Found because LibreSSL has OPENSSL_NO_COMP always-on, but this
conflicts with the way that _we_ turn off compression. Patch from
dhill, who attributes it to "OpenBSD". Fixes bug 12602; bugfix on
0.2.1.1-alpha, which introduced this turn-compression-off code.
Long ago we supported systems where there was no support for
threads, or where the threading library was broken. We shouldn't
have do that any more: on every OS that matters, threads exist, and
the OS supports running threads across multiple CPUs.
This resolves tickets 9495 and 12439. It's a prerequisite to making
our workqueue code work better, since sensible workqueue
implementations don't split across multiple processes.
When we create a process yourself with CreateProcess, we get a
handle to the process in the PROCESS_INFO output structure. But
instead of using that handle, we were manually looking up a _new_
handle based on the process ID, which is a poor idea, since the
process ID might refer to a new process later on, but the handle
can't.
This lets us avoid sending SIGTERM to something that has already
died, since we realize it has already died, and is a fix for the
unix version of #8746.
This function is supposed to construct a list of all the ciphers in
the "v2 link protocol cipher list" that are supported by Tor's
openssl. It does this by invoking ssl23_get_cipher_by_char on each
two-byte ciphersuite ID to see which ones give a match. But when
ssl23_get_cipher_by_char cannot find a match for a two-byte SSL3/TLS
ciphersuite ID, it checks to see whether it has a match for a
three-byte SSL2 ciphersuite ID. This was causing a read off the end
of the 'cipherid' array.
This was probably harmless in practice, but we shouldn't be having
any uninitialized reads.
(Using ssl23_get_cipher_by_char in this way is a kludge, but then
again the entire existence of the v2 link protocol is kind of a
kludge. Once Tor 0.2.2 clients are all gone, we can drop this code
entirely.)
Found by starlight. Fix on 0.2.4.8-alpha. Fixes bug 12227.
The old cache had problems:
* It needed to be manually preloaded. (It didn't remember any
address you didn't tell it to remember)
* It was AF_INET only.
* It looked at its cache even if the sandbox wasn't turned on.
* It couldn't remember errors.
* It had some memory management problems. (You can't use memcpy
to copy an addrinfo safely; it has pointers in.)
This patch fixes those issues, and moves to a hash table.
Fixes bug 11970; bugfix on 0.2.5.1-alpha.
These are needed under some circumstances if we are running with
expensive-hardening and sandbox at the same time.
fixes 11477, bugfix on 0.2.5.4-alpha (where we introduced
expensive-hardening)
None of the things we might exec() can possibly run under the
sanbox, so rather than crash later, we have to refuse to accept the
configuration nice and early.
The longer-term solution is to have an exec() helper, but wow is
that risky.
fixes 12043; bugfix on 0.2.5.1-alpha
When running with User set, we frequently try to look up our
information in the user database (e.g., /etc/passwd). The seccomp2
sandbox setup doesn't let us open /etc/passwd, and probably
shouldn't.
To fix this, we have a pair of wrappers for getpwnam and getpwuid.
When a real call to getpwnam or getpwuid fails, they fall back to a
cached value, if the uid/gid matches.
(Granting access to /etc/passwd isn't possible with the way we
handle opening files through the sandbox. It's not desirable either.)
On OpenBSD 5.4, time_t is a 32-bit integer. These instances contain
implicit treatment of long and time_t as comparable types, so explicitly
cast to time_t.
The memarea_strndup() function would have hit undefined behavior by
creating an 'end' pointer off the end of a string if it had ever been
given an 'n' argument bigger than the length of the memory ares that
it's scanning. Fortunately, we never did that except in the unit
tests. But it's not a safe behavior to leave lying around.
If we had an address of the form "1.2.3.4" and we tried to pass it to
tor_inet_pton with AF_INET6, it was possible for our 'eow' pointer to
briefly move backwards to the point before the start of the string,
before we moved it right back to the start of the string. C doesn't
allow that, and though we haven't yet hit a compiler that decided to
nuke us in response, it's best to fix.
So, be more explicit about requiring there to be a : before any IPv4
address part of the IPv6 address. We would have rejected addresses
without a : for not being IPv6 later on anyway.
Previously we said "Sandbox is not implemented on this platform" on
Linux boxes without libseccomp. Now we say that you need to build
Tor built with libseccomp. Fixes bug 11543; bugfix on 0.2.5.1-alpha.
