This function packs P7 messages. At least RX_data.indication messages
might be much bigger than the global buffer that is used here. Allocate
the buffer on the stack, and make it bigger. Do not use the global
buffer, it's simply not necessary; increasing the global tx buffer size
might have negative effects elsewhere in the code.
That change might make the function reentrant. The mutex seems to only
guards the codec config. However, leave it to not introduce any
regressions as of now.
E1 Setup enc/dec library and unit test
This MR is adding a library for E1 Setup enc/dec functions:
- E1 CU-UP Setup Request
- E1 CU-UP Setup Response
- E1 Setup Failure
and relevant unit tests
F1AP IDs: Add update, fix concurrency bug
I suspect a concurrency bug: For instance, during reestablishment, the
CU needs to update the DU UE ID under which the UE is identified in the
DU. Previously, the CU would remove, then add the DU UE ID info. At the
same time, the PDCP thread might look up the information. This can lead
to asserts. Modify the code to do the update under (one) lock.
Also, refactor some code.
Similar to the parent commit, make the numerology in the VNF
configurable. Unlike for the PNF change, the VNF frame/slot info is in a
per-PNF connection structure to which the "oai_integration" code has no
access. So we need to hack the nfapi_vnf_p7_add_pnf() function signature
to take the numerology, to be able to save this on a per-PNF basis.
The fact that we store this on a per-PNF basis is not strictly required,
as the VNF cannot handle multiple numerologies as of now, either.
However, we might want to extend this in the future, so it seems prudent
to store this on a per-PNF basis, so that we could start handling
multiple numerologies at the L2 without the need to change the L1 or
nFAPI code itself.
Also, the numerology is not needed except for some code that deals with
timing over nFAPI. As of now, we don't use this code (nFAPI gets a
trigger every slot, as per the VNF request, see an earlier commit in
this series), but also here, let's not make the future more complicated
by storing the numerology for each PNF now (this could always be reverted).
For 4G, put numerology 0 (15kHZ SCS), which is the SCS that LTE uses.
Fill in the dynamically received numerology, and make the numerology at
the PNF configurable, following the changes in the parent commit.
The change for the reading is somewhat of a hack, because nFAPI, as of
now, does not really foresee to store the numerology, so we fill the
numerology during the start request, when the numerlogy has been
received in the config request, but at that time, the structure for the
time information (frame/slot) does not exist yet.
At both the PNF and the VNF, introduce a separate numerology field, and
pass the numerology when converting time information using macros from
nfapi_interface.h. The actually numerology is still hardcoded to 1 (as it
was before), but the follow-up commits will put a suitable numerology.
In both cases, but the numerology next to the frame/slot information.
Similar as the parent comment, remove time macros. In this particular
case, we can just pass the exact frame/slot information, instead of
squeezing it in a single integer.
The next commits will modify the time macros used in the 5G nFAPI code
to handle different numerologies. As a preparation, remove all the
usages of these macros where they really don't matter (e.g., don't
convert if we don't need to).
the L1 needs both a PDU (in TX_data.req) as well instructions for
the actual transmission (in DL_TTI.req) to encode a PDSCH message. In
nFAPI, it can happen that a DL_TTI.request message has been received
(configuring the PDSCH transmission), without the TX_data.request having
reached the PNF. The L1 assumes to have the PDU.
To avoid problems, ensure that only the pair of DL_TTI.req/TX_data.req
is delivered. Otherwise, drop either message.
nFAPI has a mechanism to segment messages that are too big for transport
over a given medium (see e.g. SCF 225, section 2.3.2).
The maximum segment size of 10000 makes that for larger payloads, e.g.
TX_data.request, many small segments are to be sent. This can create or
increase delays on the transport. On the other hand, the currently only
available transport mechanism, UDP, allows to transport packets of up to
almost 65535 bytes. Correspondingly, increase the maximum segment size
so that less segments are to be created, and potentially, less delay is
to be incurred. # Please enter the commit message for your changes.
Lines starting
Avoid delays in tx_data.request handling by avoiding big malloc()s and
copy operations. Reimplement function to (1) peek the frame/slot
numbers, (2) decide on buffer, and (3) unpack directly into
pre-allocated memory.
