Previously, I had used integers encoded as pointers. This
introduced a flaw: NULL represented both the integer zero, and the
absence of a setting. This in turn made the checks in
cfg_msg_set_{type,chan}() not actually check for an altered value if
the previous value had been set to zero.
Also, I had previously kept a pointer to a dispatch_fypefns_t rather
than making a copy of it. This meant that if the dispatch_typefns_t
were changed between defining the typefns and creating the
dispatcher, we'd get the modified version.
Found while investigating coverage in pubsub_add_{pub,sub}_()
We want the DISPATCH_ADD_PUB() macro to count as making a
DECLARE_PUBLISH() invocation "used", so let's try a new approach
that preserves that idea. The old one apparently did not work for
some versions of osx clang.
This code tries to prevent a large number of possible errors by
enforcing different restrictions on the messages that different
modules publish and subscribe to.
Some of these rules are probably too strict, and some too lax: we
should feel free to change them as needed as we move forward and
learn more.
This "publish/subscribe" layer sits on top of lib/dispatch, and
tries to provide more type-safety and cross-checking for the
lower-level layer.
Even with this commit, we're still not done: more checking will come
in the next commit, and a set of usability/typesafety macros will
come after.
This module implements a way to send messages from one module to
another, with associated data types. It does not yet do anything to
ensure that messages are correct, that types match, or that other
forms of consistency are preserved.
We already do this in our log_debug() macro, but there are times
when we'd like to avoid allocating or precomputing something that we
are only going to log if debugging is on.
When NULL is given to lpApplicationName we enable Windows' "magical"
path interpretation logic, which makes Tor 0.4.x behave in the same way
as previous Tor versions did when it comes to executing binaries in
different system paths.
For more information about this have a look at the CreateProcessA()
documentation on MSDN -- especially the string interpretation example is
useful to understand this issue.
This bug was introduced in commit bfb94dd2ca.
See: https://bugs.torproject.org/29874
There's an incorrect comment in compat_time.c that suggests we call
FreeLibrary() before we're done using the library's functions.
See 29642 for background.
Closes ticket 29643.
* Move out code that depends on NSS to crypto_digest_nss.c
* Move out code that depends on OpenSSL to crypto_digest_openssl.c
* Keep the general code that is not specific to any of the above in
crypto_digest.c
This test was previously written to use the contents of the system
headers to decide whether INHERIT_NONE or INHERIT_ZERO was going to
work. But that won't work across different environments, such as
(for example) when the kernel doesn't match the headers. Instead,
we add a testing-only feature to the code to track which of these
options actually worked, and verify that it behaved as we expected.
Closes ticket 29541; bugfix not on any released version of Tor.
This module is currently implemented to use the same technique as
libottery (later used by the bsds' arc4random replacement), using
AES-CTR-256 as its underlying stream cipher. It's backtracking-
resistant immediately after each call, and prediction-resistant
after a while.
Here's how it works:
We generate psuedorandom bytes using AES-CTR-256. We generate BUFLEN bytes
at a time. When we do this, we keep the first SEED_LEN bytes as the key
and the IV for our next invocation of AES_CTR, and yield the remaining
BUFLEN - SEED_LEN bytes to the user as they invoke the PRNG. As we yield
bytes to the user, we clear them from the buffer.
Every RESEED_AFTER times we refill the buffer, we mix in an additional
SEED_LEN bytes from our strong PRNG into the seed.
If the user ever asks for a huge number of bytes at once, we pull SEED_LEN
bytes from the PRNG and use them with our stream cipher to fill the user's
request.
Using an anonymous mmap() is a good way to get pages that we can set
kernel-level flags on, like minherit() or madvise() or mlock().
We're going to use that so that we can make uninheritable locked
pages to store PRNG data.
The code checked for sysctl being available and HW_PHYSMEM being
defined, but HW_USERMEM was actually being used with sysctl instead
of HW_PHYSMEM.
The case for OpenBSD, etc. use HW_PHYSMEM64 (which is obviously a
64-bit variant of HW_PHYSMEM) and the case for OSX uses HW_MEMSIZE
(which appears to be a 64-bit variant of HW_PHYSMEM).
Signed-off-by: Kris Katterjohn <katterjohn@gmail.com>
We log these messages at INFO level, except when we are reading a
private key from a file, in which case we log at WARN.
