Provide a common implementation of use counting that supports naming each user
as well as counting more than just one use per user, depending on the rules the
caller implies.
In osmo-msc, we were originally using a simple int counter to see whether a
connection is still in use or should be discarded. For clarity, we later added
names to each user in the form of a bitmask of flags, to figure out exactly
which users are still active: for logging and to debug double get / double put
bugs. This however is still not adequate, since there may be more than one CM
Service Request pending. Also, it is a specialized implementation that is not
re-usable.
With this generalized implementation, we can:
- fix the problem of inadequate counting of multiple concurrent CM Service
Requests (more than one use count per user category),
- directly use arbitrary names for uses like __func__ or "foo" (no need to
define enums and value_string[]s),
- re-use the same code for e.g. vlr_subscr and get fairly detailed VLR
susbscriber usage logging for free.
Change-Id: Ife31e6798b4e728a23913179e346552a7dd338c0
For handling RTP IP addresses and ports, osmo-mgw, osmo-bsc and osmo-msc
so far have their own separate shims and code duplication around
inet_ntoa(), htons(), sockaddr conversions etc. Unify and standardize
with this common API.
In the MGW endpoint FSM that was introduced in osmo-bsc and which I
would like to re-use for osmo-msc (upcoming patch moving that to
osmo-mgw), it has turned out that using char* IP address and uint16_t
port number types are a convenient common denominator for logging,
MGCP message composition and GSM48. Ongoing osmo-msc work also uses this
for MNCC.
This is of course potentially useful for any other IP+port combinations
besides RTP stream handling.
Needless to say that most current implementations will probably stay
with their current own conversion code for a long time; for current
osmo-{bsc,msc,mgw} work (MGW endpoint FSM) though, I would like to move
to this API here.
Change-Id: Id617265337f09dfb6ddfe111ef5e578cd3dc9f63
This ensures that the rpath of the generated binaries is set to
use only the just-compiled libosmo{core,gsm,vty}.so and not any
system-wide installed libraries while avoiding the ugly shell script
wrapper.
Change-Id: I9b9ae0ed277ba71519661a66a70b7f86971e4511
Run INT_MAX and ULONG_MAX related tests only manually, remove from automatic
testing. This will hopefully fix recent build failures on various platforms.
Add a 64 bit output example for expected results when invoking
`./tdef_test range'. This is not checked automatically and merely serves for
manual reference.
For vty tests, use 32bit max values instead of INT_MAX and ULONG_MAX.
Change-Id: I6242243bde1d7ddebb858512a1f0b07f4ec3e5c2
Move T_def from osmo-bsc to libosmocore as osmo_tdef. Adjust naming to be more
consistent. Upgrade to first class API:
- add timer grouping
- add generic vty support
- add mising API doc
- add C test
- add VTY transcript tests, also as examples for using the API
From osmo_fsm_inst_state_chg() API doc, cross reference to osmo_tdef API.
The root reason for moving to libosmocore is that I want to use the
mgw_endpoint_fsm in osmo-msc for inter-MSC handover, and hence want to move the
FSM to libosmo-mgcp-client. This FSM uses the T_def from osmo-bsc. Though the
mgw_endpoint_fsm's use of T_def is minimal, I intend to use the osmo_tdef API
in osmo-msc (and probably elsewhere) as well. libosmocore is the most sensible
place for this.
osmo_tdef provides:
- a list of Tnnnn (GSM) timers with description, unit and default value.
- vty UI to allow users to configure non-default timeouts.
- API to tie T timers to osmo_fsm states and set them on state transitions.
- a few standard units (minute, second, millisecond) as well as a custom unit
(which relies on the timer's human readable description to indicate the
meaning of the value).
- conversion for standard units: for example, some GSM timers are defined in
minutes, while our FSM definitions need timeouts in seconds. Conversion is
for convenience only and can be easily avoided via the custom unit.
By keeping separate osmo_tdef arrays, several groups of timers can be kept
separately. The VTY tests in tests/tdef/ showcase different schemes:
- tests/vty/tdef_vty_test_config_root.c:
Keep several timer definitions in separately named groups: showcase the
osmo_tdef_vty_groups*() API. Each timer group exists exactly once.
- tests/vty/tdef_vty_test_config_subnode.c:
Keep a single list of timers without separate grouping.
