When the VLR requests a Ciphering Mode with vlr_ops.set_ciph_mode(), and if we
need a ciph algo flag from a Classmark information that is not yet known
(usually CM 2 during LU), send a BSSMAP Classmark Request to get it.
To manage the intermission of the Classmark Request, add
- msc_classmark_request_then_cipher_mode_cmd(),
- state SUBSCR_CONN_S_WAIT_CLASSMARK_UPDATE,
- event SUBSCR_CONN_E_CLASSMARK_UPDATE.
From state AUTH_CIPH, switch to state WAIT_CLASSMARK_UPDATE. Once the BSSMAP
Classmark Response, is received, switch back to SUBSCR_CONN_S_AUTH_CIPH and
re-initiate Ciphering Mode.
To be able to re-enter the Ciphering Mode algo decision, factor it out into
msc_geran_set_cipher_mode().
Rationale:
In the following commit, essentially we stopped supporting A5/3 ciphering:
commit 71330720b6
"MSC: Intersect configured A5 algorithms with MS-supported ones"
Change-Id: Id124923ee52a357cb7d3e04d33f585214774f3a3
A5/3 was no longer supported because from that commit on, we strictly checked
the MS-supported ciphers, but we did not have Classmark 2 available during
Location Updating.
This patch changes that: when Classmark 2 is missing, actively request it by a
BSSMAP Classmark Request; continue Ciphering only after the Response. Always
request missing Classmark, even if a lesser cipher were configured available.
If the Classmark Update response fails to come in, cause an attach failure.
Instead, we could attempt to use a lesser cipher that is also enabled. That is
left as a future feature, should that become relevant. I think it's unlikely.
Technically, we could now end up requesting a Classmark Updating both during LU
(vlr_lu_fsm) and CM Service/Paging Response (proc_arq_fsm), but in practice the
only time we lack a Classmark is: during Location Updating with A5/3 enabled.
A5/1 support is indicated in CM1 which is always available, and A5/3 support is
indicated in CM2, which is always available during CM Service Request as well
as Paging Response. So this patch has practical relevance only for Location
Updating. For networks that permit only A5/3, this patch fixes Location
Updating. For networks that support A5/3 and A5/1, so far we always used A5/1
during LU, and after this patch we request CM2 and likely use A5/3 instead.
In msc_vlr_test_gsm_ciph, verify that requesting Classmark 2 for A5/3 works
during LU. Also verify that the lack of a Classmark Response results in attach
failure.
In msc_vlr_test_gsm_ciph, a hacky unit test fakes a situation where a CM2 is
missing during proc_arq_fsm and proves that that code path works, even though
the practical relevance is currently zero. It would only become interesting if
ciphering algorithms A5/4 and higher became relevant, because support of those
would be indicated in Classmark 3, which would always require a Classmark
Request.
Related: OS#3043
Depends: I4a2e1d3923e33912579c4180aa1ff8e8f5abb7e7 (libosmocore)
Change-Id: I73c7cb6a86624695bd9c0f59abb72e2fdc655131
The intention was to use the file's basename, but __BASE_FILE__ means "the root
file that is being parsed and contains #include statements".
If we had a function using __BASE_FILE__ and that was defined in an #included
file, __BASE_FILE__ would indicate the first file where the #include is, and
not the file where the function is defined. __BASE_FILE__ works for us because
we don't ever include function definitions that log something, so __BASE_FILE__
always coincides with __FILE__ for our logging; but still __BASE_FILE__ is
semantically the wrong constant.
Related: OS#2740
Change-Id: I1c8122c909938daaf782468c1c5b0262d555c3ce
Previously the MSC_CONN_USE token for Supplementary Services was
called 'MSC_CONN_USE_TRANS_USSD'. Non-call related Supplementary
Services is not only about USSD, so let's rename it.
Change-Id: I5b3517c87a32fa64dea6b0c912f2b76c5c25a112
When sending a BSSMAP Clear or Iu Release, do not immediately discard the conn,
but wait until a BSSMAP Clear Complete / Iu Release Complete has been received.
Hence we will no longer show in the log that an incoming Release/Clear Complete
belongs to an unknown subscriber, but will still be around to properly log the
release.
Related: OS#3122
Change-Id: Ie4c6aaba3866d6e5b98004e8870a215e8cf8ffc1
Refactor:
1. Glue the gsm_subscriber_connection alloc to the subscr_conn_fsm.
2. Add separate AUTH_CIPH state to the FSM.
3. Use conn->use_count to trigger conn release.
4. Add separate RELEASING state to the FSM.
5. Add rate counters for each of the three Complete Layer 3 types.
Details:
1. Glue the gsm_subscriber_connection alloc to the subscr_conn_fsm.
Historically, a gsm_subscriber_connection was allocated in libbsc land, and
only upon Complete Layer 3 did libmsc add the fsm instance. After splitting
openbsc.git into a separate osmo-msc, this is no longer necessary, hence:
Closely tie gsm_subscriber_connection allocation to the subscr_conn_fsm
instance: talloc the conn as a child of the FSM instance, and discard the conn
as soon as the FSM terminates.
2. Add separate AUTH_CIPH state to the FSM.
Decoding the Complete Layer 3 message is distinctly separate from waiting for
the VLR FSMs to conclude. Use the NEW state as "we don't know if this is a
valid message yet", and the AUTH_CIPH state as "evaluating, don't release".
A profound effect of this: should we for any odd reason fail to leave the FSM's
NEW state, the conn will be released right at the end of msc_compl_l3(),
without needing to trigger release in each code path.
3. Use conn->use_count to trigger conn release.
Before, the FSM itself would hold a use count on the conn, and hence we would
need to ask it whether it is ready to release the conn yet by dispatching
events, to achieve a use_count decrement.
