The rule is now: take the value from the CircuitPriorityHalflife
config option if it is set. If it zero, disable the cell_ewma
algorithm. If it is set, use it to calculate the scaling factor.
If it is not set, look for a CircPriorityHalflifeMsec parameter in the
consensus networkstatus. If *that* is zero, then disable the cell_ewma
algorithm; if it is set, use it to calculate the scaling factor.
If it is not set at all, disable the algorithm.
In connection_dir_client_reached_eof, we make sure that we either
return when we get an http status code of 503 or handle the problem
and set it to 200. Later we check if the status code is 503. Remove
that check.
There are two big changes here:
- We store active circuits in a priority queue for each or_conn,
rather than doing a linear search over all the active circuits
before we send each cell.
- Rather than multiplying every circuit's cell-ewma by a decay
factor every time we send a cell (thus normalizing the value of a
current cell to 1.0 and a past cell to alpha^t), we instead
only scale down the cell-ewma every tick (ten seconds atm),
normalizing so that a cell sent at the start of the tick has
value 1.0).
Each circuit is ranked in terms of how many cells from it have been
relayed recently, using a time-weighted average.
This patch has been tested this on a private Tor network on PlanetLab,
and gotten improvements of 12-35% in time it takes to fetch a small
web page while there's a simultaneous large data transfer going on
simultaneously.
[Commit msg by nickm based on mail from Ian Goldberg.]
This changes the pqueue API by requiring an additional int in every
structure that we store in a pqueue to hold the index of that structure
within the heap.
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.
It turns out that OpenSSL 0.9.8m is likely to take a completely
different approach for reenabling renegotiation than OpenSSL 0.9.8l
did, so we need to work with both. :p Fixes bug 1158.
(patch by coderman; commit message by nickm)
Do not segfault when writing buffer stats when we haven't observed a
single circuit to report about. This is a minor bug that would only show
up in testing environments with no traffic and with reduced stats
intervals.
Avoid crashing if the client is trying to upload many bytes and the
circuit gets torn down at the same time, or if the flip side
happens on the exit relay. Bugfix on 0.2.0.1-alpha; fixes bug 1150.
New config option "CircuitStreamTimeout" to override our internal
timeout schedule for how many seconds until we detach a stream from
a circuit and try a new circuit. If your network is particularly
slow, you might want to set this to a number like 60.
On this OSX version, there is a stub mlockall() function
that doesn't work, *and* the declaration for it is hidden by
an '#ifdef _P1003_1B_VISIBLE'. This would make autoconf
successfully find the function, but our code fail to build
when no declaration was found.
This patch adds an additional test for the declaration.
This fixes bug 1147:
bionic doesn't have an actual implementation of mlockall();
mlockall() is merely in the headers but not actually in the library.
This prevents Tor compilation with the bionic libc for Android handsets.
To fix a major security problem related to incorrect use of
SSL/TLS renegotiation, OpenSSL has turned off renegotiation by
default. We are not affected by this security problem, however,
since we do renegotiation right. (Specifically, we never treat a
renegotiated credential as authenticating previous communication.)
Nevertheless, OpenSSL's new behavior requires us to explicitly
turn renegotiation back on in order to get our protocol working
again.
Amusingly, this is not so simple as "set the flag when you create
the SSL object" , since calling connect or accept seems to clear
the flags.
For belt-and-suspenders purposes, we clear the flag once the Tor
handshake is done. There's no way to exploit a second handshake
either, but we might as well not allow it.
This commit implements a new config option: 'DisableAllSwap'
This option probably only works properly when Tor is started as root.
We added two new functions: tor_mlockall() and tor_set_max_memlock().
tor_mlockall() attempts to mlock() all current and all future memory pages.
For tor_mlockall() to work properly we set the process rlimits for memory to
RLIM_INFINITY (and beyond) inside of tor_set_max_memlock().
We behave differently from mlockall() by only allowing tor_mlockall() to be
called one single time. All other calls will result in a return code of 1.
It is not possible to change DisableAllSwap while running.
A sample configuration item was added to the torrc.complete.in config file.
A new item in the man page for DisableAllSwap was added.
Thanks to Moxie Marlinspike and Chris Palmer for their feedback on this patch.
Please note that we make no guarantees about the quality of your OS and its
mlock/mlockall implementation. It is possible that this will do nothing at all.
It is also possible that you can ulimit the mlock properties of a given user
such that root is not required. This has not been extensively tested and is
unsupported. I have included some comments for possible ways we can handle
this on win32.
If your relay can't keep up with the number of incoming create cells, it
would log one warning per failure into your logs. Limit warnings to 1 per
minute.
In 5e4d53d535 we made it so that
crypto_cipher_set_key cannot fail. The call will now
always succeed, to returning a boolean for success/failure makes
no sense.
