash
# will be freed only if the count decrements to zero.
stashes = {}
stashed_blocks = 0
max_stashed_blocks = 0
for xf in self.transfers:
for _, sr in xf.stash_before:
sh = self.src.RangeSha1(sr)
if sh in stashes:
stashes[sh] += 1
else:
stashes[sh] = 1
stashed_blocks += sr.size()
self.touched_src_ranges = self.touched_src_ranges.union(sr)
out.append("stash %s %s\n" % (sh, sr.to_string_raw()))
if stashed_blocks > max_stashed_blocks:
max_stashed_blocks = stashed_blocks
free_string = []
free_size = 0
# <# blocks> <src ranges>
# OR
# <# blocks> <src ranges> <src locs> <stash refs...>
# OR
# <# blocks> - <stash refs...>
size = xf.src_ranges.size()
src_str = [str(size)]
unstashed_src_ranges = xf.src_ranges
mapped_stashes = []
for _, sr in xf.use_stash:
unstashed_src_ranges = unstashed_src_ranges.subtract(sr)
sh = self.src.RangeSha1(sr)
sr = xf.src_ranges.map_within(sr)
mapped_stashes.append(sr)
assert sh in stashes
src_str.append("%s:%s" % (sh, sr.to_string_raw()))
stashes[sh] -= 1
if stashes[sh] == 0:
free_string.append("free %s\n" % (sh,))
free_size += sr.size()
stashes.pop(sh)
if unstashed_src_ranges:
src_str.insert(1, unstashed_src_ranges.to_string_raw())
if xf.use_stash:
mapped_unstashed = xf.src_ranges.map_within(unstashed_src_ranges)
src_str.insert(2, mapped_unstashed.to_string_raw())
mapped_stashes.append(mapped_unstashed)
self.AssertPartition(RangeSet(data=(0, size)), mapped_stashes)
else:
src_str.insert(1, "-")
self.AssertPartition(RangeSet(data=(0, size)), mapped_stashes)
src_str = " ".join(src_str)
# version 3+:
# zero <rangeset>
# new <rangeset>
# erase <rangeset>
# bsdiff patchstart patchlen srchash tgthash <tgt rangeset> <src_str>
# imgdiff patchstart patchlen srchash tgthash <tgt rangeset> <src_str>
# move hash <tgt rangeset> <src_str>
tgt_size = xf.tgt_ranges.size()
if xf.style == "new":
assert xf.tgt_ranges
assert tgt_size == WriteSplitTransfers(out, xf.style, xf.tgt_ranges)
total += tgt_size
elif xf.style == "move":
assert xf.tgt_ranges
assert xf.src_ranges.size() == tgt_size
if xf.src_ranges != xf.tgt_ranges:
# take into account automatic stashing of overlapping blocks
if xf.src_ranges.overlaps(xf.tgt_ranges):
temp_stash_usage = stashed_blocks + xf.src_ranges.size()
if temp_stash_usage > max_stashed_blocks:
max_stashed_blocks = temp_stash_usage
self.touched_src_ranges = self.touched_src_ranges.union(
xf.src_ranges)
out.append("%s %s %s %s\n" % (
xf.style,
xf.tgt_sha1,
xf.tgt_ranges.to_string_raw(), src_str))
total += tgt_size
elif xf.style in ("bsdiff", "imgdiff"):
assert xf.tgt_ranges
assert xf.src_ranges
# take into account automatic stashing of overlapping blocks
if xf.src_ranges.overlaps(xf.tgt_ranges):
temp_stash_usage = stashed_blocks + xf.src_ranges.size()
if temp_stash_usage > max_stashed_blocks:
max_stashed_blocks = temp_stash_usage
self.touched_src_ranges = self.touched_src_ranges.union(xf.src_ranges)
out.append("%s %d %d %s %s %s %s\n" % (
xf.style,
xf.patch_start, xf.patch_len,
xf.src_sha1,
xf.tgt_sha1,
xf.tgt_ranges.to_string_raw(), src_str))
total += tgt_size
elif xf.style == "zero":
assert xf.tgt_ranges
to_zero = xf.tgt_ranges.subtract(xf.src_ranges)
assert WriteSplitTransfers(out, xf.style, to_zero) == to_zero.size()
total += to_zero.size()
else:
raise ValueError("unknown transfer style '%s'\n" % xf.style)
if free_string:
out.append("".join(free_string))
stashed_blocks -= free_size
if common.OPTIONS.cache_size is not None:
# Sanity check: abort if we're going to need more stash space than
# the allowed size (cache_size * threshold). There are two purposes
# of having a threshold here. a) Part of the cache may have been
# occupied by some recovery logs. b) It will buy us some time to deal
# with the oversize issue.
cache_size = common.OPTIONS.cache_size
stash_threshold = common.OPTIONS.stash_threshold
max_allowed = cache_size * stash_threshold
assert max_stashed_blocks * self.tgt.blocksize <= max_allowed, \
'Stash size %d (%d * %d) exceeds the limit %d (%d * %.2f)' % (
max_stashed_blocks * self.tgt.blocksize, max_stashed_blocks,
self.tgt.blocksize, max_allowed, cache_size,
stash_threshold)
self.touched_src_sha1 = self.src.RangeSha1(self.touched_src_ranges)
# Zero out extended blocks as a workaround for bug 20881595.
if self.tgt.extended:
assert (WriteSplitTransfers(out, "zero", self.tgt.extended) ==
self.tgt.extended.size())
total += self.tgt.extended.size()
# We erase all the blocks on the partition that a) don't contain useful
# data in the new image; b) will not be touched by dm-verity. Out of those
# blocks, we erase the ones that won't be used in this update at the
# beginning of an update. The rest would be erased at the end. This is to
# work around the eMMC issue observed on some devices, which may otherwise
# get starving for clean blocks and thus fail the update. (b/28347095)
all_tgt = RangeSet(data=(0, self.tgt.total_blocks))
all_tgt_minus_extended = all_tgt.subtract(self.tgt.extended)
new_dontcare = all_tgt_minus_extended.subtract(self.tgt.care_map)
erase_first = new_dontcare.subtract(self.touched_src_ranges)
if erase_first:
out.insert(0, "erase %s\n" % (erase_first.to_string_raw(),))
erase_last = new_dontcare.subtract(erase_first)
if erase_last:
out.append("erase %s\n" % (erase_last.to_string_raw(),))
out.insert(0, "%d\n" % (self.version,)) # format version number
out.insert(1, "%d\n" % (total,))
# v3+: the number of stash slots is unused.
out.insert(2, "0\n")
out.insert(3, str(max_stashed_blocks) + "\n")
with open(prefix + ".transfer.list", "wb") as f:
for i in out:
f.write(i)
self._max_stashed_size = max_stashed_blocks * self.tgt.blocksize
OPTIONS = common.OPTIONS
if OPTIONS.cache_size is not None:
max_allowed = OPTIONS.cache_size * OPTIONS.stash_threshold
print("max stashed blocks: %d (%d bytes), "
"limit: %d bytes (%.2f%%)\n" % (
max_stashed_blocks, self._max_stashed_size, max_allowed,
self._max_stashed_size * 100.0 / max_allowed))
else:
print("max stashed blocks: %d (%d bytes), limit: <unknown>\n" % (
max_stashed_blocks, self._max_stashed_size))
def test_CanUseImgdiff_ineligible(self):
# Disabled by caller.
block_image_diff = BlockImageDiff(EmptyImage(),
EmptyImage(),
disable_imgdiff=True)
self.assertFalse(
block_image_diff.CanUseImgdiff("/system/app/app1.apk",
RangeSet("10-15"), RangeSet("0-5")))
# Unsupported file type.
block_image_diff = BlockImageDiff(EmptyImage(), EmptyImage())
self.assertFalse(
block_image_diff.CanUseImgdiff("/system/bin/gzip",
RangeSet("10-15"), RangeSet("0-5")))
# At least one of the ranges is in non-monotonic order.
self.assertFalse(
block_image_diff.CanUseImgdiff("/system/app/app2.apk",
RangeSet("10-15"),
RangeSet("15-20 30 10-14")))
# At least one of the ranges has been modified.
src_ranges = RangeSet("0-5")
src_ranges.extra['trimmed'] = True
self.assertFalse(
block_image_diff.CanUseImgdiff("/vendor/app/app3.apk",
RangeSet("10-15"), src_ranges))
# At least one of the ranges is incomplete.
src_ranges = RangeSet("0-5")
src_ranges.extra['incomplete'] = True
self.assertFalse(
block_image_diff.CanUseImgdiff("/vendor/app/app4.apk",
RangeSet("10-15"), src_ranges))
# The stats are correctly logged.
self.assertDictEqual(
{
ImgdiffStats.SKIPPED_NONMONOTONIC: {'/system/app/app2.apk'},
ImgdiffStats.SKIPPED_TRIMMED: {'/vendor/app/app3.apk'},
ImgdiffStats.SKIPPED_INCOMPLETE: {'/vendor/app/app4.apk'},
}, block_image_diff.imgdiff_stats.stats)
def test_equality(self):
self.assertTrue(RangeSet("") == RangeSet(""))
self.assertTrue(RangeSet("3") == RangeSet("3"))
self.assertTrue(RangeSet("3 5") == RangeSet("5 3"))
self.assertTrue(
RangeSet("10-19 30-39") == RangeSet("30-32 10-14 33-39 15-19"))
self.assertTrue(RangeSet("") != RangeSet("3"))
self.assertTrue(RangeSet("10-19") != RangeSet("10-19 20"))
self.assertFalse(RangeSet(""))
self.assertTrue(RangeSet("3"))
def ParseTransferList(self, name):
"""Simulate the transfer commands and calculate the amout of I/O."""
logging.info("\nSimulating commands in '{}':".format(name))
lines = self.package.read(name).strip().splitlines()
assert len(lines) >= 4, "{} is too short; Transfer list expects at least" \
"4 lines, it has {}".format(name, len(lines))
assert int(lines[0]) >= 3
logging.info("(version: {})".format(lines[0]))
blocks_written = 0
my_stash = Stash()
for line in lines[4:]:
cmd_list = line.strip().split(" ")
cmd_name = cmd_list[0]
try:
if cmd_name == "new" or cmd_name == "zero":
assert len(cmd_list) == 2, "command format error: {}".format(line)
target_range = RangeSet.parse_raw(cmd_list[1])
blocks_written += target_range.size()
elif cmd_name == "move":
# Example: move <onehash> <tgt_range> <src_blk_count> <src_range>
# [<loc_range> <stashed_blocks>]
assert len(cmd_list) >= 5, "command format error: {}".format(line)
target_range = RangeSet.parse_raw(cmd_list[2])
blocks_written += target_range.size()
if cmd_list[4] == '-':
continue
SHA1 = cmd_list[1]
source_range = RangeSet.parse_raw(cmd_list[4])
if target_range.overlaps(source_range):
my_stash.HandleOverlapBlocks(SHA1, source_range)
elif cmd_name == "bsdiff" or cmd_name == "imgdiff":
# Example: bsdiff <offset> <len> <src_hash> <tgt_hash> <tgt_range>
# <src_blk_count> <src_range> [<loc_range> <stashed_blocks>]
assert len(cmd_list) >= 8, "command format error: {}".format(line)
target_range = RangeSet.parse_raw(cmd_list[5])
blocks_written += target_range.size()
if cmd_list[7] == '-':
continue
source_SHA1 = cmd_list[3]
source_range = RangeSet.parse_raw(cmd_list[7])
if target_range.overlaps(source_range):
my_stash.HandleOverlapBlocks(source_SHA1, source_range)
elif cmd_name == "stash":
assert len(cmd_list) == 3, "command format error: {}".format(line)
SHA1 = cmd_list[1]
source_range = RangeSet.parse_raw(cmd_list[2])
my_stash.StashBlocks(SHA1, source_range)
elif cmd_name == "free":
assert len(cmd_list) == 2, "command format error: {}".format(line)
SHA1 = cmd_list[1]
my_stash.FreeBlocks(SHA1)
except:
logging.error("failed to parse command in: " + line)
raise
self.block_written += blocks_written
self.block_stashed += my_stash.blocks_stashed
logging.info("blocks written: {} (expected: {})".format(
blocks_written, lines[1]))
logging.info("max blocks stashed simultaneously: {} (expected: {})".
