This includes changes required for
* shared main() function accross all BTS models
* use of the new phy_link / phy_instance infrastructure as the basis
for true multi-TRX operation
by using a talloc pool, we avoid having to go back to the libc
malloc pool all the time. The msgb allocations and libortp allocations
happen quite frequently during processing and show up as one of the
high priority items in osmo-bts profiles on sysmoBTS with 14 concurrent
TCH/H calls (highest load scenario).
talloc still consumes significant CPU, this is mostly due to the
zero-initialization of all the associated buffers. Strictly speaking
we shouldn't need this, but any change there would require lots of
testing, as there might be hidden assumptions in the code?
In some percentage of cases, talloc still seems to fall back on malloc
for msgb allocations, which is currently a bit of a mystery. The pools
certainly are large enough, talloc_reprt() rarely reports more than a
few tens of kilobytes used by the msgb pool.
From 2ecbf87130
This commit adds basic support for the Litecell 1.5. Multi-TRX is not
supported yet. Instead, multiple instances of the BTS can be launched
using command line parameter -n <HW_TRX_NR> to specify if TRX 1 or
2 must be used by the bts. Note that only TRX 1 opens a connection to
the PCU. Full support for GPRS on both TRX will come at the same time
than the multi-TRX support.
The BTS manager has been adapted to match the new hardware but otherwise
it has not been improved or changed compared to the one used on the
SuperFemto/Litecell (sysmobts).
When the OML signalling link is lost, first set bts->oml_link = NULL,
then iterate over the RSL links and close them. Closing the RSL link
will cause a OML state change message to be sent, which in turn tries
to use the no-longer-existing OML link.
The code should be cleaned up further to distinguish which signalling
link was lost, and actually communicate a RSL(only) loss to OML.
But for now, it's best to simply close down all links and terminate
osmo-bts to ensure all state is properly reset and recovered.
It seems the right thing to do: Once we know a PHY link is established,
the associated OML managed objects should change their state
accordingly. However, given all the hackery we do with MO states, this
actually breaks things, rather than helping. So I'm disabling that part
for now, but this needs to be re-visited at some point.
the backend is performing the actual encoding and decoding functions,
while the generic part constsits of the TDMA structures and generating
the RTS.ind
The L1 scheduler is a generally useful component that is unfortunately
tied quite a bit into the OsmoTRX support. Let's try to separate it out
by having separate per-trx/per-ts/per-chan data structures pre-fixed
with l1sched_
Using this patch it should be one step easier to use the scheduler for
other BTS models, such as the intended upcoming virtual BTS.
During the L1SAP related changes, somehow an old version of
check_for_ciph_cmd() was re-introduced, which didn't store the N(s) as
part of the lchan. To make things worse, the old code was still present
in the sysmobts specific part, but never executed.
When the oml_link is down or not yet established, we currently lost
any OML messages that were scheduled for transmission to the BSC. Let's
prevent that by keeping a queue of OML messages, which is drained at the
time the OML link comes up again.
Use the right identifier for the timeslot and not the trx number
which would always use ts==0 on the first trx. This should fix
ciphering issues for TS>0 (e.g. SDCCH8 on TS==1)
It seems that once we start to respect the T200 values as specified by
the BSC, we run into all kinds of issues with LAPDm re-transmissions,
REJ frames, unexpected supervisory frames and the like.
The libosmogsm LAPDm T200 defaults of 1s/2s are proven to "work" (i.e.
not expose the above behavior), so let's revert to them until the root
cause of this problem is determined.
The T200 default values should be in milli-seconds (as the variable name
indicates). They are not expected to be divided by the TS 12.21 OML
dividers for T200.
This change doesn't really make a difference with OpenBSC, as the BSC
always sets its own T200 values via OML, overwriting the defaults here.
There's no need to use memcpy(), which adds the risk that the types of
source and destination are not the same (see previous commit). Iterating
over the array and assigning each element is more robust.
t200_ms is an unsigned int [7] array, while the oml_default_t200_ms was
an uint8_t[7] array, which we memcpy() to the former as default
initializer. Fix this by turning oml_default_t200_ms into unsigned int,
too.
We re-use the 'wait_l1_conf' structure for implementing the
unacknowledge command window towards the PHY. This means that thre will
unconditionally be a 'wait_l1_conf' now, even for requests where the
caller didn't provide a call-back.
When re-starting OsmoBTS after unclean shutdown, the PHY is already
sending notifications (PH-DATA.ind, PH-TIME.ind, etc.) for the previous
physical channel / timeslot configuration. At the point those messages
are received, OsmoBTS might not even have A-bis OML up yet, and thus has
no clue how to process such messages (and subsequently likely crashes).
Let's block such primitives from passing further up the code until we
have received the TRX-OPEN response.