This was left over from an early draft of the microdescriptor code; it
began to populate the signatures array of a networkstatus vote, even
though there's no actual need to do that for a vote.
In its zeal to keep me from saying memset(x, '0', sizeof(x)), Coverity
disallows memset(x, 48, sizeof(x)). Fine. I'll choose a different
magic number, see if I care!
In C, the code "char x[10]; if (x) {...}" always takes the true branch of
the if statement. Coverity notices this now.
In some cases, we were testing arrays to make sure that an operation
we wanted to do would suceed. Those cases are now always-true.
In some cases, we were testing arrays to see if something was _set_.
Those caes are now tests for strlen(s), or tests for
!tor_mem_is_zero(d,len).
If all authorities restart at once right before a consensus vote, nobody
will vote about "Running", and clients will get a consensus with no usable
relays. Instead, authorities refuse to build a consensus if this happens.
The first happens on an error case when a controller wants an
impossible directory object. The second happens when we can't write
our fingerprint file.
The code for these was super-wrong, but will only break things when we
reset an option on a platform where sizeof(time_t) is different from
sizeof(int).
See task 1114. The most plausible explanation for someone sending us weak
DH keys is that they experiment with their Tor code or implement a new Tor
client. Usually, we don't care about such events, especially not on warn
level. If we really care about someone not following the Tor protocol, we
can set ProtocolWarnings to 1.
This patch introduces a new type called document_signature_t to represent the
signature of a consensus document. Now, each consensus document can have up
to one document signature per voter per digest algorithm. Also, each
detached-signatures document can have up to one signature per <voter,
algorithm, flavor>.
Previously, we insisted that a valid signature must be a signature of
the expected digest. Now we accept anything that starts with the
expected digest. This lets us include another digest later.
When we tried to use the deprecated non-threadsafe evdns
interfaces in Libevent 2 without using the also-deprecated
event_init() interface, Libevent 2 would sensibly crash, since it
has no guess where to find the Libevent library.
Here we use the evdns_base_*() functions instead if they're
present, and fake them if they aren't.
This is a possible fix for bug 1023, where if we vote (or make a v2
consensus networkstatus) right after we come online, we can call
rep_hist_note_router_unreachable() on every router we haven't connected
to yet, and thereby make all their uptime values reset.
This seems to be happening to me a lot on a garbage DSL line.
We may need to come up with 2 threshholds: a high short onehop
count and a lower longer count.
Don't count one-hop circuits when we're estimating how long it
takes circuits to build on average. Otherwise we'll set our circuit
build timeout lower than we should. Bugfix on 0.2.2.2-alpha.
Directory authorities now reject Tor relays with versions less than
0.1.2.14. This step cuts out four relays from the current network,
none of which are very big.
Previously, tor-gencert would call RSA_generate_key() directly.
This won't work on Android, which removes the (deprecated since
OpenSSL 0.9.8) function. We can't call RSA_generate_key_ex()
unconditionally either, since that didn't exist before 0.9.8.
Instead, we must call our own crypto_pk_generate_key_with_bits,
which knows how to call RSA_generate_key or RSA_generate_key_ex as
appropriate.
[Based on patch by Nathan Freitas]
Apparently the Android developers dumped OpenSSL's support for hardware
acceleration in order to save some memory, so you can't build programs using
engines on Android.
[Patch revised by nickm]
This shouldn't be necessary, but apparently the Android cross-compiler
doesn't respect -I as well as it should. (-I is supposed to add to the
*front* of the search path. Android's gcc wrapper apparently likes to add to
the end. This is broken, but we need to work around it.)
Found by coverity
test_mem_op_hex was leaking memory, which showed up in a few
tests.
Also, the dir_param test had a memleak of its own.
Found by Coverity
- Avoid memmoving 0 bytes which might lead to compiler warnings.
- Don't require relays to be entry node AND bridge at the same to time to
record clients.
- Fix a memory leak when writing dirreq-stats.
- Don't say in the stats files that measurement intervals are twice as long
as they really are.
- Reduce minimum observation time for requests to 12 hours, or we might
never record usage.
- Clear exit stats correctly after writing them, or we accumulate old stats
over time.
- Reset interval start for buffer stats, too.
The big change is to add a function to display the current SSL handshake
state, and to log it everywhere reasonable. (A failure in
SSL23_ST_CR_SRVR_HELLO_A is different from one in
SSL3_ST_CR_SESSION_TICKET_A.)
This patch also adds a new log domain for OR handshaking, so you can pull out
all the handshake log messages without having to run at debug for everything.
For example, you'd just say "log notice-err [handshake]debug-err file
tor.log".
This was the only log notice that happened during other
tor invocations, like --verify-config and --list-fingerprint.