format(my_stash.max_stash_needed, lines[3]))
logging.info("total blocks stashed: {}".format(my_stash.blocks_stashed))
logging.info("blocks stashed implicitly: {}".format(
my_stash.overlap_blocks_stashed))
def test_size(self):
self.assertEqual(RangeSet("10-19 30-34").size(), 15)
self.assertEqual(RangeSet("").size(), 0)
def test_parse_raw(self):
self.assertEqual(RangeSet.parse_raw(RangeSet("0-9").to_string_raw()),
RangeSet("0-9"))
self.assertEqual(
RangeSet.parse_raw(RangeSet("2-10 12").to_string_raw()),
RangeSet("2-10 12"))
self.assertEqual(
RangeSet.parse_raw(RangeSet("11 2-10 12 1 0").to_string_raw()),
RangeSet("11 2-10 12 1 0"))
with self.assertRaises(AssertionError):
RangeSet.parse_raw("4,0,10")
def test_map_within(self):
self.assertEqual(
RangeSet("0-9").map_within(RangeSet("3-4")), RangeSet("3-4"))
self.assertEqual(
RangeSet("10-19").map_within(RangeSet("13-14")), RangeSet("3-4"))
self.assertEqual(
RangeSet("10-19 30-39").map_within(RangeSet("17-19 30-32")),
RangeSet("7-12"))
self.assertEqual(
RangeSet("10-19 30-39").map_within(RangeSet("12-13 17-19 30-32")),
RangeSet("2-3 7-12"))
def test_first(self):
self.assertEqual(RangeSet("0-9").first(1), RangeSet("0"))
self.assertEqual(RangeSet("10-19").first(5), RangeSet("10-14"))
self.assertEqual(RangeSet("10-19").first(15), RangeSet("10-19"))
self.assertEqual(RangeSet("10-19 30-39").first(3), RangeSet("10-12"))
self.assertEqual(
RangeSet("10-19 30-39").first(15), RangeSet("10-19 30-34"))
self.assertEqual(
RangeSet("10-19 30-39").first(30), RangeSet("10-19 30-39"))
self.assertEqual(RangeSet("0-9").first(0), RangeSet(""))
class BlockImageDiff(object):
def __init__(self, tgt, src=None, threads=None, version=4,
disable_imgdiff=False):
if threads is None:
threads = multiprocessing.cpu_count() // 2
if threads == 0:
threads = 1
self.threads = threads
self.version = version
self.transfers = []
self.src_basenames = {}
self.src_numpatterns = {}
self._max_stashed_size = 0
self.touched_src_ranges = RangeSet()
self.touched_src_sha1 = None
self.disable_imgdiff = disable_imgdiff
assert version in (1, 2, 3, 4)
self.tgt = tgt
if src is None:
src = EmptyImage()
self.src = src
# The updater code that installs the patch always uses 4k blocks.
assert tgt.blocksize == 4096
assert src.blocksize == 4096
# The range sets in each filemap should comprise a partition of
# the care map.
self.AssertPartition(src.care_map, src.file_map.values())
self.AssertPartition(tgt.care_map, tgt.file_map.values())
@property
def max_stashed_size(self):
return self._max_stashed_size
def Compute(self, prefix):
# When looking for a source file to use as the diff input for a
# target file, we try:
# 1) an exact path match if available, otherwise
# 2) a exact basename match if available, otherwise
# 3) a basename match after all runs of digits are replaced by
# "#" if available, otherwise
# 4) we have no source for this target.
self.AbbreviateSourceNames()
self.FindTransfers()
# Find the ordering dependencies among transfers (this is O(n^2)
# in the number of transfers).
self.GenerateDigraph()
# Find a sequence of transfers that satisfies as many ordering
# dependencies as possible (heuristically).
self.FindVertexSequence()
# Fix up the ordering dependencies that the sequence didn't
# satisfy.
if self.version == 1:
self.RemoveBackwardEdges()
else:
self.ReverseBackwardEdges()
self.ImproveVertexSequence()
# Ensure the runtime stash size is under the limit.
if self.version >= 2 and common.OPTIONS.cache_size is not None:
self.ReviseStashSize()
# Double-check our work.
self.AssertSequenceGood()
self.ComputePatches(prefix)
self.WriteTransfers(prefix)
def HashBlocks(self, source, ranges): # pylint: disable=no-self-use
data = source.ReadRangeSet(ranges)
ctx = sha1()
for p in data:
ctx.update(p)
return ctx.hexdigest()
def WriteTransfers(self, prefix):
def WriteSplitTransfers(out, style, target_blocks):
"""Limit the size of operand in command 'new' and 'zero' to 1024 blocks.
This prevents the target size of one command from being too large; and
might help to avoid fsync errors on some devices."""
assert (style == "new" or style == "zero")
blocks_limit = 1024
total = 0
while target_blocks:
blocks_to_write = target_blocks.first(blocks_limit)
out.append("%s %s\n" % (style, blocks_to_write.to_string_raw()))
total += blocks_to_write.size()
target_blocks = target_blocks.subtract(blocks_to_write)
return total
out = []
total = 0
stashes = {}
stashed_blocks = 0
max_stashed_blocks = 0
free_stash_ids = []
next_stash_id = 0
for xf in self.transfers:
if self.version < 2:
assert not xf.stash_before
assert not xf.use_stash
for s, sr in xf.stash_before:
assert s not in stashes
if free_stash_ids:
sid = heapq.heappop(free_stash_ids)
else:
sid = next_stash_id
next_stash_id += 1
stashes[s] = sid
if self.version == 2:
stashed_blocks += sr.size()
out.append("stash %d %s\n" % (sid, sr.to_string_raw()))
else:
sh = self.HashBlocks(self.src, sr)
if sh in stashes:
stashes[sh] += 1
else:
stashes[sh] = 1
stashed_blocks += sr.size()
self.touched_src_ranges = self.touched_src_ranges.union(sr)
out.append("stash %s %s\n" % (sh, sr.to_string_raw()))
if stashed_blocks > max_stashed_blocks:
max_stashed_blocks = stashed_blocks
free_string = []
free_size = 0
if self.version == 1:
src_str = xf.src_ranges.to_string_raw() if xf.src_ranges else ""
elif self.version >= 2:
# <# blocks> <src ranges>
# OR
# <# blocks> <src ranges> <src locs> <stash refs...>
# OR
# <# blocks> - <stash refs...>
size = xf.src_ranges.size()
src_str = [str(size)]
unstashed_src_ranges = xf.src_ranges
mapped_stashes = []
for s, sr in xf.use_stash:
# TODO: We don't need 'sid' (nor free_stash_ids) in BBOTA v3+.
sid = stashes.pop(s)
unstashed_src_ranges = unstashed_src_ranges.subtract(sr)
sh = self.HashBlocks(self.src, sr)
sr = xf.src_ranges.map_within(sr)
mapped_stashes.append(sr)
if self.version == 2:
src_str.append("%d:%s" % (sid, sr.to_string_raw()))
# A stash will be used only once. We need to free the stash
# immediately after the use, instead of waiting for the automatic
# clean-up at the end. Because otherwise it may take up extra space
# and lead to OTA failures.
# Bug: 23119955
free_string.append("free %d\n" % (sid,))
free_size += sr.size()
else:
assert sh in stashes
src_str.append("%s:%s" % (sh, sr.to_string_raw()))
stashes[sh] -= 1
if stashes[sh] == 0:
free_size += sr.size()
free_string.append("free %s\n" % (sh,))
stashes.pop(sh)
heapq.heappush(free_stash_ids, sid)
if unstashed_src_ranges:
src_str.insert(1, unstashed_src_ranges.to_string_raw())
if xf.use_stash:
mapped_unstashed = xf.src_ranges.map_within(unstashed_src_ranges)
src_str.insert(2, mapped_unstashed.to_string_raw())
mapped_stashes.append(mapped_unstashed)
self.AssertPartition(RangeSet(data=(0, size)), mapped_stashes)
else:
src_str.insert(1, "-")
self.AssertPartition(RangeSet(data=(0, size)), mapped_stashes)
src_str = " ".join(src_str)
# all versions:
# zero <rangeset>
# new <rangeset>
# erase <rangeset>
#
# version 1:
# bsdiff patchstart patchlen <src rangeset> <tgt rangeset>
# imgdiff patchstart patchlen <src rangeset> <tgt rangeset>
# move <src rangeset> <tgt rangeset>
#
# version 2:
# bsdiff patchstart patchlen <tgt rangeset> <src_str>
# imgdiff patchstart patchlen <tgt rangeset> <src_str>
# move <tgt rangeset> <src_str>
#
# version 3:
# bsdiff patchstart patchlen srchash tgthash <tgt rangeset> <src_str>
# imgdiff patchstart patchlen srchash tgthash <tgt rangeset> <src_str>
# move hash <tgt rangeset> <src_str>
tgt_size = xf.tgt_ranges.size()
if xf.style == "new":
assert xf.tgt_ranges
assert tgt_size == WriteSplitTransfers(out, xf.style, xf.tgt_ranges)
total += tgt_size
elif xf.style == "move":
assert xf.tgt_ranges
assert xf.src_ranges.size() == tgt_size
if xf.src_ranges != xf.tgt_ranges:
if self.version == 1:
out.append("%s %s %s\n" % (
xf.style,
xf.src_ranges.to_string_raw(), xf.tgt_ranges.to_string_raw()))
elif self.version == 2:
out.append("%s %s %s\n" % (
xf.style,
xf.tgt_ranges.to_string_raw(), src_str))
elif self.version >= 3:
# take into account automatic stashing of overlapping blocks
if xf.src_ranges.overlaps(xf.tgt_ranges):
temp_stash_usage = stashed_blocks + xf.src_ranges.size()
if temp_stash_usage > max_stashed_blocks:
max_stashed_blocks = temp_stash_usage
self.touched_src_ranges = self.touched_src_ranges.union(
xf.src_ranges)
out.append("%s %s %s %s\n" % (
xf.style,
self.HashBlocks(self.tgt, xf.tgt_ranges),
xf.tgt_ranges.to_string_raw(), src_str))
total += tgt_size
elif xf.style in ("bsdiff", "imgdiff"):
assert xf.tgt_ranges
assert xf.src_ranges
if self.version == 1:
out.append("%s %d %d %s %s\n" % (
xf.style, xf.patch_start, xf.patch_len,
xf.src_ranges.to_string_raw(), xf.tgt_ranges.to_string_raw()))
elif self.version == 2:
out.append("%s %d %d %s %s\n" % (
xf.style, xf.patch_start, xf.patch_len,
xf.tgt_ranges.to_string_raw(), src_str))
elif self.version >= 3:
# take into account automatic stashing of overlapping blocks
if xf.src_ranges.overlaps(xf.tgt_ranges):
temp_stash_usage = stashed_blocks + xf.src_ranges.size()
if temp_stash_usage > max_stashed_blocks:
max_stashed_blocks = temp_stash_usage
self.touched_src_ranges = self.touched_src_ranges.union(
xf.src_ranges)
out.append("%s %d %d %s %s %s %s\n" % (
xf.style,
xf.patch_start, xf.patch_len,
self.HashBlocks(self.src, xf.src_ranges),
self.HashBlocks(self.tgt, xf.tgt_ranges),
xf.tgt_ranges.to_string_raw(), src_str))
total += tgt_size
elif xf.style == "zero":
assert xf.tgt_ranges
to_zero = xf.tgt_ranges.subtract(xf.src_ranges)
assert WriteSplitTransfers(out, xf.style, to_zero) == to_zero.size()
total += to_zero.size()
else:
raise ValueError("unknown transfer style '%s'\n" % xf.style)
if free_string:
out.append("".join(free_string))
stashed_blocks -= free_size
if self.version >= 2 and common.OPTIONS.cache_size is not None:
# Sanity check: abort if we're going to need more stash space than
# the allowed size (cache_size * threshold). There are two purposes
# of having a threshold here. a) Part of the cache may have been
# occupied by some recovery logs. b) It will buy us some time to deal
# with the oversize issue.
cache_size = common.OPTIONS.cache_size
stash_threshold = common.OPTIONS.stash_threshold
max_allowed = cache_size * stash_threshold
assert max_stashed_blocks * self.tgt.blocksize < max_allowed, \
'Stash size %d (%d * %d) exceeds the limit %d (%d * %.2f)' % (
max_stashed_blocks * self.tgt.blocksize, max_stashed_blocks,
self.tgt.blocksize, max_allowed, cache_size,
stash_threshold)
if self.version >= 3:
self.touched_src_sha1 = self.HashBlocks(
self.src, self.touched_src_ranges)
# Zero out extended blocks as a workaround for bug 20881595.
if self.tgt.extended:
assert (WriteSplitTransfers(out, "zero", self.tgt.extended) ==
self.tgt.extended.size())
total += self.tgt.extended.size()
# We erase all the blocks on the partition that a) don't contain useful
# data in the new image; b) will not be touched by dm-verity. Out of those
# blocks, we erase the ones that won't be used in this update at the
# beginning of an update. The rest would be erased at the end. This is to
# work around the eMMC issue observed on some devices, which may otherwise
# get starving for clean blocks and thus fail the update. (b/28347095)
all_tgt = RangeSet(data=(0, self.tgt.total_blocks))
all_tgt_minus_extended = all_tgt.subtract(self.tgt.extended)
new_dontcare = all_tgt_minus_extended.subtract(self.tgt.care_map)
erase_first = new_dontcare.subtract(self.touched_src_ranges)
if erase_first:
out.insert(0, "erase %s\n" % (erase_first.to_string_raw(),))
erase_last = new_dontcare.subtract(erase_first)
if erase_last:
out.append("erase %s\n" % (erase_last.to_string_raw(),))
out.insert(0, "%d\n" % (self.version,)) # format version number
out.insert(1, "%d\n" % (total,))
if self.version >= 2:
# version 2 only: after the total block count, we give the number
# of stash slots needed, and the maximum size needed (in blocks)
out.insert(2, str(next_stash_id) + "\n")
out.insert(3, str(max_stashed_blocks) + "\n")
with open(prefix + ".transfer.list", "wb") as f:
for i in out:
f.write(i)
if self.version >= 2:
self._max_stashed_size = max_stashed_blocks * self.tgt.blocksize
OPTIONS = common.OPTIONS
if OPTIONS.cache_size is not None:
max_allowed = OPTIONS.cache_size * OPTIONS.stash_threshold
print("max stashed blocks: %d (%d bytes), "
"limit: %d bytes (%.2f%%)\n" % (
max_stashed_blocks, self._max_stashed_size, max_allowed,
self._max_stashed_size * 100.0 / max_allowed))
else:
print("max stashed blocks: %d (%d bytes), limit: <unknown>\n" % (
max_stashed_blocks, self._max_stashed_size))
def ReviseStashSize(self):
print("Revising stash size...")
stash_map = {}
# Create the map between a stash and its def/use points. For example, for a
# given stash of (idx, sr), stashes[idx] = (sr, def_cmd, use_cmd).
for xf in self.transfers:
# Command xf defines (stores) all the stashes in stash_before.
for idx, sr in xf.stash_before:
stash_map[idx] = (sr, xf)
# Record all the stashes command xf uses.
for idx, _ in xf.use_stash:
stash_map[idx] += (xf,)
# Compute the maximum blocks available for stash based on /cache size and
# the threshold.
cache_size = common.OPTIONS.cache_size
stash_threshold = common.OPTIONS.stash_threshold
max_allowed = cache_size * stash_threshold / self.tgt.blocksize
stashes = {}
stashed_blocks = 0
new_blocks = 0
free_stash_ids = []
next_stash_id = 0
# Now go through all the commands. Compute the required stash size on the
# fly. If a command requires excess stash than available, it deletes the
# stash by replacing the command that uses the stash with a "new" command
# instead.
for xf in self.transfers:
replaced_cmds = []
# xf.stash_before generates explicit stash commands.
for idx, sr in xf.stash_before:
assert idx not in stashes
if free_stash_ids:
sid = heapq.heappop(free_stash_ids)
else:
sid = next_stash_id
next_stash_id += 1
stashes[idx] = sid
# Check the post-command stashed_blocks.
stashed_blocks_after = stashed_blocks
if self.version == 2:
stashed_blocks_after += sr.size()
else:
sh = self.HashBlocks(self.src, sr)
if sh in stashes:
stashes[sh] += 1
else:
stashes[sh] = 1
stashed_blocks_after += sr.size()
if stashed_blocks_after > max_allowed:
# We cannot stash this one for a later command. Find out the command
# that will use this stash and replace the command with "new".
use_cmd = stash_map[idx][2]
replaced_cmds.append(use_cmd)
print("%10d %9s %s" % (sr.size(), "explicit", use_cmd))
else:
stashed_blocks = stashed_blocks_after
# "move" and "diff" may introduce implicit stashes in BBOTA v3. Prior to
# ComputePatches(), they both have the style of "diff".
if xf.style == "diff" and self.version >= 3:
assert xf.tgt_ranges and xf.src_ranges
if xf.src_ranges.overlaps(xf.tgt_ranges):
if stashed_blocks + xf.src_ranges.size() > max_allowed:
replaced_cmds.append(xf)
print("%10d %9s %s" % (xf.src_ranges.size(), "implicit", xf))
# Replace the commands in replaced_cmds with "new"s.
for cmd in replaced_cmds:
# It no longer uses any commands in "use_stash". Remove the def points
# for all those stashes.
for idx, sr in cmd.use_stash:
def_cmd = stash_map[idx][1]
assert (idx, sr) in def_cmd.stash_before
def_cmd.stash_before.remove((idx, sr))
# Add up blocks that violates space limit and print total number to
# screen later.
new_blocks += cmd.tgt_ranges.size()
cmd.ConvertToNew()
# xf.use_stash generates free commands.
for idx, sr in xf.use_stash:
sid = stashes.pop(idx)
if self.version == 2:
stashed_blocks -= sr.size()
else:
sh = self.HashBlocks(self.src, sr)
assert sh in stashes
stashes[sh] -= 1
if stashes[sh] == 0:
stashed_blocks -= sr.size()
stashes.pop(sh)
heapq.heappush(free_stash_ids, sid)
num_of_bytes = new_blocks * self.tgt.blocksize
print(" Total %d blocks (%d bytes) are packed as new blocks due to "
"insufficient cache size." % (new_blocks, num_of_bytes))
def ComputePatches(self, prefix):
print("Reticulating splines...")
diff_q = []
patch_num = 0
with open(prefix + ".new.dat", "wb") as new_f:
for xf in self.transfers:
if xf.style == "zero":
tgt_size = xf.tgt_ranges.size() * self.tgt.blocksize
print("%10d %10d (%6.2f%%) %7s %s %s" % (
tgt_size, tgt_size, 100.0, xf.style, xf.tgt_name,
str(xf.tgt_ranges)))
elif xf.style == "new":
for piece in self.tgt.ReadRangeSet(xf.tgt_ranges):
new_f.write(piece)
tgt_size = xf.tgt_ranges.size() * self.tgt.blocksize
print("%10d %10d (%6.2f%%) %7s %s %s" % (
tgt_size, tgt_size, 100.0, xf.style,
xf.tgt_name, str(xf.tgt_ranges)))
elif xf.style == "diff":
src = self.src.ReadRangeSet(xf.src_ranges)
tgt = self.tgt.ReadRangeSet(xf.tgt_ranges)
# We can't compare src and tgt directly because they may have
# the same content but be broken up into blocks differently, eg:
#
# ["he", "llo"] vs ["h", "ello"]
#
# We want those to compare equal, ideally without having to
# actually concatenate the strings (these may be tens of
# megabytes).
src_sha1 = sha1()
for p in src:
src_sha1.update(p)
tgt_sha1 = sha1()
tgt_size = 0
for p in tgt:
tgt_sha1.update(p)
tgt_size += len(p)
if src_sha1.digest() == tgt_sha1.digest():
# These are identical; we don't need to generate a patch,
# just issue copy commands on the device.
xf.style = "move"
if xf.src_ranges != xf.tgt_ranges:
print("%10d %10d (%6.2f%%) %7s %s %s (from %s)" % (
tgt_size, tgt_size, 100.0, xf.style,
xf.tgt_name if xf.tgt_name == xf.src_name else (
xf.tgt_name + " (from " + xf.src_name + ")"),
str(xf.tgt_ranges), str(xf.src_ranges)))
else:
# For files in zip format (eg, APKs, JARs, etc.) we would
# like to use imgdiff -z if possible (because it usually
# produces significantly smaller patches than bsdiff).
# This is permissible if:
#
# - imgdiff is not disabled, and
# - the source and target files are monotonic (ie, the
# data is stored with blocks in increasing order), and
# - we haven't removed any blocks from the source set.
#
# If these conditions are satisfied then appending all the
# blocks in the set together in order will produce a valid
# zip file (plus possibly extra zeros in the last block),
# which is what imgdiff needs to operate. (imgdiff is
# fine with extra zeros at the end of the file.)
imgdiff = (not self.disable_imgdiff and xf.intact and
xf.tgt_name.split(".")[-1].lower()
in ("apk", "jar", "zip"))
xf.style = "imgdiff" if imgdiff else "bsdiff"
diff_q.append((tgt_size, src, tgt, xf, patch_num))
patch_num += 1
else:
assert False, "unknown style " + xf.style
if diff_q:
if self.threads > 1:
print("Computing patches (using %d threads)..." % (self.threads,))
else:
print("Computing patches...")
diff_q.sort()
patches = [None] * patch_num
# TODO: Rewrite with multiprocessing.ThreadPool?
lock = threading.Lock()
def diff_worker():
while True:
with lock:
if not diff_q:
return
tgt_size, src, tgt, xf, patchnum = diff_q.pop()
patch = compute_patch(src, tgt, imgdiff=(xf.style == "imgdiff"))
size = len(patch)
with lock:
patches[patchnum] = (patch, xf)
print("%10d %10d (%6.2f%%) %7s %s %s %s" % (
size, tgt_size, size * 100.0 / tgt_size, xf.style,
xf.tgt_name if xf.tgt_name == xf.src_name else (
xf.tgt_name + " (from " + xf.src_name + ")"),
str(xf.tgt_ranges), str(xf.src_ranges)))
threads = [threading.Thread(target=diff_worker)
for _ in range(self.threads)]
for th in threads:
th.start()
while threads:
threads.pop().join()
else:
patches = []
p = 0
with open(prefix + ".patch.dat", "wb") as patch_f:
for patch, xf in patches:
xf.patch_start = p
xf.patch_len = len(patch)
patch_f.write(patch)
p += len(patch)
def AssertSequenceGood(self):
# Simulate the sequences of transfers we will output, and check that:
# - we never read a block after writing it, and
# - we write every block we care about exactly once.
# Start with no blocks having been touched yet.
touched = array.array("B", "\0" * self.tgt.total_blocks)
# Imagine processing the transfers in order.
for xf in self.transfers:
# Check that the input blocks for this transfer haven't yet been touched.
x = xf.src_ranges
if self.version >= 2:
for _, sr in xf.use_stash:
x = x.subtract(sr)
for s, e in x:
# Source image could be larger. Don't check the blocks that are in the
# source image only. Since they are not in 'touched', and won't ever
# be touched.
for i in range(s, min(e, self.tgt.total_blocks)):
assert touched[i] == 0
# Check that the output blocks for this transfer haven't yet
# been touched, and touch all the blocks written by this
# transfer.
for s, e in xf.tgt_ranges:
for i in range(s, e):
assert touched[i] == 0
touched[i] = 1
# Check that we've written every target block.
for s, e in self.tgt.care_map:
for i in range(s, e):
assert touched[i] == 1
def ImproveVertexSequence(self):
print("Improving vertex order...")