This fixes a regression from when we re-wrote our PEM code to be
generic between nss and openssl.
Fixes bug 29042, bugfix on 0.3.5.1-alpha.
When cleaning up after an error in process_unix_exec, the stdin
pipe was being double closed instead of closing both the stdin
and stdout pipes. This occurred in two places.
Signed-off-by: Kris Katterjohn <katterjohn@gmail.com>
NOTE: This commit breaks the build, because there was a mistake in an
earlier change of exactly the sort that this is meant to detect! I'm
leaving it broken for illustration.
Test exactly what the geometric sampler returns, because that's what
the downstream callers of it are going to use.
While here, also assert that the geometric sampler returns a positive
integer. (Our geometric distribution is the one suported on {1, 2,
3, ...} that returns the number of trials before the first success,
not the one supported on {0, 1, 2, ...} that returns the number of
failures before the first success.)
Prior to this commit, the testsuite was failing on OpenBSD. After
this commit the testsuite runs fine on OpenBSD.
It was previously decided to test for the OpenBSD macro (rather than
__OpenBSD__, etc.) because OpenBSD forks seem to have the former
macro defined. sys/param.h must be included for the OpenBSD macro
definition; however, many files tested for the OpenBSD macro without
having this header included.
This commit includes sys/param.h in the files where the OpenBSD macro
is used (and sys/param.h is not already included), and it also
changes some instances of the __OpenBSD__ macro to OpenBSD.
See commit 27df23abb6 which changed
everything to use OpenBSD instead of __OpenBSD__ or OPENBSD. See
also tickets #6982 and #20980 (the latter ticket is where it was
decided to use the OpenBSD macro).
Signed-off-by: Kris Katterjohn <katterjohn@gmail.com>
In get_local_listener used by tor_ersatz_socketpair, the address
family used when binding the IPv6 socket was AF_INET instead of
AF_INET6.
Fixes bug 28995; bugfix on 0.3.5.1-alpha.
Signed-off-by: Kris Katterjohn <katterjohn@gmail.com>
Reported on tor-dev by Gisle Vanem. Bug not in any released Tor
(The suggested patch used _MSC_VER, but that's not how we do stuff
with autoconf. With autoconf, you detect the feature you want,
rather than trying to list all the systems that do or do not have
it.)
This project introduces the prob_distr.c subsystem which implements all the
probability distributions that WTF-PAD needs. It also adds unittests for all of
them.
Code and tests courtesy of Riastradh.
Co-authored-by: Taylor R Campbell <campbell+tor@mumble.net>
Co-authored-by: Mike Perry <mikeperry-git@torproject.org>
We need this for padding negotiation so that we can have later machine
revisions supercede earlier ones.
Co-authored-by: George Kadianakis <desnacked@riseup.net>
This patch changes the CancelIoEx() example code to use CancelIo(),
which is available for older versions of Windows too. I still think the
kernel handles this nicely by sending broken pipes if either side
closes the pipe while I/O operations are pending.
See: https://bugs.torproject.org/28179
Handle `ERROR_BROKEN_PIPE` from ReadFileEx() and WriteFileEx() in
process_win32_stdin_write_done() and
process_win32_handle_read_completion() instead of in the early handler.
This most importantmly makes sure that `reached_eof` is set to true when
these errors appears.
See: https://bugs.torproject.org/28179
This patch adds some missing calls to set `reached_eof` of our handles
when various error conditions happens or when we close our handle (which
happens at `process_terminate()`.
See: https://bugs.torproject.org/28179
This patch adds some additional error checking after calls to
ReadFileEx() and WriteFileEx(). I have not managed to get this code to
reach the branch where `error_code` is NOT `ERROR_SUCCESS`, but MSDN
says one should check for this condition so we do so just to be safe.
See: https://bugs.torproject.org/28179
This patch makes us delay checking for whether we have an exit code
value (via GetExitCodeProcess()) until both stdout and stderr have been
closed by the operating system either by the process itself or by
process cleanup after termination.
See: https://bugs.torproject.org/28179
This patch makes sure that we terminate the event loop from the event
loop timer instead of directly in the process' exit handler. This allows
us to run the event loop an additional time to ensure that the SleepEx()
call on Windows is called and the data from stdout/stderr is delivered
to us.
Additionally we ensure that we don't try to read or write data from a
Unix process that have been terminated in the main loop, since its file
descriptors are closed at that time.
