Add missing REQUIREMENT_A_TCHF to form a full REQUIREMENT_TCHF_MASK.
That allows detecting TCH/F -> TCH/H upgrades also in the requirement A
flags, where we look for any viable lchan.
Prepares for upcoming patch Id40d1cf8b58410c7d4eb87407fe8b8106e352438
which implements TCH/H to TCH/F upgrades within the same cell.
Related: SYS#5198 SYS#5365
Change-Id: Ic44615b314782423bab0ceef5810311776f92754
Both VAMOS- and non-VAMOS speech modes should result in indentical voice
handling. So make sure that all chan_modes are converted to non-vamos
before comparing / evaluating in switch statements.
Change-Id: I791e7966b1f8eaa3299a8a46abeb313cf5136e0b
So far we have a couple of macros iterating a specific number of lchans,
depending on dynamic timeslot state etc. With addition of VAMOS lchans,
this would become more complex and bloated.
Instead of separate iteration macros for each situation, only have one
that takes a number of lchans as argument. That allows to more clearly
pick the number of lchans, especially for non-trivial VAMOS scenarios.
Related: SYS#5315 OS#4940
Change-Id: Ib2c6baf73a81ba371143ba5adc912aef6f79238d
So far we do all channel reassignments by Handover Command. Since
osmo-bsc now supports rassignment of ongoing voice calls, do intra-cell
congestion resolution by Assignment Command.
In effect, add support for expecting an Assignment Command in
handover_test, and expect assignments instead of handovers for
intra-cell congestion resolution test cases.
Related: SYS#5330 OS#3277
Change-Id: Id56a890106b93fcee67ac9401b890e7b63bba421
I noticed during testing that an lchan used as TCH/F in fact still had
its channel mode set to Signalling -- because on Assignment, the Speech
mode used to be placed in the *previous* lchan and the new lchan was
never updated after the Activ ACK. This is unbearable confusion which I
complained about numerous times, so far mostly for cosmetic reasons. But
implementing re-assignment properly actually requires this to be cleaned
up.
Keep all volatile chan mode settings in lchan->activate.* or
lchan->modify.*, and only update lchan->* members when an ACK has been
received for those settings. So a failed request keeps a sane state.
Make sure that those settings are in fact updated in the proper lchan,
upon an ACK, so that subsequent re-assignment or mode-modify know the
accurate lchan state.
Related are upcoming patches that sort out the AMR multirate
configuration in a similar fashion, see
Iebac2dc26412d877e5364f90d6f2ed7a7952351e
Ia7519d2fa9e7f0b61b222d27d077bde4660c40b9
Ie57f9d0e3912632903d9740291225bfd1634ed47.
Related: SYS#5315 OS#4940 OS#3787 OS#3833
Change-Id: Ie0da36124d73efc28a8809b63d7c96e2167fc412
The neighbor configuration storage is fundamentally broken: it requires
all local cells to be configured before being able to list them as
neighbors of each other. Upon config write-back, the neighbor config
however is placed back inline with the other config, and hence a
written-out neighbor config no longer works on program restart.
The cause of this problem is that the config is stored as explicit
pointers between local cells (struct gsm_bts), which of course requires
the pointer to exist before being able to reference it.
Instead, store the actual configuration that the user entered as-is,
without pointers or references to objects that need to be ready. Resolve
the neighbors every time a neighbor is needed.
Hence the user may enter any config at any place in the config file,
even non-working config (like a BTS number that doesn't exist), and the
relation to actual local or remote neighbor cells is made at runtime.
Abort program startup if the initial neighbor configuration contains
errors.
Related: OS#5018
Change-Id: I9ed992f8bfff888b3933733c0576f92d50f2625b
So far the list of penalty timers was stored for an opaque target
pointer. That was either a gsm_bts pointer for a local BTS, or a cell
identifier list pointer for a remote-BSS cell.
Reasons to refactor penalty timers:
- The cell identifier list pointer came from the neighbor configuration
storage, but the way cell neighbor config is stored will change in a
subsequent patch. There will be no more cell identifier lists there.
- Storing object pointers is inherently unsafe -- if an object gets
removed and another gets allocated, the penalty timer could
theoretically remain in force for an unrelated object.
Rather store penalty timers for specific Cell IDs. Since remote-BSS
neighbors can be requested by a cell identifier *list*, use a
gsm0808_cell_id_list2 as key in the list of penalty timers.
Fix handover_test.c: have different CI for each local BTS. So far it was
the same LAC+CI for all BTSes, which now would make the test fail,
because any penalty timer would appear to apply to all local cells.