Put this list on a specific subnode below the CONFIG_NODE.
There could be several separate subnodes with timers like this, i.e.
continuing from this example, sets timers could be separated by placing
timers in specific config subnodes instead of using the global group name.
- tests/vty/tdef_vty_test_dynamic.c:
Dynamically allocate timer definitions per each new created object.
Thus there can be an arbitrary number of independent timer definitions, one
per allocated object.
T_def was introduced during the recent osmo-bsc refactoring for inter-BSC
handover, and has proven useful:
- without osmo_tdef, each invocation of osmo_fsm_inst_state_chg() needs to be
programmed with the right timeout value, for all code paths that invoke this
state change. It is a likely source of errors to get one of them wrong. By
defining a T timer exactly for an FSM state, the caller can merely invoke the
state change and trust on the original state definition to apply the correct
timeout.
- it is helpful to have a standardized config file UI to provide user
configurable timeouts, instead of inventing new VTY commands for each
separate application of T timer numbers.
Change-Id: Ibd6b1ed7f1bd6e1f2e0fde53352055a4468f23e5
I want to tweak general VTY features and need to cover with a transcript test
to show the differences. Start by showing the current situation of optional
and multi-choice arguments.
Change-Id: I5a79c83fabd02aba6406b6e0d620969c4bd0cc1d
It was introduced and forgotten to add to EXTRA_DIST in:
"logging vty: add VTY transcript test"
commit 3a9ff11e57
change-Id I948e832a33131f8eab98651d6010ceb0ccbc9a9c
Change-Id: I1bcedf3097f02b2adc679560d1cbceb27dbc345e
Add functions to encode and decode Global Call Reference as per
3GPP TS 29.205 Table B 2.1.9.1 add corresponding tests.
Change-Id: Iee95aa4e5c056645b6cb5667e4a067097d52dfbf
Related: OS#2487
I am setting out to refactor various details about logging. To show the effect,
I am first adding this new test to illustrate the exact effects on the various
osmo programs.
Add logging_vty_test.c as a standalone program that simply defines a few
logging categories and opens a telnet vty to play with.
Add logging_vty_test.vty, as an osmo_verify_transcript_vty.py test script.
Add --enable-external-tests to configure.ac, to enable running
logging_vty_test.vty during 'make check'.
Also allow running 'make vty-test' without the need to first configure with
--enable-external-tests (a flexibility I've missed many times over in the other
osmo source trees).
Add a Makefile.am stub for external CTRL tests, basically a copy-paste from
osmo-msc.git. I doubt that libosmocore will get python driven CTRL interface
testing any time soon, but if so we will know to not run it concurrently.
Change-Id: I948e832a33131f8eab98651d6010ceb0ccbc9a9c
Do not link against the system-wide installed libosmo* libs when building the
regression test programs. Always use the locally built ones.
Linking some libosmo libraries causes libtool to pull in other libosmo libs
even though they were not explicitly named. For example, ctrl_test explicitly
links libosmoctrl, but this also has dependencies to libosmovty and libosmogsm:
ldd src/ctrl/.libs/libosmoctrl.so | grep osmo
libosmocore.so.11 => /usr/local/lib/libosmocore.so.11 (0x00007f26c26d4000)
libosmogsm.so.10 => /usr/local/lib/libosmogsm.so.10 (0x00007f26c22bb000)
libosmovty.so.4 => /usr/local/lib/libosmovty.so.4 (0x00007f26c2171000)
If we omit explicit LDADD of these dependencies in the Makefile.am, libtool
will take the first canonical place to find them, which may just be the already
installed older versions of the same libs, which may or may not be compatible
with the current build. In any case, it is never intended to link installed
libs.
