class django_bcrypt_sha256_test(HandlerCase, _DjangoHelper):
"""test django_bcrypt_sha256"""
handler = hash.django_bcrypt_sha256
forbidden_characters = None
fuzz_salts_need_bcrypt_repair = True
known_correct_hashes = [
#
# custom - generated via django 1.6 hasher
#
(
"",
"bcrypt_sha256$$2a$06$/3OeRpbOf8/l6nPPRdZPp.nRiyYqPobEZGdNRBWihQhiFDh1ws1tu",
),
(
UPASS_LETMEIN,
"bcrypt_sha256$$2a$08$NDjSAIcas.EcoxCRiArvT.MkNiPYVhrsrnJsRkLueZOoV1bsQqlmC",
),
(
UPASS_TABLE,
"bcrypt_sha256$$2a$06$kCXUnRFQptGg491siDKNTu8RxjBGSjALHRuvhPYNFsa4Ea5d9M48u",
),
# test >72 chars is hashed correctly -- under bcrypt these hash the same.
(
repeat_string("abc123", 72),
"bcrypt_sha256$$2a$06$Tg/oYyZTyAf.Nb3qSgN61OySmyXA8FoY4PjGizjE1QSDfuL5MXNni",
),
(
repeat_string("abc123", 72) + "qwr",
"bcrypt_sha256$$2a$06$Tg/oYyZTyAf.Nb3qSgN61Ocy0BEz1RK6xslSNi8PlaLX2pe7x/KQG",
),
(
repeat_string("abc123", 72) + "xyz",
"bcrypt_sha256$$2a$06$Tg/oYyZTyAf.Nb3qSgN61OvY2zoRVUa2Pugv2ExVOUT2YmhvxUFUa",
),
]
known_malformed_hashers = [
# data in django salt field
"bcrypt_sha256$xyz$2a$06$/3OeRpbOf8/l6nPPRdZPp.nRiyYqPobEZGdNRBWihQhiFDh1ws1tu"
]
# NOTE: the following have been cloned from _bcrypt_test()
def populate_settings(self, kwds):
# speed up test w/ lower rounds
kwds.setdefault("rounds", 4)
super(django_bcrypt_sha256_test, self).populate_settings(kwds)
class FuzzHashGenerator(HandlerCase.FuzzHashGenerator):
def random_rounds(self):
# decrease default rounds for fuzz testing to speed up volume.
return self.randintgauss(5, 8, 6, 1)
def random_ident(self):
# omit multi-ident tests, only $2a$ counts for this class
# XXX: enable this to check 2a / 2b?
return None
class django_bcrypt_sha256_test(HandlerCase, _DjangoHelper):
"test django_bcrypt_sha256"
handler = hash.django_bcrypt_sha256
min_django_version = (1,6)
forbidden_characters = None
fuzz_salts_need_bcrypt_repair = True
known_correct_hashes = [
#
# custom - generated via django 1.6 hasher
#
('',
'bcrypt_sha256$$2a$06$/3OeRpbOf8/l6nPPRdZPp.nRiyYqPobEZGdNRBWihQhiFDh1ws1tu'),
(UPASS_LETMEIN,
'bcrypt_sha256$$2a$08$NDjSAIcas.EcoxCRiArvT.MkNiPYVhrsrnJsRkLueZOoV1bsQqlmC'),
(UPASS_TABLE,
'bcrypt_sha256$$2a$06$kCXUnRFQptGg491siDKNTu8RxjBGSjALHRuvhPYNFsa4Ea5d9M48u'),
# test >72 chars is hashed correctly -- under bcrypt these hash the same.
(repeat_string("abc123",72),
'bcrypt_sha256$$2a$06$Tg/oYyZTyAf.Nb3qSgN61OySmyXA8FoY4PjGizjE1QSDfuL5MXNni'),
(repeat_string("abc123",72)+"qwr",
'bcrypt_sha256$$2a$06$Tg/oYyZTyAf.Nb3qSgN61Ocy0BEz1RK6xslSNi8PlaLX2pe7x/KQG'),
(repeat_string("abc123",72)+"xyz",
'bcrypt_sha256$$2a$06$Tg/oYyZTyAf.Nb3qSgN61OvY2zoRVUa2Pugv2ExVOUT2YmhvxUFUa'),
]
known_malformed_hashers = [
# data in django salt field
'bcrypt_sha256$xyz$2a$06$/3OeRpbOf8/l6nPPRdZPp.nRiyYqPobEZGdNRBWihQhiFDh1ws1tu',
]
def test_30_HasManyIdents(self):
raise self.skipTest("multiple idents not supported")
def test_30_HasOneIdent(self):
# forbidding ident keyword, django doesn't support configuring this
handler = self.handler
handler(use_defaults=True)
self.assertRaises(TypeError, handler, ident="$2a$", use_defaults=True)
# NOTE: the following have been cloned from _bcrypt_test()
def populate_settings(self, kwds):
# speed up test w/ lower rounds
kwds.setdefault("rounds", 4)
super(django_bcrypt_sha256_test, self).populate_settings(kwds)
def fuzz_setting_rounds(self):
# decrease default rounds for fuzz testing to speed up volume.
return randintgauss(5, 8, 6, 1)
def fuzz_setting_ident(self):
# omit multi-ident tests, only $2a$ counts for this class
return None
def _calc_checksum(self, secret):
"""common backend code"""
# make sure it's unicode
if isinstance(secret, unicode):
secret = secret.encode("utf-8")
# NOTE: especially important to forbid NULLs for bcrypt, since many
# backends (bcryptor, bcrypt) happily accept them, and then
# silently truncate the password at first NULL they encounter!
if _BNULL in secret:
raise uh.exc.NullPasswordError(self)
# ensure backend is loaded before workaround detection
self.get_backend()
# protect from wraparound bug by truncating secret before handing it to the backend.
# bcrypt only uses first 72 bytes anyways.
if self._has_wraparound_bug and len(secret) >= 255:
secret = secret[:72]
# special case handling for variants (ordered most common first)
ident = self.ident
if ident == IDENT_2A:
# fall through and use backend w/o hacks
pass
elif ident == IDENT_2B:
if self._lacks_2b_support:
# handle $2b$ hash format even if backend is too old.
# have it generate a 2A digest, then return it as a 2B hash.
ident = IDENT_2A
elif ident == IDENT_2Y:
if self._lacks_2y_support:
# handle $2y$ hash format (not supported by BSDs, being phased out on others)
# have it generate a 2A digest, then return it as a 2Y hash.
ident = IDENT_2A
elif ident == IDENT_2:
if self._lacks_20_support:
# handle legacy $2$ format (not supported by most backends except BSD os_crypt)
# we can fake $2$ behavior using the $2a$ algorithm
# by repeating the password until it's at least 72 chars in length.
if secret:
secret = repeat_string(secret, 72)
ident = IDENT_2A
elif ident == IDENT_2X:
# NOTE: shouldn't get here.
# XXX: could check if backend does actually offer 'support'
raise RuntimeError("$2x$ hashes not currently supported by passlib")
else:
raise AssertionError("unexpected ident value: %r" % ident)
# invoke backend
config = self._get_config(ident)
return self._calc_checksum_backend(secret, config)
def check_bcryptor(secret, hash):
"bcryptor"
secret = to_native_str(secret, self.fuzz_password_encoding)
if hash.startswith(IDENT_2Y):
hash = IDENT_2A + hash[4:]
elif hash.startswith(IDENT_2):
# bcryptor doesn't support $2$ hashes; but we can fake it
# using the $2a$ algorithm, by repeating the password until
# it's 72 chars in length.
hash = IDENT_2A + hash[3:]
if secret:
secret = repeat_string(secret, 72)
return Engine(False).hash_key(secret, hash) == hash
def get_fuzz_settings(self):
secret, other, kwds = super(_bcrypt_test,self).get_fuzz_settings()
from passlib.handlers.bcrypt import IDENT_2, IDENT_2X
from passlib.utils import to_bytes
ident = kwds.get('ident')
if ident == IDENT_2X:
# 2x is just recognized, not supported. don't test with it.
del kwds['ident']
elif ident == IDENT_2 and other and repeat_string(to_bytes(other), len(to_bytes(secret))) == to_bytes(secret):
# avoid false failure due to flaw in 0-revision bcrypt:
# repeated strings like 'abc' and 'abcabc' hash identically.
other = self.get_fuzz_password()
return secret, other, kwds
def check_bcryptor(secret, hash):
"""bcryptor"""
secret = to_native_str(secret, self.FuzzHashGenerator.password_encoding)
if hash.startswith((IDENT_2B, IDENT_2Y)):
hash = IDENT_2A + hash[4:]
elif hash.startswith(IDENT_2):
