Setting the dh_flag to 1 used to indicate that the caller requests safe
primes from mbedtls_mpi_gen_prime. We generalize the functionality to
make room for more flags in that parameter.
If some algorithms are excluded in the build, it's ok for the corresponding
macros not to give the correct results. Therefore the corresponding test cases
should depend on the implementation of the algorithm. For example, it's ok for
PSA_HASH_MAX_SIZE to be less than PSA_HASH_SIZE(PSA_ALG_SHA_512) if we build
without SHA-512 support, and we indeed do this. It's even ok for an
implementation to return 0 for PSA_ALG_IS_HASH(PSA_ALG_SHA_512) if it doesn't
support SHA-512; we return 1 anyway but the tests are less
implementation-specific if we don't enforce it.
This commit adds dependencies on symbols that don't exist in Mbed TLS,
for algorithms that Mbed TLS doesn't implement. These are:
MBEDTLS_SHA512_256 for SHA-512/256, MBEDTLS_SHA3_C for SHA-3,
MBEDTLS_DSA_C and MBEDTLS_DSA_DETERMINISTIC for DSA, and
MBEDTLS_ECP_DP_xxx_ENABLED for elliptic curves that have a PSA
encoding but are not supported in Mbed TLS.
For all key types, validate feature test macros (PSA_KEY_TYPE_IS_xxx).
For asymmetric keys (public key or key pair), validate the
corresponding public/pair type.
For ECC keys, validate GET_CURVE.
For all algorithms, validate feature test macros (PSA_ALG_IS_xxx).
For hash algorithms, validate the exact hash size, and validate
xxx_GET_HASH macros on dependent algorithms.
For MAC algorithms, validate the MAC size. For AEAD algorithms,
validate the tag size.
There is a separate test case for each HMAC algorithm, which is
necessary because each has its own MAC size. For other hash-dependent
algorithms, there is no interesting variation to test here, so only
one hash gets tested.
OFB and CFB are streaming modes. XTS is a not a cipher mode but it
doesn't use a separate padding step. This leaves only CBC as a block
cipher mode that needs a padding step.
Since CBC is the only mode that uses a separate padding step, and is
likely to remain the only mode in the future, encode the padding mode
directly in the algorithm constant, rather than building up an
algorithm value from a chaining mode and a padding mode. This greatly
simplifies the interface as well as some parts of the implementation.
Don't rely on static initialization of a flexible array member, that's
a GNU extension. The previous code also triggered a Clang warning
"suggest braces around initialization of subobject" (-Wmissing-braces)
for `struct {char a[]} = {"foo"}`.
This is not useful to validate the implementation when importing
canonical input, which is the case for most import/export test cases,
but it helps validate the sanity checks themselves.
Implement sanity checks of exported public keys, using ASN.1 parsing.
Rewrite the sanity checks of key pairs using ASN.1 parsing, so as to
check more things with simpler code.
Move the code to perform sanity checks on the exported key from
generate_key to exercise_key. This way the sanity checks can be
performed after importing or deriving a key as well.
In addition to checking the exported key if its usage allows it, check
the exported public key if the key is asymmetric.
The last slot in the array was not freed due to an off-by-one error.
Amend the fill_slots test to serve as a non-regression test for this
issue: without this bug fix, it would cause a memory leak.
In psa_generator_import_key, if generating a DES or 3DES key, set the
parity bits.
Add tests for deriving a DES key. Also test deriving an AES key while
I'm at it.
In psa_generator_hkdf_read, return BAD_STATE if we're trying to
construct more output than the algorithm allows. This can't happen
through the API due to the capacity limit, but it could potentially
happen in an internal call.
Also add a test case that verifies that we can set up HKDF with its
maximum capacity and read up to the maximum capacity.
New key type PSA_KEY_TYPE_DERIVE. New usage flag PSA_KEY_USAGE_DERIVE.
New function psa_key_derivation.
No key derivation algorithm is implemented yet. The code may not
compile with -Wunused.