# At this point our digraph is acyclic; we reversed any edges that
# were backwards in the heuristically-generated sequence. The
# previously-generated order is still acceptable, but we hope to
# find a better order that needs less memory for stashed data.
# Now we do a topological sort to generate a new vertex order,
# using a greedy algorithm to choose which vertex goes next
# whenever we have a choice.
# Make a copy of the edge set; this copy will get destroyed by the
# algorithm.
for xf in self.transfers:
xf.incoming = xf.goes_after.copy()
xf.outgoing = xf.goes_before.copy()
L = [] # the new vertex order
# S is the set of sources in the remaining graph; we always choose
# the one that leaves the least amount of stashed data after it's
# executed.
S = [(u.NetStashChange(), u.order, u) for u in self.transfers
if not u.incoming]
heapq.heapify(S)
while S:
_, _, xf = heapq.heappop(S)
L.append(xf)
for u in xf.outgoing:
del u.incoming[xf]
if not u.incoming:
heapq.heappush(S, (u.NetStashChange(), u.order, u))
# if this fails then our graph had a cycle.
assert len(L) == len(self.transfers)
self.transfers = L
for i, xf in enumerate(L):
xf.order = i
def RemoveBackwardEdges(self):
print("Removing backward edges...")
in_order = 0
out_of_order = 0
lost_source = 0
for xf in self.transfers:
lost = 0
size = xf.src_ranges.size()
for u in xf.goes_before:
# xf should go before u
if xf.order < u.order:
# it does, hurray!
in_order += 1
else:
# it doesn't, boo. trim the blocks that u writes from xf's
# source, so that xf can go after u.
out_of_order += 1
assert xf.src_ranges.overlaps(u.tgt_ranges)
xf.src_ranges = xf.src_ranges.subtract(u.tgt_ranges)
xf.intact = False
if xf.style == "diff" and not xf.src_ranges:
# nothing left to diff from; treat as new data
xf.style = "new"
lost = size - xf.src_ranges.size()
lost_source += lost
print((" %d/%d dependencies (%.2f%%) were violated; "
"%d source blocks removed.") %
(out_of_order, in_order + out_of_order,
(out_of_order * 100.0 / (in_order + out_of_order))
if (in_order + out_of_order) else 0.0,
lost_source))
def ReverseBackwardEdges(self):
print("Reversing backward edges...")
in_order = 0
out_of_order = 0
stashes = 0
stash_size = 0
for xf in self.transfers:
for u in xf.goes_before.copy():
# xf should go before u
if xf.order < u.order:
# it does, hurray!
in_order += 1
else:
# it doesn't, boo. modify u to stash the blocks that it
# writes that xf wants to read, and then require u to go
# before xf.
out_of_order += 1
overlap = xf.src_ranges.intersect(u.tgt_ranges)
assert overlap
u.stash_before.append((stashes, overlap))
xf.use_stash.append((stashes, overlap))
stashes += 1
stash_size += overlap.size()
# reverse the edge direction; now xf must go after u
del xf.goes_before[u]
del u.goes_after[xf]
xf.goes_after[u] = None # value doesn't matter
u.goes_before[xf] = None
print((" %d/%d dependencies (%.2f%%) were violated; "
"%d source blocks stashed.") %
(out_of_order, in_order + out_of_order,
(out_of_order * 100.0 / (in_order + out_of_order))
if (in_order + out_of_order) else 0.0,
stash_size))
def FindVertexSequence(self):
print("Finding vertex sequence...")
# This is based on "A Fast & Effective Heuristic for the Feedback
# Arc Set Problem" by P. Eades, X. Lin, and W.F. Smyth. Think of
# it as starting with the digraph G and moving all the vertices to
# be on a horizontal line in some order, trying to minimize the
# number of edges that end up pointing to the left. Left-pointing
# edges will get removed to turn the digraph into a DAG. In this
# case each edge has a weight which is the number of source blocks
# we'll lose if that edge is removed; we try to minimize the total
# weight rather than just the number of edges.
# Make a copy of the edge set; this copy will get destroyed by the
# algorithm.
for xf in self.transfers:
xf.incoming = xf.goes_after.copy()
xf.outgoing = xf.goes_before.copy()
xf.score = sum(xf.outgoing.values()) - sum(xf.incoming.values())
# We use an OrderedDict instead of just a set so that the output
# is repeatable; otherwise it would depend on the hash values of
# the transfer objects.
G = OrderedDict()
for xf in self.transfers:
G[xf] = None
s1 = deque() # the left side of the sequence, built from left to right
s2 = deque() # the right side of the sequence, built from right to left
heap = []
for xf in self.transfers:
xf.heap_item = HeapItem(xf)
heap.append(xf.heap_item)
heapq.heapify(heap)
# Use OrderedDict() instead of set() to preserve the insertion order. Need
# to use 'sinks[key] = None' to add key into the set. sinks will look like
# { key1: None, key2: None, ... }.
sinks = OrderedDict.fromkeys(u for u in G if not u.outgoing)
sources = OrderedDict.fromkeys(u for u in G if not u.incoming)
def adjust_score(iu, delta):
iu.score += delta
iu.heap_item.clear()
iu.heap_item = HeapItem(iu)
heapq.heappush(heap, iu.heap_item)
while G:
# Put all sinks at the end of the sequence.
while sinks:
new_sinks = OrderedDict()
for u in sinks:
if u not in G: continue
s2.appendleft(u)
del G[u]
for iu in u.incoming:
adjust_score(iu, -iu.outgoing.pop(u))
if not iu.outgoing:
new_sinks[iu] = None
sinks = new_sinks
# Put all the sources at the beginning of the sequence.
while sources:
new_sources = OrderedDict()
for u in sources:
if u not in G: continue
s1.append(u)
del G[u]
for iu in u.outgoing:
adjust_score(iu, +iu.incoming.pop(u))
if not iu.incoming:
new_sources[iu] = None
sources = new_sources
if not G: break
# Find the "best" vertex to put next. "Best" is the one that
# maximizes the net difference in source blocks saved we get by
# pretending it's a source rather than a sink.
while True:
u = heapq.heappop(heap)
if u and u.item in G:
u = u.item
break
s1.append(u)
del G[u]
for iu in u.outgoing:
adjust_score(iu, +iu.incoming.pop(u))
if not iu.incoming:
sources[iu] = None
for iu in u.incoming:
adjust_score(iu, -iu.outgoing.pop(u))
if not iu.outgoing:
sinks[iu] = None
# Now record the sequence in the 'order' field of each transfer,
# and by rearranging self.transfers to be in the chosen sequence.
new_transfers = []
for x in itertools.chain(s1, s2):
x.order = len(new_transfers)
new_transfers.append(x)
del x.incoming
del x.outgoing
self.transfers = new_transfers
def GenerateDigraph(self):
print("Generating digraph...")
# Each item of source_ranges will be:
# - None, if that block is not used as a source,
# - an ordered set of transfers.
source_ranges = []
for b in self.transfers:
for s, e in b.src_ranges:
if e > len(source_ranges):
source_ranges.extend([None] * (e-len(source_ranges)))
for i in range(s, e):
if source_ranges[i] is None:
source_ranges[i] = OrderedDict.fromkeys([b])
else:
source_ranges[i][b] = None
for a in self.transfers:
intersections = OrderedDict()
for s, e in a.tgt_ranges:
for i in range(s, e):
if i >= len(source_ranges): break
# Add all the Transfers in source_ranges[i] to the (ordered) set.
if source_ranges[i] is not None:
for j in source_ranges[i]:
intersections[j] = None
for b in intersections:
if a is b: continue
# If the blocks written by A are read by B, then B needs to go before A.
i = a.tgt_ranges.intersect(b.src_ranges)
if i:
if b.src_name == "__ZERO":
# the cost of removing source blocks for the __ZERO domain
# is (nearly) zero.
size = 0
else:
size = i.size()
b.goes_before[a] = size
a.goes_after[b] = size
def FindTransfers(self):
"""Parse the file_map to generate all the transfers."""
def AddSplitTransfers(tgt_name, src_name, tgt_ranges, src_ranges,
style, by_id):
"""Add one or multiple Transfer()s by splitting large files.
For BBOTA v3, we need to stash source blocks for resumable feature.
However, with the growth of file size and the shrink of the cache
partition source blocks are too large to be stashed. If a file occupies
too many blocks, we split it into smaller pieces by getting multiple
Transfer()s.
The downside is that after splitting, we may increase the package size
since the split pieces don't align well. According to our experiments,
1/8 of the cache size as the per-piece limit appears to be optimal.
Compared to the fixed 1024-block limit, it reduces the overall package
size by 30% for volantis, and 20% for angler and bullhead."""
# Possibly split large files into smaller chunks.
pieces = 0
cache_size = common.OPTIONS.cache_size
split_threshold = 0.125
max_blocks_per_transfer = int(cache_size * split_threshold /
self.tgt.blocksize)
# Change nothing for small files.
if (tgt_ranges.size() <= max_blocks_per_transfer and
src_ranges.size() <= max_blocks_per_transfer):
Transfer(tgt_name, src_name, tgt_ranges, src_ranges, style, by_id)
return
while (tgt_ranges.size() > max_blocks_per_transfer and
src_ranges.size() > max_blocks_per_transfer):
tgt_split_name = "%s-%d" % (tgt_name, pieces)
src_split_name = "%s-%d" % (src_name, pieces)
tgt_first = tgt_ranges.first(max_blocks_per_transfer)
src_first = src_ranges.first(max_blocks_per_transfer)
Transfer(tgt_split_name, src_split_name, tgt_first, src_first, style,
by_id)
tgt_ranges = tgt_ranges.subtract(tgt_first)
src_ranges = src_ranges.subtract(src_first)
pieces += 1
# Handle remaining blocks.
if tgt_ranges.size() or src_ranges.size():
# Must be both non-empty.
assert tgt_ranges.size() and src_ranges.size()
tgt_split_name = "%s-%d" % (tgt_name, pieces)
src_split_name = "%s-%d" % (src_name, pieces)
Transfer(tgt_split_name, src_split_name, tgt_ranges, src_ranges, style,
by_id)
def AddTransfer(tgt_name, src_name, tgt_ranges, src_ranges, style, by_id,
split=False):
"""Wrapper function for adding a Transfer()."""
# We specialize diff transfers only (which covers bsdiff/imgdiff/move);
# otherwise add the Transfer() as is.
if style != "diff" or not split:
Transfer(tgt_name, src_name, tgt_ranges, src_ranges, style, by_id)
return
# Handle .odex files specially to analyze the block-wise difference. If
# most of the blocks are identical with only few changes (e.g. header),
# we will patch the changed blocks only. This avoids stashing unchanged
# blocks while patching. We limit the analysis to files without size
# changes only. This is to avoid sacrificing the OTA generation cost too
# much.
if (tgt_name.split(".")[-1].lower() == 'odex' and
tgt_ranges.size() == src_ranges.size()):
# 0.5 threshold can be further tuned. The tradeoff is: if only very
# few blocks remain identical, we lose the opportunity to use imgdiff
# that may have better compression ratio than bsdiff.
crop_threshold = 0.5
tgt_skipped = RangeSet()
src_skipped = RangeSet()
tgt_size = tgt_ranges.size()
tgt_changed = 0
for src_block, tgt_block in zip(src_ranges.next_item(),
tgt_ranges.next_item()):
src_rs = RangeSet(str(src_block))
tgt_rs = RangeSet(str(tgt_block))
if self.src.ReadRangeSet(src_rs) == self.tgt.ReadRangeSet(tgt_rs):
tgt_skipped = tgt_skipped.union(tgt_rs)
src_skipped = src_skipped.union(src_rs)
else:
tgt_changed += tgt_rs.size()
# Terminate early if no clear sign of benefits.
if tgt_changed > tgt_size * crop_threshold:
break
if tgt_changed < tgt_size * crop_threshold:
assert tgt_changed + tgt_skipped.size() == tgt_size
print('%10d %10d (%6.2f%%) %s' % (tgt_skipped.size(), tgt_size,
tgt_skipped.size() * 100.0 / tgt_size, tgt_name))
AddSplitTransfers(
"%s-skipped" % (tgt_name,),
"%s-skipped" % (src_name,),
tgt_skipped, src_skipped, style, by_id)
# Intentionally change the file extension to avoid being imgdiff'd as
# the files are no longer in their original format.
tgt_name = "%s-cropped" % (tgt_name,)
src_name = "%s-cropped" % (src_name,)
tgt_ranges = tgt_ranges.subtract(tgt_skipped)
src_ranges = src_ranges.subtract(src_skipped)
# Possibly having no changed blocks.
if not tgt_ranges:
return
# Add the transfer(s).
AddSplitTransfers(
tgt_name, src_name, tgt_ranges, src_ranges, style, by_id)
print("Finding transfers...")