# bcryptor doesn't support $2$ hashes; but we can fake it
# using the $2a$ algorithm, by repeating the password until
# it's 72 chars in length.
hash = IDENT_2A + hash[3:]
if secret:
secret = repeat_string(secret, 72)
return Engine(False).hash_key(secret, hash) == hash
def get_fuzz_settings(self):
secret, other, kwds = super(_bcrypt_test, self).get_fuzz_settings()
from passlib.handlers.bcrypt import IDENT_2, IDENT_2X
from passlib.utils import to_bytes
ident = kwds.get('ident')
if ident == IDENT_2X:
# 2x is just recognized, not supported. don't test with it.
del kwds['ident']
elif ident == IDENT_2 and other and repeat_string(
to_bytes(other), len(to_bytes(secret))) == to_bytes(secret):
# avoid false failure due to flaw in 0-revision bcrypt:
# repeated strings like 'abc' and 'abcabc' hash identically.
other = self.get_fuzz_password()
return secret, other, kwds
def key_to_words(data, size=18):
"""convert data to tuple of <size> 4-byte integers, repeating or
truncating data as needed to reach specified size"""
assert isinstance(data, bytes)
dlen = len(data)
if not dlen:
# return all zeros - original C code would just read the NUL after
# the password, so mimicing that behavior for this edge case.
return [0] * size
# repeat data until it fills up 4*size bytes
data = repeat_string(data, size << 2)
# unpack
return struct.unpack(">%dI" % (size, ), data)
def key_to_words(data, size=18):
"""convert data to tuple of <size> 4-byte integers, repeating or
truncating data as needed to reach specified size"""
assert isinstance(data, bytes)
dlen = len(data)
if not dlen:
# return all zeros - original C code would just read the NUL after
# the password, so mimicing that behavior for this edge case.
return [0]*size
# repeat data until it fills up 4*size bytes
data = repeat_string(data, size<<2)
# unpack
return struct.unpack(">%dI" % (size,), data)
def check_bcrypt(secret, hash):
"""bcrypt"""
secret = to_bytes(secret, self.fuzz_password_encoding)
# if hash.startswith(IDENT_2Y):
# hash = IDENT_2A + hash[4:]
if hash.startswith(IDENT_2):
# bcryptor doesn't support $2$ hashes; but we can fake it
# using the $2a$ algorithm, by repeating the password until
# it's 72 chars in length.
hash = IDENT_2A + hash[3:]
if secret:
secret = repeat_string(secret, 72)
hash = to_bytes(hash)
try:
return bcrypt.hashpw(secret, hash) == hash
except ValueError:
raise ValueError("bcrypt rejected hash: %r" % (hash,))
def _calc_checksum_bcryptor(self, secret):
# bcryptor behavior:
# py2: unicode secret/hash encoded as ascii bytes before use,
# bytes taken as-is; returns ascii bytes.
# py3: not supported
if self.ident == IDENT_2:
# bcryptor doesn't support $2$ hashes; but we can fake $2$ behavior
# using the $2a$ algorithm, by repeating the password until
# it's at least 72 chars in length.
if secret:
secret = repeat_string(secret, 72)
config = self._get_config(IDENT_2A)
else:
config = self._get_config()
hash = _bcryptor_engine(False).hash_key(secret, config)
assert hash.startswith(config) and len(hash) == len(config)+31
return str_to_uascii(hash[-31:])
def _calc_checksum_bcryptor(self, secret):
# bcryptor behavior:
# py2: unicode secret/hash encoded as ascii bytes before use,
# bytes taken as-is; returns ascii bytes.
# py3: not supported
if self.ident == IDENT_2:
# bcryptor doesn't support $2$ hashes; but we can fake $2$ behavior
# using the $2a$ algorithm, by repeating the password until
# it's at least 72 chars in length.
if secret:
secret = repeat_string(secret, 72)
config = self._get_config(IDENT_2A)
else:
config = self._get_config()
hash = _bcryptor_engine(False).hash_key(secret, config)
assert hash.startswith(config) and len(hash) == len(config) + 31
return str_to_uascii(hash[-31:])
def check_bcrypt(secret, hash):
"bcrypt"
secret = to_bytes(secret, self.fuzz_password_encoding)
#if hash.startswith(IDENT_2Y):
# hash = IDENT_2A + hash[4:]
if hash.startswith(IDENT_2):
# bcryptor doesn't support $2$ hashes; but we can fake it
# using the $2a$ algorithm, by repeating the password until
# it's 72 chars in length.
hash = IDENT_2A + hash[3:]
if secret:
secret = repeat_string(secret, 72)
hash = to_bytes(hash)
try:
return bcrypt.hashpw(secret, hash) == hash
except ValueError:
raise ValueError("bcrypt rejected hash: %r" % (hash, ))
def generate(self):
opts = super(_bcrypt_test.FuzzHashGenerator, self).generate()
secret = opts['secret']
other = opts['other']
settings = opts['settings']
ident = settings.get('ident')
if ident == IDENT_2X:
# 2x is just recognized, not supported. don't test with it.
del settings['ident']
elif ident == IDENT_2 and other and repeat_string(to_bytes(other), len(to_bytes(secret))) == to_bytes(secret):
# avoid false failure due to flaw in 0-revision bcrypt:
# repeated strings like 'abc' and 'abcabc' hash identically.
opts['secret'], opts['other'] = self.random_password_pair()
return opts
def _calc_checksum_bcrypt(self, secret):
# bcrypt behavior:
# hash must be ascii bytes
# secret must be bytes
# returns bytes
if self.ident == IDENT_2:
# bcrypt doesn't support $2$ hashes; but we can fake $2$ behavior
# using the $2a$ algorithm, by repeating the password until
# it's at least 72 chars in length.
if secret:
secret = repeat_string(secret, 72)
config = self._get_config(IDENT_2A)
else:
config = self._get_config()
if isinstance(config, unicode):
config = config.encode("ascii")
hash = _bcrypt.hashpw(secret, config)
assert hash.startswith(config) and len(hash) == len(config)+31
assert isinstance(hash, bytes)
return hash[-31:].decode("ascii")
def _calc_checksum_bcrypt(self, secret):
# bcrypt behavior:
# hash must be ascii bytes
# secret must be bytes
# returns bytes
if self.ident == IDENT_2:
# bcrypt doesn't support $2$ hashes; but we can fake $2$ behavior
# using the $2a$ algorithm, by repeating the password until
# it's at least 72 chars in length.
if secret:
secret = repeat_string(secret, 72)
config = self._get_config(IDENT_2A)
else:
config = self._get_config()
if isinstance(config, unicode):
config = config.encode("ascii")
hash = _bcrypt.hashpw(secret, config)
assert hash.startswith(config) and len(hash) == len(config) + 31
assert isinstance(hash, bytes)
return hash[-31:].decode("ascii")
def generate(self):
opts = super(_bcrypt_test.FuzzHashGenerator, self).generate()
secret = opts['secret']
other = opts['other']
settings = opts['settings']
ident = settings.get('ident')
if ident == IDENT_2X:
# 2x is just recognized, not supported. don't test with it.
del settings['ident']
elif ident == IDENT_2 and other and repeat_string(
to_bytes(other), len(
to_bytes(secret))) == to_bytes(secret):
# avoid false failure due to flaw in 0-revision bcrypt:
# repeated strings like 'abc' and 'abcabc' hash identically.
opts['secret'], opts['other'] = self.random_password_pair()
return opts
def check_bcrypt(secret, hash):
"""bcrypt"""
secret = to_bytes(secret, self.FuzzHashGenerator.password_encoding)
if hash.startswith(IDENT_2B):
# bcrypt <1.1 lacks 2B support
hash = IDENT_2A + hash[4:]
elif hash.startswith(IDENT_2):
# bcrypt doesn't support $2$ hashes; but we can fake it
# using the $2a$ algorithm, by repeating the password until
# it's 72 chars in length.
hash = IDENT_2A + hash[3:]
if secret:
secret = repeat_string(secret, 72)
elif hash.startswith(IDENT_2Y) and bcrypt.__version__ == "3.0.0":
hash = IDENT_2B + hash[4:]
hash = to_bytes(hash)
try:
return bcrypt.hashpw(secret, hash) == hash
except ValueError:
raise ValueError("bcrypt rejected hash: %r (secret=%r)" % (hash, secret))
def check_bcrypt(secret, hash):
"""bcrypt"""
secret = to_bytes(secret, self.FuzzHashGenerator.password_encoding)
if hash.startswith(IDENT_2B):
# bcrypt <1.1 lacks 2B support
hash = IDENT_2A + hash[4:]
elif hash.startswith(IDENT_2):
# bcrypt doesn't support $2$ hashes; but we can fake it
# using the $2a$ algorithm, by repeating the password until
# it's 72 chars in length.
hash = IDENT_2A + hash[3:]
if secret:
secret = repeat_string(secret, 72)
elif hash.startswith(IDENT_2Y) and bcrypt.__version__ == "3.0.0":
hash = IDENT_2B + hash[4:]
hash = to_bytes(hash)
try:
return bcrypt.hashpw(secret, hash) == hash
except ValueError:
raise ValueError("bcrypt rejected hash: %r (secret=%r)" %
(hash, secret))
def _calc_checksum_bcryptor(self, secret):