The server cipher list is (thanks to #11513) chosen systematically to
put the best choices for Tor first. The client cipher list is chosen
to resemble a browser. So let's set SSL_OP_CIPHER_SERVER_PREFERENCE
to have the servers pick according to their own preference order.
Older versions of Libevent are happy to open SOCK_DGRAM sockets
non-cloexec and non-nonblocking, and then set those flags
afterwards. It's nice to be able to allow a flag to be on or off in
the sandbox without having to enumerate all its values.
Also, permit PF_INET6 sockets. (D'oh!)
Libevent uses an arc4random implementation (I know, I know) to
generate DNS transaction IDs and capitalization. But it liked to
initialize it either with opening /dev/urandom (which won't work
under the sandbox if it doesn't use the right pointer), or with
sysctl({CTL_KERN,KERN_RANDOM,RANDOM_UUIC}). To make _that_ work, we
were permitting sysctl unconditionally. That's not such a great
idea.
Instead, we try to initialize the libevent PRNG _before_ installing
the sandbox, and make sysctl always fail with EPERM under the
sandbox.
The compiler doesn't warn about this code:
rc = seccomp_rule_add(ctx, SCMP_ACT_ALLOW, SCMP_SYS(openat), 1,
SCMP_CMP(0, SCMP_CMP_EQ, AT_FDCWD),
SCMP_CMP(1, SCMP_CMP_EQ, param->value),
SCMP_CMP(2, SCMP_CMP_EQ, O_RDONLY|...));
but note that the arg_cnt argument above is only 1. This means that
only the first filter (argument 0 == AT_FDCWD) is actually checked!
This patch also fixes the above error in the openat() filter.
Earlier I fixed corresponding errors in filters for rename() and
mprotect().
Appearently, the majority of the filenames we pass to
sandbox_cfg_allow() functions are "freeable right after". So, consider
_all_ of them safe-to-steal, and add a tor_strdup() in the few cases
that aren't.
(Maybe buggy; revise when I can test.)
(If we don't restrict rename, there's not much point in restricting
open, since an attacker could always use rename to make us open
whatever they want.)
A new set of unit test cases are provided, as well as introducing
an alternative paradigm and macros to support it. Primarily, each test
case is given its own namespace, in order to isolate tests from each
other. We do this by in the usual fashion, by appending module and
submodule names to our symbols. New macros assist by reducing friction
for this and other tasks, like overriding a function in the global
namespace with one in the current namespace, or declaring integer
variables to assist tracking how many times a mock has been called.
A set of tests for a small-scale module has been included in this
commit, in order to highlight how the paradigm can be used. This
suite gives 100% coverage to status.c in test execution.
Back in 175b2678, we allowed servers to recognize clients who are
telling them the truth about their ciphersuites, and select the best
cipher from on that list. This implemented the server side of proposal
198.
In bugs 11492, 11498, and 11499, cypherpunks found a bunch of mistakes
and omissions and typos in the UNRESTRICTED_SERVER_CIPHER_LIST we had.
In #11513, I found a couple more.
Rather than try to hand-edit this list, I wrote a short python script
to generate our ciphersuite preferences from the openssl headers.
The new rules are:
* Require forward secrecy.
* Require RSA (since our servers only configure RSA keys)
* Require AES or 3DES. (This means, reject RC4, DES, SEED, CAMELLIA,
and NULL.)
* No export ciphersuites.
Then:
* Prefer AES to 3DES.
* If both suites have the same cipher, prefer ECDHE to DHE.
* If both suites have the same DHE group type, prefer GCM to CBC.
* If both suites have the same cipher mode, prefer SHA384 to SHA256
to SHA1.
* If both suites have the same digest, prefer AES256 to AES128.
This involves some duplicate code between backtrace.c and sandbox.c,
but I don't see a way around it: calling more functions would mean
adding more steps to our call stack, and running clean_backtrace()
against the wrong point on the stack.
The major changes are to re-order some ciphers, to drop the ECDH suites
(note: *not* ECDHE: ECDHE is still there), to kill off some made-up
stuff (like the SSL_RSA_FIPS_WITH_3DES_EDE_CBC_SHA suite), to drop
some of the DSS suites... *and* to enable the ECDHE+GCM ciphersuites.
This change is autogenerated by get_mozilla_ciphers.py from
Firefox 28 and OpenSSL 1.0.1g.
Resolves ticket 11438.
