There are currently 4 tests in ssl-opt.sh with either -C "resend" or -S
"resend", that is, asserting that no retransmission will occur. They sometimes
fail on loaded CI machines as one side doesn't send a message fast enough,
causing the other side to retransmit, causing the test to fail.
(For the "reconnect" test there was an other issue causing random failures,
fixed in a previous commit, but even after that fix the test would still
sometimes randomly fail, even if much more rarely.)
While it's a hard problem to fix in a general and perfect way, in practice the
probability of failures can be drastically reduced by making the timeout
values much larger.
For some tests, where retransmissions are actually expected, this would have
the negative effect of increasing the average running time of the test, as
each side would wait for longer before it starts retransmission, so we have a
trade-off between average running time and probability of spurious failures.
But for tests where retransmission is not expected, there is no such trade-off
as the expected running time of the test (assuming the code is correct most of
the time) is not impacted by the timeout value. So the only negative effect of
increasing the timeout value is on the worst-case running time on the test,
which is much less important, as test should only fail quite rarely.
This commit addresses the easy case of tests that don't expect retransmission
by increasing the value of their timeout range to 10s-20s. This value
corresponds to the value used for tests that assert `-S "autoreduction"` which
are in the same case and where the current value seems acceptable so far.
It also represents an increase, compared to the values before this commit, of
a factor 20 for the "reconnect" tests which were frequently observed to fail
in the CI, and of a factor 10 for the first two "DTLS proxy" tests, which were
observed to fail much less frequently, so hopefully the new values are enough
to reduce the probability of spurious failures to an acceptable level.
Signed-off-by: Manuel Pégourié-Gonnard <manuel.pegourie-gonnard@arm.com>
The server must check client reachability (we chose to do that by checking a
cookie) before destroying the existing association (RFC 6347 section 4.2.8).
Let's make sure we do, by having a proxy-in-the-middle inject a ClientHello -
the server should notice, but not destroy the connection.
Signed-off-by: Manuel Pégourié-Gonnard <manuel.pegourie-gonnard@arm.com>
- "Default" should only be used for tests that actually use the defaults (ie,
not passing options on the command line, except maybe debug/dtls)
- All tests in the "Encrypt then MAC" group should start with that string as a
common prefix
Signed-off-by: Manuel Pégourié-Gonnard <manuel.pegourie-gonnard@arm.com>
This commit only addresses the timeouts in the "DTLS proxy: 3d, ..." tests.
The discrepancy with the 2.16 branch became apparent for some of these tests
when backporting the previous commit (skip_close_nofity), so let's align the
whole series for consistency and to make future backporting easier.
Signed-off-by: Manuel Pégourié-Gonnard <manuel.pegourie-gonnard@arm.com>
The ssl-opt.sh test cases using session resumption tend to fail occasionally
on the CI due to a race condition in how ssl_server2 and ssl_client2 handle
the reconnection cycle.
The server does the following in order:
- S1 send application data
- S2 send a close_notify alert
- S3 close the client socket
- S4 wait for a "new connection" (actually a new datagram)
- S5 start a handshake
The client does the following in order:
- C1 wait for and read application data from the server
- C2 send a close_notify alert
- C3 close the server socket
- C4 reset session data and re-open a server socket
- C5 start a handshake
If the client has been able to send the close_notify (C2) and if has been
delivered to the server before if closes the client socket (S3), when the
server reaches S4, the datagram that we start the new connection will be the
ClientHello and everything will be fine.
However if S3 wins the race and happens before the close_notify is delivered,
in S4 the close_notify is what will be seen as the first datagram in a new
connection, and then in S5 this will rightfully be rejected as not being a
valid ClientHello and the server will close the connection (and go wait for
another one). The client will then fail to read from the socket and exit
non-zero and the ssl-opt.sh harness will correctly report this as a failure.
In order to avoid this race condition in test using ssl_client2 and
ssl_server2, this commits introduces a new command-line option
skip_close_notify to ssl_client2 and uses it in all ssl-opt.sh tests that use
session resumption with DTLS and ssl_server2.
This works because ssl_server2 knows how many messages it expects in each
direction and in what order, and closes the connection after that rather than
relying on close_notify (which is also why there was a race in the first
place).
Tests that use another server (in practice there are two of them, using
OpenSSL as a server) wouldn't work with skip_close_notify, as the server won't
close the connection until the client sends a close_notify, but for the same
reason they don't need it (there is no race between receiving close_notify and
closing as the former is the cause of the later).
An alternative approach would be to make ssl_server2 keep the connection open
until it receives a close_notify. Unfortunately it creates problems for tests
where we simulate a lossy network, as the close_notify could be lost (and the
client can't retransmit it). We could modify udp_proxy with an option to never
drop alert messages, but when TLS 1.3 comes that would no longer work as the
type of messages will be encrypted.