Related: OS#5018
Change-Id: I72dd6226a6d69c3f653a3174c6f55bf4eecc6885
Originally, the lchan stored only the Timing Advance from the initial
channel request, hence it was called rqd_ta.
Since quite a while now, rqd_ta also stores the most recent Timing
Advance from each received Measurement Report. So rename to last_ta.
This is cosmetic preparation for an upcoming patch that clarifies
whether the Timing Advance is already known for Channel Activation.
Change-Id: I1049526a173819baeb4978db5bf018ba3f1006a0
When balancing congestion, not only look at TCH/F or TCH/H separately,
but also to take into account the effects on the other TCH kind from
using/freeing dynamic TS.
Related: OS#5298
Change-Id: I433df6f343650f9056b1bab926bc19ac1d867ad5
Add bool log argument to lchan_avail_by_type() and omit logging when
passed as false. From handover_decision_2.c, pass 'log' as false, from
all other callers pass true, i.e. for unchanged behavior.
Rationale:
Usually, we use lchan_avail_by_type() to select a new lchan to initiate
actual service. For that, it is interesting to see how osmo-bsc decides
which lchan will be used.
For handover decision 2, we since recently call lchan_avail_by_type()
for each and every handover candidate, to determine whether it will
occupy a dynamic timeslot or not (to know whether we would congest the
other TCH kind). So this happens for each permutation of source lchan
and target cell. That produces a lot of logging, out of proportion of
being useful to the maintainer.
Change-Id: Ia403f8fc853ca9ea9e81f7a7395df6b23845ebed
For handover algorithm 2, properly figure out what effects the target
cell will see for the *other* TCH kind when a handover would occupy a
dynamic timeslot.
Before this, only TCH/F or TCH/H would be regarded at a time. This
introduces detection of whether a dynamic timeslot would be occupied by
a handover, and how losing one unused dynamic timeslot affects the
congestion situation for the TCH kind that is not targeted by the
handover.
In other words, if a handover to TCH/F causes congestion in TCH/H
because of a dynamic timeslot becoming occupied, the handover will not
be performed. Before this, oscillation situations could occur.
A subsequent patch will do the same for congestion balancing.
Related: SYS#5297
Change-Id: I1536b60f03cb0aeb6ba14a72b518aec82fa572fe
For balancing load across congested cells and across congested TCH/*
kinds, instead of comparing the number of lchans above the configured
congestion threshold, compare the percent of lchans of overload.
In short, using a percentage prevents cells with less min-free-slots to
fill up 100% while neighbor cells still may have several free lchans
available.
An obvious example of why this is desirable is illustrated by
test_balance_congestion_by_percentage.ho_vty:
Cell A has min-free-slots 2, and has all slots occupied.
Cell B has min-free-slots 4, and has 2 slots remaining free.
If we count congested lchans as in current master: cell A has a
congestion count of 2: two more lchans in use than "allowed". If we move
one lchan over to cell B, it ends up with a congestion count of 3, which
is worse than 2. So when counting lchans, we decide that cell A should
remain full.
Instead, when comparing percentage of remaining lchans, we would see
that cell A is loaded 100% above congestion (2 of 2 remaining lchans in
use), but when moving one lchan to cell B, that would only be 75% loaded
above its treshold (3 of 4 remaining lchans in use). So a percentage
comparison would cause a handover to cell B.
Related: SYS#5259
Change-Id: I55234c6c99eb02ceee52be0d7388bea14304930f
When evenly distributing congestion across cells, count the number of
occupied lchans surpassing congestion, and not the overall number of
free lchans -- which disregards congestion thresholds.
Fix the bugs shown by
test_congestion_no_oscillation.ho_vty
test_balance_congestion_tchf_tchh.ho_vty
This implements a simple calculation for congestion load by counting
lchans in use above congestion. An improvement of this calculation
using percent follows in I55234c6c99eb02ceee52be0d7388bea14304930f.
Related: SYS#5259
Change-Id: Icb373dc6bfc9819446db5e96f71921781fe2026d
Store the number of free lchans and the min-free-slots settings in
ho_candidate, instead of figuring those out in various places, to make
it easier to read.
Prepare for upcoming patch which also requires these values to fix a
bug.
Change-Id: Ie6ca5af5e8d0ebb8deaaaa637e2728008ecba517
So far it is often confusing which cell a specific member refers to.
Clarify lchan, bts, ... to current.lchan, target.bts, ...
Also move the rxlev_{current,target} to {current,target}.rxlev.
Eliminate numerous local variables to make it easier to read which side
is being used (e.g. "c->target.bts" instead of just "bts").