All library dependencies are listed by this quick script:
cd libosmocore
for l in $(find . -name "*.so") ; do echo; echo "$l"; ldd $l | grep libosmo; done
./.libs/libosmocore.so
./coding/.libs/libosmocoding.so
libosmocore.so.11 => /usr/local/lib/libosmocore.so.11 (0x00007f25fc3c2000)
libosmogsm.so.10 => /usr/local/lib/libosmogsm.so.10 (0x00007f25fbfa9000)
libosmocodec.so.0 => /usr/local/lib/libosmocodec.so.0 (0x00007f25fbf9b000)
./codec/.libs/libosmocodec.so
libosmocore.so.11 => /usr/local/lib/libosmocore.so.11 (0x00007fb4c900d000)
./ctrl/.libs/libosmoctrl.so
libosmocore.so.11 => /usr/local/lib/libosmocore.so.11 (0x00007f5df5129000)
libosmogsm.so.10 => /usr/local/lib/libosmogsm.so.10 (0x00007f5df4d10000)
libosmovty.so.4 => /usr/local/lib/libosmovty.so.4 (0x00007f5df4bc6000)
./gb/.libs/libosmogb.so
libosmocore.so.11 => /usr/local/lib/libosmocore.so.11 (0x00007f788e536000)
libosmovty.so.4 => /usr/local/lib/libosmovty.so.4 (0x00007f788e3ec000)
libosmogsm.so.10 => /usr/local/lib/libosmogsm.so.10 (0x00007f788dfd3000)
./vty/.libs/libosmovty.so
libosmocore.so.11 => /usr/local/lib/libosmocore.so.11 (0x00007f3b7ed21000)
./gsm/.libs/libosmogsm.so
libosmocore.so.11 => /usr/local/lib/libosmocore.so.11 (0x00007fc69472e000)
./sim/.libs/libosmosim.so
libosmocore.so.11 => /usr/local/lib/libosmocore.so.11 (0x00007f2f6412d000)
libosmogsm.so.10 => /usr/local/lib/libosmogsm.so.10 (0x00007f2f63d14000)
Add all explicit linking of all required library dependencies in all regression
test programs, as shown by above listing.
Example for reproducing a problem:
In libosmocore.a, introduce a new function, and call that from libosmovty code.
For example, I made loglevel_strs non-static in logging.c, and used that in
logging_vty.c. Build and install this in a place where libtool can find it.
Then go back to before this change and rebuild. You will see that linking
ctrl_test (before this patch) then complains about libosmovty requiring the
loglevel_strs symbol which it cannot find in libosmocore.so.
Change-Id: Id084e6e6efd25cd62b1bd7a4fc7c5985c39130c6
As oap_client has moved from osmo-sgsn to libosmogsm, it is only fair
that the related unit test shall also be moved here.
Change-Id: I9d64e10b4bacac9b530cf077841bad762fc6d558
Recent OS#3407 shows that we should verify stderr to catch sanitizer failures.
(They might not always be ignorable like that one.)
Change-Id: Ic9e437a1cc96ae081e0fd6a9b6e3156987e14c0c
The conv_gen.py utility was tested against both Python 2 and 3,
so there is no need to enforce Python 2. Also, having:
#!/usr/local/bin/python{2|3}
is a bad idea, because Python may be installed in a different location.
Change-Id: I6007d481047b584db13d6eda70fb99f11f9ddaa1
Sometimes the library probiding dlopen is not the same one providing
dlsym.
This is the case when compiling with AddressSanitizer enabled. In this
case, AC_SEARCH_LIBS([dlopen]...) reports no lib is required, but tests
using dlsym still require to link against -ldl.
Change-Id: Ic619b0885688066b60c97caf1e2c7e5402c1d9f7
The sercomm functions are unavailable in case of embedded build. Add
stub and link the tests against it.
Change-Id: I9bc5cb2f822b1a3ffdc6ec29f46b6bac8288314e
As of 67bdd80a96 the stats.c is
effectively disable so we should disable the corresponding tests as
well.
Change-Id: I42ff7a6619c0a5926fdc2ec779cf04689c567e15
When a bad GSM voice frame is received, it's being replaced
by a silence frame. This may cause unpleasant audio effects.
This change implements a functionality to craft a replacement
frame from the last known good frame. Currently, only FR is
supported, support for other codecs may be added latter.
Change-Id: I06a21f60db01bfe1c2b838f93866fad1d53fdcd1
In case of embedded build some tests are failing to link properly. Fix
it:
* do not run fsm_test unless CTRL is enabled
* do not run fr_test unless GB is enabled
* do not link loggingrb_test with libosmovty
Change-Id: Icedad5ba3ed311ccdb97fa3ccd3002f5fda8be68
* remove duplicate code: use function from libosmocore
* use utility function to dump ubits
* reformat for easier reading
* link against libosmocore
Change-Id: I8c31b0954176a2c53305936a025c92a793b6d9b6
The testsuite fails on some specific build machines in the OBS
build cluster. Let's try to figure out which CPU flags they have
to narrow down the cause of this.