# bcryptor behavior:
# py2: unicode secret/hash encoded as ascii bytes before use,
# bytes taken as-is; returns ascii bytes.
# py3: not supported
if isinstance(secret, unicode):
secret = secret.encode("utf-8")
if _BNULL in secret:
# NOTE: especially important to forbid NULLs for bcryptor,
# since it happily accepts them, and then silently truncates
# the password at first one it encounters :(
raise uh.exc.NullPasswordError(self)
if self.ident == IDENT_2:
# bcryptor doesn't support $2$ hashes; but we can fake $2$ behavior
# using the $2a$ algorithm, by repeating the password until
# it's at least 72 chars in length.
if secret:
secret = repeat_string(secret, 72)
config = self._get_config(IDENT_2A)
else:
config = self._get_config()
hash = bcryptor_engine(False).hash_key(secret, config)
assert hash.startswith(config) and len(hash) == len(config)+31
return str_to_uascii(hash[-31:])
def _calc_checksum_bcryptor(self, secret):
# bcryptor behavior:
# py2: unicode secret/hash encoded as ascii bytes before use,
# bytes taken as-is; returns ascii bytes.
# py3: not supported
if isinstance(secret, unicode):
secret = secret.encode("utf-8")
if _BNULL in secret:
# NOTE: especially important to forbid NULLs for bcryptor,
# since it happily accepts them, and then silently truncates
# the password at first one it encounters :(
raise uh.exc.NullPasswordError(self)
if self.ident == IDENT_2:
# bcryptor doesn't support $2$ hashes; but we can fake $2$ behavior
# using the $2a$ algorithm, by repeating the password until
# it's at least 72 chars in length.
if secret:
secret = repeat_string(secret, 72)
config = self._get_config(IDENT_2A)
else:
config = self._get_config()
hash = bcryptor_engine(False).hash_key(secret, config)
assert hash.startswith(config) and len(hash) == len(config) + 31
return str_to_uascii(hash[-31:])
def _raw_sha2_crypt(pwd, salt, rounds, use_512=False):
"""perform raw sha256-crypt / sha512-crypt
this function provides a pure-python implementation of the internals
for the SHA256-Crypt and SHA512-Crypt algorithms; it doesn't
handle any of the parsing/validation of the hash strings themselves.
:arg pwd: password chars/bytes to hash
:arg salt: salt chars to use
:arg rounds: linear rounds cost
:arg use_512: use sha512-crypt instead of sha256-crypt mode
:returns:
encoded checksum chars
"""
# ===================================================================
# init & validate inputs
# ===================================================================
# NOTE: the setup portion of this algorithm scales ~linearly in time
# with the size of the password, making it vulnerable to a DOS from
# unreasonably large inputs. the following code has some optimizations
# which would make things even worse, using O(pwd_len**2) memory
# when calculating digest P.
#
# to mitigate these two issues: 1) this code switches to a
# O(pwd_len)-memory algorithm for passwords that are much larger
# than average, and 2) Passlib enforces a library-wide max limit on
# the size of passwords it will allow, to prevent this algorithm and
# others from being DOSed in this way (see passlib.exc.PasswordSizeError
# for details).
# validate secret
if isinstance(pwd, unicode):
# XXX: not sure what official unicode policy is, using this as default
pwd = pwd.encode("utf-8")
assert isinstance(pwd, bytes)
if _BNULL in pwd:
raise uh.exc.NullPasswordError(
sha512_crypt if use_512 else sha256_crypt)
pwd_len = len(pwd)
# validate rounds
assert 1000 <= rounds <= 999999999, "invalid rounds"
# NOTE: spec says out-of-range rounds should be clipped, instead of
# causing an error. this function assumes that's been taken care of
# by the handler class.
# validate salt
assert isinstance(salt, unicode), "salt not unicode"
salt = salt.encode("ascii")
salt_len = len(salt)
assert salt_len < 17, "salt too large"
# NOTE: spec says salts larger than 16 bytes should be truncated,
# instead of causing an error. this function assumes that's been
# taken care of by the handler class.
# load sha256/512 specific constants
if use_512:
hash_const = hashlib.sha512
transpose_map = _512_transpose_map
else:
hash_const = hashlib.sha256
transpose_map = _256_transpose_map
# ===================================================================
# digest B - used as subinput to digest A
# ===================================================================
db = hash_const(pwd + salt + pwd).digest()
# ===================================================================
# digest A - used to initialize first round of digest C
# ===================================================================
# start out with pwd + salt
a_ctx = hash_const(pwd + salt)
a_ctx_update = a_ctx.update
# add pwd_len bytes of b, repeating b as many times as needed.
a_ctx_update(repeat_string(db, pwd_len))
# for each bit in pwd_len: add b if it's 1, or pwd if it's 0
i = pwd_len
while i:
a_ctx_update(db if i & 1 else pwd)
i >>= 1
# finish A
da = a_ctx.digest()
# ===================================================================
# digest P from password - used instead of password itself
# when calculating digest C.
# ===================================================================
if pwd_len < 96:
# this method is faster under python, but uses O(pwd_len**2) memory;
# so we don't use it for larger passwords to avoid a potential DOS.
dp = repeat_string(hash_const(pwd * pwd_len).digest(), pwd_len)
else:
# this method is slower under python, but uses a fixed amount of memory.
tmp_ctx = hash_const(pwd)
tmp_ctx_update = tmp_ctx.update
i = pwd_len - 1
while i:
tmp_ctx_update(pwd)
i -= 1
dp = repeat_string(tmp_ctx.digest(), pwd_len)
assert len(dp) == pwd_len
# ===================================================================
# digest S - used instead of salt itself when calculating digest C
# ===================================================================
ds = hash_const(salt * (16 + byte_elem_value(da[0]))).digest()[:salt_len]
assert len(ds) == salt_len, "salt_len somehow > hash_len!"
# ===================================================================
# digest C - for a variable number of rounds, combine A, S, and P
# digests in various ways; in order to burn CPU time.
# ===================================================================
# NOTE: the original SHA256/512-Crypt specification performs the C digest
# calculation using the following loop:
#
##dc = da
##i = 0
# while i < rounds:
## tmp_ctx = hash_const(dp if i & 1 else dc)
# if i % 3:
# tmp_ctx.update(ds)
# if i % 7:
# tmp_ctx.update(dp)
## tmp_ctx.update(dc if i & 1 else dp)
## dc = tmp_ctx.digest()
## i += 1
#
# The code Passlib uses (below) implements an equivalent algorithm,
# it's just been heavily optimized to pre-calculate a large number
# of things beforehand. It works off of a couple of observations
# about the original algorithm:
#
# 1. each round is a combination of 'dc', 'ds', and 'dp'; determined
# by the whether 'i' a multiple of 2,3, and/or 7.
# 2. since lcm(2,3,7)==42, the series of combinations will repeat
# every 42 rounds.
# 3. even rounds 0-40 consist of 'hash(dc + round-specific-constant)';
# while odd rounds 1-41 consist of hash(round-specific-constant + dc)
#
# Using these observations, the following code...
# * calculates the round-specific combination of ds & dp for each round 0-41
# * runs through as many 42-round blocks as possible
# * runs through as many pairs of rounds as possible for remaining rounds
# * performs once last round if the total rounds should be odd.
#
# this cuts out a lot of the control overhead incurred when running the
# original loop 40,000+ times in python, resulting in ~20% increase in
# speed under CPython (though still 2x slower than glibc crypt)
# prepare the 6 combinations of ds & dp which are needed
# (order of 'perms' must match how _c_digest_offsets was generated)
dp_dp = dp + dp
dp_ds = dp + ds
perms = [dp, dp_dp, dp_ds, dp_ds + dp, ds + dp, ds + dp_dp]
# build up list of even-round & odd-round constants,
# and store in 21-element list as (even,odd) pairs.
data = [(perms[even], perms[odd]) for even, odd in _c_digest_offsets]
# perform as many full 42-round blocks as possible
dc = da
blocks, tail = divmod(rounds, 42)
while blocks:
for even, odd in data:
dc = hash_const(odd + hash_const(dc + even).digest()).digest()
blocks -= 1
# perform any leftover rounds
if tail:
# perform any pairs of rounds
pairs = tail >> 1
for even, odd in data[:pairs]:
dc = hash_const(odd + hash_const(dc + even).digest()).digest()
# if rounds was odd, do one last round (since we started at 0,
# last round will be an even-numbered round)
if tail & 1:
dc = hash_const(dc + data[pairs][0]).digest()
# ===================================================================
# encode digest using appropriate transpose map
# ===================================================================
return h64.encode_transposed_bytes(dc, transpose_map).decode("ascii")
def _calc_checksum(self, secret):
"""
This function implements the "encrypted" hash format used by Cisco
PIX & ASA. It's behavior has been confirmed for ASA 9.6,
but is presumed correct for PIX & other ASA releases,
as it fits with known test vectors, and existing literature.
While nearly the same, the PIX & ASA hashes have slight differences,
so this function performs differently based on the _is_asa class flag.