Most of these are simple. The only nontrivial part is that our
pattern for using ENUM_BF was confusing doxygen by making declarations
that didn't look like declarations.
ubsan doesn't like us to do (1u<<32) when 32 is wider than
unsigned. Fortunately, we already special-case
addr_mask_get_bits(0), so we can just change the loop bounds.
This make clang's memory sanitizer happier that we aren't reading
off the end of a char[1]. We hadn't replaced the char[1] with a
char[FLEXIBLE_ARRAY_MEMBER] before because we were doing a union
trick to force alignment. Now we use __attribute__(aligned) where
available, and we do the union trick elsewhere.
Most of this patch is just replacing accesses to (x)->u.mem with
(x)->U_MEM, where U_MEM is defined as "u.mem" or "mem" depending on
our implementation.
This contains the obvious implementation using the circuitmux data
structure. It also runs the old (slow) algorithm and compares
the results of the two to make sure that they're the same.
Needs review and testing.
In a couple of places, to implement the OOM-circuit-killer defense
against sniper attacks, we have counters to remember the age of
cells or data chunks. These timers were based on wall clock time,
which can move backwards, thus giving roll-over results for our age
calculation. This commit creates a low-budget monotonic time, based
on ratcheting gettimeofday(), so that even in the event of a time
rollback, we don't do anything _really_ stupid.
A future version of Tor should update this function to do something
even less stupid here, like employ clock_gettime() or its kin.
It's possible for two threads to hit assertion failures at the same
time. If that happens, let's keep them from stomping on the same
cb_buf field.
Fixes bug 11048; bugfix on 0.2.5.2-alpha. Reported by "cypherpunks".
clang 3.4 introduced a new by-default warning about unused static
functions, which we triggered heavily for the hashtable and map function
generating macros. We can use __attribute__ ((unused)) (thanks nickm for
the suggestion :-) ) to silence these warnings.
It's increasingly apparent that we want to make sure we initialize our
PRNG nice and early, or else OpenSSL will do it for us. (OpenSSL
doesn't do _too_ bad a job, but it's nice to do it ourselves.)
We'll also need this for making sure we initialize the siphash key
before we do any hashes.
I've made an exception for cases where I'm sure that users can't
influence the inputs. This is likely to cause a slowdown somewhere,
but it's safer to siphash everything and *then* look for cases to
optimize.
This patch doesn't actually get us any _benefit_ from siphash yet,
since we don't really randomize the key at any point.
This time, we use a pthread_attr to make sure that if pthread_create
succeeds, the thread is successfully detached.
This probably isn't the big thing going on with 4345, since it'd be
a bit weird for pthread_detach to be failing. But it's worth
getting it right.
This fixes bug 10402, where the rdrand engine would use the rdrand
instruction, not as an additional entropy source, but as a replacement
for the entire userspace PRNG. That's obviously stupid: even if you
don't think that RDRAND is a likely security risk, the right response
to an alleged new alleged entropy source is never to throw away all
previously used entropy sources.
Thanks to coderman and rl1987 for diagnosing and tracking this down.
This flag prevents the creation of a console window popup on Windows. We
need it for pluggable transport executables--otherwise you get blank
console windows when you launch the 3.x browser bundle with transports
enabled.
http://msdn.microsoft.com/en-us/library/ms684863.aspx#CREATE_NO_WINDOW
The browser bundles that used Vidalia used to set this flag when
launching tor itself; it was apparently inherited by the pluggable
transports launched by tor. In the 3.x bundles, tor is launched by some
JavaScript code, which doesn't have the ability to set CREATE_NO_WINDOW.
tor itself is now being compiled with the -mwindows option, so that it
is a GUI application, not a console application, and doesn't show a
console window in any case. This workaround doesn't work for pluggable
transports, because they need to be able to write control messages to
stdout.
https://trac.torproject.org/projects/tor/ticket/9444#comment:30
The previous commit from piet would have backed out some of proposal
198 and made servers built without the V2 handshake not use the
unrestricted cipher list from prop198.
Bug not in any released Tor.
We had accidentially grown two fake ones: one for backtrace.c, and one
for sandbox.c. Let's do this properly instead.
Now, when we configure logs, we keep track of fds that should get told
about bad stuff happening from signal handlers. There's another entry
point for these that avoids using non-signal-handler-safe functions.
On platforms with the backtrace/backtrace_symbols_fd interface, Tor
can now dump stack traces on assertion failure. By default, I log
them to DataDir/stack_dump and to stderr.
