The timer is used for various timeouts and there is still external
client code that is calling it. In a perfect world the client code
would indicate that an event has happened and the internal timer
will be stopped. The best compromise is the "stop_t3191" method. It
allows to add semantic verification that the timer has been running.
Move the method into the PDCH. Extract the finding of TLLI into a
new class called Decoding. Move the assemble and forward LLC frames
into the TBF as it is poking in the internals of the TBF.
All dispatching will go through the PDCH. This will clean a lot
of the look-ups inside the gprs_rlcmac_data.c and continue with
adding structure to the pcu code.
Simplify the reset code now that the PDCH can know where it is
located. Rename the variables in the sba to trx_no and ts_no as
it stores the number and not the actual thing.
The code in gprs_rlcmac_data should ask the TBF for help in packing
the frames but it really shouldn't poke in the internals of the
tbf structure. This is very bad capsuling and has plenty of copy
and paste.
At the same thime this will be the most dangerous refactoring of
the code base.
The list belongs to the BTS. This makes cleaning this up more easy
and establishes a hierachy of resources that start from the BTS. The
debug_diagram code is now broken.
This is preparing the next commits and will ease working with
the debug_diagram when the global lists are killed. It will
simplify the APIin the long run.
The PollController is a friend of the SBAController and is allowed
to access the internal list. The list is hidden from everyone else.
This is done because the calculation of timeout should belong into
the PollController and not into the SBAController.
Only the gprs_rlcmac_pdch will manipulate the paging list now. There
can be various more refactorings of the code but they can be done
later. E.g. on memory allocation failure we can continue instead
of leaving the code, we should also set any_tbf only after things
have been paged.
Rely on packet_paging_request returning NULL in case the queue
is empty. We should move the write_packet_paging_request into
a separate file/object as well.
When a PDCH is disabled all resources should be freed. This is
currently not possible as the PDCH does not know where it belongs
to. On top of that the list (and other resources) should be
properly initialized on construction so that disable() is idempotent
and does not check if it was disabled. During the re-factoring I
noticed that during a sysmobts re-start some resources are not
freed. I left a warning in the code to resolve this issue later.
The current_frame is an attribute of the BTS. Move it from the
pcu_l1_if.cpp into the BTS. As the next step we can trigger
actions depending on the change of the frame.
Compared to the previous code there will be a branch to get the
global pointer so the code will be slightly slower than the previous
version but it allows us to start creating objects but still use
the code from C. It is best approach I have found so far.
One downside of C++ is that by default talloc will not be used
(unless we override the new operator to use talloc. Right now
we need to memset the C data structure by hand. The benefit of
enforcing a better structure should is more important though.
Remove the global state from gprs_rlcmac_trigger_downlink_assignment
and walk up to the pcu_l1_if.cpp where I find the timeout handling
that should be part of the SBA and TBF functionality. In terms of
hierachies things start to be more clear.
There should be the BTS object. That holds the SBA and TBF Controllers
that can allocate TBFs and SBAs and will handle the timeout polling
for a BTS.
Global state makes mocking/writing tests more difficult. Continue
by removing direct usage of the bts and adding it as pointer to
the function calls (showing the true dependencies of the functions)
There was no difference but there is no point in poking in the
internals of the tbf, pdch and trx from various places. Information
hiding has a real purpose, e.g. compare with reading "the mythical
man month".
Vladimir Rolbin pointed out that in case of the alloc_algorithm_a
the usf has not been assigned. For now move all the assignment into
a method and call it from the A and B algorithm.
There really shouldn't be too many callers of state. Instead the
tbf should dispatch depending on the internal state. For now
introduce state_is and state_is_not accessor functions so we can
start to see who is using the internal state.
This is the begin of a long march of turning tbf into a C++ class
and properly hiding the secrets inside this implementation instead
of having it spread across various different files.
The TBF should use the IMSI to identify a block flow but all
handling is spread across the entire code. Start to clean this
up by moving relevant code into the tbf file. Afterwards one
can clean up and add more internal structure.
Do not claim that the payload is not known. Add the missing break.
Do not print:
<0002> gprs_rlcmac.cpp:1174 GPRS_RLCMAC_CONTROL_BLOCK_OPT block payload is not supported.
<0002> gprs_rlcmac.cpp:1176 Unknown RLCMAC block payload.
For mocking/unit-testing/emulation (and a dual trx-systems) having
global state is quite bad. Cut back on the usage of the global
struct gprs_rlcmac_bts. It also makes the complexity of certain
routines more clear.
Fixes:
bitvector.cpp: In function 'int bitvec_pack(bitvec*, uint8_t*)':
bitvector.cpp:53:22: warning: comparison between signed and unsigned integer expressions [-Wsign-compare]
bitvector.cpp: In function 'int bitvec_unpack(bitvec*, uint8_t*)':
bitvector.cpp:63:22: warning: comparison between signed and unsigned integer expressions [-Wsign-compare]
bitvector.cpp: In function 'uint64_t bitvec_read_field(bitvec*, unsigned int&, unsigned int)':
bitvector.cpp:91:18: warning: comparison between signed and unsigned integer expressions [-Wsign-compare]
bitvector.cpp: In function 'int bitvec_write_field(bitvec*, unsigned int&, uint64_t, unsigned int)':
bitvector.cpp:108:18: warning: comparison between signed and unsigned integer expressions [-Wsign-compare]
femtobts.c:250:2: warning: excess elements in array initializer [enabled by default]
{ SuperFemto_ClkSrcId_NetList, "nwl" },
^
femtobts.c:250:2: warning: (near initialization for ‘femtobts_clksrc_names’) [enabled by default]
femtobts.c:251:2: warning: excess elements in array initializer [enabled by default]
{ 0, NULL }
^
femtobts.c:251:2: warning: (near initialization for ‘femtobts_clksrc_names’) [enabled by default]
This might clash with C++11 and literal values but we will
see that once the compilers enable that by default.