Write some unit test code for psa_key_derivation. Most of it cannot be
used yet due to the lack of a key derivation algorithm.
In asymmetric_encrypt_decrypt, use the buffer size advertized by the
library for the ciphertext, and the length of the plaintext for the
re-decrypted output.
Test the output length if known. Require it to be 0 on error for
encrypt/decrypt functions. If the output length is unknown, test at
least that it's within the buffer limits.
Add a label argument to all asymmetric encryption test functions
(currently empty in all tests, but that will change soon).
In asymmetric_encrypt and asymmetric_decrypt, with an empty label,
test with both a null pointer and a non-null pointer.
Although RSASSA-PSS defines its input as a message to be hashed, we
implement a sign-the-hash function. This function can take an input
which isn't a hash, so don't restrict the size of the input, any more
than Mbed TLS does.
Remove a redundant check that hash_length fits in unsigned int for the
sake of Mbed TLS RSA functions.
Test that PSS accepts inputs of various lengths. For PKCS#1 v1.5
signature in raw mode, test the maximum input length.
This required tweaking exercise_signature_key to use a payload size
for the signature based on the algorithm, since our implementation of
PSS requires that the input size matches the hash size. This would
also be the case for PKCS#1 v1.5 with a specified hash.
* No test depends on MBEDTLS_PK_C except via MBEDTLS_PK_PARSE_C, so
remove MBEDTLS_PK_C and keep only MBEDTLS_PK_PARSE_C.
* Add MBEDTLS_PK_WRITE_C for pk export tests.
* Add MBEDTLS_GENPRIME for RSA key generation tests.
* Add dependencies to AEAD tests.
* Add missing dependencies to many RSA tests.
* Add a test for decryption with invalid padding.
* Add a test for encryption with input too large.
* In negative tests, pass input whose length matches the key length,
unless that's what the test is about.
Make function names for multipart operations more consistent (cipher
edition).
Rename symmetric cipher multipart operation functions so that they all
start with psa_cipher_:
* psa_encrypt_setup -> psa_cipher_encrypt_setup
* psa_decrypt_setup -> psa_cipher_decrypt_setup
* psa_encrypt_set_iv -> psa_cipher_set_iv
* psa_encrypt_generate_iv -> psa_cipher_generate_iv
Change most asymmetric_verify to use public keys (they were all using
key pairs before). Keep one test with an RSA key pair and one with an
EC key pair.
Revise the test function asymmetric_encrypt_fail into
asymmetric_encrypt and use it for positive tests as well. Get the
expected output length from PSA_ASYMMETRIC_ENCRYPT_OUTPUT_SIZE. Check
the actual output length against test data.
Add positive test cases for encryption: one with an RSA public
key (this is the only test for encryption with a public key rather
than a key pair) and one with a key pair.
Make function names for multipart operations more consistent (MAC
setup edition).
Split psa_mac_setup into two functions psa_mac_sign_setup and
psa_mac_verify_setup. These functions behave identically except that
they require different usage flags on the key. The goal of the split
is to enforce the key policy during setup rather than at the end of
the operation (which was a bit of a hack).
In psa_mac_sign_finish and psa_mac_verify_finish, if the operation is
of the wrong type, abort the operation before returning BAD_STATE.
No common signature algorithm uses a salt (RSA-PKCS#1v1.5, RSA-PSS,
DSA, ECDSA, EdDSA). We don't even take an IV for MAC whereas MAC
algorithms with IV are uncommon but heard of. So remove the salt
parameter from psa_asymmetric_sign and psa_asymmetric_verify.
Add tests of key policy checks for MAC, cipher, AEAD, asymmetric
encryption and asymmetric signature. For each category, test
with/without the requisite usage flag in each direction, and test
algorithm mismatch.
At this point it fixes memory leaks as well. These memory leaks are the
fault of the 'psa_cipher_finish()' function and the calls fixed in this
commit (among with many others in the test suite) will become obsolete
after fixing 'psa_cipher_finish()'.