No functional change.
Change-Id: I06898eb745a5be548df0b76fa760ce790cfab3ed
Fix flaws in picking a candidate for congestion resolution, shown in
recently added tests.
- For TCH/H->TCH/F upgrading, do not favor moving to a weaker neighbor
cell.
- When comparing dynamic timeslots on the same cell, favor a dynamic
timeslot that frees an entire dyn TS even though the target rxlev
differs.
Do not separate the passes for inter-cell and intra-cell candidates:
before, the inter-cell pass would already pick a candidate and start
handover, even though the subsequent intra-cell pass would have revealed
a better candidate. Join the intra-cell considerations into
pick_better_lchan_to_move().
The intra-cell pass was separate, because it would find the *weakest*
current rxlev, to give a TCH/H to TCH/F upgrade to the currently weakest
lchan.
Instead of the separate pass for weakest rxlev, in addition to the
target cell's rxlev, also consider the rxlev *change* in
pick_better_lchan_to_move(): For candidates that do not change the rxlev
(usually those that stay in the same cell) and that upgrade to TCH/F,
favor a candidate with weaker current rxlev.
Completely revisit the conditions in pick_better_lchan_to_move() to
yield the desired prioritization of candidate preferences.
In three handover tests, remove the "FAIL" comments and adjust to now
expect the actually desired behavior.
Related: SYS#5032
Change-Id: I2704899c85c35dfd4eba43468452483f40016ca2
Instead of passing the TCH/H -> TCH/F bias (AFS bias) in local
variables, rather store it in the ho_candidate struct next to the other
rxlev related values.
Add the AFS bias to the compared rxlev in pick_better_lchan_to_move().
Modify pick_better_lchan_to_move() to simpler semantics of returning
either a or b.
No functional change.
Change-Id: I73860abdf2a77270ca4851ad58c09767d1bb08f1
Store the rxlev of the current lchan and the target BTS in the
ho_candidate, to clarify the code.
No functional change, cosmetically prepare for
I2704899c85c35dfd4eba43468452483f40016ca2.
Change-Id: Ie6c165e17bb3c99eebc967a6bb02529db8bdfc98
In handover_decision_2.c, instead of repeating a similar loop four
times, put that loop in a function and parameterize it.
Prepare for a fix of two problems in handover decision 2, see
I2704899c85c35dfd4eba43468452483f40016ca2.
Change-Id: I7c32d08e490a88a7f044b0a71dc4b07d748dd572
When a used timeslot gets moved to another timeslot for load balancing, prefer
moving a dynamic timeslot, as illustrated by handover_test.c test 30.
Rationale: freeing up a dynamic timeslot is better for PDCH availability, as
well as for flexibility in timeslots. Test 30 shows that when freeing a static
TCH/F even though a dynamic one with identical ratings is in use, later
handovers to a TCH/H may become impossible, because no more dynamic timeslots
are available to switch to TCH/H. A freed dynamic timeslot allows congestion
resolution to continue and reduce more TCH/F to TCH/H.
The scope of this preference is per-TRX, where the RXLEV ratings used for
picking a target lchan are the same by definition. In other words, this never
overrules picking another lchan that has better RXLEV.
Among lchans on dynamic timeslots that could be moved, this code favors moving
later lchans; mainly because it makes for a simpler condition in the code.
Change-Id: Ic221b8d2687cdec0bf94410c84a4da43853f0900
There are four places deciding which of 2 lchans to move, depending on average
db ratings. Upcoming patches will enrich that decision for better handling of
dynamic timeslots, so have one common function for these to avoid code dup.
Change-Id: I745dc95cf564dd330295cecb4d64dccebf55163f
Instead of logging a hex value for the met requirements, fully expand the "ABC"
flags for both TCH/F and TCH/H.
From HO_CANDIDATE_FMT/_ARGS, split off into REQUIREMENTS_FMT/_ARGS and use that
when logging the chosen HO candidates.
Also change the RX level to dBm, to match general logging and reduce confusion
between rxlev number variants in the log.
Change-Id: I1b30a6e98bdb4bd92e72864fafdd2f4f3ae3134c
When check_requirements() returns zero, do not keep such an entry in the
candidates list at all. This removes logging confusion, where some "candidates"
are still listed even though not meeting any handover requirements.
Change-Id: I12e48292d5731cb601165c870b9570003bc488ec
Place all code related to the object into the related file.
Having all the data model in one file made sense in early stage of
development to make progress quickly, but nowadays it hurts more than
helps, due to constantly growing size and more and more bits being
added to the model, gaining in complexity.