Co-authored-by: hsum <ming-hong.hsu@eurecom.fr>
Co-authored-by: chenyi <Yi-Quan.Chen@eurecom.fr>
Co-authored-by: Rúben Soares Silva <rsilva@allbesmart.pt>
The time that a message between PNF/VNF to arrive depends mostly on the
transport. To allow for some delays, there is a slot_ahead time, during
which the VNF is allowed to schedule and send instructions to the PNF.
This can be multiple slots; the 1ms hitherto given might typically too
short. Increase to 10ms, to encompass a wider range of slot_ahead times.
Make the corresponding log message of when old (stale) message are
removed a bit clearer with respect to times.
This is the first commit in which nFAPI works. Follow-up commits improve
performance.
The usage of this flag in the L1 is simply wrong: either, the NFAPI mode
runs L1, or not. Correspondingly, add an assert to check this invariant.
In the NR_IF_module, the correct check to do is for VNF (implying no L1,
or "emulation of L1" [which is not true]), in which case we use
queues.
The L1 receives a pointer to the L1 payload to use for PDSCH generation.
This would overwrite an existing pre-allocated pointer. Remove this
preallocation. Instead, allocate it in the simulator (the only place
that actively uses this buffer allocation; in nr-softmodem, it's lost by
assigning a pointer). Allocate with an extra byte, because the payload
(in bits) might not be a multiple of 8 bits/1 byte, so the CRC is in the
last 3 bytes.
Also, improve the log with frame/slot information if the PDSCH payload
pointer is NULL.
Since commit defc97ec16 ("Msg2 and Msg3 at gNB: check allocation before
committing to Msg2"), the scheduler calculate Msg2/Msg3 frame/slot
before allocation of Msg2, and not when receiving Msg1. It also respects
the RA response window correctly.
Correspondingly, we don't need to calculate here if the RACH.indication
is within the time window. The scheduler will do this.
Reset the counters. In monolithic, this happens in the scheduler; in the
PNF, without this change, it would trigger PDCCH again.
Co-authored-by: hsum <ming-hong.hsu@eurecom.fr>
Refactor pnf_phy_ul_dci_req(), pnf_phy_dl_tti_req(),
pnf_phy_tx_data_req(), pnf_phy_ul_tti_req(), to use L1 functions to
"load" L1 jobs to be executed, via L1 functions.
In monolithic, the sched_response mechanism ensures that FAPI messages
for one slot are not overwritten by scheduler instructions for the
following slots (i.e., it is like a memory allocation pool).
The allocation happens in nr_schedule_response(), not used by the PNF.
Hence, only deallocate when in monolithic.
The PNF uses a big ring buffer over all slots in a frame to ensure the
same.
This function applied an sf_ahead, bigger than the actual slot
indication message sending (so a message could NEVER arrive on time!).
Reduce to zero, because this is when it should arrive for us.
FAPI says these lists should be for MAX_TDD_PERIODICITY (which is maybe
a frame, or a TDD period? but likely not two frames). Reduce to one
frame.
Also, the nFAPI code assumes one entire frame, so use that (this commit
necessary to actually make the PNF work, otherwise, the
packing/unpacking of the FAPI message would need to do 2 frames).
Closes: #876
Upon reception of FAPI UL messages (e.g., RACH.indication), those
messages are put into queues, to be delivered in a single call,
NR_UL_indication(). Call the scheduler to trigger processing in UL.
Note that ideally, these queues should not exist at all, and messages
should just be delivered asynchronously, as this would lead to less
copying of messages. Currently, we fulfill the scheduler interface
instead.
Important! Trigger UL after having run DL scheduler and sent the
messages, to ensure short delay between Slot.indication, and the
response to those messages.
The previous designs seems to do:
loop {
poll_ind_queue()
if (msg)
scheduler;
pselect() on messages;
handle_message {
if slot_ind
put_ind_queue()
}
}
So basically, we artificially put a queue for slot indications in the
middle, but still handle it in the same thread(!). This for some reason
results in a big slow down if the PNF runs faster.
Now, we just do pselect(), waiting for messages. We handle the slot
indication immediately (the schedule should take some microseconds),
then return to pselect().
Use this function to implement distinct behavior for monolithic or PNF:
- monolithic: run the scheduler, as normal
- PNF: send a slot indication; the VNF will send an answer which the PNF
will handle elsewhere.