When writing the config file from the command line, we must not forget
to write the phy-netdev parameter, otherwise the program will fail to
re-start later :/
by using a talloc pool, we avoid having to go back to the libc
malloc pool all the time. The msgb allocations and libortp allocations
happen quite frequently during processing and show up as one of the
high priority items in osmo-bts profiles on sysmoBTS with 14 concurrent
TCH/H calls (highest load scenario).
talloc still consumes significant CPU, this is mostly due to the
zero-initialization of all the associated buffers. Strictly speaking
we shouldn't need this, but any change there would require lots of
testing, as there might be hidden assumptions in the code?
In some percentage of cases, talloc still seems to fall back on malloc
for msgb allocations, which is currently a bit of a mystery. The pools
certainly are large enough, talloc_reprt() rarely reports more than a
few tens of kilobytes used by the msgb pool.
In some situations, a PHY might send us a primitive for a logical
channel that is not (or no longer) active. Passing such primitives
higher up the stack is asking for trouble. Specifically, LAPDm
instances cannot accept messages once their instance has been released.
We introduce two new helper functions: get_lchan_by_chan_nr() as well as
get_active_lchan_by_chan_nr(). The former just centralizes the look-up
of the lchan by timeslot number and sub-slot number. The latter also
checks to ensure the lchan is active, which is used for PH-DATA / PH-RTS
primitives. To the contrary, MPH primitives generally don't require the
cahnnel to be active for processing.
The way we recycle the msgb with a l1sap header when transforming a
PH-DATA.req L1SAP primitive into a PHY/L1 primitive was flawed in
several ways:
1) the way the L1SAP header was stored in the buffer didn't provide
sufficient tailroom for the L1 primitive
2) the alignment of the data in L1SAP is at a 32bit bounadry, but not
in the L1 primitive, causing unaligned accesses.
OpenBSC introduced a naming change in
615ed46a6ab25f71a7ab0d8201d33b4dbf8fc5b0 but osmo-bts fixes were only
about osmo-bts-sysmo, not osmo-bts-trx. This updates osmo-bts-trx
accordingly.
A known issue with this code is that BER is not updated for lost TCH frames,
because osmo-trx doesn't send any indication for them and we don't have
a callback to handle this.
Otherwise the code seem to work fine.
3GPP TS 05.03 "Channel coding" specifies the puncturing matrix (1,0,1)
for class 1 information bits and tail bits valued u(0) to u(103) for a
maximum puncturing index of 311. The puncturing index 313 exceeds the
maximum index and causes osmo_conv_get_output_length() to output the
improper length of 210 instead of 211.
Signed-off-by: Thomas Tsou <tom@tsou.cc>
If frame number is out of range (>= 2715648), the scheduler's process
would end up in an infinite loop. This is because the loop would schedule
bursts until the indicated frame number is reached, which would not be
possible.
The openbts, calypso-bts and osmo-trx might send out out of range clock
indications every 3.5 hour.
RTS (ready-to-send) must be issued in advance, so BTS core and especially
osmo-pcu can provide downlink data frames early enough. In some cases PCU
might provide frames too late, so they must be dropped. If PCU provides
frames too late, due to high system load, this "RTS advance" setting must
be increased.
Instead of limiting the number of TRX at VTY to the actual number of
supported TRX, VTY allows to configure any possible number of TRX. If a
TRX is configured, which is not supported by BTS model, an error message is
returned, which states that the given TRX is not supported.
Handover and assignment may activate channels with ciphering already set,
so we need to tell scheduler to enable/disable ciphering and set the
correct cipher state.
Only if transceiver becomes available, control commands are sent. If
tranceiver is gone, reset scheduler.
The current availability state is sent to BSC via OML state change
commands.
MS uplink power control is required in pretty much any BTS, and we
cannot assume that they PHY / L1 will always take care of it by
itself. So the correspondign code is moved to common/power_control.c
and called from the generic part of L1SAP.
The corresponding VTY paramter has been moved from the sysmobts-specific
trx VTY node to the common BTS VTY node.
There are three transitions:
1. LCHAN_CIPH_NONE -> LCHAN_CIPH_RX_REQ -> LCHAN_CIPH_RX_CONF
It is used to enable ciphering in RX (uplink) direction only.
2. LCHAN_CIPH_RX_CONF -> LCHAN_CIPH_RX_CONF_TX_REQ -> LCHAN_CIPH_RXTX_CONF
It is used to additionally enable ciphering in TX (downlink) direction.
3. LCHAN_CIPH_NONE -> LCHAN_CIPH_RXTX_REQ -> LCHAN_CIPH_RX_CONF_TX_REQ
-> LCHAN_CIPH_RXTX_CONF
It is used to enable ciphering in both TX and RX directions. This is used
when the channel is activated with encryption already enabled. (assignment
or handover)
In order to follow the order of these transitions, the RX direction must
always be set before the TX direction.