Plus, now we think it works, so no need to hear about it.
"Tinytest" is a minimalist C unit testing framework I wrote for
Libevent. It supports some generally useful features, like being able
to run separate unit tests in their own processes.
I tried to do the refactoring to change test.c as little as possible.
Thus, we mostly don't call the tinytest macros directly. Instead, the
test.h header is now a wrapper on tinytest.h to make our existing
test_foo() macros work.
The next step(s) here will be:
- To break test.c into separate files, each with its own test group.
- To look into which things we can test
- To refactor the more fiddly tests to use the tinytest macros
directly and/or run forked.
- To see about writing unit tests for things we couldn't previously
test without forking.
If the networkstatus consensus tells us that we should use a
negative circuit package window, ignore it. Otherwise we'll
believe it and then trigger an assert.
Also, change the interface for networkstatus_get_param() so we
don't have to lookup the consensus beforehand.
A) We were considering a circuit had timed out in the special cases
where we close rendezvous circuits because the final rendezvous
circuit couldn't be built in time.
B) We were looking at the wrong timestamp_created when considering
a timeout.
Don't discard all circuits every MaxCircuitDirtiness, because the
user might legitimately have set that to a very lower number.
Also don't use up all of our idle circuits with testing circuits,
since that defeats the point of preemptive circuits.
We want it to be under our control so it doesn't mess
up initialization. This is likely the cause for
the bug the previous assert-adding commit (09a75ad) was
trying to address.
Using CircuitBuildTimeout is prone to issues with SIGHUP, etc.
Also, shuffle the circuit build times array after loading it
in so that newer measurements don't replace chunks of
similarly timed measurements.
To further attempt to fix bug 1090, make sure connection_ap_can_use_exit
always returns 0 when the chosen exit router is excluded. This should fix
bug1090.
When we excluded some Exits, we were sometimes warning the user that we
were going to use the node regardless. Many of those warnings were in
fact bogus, because the relay in question was not used to connect to
the outside world.
Based on patch by Rotor, thanks!
Tor now reads the "circwindow" parameter out of the consensus,
and uses that value for its circuit package window rather than the
default of 1000 cells. Begins the implementation of proposal 168.
This code adds a new field to vote on: "params". It consists of a list of
sorted key=int pairs. The output is computed as the median of all the
integers for any key on which anybody voted.
Improved with input from Roger.
Adding the same vote to a networkstatus consensus leads to a memory leak
on the client side. Fix that by only using the first vote from any given
voter, and ignoring the others.
Problem found by Rotor, who also helped writing the patch. Thanks!
A vote may only contain exactly one signature. Make sure we reject
votes that violate this.
Problem found by Rotor, who also helped writing the patch. Thanks!
Fix an obscure bug where hidden services on 64-bit big-endian
systems might mis-read the timestamp in v3 introduce cells, and
refuse to connect back to the client. Discovered by "rotor".
Bugfix on 0.2.1.6-alpha.
(Given that we're pretty much assuming that int is 32 bits, and given that
hex values are always unsigned, taking out the "ul" from 0xff000000 should
be fine.)
Add a "getinfo status/accepted-server-descriptor" controller
command, which is the recommended way for controllers to learn
whether our server descriptor has been successfully received by at
least on directory authority. Un-recommend good-server-descriptor
getinfo and status events until we have a better design for them.
We were telling the controller about CHECKING_REACHABILITY and
REACHABILITY_FAILED status events whenever we launch a testing
circuit or notice that one has failed. Instead, only tell the
controller when we want to inform the user of overall success or
overall failure. Bugfix on 0.1.2.6-alpha. Fixes bug 1075. Reported
by SwissTorExit.
When we added support for fractional units (like 1.5 MB) I broke
support for giving units with no space (like 2MB). This patch should
fix that. It also adds a propoer tor_parse_double().
Fix for bug 1076. Bugfix on 0.2.2.1-alpha.
We were triggering a CLOCK_SKEW controller status event whenever
we connect via the v2 connection protocol to any relay that has
a wrong clock. Instead, we should only inform the controller when
it's a trusted authority that claims our clock is wrong. Bugfix
on 0.2.0.20-rc; starts to fix bug 1074. Reported by SwissTorExit.
- Refactor geoip.c by moving duplicate code into rotate_request_period().
- Don't leak memory when cleaning up cell queues.
- Make sure that exit_(streams|bytes_(read|written)) are initialized in all
places accessing these arrays.
- Read only the last block from *stats files and ensure that its timestamp
is not more than 25 hours in the past and not more than 1 hour in the
future.
- Stop truncating the last character when reading *stats files.
The only thing that's left now is to avoid reading whole *stats files into
memory.