empty = RangeSet()
for tgt_fn, tgt_ranges in self.tgt.file_map.items():
if tgt_fn == "__ZERO":
# the special "__ZERO" domain is all the blocks not contained
# in any file and that are filled with zeros. We have a
# special transfer style for zero blocks.
src_ranges = self.src.file_map.get("__ZERO", empty)
AddTransfer(tgt_fn, "__ZERO", tgt_ranges, src_ranges,
"zero", self.transfers)
continue
elif tgt_fn == "__COPY":
# "__COPY" domain includes all the blocks not contained in any
# file and that need to be copied unconditionally to the target.
AddTransfer(tgt_fn, None, tgt_ranges, empty, "new", self.transfers)
continue
elif tgt_fn in self.src.file_map:
# Look for an exact pathname match in the source.
AddTransfer(tgt_fn, tgt_fn, tgt_ranges, self.src.file_map[tgt_fn],
"diff", self.transfers, self.version >= 3)
continue
b = os.path.basename(tgt_fn)
if b in self.src_basenames:
# Look for an exact basename match in the source.
src_fn = self.src_basenames[b]
AddTransfer(tgt_fn, src_fn, tgt_ranges, self.src.file_map[src_fn],
"diff", self.transfers, self.version >= 3)
continue
b = re.sub("[0-9]+", "#", b)
if b in self.src_numpatterns:
# Look for a 'number pattern' match (a basename match after
# all runs of digits are replaced by "#"). (This is useful
# for .so files that contain version numbers in the filename
# that get bumped.)
src_fn = self.src_numpatterns[b]
AddTransfer(tgt_fn, src_fn, tgt_ranges, self.src.file_map[src_fn],
"diff", self.transfers, self.version >= 3)
continue
AddTransfer(tgt_fn, None, tgt_ranges, empty, "new", self.transfers)
def AbbreviateSourceNames(self):
for k in self.src.file_map.keys():
b = os.path.basename(k)
self.src_basenames[b] = k
b = re.sub("[0-9]+", "#", b)
self.src_numpatterns[b] = k
@staticmethod
def AssertPartition(total, seq):
"""Assert that all the RangeSets in 'seq' form a partition of the
'total' RangeSet (ie, they are nonintersecting and their union
equals 'total')."""
so_far = RangeSet()
for i in seq:
assert not so_far.overlaps(i)
so_far = so_far.union(i)
assert so_far == total
def test_extend(self):
self.assertEqual(RangeSet("0-9").extend(1), RangeSet("0-10"))
self.assertEqual(RangeSet("10-19").extend(15), RangeSet("0-34"))
self.assertEqual(
RangeSet("10-19 30-39").extend(4), RangeSet("6-23 26-43"))
self.assertEqual(RangeSet("10-19 30-39").extend(10), RangeSet("0-49"))
def WriteTransfers(self, prefix):
def WriteSplitTransfers(out, style, target_blocks):
"""Limit the size of operand in command 'new' and 'zero' to 1024 blocks.
This prevents the target size of one command from being too large; and
might help to avoid fsync errors on some devices."""
assert (style == "new" or style == "zero")
blocks_limit = 1024
total = 0
while target_blocks:
blocks_to_write = target_blocks.first(blocks_limit)
out.append("%s %s\n" % (style, blocks_to_write.to_string_raw()))
total += blocks_to_write.size()
target_blocks = target_blocks.subtract(blocks_to_write)
return total
out = []
total = 0
stashes = {}
stashed_blocks = 0
max_stashed_blocks = 0
free_stash_ids = []
next_stash_id = 0
for xf in self.transfers:
if self.version < 2:
assert not xf.stash_before
assert not xf.use_stash
for s, sr in xf.stash_before:
assert s not in stashes
if free_stash_ids:
sid = heapq.heappop(free_stash_ids)
else:
sid = next_stash_id
next_stash_id += 1
stashes[s] = sid
if self.version == 2:
stashed_blocks += sr.size()
out.append("stash %d %s\n" % (sid, sr.to_string_raw()))
else:
sh = self.HashBlocks(self.src, sr)
if sh in stashes:
stashes[sh] += 1
else:
stashes[sh] = 1
stashed_blocks += sr.size()
self.touched_src_ranges = self.touched_src_ranges.union(sr)
out.append("stash %s %s\n" % (sh, sr.to_string_raw()))
if stashed_blocks > max_stashed_blocks:
max_stashed_blocks = stashed_blocks
free_string = []
free_size = 0
if self.version == 1:
src_str = xf.src_ranges.to_string_raw() if xf.src_ranges else ""
elif self.version >= 2:
# <# blocks> <src ranges>
# OR
# <# blocks> <src ranges> <src locs> <stash refs...>
# OR
# <# blocks> - <stash refs...>
size = xf.src_ranges.size()
src_str = [str(size)]
unstashed_src_ranges = xf.src_ranges
mapped_stashes = []
for s, sr in xf.use_stash:
# TODO: We don't need 'sid' (nor free_stash_ids) in BBOTA v3+.
sid = stashes.pop(s)
unstashed_src_ranges = unstashed_src_ranges.subtract(sr)
sh = self.HashBlocks(self.src, sr)
sr = xf.src_ranges.map_within(sr)
mapped_stashes.append(sr)
if self.version == 2:
src_str.append("%d:%s" % (sid, sr.to_string_raw()))
# A stash will be used only once. We need to free the stash
# immediately after the use, instead of waiting for the automatic
# clean-up at the end. Because otherwise it may take up extra space
# and lead to OTA failures.
# Bug: 23119955
free_string.append("free %d\n" % (sid,))
free_size += sr.size()
else:
assert sh in stashes
src_str.append("%s:%s" % (sh, sr.to_string_raw()))
stashes[sh] -= 1
if stashes[sh] == 0:
free_size += sr.size()
free_string.append("free %s\n" % (sh,))
stashes.pop(sh)
heapq.heappush(free_stash_ids, sid)
if unstashed_src_ranges:
src_str.insert(1, unstashed_src_ranges.to_string_raw())
if xf.use_stash:
mapped_unstashed = xf.src_ranges.map_within(unstashed_src_ranges)
src_str.insert(2, mapped_unstashed.to_string_raw())
mapped_stashes.append(mapped_unstashed)
self.AssertPartition(RangeSet(data=(0, size)), mapped_stashes)
else:
src_str.insert(1, "-")
self.AssertPartition(RangeSet(data=(0, size)), mapped_stashes)
src_str = " ".join(src_str)
# all versions:
# zero <rangeset>
# new <rangeset>
# erase <rangeset>
#
# version 1:
# bsdiff patchstart patchlen <src rangeset> <tgt rangeset>
# imgdiff patchstart patchlen <src rangeset> <tgt rangeset>
# move <src rangeset> <tgt rangeset>
#
# version 2:
# bsdiff patchstart patchlen <tgt rangeset> <src_str>
# imgdiff patchstart patchlen <tgt rangeset> <src_str>
# move <tgt rangeset> <src_str>
#
# version 3:
# bsdiff patchstart patchlen srchash tgthash <tgt rangeset> <src_str>
# imgdiff patchstart patchlen srchash tgthash <tgt rangeset> <src_str>
# move hash <tgt rangeset> <src_str>
tgt_size = xf.tgt_ranges.size()
if xf.style == "new":
assert xf.tgt_ranges
assert tgt_size == WriteSplitTransfers(out, xf.style, xf.tgt_ranges)
total += tgt_size
elif xf.style == "move":
assert xf.tgt_ranges
assert xf.src_ranges.size() == tgt_size
if xf.src_ranges != xf.tgt_ranges:
if self.version == 1:
out.append("%s %s %s\n" % (
xf.style,
xf.src_ranges.to_string_raw(), xf.tgt_ranges.to_string_raw()))
elif self.version == 2:
out.append("%s %s %s\n" % (
xf.style,
xf.tgt_ranges.to_string_raw(), src_str))
elif self.version >= 3:
# take into account automatic stashing of overlapping blocks
if xf.src_ranges.overlaps(xf.tgt_ranges):
temp_stash_usage = stashed_blocks + xf.src_ranges.size()
if temp_stash_usage > max_stashed_blocks:
max_stashed_blocks = temp_stash_usage
self.touched_src_ranges = self.touched_src_ranges.union(
xf.src_ranges)
out.append("%s %s %s %s\n" % (
xf.style,
self.HashBlocks(self.tgt, xf.tgt_ranges),
xf.tgt_ranges.to_string_raw(), src_str))
total += tgt_size
elif xf.style in ("bsdiff", "imgdiff"):
assert xf.tgt_ranges
assert xf.src_ranges
if self.version == 1:
out.append("%s %d %d %s %s\n" % (
xf.style, xf.patch_start, xf.patch_len,
xf.src_ranges.to_string_raw(), xf.tgt_ranges.to_string_raw()))
elif self.version == 2:
out.append("%s %d %d %s %s\n" % (
xf.style, xf.patch_start, xf.patch_len,
xf.tgt_ranges.to_string_raw(), src_str))
elif self.version >= 3:
# take into account automatic stashing of overlapping blocks
if xf.src_ranges.overlaps(xf.tgt_ranges):
temp_stash_usage = stashed_blocks + xf.src_ranges.size()
if temp_stash_usage > max_stashed_blocks:
max_stashed_blocks = temp_stash_usage
self.touched_src_ranges = self.touched_src_ranges.union(
xf.src_ranges)
out.append("%s %d %d %s %s %s %s\n" % (
xf.style,
xf.patch_start, xf.patch_len,
self.HashBlocks(self.src, xf.src_ranges),
self.HashBlocks(self.tgt, xf.tgt_ranges),
xf.tgt_ranges.to_string_raw(), src_str))
total += tgt_size
elif xf.style == "zero":
assert xf.tgt_ranges
to_zero = xf.tgt_ranges.subtract(xf.src_ranges)
assert WriteSplitTransfers(out, xf.style, to_zero) == to_zero.size()
total += to_zero.size()
else:
raise ValueError("unknown transfer style '%s'\n" % xf.style)
if free_string:
out.append("".join(free_string))
stashed_blocks -= free_size
if self.version >= 2 and common.OPTIONS.cache_size is not None:
# Sanity check: abort if we're going to need more stash space than
# the allowed size (cache_size * threshold). There are two purposes
# of having a threshold here. a) Part of the cache may have been
# occupied by some recovery logs. b) It will buy us some time to deal
# with the oversize issue.
cache_size = common.OPTIONS.cache_size
stash_threshold = common.OPTIONS.stash_threshold
max_allowed = cache_size * stash_threshold
assert max_stashed_blocks * self.tgt.blocksize < max_allowed, \
'Stash size %d (%d * %d) exceeds the limit %d (%d * %.2f)' % (
max_stashed_blocks * self.tgt.blocksize, max_stashed_blocks,
self.tgt.blocksize, max_allowed, cache_size,
stash_threshold)
if self.version >= 3:
self.touched_src_sha1 = self.HashBlocks(
self.src, self.touched_src_ranges)
# Zero out extended blocks as a workaround for bug 20881595.
if self.tgt.extended:
assert (WriteSplitTransfers(out, "zero", self.tgt.extended) ==
self.tgt.extended.size())
total += self.tgt.extended.size()
# We erase all the blocks on the partition that a) don't contain useful
# data in the new image; b) will not be touched by dm-verity. Out of those
# blocks, we erase the ones that won't be used in this update at the
# beginning of an update. The rest would be erased at the end. This is to
# work around the eMMC issue observed on some devices, which may otherwise
# get starving for clean blocks and thus fail the update. (b/28347095)
all_tgt = RangeSet(data=(0, self.tgt.total_blocks))
all_tgt_minus_extended = all_tgt.subtract(self.tgt.extended)
new_dontcare = all_tgt_minus_extended.subtract(self.tgt.care_map)
erase_first = new_dontcare.subtract(self.touched_src_ranges)
if erase_first:
out.insert(0, "erase %s\n" % (erase_first.to_string_raw(),))
erase_last = new_dontcare.subtract(erase_first)
if erase_last:
out.append("erase %s\n" % (erase_last.to_string_raw(),))
out.insert(0, "%d\n" % (self.version,)) # format version number
out.insert(1, "%d\n" % (total,))
if self.version >= 2:
# version 2 only: after the total block count, we give the number
# of stash slots needed, and the maximum size needed (in blocks)
out.insert(2, str(next_stash_id) + "\n")
out.insert(3, str(max_stashed_blocks) + "\n")
with open(prefix + ".transfer.list", "wb") as f:
for i in out:
f.write(i)
if self.version >= 2:
self._max_stashed_size = max_stashed_blocks * self.tgt.blocksize
OPTIONS = common.OPTIONS
if OPTIONS.cache_size is not None:
max_allowed = OPTIONS.cache_size * OPTIONS.stash_threshold
print("max stashed blocks: %d (%d bytes), "
"limit: %d bytes (%.2f%%)\n" % (
max_stashed_blocks, self._max_stashed_size, max_allowed,
self._max_stashed_size * 100.0 / max_allowed))
else:
print("max stashed blocks: %d (%d bytes), limit: <unknown>\n" % (
max_stashed_blocks, self._max_stashed_size))
class BlockImageDiff(object):
def __init__(self, tgt, src=None, threads=None, version=4,
disable_imgdiff=False):
if threads is None:
threads = multiprocessing.cpu_count() // 2
if threads == 0:
threads = 1
self.threads = threads
self.version = version
self.transfers = []
self.src_basenames = {}
self.src_numpatterns = {}
self._max_stashed_size = 0
self.touched_src_ranges = RangeSet()
self.touched_src_sha1 = None
self.disable_imgdiff = disable_imgdiff
assert version in (3, 4)
self.tgt = tgt
if src is None:
src = EmptyImage()
