There is a bug causing busy loops in Libevent and infinite loops in
the Shadow simulator. A connection that is marked for close, wants
to flush, is held open to flush, but is rate limited (the token
bucket is empty) triggers the bug.
This commit fixes the bug. Details are below.
This currently happens on read and write callbacks when the active
socket is marked for close. In this case, Tor doesn't actually try
to complete the read or write (it returns from those methods when
marked), but instead tries to clear the connection with
conn_close_if_marked(). Tor will not close a marked connection that
contains data: it must be flushed first. The bug occurs when this
flush operation on the marked connection can not occur because the
connection is rate-limited (its write token bucket is empty).
The fix is to detect when rate limiting is preventing a marked
connection from properly flushing. In this case, it should be
flagged as read/write_blocked_on_bandwidth and the read/write events
de-registered from Libevent. When the token bucket gets refilled, it
will check the associated read/write_blocked_on_bandwidth flag, and
add the read/write event back to Libevent, which will cause it to
fire. This time, it will be properly flushed and closed.
The reason that both read and write events are both de-registered
when the marked connection can not flush is because both result in
the same behavior. Both read/write events on marked connections will
never again do any actual reads/writes, and are only useful to
trigger the flush and close the connection. By setting the
associated read/write_blocked_on_bandwidth flag, we ensure that the
event will get added back to Libevent, properly flushed, and closed.
Why is this important? Every Shadow event occurs at a discrete time
instant. If Tor does not properly deregister Libevent events that
fire but result in Tor essentially doing nothing, Libevent will
repeatedly fire the event. In Shadow this means infinite loop,
outside of Shadow this means wasted CPU cycles.
This is a feature removal: we no longer fake any ciphersuite other
than the not-really-standard SSL_RSA_FIPS_WITH_3DES_EDE_CBC_SHA
(0xfeff). This change will let servers rely on our actually
supporting what we claim to support, and thereby let Tor migrate to
better TLS ciphersuites.
As a drawback, Tor instances that use old openssl versions and
openssl builds with ciphers disabled will no longer give the
"firefox" cipher list.
Manually removed range 0.116.0.0 to 0.119.255.255 which Maxmind says is
assigned to AT. This is very likely a bug in their database, because
0.0.0.0/8 is a reserved range.
From what I can tell, this configuration is usually a mistake, and
leads people to think that all their traffic is getting proxied when
in fact practically none of it is. Resolves the issue behind "bug"
4663.
The function is not guaranteed to NUL-terminate its output. It
*is*, however, guaranteed not to generate more than two bytes per
multibyte character (plus terminating nul), so the general approach
I'm taking is to try to allocate enough space, AND to manually add a
NUL at the end of each buffer just in case I screwed up the "enough
space" thing.
Fixes bug 5909.
This feature can make Tor relays less identifiable by their use of the
mod_ssl DH group, but at the cost of some usability (#4721) and bridge
tracing (#6087) regressions.
We should try to turn this on by default again if we find that the
mod_ssl group is uncommon and/or we move to a different DH group size
(see #6088). Before we can do so, we need a fix for bugs #6087 and
Resolves ticket #5598 for now.
These include:
- Having a weird in_addr that can't be initialized with {0}
- Needing INVALID_HANDLE_VALUE instead of -1 for file handles.
- Having a weird dependent definition for struct stat.
- pid is signed, not unsigned.
These stats are currently discarded, but we might as well
hard-disable them on bridges, to be clean.
Fix for bug 5824; bugfix on 0.2.1.17-rc.
Patch originally by Karsten Loesing.
Also, try to resolve some doxygen issues. First, define a magic
"This is doxygen!" macro so that we take the correct branch in
various #if/#else/#endifs in order to get the right documentation.
Second, add in a few grouping @{ and @} entries in order to get some
variables and fields to get grouped together.
This code shouldn't have any effect in 0.2.3, since we already accept
(and handle) data received while we are expecting a renegotiation.
(That's because the 0.2.3.x handshake _does_ have data there instead of
the renegotiation.)