Note that unlike subversion revision numbers, it isn't meaningful to
compare these for anything but equality. We define a sort-order anyway,
in case one of these accidentally slips into a recommended-versions
list.
If any the v3 certs we download are unparseable, we should actually
notice the failure so we don't retry indefinitely. Bugfix on 0.2.0.x;
reported by "rotator".
Once we had called log_free_all(), anything that tried to log a
message (like a failed tor_assert()) would fail like this:
1. The logging call eventually invokes the _log() function.
2. _log() calls tor_mutex_lock(log_mutex).
3. tor_mutex_lock(m) calls tor_assert(m).
4. Since we freed the log_mutex, tor_assert() fails, and tries to
log its failure.
5. GOTO 1.
Now we allocate the mutex statically, and never destroy it on
shutdown.
Bugfix on 0.2.0.16-alpha, which introduced the log mutex.
This bug was found by Matt Edman.
(This would be everywhere running OpenSSL 0.9.7x and earlier, including
all current Macintosh users.)
The code is based on Tom St Denis's LibTomCrypt implementation,
modified to be way less general and use Tor's existing facilities. I
picked this one because it was pretty fast and pretty free, and
because Python uses it too.
The more verbose logs that were added in ee58153 also include a string
that might not have been initialized. This can lead to segfaults, e.g.,
when setting up private Tor networks. Initialize this string with NULL.
Send circuit or stream sendme cells when our window has decreased
by 100 cells, not when it has decreased by 101 cells. Bug uncovered
by Karsten when testing the "reduce circuit window" performance
patch. Bugfix on the 54th commit on Tor -- from July 2002,
before the release of Tor 0.0.0. This is the new winner of the
oldest-bug prize.
I don't think we actually use (or plan to use) strtok_r in a reentrant
way anywhere in our code, but would be nice not to have to think about
whether we're doing it.
Relays no longer publish a new server descriptor if they change
their MaxAdvertisedBandwidth config option but it doesn't end up
changing their advertised bandwidth numbers. Bugfix on 0.2.0.28-rc;
fixes bug 1026. Patch from Sebastian.
Specifically, every time we get a create cell but we have so many already
queued that we refuse it.
Bugfix on 0.2.0.19-alpha; fixes bug 1034. Reported by BarkerJr.
The problem is that clients and hidden services are receiving
relay_early cells, and they tear down the circuit.
Hack #1 is for rendezvous points to rewrite relay_early cells to
relay cells. That way there are never any incoming relay_early cells.
Hack #2 is for clients and hidden services to never send a relay_early
cell on an established rendezvous circuit. That works around rendezvous
points that haven't upgraded yet.
Hack #3 is for clients and hidden services to not tear down the circuit
when they receive an inbound relay_early cell. We already refuse extend
cells at clients.
When determining how long directory requests take or how long cells spend
in queues, we were comparing timestamps on microsecond detail only to
convert results to second or millisecond detail later on. But on 32-bit
architectures this means that 2^31 microseconds only cover time
differences of up to 36 minutes. Instead, compare timestamps on
millisecond detail.
Now that we require EntryStatistics to be 1 for counting connecting
clients, unit tests need to set that config option, too.
Reported by Sebastian Hahn.
Changes to directory request statistics:
- Rename GEOIP statistics to DIRREQ statistics, because they now include
more than only GeoIP-based statistics, whereas other statistics are
GeoIP-dependent, too.
- Rename output file from geoip-stats to dirreq-stats.
- Add new config option DirReqStatistics that is required to measure
directory request statistics.
- Clean up ChangeLog.
Also ensure that entry guards statistics have access to a local GeoIP
database.
This new option will allow clients to download the newest fresh consensus
much sooner than they normally would do so, even if they previously set
FetchDirInfoEarly. This includes a proper ChangeLog entry and an updated man
page.
Introduce a threshold of 0.01% of bytes that must be read and written per
port in order to be included in the statistics. Otherwise we cannot include
these statistics in extra-info documents, because they are too big.
Change the labels "-written" and "-read" so that the meanings are as
intended.
The internal error "could not find intro key" occurs when we want to send
an INTRODUCE1 cell over a recently finished introduction circuit and think
we built the introduction circuit with a v2 hidden service descriptor, but
cannot find the introduction key in our descriptor.
My first guess how we can end up in this situation is that we are wrong in
thinking that we built the introduction circuit based on a v2 hidden
service descriptor. This patch checks if we have a v0 descriptor, too, and
uses that instead.
o Minor features:
- If we're a relay and we change our IP address, be more verbose
about the reason that made us change. Should help track down
further bugs for relays on dynamic IP addresses.
when we write out our stability info, detect relays that have slipped
through the cracks. log about them and correct the problem.
if we continue to see a lot of these over time, it means there's another
spot where relays fall out of the routerlist without being marked as
unreachable.