See: https://bugs.torproject.org/28179
This patch changes the API of the Windows backend of the Process
subsystem to allow the dormant interface to disable the Process event
timer.
See: https://bugs.torproject.org/28179
This patch changes our process_t's exit_callback to return a boolean
value. If the returned value is true, the process subsystem will call
process_free() on the given process_t.
See: https://bugs.torproject.org/28179
This patch moves the remaining code from subprocess.{h,c} to more
appropriate places in the process.c and process_win32.c module.
We also delete the now empty subprocess module files.
See: https://bugs.torproject.org/28179
This patch adds a new function that allows us to reset the environment
of a given process_t with a list of key/value pairs.
See: https://bugs.torproject.org/28179
This patch makes sure that we call process_notify_event_exit() after we
have done any modifications we need to do to the state of a process_t.
This allows application developers to call process_free() in the
exit_callback of the process.
See: https://bugs.torproject.org/28179
This patch adds support for getting the unique process identifier from a
given process_t. This patch implements both support for both the Unix
and Microsoft Windows backend.
See: https://bugs.torproject.org/28179
This patch adds support for Microsoft Windows in the Process subsystem.
Libevent does not support mixing different types of handles (sockets,
named pipes, etc.) on Windows in its core event loop code. This have
historically meant that Tor have avoided attaching any non-networking
handles to the event loop. This patch uses a slightly different approach
to roughly support the same features for the Process subsystem as we do
with the Unix backend.
In this patch we use Windows Extended I/O functions (ReadFileEx() and
WriteFileEx()) which executes asynchronously in the background and
executes a completion routine when the scheduled read or write operation
have completed. This is much different from the Unix backend where the
operating system signals to us whenever a file descriptor is "ready" to
either being read from or written to.
To make the Windows operating system execute the completion routines of
ReadFileEx() and WriteFileEx() we must get the Tor process into what
Microsoft calls an "alertable" state. To do this we execute SleepEx()
with a zero millisecond sleep time from a main loop timer that ticks
once a second. This moves the process into the "alertable" state and
when we return from the zero millisecond timeout all the outstanding I/O
completion routines will be called and we can schedule the next reads
and writes.
The timer loop is also responsible for detecting whether our child
processes have terminated since the last timer tick.
See: https://bugs.torproject.org/28179
This patch adds the Unix backend for the Process subsystem. The Unix
backend attaches file descriptors from the child process's standard in,
out and error to Tor's libevent based main loop using traditional Unix
pipes. We use the already available `waitpid` module to get events
whenever the child process terminates.
See: https://bugs.torproject.org/28179
This patch adds a new Process subsystem for running external programs in
the background of Tor. The design is focused around a new type named
`process_t` which have an API that allows the developer to easily write
code that interacts with the given child process. These interactions
includes:
- Easy API for writing output to the child process's standard input
handle.
- Receive callbacks whenever the child has output on either its standard
output or standard error handles.
- Receive callback when the child process terminates.
We also support two different "protocols" for handling output from the
child process. The default protocol is the "line" protocol where the
process output callbacks will be invoked only when there is complete
lines (either "\r\n" or "\n" terminated). We also support the "raw"
protocol where the read callbacks will get whatever the operating system
delivered to us in a single read operation.
This patch does not include any operating system backends, but the Unix
and Windows backends will be included in separate commits.
See: https://bugs.torproject.org/28179
The point of this function is to make sure that the ed25519-based
implementation of curve25519_basepoint() actually works when we
start tor, and use the regular fallback implementation if it
doesn't. But it accounts for 9% of our startup time in the case
when we have directory information, and I think it's safe to make
the test shorter. After all, it has yet to find any actual bugs in
curved25519_scalarmult_basepoint_donna() on any platforms.
Closes ticket 28838.
When the clock jumps, and we have a record of last user activity,
adjust that record. This way if I'm inactive for 10 minutes and
then the laptop is sleeping for an hour, I'll still count as having
been inactive for 10 minutes.
Previously, we treat every jump as if it were activity, which is
ridiculous, and would prevent a Tor instance with a jumpy clock from
ever going dormant.
This patch adds two new functions: buf_flush_to_pipe() and
buf_read_from_pipe(), which makes use of our new buf_flush_to_fd() and
buf_read_from_fd() functions.