Noteable changes from PIX to ASA include password size limit
increased from 16 -> 32, and other internal changes.
"""
# select PIX vs or ASA mode
asa = self._is_asa
#
# encode secret
#
# per ASA 8.4 documentation,
# http://www.cisco.com/c/en/us/td/docs/security/asa/asa84/configuration/guide/asa_84_cli_config/ref_cli.html#Supported_Character_Sets,
# it supposedly uses UTF-8 -- though some double-encoding issues have
# been observed when trying to actually *set* a non-ascii password
# via ASDM, and access via SSH seems to strip 8-bit chars.
#
if isinstance(secret, unicode):
secret = secret.encode("utf-8")
#
# check if password too large
#
# Per ASA 9.6 changes listed in
# http://www.cisco.com/c/en/us/td/docs/security/asa/roadmap/asa_new_features.html,
# prior releases had a maximum limit of 32 characters.
# Testing with an ASA 9.6 system bears this out --
# setting 32-char password for a user account,
# and logins will fail if any chars are appended.
# (ASA 9.6 added new PBKDF2-based hash algorithm,
# which supports larger passwords).
#
# Per PIX documentation
# http://www.cisco.com/en/US/docs/security/pix/pix50/configuration/guide/commands.html,
# it would not allow passwords > 16 chars.
#
# Thus, we unconditionally throw a password size error here,
# as nothing valid can come from a larger password.
# NOTE: assuming PIX has same behavior, but at 16 char limit.
#
spoil_digest = None
if len(secret) > self.truncate_size:
if self.use_defaults:
# called from hash()
msg = "Password too long (%s allows at most %d bytes)" % (
self.name,
self.truncate_size,
)
raise uh.exc.PasswordSizeError(self.truncate_size, msg=msg)
else:
# called from verify() --
# We don't want to throw error, or return early,
# as that would let attacker know too much. Instead, we set a
# flag to add some dummy data into the md5 digest, so that
# output won't match truncated version of secret, or anything
# else that's fixed and predictable.
spoil_digest = secret + _DUMMY_BYTES
#
# append user to secret
#
# Policy appears to be:
#
# * Nothing appended for enable password (user = "")
#
# * ASA: If user present, but secret is >= 28 chars, nothing appended.
#
# * 1-2 byte users not allowed.
# DEVIATION: we're letting them through, and repeating their
# chars ala 3-char user, to simplify testing.
# Could issue warning in the future though.
#
# * 3 byte user has first char repeated, to pad to 4.
# (observed under ASA 9.6, assuming true elsewhere)
#
# * 4 byte users are used directly.
#
# * 5+ byte users are truncated to 4 bytes.
#
user = self.user
if user:
if isinstance(user, unicode):
user = user.encode("utf-8")
if not asa or len(secret) < 28:
secret += repeat_string(user, 4)
#
# pad / truncate result to limit
#
# While PIX always pads to 16 bytes, ASA increases to 32 bytes IFF
# secret+user > 16 bytes. This makes PIX & ASA have different results
# where secret size in range(13,16), and user is present --
# PIX will truncate to 16, ASA will truncate to 32.
#
if asa and len(secret) > 16:
pad_size = 32
else:
pad_size = 16
secret = right_pad_string(secret, pad_size)
#
# md5 digest
#
if spoil_digest:
# make sure digest won't match truncated version of secret
secret += spoil_digest
digest = md5(secret).digest()
#
# drop every 4th byte
# NOTE: guessing this was done because it makes output exactly
# 16 bytes, which may have been a general 'char password[]'
# size limit under PIX
#
digest = join_byte_elems(c for i, c in enumerate(digest)
if (i + 1) & 3)
#
# encode using Hash64
#
return h64.encode_bytes(digest).decode("ascii")
def _norm_digest_args(cls, secret, ident, new=False):
# make sure secret is unicode
if isinstance(secret, unicode):
secret = secret.encode("utf-8")
# check max secret size
uh.validate_secret(secret)
# check for truncation (during .hash() calls only)
if new:
cls._check_truncate_policy(secret)
# NOTE: especially important to forbid NULLs for bcrypt, since many
# backends (bcryptor, bcrypt) happily accept them, and then
# silently truncate the password at first NULL they encounter!
if _BNULL in secret:
raise uh.exc.NullPasswordError(cls)
# TODO: figure out way to skip these tests when not needed...
# protect from wraparound bug by truncating secret before handing it to the backend.
# bcrypt only uses first 72 bytes anyways.
# NOTE: not needed for 2y/2b, but might use 2a as fallback for them.
if cls._has_2a_wraparound_bug and len(secret) >= 255:
secret = secret[:72]
# special case handling for variants (ordered most common first)
if ident == IDENT_2A:
# nothing needs to be done.
pass
elif ident == IDENT_2B:
if cls._lacks_2b_support:
# handle $2b$ hash format even if backend is too old.
# have it generate a 2A/2Y digest, then return it as a 2B hash.
# 2a-only backend could potentially exhibit wraparound bug --
# but we work around that issue above.
ident = cls._fallback_ident
elif ident == IDENT_2Y:
if cls._lacks_2y_support:
# handle $2y$ hash format (not supported by BSDs, being phased out on others)
# have it generate a 2A/2B digest, then return it as a 2Y hash.
ident = cls._fallback_ident
elif ident == IDENT_2:
if cls._lacks_20_support:
# handle legacy $2$ format (not supported by most backends except BSD os_crypt)
# we can fake $2$ behavior using the 2A/2Y/2B algorithm
# by repeating the password until it's at least 72 chars in length.
if secret:
secret = repeat_string(secret, 72)
ident = cls._fallback_ident
elif ident == IDENT_2X:
# NOTE: shouldn't get here.
# XXX: could check if backend does actually offer 'support'
raise RuntimeError(
"$2x$ hashes not currently supported by passlib")
else:
raise AssertionError("unexpected ident value: %r" % ident)
return secret, ident
def _raw_md5_crypt(pwd, salt, use_apr=False):
"""perform raw md5-crypt calculation
this function provides a pure-python implementation of the internals
for the MD5-Crypt algorithms; it doesn't handle any of the
parsing/validation of the hash strings themselves.
:arg pwd: password chars/bytes to encrypt
:arg salt: salt chars to use
:arg use_apr: use apache variant
:returns:
encoded checksum chars
"""
# NOTE: regarding 'apr' format:
# really, apache? you had to invent a whole new "$apr1$" format,
# when all you did was change the ident incorporated into the hash?
# would love to find webpage explaining why just using a portable
# implementation of $1$ wasn't sufficient. *nothing else* was changed.
#===================================================================
# init & validate inputs
#===================================================================
# validate secret
# XXX: not sure what official unicode policy is, using this as default
if isinstance(pwd, unicode):
pwd = pwd.encode("utf-8")
assert isinstance(pwd, bytes), "pwd not unicode or bytes"
if _BNULL in pwd:
raise uh.exc.NullPasswordError(md5_crypt)
pwd_len = len(pwd)
# validate salt - should have been taken care of by caller
assert isinstance(salt, unicode), "salt not unicode"
salt = salt.encode("ascii")
assert len(salt) < 9, "salt too large"
# NOTE: spec says salts larger than 8 bytes should be truncated,
# instead of causing an error. this function assumes that's been
# taken care of by the handler class.
# load APR specific constants
if use_apr:
magic = _APR_MAGIC
else:
magic = _MD5_MAGIC
#===================================================================
# digest B - used as subinput to digest A
#===================================================================
db = md5(pwd + salt + pwd).digest()
#===================================================================
# digest A - used to initialize first round of digest C
#===================================================================
# start out with pwd + magic + salt
a_ctx = md5(pwd + magic + salt)
a_ctx_update = a_ctx.update
# add pwd_len bytes of b, repeating b as many times as needed.
a_ctx_update(repeat_string(db, pwd_len))
# add null chars & first char of password
# NOTE: this may have historically been a bug,
# where they meant to use db[0] instead of B_NULL,
# but the original code memclear'ed db,
# and now all implementations have to use this.
i = pwd_len
evenchar = pwd[:1]
while i:
a_ctx_update(_BNULL if i & 1 else evenchar)
i >>= 1
# finish A
da = a_ctx.digest()
#===================================================================
# digest C - for a 1000 rounds, combine A, S, and P
# digests in various ways; in order to burn CPU time.
#===================================================================
# NOTE: the original MD5-Crypt implementation performs the C digest
# calculation using the following loop:
#
##dc = da
##i = 0
##while i < rounds:
## tmp_ctx = md5(pwd if i & 1 else dc)
## if i % 3:
## tmp_ctx.update(salt)
## if i % 7:
## tmp_ctx.update(pwd)
## tmp_ctx.update(dc if i & 1 else pwd)
## dc = tmp_ctx.digest()
## i += 1
#
# The code Passlib uses (below) implements an equivalent algorithm,
# it's just been heavily optimized to pre-calculate a large number
# of things beforehand. It works off of a couple of observations
# about the original algorithm:
#
# 1. each round is a combination of 'dc', 'salt', and 'pwd'; and the exact
# combination is determined by whether 'i' a multiple of 2,3, and/or 7.
# 2. since lcm(2,3,7)==42, the series of combinations will repeat
# every 42 rounds.
# 3. even rounds 0-40 consist of 'hash(dc + round-specific-constant)';
# while odd rounds 1-41 consist of hash(round-specific-constant + dc)
#
# Using these observations, the following code...
# * calculates the round-specific combination of salt & pwd for each round 0-41
# * runs through as many 42-round blocks as possible (23)
# * runs through as many pairs of rounds as needed for remaining rounds (17)
# * this results in the required 42*23+2*17=1000 rounds required by md5_crypt.
#
# this cuts out a lot of the control overhead incurred when running the
# original loop 1000 times in python, resulting in ~20% increase in
# speed under CPython (though still 2x slower than glibc crypt)
# prepare the 6 combinations of pwd & salt which are needed
# (order of 'perms' must match how _c_digest_offsets was generated)
pwd_pwd = pwd + pwd
pwd_salt = pwd + salt
perms = [
pwd, pwd_pwd, pwd_salt, pwd_salt + pwd, salt + pwd, salt + pwd_pwd
]
# build up list of even-round & odd-round constants,
# and store in 21-element list as (even,odd) pairs.
data = [(perms[even], perms[odd]) for even, odd in _c_digest_offsets]
# perform 23 blocks of 42 rounds each (for a total of 966 rounds)
dc = da
blocks = 23
while blocks:
for even, odd in data:
dc = md5(odd + md5(dc + even).digest()).digest()
blocks -= 1
# perform 17 more pairs of rounds (34 more rounds, for a total of 1000)
for even, odd in data[:17]:
dc = md5(odd + md5(dc + even).digest()).digest()
#===================================================================
# encode digest using appropriate transpose map
#===================================================================
return h64.encode_transposed_bytes(dc, _transpose_map).decode("ascii")
def _calc_checksum(self, secret):
"""
This function implements the "encrypted" hash format used by Cisco
PIX & ASA. It's behavior has been confirmed for ASA 9.6,
but is presumed correct for PIX & other ASA releases,
as it fits with known test vectors, and existing literature.
While nearly the same, the PIX & ASA hashes have slight differences,
so this function performs differently based on the _is_asa class flag.
Noteable changes from PIX to ASA include password size limit
increased from 16 -> 32, and other internal changes.
"""
# select PIX vs or ASA mode
asa = self._is_asa
#
# encode secret
#
# per ASA 8.4 documentation,
# http://www.cisco.com/c/en/us/td/docs/security/asa/asa84/configuration/guide/asa_84_cli_config/ref_cli.html#Supported_Character_Sets,
# it supposedly uses UTF-8 -- though some double-encoding issues have
# been observed when trying to actually *set* a non-ascii password
# via ASDM, and access via SSH seems to strip 8-bit chars.
#
if isinstance(secret, unicode):
secret = secret.encode("utf-8")
#
# check if password too large
#
# Per ASA 9.6 changes listed in
# http://www.cisco.com/c/en/us/td/docs/security/asa/roadmap/asa_new_features.html,
# prior releases had a maximum limit of 32 characters.
# Testing with an ASA 9.6 system bears this out --
# setting 32-char password for a user account,
# and logins will fail if any chars are appended.
# (ASA 9.6 added new PBKDF2-based hash algorithm,
# which supports larger passwords).
#
# Per PIX documentation
# http://www.cisco.com/en/US/docs/security/pix/pix50/configuration/guide/commands.html,
# it would not allow passwords > 16 chars.
#
# Thus, we unconditionally throw a password size error here,
# as nothing valid can come from a larger password.
# NOTE: assuming PIX has same behavior, but at 16 char limit.
#
spoil_digest = None
if len(secret) > self.truncate_size:
if self.use_defaults:
# called from hash()
msg = "Password too long (%s allows at most %d bytes)" % \
(self.name, self.truncate_size)
raise uh.exc.PasswordSizeError(self.truncate_size, msg=msg)
else:
# called from verify() --
# We don't want to throw error, or return early,
# as that would let attacker know too much. Instead, we set a
# flag to add some dummy data into the md5 digest, so that
# output won't match truncated version of secret, or anything
# else that's fixed and predictable.
spoil_digest = secret + _DUMMY_BYTES
#
# append user to secret
#
# Policy appears to be:
#
# * Nothing appended for enable password (user = "")
#
# * ASA: If user present, but secret is >= 28 chars, nothing appended.
#
# * 1-2 byte users not allowed.
# DEVIATION: we're letting them through, and repeating their
# chars ala 3-char user, to simplify testing.
# Could issue warning in the future though.
#
# * 3 byte user has first char repeated, to pad to 4.
# (observed under ASA 9.6, assuming true elsewhere)
#
# * 4 byte users are used directly.
#
# * 5+ byte users are truncated to 4 bytes.
#
user = self.user
if user:
if isinstance(user, unicode):
user = user.encode("utf-8")
if not asa or len(secret) < 28:
secret += repeat_string(user, 4)
#
# pad / truncate result to limit
#
# While PIX always pads to 16 bytes, ASA increases to 32 bytes IFF
# secret+user > 16 bytes. This makes PIX & ASA have different results
# where secret size in range(13,16), and user is present --
# PIX will truncate to 16, ASA will truncate to 32.
#
if asa and len(secret) > 16:
pad_size = 32
else:
pad_size = 16
secret = right_pad_string(secret, pad_size)
#
# md5 digest
#
if spoil_digest:
# make sure digest won't match truncated version of secret
secret += spoil_digest
digest = md5(secret).digest()
#
# drop every 4th byte
# NOTE: guessing this was done because it makes output exactly
# 16 bytes, which may have been a general 'char password[]'
# size limit under PIX
#
digest = join_byte_elems(c for i, c in enumerate(digest) if (i + 1) & 3)
#
# encode using Hash64
#
return h64.encode_bytes(digest).decode("ascii")
def _raw_sha2_crypt(pwd, salt, rounds, use_512=False):
"""perform raw sha256-crypt / sha512-crypt
this function provides a pure-python implementation of the internals
for the SHA256-Crypt and SHA512-Crypt algorithms; it doesn't
handle any of the parsing/validation of the hash strings themselves.
:arg pwd: password chars/bytes to encrypt
:arg salt: salt chars to use
:arg rounds: linear rounds cost
:arg use_512: use sha512-crypt instead of sha256-crypt mode
:returns:
encoded checksum chars
"""
#===================================================================
# init & validate inputs
#===================================================================
# NOTE: the setup portion of this algorithm scales ~linearly in time
# with the size of the password, making it vulnerable to a DOS from
# unreasonably large inputs. the following code has some optimizations
# which would make things even worse, using O(pwd_len**2) memory
# when calculating digest P.
#
# to mitigate these two issues: 1) this code switches to a
# O(pwd_len)-memory algorithm for passwords that are much larger
# than average, and 2) Passlib enforces a library-wide max limit on
# the size of passwords it will allow, to prevent this algorithm and
# others from being DOSed in this way (see passlib.exc.PasswordSizeError
# for details).
# validate secret
if isinstance(pwd, unicode):
# XXX: not sure what official unicode policy is, using this as default
pwd = pwd.encode("utf-8")
assert isinstance(pwd, bytes)
if _BNULL in pwd:
raise uh.exc.NullPasswordError(sha512_crypt if use_512 else sha256_crypt)
pwd_len = len(pwd)
# validate rounds
assert 1000 <= rounds <= 999999999, "invalid rounds"
# NOTE: spec says out-of-range rounds should be clipped, instead of
# causing an error. this function assumes that's been taken care of
# by the handler class.
# validate salt
assert isinstance(salt, unicode), "salt not unicode"
salt = salt.encode("ascii")
salt_len = len(salt)
assert salt_len < 17, "salt too large"
# NOTE: spec says salts larger than 16 bytes should be truncated,
# instead of causing an error. this function assumes that's been
# taken care of by the handler class.
# load sha256/512 specific constants
if use_512:
hash_const = hashlib.sha512
hash_len = 64
transpose_map = _512_transpose_map
else:
hash_const = hashlib.sha256
hash_len = 32
transpose_map = _256_transpose_map
#===================================================================
# digest B - used as subinput to digest A
#===================================================================
db = hash_const(pwd + salt + pwd).digest()
#===================================================================
# digest A - used to initialize first round of digest C
#===================================================================
# start out with pwd + salt
a_ctx = hash_const(pwd + salt)
a_ctx_update = a_ctx.update
# add pwd_len bytes of b, repeating b as many times as needed.
a_ctx_update(repeat_string(db, pwd_len))
# for each bit in pwd_len: add b if it's 1, or pwd if it's 0
i = pwd_len
while i:
a_ctx_update(db if i & 1 else pwd)
i >>= 1
# finish A
da = a_ctx.digest()
#===================================================================
# digest P from password - used instead of password itself
# when calculating digest C.
#===================================================================
if pwd_len < 96:
# this method is faster under python, but uses O(pwd_len**2) memory;
# so we don't use it for larger passwords to avoid a potential DOS.
dp = repeat_string(hash_const(pwd * pwd_len).digest(), pwd_len)
else:
# this method is slower under python, but uses a fixed amount of memory.
tmp_ctx = hash_const(pwd)
tmp_ctx_update = tmp_ctx.update
i = pwd_len-1
while i:
tmp_ctx_update(pwd)
i -= 1
dp = repeat_string(tmp_ctx.digest(), pwd_len)
assert len(dp) == pwd_len
#===================================================================
# digest S - used instead of salt itself when calculating digest C
#===================================================================
ds = hash_const(salt * (16 + byte_elem_value(da[0]))).digest()[:salt_len]
assert len(ds) == salt_len, "salt_len somehow > hash_len!"