Conflicts:
src/or/or.h
src/or/relay.c
Conflicts were simple to resolve. More fixes were needed for
compilation, including: reinstating the tv_to_msec function, and renaming
*_conn_cells to *_chan_cells.
According to the manpage, bridges use P256 for conformity and relays
use P224 for speed. But skruffy points out that we've gotten it
backwards in the code.
In this patch, we make the default P256 for everybody.
Fixes bug 9780; bugfix on 0.2.4.8-alpha.
This is probably not an exploitable bug, since you would need to have
errno be a large negative value in the unix pluggable-transport launcher
case. Still, best avoided.
Fixes bug 9928; bugfix on 0.2.3.18-rc.
To fix#6033, we disabled TLS 1.1 and 1.2. Eventually, OpenSSL fixed
the bug behind #6033.
I've considered alternate implementations that do more testing to see
if there's secretly an OpenSSL 1.0.1c or something that secretly has a
backport of the OpenSSL 1.0.1e fix, and decided against it on the
grounds of complexity.
tor_malloc returns void *; in C, it is not necessary to cast a
void* to another pointer type before assigning it.
tor_malloc fails with an error rather than returning NULL; it's not
necessary to check its output. (In one case, doing so annoyed Coverity.)
SCMP_CMP(a,b,c) leaves the fourth field of the structure undefined,
giving a missing-initializer error. All of our uses are
three-argument, so I'm overriding the default.
Incidentally, this business here where I make crypto_rand mockable:
this is exactly the kind of thing that would make me never want to
include test-support stuff in production builds.
This way, we don't have to use snprintf, which is not guaranteed to
be signal-safe.
(Technically speaking, strlen() and strlcpy() are not guaranteed to
be signal-safe by the POSIX standard. But I claim that they are on
every platform that supports libseccomp2, which is what matters
here.)
Better tests for upper bounds, and for failing cases.
Also, change the function's interface to take a buffer length rather
than a maximum length, and then NUL-terminate: functions that don't
NUL-terminate are trouble waiting to happen.
The only thing that used format_helper_exit_status on win32 was the
unit tests. This caused an error when we tried to leave a static
format_helper_exit_status lying around in a production object file.
The easiest solution is to admit that this way of dealing with process
exit status is Unix-only.
This is not the most beautiful possible implementation (it requires
decorating mockable functions with ugly macros), but it actually
works, and is portable across multiple compilers and architectures.
If you pass the --enable-coverage flag on the command line, we build
our testing binaries with appropriate options eo enable coverage
testing. We also build a "tor-cov" binary that has coverage enabled,
for integration tests.
On recent OSX versions, test coverage only works with clang, not gcc.
So we warn about that.
Also add a contrib/coverage script to actually run gcov with the
appropriate options to generate useful .gcov files. (Thanks to
automake, the .o files will not have the names that gcov expects to
find.)
Also, remove generated gcda and gcno files on clean.
We previously used FILENAME_PRIVATE identifiers mostly for
identifiers exposed only to the unit tests... but also for
identifiers exposed to the benchmarker, and sometimes for
identifiers exposed to a similar module, and occasionally for no
really good reason at all.
Now, we use FILENAME_PRIVATE identifiers for identifiers shared by
Tor and the unit tests. They should be defined static when we
aren't building the unit test, and globally visible otherwise. (The
STATIC macro will keep us honest here.)
For identifiers used only by the unit tests and never by Tor at all,
on the other hand, we wrap them in #ifdef TOR_UNIT_TESTS.
This is not the motivating use case for the split test/non-test
build system; it's just a test example to see how it works, and to
take a chance to clean up the code a little.
This is mainly a matter of automake trickery: we build each static
library in two versions now: one with the TOR_UNIT_TESTS macro
defined, and one without. When TOR_UNIT_TESTS is defined, we can
enable mocking and expose more functions. When it's not defined, we
can lock the binary down more.
The alternatives would be to have alternate build modes: a "testing
configuration" for building the libraries with test support, and a
"production configuration" for building them without. I don't favor
that approach, since I think it would mean more people runnning
binaries build for testing, or more people not running unit tests.
This implements "algorithm 1" from my discussion of bug #9072: on OOM,
find the circuits with the longest queues, and kill them. It's also a
fix for #9063 -- without the side-effects of bug #9072.
The memory bounds aren't perfect here, and you need to be sure to
allow some slack for the rest of Tor's usage.
This isn't a perfect fix; the rest of the solutions I describe on
codeable.