Change-Id: Ib23e5bfb3c894206fad62d6cc6151583b1bb75a6
According to
https://www.gnu.org/software/automake/manual/automake.html#Libtool-Flags
the libraries supposed to be added to *_LDADD or *_LIBADD
while *_LDFLAGS should contain additional libtool linking
flags. Previously we used both. Let's unify this and move all the
libraries into proper automake variable. While at it - also add
libosmocore.la for tests to LDADD since all the tests link against it
anyway.
Change-Id: Ia657a66db75df831421af5df1175a992da5ba80f
This function is actively used by OsmoPCU but have not been covered by
tests so far. The test code is based on
Minh-Quang Nguyen <minh-quang.nguyen@nutaq.com> submission with some
modifications.
The test's FIXME will be addressed in follow-up patches.
Change-Id: I2ee544256b8675bc62a42493aab66a8eeee54f90
Related: OS#1526
Add GSM23003_IMSI_MIN_DIGITS definition.
Add regression test gsm23003_test.c to test the two new functions.
Will be used by OsmoHLR to validate VTY and CTRL input.
Change-Id: I1e94f5b0717b947d2a7a7d36bacdf04a75cb3522
The recent exit-by-indent patch breaks a VTY case where a node is entered but
directly followed by a sibling or ancestor without listing any child nodes.
Regression introduced by I24cbb3f6de111f2d31110c3c484c066f1153aac9.
An example is a common usage in osmo-bts, where 'phy N' / 'instance N' is a
parent node that is commonly left empty:
phy 0
instance 0
bts 0
band 1800
Before this patch, this case produces the error:
There is no such command.
Error occurred during reading the below line:
bts 0
Fix indentation parsing logic in command.c to accomodate this case.
Add a unit test for empty parent node.
Change-Id: Ia0880a17ae55accb092ae8585cc3a1bec9986891
Note: This will break users' config files if they do not use consistent
indenting. (see below for a definition of "consistent".)
When reading VTY commands from a file, use indenting as means to implicitly
exit child nodes. Do not look for commands in the parent node implicitly.
The VTY so far implies 'exit' commands if a VTY line cannot be parsed on the
current node, but succeeds on the parent node. That is the mechanism by which
our VTY config files do not need 'exit' at the end of each child node.
We've hit problems with this in the following scenarios, which will show
improved user experience after this patch:
*) When both a parent and its child node have commands with identical names:
cs7 instace 0
point-code 1.2.3
sccp-address osmo-msc
point-code 0.0.1
If I put the parent's command below the child, it is still interpreted in the
context of the child node:
cs7 instace 0
sccp-address osmo-msc
point-code 0.0.1
point-code 1.2.3
Though the indenting lets me assume I am setting the cs7 instance's global PC
to 1.2.3, I'm actually overwriting osmo-msc's PC with 1.2.3 and discarding the
0.0.1.
*) When a software change moves a VTY command from a child to a parent. Say
'timezone' moved from 'bts' to 'network' level:
network
timezone 1 2
Say a user still has an old config file with 'timezone' on the child level:
network
bts 0
timezone 1 2
trx 0
The user would expect an error message that 'timezone' is invalid on the 'bts'
level. Instead, the VTY finds the parent node's 'timezone', steps out of 'bts'
to the 'network' level, and instead says that the 'trx' command does not exist.
Format:
Consistent means that two adjacent indenting lines have the exact
same indenting characters for the common length:
Weird mix if you ask me, but correct and consistent:
ROOT
<space>PARENT
<space><tab><space>CHILD
<space><tab><space><tab><tab>GRANDCHILD
<space><tab><space><tab><tab>GRANDCHILD2
<space>SIBLING
Inconsistent:
ROOT
<space>PARENT
<tab><space>CHILD
<space><space><tab>GRANDCHILD
<space><tab><tab>GRANDCHILD2
<tab>SIBLING
Also, when going back to a parent level, the exact same indenting must be used
as before in that node:
Incorrect:
ROOT
<tab>PARENT
<tab><tab><tab>CHILD
<tab><tab>SIBLING
As not really intended side effect, it is also permitted to indent the entire
file starting from the root level. We could guard against it but there's no
harm:
Correct and consistent:
<tab>ROOT
<tab><tab>PARENT
<tab><tab><tab><tab>CHILD
<tab><tab>SIBLING
Implementation:
Track parent nodes state: whenever a command enters a child node, push a parent
node onto an llist to remember the exact indentation characters used for that
level.