If no cipher key is set (A5/0), ciphering is set to ALG 0, but lchan cipher
state remains at LCHAN_CIPH_NONE.
This part moves control channel message primitives from osmo-bts-sysmo to
common part.
In order to control ciphering fo BTS model, CIPHER (MPH_INFO) messages are
used.
The original code handled both the fact where a TIME indication would be
missed (and thus the frame number be higher than previous + 1), as well
as the two cases for combined / non-combined CCCH.
The L1SAP code removed some of those bits, which I'm re-introducing
here.
This is a regression of the code compared to the existing
sysmoBTS code, where the L1 tells us whether its AGCH or
PCH. However, it was not used even in the old code, so
we can afford to simply put a #warning here.
In case of a RACH INDICATION on CCCH, we need to set CHAN_NR to
0x88 (RSL_CHAN_RACH). In other cases, chan_nr needs to reflect
the actual logical channel (TCH/SDCCH) on whcih the handover happened.
... in an effort to avoid introducing new/more spaghetti code
Also, use offsetof() instead of pointer calculation to determine
the start of GsmL1_Prim_t.u.phDataReq.msgUnitParam.u8Buffer
This first part moves BCCH message primitives from osmo-bts-sysmo to common
part. A new file "common/l1sap.c" is introduced to implement handling of
layer 1 messages from/to BTS model.
Spotted by Ciaby while debugging an audio issue. Do not
send anything to port==0 to the BSC/NITB. Look at the
upper bits of the speech_mode to determine if sending is
allowed. 0x1 means recv_only and all other modes allow us
to send.
Manually verified with a single phone call with LCR bridge
mode to send a CRCX early but a MDCX sendrecv later. The
audio starts to flow after the MDCX message. Virtual Addr
space didn't increase over 10 calls. The l1p_msg is freed
by the caller.
The code might not re-set speech_mode from one call to
another but if it is ever != 0 it can be expected that
the BSC will always set it. This is because we do not
(and don't want to) allocate the lchan dynamically on
every usage.
Fixes: SYS#2111
Use the new libosmo-abis API to query the session for the
statistics and then send it as a TLV element to the BSC.
This can be used to do post processing about the call. E.g
to figure out if no audio arrived at all.
The RSL_IE_MEAS_RES_NR is mandatory element with a minimum
of 5 octets (two for TL and three for the value). When we
establish a new channel we might not have had enough time
in a TDMA frame to calculate the average. The issue is not
easy to reproduce. At the point we receive the measurement
report we have two uplink measurements queued. As it is not
easy to reproduce and only occurs when a channel is new
I have decided to drop the message instead of sending made
up uplink measurement reports.
As of now lchan_build_rsl_ul_meas will always return 3 and
the condition will never be false.
Avoids: SYS#1781
The write_queue is designed to have a maximum amount of pending
messages and will refuse to take new messages when it has been
reached. The caller can decide if it wants to flush the queue
and add the message again, create a log. But in all cases the
ownership of the msgb has not been transferred. Fix the potential
memory leak in the failure situation.
Use the standard RSL commands to order a logical channel
to use a fixed power level.
The code is not fully verified and there was a last minute
change to invoke bts_model_adjst_ms_pwr.
Currently the DSP is instructed to achieve a given uplink
power target but there are circumstances (e.g. EMV testing)
where we need more control over it. The "manual/software/osmo"
power control can only be implemented per TRX and not per
lchan. Add a very very basic control that checks the MS Power
used by the phone, the actual receive level and then adjust
the power.
The code doesn't take the history into account, if the phone
can not reach the requested power level the code will be stuck
(e.g. no timeout based on multiframes). It has a mode for a
fixed power control but no way to set it yet.
The change of the mode requires a restart of the software.
* Print the GPS FD that was opened (e.g. to see if it was
closed again)
* Print the state changes/expectations
* Print the correct to be applied. I wondered if I shouldo do
a cor = cor * -1.. cor = -cor.. or add CLOCK_CORR(err) macro
to use it inside the printf and correction and decided the
gain is not worth the risk.
Continously run the calibration process. Everytime we call the
reset function classify the outcome. In case of a failure schedule
the next command soon and otherwise wait several hours.
Remember if the process was started through the VTY or the run
loop. In case it can't be started immediately reset and schedule
a new run.
After a reboot the system might have been off for a long time
and the currently used value might be wrong. Remember that we
never ran the calibration and execute it on start.
We should only calibrate the clock if there is a GPS fix. Start
gpsd to determine if there is a fix or not. Work around trimble
decoding issues (sent an email upstream). We need to gain some
more experience to see if there memory leaks. We also need to
re-schedule the calibration depending on the outcome.
Change the sign before passing it as correction value. The error
is the difference between the TCXO and GPS. We need to correct by
the reverse of the error. This seems to be different depending on
the clock source we have.