Add required includes in tests and psa_crypto.c file in order to be able to compilef for the SPM solution.
Some functions needed to be deprecated from psa_crypto.c since they already implemented in the SPM.
Use PSA_BLOCK_CIPHER_BLOCK_SIZE() macro to get the cipher block size instead of accessing the operation struct
additionally, for SPM case, the 'block_size' member is not a member in the operation struct
In tests that had a hard-coded buffer size, use PSA_MAC_MAX_SIZE or
PSA_ASYMMETRIC_SIGNATURE_MAX_SIZE as appropriate.
Test that PSA_xxx_MAX_SIZE is larger than the size used in tests that
expect a specific output.
Change the representation of an ECDSA signature from the ASN.1 DER
encoding used in TLS and X.509, to the concatenation of r and s
in big-endian order with a fixed size. A fixed size helps memory and
buffer management and this representation is generally easier to use
for anything that doesn't require the ASN.1 representation. This is
the same representation as PKCS#11 (Cryptoki) except that PKCS#11
allows r and s to be truncated (both to the same length), which
complicates the implementation and negates the advantage of a
fixed-size representation.
* Distinguish randomized ECDSA from deterministic ECDSA.
* Deterministic ECDSA needs to be parametrized by a hash.
* Randomized ECDSA only uses the hash for the initial hash step,
but add ECDSA(hash) algorithms anyway so that all the signature
algorithms encode the initial hashing step.
* Add brief documentation for the ECDSA signature mechanisms.
* Also define DSA signature mechanisms while I'm at it. There were
already key types for DSA.
* PSS needs to be parametrized by a hash.
* Don't use `_MGF1` in the names of macros for OAEP and PSS. No one
ever uses anything else.
* Add brief documentation for the RSA signature mechanisms.
Add a negative test for import where the expected key is an EC key
with the correct key size, but the wrong curve. Change the test that
tries to import an RSA key when an EC key is expected to have the
expected key size.
Because exporting-public a symmetric key fails, we have no reasonable
expectation that the exported key length has any value at all other than
something obviously incorrect or "empty", like a key with a length of 0.
Our current implementation explicitly sets the exported key length to 0
on errors, so test for this. Fix the "PSA import/export-public: cannot
export-public a symmetric key" test to expect a key length of 0 instead
of 162.
Make psa_export_key() always set a valid data_length when exporting,
even when there are errors. This makes the API easier to use for buggy
programs (like our test code).
Our test code previously used exported_length uninitialized when
checking to see that the buffer returned was all zero in import_export()
in the case where an error was returned from psa_export_key().
Initialize exported_length to an invalid length, and check that it gets
set properly by psa_export_key(), to avoid this using export_length
uninitialized. Note that the mem_is_zero() check is still valid when
psa_export_key() returns an error, e.g. where exported_length is 0, as
we want to check that nothing was written to the buffer on error.
Out test code also previous passed NULL for the data_length parameter of
psa_export_key() when it expected a failure (in key_policy_fail()).
However, data_length is not allowed to be NULL, especially now that we
write to data_length from psa_export_key() even when there are errors.
Update the test code to not pass in a NULL data_length.
It isn't used to define other macros and it doesn't seem that useful
for users. Remove it, we can reintroduce it if needed.
Define a similar function key_type_is_raw_bytes in the implementation
with a clear semantics: it's a key that's represented as a struct
raw_data.
In the test generate_random, focus on testing that psa_generate_random
is writing all the bytes of the output buffer and no more. Add a check
that it is writing to each byte of the output buffer. Do not try to
look for repeating output as the structure of a unit test isn't likely
to catch that sort of problem anyway.
Also add what was missing in the test suite to support block ciphers
with a block size that isn't 16.
Fix some buggy test data that passed only due to problems with DES
support in the product.
In psa_hash_start, psa_mac_start and psa_cipher_setup, return
PSA_ERROR_INVALID_ARGUMENT rather than PSA_ERROR_NOT_SUPPORTED when
the algorithm parameter is not the right category.