See: https://bugs.torproject.org/28179
This patch refactors buf_read_from_socket() into buf_read_from_fd(), and
creates a specialized function for buf_read_from_socket(), which uses
buf_read_from_fd().
See: https://bugs.torproject.org/28179
This patch refactors buf_flush_to_socket() into buf_flush_to_fd() and
creates a specialization function for buf_flush_to_socket() that makes
use of buf_flush_to_fd().
See: https://bugs.torproject.org/28179
Nothing should ever look at them on failure, but in some cases,
the unit tests don't check for failure, and then GCC-LTO freaks out.
Fixes part of 27772.
OpenSolaris apparently doesn't have timeradd(), so we added a
replacement, but we weren't including it here after the big
refactoring in 0.3.5.1-alpha.
Fixes bug 27963; bugfix on 0.3.5.1-alpha.
Allowing this didn't do any actual harm, since there aren't any
shared structures or leakable objects here. Still, it's bad style
and might cause trouble in the future.
Closes ticket 27856.
The trunnel functions are written under the assumption that their
allocators can fail, so GCC LTO thinks they might return NULL. In
point of fact, they're using tor_malloc() and friends, which can't
fail, but GCC won't necessarily figure that out.
Fixes part of #27772.
Instead, have it call a mockable function. We don't want
crypto_strongest_rand() to be mockable, since doing so creates a
type error when we call it from ed25519-donna, which we do not build
in a test mode.
Fixes bug 27728; bugfix on 0.3.5.1-alpha
This shouldn't be a user-visible change: nobody has a 16 MB RSA
key that they're trying to use with Tor.
I'm doing this to fix CID 1439330 / ticket 27730, where coverity
complains (on 64-bit) that we are making a comparison that is never
true.
This is harder than with OpenSSL, since OpenSSL counts the bytes on
its own and NSS doesn't. To fix this, we need to define a new
PRFileDesc layer that has its own byte-counting support.
Closes ticket 27289.
This function tells the underlying TLS object that it shouldn't
close the fd on exit. Mostly, we hope not to have to use it, since
the NSS implementation is kludgey, but it should allow us to fix
This is an attempt to work around what I think may be a bug in
OSS-Fuzz, which thinks that uninitialized data might be passed to
the curve25519 functions.
This reverts commit b5fddbd241.
The commit here was supposed to be a solution for #27451 (fd
management with NSS), but instead it caused an assertion failure.
Fixes bug 27500; but not in any released Tor.
On new glibc versions, there's an explicit_bzero(). With openssl,
there's openssl_memwipe().
When no other approach works, use memwipe() and a memory barrier.
This function was a wrapper around RSA_check_key() in openssl, which
checks for invalid RSA private keys (like those where p or q are
composite, or where d is not the inverse of e, or where n != p*q).
We don't need a function like this in NSS, since unlike OpenSSL, NSS
won't let you import a bogus private key.
I've renamed the function and changed its return type to make it
more reasonable, and added a unit test for trying to read a key
where n != p*q.
This function was supposed to implement a half-duplex mode for our
TLS connections. However, nothing in Tor actually uses it (besides
some unit tests), and the implementation looks really questionable
to me. It's probably best to remove it. We can add a tested one
later if we need one in the future.
We used to link both libraries at once, but now that I'm working on
TLS, there's nothing left to keep OpenSSL around for when NSS is
enabled.
Note that this patch causes a couple of places that still assumed
OpenSSL to be disabled when NSS is enabled
- tor-gencert
- pbkdf2
Also, add a stubbed-out nss version of the modules. The tests won't
pass with NSS yet since the NSS modules don't do anything.
This is a good patch to read with --color-moved.
This cleans up a lot of junk from crypto_rsa_openssl, and will
save us duplicated code in crypto_rsa_nss (when it exists).
(Actually, it already exists, but I am going to use git rebase so
that this commit precedes the creation of crypto_rsa_nss.)
This change also makes tor_ersatz_socketpair() follow the same
interface as socketpair() rather than tor_socketpair(), so it now
needs to be wrapped in the same code as socketpair() does.
This is comparatively straightforward too, except for a couple of
twists:
* For as long as we're building with two crypto libraries, we
want to seed _both_ their RNGs, and use _both_ their RNGs to
improve the output of crypto_strongest_rand()
* The NSS prng will sometimes refuse to generate huge outputs.
When it does, we stretch the output with SHAKE. We only need
this for the tests.