Instead, unite the FSM instance and conn, and do not hold a use count by the
FSM. Hence, trigger an FSM "UNUSED" event only when the use_count reaches zero.
As long as use counts are done correctly, the FSM will terminate correctly.
These exceptions:
- The new AUTH_CIPH state explicitly ignores UNUSED events, since we expect the
use count to reach zero while evaluating Authentication and Ciphering. (I
experimented with holding a use count by AUTH_CIPH onenter() and releasing by
onleave(), but the use count and thus the conn are released before the next
state can initiate transactions that would increment the use count again.
Same thing for the VLR FSMs holding a use count, they should be done before
we advance to the next state. The easiest is to simply expect zero use count
during the AUTH_CIPH state.)
- A CM Service Request means that even though the MSC would be through with all
it wants to do, we shall still wait for a request to follow from the MS.
Hence the FSM holds a use count on itself while a CM Service is pending.
- While waiting for a Release/Clear Complete, the FSM holds a use count on
itself.
4. Add separate RELEASING state to the FSM.
If we decide to release for other reasons than a use count reaching zero, we
still need to be able to wait for the msc_dtap() use count on the conn to
release.
(An upcoming patch will further use the RELEASING state to properly wait for
Clear Complete / Release Complete messages.)
5. Add rate counters for each of the three Complete Layer 3 types.
Besides LU, also count CM Service Request and Paging Response
acceptance/rejections. Without these counters, only very few of the auth+ciph
outcomes actually show in the counters.
Related: OS#3122
Change-Id: I55feb379e176a96a831e105b86202b17a0ffe889
The current msc_subscr_con_allocate() was in fact only used by msc_vlr_tests,
while both a_iface_bssap.c and iucs.c did their own duplicate code of
allocating the gsm_subscriber_connection struct. Unify.
Drop the old msc_subscr_con_allocate(), instead add msc_subscr_conn_alloc().
The new function also takes via_ran and lac arguments directly.
The conn allocation will soon be closely tied to the subscr_conn_fsm instance
allocation, so place the new function definition alongside the other
subscr_conn_fsm API, and match its naming ("conn").
Related: OS#3122
Change-Id: Ia57b42a149a43f9c370b1310e2e1f512183993ea
Instead of jumping through hoops to pass the Complete Layer 3 operation that
created this conn via FSM event dispatch parameters, put it right in the
gsm_subscriber_connection struct, where it always belonged.
Move definition of the enum complete_layer3_type to gsm_data.h, where
gsm_subscriber_connection is defined.
Introduce msc_subscr_conn_update_id() to set the complete_layer3_type of the
conn as soon as a Complete Layer 3 message is received.
In msc_subscr_conn_update_id(), already include an mi_string argument to
prepare for an upcoming patch where the FSM will be allocated much earlier when
the Mobile Identity is not known yet, and we'll also update the fi->id here.
The odd logging change in the msc_vlr_tests output uncovers a wrong use of the
osmo_fsm_inst_dispatch() data argument for SUBSCR_CONN_E_CN_CLOSE events: if a
child FSM signals unsuccessful result, instead of the failure cause, it passed
the complete_layer3_type, as requested upon FSM allocation, which was then
misinterpreted as a failure cause. Now a child FSM failure will pass NULL
instead, while other SUBSCR_CONN_E_CN_CLOSE events may still pass a valid cause
value.
Related: OS#3122
Change-Id: Iae30dd57a8861c4eaaf56999f872d4e635ba97fb
'subscr_conn_from' could mean anything: from what, RAN type? BSS identifier? MM
action? Clearly name it as the Complete Layer 3 kind it represents.
Related: OS#3122
Change-Id: I6263a80e6db01c2ca48df6c58b05e2fd19347057
In trans_free(), call subscr_conn_release_when_unused(), so that we are sure to
clean up after the last transaction is done.
This fixes an error where a conn lingered after a CC failure, because that code
path forgot to trigger cleanup.
Rationale: so far we were triggering the release check after each DTAP dispatch
(compl_l3 and "normal" DTAP), which is sufficient for properly closed
transactions. We also need a check for when a timeout clears an erratic trans.
Adjust test expectation of test_call_mo_to_unknown_timeout to show that the
error is now fixed.
msc_vlr_test_reject_concurrency now sees an additional release checking event
when the SMS transaction is done, which is expected and does not affect the
test otherwise.
Related: OS#2779
Change-Id: I46ff2e9b09b67e4e0d79cccf8c04936f17281fcb
The naming of "bump" was short and made sense to me at the time of writing, but
it is keeping pretty much everyone else at a distance, no-one intuitively gets
what it is supposed to mean.
Clarify by renaming to "release_when_unused".
Adjust test expectations.
Change-Id: I4dcc55f536f63b13a3da29fff1df5fe16751f83a
The ID will include the type of connection (GERAN_A, UTRAN_IU) followed
by the SCCP conn_id.
This can be used for the fsm instance ID before we know the IMSI.
Change-Id: I4b875772e3994ad3458ee60dbf880604486d9afd
When hunting a conn use count bug, it was very hard to figure out who's (not)
using the conn. To ease tracking down this bug and future bugs, explicitly name
what a conn is being reserved for, and track in a bit mask.
Show in the DREF logs what uses and un-uses a conn. See the test expectation
updates, which nicely show how that clarifies the state of the conn in the
logs.
On errors, log them, but don't fail hard: if one conn use/un-use fails, we
don't want to crash the entire MSC before we have to.
Change-Id: I259aa0eec41efebb4c8221275219433eafaa549b
Philipp Maier
Neels Hofmeyr
Daniel Willmann
Harald Welte