As soon as the first line on a child node is parsed, remember this new
indentation (which must have a longer strlen() than its parent level) to apply
to all remaining child siblings and grandchildren.
If the amount of spaces that indent a following VTY command are less than this
expected indentation, call vty_go_parent() until it matches up.
At any level, if the common length of indentation characters mismatch, abort
parsing in error.
Transitions to child node are spread across VTY implementations and are hard to
change. But transitions to the parent node are all handled by vty_go_parent().
By popping a parent from the list of parents in vty_go_parent(), we can also
detect that a command has changed the node without changing the parent, hence
it must have stepped into a child node, and we can push a parent frame.
The behavior on the interactive telnet VTY remains unchanged.
Change-Id: I24cbb3f6de111f2d31110c3c484c066f1153aac9
These PRBS sequences are specified in ITU-T O.150. They are typically
used as test data to be transmitted for BER (bit error rate) testing.
Change-Id: I227b6a6e86a251460ecb816afa9a7439d5fb94d1
We need to have an architecture-independend way of endian conversion /
byte swapping functions which will also work on embedded (bare iron)
builds. Let's introduce osmocom/core/bytesawp.h for this purpose.
Change-Id: Ibc0cc1e36d4ed63a35cf8ceff3af0f26e5ac7a3d
This change extends the convolutional code test coverage, adding
the GSM 05.03 specific test vectors, generated by the conv_gen.py.
Inspired by Tom's patch:
http://lists.osmocom.org/pipermail/openbsc/2014-April/007364.html
Change-Id: I76d1cd4032d2f74c5bb93bde4fab99aa655b7f1a
* data structure representing 3GPP TS 52.021 §9.4.62 SW Description
* function to serialize it into msgb
* function to deserialize it from buffer
* functions to extract/estimate buffer size for SW Description
* test harness (partially taken from OpenBSC)
There are several similar functions to deal with SW Description in
OpenBSC, there's also need to use similar functionality in
OsmoBTS. Hence it's better to put the code into common library with
proper tests and documentation.
Change-Id: Ib63b6b5e83b8914864fc7edd789f8958cdc993cd
Related: OS#1614
Sometimes (particularly when testing), we may want to parse+execute an
arbitrary control command simply form a string buffer, rather than from
a msgb. Let's add a helper for that.
Change-Id: Iaca748e0d942bb2a1ee7c2776b37485e1439eb0c
Add test for osmo-auc-gen invocations to ensure stability across upcoming SQN
increment scheme changes.
The test comprises of a shell script that invokes the osmo-auc-gen binary with
various milenage parameters, of which the stdout/stderr are verified.
More osmo-auc-gen invocations could be added, but my main focus is on the SEQ
changes. Instead of manually testing that it still works for each SQN patch, I
want this test to do it for me.
To make sure that osmo-auc-gen is build before the tests are launched, place
'utils' before 'tests' in the root Makefile.am.
Related: OS#1968
Change-Id: Ib4af34201cd2e7d76037bcd31dd89ef18c1a9aec
There are some projects, such as GR-GSM and OsmocomBB, which would
benefit from using one shared implementation of GSM 05.03 code. So,
this commit introduces a new sub-library called libosmocoding, which
(for now) provides GSM, GPRS and EDGE transcoding routines, migrated
from OsmoBTS.
The original GSM 05.03 code from OsmoBTS was relicensed under
GPLv2-or-later with permission of copyright holders (Andreas Eversberg,
Alexander Chemeris and Tom Tsou).
The following data types are currently supported:
- xCCH
- PDTCH (CS 1-4 and MCS 1-9)
- TCH/FR
- TCH/HR
- TCH/AFS
- RCH/AHS
- RACH
- SCH
Change-Id: I0c3256b87686d878e4e716d12393cad5924fdfa1
To be able to add some more tests, related to convolutional coding,
without duplication of code, the test logic was separated from the
conv_test.c into conv.c and conv.h.
Change-Id: Idbdc7e19cb9b9a36cd1fccd621cd858e87530d98
Use value_string for enum ctrl_type instead of custom code. Add
corresponding unit tests.