I'm leaving it in anyway, since if it breaks anything, we'll want it
broken in master too so we can find out about it. I added an XXX023
comment so that we can come back later and fix that.
This fixes a DoS issue where a client could send so much data in 5
minutes that they exhausted the server's RAM. Fix for bug 5934 and
6007. Bugfix on 0.2.0.20-rc, which enabled the v2 handshake.
This fixes a warning in efb8a09f, where Debain Lenny's GCC doesn't get
that
for (i=0; i<3; ++i) {
const char *p;
switch(i) {
case 0:
p="X"; break;
case 1:
p="Y"; break;
case 2:
p="Z"; break;
}
printf("%s\n", p);
}
will never try to print an uninitialezed value.
Found by buildbots. Bug in no released versions of Tor.
It appears that when OpenSSL negotiates a 1.1 or 1.2 connection, and it
decides to renegotiate, the client will send a record with version "1.0"
rather than with the current TLS version. This would cause the
connection to fail whenever both sides had OpenSSL 1.0.1, and the v2 Tor
handshake was in use.
As a workaround, disable TLS 1.1 and TLS 1.2. When a later version of
OpenSSL is released, we can make this conditional on running a fixed
version of OpenSSL.
Alternatively, we could disable TLS 1.1 and TLS 1.2 only on the client
side. But doing it this way for now means that we not only fix TLS with
patched clients; we also fix TLS when the server has this patch and the
client does not. That could be important to keep the network running
well.
Fixes bug 6033.
This solves bug 5283, where client traffic could get sent over the
same circuit as an anonymized connection to a directory, even if
that circuit used an exit node unsuitable for clients. By marking
the directory connection as needs_internal, we ensure that the
(non-internal!) client-traffic connection won't be sent over the
same circuit.
Conflicts:
src/test/test_util.c
Merge the unit tests; I added some when I did this branch against
0.2.2, and then the test format changed and master added more tests.
Conflicts:
src/common/compat.h
Conflict was between replacement of MS_WINDOWS with _WIN32 in
master, and with removal of file_handle from tor_mmap_t struct in
close_file_mapping branch (for bug 5951 fix).
It turns out that if you set the third argument of
__attribute__(format) to 0, GCC and Clang will check the format
argument without expecting to find variadic arguments. This is the
correct behavior for vsnprintf, vasprintf, and vscanf.
I'm hoping this will fix bug 5969 (a clang warning) by telling clang that
the format argument to tor_vasprintf is indeed a format string.
On Windows, getsockname() on a nonblocking apparently won't work
until the connection is done connecting. On XP, it seems to fail by
reporting success and declaring that your address is INADDR_ANY. On the
Win8 preview, though, it fails more loudly and says WSAEINVAL.
Fix for bug 5374; bugfix on 0.1.1.14-alpha.
The parent of "/foo" is "/"; and "/" is its own parent.
This would cause Tor to fail if you tried to have a PF_UNIX control
socket in the root directory. That would be a stupid thing to do
for other reasons, but there's no reason to fail like _this_.
Bug found by Esteban Manchado Velázquez. Fix for bug 5089; bugfix on
Tor 0.2.2.26-beta. Unit test included.
Roger explains at
http://archives.seul.org/tor/talk/Nov-2011/msg00209.html :
"If you list your bridge as part of your family in the relay
descriptor, then everybody can learn your bridge fingerprint, and
they can look up your bridge's descriptor (and thus location) at
the bridge directory authority."
Now, we can't stop relays from listing bridges, but we can warn when
we notice a bridge listing anybody, which might help some.
This fixes bug 4657; it's a fix on 0.2.0.3-alpha, where bridges were
first introduced.
To hit this leak, you need to be a relay that gets a RESOLVE request
or an exit node getting a BEGIN or RESOLVE request. You must either
have unconfigured (and unconfigurable) nameservers, or you must have
somehow set DisableNetwork after a network request arrived but
before you managed to process it.