When psa_mac_start(), psa_encrypt_setup() or psa_cipher_setup()
failed, depending on when the failure happened, it was possible that
psa_mac_abort() or psa_cipher_abort() would crash because it would try
to call a free() function uninitialized data in the operation
structure. Refactor the functions so that they initialize the
operation structure before doing anything else.
Add non-regression tests and a few more positive and negative unit
tests for psa_mac_start() and psa_cipher_setup() (the latter via
psa_encrypt_setip()).
In some configurations (like config-mini-tls1_1.h), size is unused. This
leads to failures when building with CMake Asan, because that build
doesn't use "-Wno-unused-value".
Fixes: e01822299624 ("New function mbedtls_rsa_get_bitlen")
In psa_export_key, ensure that each byte of the output buffer either
contains its original value, is zero, or is part of the actual output.
Specifically, don't risk having partial output on error, and don't
leave extra data at the end of the buffer when exporting an asymmetric
key.
Test that exporting to a previously zeroed buffer leaves the buffer
zeroed outside the actual output if any.
Exporting an asymmetric key only worked if the target buffer had
exactly the right size, because psa_export_key uses
mbedtls_pk_write_key_der or mbedtls_pk_write_pubkey_der and these
functions write to the end of the buffer, which psa_export_key did not
correct for. Fix this by moving the data to the beginning of the
buffer if necessary.
Add non-regression tests.
Switch the default config.h back to the upstream version, plus the new
feature from this branch MBEDTLS_PSA_CRYPTO_C, plus MBEDTLS_CMAC_C
because it's a features we're using to explore the API design but
that's off by default in Mbed TLS.
Having a crypto-only version saved a bit of developer time, and it's
something we want to ship, but we also need a full build with TLS to
work, and the CI scripts assume that the default build includes TLS.
As a consequence, list-macros.sh no longer needs a special case to
pass check-names.sh.
* init-deinit
* import-export
* policies
* lifetime
* hash
* MAC
* cipher
* AEAD
* asymmetric sign
* asymmetric verify
* asymmetric encrypt-decrypt
This commit only moves test functions and test cases around. It does
not modify, add or remove tests.
psa_import_key must check that the imported key data matches the
expected key type. Implement the missing check for EC keys that the
curve is the expected one.
Instead of rolling our own list of elliptic curve identifiers, use one
from somewhere. Pick TLS because it's the right size (16 bits) and
it's as good as any.
Get rid of many redundant casts. In particular, it is not useful to
cast uint32_t values to size_t before performing arithmetic or
comparisons on them.
Rewrap a number of function calls, many of which now have narrower
arguments thanks to the removed casts. When a function call doesn't
fit on a single line, avoid grouping unrelated parameters together,
but do try to group a buffer pointer and the associated size.
Define more auxiliary variables xxx of a particular integer
type (psa_algorithm_t, psa_key_usage_t, etc.) corresponding to a test
function xxx_arg which has the type int. This avoids the need to cast
xxx_arg to an unsigned type sometimes in the code.
Avoid lines longer than 80 columns.
Remove some redundant parentheses, e.g. change
if( ( a == b ) && ( c == d ) )
to
if( a == b && c == d )
which makes lines less long and makes the remaining parentheses more
relevant.
Add missing parentheses around return statements.
There should be no semantic change in this commit.
