When a directory request fails, we flag the relay as non Running so we
don't use it anymore.
This can be problematic with onion services because there are cases
where a tor instance could have a lot of services, ephemeral ones, and
keeps failing to upload descriptors, let say due to a bad network, and
thus flag a lot of nodes as non Running which then in turn can not be
used for circuit building.
This commit makes it that we never flag nodes as non Running on a onion
service directory request (upload or fetch) failure as to keep the
hashring intact and not affect other parts of tor.
Fortunately, the onion service hashring is _not_ selected by looking at
the Running flag but since we do a 3-hop circuit to the HSDir, other
services on the same instance can influence each other by removing nodes
from the consensus for path selection.
This was made apparent with a small network that ran out of nodes to
used due to rapid succession of onion services uploading and failing.
See #40434 for details.
Fixes#40434
Signed-off-by: David Goulet <dgoulet@torproject.org>
Fixes bug 40078.
As reported by hdevalence our batch verification logic can cause an assert
crash.
The assert happens because when the batch verification of ed25519-donna fails,
the code in `ed25519_checksig_batch()` falls back to doing a single
verification for each signature.
The crash occurs because batch verification failed, but then all signatures
individually verified just fine.
That's because batch verification and single verification use a different
equation which means that there are sigs that can pass single verification
but fail batch verification.
Fixing this would require modding ed25519-donna which is not in scope for
this ticket, and will be soon deprecated in favor of arti and
ed25519-dalek, so my branch instead removes batch verification.
Continue having a tor_gmtime_impl() unit test so that we can detect
any problems in our replacement function; add a new test function to
make sure that gmtime<->timegm are a round-trip on now-ish times.
This is a fix for bug #40383, wherein we ran into trouble because
tor_timegm() does not believe that time_t should include a count of
leap seconds, but FreeBSD's gmtime believes that it should. This
disagreement meant that for a certain amount of time each day,
instead of calculating the most recent midnight, our voting-schedule
functions would calculate the second-most-recent midnight, and lead
to an assertion failure.
I am calling this a bugfix on 0.2.0.3-alpha when we first started
calculating our voting schedule in this way.
This patch enables the deterministic RNG for address set tests,
including the tests which uses address set indirectly via the nodelist
API.
This should prevent random test failures in the highly unlikely case of
a false positive which was seen in tor#40419.
See: tpo/core/tor#40419.
This issue was reported by Jann Horn part of Google's Project Zero.
Jann's one-sentence summary: entry/middle relays can spoof RELAY_END cells on
half-closed streams, which can lead to stream confusion between OP and
exit.
Fixes#40389
Previously, we would detect errors from a missing RNG
implementation, but not failures from the RNG code itself.
Fortunately, it appears those failures do not happen in practice
when Tor is using OpenSSL's default RNG implementation. Fixes bug
40390; bugfix on 0.2.8.1-alpha. This issue is also tracked as
TROVE-2021-004. Reported by Jann Horn at Google's Project Zero.
Cached_dir_t is a somewhat "legacy" kind of storage when used for
consensus documents, and it appears that there are cases when
changing our settings causes us to stop updating those entries.
This can cause trouble, as @arma found out in #40375, where he
changed his settings around, and consensus diff application got
messed up: consensus diffs were being _requested_ based on the
latest consensus, but were being (incorrectly) applied to a
consensus that was no longer the latest one.
This patch is a minimal fix for backporting purposes: it has Tor do
the same search when applying consensus diffs as we use to request
them. This should be sufficient for correct behavior.
There's a similar case in GETINFO handling; I've fixed that too.
Fixes#40375; bugfix on 0.3.1.1-alpha.
When seccomp sandbox is active, SAVECONF failed because it was not
able to save the backup files for torrc. This commit simplifies
the implementation of SAVECONF and sandbox by making it keep only
one backup of the configuration file.
The connection type for the listener part was missing from the "is
connection a listener" function.
This lead to our periodic event that retries our listeners to keep
trying to bind() again on an already opened MetricsPort.
Closes#40370
Signed-off-by: David Goulet <dgoulet@torproject.org>
This change permits the newfstatat() system call, and fixes issues
40382 (and 40381).
This isn't a free change. From the commit:
// Libc 2.33 uses this syscall to implement both fstat() and stat().
//
// The trouble is that to implement fstat(fd, &st), it calls:
// newfstatat(fs, "", &st, AT_EMPTY_PATH)
// We can't detect this usage in particular, because "" is a pointer
// we don't control. And we can't just look for AT_EMPTY_PATH, since
// AT_EMPTY_PATH only has effect when the path string is empty.
//
// So our only solution seems to be allowing all fstatat calls, which
// means that an attacker can stat() anything on the filesystem. That's
// not a great solution, but I can't find a better one.