So, I doubt this is reached often. Still, a leak's a leak. Fix for
bug 5916; bugfix on 0.2.3.9-alpha and 0.1.2.1-alpha.
%f is correct; %lf is only needed with scanf. Apparently, on some
old BSDs, %lf is deprecated.
Didn't we do this before? Yes, we did. But we only got the
instances of %lf, not more complicated things like %.5lf . This
patch tries to get everything.
Based on a patch for 3894 by grarpamp.
These errors usually mean address exhaustion; reporting them as such
lets clients adjust their load to try other exits.
Fix for bug 4710; bugfix on 0.1.0.1-rc, which started using
END_STREAM_REASON_RESOURCELIMIT.
(When the correct answer is given in terms of seconds since the
epoch, it's hard to be sure that it really is the right answer
just by reading the code.)
* It seems parse_http_time wasn't parsing correctly any date with commas (RFCs
1123 and 850). Fix that.
* It seems parse_http_time was reporting the wrong month (they start at 0, not
1). Fix that.
* Add some tests for parse_http_time, covering all three formats.
Silences the log message:
[warn] {BUG} _connection_mark_unattached_ap(): Bug: stream (marked at connection_edge.c:2224) sending two socks replies?
after the client triggered the "Tor is not an HTTP Proxy" response.
No additional socks reply was sent, though.
Previously, we only did this check at startup, which could lead to
us holding a guard indefinitely, and give weird results. Fixes bug
5380; bugfix on 0.2.1.14-rc.
(Patch by Roger; changes file and commit message by Nick)
Previously, we skipped everything that got invoked from
options_init_from_torrc. But some of the stuff in
options_act_reversible and options_act is actually important, like
reopening the logs.
Now, a SIGHUP always makes the effects of an options_set() happen,
even though the options haven't changed.
Fix for bug 5095; bugfix on 0.2.1.9-alpha, which introduced
__ReloadTorrcOnSIGHUP.
This would happen if the deliver window could become negative
because of an nonexistent connection. (Fortunately, _that_ can't
occur, thanks to circuit_consider_sending_sendme. Still, if we
change our windowing logic at all, we won't want this to become
triggerable.) Fix for bug 5541. Bugfix on 4a66865d, back from
0.0.2pre14. asn found this. Nice catch, asn!
We've been only treating SW_SERVER_HELLO_A as meaning that an SSL
handshake was happening. But that's not right: if the initial
attempt to write a ServerHello fails, we would get a callback in
state SW_SERVER_HELLO_B instead.
(That's "instead" and not "in addition": any failed attempt to write
the hello will fail and cause the info callback not to get written.)
Fix for bug 4592; bugfix on 0.2.0.13-alpha.
This tells the windows headers to give us definitions that didn't
exist before XP -- like the ones that we need for IPv6 support.
See bug #5861. We didn't run into this issue with mingw, since
mingw doesn't respect _WIN32_WINNT as well as it should for some of
its definitions.
MSVC warns if you declare a function as having a "int foo" argument
and then implement it with a "const int foo" argument, even though
the latter "const" is not a part of the function's interface.
Instead, allow packagers to put a 'TOR_BUILD_TAG' field in the
server descriptor to indicate a platform-specific value, if they
need to. (According to weasel, this was his use for the git- tag
previously.)
This is part of 2988
For uname-based detection, we now give only the OS name (e.g.,
"Darwin", "Linux".) For Windows, we give only the Operating System
name as inferred from dw(Major|Minor)version, (e.g., "Windows XP",
"Windows 7"), and whether the VER_NT_SERVER flag is set.
For ticket 2988.
This time, I follow grarpamp's suggestion and move the check for
.exit+AllowDotExit 0 to the top of connection_ap_rewrite_and_attach,
before any rewriting occurs. This way, .exit addresses are
forbidden as they arrive from a socks connection or a DNSPort
request, and not otherwise.