* development-restricted: (578 commits)
Update library version number to 2.13.1
Don't define _POSIX_C_SOURCE in header file
Don't declare and define gmtime()-mutex on Windows platforms
Correct preprocessor guards determining use of gmtime()
Correct documentation of mbedtls_platform_gmtime_r()
Correct typo in documentation of mbedtls_platform_gmtime_r()
Correct POSIX version check to determine presence of gmtime_r()
Improve documentation of mbedtls_platform_gmtime_r()
platform_utils.{c/h} -> platform_util.{c/h}
Don't include platform_time.h if !MBEDTLS_HAVE_TIME
Improve wording of documentation of MBEDTLS_PLATFORM_GMTIME_R_ALT
Fix typo in documentation of MBEDTLS_PLATFORM_GMTIME_R_ALT
Replace 'thread safe' by 'thread-safe' in the documentation
Improve documentation of MBEDTLS_HAVE_TIME_DATE
ChangeLog: Add missing renamings gmtime -> gmtime_r
Improve documentation of MBEDTLS_HAVE_TIME_DATE
Minor documentation improvements
Style: Add missing period in documentation in threading.h
Rename mbedtls_platform_gmtime() to mbedtls_platform_gmtime_r()
Guard decl and use of gmtime mutex by HAVE_TIME_DATE and !GMTIME_ALT
...
Always adding things at the end tends to create merge conflicts.
Adding in the middle in this way makes the order more logical in
addition to avoiding conflicts.
* PKCS#7 padding always adds at least one byte of padding, so test
data with plaintext length = ciphertext length could not have been
correct.
* CTR has plaintext length = ciphertext length, so test data with
differing lengths could not have been correct.
In cipher_test_verify_output_multpart, tweak the ways chunk sizes are
added in order to get rid of the variable temp. In other functions,
this commit does not change the logic at all.
Only whitespace changes.
* Remove tabs.
* Remove trailing whitespace.
* Correct some misindented lines.
* Normalize whitespace around some punctuation.
* Split some lines to avoid going over 80 columns.
cipher_test_positive was never compiled due to a syntax error in the
BEGIN_CASE magic comment. It has now been duplicated as
cipher_test_encrypt. Remove the copy that was never compiled.
Conflicts:
library/psa_crypto.c
tests/suites/test_suite_psa_crypto.data
tests/suites/test_suite_psa_crypto.function
All the conflicts are concurrent additions where the order doesn't
matter. I put the code from feature-psa (key policy) before the code
from PR #13 (key lifetime).
Conflict resolution:
* `tests/suites/test_suite_psa_crypto.data`: in the new tests from PR #14,
rename `PSA_ALG_RSA_PKCS1V15_RAW` to `PSA_ALG_RSA_PKCS1V15_SIGN_RAW` as
was done in PR #15 in the other branch.
New header file crypto_struct.h. The main file crypto.sh declares
structures which are implementation-defined. These structures must be
defined in crypto_struct.h, which is included at the end so that the
structures can use types defined in crypto.h.
Implement psa_hash_start, psa_hash_update and psa_hash_final. This
should work for all hash algorithms supported by Mbed TLS, but has
only been smoke-tested for SHA-256, and only in the nominal case.
Don't use the pk module except as required for pkparse/pkwrite. The
PSA crypto layer is meant to work alongside pk, not on top of it.
Fix the compile-time dependencies on RSA/ECP handling in
psa_export_key, psa_destroy_key and psa_get_key_information.
Define psa_key_type_t and a first stab at a few values.
New functions psa_import_key, psa_export_key, psa_destroy_key,
psa_get_key_information. Implement them for raw data and RSA.
Under the hood, create an in-memory, fixed-size keystore with room
for MBEDTLS_PSA_KEY_SLOT_COUNT - 1 keys.
Add a new function mbedtls_rsa_get_bitlen which returns the RSA key
size, i.e. the bit size of the modulus. In the pk module, call
mbedtls_rsa_get_bitlen instead of mbedtls_rsa_get_len, which gave the
wrong result for key sizes that are not a multiple of 8.
This commit adds one non-regression test in the pk suite. More tests
are needed for RSA key sizes that are a multiple of 8.
This commit does not address RSA alternative implementations, which
only provide an interface that return the modulus size in bytes.
New module psa_crypto.c (MBEDTLS_PSA_CRYPTO_C):
Platform Security Architecture compatibility layer on top of
libmedcrypto.
Implement psa_crypto_init function which sets up a RNG.