self.src = src
# The updater code that installs the patch always uses 4k blocks.
assert tgt.blocksize == 4096
assert src.blocksize == 4096
# The range sets in each filemap should comprise a partition of
# the care map.
self.AssertPartition(src.care_map, src.file_map.values())
self.AssertPartition(tgt.care_map, tgt.file_map.values())
@property
def max_stashed_size(self):
return self._max_stashed_size
def Compute(self, prefix):
# When looking for a source file to use as the diff input for a
# target file, we try:
# 1) an exact path match if available, otherwise
# 2) a exact basename match if available, otherwise
# 3) a basename match after all runs of digits are replaced by
# "#" if available, otherwise
# 4) we have no source for this target.
self.AbbreviateSourceNames()
self.FindTransfers()
# Find the ordering dependencies among transfers (this is O(n^2)
# in the number of transfers).
self.GenerateDigraph()
# Find a sequence of transfers that satisfies as many ordering
# dependencies as possible (heuristically).
self.FindVertexSequence()
# Fix up the ordering dependencies that the sequence didn't
# satisfy.
self.ReverseBackwardEdges()
self.ImproveVertexSequence()
# Ensure the runtime stash size is under the limit.
if common.OPTIONS.cache_size is not None:
self.ReviseStashSize()
# Double-check our work.
self.AssertSequenceGood()
self.ComputePatches(prefix)
self.WriteTransfers(prefix)
def WriteTransfers(self, prefix):
def WriteSplitTransfers(out, style, target_blocks):
"""Limit the size of operand in command 'new' and 'zero' to 1024 blocks.
This prevents the target size of one command from being too large; and
might help to avoid fsync errors on some devices."""
assert style == "new" or style == "zero"
blocks_limit = 1024
total = 0
while target_blocks:
blocks_to_write = target_blocks.first(blocks_limit)
out.append("%s %s\n" % (style, blocks_to_write.to_string_raw()))
total += blocks_to_write.size()
target_blocks = target_blocks.subtract(blocks_to_write)
return total
out = []
total = 0
# In BBOTA v3+, it uses the hash of the stashed blocks as the stash slot
# id. 'stashes' records the map from 'hash' to the ref count. The stash
# will be freed only if the count decrements to zero.
stashes = {}
stashed_blocks = 0
max_stashed_blocks = 0
for xf in self.transfers:
for _, sr in xf.stash_before:
sh = self.src.RangeSha1(sr)
if sh in stashes:
stashes[sh] += 1
else:
stashes[sh] = 1
stashed_blocks += sr.size()
self.touched_src_ranges = self.touched_src_ranges.union(sr)
out.append("stash %s %s\n" % (sh, sr.to_string_raw()))
if stashed_blocks > max_stashed_blocks:
max_stashed_blocks = stashed_blocks
free_string = []
free_size = 0
# <# blocks> <src ranges>
# OR
# <# blocks> <src ranges> <src locs> <stash refs...>
# OR
# <# blocks> - <stash refs...>
size = xf.src_ranges.size()
src_str = [str(size)]
unstashed_src_ranges = xf.src_ranges
mapped_stashes = []
for _, sr in xf.use_stash:
unstashed_src_ranges = unstashed_src_ranges.subtract(sr)
sh = self.src.RangeSha1(sr)
sr = xf.src_ranges.map_within(sr)
mapped_stashes.append(sr)
assert sh in stashes
src_str.append("%s:%s" % (sh, sr.to_string_raw()))
stashes[sh] -= 1
if stashes[sh] == 0:
free_string.append("free %s\n" % (sh,))
free_size += sr.size()
stashes.pop(sh)
if unstashed_src_ranges:
src_str.insert(1, unstashed_src_ranges.to_string_raw())
if xf.use_stash:
mapped_unstashed = xf.src_ranges.map_within(unstashed_src_ranges)
src_str.insert(2, mapped_unstashed.to_string_raw())
mapped_stashes.append(mapped_unstashed)
self.AssertPartition(RangeSet(data=(0, size)), mapped_stashes)
else:
src_str.insert(1, "-")
self.AssertPartition(RangeSet(data=(0, size)), mapped_stashes)
src_str = " ".join(src_str)
# version 3+:
# zero <rangeset>
# new <rangeset>
# erase <rangeset>
# bsdiff patchstart patchlen srchash tgthash <tgt rangeset> <src_str>
# imgdiff patchstart patchlen srchash tgthash <tgt rangeset> <src_str>
# move hash <tgt rangeset> <src_str>
tgt_size = xf.tgt_ranges.size()
if xf.style == "new":
assert xf.tgt_ranges
assert tgt_size == WriteSplitTransfers(out, xf.style, xf.tgt_ranges)
total += tgt_size
elif xf.style == "move":
assert xf.tgt_ranges
assert xf.src_ranges.size() == tgt_size
if xf.src_ranges != xf.tgt_ranges:
# take into account automatic stashing of overlapping blocks
if xf.src_ranges.overlaps(xf.tgt_ranges):
temp_stash_usage = stashed_blocks + xf.src_ranges.size()
if temp_stash_usage > max_stashed_blocks:
max_stashed_blocks = temp_stash_usage
self.touched_src_ranges = self.touched_src_ranges.union(
xf.src_ranges)
out.append("%s %s %s %s\n" % (
xf.style,
xf.tgt_sha1,
xf.tgt_ranges.to_string_raw(), src_str))
total += tgt_size
elif xf.style in ("bsdiff", "imgdiff"):
assert xf.tgt_ranges
assert xf.src_ranges
# take into account automatic stashing of overlapping blocks
if xf.src_ranges.overlaps(xf.tgt_ranges):
temp_stash_usage = stashed_blocks + xf.src_ranges.size()
if temp_stash_usage > max_stashed_blocks:
max_stashed_blocks = temp_stash_usage
self.touched_src_ranges = self.touched_src_ranges.union(xf.src_ranges)
out.append("%s %d %d %s %s %s %s\n" % (
xf.style,
xf.patch_start, xf.patch_len,
xf.src_sha1,
xf.tgt_sha1,
xf.tgt_ranges.to_string_raw(), src_str))
total += tgt_size
elif xf.style == "zero":
assert xf.tgt_ranges
to_zero = xf.tgt_ranges.subtract(xf.src_ranges)
assert WriteSplitTransfers(out, xf.style, to_zero) == to_zero.size()
total += to_zero.size()
else:
raise ValueError("unknown transfer style '%s'\n" % xf.style)
if free_string:
out.append("".join(free_string))
stashed_blocks -= free_size
if common.OPTIONS.cache_size is not None:
# Sanity check: abort if we're going to need more stash space than
# the allowed size (cache_size * threshold). There are two purposes
# of having a threshold here. a) Part of the cache may have been
# occupied by some recovery logs. b) It will buy us some time to deal
# with the oversize issue.
cache_size = common.OPTIONS.cache_size
stash_threshold = common.OPTIONS.stash_threshold
max_allowed = cache_size * stash_threshold
assert max_stashed_blocks * self.tgt.blocksize <= max_allowed, \
'Stash size %d (%d * %d) exceeds the limit %d (%d * %.2f)' % (
max_stashed_blocks * self.tgt.blocksize, max_stashed_blocks,
self.tgt.blocksize, max_allowed, cache_size,
stash_threshold)
self.touched_src_sha1 = self.src.RangeSha1(self.touched_src_ranges)
# Zero out extended blocks as a workaround for bug 20881595.
if self.tgt.extended:
assert (WriteSplitTransfers(out, "zero", self.tgt.extended) ==
self.tgt.extended.size())
total += self.tgt.extended.size()
# We erase all the blocks on the partition that a) don't contain useful
# data in the new image; b) will not be touched by dm-verity. Out of those
# blocks, we erase the ones that won't be used in this update at the
# beginning of an update. The rest would be erased at the end. This is to
# work around the eMMC issue observed on some devices, which may otherwise
# get starving for clean blocks and thus fail the update. (b/28347095)
all_tgt = RangeSet(data=(0, self.tgt.total_blocks))
all_tgt_minus_extended = all_tgt.subtract(self.tgt.extended)
new_dontcare = all_tgt_minus_extended.subtract(self.tgt.care_map)
erase_first = new_dontcare.subtract(self.touched_src_ranges)
if erase_first:
out.insert(0, "erase %s\n" % (erase_first.to_string_raw(),))
erase_last = new_dontcare.subtract(erase_first)
if erase_last:
out.append("erase %s\n" % (erase_last.to_string_raw(),))
out.insert(0, "%d\n" % (self.version,)) # format version number
out.insert(1, "%d\n" % (total,))
# v3+: the number of stash slots is unused.
out.insert(2, "0\n")
out.insert(3, str(max_stashed_blocks) + "\n")
with open(prefix + ".transfer.list", "wb") as f:
for i in out:
f.write(i)
self._max_stashed_size = max_stashed_blocks * self.tgt.blocksize
OPTIONS = common.OPTIONS
if OPTIONS.cache_size is not None:
max_allowed = OPTIONS.cache_size * OPTIONS.stash_threshold
print("max stashed blocks: %d (%d bytes), "
"limit: %d bytes (%.2f%%)\n" % (
max_stashed_blocks, self._max_stashed_size, max_allowed,
self._max_stashed_size * 100.0 / max_allowed))
else:
print("max stashed blocks: %d (%d bytes), limit: <unknown>\n" % (
max_stashed_blocks, self._max_stashed_size))
def ReviseStashSize(self):
print("Revising stash size...")
stash_map = {}
# Create the map between a stash and its def/use points. For example, for a
# given stash of (raw_id, sr), stash_map[raw_id] = (sr, def_cmd, use_cmd).
for xf in self.transfers:
# Command xf defines (stores) all the stashes in stash_before.
for stash_raw_id, sr in xf.stash_before:
stash_map[stash_raw_id] = (sr, xf)
# Record all the stashes command xf uses.
for stash_raw_id, _ in xf.use_stash:
stash_map[stash_raw_id] += (xf,)
# Compute the maximum blocks available for stash based on /cache size and
# the threshold.
cache_size = common.OPTIONS.cache_size
stash_threshold = common.OPTIONS.stash_threshold
max_allowed = cache_size * stash_threshold / self.tgt.blocksize
# See the comments for 'stashes' in WriteTransfers().
stashes = {}
stashed_blocks = 0
new_blocks = 0
# Now go through all the commands. Compute the required stash size on the
# fly. If a command requires excess stash than available, it deletes the
# stash by replacing the command that uses the stash with a "new" command
# instead.
for xf in self.transfers:
replaced_cmds = []
# xf.stash_before generates explicit stash commands.
for stash_raw_id, sr in xf.stash_before:
# Check the post-command stashed_blocks.
stashed_blocks_after = stashed_blocks
sh = self.src.RangeSha1(sr)
if sh not in stashes:
stashed_blocks_after += sr.size()
if stashed_blocks_after > max_allowed:
# We cannot stash this one for a later command. Find out the command
# that will use this stash and replace the command with "new".
use_cmd = stash_map[stash_raw_id][2]
replaced_cmds.append(use_cmd)
print("%10d %9s %s" % (sr.size(), "explicit", use_cmd))
else:
# Update the stashes map.
if sh in stashes:
stashes[sh] += 1
else:
stashes[sh] = 1
stashed_blocks = stashed_blocks_after
# "move" and "diff" may introduce implicit stashes in BBOTA v3. Prior to
# ComputePatches(), they both have the style of "diff".
if xf.style == "diff":
assert xf.tgt_ranges and xf.src_ranges
if xf.src_ranges.overlaps(xf.tgt_ranges):
if stashed_blocks + xf.src_ranges.size() > max_allowed:
replaced_cmds.append(xf)
print("%10d %9s %s" % (xf.src_ranges.size(), "implicit", xf))
# Replace the commands in replaced_cmds with "new"s.
for cmd in replaced_cmds:
# It no longer uses any commands in "use_stash". Remove the def points
# for all those stashes.
for stash_raw_id, sr in cmd.use_stash:
def_cmd = stash_map[stash_raw_id][1]
assert (stash_raw_id, sr) in def_cmd.stash_before
def_cmd.stash_before.remove((stash_raw_id, sr))
# Add up blocks that violates space limit and print total number to
# screen later.
new_blocks += cmd.tgt_ranges.size()
cmd.ConvertToNew()
# xf.use_stash may generate free commands.
for _, sr in xf.use_stash:
sh = self.src.RangeSha1(sr)
assert sh in stashes
stashes[sh] -= 1
if stashes[sh] == 0:
stashed_blocks -= sr.size()
stashes.pop(sh)
num_of_bytes = new_blocks * self.tgt.blocksize
print(" Total %d blocks (%d bytes) are packed as new blocks due to "
"insufficient cache size." % (new_blocks, num_of_bytes))
return new_blocks
def ComputePatches(self, prefix):
print("Reticulating splines...")