#===================================================================
# digest C - for a variable number of rounds, combine A, S, and P
# digests in various ways; in order to burn CPU time.
#===================================================================
# NOTE: the original SHA256/512-Crypt specification performs the C digest
# calculation using the following loop:
#
##dc = da
##i = 0
##while i < rounds:
## tmp_ctx = hash_const(dp if i & 1 else dc)
## if i % 3:
## tmp_ctx.update(ds)
## if i % 7:
## tmp_ctx.update(dp)
## tmp_ctx.update(dc if i & 1 else dp)
## dc = tmp_ctx.digest()
## i += 1
#
# The code Passlib uses (below) implements an equivalent algorithm,
# it's just been heavily optimized to pre-calculate a large number
# of things beforehand. It works off of a couple of observations
# about the original algorithm:
#
# 1. each round is a combination of 'dc', 'ds', and 'dp'; determined
# by the whether 'i' a multiple of 2,3, and/or 7.
# 2. since lcm(2,3,7)==42, the series of combinations will repeat
# every 42 rounds.
# 3. even rounds 0-40 consist of 'hash(dc + round-specific-constant)';
# while odd rounds 1-41 consist of hash(round-specific-constant + dc)
#
# Using these observations, the following code...
# * calculates the round-specific combination of ds & dp for each round 0-41
# * runs through as many 42-round blocks as possible
# * runs through as many pairs of rounds as possible for remaining rounds
# * performs once last round if the total rounds should be odd.
#
# this cuts out a lot of the control overhead incurred when running the
# original loop 40,000+ times in python, resulting in ~20% increase in
# speed under CPython (though still 2x slower than glibc crypt)
# prepare the 6 combinations of ds & dp which are needed
# (order of 'perms' must match how _c_digest_offsets was generated)
dp_dp = dp+dp
dp_ds = dp+ds
perms = [dp, dp_dp, dp_ds, dp_ds+dp, ds+dp, ds+dp_dp]
# build up list of even-round & odd-round constants,
# and store in 21-element list as (even,odd) pairs.
data = [ (perms[even], perms[odd]) for even, odd in _c_digest_offsets]
# perform as many full 42-round blocks as possible
dc = da
blocks, tail = divmod(rounds, 42)
while blocks:
for even, odd in data:
dc = hash_const(odd + hash_const(dc + even).digest()).digest()
blocks -= 1
# perform any leftover rounds
if tail:
# perform any pairs of rounds
pairs = tail>>1
for even, odd in data[:pairs]:
dc = hash_const(odd + hash_const(dc + even).digest()).digest()
# if rounds was odd, do one last round (since we started at 0,
# last round will be an even-numbered round)
if tail & 1:
dc = hash_const(dc + data[pairs][0]).digest()
#===================================================================
# encode digest using appropriate transpose map
#===================================================================
return h64.encode_transposed_bytes(dc, transpose_map).decode("ascii")
def _norm_digest_args(cls, secret, ident, new=False):
# make sure secret is unicode
if isinstance(secret, unicode):
secret = secret.encode("utf-8")
# check max secret size
uh.validate_secret(secret)
# check for truncation (during .hash() calls only)
if new:
cls._check_truncate_policy(secret)
# NOTE: especially important to forbid NULLs for bcrypt, since many
# backends (bcryptor, bcrypt) happily accept them, and then
# silently truncate the password at first NULL they encounter!
if _BNULL in secret:
raise uh.exc.NullPasswordError(cls)
# TODO: figure out way to skip these tests when not needed...
# protect from wraparound bug by truncating secret before handing it to the backend.
# bcrypt only uses first 72 bytes anyways.
# NOTE: not needed for 2y/2b, but might use 2a as fallback for them.
if cls._has_2a_wraparound_bug and len(secret) >= 255:
secret = secret[:72]
# special case handling for variants (ordered most common first)
if ident == IDENT_2A:
# nothing needs to be done.
pass
elif ident == IDENT_2B:
if cls._lacks_2b_support:
# handle $2b$ hash format even if backend is too old.
# have it generate a 2A/2Y digest, then return it as a 2B hash.
# 2a-only backend could potentially exhibit wraparound bug --
# but we work around that issue above.
ident = cls._fallback_ident
elif ident == IDENT_2Y:
if cls._lacks_2y_support:
# handle $2y$ hash format (not supported by BSDs, being phased out on others)
# have it generate a 2A/2B digest, then return it as a 2Y hash.
ident = cls._fallback_ident
elif ident == IDENT_2:
if cls._lacks_20_support:
# handle legacy $2$ format (not supported by most backends except BSD os_crypt)
# we can fake $2$ behavior using the 2A/2Y/2B algorithm
# by repeating the password until it's at least 72 chars in length.
if secret:
secret = repeat_string(secret, 72)
ident = cls._fallback_ident
elif ident == IDENT_2X:
# NOTE: shouldn't get here.
# XXX: could check if backend does actually offer 'support'
raise RuntimeError("$2x$ hashes not currently supported by passlib")
else:
raise AssertionError("unexpected ident value: %r" % ident)
return secret, ident
class _bcrypt_sha256_test(HandlerCase):
"base for BCrypt-SHA256 test cases"
handler = hash.bcrypt_sha256
reduce_default_rounds = True
forbidden_characters = None
fuzz_salts_need_bcrypt_repair = True
fallback_os_crypt_handler = hash.bcrypt
known_correct_hashes = [
#
# custom test vectors
#
# empty
("",
'$bcrypt-sha256$2a,5$E/e/2AOhqM5W/KJTFQzLce$F6dYSxOdAEoJZO2eoHUZWZljW/e0TXO'
),
# ascii
("password",
'$bcrypt-sha256$2a,5$5Hg1DKFqPE8C2aflZ5vVoe$12BjNE0p7axMg55.Y/mHsYiVuFBDQyu'
),
# unicode / utf8
(UPASS_TABLE,
'$bcrypt-sha256$2a,5$.US1fQ4TQS.ZTz/uJ5Kyn.$QNdPDOTKKT5/sovNz1iWg26quOU4Pje'
),
(UPASS_TABLE.encode("utf-8"),
'$bcrypt-sha256$2a,5$.US1fQ4TQS.ZTz/uJ5Kyn.$QNdPDOTKKT5/sovNz1iWg26quOU4Pje'
),
# test >72 chars is hashed correctly -- under bcrypt these hash the same.
# NOTE: test_60_secret_size() handles this already, this is just for overkill :)
(repeat_string("abc123", 72),
'$bcrypt-sha256$2a,5$X1g1nh3g0v4h6970O68cxe$r/hyEtqJ0teqPEmfTLoZ83ciAI1Q74.'
),
(repeat_string("abc123", 72) + "qwr",
'$bcrypt-sha256$2a,5$X1g1nh3g0v4h6970O68cxe$021KLEif6epjot5yoxk0m8I0929ohEa'
),
(repeat_string("abc123", 72) + "xyz",
'$bcrypt-sha256$2a,5$X1g1nh3g0v4h6970O68cxe$7.1kgpHduMGEjvM3fX6e/QCvfn6OKja'
),
]
known_correct_configs = [
('$bcrypt-sha256$2a,5$5Hg1DKFqPE8C2aflZ5vVoe', "password",
'$bcrypt-sha256$2a,5$5Hg1DKFqPE8C2aflZ5vVoe$12BjNE0p7axMg55.Y/mHsYiVuFBDQyu'
),
]
known_malformed_hashes = [
# bad char in otherwise correct hash
# \/
'$bcrypt-sha256$2a,5$5Hg1DKF!PE8C2aflZ5vVoe$12BjNE0p7axMg55.Y/mHsYiVuFBDQyu',
# unrecognized bcrypt variant
'$bcrypt-sha256$2c,5$5Hg1DKFqPE8C2aflZ5vVoe$12BjNE0p7axMg55.Y/mHsYiVuFBDQyu',
# unsupported bcrypt variant
'$bcrypt-sha256$2x,5$5Hg1DKFqPE8C2aflZ5vVoe$12BjNE0p7axMg55.Y/mHsYiVuFBDQyu',
# rounds zero-padded
'$bcrypt-sha256$2a,05$5Hg1DKFqPE8C2aflZ5vVoe$12BjNE0p7axMg55.Y/mHsYiVuFBDQyu',
# config string w/ $ added
'$bcrypt-sha256$2a,5$5Hg1DKFqPE8C2aflZ5vVoe$',
]
#===================================================================
# override some methods -- cloned from bcrypt
#===================================================================
def setUp(self):