Add a mbedtls_psa_crypto_free function which deinitializes the
library.
Define a first batch of error codes.
While these tests and the issue with it are pre-existing:
- we previously didn't understand that the issue was an openssl bug
- failures seem to have become more frequent since the recent changes
So let's disable these fragile tests in order to get a clean CI. We still have
the tests against gnutls-serv for interop testing.
While making the initial commit, I thought $OPENSSL_LEGACY was not affect by
this bug, but it turns out I was wrong. All versions of OpenSSL installed on
the CI are. Therefore, the corresponding tests are disabled for the same
reason as the gnutls-cli tests above it.
This commit is only about the tests that were added in the recent
fragmentation work. One of those two tests had a particularly
annoying mode of failure: it failed consistently with seed=1 (use in the
release version of all.sh), once #1951 was applied. This has nothing
particular to do with #1951, except that by changing retransmission behaviour
1951 made the proxy run into a path that triggered the OpenSSL bug with this
seed, while it previously did that only with other seeds.
Other 3d interop test are also susceptible to triggering this OpenSSL bug or
others (or bugs in GnuTLS), but they are left untouched by this commit as:
- they were pre-existing to the recent DTLS branches;
- they don't seem to have the particularly annoying seed=1 mode of failure.
However it's probably desirable to do something about them at some point in
the future.
previously a single function was used for most test cases (ctr_drbg_validate) making it harder to understand what the exact scenario is as a result it was split into easier to understand functions.
the testing functions were re-factored so that the common code was extracted to a single static function (removing the need for unclear goto statements).
As part of the re-factor the test functions now use data_t for parameters (support for this was introduced in previous rebase),
the change is designed to make configuring 128bit keys for ctr_drbg more similar to other configuration options. Tests have been updated accordingly.
also clarified test naming.
This commit adds a test to ssl-opt.sh which exercises the behavior
of the library in the situation where a single proper fragment
of a future handshake message is received prior to the next
expected handshake message (concretely, the client receives
the first fragment of the server's Certificate message prior
to the server's ServerHello).
This commit adds two builds to all.sh which use a value of
MBEDTLS_SSL_DTLS_MAX_BUFFERING that allows to run the
reordering tests in ssl-opt.sh introduced in the last commit.
This commit adds tests to ssl-opt.sh which trigger code-paths
responsible for freeing future buffered messages when the buffering
limitations set by MBEDTLS_SSL_DTLS_MAX_BUFFERING don't allow the
next expected message to be reassembled.
These tests only work for very specific ranges of
MBEDTLS_SSL_DTLS_MAX_BUFFERING and will therefore be skipped
on a run of ssl-opt.sh in ordinary configurations.
This commit adds functions requires_config_value_at_most()
and requires_config_value_at_least() which can be used to
only run tests when a numerical value from config.h
(e.g. MBEDTLS_SSL_IN_CONTENT_LEN) is within a certain range.
The negotiated MFL is always the one suggested by the client, even
if the server has a smaller MFL configured locally. Hence, in the test
where the client asks for an MFL of 4096 bytes while the server locally
has an MFL of 512 bytes configured, the client will still send datagrams
of up to ~4K size.
Depending on the settings of the local machine, gnutls-cli will either try
IPv4 or IPv6 when trying to connect to localhost. With TLS, whatever it tries
first, it will notice if any failure happens and try the other protocol if
necessary. With DTLS it can't do that. Unfortunately for now there isn't
really any good way to specify an address and hostname independently, though
that might come soon: https://gitlab.com/gnutls/gnutls/issues/344
A work around is to specify an address directly and then use --insecure to
ignore certificate hostname mismatch; that is OK for tests that are completely
unrelated to certificate verification (such as the recent fragmenting tests)
but unacceptable for others.
For that reason, don't specify a default hostname for gnutls-cli, but instead
let each test choose between `--insecure 127.0.0.1` and `localhost` (or
`--insecure '::1'` if desired).