Currently, having lots of different objects mixed up in gsm_data.h is a hole
of despair, where nobody can make any sense were to properly put new stuff
in, ending up with functions related to same object in different files
or with wrong prefixes, declarations of non-existing functions, etc.
because people cannot make up their mind on strict relation to objects
in the data model.
Splitting them in files really helps finding code operating on a
specific object and helping with logically splitting in the future.
Change-Id: I00c15f5285b5c1a0109279b7ab192d5467a04ece
Fix neighbor config to match OsmoBSC manual: implement the plan for neighbor
configuration that was so far only described in the manual without actually
being in operation.
This first allows re-using ARFCN+BSIC pairs in and across BSS.
So far the handover_start() code always looked for handover target cells across
*all* local cells, even if they were not listed as neighbors to a source cell.
Imply all cells as neighbors only as long as there are no explicit neighbors
configured. As soon as the first 'neighbor' line appears in a 'bts' config,
only the listed neighbors are regarded as handover target cells. (The
'neighbor-list' commands are not related to this, only the relatively new
'neighbor (bts|lac|cgi|...)' commands affect actual handover procedures.)
TTCN3 tests TC_ho_neighbor_config_1 thru _7 play through the various aspects of
neighbor configuration: both the legacy implicit all-cells-are-neighbors as
well as allowing only explicit neighbors by config.
Related: OS#4056
Related: osmo-ttcn3-hacks Ia4ba0e75abd3d45a3422b2525e5f938cdc5a04cc
Change-Id: I29bca59ab232eddc74e0d4698efb9c9992443983
Implement basic support for inter-BSC HO from handover_decision_2: do inter-BSC
handover only when rxlev / rxqual / ta drop below the minimum requirements.
I considered adding a vty config flag to disable/enable remote-BSS handover,
but to avoid inter-BSC HO the user can simply refrain from configuring
neighbors for a particular cell.
In collect_assignment_candidate(), it is important to clear out any new
candidate entry. Hence adopt the same pattern as below: first compose a new
(cleared) candidate, then write the entry into the list.
Related: OS#3638
Change-Id: Id78ac1b2016998a2931a23d62ec7a3f37bb764c6
a) Prepare for triggering handover for any candidate, remote or local.
b) drop redundant arguments.
Change-Id: I5ba8b556703010c8e232b516285a837c999f87ef
Add FSMs:
- timeslot_fsm: handle dynamic timeslots and OML+RSL availability.
- lchan_fsm: handle an individual lchan activation, RTP stream and release,
signal the appropriate calling FSMs on success, failure, release.
- mgw_endpoint_fsm: handle one entire endpoint with several CI.
- assignment_fsm: BSSMAP Assignment Request.
- handover_fsm: all of intra, inter-MO and inter-MT handover.
Above FSMs absorb large parts of the gscon FSM. The gscon FSM was surpassing
the maximum amount events (32), and it is more logical to treat assignment,
handover and MGW procedures in separate FSMs.
- Add logging macros for each FSM type:
- LOG_TS()
- LOG_LCHAN()
- LOG_MGWEP(), LOG_CI()
- LOG_ASSIGNMENT()
- LOG_HO()
These log with the osmo_fsm_inst where present.
New style decision: logging without a final newline char is awkward,
especially for gsmtap logging and when other logs interleave LOGPC() calls;
we have various cases where the final \n goes missing, and also this invokes
the log category checking N times instead of once.
So I decided to make these macros *always* append a newline, but only if
there is no final newline yet. I hope that the compiler optimizes the
strlen() of the constant format strings away. Thus I can log with or without
typing "\n" and always get an \n termination anyway.
General:
- replace osmo_timers, state enums and program-wide osmo_signal_dispatch()
with dedicated FSM timeouts, states and events.
- introduce a common way to handle Tnnn timers: gsm_timers.h/.c: struct T_def.
These can be used (with some macro magic) to define a state's timeout once,
and not make mistakes for each osmo_fsm_inst_state_chg().
Details:
bsc_subscr_conn_fsm.c:
- move most states of this FSM to lchan_fsm, assignment_fsm, handover_fsm and
mgw_endpoint_fsm.
- There is exactly one state for an ongoing Assignment, with all details
handled in conn->assignment.fi. The state relies on the assignment_fsm's
timeout.
- There is one state for an ongoing Handover; except for an incoming Handover
from a remote BSS, the gscon remains in ST_INIT until the new lchan and conn
are both established.
