We really should ignore any timeouts that have *no* network activity for their
entire measured lifetime, now that we have the 95th percentile measurement
changes. Usually this is up to a minute, even on fast connections.
If we really want all this complexity for these stages here, we need to handle
it better for people with large timeouts. It should probably go away, though.
Rechecking the timeout condition was foolish, because it is checked on the
same codepath. It was also wrong, because we didn't round.
Also, the liveness check itself should be <, and not <=, because we only have
1 second resolution.
Specifically, a circ attempt that we'd launched while the network was
down could timeout after we've marked our entrynodes up, marking them
back down again. The fix is to annotate as bad the OR conns that were
around before we did the retry, so if a circuit that's attached to them
times out we don't do anything about it.
We frequently add cells to stream-blocked queues for valid reasons
that don't mean we need to block streams. The most obvious reason
is if the cell arrives over a circuit rather than from an edge: we
don't block circuits, no matter how full queues get. The next most
obvious reason is that we allow CONNECTED cells from a newly created
stream to get delivered just fine.
This patch changes the behavior so that we only iterate over the
streams on a circuit when the cell in question came from a stream,
and we only block the stream that generated the cell, so that other
streams can still get their CONNECTEDs in.
Right now it says "552 internal error" because there's no way for
getinfo_helper_*() countries to specify an error message. This
patch changes the getinfo_helper_*() interface, and makes most of the
getinfo helpers give useful error messages in response to failures.
This should prevent recurrences of bug 1699, where a missing GeoIPFile
line in the torrc made GETINFO ip-to-county/* fail in a "not obvious
how to fix" way.
We now record large times as abandoned, to prevent a filter step from
happening and skewing our results.
Also, issue a warn for a rare case that can happen for funky values of Xm or
too many abandoned circuits. Can happen (very rarely) during unit tests, but
should not be possble during live operation, due to network liveness filters
and discard logic.
Many friendly operating systems have 64-bit times, and it's not nice
to pass them to an %ld format.
It's also extremely not-nice to write a time to the log as an
integer. Most people think it's 2010 June 29 23:57 UTC+epsilon, not
1277855805+epsilon.
We need to record different statistics at point of timeout, vs the point
of forcible closing.
Also, give some better names to constants and state file variables
to indicate they are not dealing with timeouts, but abandoned circuits.
There are now four ways that CBT can be disabled:
1. Network-wide, with the cbtdisabled consensus param.
2. Via config, with "LearnCircuitBuildTimeout 0"
3. Via config, with "AuthoritativeDirectory 1"
4. Via a state file write failure.
This should prevent some asserts and storage of incorrect build times
for the cases where Tor is suspended during a circuit construction, or
just after completing a circuit. The idea is that if the circuit
build time is much greater than we would have cut it off at, we probably
had a suspend event along this codepath, and we should discard the
value.
In case we decide that the timeout rate is now too high due to our
change of the max synthetic quantile value, this consensus parameter
will allow us to restore it to the previous value.
This is for the other issue we saw in Bug 1335. A large number of high
timeouts were causing the timeout calculation to slowly drift upwards,
especially in conditions of load. This fix repeatedly regenerates all of
our synthetic timeouts whenever the timeout changes, to try to prevent
drift.
It also lowers the timeout cap to help for some cases of Bug 1245, where
some timeout values were so large that we ended up allocating a ton of
scratch memory to count the histogram bins.
The downside is that lowering this cap is affecting our timeout rate.
Unfortunately, the buildtimeout quantile is now higher than the actual
completion rate by what appears to be about 7-10%, which probably
represents the skew in the distribution due to lowering this synthetic
cap.
In my state files, I was seeing several peaks, probably due to different
guards having different latency. This change is meant to better capture
this behavior and generate more reasonable timeouts when it happens. It
is improving the timeout values for my collection of state files.
Another dereference-then-NULL-check sequence. No reports of this bug
triggered in the wild. Fixes bugreport 1256.
Thanks to ekir for discovering and reporting this bug.
asprintf() is a GNU extension that some BSDs have picked up: it does a printf
into a newly allocated chunk of RAM.
Our tor_asprintf() differs from standard asprintf() in that:
- Like our other malloc functions, it asserts on OOM.
- It works on windows.
- It always sets its return-field.
Specifically, there are two cases: a) are we willing to start a new
circuit at a node not in your ExitNodes config option, and b) are we
willing to make use of a circuit that's already established but has an
unsuitable exit.
Now we discard all your circuits when you set ExitNodes, so the only
way you could end up with an exit circuit that ends at an unsuitable
place is if we explicitly ran out of exit nodes, StrictNodes was 0,
and we built this circuit to solve a stream that needs solving.
Fixes bug in dc322931, which would ignore the just-built circuit because
it has an unsuitable exit.
Before it would prepend your requested entrynodes to your list of guard
nodes, but feel free to use others after that. Now it chooses only
from your EntryNodes if any of those are available, and only falls back
to others if a) they're all down and b) StrictNodes is not set.
Also, now we refresh your entry guards from EntryNode at each consensus
fetch (rather than just at startup and then they slowly rot as the
network changes).
The goal here is to make users less likely to set StrictNodes, since
it's doing closer to what they expect it should be doing.
The new rule is: safe_str_X() means "this string is a piece of X
information; make it safe to log." safe_str() on its own means
"this string is a piece of who-knows-what; make it safe to log".
Some *_free functions threw asserts when passed NULL. Now all of them
accept NULL as input and perform no action when called that way.
This gains us consistence for our free functions, and allows some
code simplifications where an explicit null check is no longer necessary.
