Using the talloc leak report we see that there are some msgb's
that are allocated for SMS but we don't have transactions or
SMS around. We need to improve the name of the messages to
uniquely dscribe where they are from but the obvious leak does
occur in this routine.
The no available transaction id is most likely the case where
we leak memory. This should not occur and shows another issue
with the smsqueue/smpp handling. It doesn't explain the subscr
reference count issue either.
Extract of the leak report:
GSM 04.11 contains 1160 bytes in 1 blocks (ref 0) 0x2517dc0
GSM 04.11 contains 1160 bytes in 1 blocks (ref 0) 0x24b56e0
GSM 04.11 contains 1160 bytes in 1 blocks (ref 0) 0x23e7930
At the rhizomatica sites some unauthorized subscribers have a
last used from long time ago but a bigger reference count. This
means that we have an unbalanced subscr_get/subscr_put that we
need to investigate. These two fixes are genuine fixes but are
unlikely to fix the issue I am not understanding yet.
In case the subscriber is currently busy we would omit the
subscr_put. This seems to be very hard to hit as the subscr
need to be active and at the same time be selected for the
purge operation.
Various paging clean-ups. I couldn't re-produce the assert that
was hit by Rhizomatica in the subscriber code. Let's clean things
up a bit to avoid some of the potential issues this code had.
Always stop paging regardless of where the subscriber is at the
time of the cancellation.
As the comment says we should not rely that the paging
occurs on the current LAC. We might page at more BTS.
Walk all the BTS to stop paging. No callbacks will be
issued by this stop operation.
In case we can't page on a BTS then stop it everywhere. The
callers of paging_request assume that this is kind of an
atomic operation and we should help with that.
Coordinate with the normal subscriber channel requests instead
of going to page ourselves. This might lead to getting a channel
that is of a different type though.
vty_interface_layer3.c:584:4: warning: format '%d' expects argument of type 'int', but argument 3 has type 'long unsigned int' [-Wformat=]
sizeof(subscr->extension)-1, VTY_NEWLINE);
When two phones use the same TMSI and no A3A8 is possible
we could end in a hard to detect issue. Assume that the IMSI
will not clash and we have issues for foreign simcards.
In case foreign simcards are used we can not do authentication
and ciphering. In case a TMSI is re-used too early and we do
page using TMSI we can't know which of the two MS is responding
to us. We could change the "secure channel" routine to ask for
the IMSI and only then stop the paging.
As we don't have ciphering there is not much use in using the
TMSI. Add a mode "no assign-tmsi" that will not assign the TMSI
during LU. Now CM Service Request and Paging Response will
work using the IMSI. There can't be a clash with that.
[ciaby fixed the vty write to use the right name]
When we can't find the TMSI then the subscriber is not in our
VLR. We have not consulted with the HLR and it is better to not
use such a severe error code.
in_address is not 'accidently' included by FreeBSD when we include
the osmocom/core/select.h header file. We need to include a bit
more.
In file included from mgcp_protocol.c:38:
../../include/openbsc/mgcp_internal.h:134:21: error: field has incomplete type 'struct sockaddr_in'
struct sockaddr_in forward;
List needs to be executed from within the right configuration
node to see if it is available or not. list on the toplevel
will uncoditionally show "smpp" as part of the logging config.
Struct osmo_msc_data contains int core_ncc, which is actually the
MNC part of the PLMN, not to be confused with the Network Colour
Code.
The following patch renames this field for clarity and consistency
with the standards.
If we have tried SMPP first and it was not routable, and then
tried the local delivery there is no point in trying SMPP with
the same parameters again. Leave early and return unknown sub
to the caller.
default-route would only be looked at after there has been
no subscriber in the local database. Depending on the setup
this is not what one wants. This has been discussed at the
OsmoDevCon and there have been hacks in some branches. Let's
introduce a VTY command to select if SMPP should be consulted
first and then fallback to the current behavior.
Even if it is using BSC/NITB types let's put it in the header
file than just declaring it at a place that could bitrot in a
way that doesn't lead a warning.