Alternatives include:
- having test certificates with 127.0.0.1 as the hostname, but having an IP as
the CN is unusual, and we would need to change our test certs;
- have our server open two sockets under the hood and listen on both IPv4 and
IPv6 (that's what gnutls-serv does, and IMO it's a good thing) but that
obviously requires development and testing (esp. for windows compatibility)
- wait for a newer version of GnuTLS to be released, install it on the CI and
developer machines, and use that in all tests - quite satisfying but can't
be done now (and puts stronger requirements on test environment).
From Hanno:
When a server replies to a cookieless ClientHello with a HelloVerifyRequest,
it is supposed to reset the connection and wait for a subsequent ClientHello
which includes the cookie from the HelloVerifyRequest.
In testing environments, it might happen that the reset of the server
takes longer than for the client to replying to the HelloVerifyRequest
with the ClientHello+Cookie. In this case, the ClientHello gets lost
and the client will need retransmit. This may happen even if the underlying
datagram transport is reliable.
This commit continues commit 47db877 by removing resend guards in the
ssl-opt.sh tests 'DTLS fragmenting: proxy MTU, XXX' which sometimes made
the tests fail in case the log showed a resend from the client.
See 47db877 for more information.
Unify the three existing validation functions (with prediction
resistance, with manual reseeding between generations, and with no
reseeding) into a single function that supports these three scenarios
plus a fourth one (reseed before the first generation).
The four supported scenarios cover the three scenarios from the
current CAVP test vectors (no reseed, reseed before generating,
prediction resistance) plus a fourth scenario used by the existing
test vectors (reseed after generating).
(cherry picked from commit cee9bedee6bc1a8e2b22fa8a31647b62ebb8a0a4)
The ctr_drbg_validate_xxx test functions had hard-coded sizes for the
entropy and the output size. Generalize the sizes.
Keep track of the current entropy size.
Unhexify the expected output and compare with the actual output,
rather than hexifying the actual output and comparing the hex.
When a server replies to a cookieless ClientHello with a HelloVerifyRequest,
it is supposed to reset the connection and wait for a subsequent ClientHello
which includes the cookie from the HelloVerifyRequest.
In testing environments, it might happen that the reset of the server
takes longer than for the client to replying to the HelloVerifyRequest
with the ClientHello+Cookie. In this case, the ClientHello gets lost
and the client will need retransmit. This may happen even if the underlying
datagram transport is reliable.
This commit removes a guard in the ssl-opt.sh test
'DTLS fragmenting: proxy MTU, resumed handshake' which made
the test fail in case the log showed a resend from the client.
We previously observed random-looking failures from this test. I think they
were caused by a race condition where the client tries to reconnect while the
server is still closing the connection and has not yet returned to an
accepting state. In that case, the server would fail to see and reply to the
ClientHello, and the client would have to resend it.
I believe logs of failing runs are compatible with this interpretation:
- the proxy logs show the new ClientHello and the server's closing Alert are
sent the same millisecond.
- the client logs show the server's closing Alert is received after the new
handshake has been started (discarding message from wrong epoch).
The attempted fix is for the client to wait a bit before reconnecting, which
should vastly enhance the probability of the server reaching its accepting
state before the client tries to reconnect. The value of 1 second is arbitrary
but should be more than enough even on loaded machines.
The test was run locally 100 times in a row on a slightly loaded machine (an
instance of all.sh running in parallel) without any failure after this fix.
Use the same values as other 3d tests: this makes the test hopefully a bit
faster than the default values, while not increasing the failure rate.
While at it:
- adjust "needs_more_time" setting for 3d interop tests (we can't set the
timeout values for other implementations, so the test might be slow)
- fix some supposedly DTLS 1.0 test that were using dtls1_2 on the command
line
Now that the UDP proxy has the ability to delay specific
handshake message on the client and server side, use
this to rewrite the reordering tests and thereby make
them independent on the choice of PRNG used by the proxy
(which is not stable across platforms).