It _is_ a little more complicated than that, though. We need to
treat any .exit addresses whose source is TrackHostExits as meaning
that we can retry without that exit. We also need to treat any
.exit address that comes from an AutomapHostsOnResolve operation as
user-provided (and thus forbidden if AllowDotExits==0), so that
transitioning from AllowDotExits==1 to AllowDotExits==0 will
actually turn off automapped .exit addresses.
This patch changes the total serverdesc threshold from 25% to 75%
and the exit threshold from 33% to 50%. The goal is to make
initially constructed circuits less horrible, and to make initial
less awful (since fetching directory information in parallel with
whatever the user is trying to do can hurt their performance).
Implements ticket 3196.
We were doing an O(n) strlen in router_get_extrainfo_hash() for
every one we tried to parse. Instead, have
router_get_extrainfo_hash() take the length of the extrainfo as an
argument, so that when it's called from
extrainfo_parse_from_string(), it doesn't do a strlen() over the
whole pile of extrainfos.
Now that the pt code logs mp->argv[0] all over the place, we need to
be sure to set up mp->argv in our tests.
Bugfix on e603692adc, not in any released version.
If the authorities agreed on a sufficiently bad bwweightscale value
(<=0 or == INT32_MAX), the bandwidth algorithm could make the voters
assert while computing the consensus.
Fix for bug5786; bugfix on 0.2.2.17-alpha
The underlying strtoX functions handle overflow by saturating and
setting errno to ERANGE. If the min/max arguments to the
tor_parse_* functions are equal to the minimum/maximum of the
underlying type, then with the old approach, we wouldn't treat a
too-large value as genuinely broken.
Found this while looking at bug 5786; bugfix on 19da1f36 (in Tor
0.0.9), which introduced these functions.
(Note: It makes sense to use tor-gencert on Windows for testing
purposes only. If you are a directory authority operator, and you
are contemplating running tor-gencert on a Windows box in an actual
production environment, you are probably making a mistake.)
We had been checking for EINVAL, but that means that SOCK_* isn't
supported, not that the syscall itself is missing.
Bugfix on 0.2.3.1-alpha, which started to use accept4.
This is how IPv6 says "0.0.0.0" and something we will have to
translate into a globally reachable address before putting it in a
descriptor.
The fix is a short term solution until a real one is implemented.
Closes#5146.
I think that the trailing __ got added in false analogy to
HAVE_MACRO__func__, HAVE_MACRO__FUNC__, and HAVE_MACRO__FUNCTION__.
But those macros actually indicate the presence of __func__,
__FUNC__, and __FUNCTION__ respectively. The __ at the end of
HAVE_EXTERN_ENVIRON_DECLARED would only be appropriate if the
environ were declared__, whatever that means.
(As a side-note, HAVE_MACRO__func__ and so on should probably be
renamed HAVE_MACRO___func__ and so on. But that can wait.)
This is an identifier renaming only.
If the client uses a v2 cipherlist on the renegotiation handshake,
it looks as if they could fail to get a good cert chain from the
server, since they server would re-disable certificate chaining.
This patch makes it so the code that make the server side of the
first v2 handshake special can get called only once.
Fix for 4591; bugfix on 0.2.0.20-rc.
They boil down to:
- MS_WINDOWS is dead and replaced with _WIN32, but we let a few
instances creep in when we merged Esteban's tests.
- Capitalizing windows header names confuses mingw.
- #ifdef 0 ain't C.
- One unit test wasn't compiled on windows, but was being listed
anyway.
- One unit test was checking for the wrong value.
Gisle Vanem found and fixed the latter 3 issues.
Fixes bug #4528 "read_to_buf_tls(): Inconsistency in code".
This check was added back in 0.1.0.3-rc, but somehow we forgot to
leave it in when we refactored read_to_buf_tls in 0.1.0.5-rc.
(patch by Arturo; commit message and changes file by nickm)
Instead of checking for 'rejected' and calling everything else okay,
let's check for 'outdated' and call everythign else a problem. This
way we don't risk missing future errors so much.
When logging a message that _looks_ like an error message at info, we
should mention that it isn't really a problem.