Related: OS#1615
Change-Id: Icd4e96dd9f00876cb70b43cfcf42ab4f10311b28
Configure logging to be deterministic and add stderr checking to testuite.at.
However, exclude the thousands of message modification log lines from the log
to not have a huge test expectation file.
Change-Id: I0dd7112967a64a168556b62e5ec15107b7608ffb
The write queue was always meant to not queue more than the
max_length messages but the implementation never rejected a
message.
Begin to log and enforce the queue size limit, add a testcase
to verify the code and initialize except_cb as part of a fix
for that new test case.
Real applications might now run into the queue limit and drop
messages where they just queued them before. It is unfortunate
but I still think it is good to implement the routine as it was
intended. We need to review osmo_wqueue_enqueue once more to
see that no msgb is leaked.
Change-Id: I1e6aef30f3e73d4bcf2967bc49f0783aa65395ae
Numerous issues caused sim_test to be attempted even though libosmosim was not
built:
In configure.ac, the ENABLE_PCSC variable lacked an AC_SUBST() to be exported.
Furthermore in configure.ac, no value 'yes'/'no' was assigned to the
ENABLE_PCSC variable, only to the enable_pcsc value.
In testsuite.at, encapsulating the sim_test in 'if ENABLE_PCSC' seems to have
no effect, regardless (not even when using a variable that should be defined
accurately).
So fix with these steps, similarly to how we do it in openbsc:
In AC_ARG_ENABLE, directly use 'ENABLE_PCSC' to assign 'yes'/'no'.
Export the same using AC_SUBST().
Add tests/atlocal.in to translate ENABLE_PCSC to enable_sim_test (also add
atlocal to AC_OUTPUT and distclean).
Use enable_sim_test in testuite.at, as seen in openbsc: use AT_CHECK() to
indicate skipping the test if enable_sim_test isn't 'yes'.
Change-Id: I9e8740c7d2dfbd272e22fee85972ef3fda7184a8
Add src/gsm/libosmogsm.la explicitly to some test linkages, because otherwise
the linker would pick the libosmogsm already installed on the system instead of
the one that was just built in the source tree.
I noticed because a libosmogsm needing some more symbols from libosmocodec was
still installed, while the patch that cause it was already removed. Thus I
caught all(?) test binaries that linked libosmogsm from $PREFIX.
Change-Id: Ie61d60e1506f16de20add70fd0f44ebfa7a00a75
Corresponding test code include both official test vectors from the
specs and data from over-the-air tests.
This obsoletes libosmo-crypt-a53 as it was last missing piece
unimplemented in libosmogsm.
Change-Id: I939e4f6b91b4a7c591ef3761fe2d46ed1c2fb2d3
Related: OS#1582
This code is supposed to formalize some of the state machine handling in
Osmocom code.
Change-Id: I0b0965a912598c1f6b84042a99fea9d522642466
Reviewed-on: https://gerrit.osmocom.org/163
Tested-by: Jenkins Builder
Reviewed-by: Harald Welte <laforge@gnumonks.org>
* add functions to encode/decode various codec paramters from RTP payload with
AMR frame according to RFC 4867
* those functions are extended version based on code from osmo-bts'
amr.c by Andreas Eversberg
* add corresponding enum types and strings for logging
* add regression tests
It's useful both to replace manual parsing in osmo-bts with fuctions
covered by test suite and as a debugging helpers for issues related to
AMR.
Change-Id: Ia217679a07d3fbc970f435e20f6eac33d34bd597
Related: OS#1562
Reviewed-on: https://gerrit.osmocom.org/118
Tested-by: Jenkins Builder
Reviewed-by: Holger Freyther <holger@freyther.de>
Move those routines from OpenBSC to libosmogsm, so they can be
re-used from other programs. I think it was a mistake to add them only
inside the openbsc repository in the first place. We need to pay more
attention to this in the future.
Add python utility to generate .c code with convolutional
encoder/decoder based on polynomial description of the code. If argument
given it'll be interpreted as intended output directory, otherwise
current working directory is used.
Codes for *CCH, CS2/3 and TCH/AFS are generated. Corresponding manual
implementations are removed from tests. This introduce build-time
dependency on python.
The main work for this patch was generously contributed by Sylvain
Munaut.
Fixes: OS#1629