- move bssmap_add_lcls_status() to osmo_bsc_lcls.c
abis_rsl.c:
- move all dynamic timeslot logic away into timeslot_fsm. Only keep plain send/receive functions in
abis_rsl.c
- reduce some rsl functions to merely send a message, rename to "_tx_".
- rsl_ipacc_mdcx(): add '_tx_' in the name; move parts that change the lchan state out into the
lchan_fsm, the lchan->abis_ip.* are now set there prior to invoking this function.
- move all timers and error/release handling away into various FSMs.
- tweak ipa_smod_s_for_lchan() and ipa_rtp_pt_for_lchan() to not require an
lchan passed, but just mode,type that they require. Rename to
ipacc_speech_mode*() and ipacc_payload_type().
- add rsl_forward_layer3_info, used for inter-BSC HO MO, to just send the RR
message received during BSSMAP Handover Command.
- move various logging to LOG_LCHAN() in order to log with the lchan FSM instance.
One drawback is that the lchan FSM is limited to one logging category, i.e. this moves some logging
from DRR to DRSL. It might actually make sense to combine those categories.
- lose LOGP...LOGPC logging cascades: they are bad for gsmtap logging and for performance.
- handle_classmark_chg(): change logging, move cm2 len check out of the cm3 condition (I hope that's
correct).
- gsm48_send_ho_cmd(): split off gsm48_make_ho_cmd() which doesn't send right away, so that during
inter-bsc HO we can make an RR Handover Command to send via the MSC to the remote BSS.
assignment_fsm.c:
- the Chan Mode Modify in case of re-using the same lchan is not implemented
yet, because this was also missing in the previous implementation (OS#3357).
osmo_bsc_api.c:
- simplify bsc_mr_config() and move to lchan_fsm.c, the only caller; rename to
lchan_mr_config(). (bsc_mr_config() used to copy the values to mr_bts_lv
twice, once by member assignment and then again with a memcpy.)
- During handover, we used to copy the MR config from the old lchan. Since we
may handover between FR and HR, rather set the MR Config anew every time, so
that FR rates are always available on FR lchans, and never on HR lchans.
Depends: I03ee7ce840ecfa0b6a33358e7385528aabd4873f (libosmocore),
I1f2918418c38918c5ac70acaa51a47adfca12b5e (libosmocore)
Change-Id: I82e3f918295daa83274a4cf803f046979f284366
Move all of libbsc/ into osmo-bsc/, and separate/move some implementations to
allow linking from utils/* and ipaccess/* without pulling in unccessary
dependencies.
Some utilities use gsm_network and gsm_bts structs, which already include data
structures for fairly advanced uses. Move initialization that only osmo-bsc
needs into new bsc_network_init() and bsc_bts_alloc_register() functions, so
that the leaner tools can use the old gsm_* versions without the need to link
everything (e.g. handover and lchan alloc code).
In some instances, there need to be stubs if to cut off linking "just before
the RSL level" and prevent dependencies from creeping in.
- abis_rsl_rcvmsg(): the only program currently interpreting RSL messages is
osmo-bsc, the utils are merely concerned with OML, if at all.
- paging_flush_bts(): ip.access nanobts models call this when the RSL link is
dropped. Only osmo-bsc actually needs to do anything there.
- on_gsm_ts_init(): the mechanism to trigger timeslot initialization is related
to OML, while this action to take on init would pull in RSL dependencies.
utils/ and ipaccess/ each have a stubs.c file to implement these stubs. Tests
implement stubs inline where required.
From src/utils/, src/ipaccess/ and tests/*/, link in .o files from osmo-bsc/.
In order for this to work, the osmo-bsc subdir must be built before the other
source trees. (An alternative would be to include the .c files as sources, but
that would re-compile them in every source tree. Not a large burden really, but
unless linking .o files gives problems, let's have the quicker build.)
Minor obvious cleanups creep in with this patch, I will not bother to name them
individually now unless code review asks me to.
Rationale:
1) libbsc has been separate to use it for osmo-nitb and osmo-bsc in the old
openbsc.git. This is no longer required, and spreading over libbsc and osmo-bsc
is distracting.
2) Recently, ridiculous linking requirements have made adding new functions
cumbersome, because libbsc has started depending on osmo-bsc/*.c
implementations: on gscon FSM and bssap functions. For example, neither
bs11_config nor ipaccess-config nor bts_test need handover_cfg or BSSMAP
message composition. It makes no sense to link the entire osmo-bsc to it, nor
do we want to keep adding stubs to each linking realm.
Change-Id: I36a586726f5818121abe54d25654819fc451d3bf
Neels Hofmeyr
Pau Espin
Martin Hauke