Add function declarations to make nr_ulsim compile, as NR_IF_Module.c is
part of the MAC sources. As of now, this is necessary; I think the right
way would be to take NR_IF_Module.c out of the MAC, and make this a
"glue" library that sticks L1 and L2 together, in monolithic (in VNF, in
that case, we should not link in L1, and for the PNF not the L2,
creating separate executables... Rome was not built in one day).
We are required to store multiple PUCCH (e.g., for multiple UEs).
However, nr_fill_pucch() automatically freed the previously active PUCCH
structure. Of course(!) this does not make sense.
However, the problem is that init_nr_transport() allocated only one
PUCCH structure. Also this does not make sense to me. The problem was
that slot_ahead was 0, which logically cannot happen in this function
(because slot_ahead is illogical, we must give the radio some time). The
reason is that it uses if_inst->sl_ahead, which, prior to this commit,
was not set.
The init_nr_transport() function is called normally from main() through
init_eNB_afterRU(). In this case, we initialize if_inst->sl_ahead to
what is set in main(). In the PNF case, however, we call
init_nr_transport() from the PNF thread, when receiving the start
request. In the main(), we wait for this to happen, before setting
if_inst->sl_ahead. Hence, initializing this variable (which we should
safely be able to do) before solves the problem.
The radio gives the clock. Therefore, we do not need to keep the time
separately in the PNF.
Also, the VNF should not run the ru_thread, as this is an L1 thread.
This is to align with monolithic/FAPI, where we get slot indications in
each slot. In the nFAPI, this is replaced by timing information, such
that the VNF ticks automatically, and gets indications from the PNF on
early/late messages. The timing information handling is not implemented
in the present nFAPI implementation. Instead, we adopt the FAPI behavior
and send timing information in every slot, abusing it to be a slot
indication.
There is no need to add the PNF in nr_start_resp_cb(), because it has
already been done in nr_param_resp_cb(). This might be wrong, because 4G
does it only in nr_start_resp_cb(), but for the moment, since it is also
used in nr_param_resp_cb(), avoid the double add.
The same function is called in nr_start_resp_cb(), why there and not here?
At reception of nr_param_response, we receive the remote PNF information
on PNF/VNF. We have to store it.
In nr_start_response, it might not be correct and/or can be taken out.
The next commit will remove it.
Prior to this this commit, the structure NR_UE_info_t is 4881632 bytes.
On some machines, e.g., one of the CI systems, this can lead to realtime
problems when doing the initial calloc() of the data structure.
Commit e586efb29d ("Enable 32 HARQ
Processes in DL, UL") introduced the real-time problems by increasing
the number of HARQ processes.
The reason NR_UE_info_t is that big is that it contains buffers for DL
HARQ data that might need to be retransmitted (the L1 is stateless, it
cannot store this for the L2). To reduce the size, dynamically allocate
the DL HARQ buffers when they are needed. This also reduces the size of
NR_UE_info_t to 15840 bytes.
Encapsulate transportBlock in byte_array and accessory functions, and
allocate in big chunks of multiple of ~32kB.
Reported-By: Cedric Roux <cedric.roux@eurecom.fr>
Fixes: e586efb29dCloses: #875
ran_func_kpm.c uses cp_ba_to_str(), but it is not defined in OAI. Define
it so we can link properly, instead of relying on FlexRIC (where it was
defined).
Introduce function cu_update_f1_ue_data() which does remove+add under
one lock. This fixes a concurrency bug. For instance, during
reestablishment, the CU needs to update the DU UE ID under which the UE
is identified in the DU. Previously, the CU would remove, then add the
DU UE ID info. At the same time, the PDCP thread might look up the
information. This can lead to asserts. Use the above addmod
functionality to do the remove-add under a mutex.
Modify test to verify this works as expected.
chore(ci): Move to new Jenkins and Openshift Cluster
- Update helm-charts with new node roles
- Add resource definition in physim helm-charts to easily schedule them
on different cluster servers
- Update entitlement creation methodology
- Add namespace for aerial pipeline
Modify xu_add_f1_ue_data() to be xu_addmod_f1_ue_data(), i.e., "addmod"
behavior in which it is possible to also update f1_ue_data. I do it this
way since an xu_updade_f1_ue_data() function would be almost the same as
the add function (only deleting existing data, whereas the original
add() was asserting in that case).