Signed-off-by: Manuel Pégourié-Gonnard <manuel.pegourie-gonnard@arm.com>
In the 2.7 branch, test-ca.crt has all the components of its Subject name
encoded as PrintableString, because it's generated with our cert_write
program, and our code writes all components that way until Mbed TLS 2.14.
But the default RSA SHA-256 certificate, server2-sha256.crt, has the O and CN
components of its Issuer name encoded as UTF8String, because it was generated
with OpenSSL and that's what OpenSSL does, regardless of how those components
were encoded in the CA's Subject name.
This triggers some overly strict behaviour in some libraries, most notably NSS
and GnuTLS (of interest to us in ssl-opt.sh) which won't recognize the trusted
root as a possible parent for the presented certificate, see for example:
https://github.com/ARMmbed/mbedtls/issues/1033
Fortunately, we have at our disposal a version of test-ca.crt with encodings
matching the ones in server2-sha256.crt, in the file test-ca_utf8.crt. So
let's append that to gnutls-cli's list of trusted roots, so that it recognizes
certs signed by this CA but with the O and CN components as UTF8String.
Note: Since https://github.com/ARMmbed/mbedtls/pull/1641 was merged (in Mbed
TLS 2.14), we changed how we encode those components, so in the 2.16 branch,
cert_write generates test-ca.crt with encodings that matches the ones used by
openssl when generating server2-sha256.crt, so the issue of gnutls-cli
rejecting server2-sha256.crt is specific to the 2.7 branch.
The splitting of this test into two versions depending on whether SHA-1 was
allowed by the server was a mistake in
5d2511c4d4 - the test has nothing to do with
SHA-1 in the first place, as the server doesn't request a certificate from
the client so it doesn't matter if the server accepts SHA-1 or not.
While the whole script makes (often implicit) assumptions about the version of
GnuTLS used, generally speaking it should work out of the box with the version
packaged on our reference testing platform, which is Ubuntu 16.04 so far.
With the update from Jan 8 2020 (3.4.10-4ubuntu1.6), the patches for rejecting
SHA-1 in certificate signatures were backported, so we should avoid presenting
SHA-1 signed certificates to a GnuTLS peer in ssl-opt.sh.
Limit log output in compat.sh and ssl-opt.sh, in case of failures with
these scripts where they may output seemingly unlimited length error
logs.
Note that ulimit -f uses units of 512 bytes, so we use 10 * 1024 * 1024
* 2 to get 10 GiB.
* origin/pr/2435:
Use certificates from data_files and refer them
Specify server certificate to use in SHA-1 test
refactor CA and SRV certificates into separate blocks
refactor SHA-1 certificate defintions and assignment
refactor server SHA-1 certificate definition into a new block
define TEST_SRV_CRT_RSA_SOME in similar logic to TEST_CA_CRT_RSA_SOME
server SHA-256 certificate now follows the same logic as CA SHA-256 certificate
add entry to ChangeLog
The test used 3DES as the suite for SSLv3, which now makes the handshake fails
with "no ciphersuite in common", failing the test as well. Use Camellia
instead (as there are not enough AES ciphersuites before TLS 1.2 to
distinguish between the 3 versions).
Document some dependencies, but not all. Just trying to avoid introducing new
issues by using a new cipher here, not trying to make it perfect, which is a
much larger task out of scope of this commit.
DTLS records from previous epochs were incorrectly checked against the
current epoch transform's minimal content length, leading to the
rejection of entire datagrams. This commit fixed that and adapts two
test cases accordingly.
Internal reference: IOTSSL-1417
1. Update the test script to un the ECC tests only if the relevant
configurations are defined in `config.h` file
2. Change the HASH of the ciphersuite from SHA1 based to SHA256
for better example
If lsof is not available, wait_server_start uses a fixed timeout,
which can trigger a race condition if the timeout turns out to be too
short. Emit a warning so that we know this is going on from the test
logs.
- Some of the CI machines don't have lsof installed yet, so rely on an sleeping
an arbitrary number of seconds while the server starts. We're seeing
occasional failures with the current delay because the CI machines are highly
loaded, which seems to indicate the current delay is not quite enough, but
hopefully not to far either, so double it.
- While at it, also double the watchdog delay: while I don't remember seeing
much failures due to client timeout, this change doesn't impact normal
running time of the script, so better err on the safe side.
These changes don't affect the test and should only affect the false positive
rate coming from the test framework in those scripts.
In wait_server_start, fork less. When lsof is present, call it on the
expected process. This saves a few percent of execution time on a
lightly loaded machine. Also, sleep for a short duration rather than
using a tight loop.