This is a last minute untested change.
This runs the entire procedure for calibration with reasonable
error and success checking. It can be triggered from the VTY
of the sysmobts-mgr right now.
What is missing is to hook up with GPSD to check if the system
has a fix and provide a mode that will continously run the
calibration command.
The CTRL code should have used/extended the l1_if calibration
code. The sysmobts-mgr code first needs to determine if the
clock adjustment is necessary at all. This is done by first
resetting the counters, then waiting, then asking for the diff
and then applying the correction value. But the reference clock
is only set by the application comand.
Copy more code of l1if_rf_clock_info_reset to set the reference
clock as value. This is leaving some todos inside the code that
will be resolved as part of SYS#835.
Related: SYS#835
In the long run we will connect to GPSD and wait for a fix and
then run the calibration. The first step is to open (and re-open)
the control connection to the BTS.
As the connection is on localhost there should not be a computation
overhead to always have the connection open. When connecting assume
that the ASYNC connect worked directly as otherwise we get no
notification of the failure.
This looks like a "bug" of libosmo-abis that should check if the
socket has been connected or not.
This should handle OML channel combinations with CBCH and activate the
CBCH SAPI towards the DSP correspondingly. What is still missing is
sending any actual information over the CBCH in respons to the
PH-RTS.ind coming up from L1.
Read the serial number once and format it as a string. In
case no serial number is present -1 will be returned.
Manually tested with a slightly modified version. serial_nr
was the expected one.
Add new power actions for the sysmoBTS2050. This allows to
switch off the secondary/slave when the system temperature
is too high and back on when the normal level is reached.
Do not allow to switch off the master (so remove the enum
value), do not check if the slave is switching itself off.
Instead of keeping state to remember what was done and needs
to be undone this patch introduces actions that will be executed
when the system is back to normal.
By design the system is considered to be in the normal state
and these actions will be only executed after the system is
coming back to the normal state.
One advantage of this scheme is that an operator can decide
that an overheated systems hould be off duty and requires manual
interaction to be allowed back in service.
The change has only been smoke tested
Fixes: SYS#833
Read the clock calibration from the place that will be read by
the BTS process. Use the standard eeprom code for doing that.
The code assumes that this and the other eeprom code don't
write/invlidate the others reason. If that assumption would not
be true calls to eeprom_free_resources should be added.
For systems without direct access to the PA the best option
is to simply switch off the bts service. This will stop the
transmission which will take load from the DSP/FPGA/RF circuit
and indirectly from the PA as well.
We should introduce "pa-on and bts-on" that can be executed
as "normal" action.
Somebody could decide to switch off the PA in the warning level
already. Support this mode of operation. This means we could have
a config that:
* Enables the PA in the normal level
* Disables it in the critical level
With kdbus or better IPC we could even have the PA and other
parts be represented as service that talk to a bts manager and
then simply execute start/stop requests. This would make the
entire TODO entry irrelevant as state would be managed by
systemd and one can see the time the service was executed.
With "show bts 0" all objects were still listed as dependency.
Once the BTS has been started.. move all the other objects into
the enabled state. Our OpenBSC OML code doesn't care but people
using the VTY to inspect state will be more happy.
One day... we will create proper statemachines inside the BTS
and the BSC instead of changing the state in the BTS impl.
Fixes: ONW#1330
The service file will prevent a re-launch of the sysmobts.service
in case the main process exits with '1'. In case the ethernet is
not available yet the routine would fail and the sysmobts process
will not be restarted.
unable to connect/bind socket: Network is unreachable
<000f> input/ipaccess.c:885 cannot open OML BTS link: Network is unreachable
unable to connect to BSC
root@sysmobts-v2:~# echo $?
2
Fixes: SYS#736
libosmo-abis doesn't make it easy to have these parameters
as const.. just declare it non-const in the api. We pass
a static string but we know it will not be modified.
We use the completion call-back from L1 to compare the instructed
TRX board output power with the actual value as reported back from
L1.
Right now we only print an error message in case the values disagree.
In the future we might want to either use that value as part of our
calculation or send an OML alarm report to the BSC.
We do not have the vty test script to do roundtrip testing. There
is no max-initinal-pout, then there was a typo inside 'initial' and
for the relative step size the unit is dB/mdB. Fix both of that.
Back in March 2013, some structures and defines related to decoded
measurement reports have been moved from openbsc to libosmocore
(libosmocore e128f4663104ed64e33e362cff2566f36d65e658) so that they can
be used also from osmo-bts. This finally follows up on that.
You need openbsc 7ff4f0e0fc692bfab829da50edb104e58b271e7e or later.
This sysmobts specific control interface allows for clock calibration
from an external program by means of the "trx.0.clock-info" and
"trx.0.clock-correction" values.
Harald Welte
Yves Godin
Alexander Huemer
Martin Hauke