Modify tests to test for this behavior.
This fixes a concurrency bug. For instance, during reestablishment, the
CU needs to update the DU UE ID under which the UE is identified in the
DU. Previously, the CU would remove, then add the DU UE ID info.
At the same time, the PDCP thread might look up the information. This
can lead to asserts. Use the above addmod functionality to do the
remove-add under a mutex.
Fix calculation of RSSI in pucch_rx.c
Fix two issues with RSSI calculation in PUCCH:
- RSSI was calculated using partial sum instead of average of squared samples.
- There was a missing parentheses pair which causes the RSSI value to underflow.
The proper way to calculate power in dBFS is
let signal_energy = average of squared samples of received signal over resource elements, antennas and symbols under measurement
then:
rssi[dBFS] = -10log10(max_sample^2 / signal_energy),
or (to avoid division):
rssi[dBFS] = -(10log10(max_sample^2) - 10log10(signal_energy))
CI: Increase number of test trials for timing tests
- Increasing a number of test trials aims to reduce variations and
improve reliability of the nr_ulsim and nr_dlsim test results.
- Disable CPU sleep states on the machine when running physims.
- Minor adjustment of timing threshold in 60 MHz phytest.
Delay compensation for PUCCH2
- Implementation of delay compensation for PUCCH2, which minimizes for
exampe logs "reported RSRP index XXX invalid"
- Adding a test to the CI to perform the delay compensation in PUCCH2
- Refactor code to use c16_t data type in PUCCH decoder
Remove spinlock from threadpool task return
The idea to use atomic aligned to cache line for counting finished jobs
is reused, but instead of polling a semaphore was added.
remove calls to clear SIMD registers, that was a historical issue
This works because gcc/clang replaces the MMX calls by SSE2 automatically.
There are some mm_empty() calls in bad places (places where we don't use
_m64) and it is lacking in places where we still use _m64. Given that
MMX calls are converted to SSE2, there is no problem, and we will remove
remaining _m64 variables.
RSSI threshold for PUSCH & PUCCH power control
To prevent ADC railing or "clipping" when the received power is too
large, introduce RSSI threshold parameters for PUSCH and PUCCH, which
are now configured in units of 0.1 dB. Here is how you can configure and
use these thresholds:
1. Configuration:
- Set the pusch_RSSI_Threshold and pucch_RSSI_Threshold parameters in
your configuration file (e.g., gnb.conf) (MACRLC section).
- These parameters define the maximum acceptable RSSI values for
PUSCH and PUCCH, respectively. The unit is either dBm or dBFS,
depending on the RSSI reporting configuration.
2. Usage:
- The nr_limit_tpc function will automatically adjust the Transmit
Power Control (TPC) commands based on the configured RSSI
thresholds.
- If the RSSI exceeds the threshold, the TPC command will be limited
to prevent further increase in power, thereby avoiding ADC
clipping.
Introduce ITTI queue in RRC-to-MAC direction
resolve ue rcc to mac race conditions by a itti queue, but it doesn't
resolve the problem of exact slot we should apply the new cell
configurtion
fhi_72: allow to provide dpdk_iova_mode to xran
- Introduce a gNB configuration option that enables users to select the
desired IOVA mode for DPDK ("PA" or "VA") with dpdk_iova_mode
parameters in fhi_72 configuration section. If not set, the default
remains PA mode to ensure backward compatibility with existing
configurations. Detailed information about IOVA modes, including their
advantages and disadvantages, can be found in the DPDK documentation:
Memory in DPDK Part 2: Deep Dive into IOVA
- Using IOVA VA mode allows DPDK to run without requiring SYS_ADMIN
capability for the docker container.
- Switch to IOVA VA mode in the FHI7.2 pipeline and reduce the required
capabilities for the docker container.
Move UL_INFO structure allocation local to a rx slot processing
UL_INFO is allocated a gNB, so require mutex and obscur information
passing function to function. Instead, make a local UL_INFO per slot,
removing the mutex and making the code easier to understand.
nrLDPC_coding: Fix naming, comments, superfluous variables and documentation typos in coding library interfaces
This MR is clarifying the interfaces to the LDPC coding libraries. It
aims to improve the readability of the interfaces and compliance of
naming with specification 38.212.
Robert Schmidt