# ensure builtin is enabled for duration of test.
if TEST_MODE("full") and self.backend == "builtin":
key = "PASSLIB_BUILTIN_BCRYPT"
orig = os.environ.get(key)
if orig:
self.addCleanup(os.environ.__setitem__, key, orig)
else:
self.addCleanup(os.environ.__delitem__, key)
os.environ[key] = "enabled"
super(_bcrypt_sha256_test, self).setUp()
def populate_settings(self, kwds):
# builtin is still just way too slow.
if self.backend == "builtin":
kwds.setdefault("rounds", 4)
super(_bcrypt_sha256_test, self).populate_settings(kwds)
#===================================================================
# override ident tests for now
#===================================================================
def test_30_HasManyIdents(self):
raise self.skipTest("multiple idents not supported")
def test_30_HasOneIdent(self):
# forbidding ident keyword, we only support "2a" for now
handler = self.handler
handler(use_defaults=True)
self.assertRaises(ValueError, handler, ident="$2y$", use_defaults=True)
#===================================================================
# fuzz testing -- cloned from bcrypt
#===================================================================
def fuzz_setting_rounds(self):
# decrease default rounds for fuzz testing to speed up volume.
return randintgauss(5, 8, 6, 1)
def _raw_md5_crypt(pwd, salt, use_apr=False):
"""perform raw md5-crypt calculation
this function provides a pure-python implementation of the internals
for the MD5-Crypt algorithms; it doesn't handle any of the
parsing/validation of the hash strings themselves.
:arg pwd: password chars/bytes to encrypt
:arg salt: salt chars to use
:arg use_apr: use apache variant
:returns:
encoded checksum chars
"""
# NOTE: regarding 'apr' format:
# really, apache? you had to invent a whole new "$apr1$" format,
# when all you did was change the ident incorporated into the hash?
# would love to find webpage explaining why just using a portable
# implementation of $1$ wasn't sufficient. *nothing else* was changed.
#===================================================================
# init & validate inputs
#===================================================================
# validate secret
# XXX: not sure what official unicode policy is, using this as default
if isinstance(pwd, unicode):
pwd = pwd.encode("utf-8")
assert isinstance(pwd, bytes), "pwd not unicode or bytes"
if _BNULL in pwd:
raise uh.exc.NullPasswordError(md5_crypt)
pwd_len = len(pwd)
# validate salt - should have been taken care of by caller
assert isinstance(salt, unicode), "salt not unicode"
salt = salt.encode("ascii")
assert len(salt) < 9, "salt too large"
# NOTE: spec says salts larger than 8 bytes should be truncated,
# instead of causing an error. this function assumes that's been
# taken care of by the handler class.
# load APR specific constants
if use_apr:
magic = _APR_MAGIC
else:
magic = _MD5_MAGIC
#===================================================================
# digest B - used as subinput to digest A
#===================================================================
db = md5(pwd + salt + pwd).digest()
#===================================================================
# digest A - used to initialize first round of digest C
#===================================================================
# start out with pwd + magic + salt
a_ctx = md5(pwd + magic + salt)
a_ctx_update = a_ctx.update
# add pwd_len bytes of b, repeating b as many times as needed.
a_ctx_update(repeat_string(db, pwd_len))
# add null chars & first char of password
# NOTE: this may have historically been a bug,
# where they meant to use db[0] instead of B_NULL,
# but the original code memclear'ed db,
# and now all implementations have to use this.
i = pwd_len
evenchar = pwd[:1]
while i:
a_ctx_update(_BNULL if i & 1 else evenchar)
i >>= 1
# finish A
da = a_ctx.digest()
#===================================================================
# digest C - for a 1000 rounds, combine A, S, and P
# digests in various ways; in order to burn CPU time.
#===================================================================
# NOTE: the original MD5-Crypt implementation performs the C digest
# calculation using the following loop:
#
##dc = da
##i = 0
##while i < rounds:
## tmp_ctx = md5(pwd if i & 1 else dc)
## if i % 3:
## tmp_ctx.update(salt)
## if i % 7:
## tmp_ctx.update(pwd)
## tmp_ctx.update(dc if i & 1 else pwd)
## dc = tmp_ctx.digest()
## i += 1
#
# The code Passlib uses (below) implements an equivalent algorithm,
# it's just been heavily optimized to pre-calculate a large number
# of things beforehand. It works off of a couple of observations
# about the original algorithm:
#
# 1. each round is a combination of 'dc', 'salt', and 'pwd'; and the exact
# combination is determined by whether 'i' a multiple of 2,3, and/or 7.
# 2. since lcm(2,3,7)==42, the series of combinations will repeat
# every 42 rounds.
# 3. even rounds 0-40 consist of 'hash(dc + round-specific-constant)';
# while odd rounds 1-41 consist of hash(round-specific-constant + dc)
#
# Using these observations, the following code...
# * calculates the round-specific combination of salt & pwd for each round 0-41
# * runs through as many 42-round blocks as possible (23)
# * runs through as many pairs of rounds as needed for remaining rounds (17)
# * this results in the required 42*23+2*17=1000 rounds required by md5_crypt.
#
# this cuts out a lot of the control overhead incurred when running the
# original loop 1000 times in python, resulting in ~20% increase in
# speed under CPython (though still 2x slower than glibc crypt)
# prepare the 6 combinations of pwd & salt which are needed
# (order of 'perms' must match how _c_digest_offsets was generated)
pwd_pwd = pwd+pwd
pwd_salt = pwd+salt
perms = [pwd, pwd_pwd, pwd_salt, pwd_salt+pwd, salt+pwd, salt+pwd_pwd]
# build up list of even-round & odd-round constants,
# and store in 21-element list as (even,odd) pairs.
data = [ (perms[even], perms[odd]) for even, odd in _c_digest_offsets]
# perform 23 blocks of 42 rounds each (for a total of 966 rounds)
dc = da
blocks = 23
while blocks:
for even, odd in data:
dc = md5(odd + md5(dc + even).digest()).digest()
blocks -= 1
# perform 17 more pairs of rounds (34 more rounds, for a total of 1000)
for even, odd in data[:17]:
dc = md5(odd + md5(dc + even).digest()).digest()
#===================================================================
# encode digest using appropriate transpose map
#===================================================================
return h64.encode_transposed_bytes(dc, _transpose_map).decode("ascii")
class _bcrypt_sha256_test(HandlerCase):
"base for BCrypt-SHA256 test cases"
handler = hash.bcrypt_sha256
reduce_default_rounds = True
forbidden_characters = None
fuzz_salts_need_bcrypt_repair = True
known_correct_hashes = [
# -------------------------------------------------------------------
# custom test vectors for old v1 format
# -------------------------------------------------------------------
# empty
("",
'$bcrypt-sha256$2a,5$E/e/2AOhqM5W/KJTFQzLce$F6dYSxOdAEoJZO2eoHUZWZljW/e0TXO'
),
# ascii
("password",
'$bcrypt-sha256$2a,5$5Hg1DKFqPE8C2aflZ5vVoe$12BjNE0p7axMg55.Y/mHsYiVuFBDQyu'
),
# unicode / utf8
(UPASS_TABLE,
'$bcrypt-sha256$2a,5$.US1fQ4TQS.ZTz/uJ5Kyn.$QNdPDOTKKT5/sovNz1iWg26quOU4Pje'
),
(UPASS_TABLE.encode("utf-8"),
'$bcrypt-sha256$2a,5$.US1fQ4TQS.ZTz/uJ5Kyn.$QNdPDOTKKT5/sovNz1iWg26quOU4Pje'
),
# ensure 2b support
("password",
'$bcrypt-sha256$2b,5$5Hg1DKFqPE8C2aflZ5vVoe$12BjNE0p7axMg55.Y/mHsYiVuFBDQyu'
),
(UPASS_TABLE,
'$bcrypt-sha256$2b,5$.US1fQ4TQS.ZTz/uJ5Kyn.$QNdPDOTKKT5/sovNz1iWg26quOU4Pje'
),
# test >72 chars is hashed correctly -- under bcrypt these hash the same.
# NOTE: test_60_truncate_size() handles this already, this is just for overkill :)
(repeat_string("abc123", 72),
'$bcrypt-sha256$2b,5$X1g1nh3g0v4h6970O68cxe$r/hyEtqJ0teqPEmfTLoZ83ciAI1Q74.'