Fixes:
csn1.cpp: In function 'gint16 csnStreamDecoder(csnStream_t*, const CSN_DESCR*, bitvec*, unsigned int&, void*)':
csn1.cpp:864:17: warning: format '%d' expects argument of type 'int', but argument 8 has type 'guint64 {aka long unsigned int}' [-Wformat]
csn1.cpp:1144:15: warning: format '%u' expects argument of type 'unsigned int', but argument 7 has type 'uint64_t {aka long long unsigned int}' [-Wformat]
csn1.cpp:1150:15: warning: format '%u' expects argument of type 'unsigned int', but argument 7 has type 'uint64_t {aka long long unsigned int}' [-Wformat]
csn1.cpp: In function 'gint16 csnStreamEncoder(csnStream_t*, const CSN_DESCR*, bitvec*, unsigned int&, void*)':
csn1.cpp:2119:17: warning: format '%d' expects argument of type 'int', but argument 8 has type 'guint64 {aka long unsigned int}' [-Wformat]
Decrease the number of lines of a single method by splitting things up.
The fewer lines of code, branches and side-effects in a method, the easier
it will be to understand. The other benefit is that one can start creating
unit tests for the some parts of the code.
One of the issues with not properly re-setting everything is that
due the global state it is not clear which variables belong together
and how long it exists. Begin with creating a osmo_pcu and moving
things into this class.
Think of an organic cell, this commit is introducing the cell wall
around it... and defines what is inside and what is outside of it.
The PCU does not properly re-set the state when the connection to the
BTS is lost (and the SGSN potentially is re-started during that). This
results in the BSSGP BVCI > 1 remaining blocked and no data will be
accepted by the SGSN.
Add the '-e' option and exit the PCU when the BSSGP/NS are getting
destroyed.
osmo_timer_del is an idempotent operation. There is no requirement
to check if it is running. If you don't want a timer to run, delete
it. Maybe one should have called the method _unschedule, _cancel to
make this more clear.
The timing advance of any TBF is stored when it ends. Whenever a new TBF
with the same TLLI is created (downlink TBF), the stored TA is recalled.
This algorithm assumes that the mobile does not move too fast during
transfer. Also the mobile must start a connection in order to get correct
initial timing advance.
This algorithm does not implement the timing advance procedure as defined
in TS 04.60. To implement the standard timing advance procedure, the BTS
must decode RACH on certain bursts, the mobile is expected to send them.
This requires much more complexity to a transceiver like USRP/UmTRX or
Calypso BTS.
The algorithm was tested at TA >= 8 and works quite well.
I ported the Holger's scheduling patch from sysmobts to osmo-pcu.
This is usefull, if PCU uses direct access to the DSP of sysmobts.
The latency to respond to a PH-READY_TO_SEND.ind may not be higher
than 18ms. Currently we are using nice to increase our priority but
for a heavily loaded cell this is not enough. Add an option to enable
realtime scheduling and use it in the screenrc.
Linux offers two realtime scheduling classes these are SCHED_FIFO
and SCHED_RR. For SCHED_FIFO the process is running as long as possible
(potentially taking all the CPU and never yielding it), for SCHED_RR
the process can still be pre-empted at the end of the timeslice.
Using SCHED_RR appears to be the more safe option as a run-a-way
sysmobts process will not be able to take all the CPU time.
The measurements include:
- DL bandwidth usage
- DL packet loss rate
- DL measurements by mobile
- UL measurements by BTS
In order to receive DL measurements from mobile, it must be enabled via
system information message at BSC.
Since we don't know the RX and TX parameters of the BTS that might be used
with PCU, the MS should not adapt the TX power from the RX level. So the
MS should always transmits with same power.
Finding an 'alpha' and 'gamma' value that will result in a constant
RX level at the BTS is a task of deployment.
If not compiled with this support, the closing function does not exists,
so it must not be called.
Removed a "#warning", because closing is already supported.
This simpliefies the allocation process.
tfi_alloc is responsible to allocate a TFI, not a time slot.
The first time slot available depends on multislot class and on other
ongoing TBF (concurrent TBFs), so it is part of the allocation
algorithm to select it.
In order to use this feature, sysmoBTS requires option "-M", otherwise
the traffic is forwarded through socket interface.
This is essential, if PCU runs on processor of sysmoBTS. The reaction
time and delay of PDTCH frames could heavily degrade proper packet flow.
The slots are choosen in a way that later selection of slots for
concurrent TBFs will have the same downlink/uplink control channel
(PACCH).
This is required for polling acknowledge of new TBF assignment on PACCH.
The MS will always have the same PACCH uplink slot while beeing in packet
transfer mode.
Original code: c7e7f6868b
(The code was committed earlier, but got lost somehow.)
I added IMSI, so the paging request is sent in correct paging group.
Also I excluded rest octets from pseudo length.
It is tested and it work.
Some MS will leave packet transfer mode, if TAI is given, but no timing
advance control messages are sent on PTCCH. This could results in
permanent loss of larger packets.
Holger Hans Peter Freyther
Harald Welte