diff_queue = []
patch_num = 0
with open(prefix + ".new.dat", "wb") as new_f:
for index, xf in enumerate(self.transfers):
if xf.style == "zero":
tgt_size = xf.tgt_ranges.size() * self.tgt.blocksize
print("%10d %10d (%6.2f%%) %7s %s" % (
tgt_size, tgt_size, 100.0, xf.style, xf.tgt_name))
elif xf.style == "new":
self.tgt.WriteRangeDataToFd(xf.tgt_ranges, new_f)
tgt_size = xf.tgt_ranges.size() * self.tgt.blocksize
print("%10d %10d (%6.2f%%) %7s %s" % (
tgt_size, tgt_size, 100.0, xf.style,
xf.tgt_name))
elif xf.style == "diff":
# We can't compare src and tgt directly because they may have
# the same content but be broken up into blocks differently, eg:
#
# ["he", "llo"] vs ["h", "ello"]
#
# We want those to compare equal, ideally without having to
# actually concatenate the strings (these may be tens of
# megabytes).
if xf.src_sha1 == xf.tgt_sha1:
# These are identical; we don't need to generate a patch,
# just issue copy commands on the device.
xf.style = "move"
tgt_size = xf.tgt_ranges.size() * self.tgt.blocksize
if xf.src_ranges != xf.tgt_ranges:
print("%10d %10d (%6.2f%%) %7s %s %s (from %s)" % (
tgt_size, tgt_size, 100.0, xf.style,
xf.tgt_name if xf.tgt_name == xf.src_name else (
xf.tgt_name + " (from " + xf.src_name + ")"),
str(xf.tgt_ranges), str(xf.src_ranges)))
else:
# For files in zip format (eg, APKs, JARs, etc.) we would
# like to use imgdiff -z if possible (because it usually
# produces significantly smaller patches than bsdiff).
# This is permissible if:
#
# - imgdiff is not disabled, and
# - the source and target files are monotonic (ie, the
# data is stored with blocks in increasing order), and
# - we haven't removed any blocks from the source set.
#
# If these conditions are satisfied then appending all the
# blocks in the set together in order will produce a valid
# zip file (plus possibly extra zeros in the last block),
# which is what imgdiff needs to operate. (imgdiff is
# fine with extra zeros at the end of the file.)
imgdiff = (not self.disable_imgdiff and xf.intact and
xf.tgt_name.split(".")[-1].lower()
in ("apk", "jar", "zip"))
xf.style = "imgdiff" if imgdiff else "bsdiff"
diff_queue.append((index, imgdiff, patch_num))
patch_num += 1
else:
assert False, "unknown style " + xf.style
if diff_queue:
if self.threads > 1:
print("Computing patches (using %d threads)..." % (self.threads,))
else:
print("Computing patches...")
diff_total = len(diff_queue)
patches = [None] * diff_total
error_messages = []
if sys.stdout.isatty():
global diff_done
diff_done = 0
# Using multiprocessing doesn't give additional benefits, due to the
# pattern of the code. The diffing work is done by subprocess.call, which
# already runs in a separate process (not affected much by the GIL -
# Global Interpreter Lock). Using multiprocess also requires either a)
# writing the diff input files in the main process before forking, or b)
# reopening the image file (SparseImage) in the worker processes. Doing
# neither of them further improves the performance.
lock = threading.Lock()
def diff_worker():
while True:
with lock:
if not diff_queue:
return
xf_index, imgdiff, patch_index = diff_queue.pop()
xf = self.transfers[xf_index]
src_ranges = xf.src_ranges
tgt_ranges = xf.tgt_ranges
# Needs lock since WriteRangeDataToFd() is stateful (calling seek).
with lock:
src_file = common.MakeTempFile(prefix="src-")
with open(src_file, "wb") as fd:
self.src.WriteRangeDataToFd(src_ranges, fd)
tgt_file = common.MakeTempFile(prefix="tgt-")
with open(tgt_file, "wb") as fd:
self.tgt.WriteRangeDataToFd(tgt_ranges, fd)
try:
patch = compute_patch(src_file, tgt_file, imgdiff)
except ValueError as e:
with lock:
error_messages.append(
"Failed to generate %s for %s: tgt=%s, src=%s:\n%s" % (
"imgdiff" if imgdiff else "bsdiff",
xf.tgt_name if xf.tgt_name == xf.src_name else
xf.tgt_name + " (from " + xf.src_name + ")",
xf.tgt_ranges, xf.src_ranges, e.message))
with lock:
patches[patch_index] = (xf_index, patch)
if sys.stdout.isatty():
global diff_done
diff_done += 1
progress = diff_done * 100 / diff_total
# '\033[K' is to clear to EOL.
print(' [%d%%] %s\033[K' % (progress, xf.tgt_name), end='\r')
sys.stdout.flush()
threads = [threading.Thread(target=diff_worker)
for _ in range(self.threads)]
for th in threads:
th.start()
while threads:
threads.pop().join()
if sys.stdout.isatty():
print('\n')
if error_messages:
print('\n'.join(error_messages))
sys.exit(1)
else:
patches = []
offset = 0
with open(prefix + ".patch.dat", "wb") as patch_fd:
for index, patch in patches:
xf = self.transfers[index]
xf.patch_len = len(patch)
xf.patch_start = offset
offset += xf.patch_len
patch_fd.write(patch)
if common.OPTIONS.verbose:
tgt_size = xf.tgt_ranges.size() * self.tgt.blocksize
print("%10d %10d (%6.2f%%) %7s %s %s %s" % (
xf.patch_len, tgt_size, xf.patch_len * 100.0 / tgt_size,
xf.style,
xf.tgt_name if xf.tgt_name == xf.src_name else (
xf.tgt_name + " (from " + xf.src_name + ")"),
xf.tgt_ranges, xf.src_ranges))
def AssertSequenceGood(self):
# Simulate the sequences of transfers we will output, and check that:
# - we never read a block after writing it, and
# - we write every block we care about exactly once.
# Start with no blocks having been touched yet.
touched = array.array("B", "\0" * self.tgt.total_blocks)
# Imagine processing the transfers in order.
for xf in self.transfers:
# Check that the input blocks for this transfer haven't yet been touched.
x = xf.src_ranges
for _, sr in xf.use_stash:
x = x.subtract(sr)
for s, e in x:
# Source image could be larger. Don't check the blocks that are in the
# source image only. Since they are not in 'touched', and won't ever
# be touched.
for i in range(s, min(e, self.tgt.total_blocks)):
assert touched[i] == 0
# Check that the output blocks for this transfer haven't yet
# been touched, and touch all the blocks written by this
# transfer.
for s, e in xf.tgt_ranges:
for i in range(s, e):
assert touched[i] == 0
touched[i] = 1
# Check that we've written every target block.
for s, e in self.tgt.care_map:
for i in range(s, e):
assert touched[i] == 1
def ImproveVertexSequence(self):
print("Improving vertex order...")
# At this point our digraph is acyclic; we reversed any edges that
# were backwards in the heuristically-generated sequence. The
# previously-generated order is still acceptable, but we hope to
# find a better order that needs less memory for stashed data.
# Now we do a topological sort to generate a new vertex order,
# using a greedy algorithm to choose which vertex goes next
# whenever we have a choice.
# Make a copy of the edge set; this copy will get destroyed by the
# algorithm.
for xf in self.transfers:
xf.incoming = xf.goes_after.copy()
xf.outgoing = xf.goes_before.copy()
L = [] # the new vertex order
# S is the set of sources in the remaining graph; we always choose
# the one that leaves the least amount of stashed data after it's
# executed.
S = [(u.NetStashChange(), u.order, u) for u in self.transfers
if not u.incoming]
heapq.heapify(S)
while S:
_, _, xf = heapq.heappop(S)
L.append(xf)
for u in xf.outgoing:
del u.incoming[xf]
if not u.incoming:
heapq.heappush(S, (u.NetStashChange(), u.order, u))
# if this fails then our graph had a cycle.
assert len(L) == len(self.transfers)
self.transfers = L
for i, xf in enumerate(L):
xf.order = i
def RemoveBackwardEdges(self):
print("Removing backward edges...")
in_order = 0
out_of_order = 0
lost_source = 0
for xf in self.transfers:
lost = 0
size = xf.src_ranges.size()
for u in xf.goes_before:
# xf should go before u
if xf.order < u.order:
# it does, hurray!
in_order += 1
else:
# it doesn't, boo. trim the blocks that u writes from xf's
# source, so that xf can go after u.
out_of_order += 1
assert xf.src_ranges.overlaps(u.tgt_ranges)
xf.src_ranges = xf.src_ranges.subtract(u.tgt_ranges)
xf.intact = False
if xf.style == "diff" and not xf.src_ranges:
# nothing left to diff from; treat as new data
xf.style = "new"
lost = size - xf.src_ranges.size()
lost_source += lost
print((" %d/%d dependencies (%.2f%%) were violated; "
"%d source blocks removed.") %
(out_of_order, in_order + out_of_order,
(out_of_order * 100.0 / (in_order + out_of_order))
if (in_order + out_of_order) else 0.0,
lost_source))
def ReverseBackwardEdges(self):
"""Reverse unsatisfying edges and compute pairs of stashed blocks.
For each transfer, make sure it properly stashes the blocks it touches and
will be used by later transfers. It uses pairs of (stash_raw_id, range) to
record the blocks to be stashed. 'stash_raw_id' is an id that uniquely
identifies each pair. Note that for the same range (e.g. RangeSet("1-5")),
it is possible to have multiple pairs with different 'stash_raw_id's. Each
'stash_raw_id' will be consumed by one transfer. In BBOTA v3+, identical
blocks will be written to the same stash slot in WriteTransfers().
"""
print("Reversing backward edges...")
in_order = 0
out_of_order = 0
stash_raw_id = 0
stash_size = 0
for xf in self.transfers:
for u in xf.goes_before.copy():
# xf should go before u
if xf.order < u.order:
# it does, hurray!
in_order += 1
else:
# it doesn't, boo. modify u to stash the blocks that it
# writes that xf wants to read, and then require u to go
# before xf.
out_of_order += 1
overlap = xf.src_ranges.intersect(u.tgt_ranges)
assert overlap
u.stash_before.append((stash_raw_id, overlap))
xf.use_stash.append((stash_raw_id, overlap))
stash_raw_id += 1
stash_size += overlap.size()
# reverse the edge direction; now xf must go after u
del xf.goes_before[u]
del u.goes_after[xf]
xf.goes_after[u] = None # value doesn't matter
u.goes_before[xf] = None
print((" %d/%d dependencies (%.2f%%) were violated; "
"%d source blocks stashed.") %
(out_of_order, in_order + out_of_order,
(out_of_order * 100.0 / (in_order + out_of_order))
if (in_order + out_of_order) else 0.0,
stash_size))
def FindVertexSequence(self):
print("Finding vertex sequence...")