),
(repeat_string("abc123", 72) + "qwr",
'$bcrypt-sha256$2b,5$X1g1nh3g0v4h6970O68cxe$021KLEif6epjot5yoxk0m8I0929ohEa'
),
(repeat_string("abc123", 72) + "xyz",
'$bcrypt-sha256$2b,5$X1g1nh3g0v4h6970O68cxe$7.1kgpHduMGEjvM3fX6e/QCvfn6OKja'
),
# -------------------------------------------------------------------
# custom test vectors for v2 format
# TODO: convert to v2 format
# -------------------------------------------------------------------
# empty
("",
'$bcrypt-sha256$v=2,t=2b,r=5$E/e/2AOhqM5W/KJTFQzLce$WFPIZKtDDTriqWwlmRFfHiOTeheAZWe'
),
# ascii
("password",
'$bcrypt-sha256$v=2,t=2b,r=5$5Hg1DKFqPE8C2aflZ5vVoe$wOK1VFFtS8IGTrGa7.h5fs0u84qyPbS'
),
# unicode / utf8
(UPASS_TABLE,
'$bcrypt-sha256$v=2,t=2b,r=5$.US1fQ4TQS.ZTz/uJ5Kyn.$pzzgp40k8reM1CuQb03PvE0IDPQSdV6'
),
(UPASS_TABLE.encode("utf-8"),
'$bcrypt-sha256$v=2,t=2b,r=5$.US1fQ4TQS.ZTz/uJ5Kyn.$pzzgp40k8reM1CuQb03PvE0IDPQSdV6'
),
# test >72 chars is hashed correctly -- under bcrypt these hash the same.
# NOTE: test_60_truncate_size() handles this already, this is just for overkill :)
(repeat_string("abc123", 72),
'$bcrypt-sha256$v=2,t=2b,r=5$X1g1nh3g0v4h6970O68cxe$zu1cloESVFIOsUIo7fCEgkdHaI9SSue'
),
(repeat_string("abc123", 72) + "qwr",
'$bcrypt-sha256$v=2,t=2b,r=5$X1g1nh3g0v4h6970O68cxe$CBF9csfEdW68xv3DwE6xSULXMtqEFP.'
),
(repeat_string("abc123", 72) + "xyz",
'$bcrypt-sha256$v=2,t=2b,r=5$X1g1nh3g0v4h6970O68cxe$zC/1UDUG2ofEXB6Onr2vvyFzfhEOS3S'
),
]
known_correct_configs = [
# v1
('$bcrypt-sha256$2a,5$5Hg1DKFqPE8C2aflZ5vVoe', "password",
'$bcrypt-sha256$2a,5$5Hg1DKFqPE8C2aflZ5vVoe$12BjNE0p7axMg55.Y/mHsYiVuFBDQyu'
),
# v2
('$bcrypt-sha256$v=2,t=2b,r=5$5Hg1DKFqPE8C2aflZ5vVoe', "password",
'$bcrypt-sha256$v=2,t=2b,r=5$5Hg1DKFqPE8C2aflZ5vVoe$wOK1VFFtS8IGTrGa7.h5fs0u84qyPbS'
),
]
known_malformed_hashes = [
# -------------------------------------------------------------------
# v1 format
# -------------------------------------------------------------------
# bad char in otherwise correct hash
# \/
'$bcrypt-sha256$2a,5$5Hg1DKF!PE8C2aflZ5vVoe$12BjNE0p7axMg55.Y/mHsYiVuFBDQyu',
# unrecognized bcrypt variant
'$bcrypt-sha256$2c,5$5Hg1DKFqPE8C2aflZ5vVoe$12BjNE0p7axMg55.Y/mHsYiVuFBDQyu',
# unsupported bcrypt variant
'$bcrypt-sha256$2x,5$5Hg1DKFqPE8C2aflZ5vVoe$12BjNE0p7axMg55.Y/mHsYiVuFBDQyu',
# rounds zero-padded
'$bcrypt-sha256$2a,05$5Hg1DKFqPE8C2aflZ5vVoe$12BjNE0p7axMg55.Y/mHsYiVuFBDQyu',
# config string w/ $ added
'$bcrypt-sha256$2a,5$5Hg1DKFqPE8C2aflZ5vVoe$',
# -------------------------------------------------------------------
# v2 format
# -------------------------------------------------------------------
# bad char in otherwise correct hash
# \/
'$bcrypt-sha256$v=2,t=2b,r=5$5Hg1DKF!PE8C2aflZ5vVoe$12BjNE0p7axMg55.Y/mHsYiVuFBDQyu',
# unsupported version (for this format)
'$bcrypt-sha256$v=1,t=2b,r=5$5Hg1DKFqPE8C2aflZ5vVoe$12BjNE0p7axMg55.Y/mHsYiVuFBDQyu',
# unrecognized version
'$bcrypt-sha256$v=3,t=2b,r=5$5Hg1DKFqPE8C2aflZ5vVoe$12BjNE0p7axMg55.Y/mHsYiVuFBDQyu',
# unrecognized bcrypt variant
'$bcrypt-sha256$v=2,t=2c,r=5$5Hg1DKFqPE8C2aflZ5vVoe$12BjNE0p7axMg55.Y/mHsYiVuFBDQyu',
# unsupported bcrypt variant
'$bcrypt-sha256$v=2,t=2a,r=5$5Hg1DKFqPE8C2aflZ5vVoe$12BjNE0p7axMg55.Y/mHsYiVuFBDQyu',
'$bcrypt-sha256$v=2,t=2x,r=5$5Hg1DKFqPE8C2aflZ5vVoe$12BjNE0p7axMg55.Y/mHsYiVuFBDQyu',
# rounds zero-padded
'$bcrypt-sha256$v=2,t=2b,r=05$5Hg1DKFqPE8C2aflZ5vVoe$12BjNE0p7axMg55.Y/mHsYiVuFBDQyu',
# config string w/ $ added
'$bcrypt-sha256$v=2,t=2b,r=5$5Hg1DKFqPE8C2aflZ5vVoe$',
]
# ===================================================================
# override some methods -- cloned from bcrypt
# ===================================================================
def setUp(self):
# ensure builtin is enabled for duration of test.
if TEST_MODE("full") and self.backend == "builtin":
key = "PASSLIB_BUILTIN_BCRYPT"
orig = os.environ.get(key)
if orig:
self.addCleanup(os.environ.__setitem__, key, orig)
else:
self.addCleanup(os.environ.__delitem__, key)
os.environ[key] = "enabled"
super(_bcrypt_sha256_test, self).setUp()
warnings.filterwarnings(
"ignore", ".*backend is vulnerable to the bsd wraparound bug.*")
def populate_settings(self, kwds):
# builtin is still just way too slow.
if self.backend == "builtin":
kwds.setdefault("rounds", 4)
super(_bcrypt_sha256_test, self).populate_settings(kwds)
# ===================================================================
# override ident tests for now
# ===================================================================
def require_many_idents(self):
raise self.skipTest("multiple idents not supported")
def test_30_HasOneIdent(self):
# forbidding ident keyword, we only support "2b" for now
handler = self.handler
handler(use_defaults=True)
self.assertRaises(ValueError, handler, ident="$2y$", use_defaults=True)
# ===================================================================
# fuzz testing -- cloned from bcrypt
# ===================================================================
class FuzzHashGenerator(HandlerCase.FuzzHashGenerator):
def random_rounds(self):
# decrease default rounds for fuzz testing to speed up volume.
return self.randintgauss(5, 8, 6, 1)
def random_ident(self):
return "2b"
# ===================================================================
# custom tests
# ===================================================================
def test_using_version(self):
# default to v2
handler = self.handler
self.assertEqual(handler.version, 2)
# allow v1 explicitly
subcls = handler.using(version=1)
self.assertEqual(subcls.version, 1)
# forbid unknown ver
self.assertRaises(ValueError, handler.using, version=999)
# allow '2a' only for v1
subcls = handler.using(version=1, ident="2a")
self.assertRaises(ValueError, handler.using, ident="2a")
def test_calc_digest_v2(self):
"""
test digest calc v2 matches bcrypt()
"""
from passlib.hash import bcrypt
from passlib.crypto.digest import compile_hmac
from passlib.utils.binary import b64encode
# manually calc intermediary digest
salt = "nyKYxTAvjmy6lMDYMl11Uu"
secret = "test"
temp_digest = compile_hmac("sha256", salt.encode("ascii"))(
secret.encode("ascii"))
temp_digest = b64encode(temp_digest).decode("ascii")
self.assertEqual(temp_digest,
"J5TlyIDm+IcSWmKiDJm+MeICndBkFVPn4kKdJW8f+xY=")
# manually final hash from intermediary
# XXX: genhash() could be useful here
bcrypt_digest = bcrypt(ident="2b", salt=salt,
rounds=12)._calc_checksum(temp_digest)
self.assertEqual(bcrypt_digest, "M0wE0Ov/9LXoQFCe.jRHu3MSHPF54Ta")
self.assertTrue(
bcrypt.verify(temp_digest, "$2b$12$" + salt + bcrypt_digest))
# confirm handler outputs same thing.
# XXX: genhash() could be useful here
result = self.handler(ident="2b", salt=salt,
rounds=12)._calc_checksum(secret)
self.assertEqual(result, bcrypt_digest)