The "default-route" for SMPP will be used after a local
subscriber look-up. Sometimes we want to route everything
to SMPP. Make this possible by changing this routine.
We just link to libosmovty and if it requires crypt internally it
needs to link to that (and not us). This looks like a left-over
from when we moved the VTY code out of OpenBSC
We don't need to consume all the entropy of the kernel but can
use libcrypto (OpenSSL) to generate random data. It is not clear
if we need to call RAND_load_file but I think we can assume that
our Unices have a /dev/urandom.
This takes less CPU time, provides good enough entropy (in theory)
and leaves some in the kernel entropy pool.
We are using the token to find the right bsc_config and
then we can use the last_rand of the bsc_connection to
calculate the expected result and try to compare it with
a time constant(???) memcmp.
Check if the NAT has sent 16 bytes of RAND and if a key
has been configured in the system and then generate a
result using milenage. The milenage res will be sent and
noth the four byte GSM SRES derivation.
Generate 16 byte of random data to be used for A3A8 by
the BSC in the response. We can't know which BSC it is
at this point and I don't want to send another message
once the token has been received so always send the data
with an undefined code. The old BSCs don't parse the
message and will happily ignore the RAND.
/dev/urandom can give short reads on Linux so loop
around it until the bytes have been read from the kernel.
Instead of doing open/read/close all the time, open the
FD in the beginning and keep it open. To scare me even
more I have seen /dev/urandom actually providing a short
read and then blocking but it seems to be the best way
to get the random byes we need for authentication.
So one should/could run the cheap random generator on
the system (e.g. haveged) or deal with the NAT process
to block.
Unfortunately the basic structure of the response is broken.
There is a two byte length followed by data. The concept of
a 'tag' happens to be the first byte of the data.
This means we want to write strlen of the token, then we
want to write the NUL and then we need to account for the
tag in front.
Introduce a flag if the new or old format should be used.
This will allow to have new BSCs talk to old NATs without
an additional change. In the long run we can clean that up.
In case the token was not correct, just close the connection.
It is not clear that forcing a new TCP connection is going to
give us any extra security here. But with the upcoming auth
handling it does make sense to have both case look similar.
In the libfilter source code, which is built regardless of --enable-nat,
headers from libosmo-sccp were used, thus causing a build failure (see
below) when building without --enable-nat, and libosmo-sccp not being
installed (or being installed in a prefix not otherwise included in the
build).
The build fails like this:
In file included from ../../../src/libfilter/bsc_msg_filter.c:27:0:
../../../include/openbsc/bsc_nat_sccp.h:27:37: fatal error: osmocom/sccp/sccp_types.h: No such file or directory
As the includes seem not to be actually needed, this change fixes the
issue by just omitting them.
Running "make distcheck" failed trying to generate ".version" into the
read-only unpacked source directory. Actually shipping ".version" in the
tarball fixes that.
There was no context for the SCCP CREF message and this means
that the msc_con was a plain NULL pointer that was dereferenced
and the application would crash.
Use the new API to pass the incoming MSC Connection which sould
be used for the SCCP CREF message as context. The code has not
been fed with an actual SCCP CR message.
The loop was used to print all returned addresses but we can
simply pick the first one. This is fixing a coverity issue that
the loop will be executed eaxactly once (and that was on
purpose).
Simplify the code and just take the first element (which might
be NULL).
Fixes: Coverity CID#1302852
We can't do much in case the fd is failing to be registered.
There should be a timeout that is catching this and it might
be able to repair it self.
Fixes: Coverity CID#1302854
The code to do that doesn't belong to the control interface, so
abstract it out to a separate function gsm_bts_set_system_infos().
[hfreyther: Fix the coding style...]
In case the query for "hostname" will fail c-ares will append the
domain name of /etc/resolv.conf and query again. We don't want that
so claim we provide a list of domain names and then don't provide
any.
I didn't intend to have pushed the c-ares code to master yet.
Holger Hans Peter Freyther