# This is based on "A Fast & Effective Heuristic for the Feedback
# Arc Set Problem" by P. Eades, X. Lin, and W.F. Smyth. Think of
# it as starting with the digraph G and moving all the vertices to
# be on a horizontal line in some order, trying to minimize the
# number of edges that end up pointing to the left. Left-pointing
# edges will get removed to turn the digraph into a DAG. In this
# case each edge has a weight which is the number of source blocks
# we'll lose if that edge is removed; we try to minimize the total
# weight rather than just the number of edges.
# Make a copy of the edge set; this copy will get destroyed by the
# algorithm.
for xf in self.transfers:
xf.incoming = xf.goes_after.copy()
xf.outgoing = xf.goes_before.copy()
xf.score = sum(xf.outgoing.values()) - sum(xf.incoming.values())
# We use an OrderedDict instead of just a set so that the output
# is repeatable; otherwise it would depend on the hash values of
# the transfer objects.
G = OrderedDict()
for xf in self.transfers:
G[xf] = None
s1 = deque() # the left side of the sequence, built from left to right
s2 = deque() # the right side of the sequence, built from right to left
heap = []
for xf in self.transfers:
xf.heap_item = HeapItem(xf)
heap.append(xf.heap_item)
heapq.heapify(heap)
# Use OrderedDict() instead of set() to preserve the insertion order. Need
# to use 'sinks[key] = None' to add key into the set. sinks will look like
# { key1: None, key2: None, ... }.
sinks = OrderedDict.fromkeys(u for u in G if not u.outgoing)
sources = OrderedDict.fromkeys(u for u in G if not u.incoming)
def adjust_score(iu, delta):
iu.score += delta
iu.heap_item.clear()
iu.heap_item = HeapItem(iu)
heapq.heappush(heap, iu.heap_item)
while G:
# Put all sinks at the end of the sequence.
while sinks:
new_sinks = OrderedDict()
for u in sinks:
if u not in G: continue
s2.appendleft(u)
del G[u]
for iu in u.incoming:
adjust_score(iu, -iu.outgoing.pop(u))
if not iu.outgoing:
new_sinks[iu] = None
sinks = new_sinks
# Put all the sources at the beginning of the sequence.
while sources:
new_sources = OrderedDict()
for u in sources:
if u not in G: continue
s1.append(u)
del G[u]
for iu in u.outgoing:
adjust_score(iu, +iu.incoming.pop(u))
if not iu.incoming:
new_sources[iu] = None
sources = new_sources
if not G: break
# Find the "best" vertex to put next. "Best" is the one that
# maximizes the net difference in source blocks saved we get by
# pretending it's a source rather than a sink.
while True:
u = heapq.heappop(heap)
if u and u.item in G:
u = u.item
break
s1.append(u)
del G[u]
for iu in u.outgoing:
adjust_score(iu, +iu.incoming.pop(u))
if not iu.incoming:
sources[iu] = None
for iu in u.incoming:
adjust_score(iu, -iu.outgoing.pop(u))
if not iu.outgoing:
sinks[iu] = None
# Now record the sequence in the 'order' field of each transfer,
# and by rearranging self.transfers to be in the chosen sequence.
new_transfers = []
for x in itertools.chain(s1, s2):
x.order = len(new_transfers)
new_transfers.append(x)
del x.incoming
del x.outgoing
self.transfers = new_transfers
def GenerateDigraph(self):
print("Generating digraph...")
# Each item of source_ranges will be:
# - None, if that block is not used as a source,
# - an ordered set of transfers.
source_ranges = []
for b in self.transfers:
for s, e in b.src_ranges:
if e > len(source_ranges):
source_ranges.extend([None] * (e-len(source_ranges)))
for i in range(s, e):
if source_ranges[i] is None:
source_ranges[i] = OrderedDict.fromkeys([b])
else:
source_ranges[i][b] = None
for a in self.transfers:
intersections = OrderedDict()
for s, e in a.tgt_ranges:
for i in range(s, e):
if i >= len(source_ranges): break
# Add all the Transfers in source_ranges[i] to the (ordered) set.
if source_ranges[i] is not None:
for j in source_ranges[i]:
intersections[j] = None
for b in intersections:
if a is b: continue
# If the blocks written by A are read by B, then B needs to go before A.
i = a.tgt_ranges.intersect(b.src_ranges)
if i:
if b.src_name == "__ZERO":
# the cost of removing source blocks for the __ZERO domain
# is (nearly) zero.
size = 0
else:
size = i.size()
b.goes_before[a] = size
a.goes_after[b] = size
def FindTransfers(self):
"""Parse the file_map to generate all the transfers."""
def AddSplitTransfers(tgt_name, src_name, tgt_ranges, src_ranges,
style, by_id):
"""Add one or multiple Transfer()s by splitting large files.
For BBOTA v3, we need to stash source blocks for resumable feature.
However, with the growth of file size and the shrink of the cache
partition source blocks are too large to be stashed. If a file occupies
too many blocks, we split it into smaller pieces by getting multiple
Transfer()s.
The downside is that after splitting, we may increase the package size
since the split pieces don't align well. According to our experiments,
1/8 of the cache size as the per-piece limit appears to be optimal.
Compared to the fixed 1024-block limit, it reduces the overall package
size by 30% for volantis, and 20% for angler and bullhead."""
# Possibly split large files into smaller chunks.
pieces = 0
cache_size = common.OPTIONS.cache_size
split_threshold = 0.125
max_blocks_per_transfer = int(cache_size * split_threshold /
self.tgt.blocksize)
# Change nothing for small files.
if (tgt_ranges.size() <= max_blocks_per_transfer and
src_ranges.size() <= max_blocks_per_transfer):
Transfer(tgt_name, src_name, tgt_ranges, src_ranges,
self.tgt.RangeSha1(tgt_ranges), self.src.RangeSha1(src_ranges),
style, by_id)
return
while (tgt_ranges.size() > max_blocks_per_transfer and
src_ranges.size() > max_blocks_per_transfer):
tgt_split_name = "%s-%d" % (tgt_name, pieces)
src_split_name = "%s-%d" % (src_name, pieces)
tgt_first = tgt_ranges.first(max_blocks_per_transfer)
src_first = src_ranges.first(max_blocks_per_transfer)
Transfer(tgt_split_name, src_split_name, tgt_first, src_first,
self.tgt.RangeSha1(tgt_first), self.src.RangeSha1(src_first),
style, by_id)
tgt_ranges = tgt_ranges.subtract(tgt_first)
src_ranges = src_ranges.subtract(src_first)
pieces += 1
# Handle remaining blocks.
if tgt_ranges.size() or src_ranges.size():
# Must be both non-empty.
assert tgt_ranges.size() and src_ranges.size()
tgt_split_name = "%s-%d" % (tgt_name, pieces)
src_split_name = "%s-%d" % (src_name, pieces)
Transfer(tgt_split_name, src_split_name, tgt_ranges, src_ranges,
self.tgt.RangeSha1(tgt_ranges), self.src.RangeSha1(src_ranges),
style, by_id)
def AddTransfer(tgt_name, src_name, tgt_ranges, src_ranges, style, by_id,
split=False):
"""Wrapper function for adding a Transfer()."""
# We specialize diff transfers only (which covers bsdiff/imgdiff/move);
# otherwise add the Transfer() as is.
if style != "diff" or not split:
Transfer(tgt_name, src_name, tgt_ranges, src_ranges,
self.tgt.RangeSha1(tgt_ranges), self.src.RangeSha1(src_ranges),
style, by_id)
return
# Handle .odex files specially to analyze the block-wise difference. If
# most of the blocks are identical with only few changes (e.g. header),
# we will patch the changed blocks only. This avoids stashing unchanged
# blocks while patching. We limit the analysis to files without size
# changes only. This is to avoid sacrificing the OTA generation cost too
# much.
if (tgt_name.split(".")[-1].lower() == 'odex' and
tgt_ranges.size() == src_ranges.size()):
# 0.5 threshold can be further tuned. The tradeoff is: if only very
# few blocks remain identical, we lose the opportunity to use imgdiff
# that may have better compression ratio than bsdiff.
crop_threshold = 0.5
tgt_skipped = RangeSet()
src_skipped = RangeSet()
tgt_size = tgt_ranges.size()
tgt_changed = 0
for src_block, tgt_block in zip(src_ranges.next_item(),
tgt_ranges.next_item()):
src_rs = RangeSet(str(src_block))
tgt_rs = RangeSet(str(tgt_block))
if self.src.ReadRangeSet(src_rs) == self.tgt.ReadRangeSet(tgt_rs):
tgt_skipped = tgt_skipped.union(tgt_rs)
src_skipped = src_skipped.union(src_rs)
else:
tgt_changed += tgt_rs.size()
# Terminate early if no clear sign of benefits.
if tgt_changed > tgt_size * crop_threshold:
break
if tgt_changed < tgt_size * crop_threshold:
assert tgt_changed + tgt_skipped.size() == tgt_size
print('%10d %10d (%6.2f%%) %s' % (tgt_skipped.size(), tgt_size,
tgt_skipped.size() * 100.0 / tgt_size, tgt_name))
AddSplitTransfers(
"%s-skipped" % (tgt_name,),
"%s-skipped" % (src_name,),
tgt_skipped, src_skipped, style, by_id)
# Intentionally change the file extension to avoid being imgdiff'd as
# the files are no longer in their original format.
tgt_name = "%s-cropped" % (tgt_name,)
src_name = "%s-cropped" % (src_name,)
tgt_ranges = tgt_ranges.subtract(tgt_skipped)
src_ranges = src_ranges.subtract(src_skipped)
# Possibly having no changed blocks.
if not tgt_ranges:
return
# Add the transfer(s).
AddSplitTransfers(
tgt_name, src_name, tgt_ranges, src_ranges, style, by_id)
print("Finding transfers...")
empty = RangeSet()
for tgt_fn, tgt_ranges in self.tgt.file_map.items():
if tgt_fn == "__ZERO":
# the special "__ZERO" domain is all the blocks not contained
# in any file and that are filled with zeros. We have a
# special transfer style for zero blocks.
src_ranges = self.src.file_map.get("__ZERO", empty)
AddTransfer(tgt_fn, "__ZERO", tgt_ranges, src_ranges,
"zero", self.transfers)
continue
elif tgt_fn == "__COPY":
# "__COPY" domain includes all the blocks not contained in any
# file and that need to be copied unconditionally to the target.
AddTransfer(tgt_fn, None, tgt_ranges, empty, "new", self.transfers)
continue
elif tgt_fn in self.src.file_map:
# Look for an exact pathname match in the source.
AddTransfer(tgt_fn, tgt_fn, tgt_ranges, self.src.file_map[tgt_fn],
"diff", self.transfers, True)
continue
b = os.path.basename(tgt_fn)
if b in self.src_basenames:
# Look for an exact basename match in the source.
src_fn = self.src_basenames[b]
AddTransfer(tgt_fn, src_fn, tgt_ranges, self.src.file_map[src_fn],
"diff", self.transfers, True)
continue
b = re.sub("[0-9]+", "#", b)
if b in self.src_numpatterns:
# Look for a 'number pattern' match (a basename match after
# all runs of digits are replaced by "#"). (This is useful
# for .so files that contain version numbers in the filename
# that get bumped.)
src_fn = self.src_numpatterns[b]
AddTransfer(tgt_fn, src_fn, tgt_ranges, self.src.file_map[src_fn],
"diff", self.transfers, True)
continue
AddTransfer(tgt_fn, None, tgt_ranges, empty, "new", self.transfers)
def AbbreviateSourceNames(self):
for k in self.src.file_map.keys():
b = os.path.basename(k)
self.src_basenames[b] = k
b = re.sub("[0-9]+", "#", b)
self.src_numpatterns[b] = k
@staticmethod
def AssertPartition(total, seq):
"""Assert that all the RangeSets in 'seq' form a partition of the
'total' RangeSet (ie, they are nonintersecting and their union
equals 'total')."""
so_far = RangeSet()
for i in seq:
assert not so_far.overlaps(i)
so_far = so_far.union(i)
assert so_far == total
