First, we introduce a flag to teach src/test/test to split its work
into chunks. Then we replace our invocation of src/test/test in our
"make check" target with a set of 8 scripts that invoke the first
8th of the tests, the second 8th, and so on.
This change makes our "make -kj4 check" target in our hardened
gitlab build more than twice as fast, since src/test/test was taking
the longest to finish.
Closes 40098.
The rend_cache/entry_free was missing the rend cache allocation increment
before freeing the object.
Without it, it had an underflow bug:
Sep 17 08:40:13.845 [warn] rend_cache_decrement_allocation(): Bug: Underflow
in rend_cache_decrement_allocation (on Tor 0.4.5.0-alpha-dev
7eef9ced61)
Fixes#40125
Signed-off-by: David Goulet <dgoulet@torproject.org>
tl;dr We were not counting cells flying from the client to the service, but we
were counting cells flying from the service to the client.
When a rendezvous cell arrives from the client to the RP, the RP forwards it to
the service.
For this to happen, the cell first passes through command_process_relay_cell()
which normally does the statistics counting. However because the `rend_circ`
circuit was not flagged with `circuit_carries_hs_traffic_stats` in
rend_mid_rendezvous(), the cell is not counted there.
Then the cell goes to circuit_receive_relay_cell() which has a special code
block based on `rend_splice` specifically for rendezvous cells, and the cell
gets directly passed to `rend_circ` via a direct call to
circuit_receive_relay_cell(). The cell never passes through
command_process_relay_cell() ever again and hence is never counted by our
rephist module.
The fix here is to flag the `rend_circ` circuit with
`circuit_carries_hs_traffic_stats` so that the cell is counted as soon as it
hits command_process_relay_cell().
Furthermore we avoid double-counting cells since the special code block of
circuit_receive_relay_cell() makes us count rendezvous cells only as they enter
the RP and not as they exit it.
Fixes#40117.
Opening a new listener connection can fail in many ways like a bind()
permission denied on a low port for instance.
And thus, we should expect to handle an error when creating a new one instead
of assert() on it.
To hit the removed assert:
ORPort 80
KeepBindCapabilities 0
Start tor. Then edit torrc:
ORPort <some-IP>:80
HUP tor and the assert is hit.
Fixes#40073
Signed-off-by: David Goulet <dgoulet@torproject.org>
This function once served to let circuits continue to be built over
version-1 link connections. But such connections are long-obsolete,
and it's time to remove this check.
Closes#40081.
We found this in #40076, after we started using buf_move_all() in
more places. Fixes bug #40076; bugfix on 0.3.3.1-alpha. As far as
I know, the crash only affects master, but I think this warrants a
backport, "just in case".
Without this fix, if an PT forgets to send a USERADDR command, that
results in a connection getting treated as local for the purposes of
rate-limiting.
If the PT _does_ use USERADDR, we still believe it.
Closes ticket 33747.
Previously we tolerated up to 1.5 connections for every relay we
were connected to, and didn't warn if we had fewer than 5
connections total.
Now we tolerate up to 1.5 connections per relay, and up to 4
connections per authority, and we don't warn at all when we have
fewer than 25 connections total.
Fixes bug 33880, which seems to have been provoked by our #17592
change in 0.3.5.
check-cocci is still a good idea -- perhaps as a cron job? But
doing it as part of our regular tests has just been confusing,
especially to volunteers who shouldn't have to become coccinelle
experts in order to get their patches through our CI.
Closes#40030.
For some reasons, Appveyor started to use the stdio printf format for 64 bit
values (PRIu64, ...). Mingw doesn't like that so force it to use the Windows
specific macros by setting D__USE_MINGW_ANSI_STDIO=0.
Fixes#40026
This patch fixes an out-of-bound memory read in
`tor_tls_cert_matches_key()` when Tor is compiled to use Mozilla's NSS
instead of OpenSSL.
The NSS library stores some length fields in bits instead of bytes, but
the comparison function found in `SECITEM_ItemsAreEqual()` needs the
length to be encoded in bytes. This means that for a 140-byte,
DER-encoded, SubjectPublicKeyInfo struct (with a 1024-bit RSA public key
in it), we would ask `SECITEM_ItemsAreEqual()` to compare the first 1120
bytes instead of 140 (140bytes * 8bits = 1120bits).
This patch fixes the issue by converting from bits to bytes before
calling `SECITEM_ItemsAreEqual()` and convert the `len`-fields back to
bits before we leave the function.
This patch is part of the fix for TROVE-2020-001.
See: https://bugs.torproject.org/33119
This is not actually a bug! It can happen for a bunch of reasons,
which all boil down to "trying to add an extrainfo for which we no
longer have the corresponding routerinfo".
Fixes#16016; bugfix on 0.2.6.3-alpha.
Resume being willing to use preemptively-built circuits when
UseEntryGuards is set to 0. We accidentally disabled this feature with
that config setting (in our fix for #24469), leading to slower load times.
Fixes bug 34303; bugfix on 0.3.3.2-alpha.
The client auth protocol allows attacker-controlled x25519 private keys being
passed around, which allows an attacker to potentially trigger the all-zeroes
assert for client_auth_sk in hs_descriptor.c:decrypt_descriptor_cookie().
We fixed that by making sure that an all-zeroes client auth key will not be
used.
There are no guidelines for validating x25519 private keys, and the assert was
there as a sanity check for code flow issues (we don't want to enter that
function with an unitialized key if client auth is being used). To avoid such
crashes in the future, we also changed the assert to a BUG-and-err.
This change broke torsocks that by default is expecting an IPv4 for hostname
resolution because it can't ask tor for a specific IP version with the SOCKS5
extension.
PreferIPv6 made it that sometimes the IPv6 could be returned to torsocks that
was expecting an IPv4.
Torsocks is probably a very unique case because the runtime flow is that it
hijacks DNS resolution (ex: getaddrinfo()), gets an IP and then sends it back
for the connect() to happen.
The libc has DNS resolution functions that allows the caller to request a
specific INET family but torsocks can't tell tor to resolve the hostname only
to an IPv4 or IPv6 and thus by default fallsback to IPv4.
Reverting this change into 0.4.3.x series but we'll keep it in the 0.4.4.x
series in the hope that we add this SOCKS5 extension to tor for DNS resolution
and then change torsocks to use that.
Fixes#33804
Signed-off-by: David Goulet <dgoulet@torproject.org>
Avoid conflicts between the fake sockets in tor's unit tests, and real
file descriptors. Resolves issues running unit tests with GitHub
Actions, where the process that embeds or launches the tests has
already opened a large number of file descriptors.
Fixes bug 33782; bugfix on 0.2.8.1-alpha.
Found and fixed by Putta Khunchalee.
When IPv6 ORPorts are set to "auto", tor relays and bridges would
advertise an incorrect port in their descriptor.
This may be a low-severity memory safety issue, because the published
port number may be derived from uninitialised or out-of-bounds memory
reads.
Fixes bug 32588; bugfix on 0.2.3.9-alpha.
