Refactor DLSCH scheduler
Refactor the DLSCH post processor similarly to what has been done in
!3521 with the goal of saving an additional iteration through all UEs,
hopefully making the DLSCH scheduler faster.
It further includes changes to make working with and debugging multi-UE
operation easier:
- add some measurements in L1 (for total TX/RX time)
- a new option MACRLCs.[0].stats_max_ue to limit the number of UEs in
the periodic output (default: 8, 0=disable)
- avoid stack overflow in the MAC
- fix the writing of stats to nrMAC_stats.log and nrL1_stats.log when
the written output reduces by truncating the files
- avoid to schedule retransmission with 0 RBs.
This slightly speeds up the scheduler (it e.g., does not need to check
for UEs that have retransmissions), and harmonizes with pf_ul(), which
evaluates (UL) BSR at a similar place within the scheduler.
Check and return, if necessary, that we have enough resources to make
retransmissions. Without, we can asserts in L1, such as
Assertion (NPRB>0 && (NPRB + RBstart <= BWPsize)) failed!
In PRBalloc_to_locationandbandwidth0() openairinterface5g/common/utils/nr/nr_common.c:506
Illegal NPRB/RBstart Configuration (0,51) for BWPsize 51
which indicates that the scheduler requested a transmission with 0 PRBs,
which does not make sense.
Refactor PRACH handling at the gNB
- Remove race conditions by correctly setting mutexes
- Fix a bug when having multiple RACH occasions
- Reduce memory footprint by removing global memory, which should also
fix potential memory data races in FHI 7.2
- Additional fixes, see commit messages
Call the post-processor when the allocation is "fixed" (i.e., nothing
changes anymore), instead of having a final loop across all UEs that
might potentially be costly.
In order to calculate the PF metric, note that the statistics "reset" is
moved into pf_dl() instead of the post-processing loop, as it has to be
done on each slot.
See also: dd98030202 ("Refactor to call post-processor in place")
Similarly to parent commit(s), refactor the use of NR_sched_pdsch.
Instead of using nr_mac->sched_ctrlCommon, put the NR_sched_pdsch_t on
the stack.
This also refactors update_rb_mcs_tbs() to update the existing
NR_sched_pdsch instead of updating individual fields.
This also reduces the extend to which sched_ctrlCommon is used.
Assign NR_sched_pdsch to sched_ctrl->sched_pdsch in one place, where the
post-processing functionality will be called in a follow-up commit. This
should result in no functional change.
It seems to be assumed that curInfo actually has the currently active
number of layers/PMI. However, I am not sure this would always be true
(e.g., UE might not have been scheduled for some slots, but number of
layers decreased as reported by CSI). Instead, recalculate the info
from scratch when necessary.
PF uses MCS and average throughput to decide which UE to schedule. Thus,
the first loop decides on MCS, whereas the second does the "main UE
allocation" after sorting by the PF metric.
A follow-up commit will remove sched_pusch from NR_UE_sched_ctrl_t.
Thus, we cannot rely on sched_ctrl to store the MCS in the first loop,
and look it up from there in the second. Instead, store it as part of
the UE iterator data, and take it from there.
See also: f3068caa37 ("Store selected MCS in UE iterator", for UL)
Introduce a struct with information on FAPI structures to save resource
and data allocation into. It's in nr_mac_gNB.h because the next commits
will reuse this for all preprocessors.
Especially in pf_dl()/pf_ul(), we use a first loop using
UE_iterator(), then iterate over a subset of UEs. It has happened a
couple of times that I was using variable "UE" (from the iterator), when
it should only be used in the first UE_iterator() loop.
Rewrite UE_iterator() to effectively limit the scope of the iterator
variables to the body of the UE_iterator. This should also make it
possible to reuse multiple UE_iterator()s with the same variables in
sequence.
Fix for scenario with no CSI report configured for CSI-RS
After changes in !3714 (merged), if we select to measure RSRP on SSB and not
CSI-RS (default option) and there is a single antenna port (so no MIMO report
either) we won't schedule any CSI-RS measurements for the UE. In that case
ideally we shouldn't set do_CSIRS = 1 in configuration file but still the
option is possible. So the fix is to not perform any measurements at UE on
that CSI-RS but not asserting.
Closes#1015
The function rx_nr_prach become autonomous, it doesn't read anymore global variables
So, it can become (later) a threadpool work, that will send a message when it finishes
rx_nr_prach_ru() still use some global variables, but it is executed in sequence, inside the ru main loop, that limits the probability of race conditions
Handle GTP receiver errors, fix memory leaks, add CU-UP test
- Fix use-after-free bugs in GTP by properly stopping the receive thread(s)
- Fix a bug from !3519 (merged): QFI 0 is a valid QFI
- Fix multiple memory leaks in CU-UP, E1, CU-UP load tester, SCTP
- Add a (conservative) functional CU-UP test that verifies that the load tester
and CU-UP work. Test only a low throughput, as it will run in parallel with
other tests.
The gNB will stop the periodical logging of UEs in MAC Info level after
a configurable number of UEs have been printed, by default 8.
Set it to 17 in some CI tests to ensure we still get all UEs for the
time being, as the CI might depend on this.
improvements in formatting and LOGs in DMRS common functions at MAC
DMRS LOGs at MAC were misleading. This MR tried to improve them (and also
adds some formatting fix).
The call to snprintf() as was used is wrong: we use the return value to
advance output, but the man page says
> The functions snprintf() and vsnprintf() do not write more than size
> bytes (including the terminating null byte ('\0')). If the output was
> truncated due to this limit, then the return value is the number of
> characters (excluding the terminating null byte) which would have been
> written to the final string if enough space had been available.
> Thus, a return value of size or more means that the output was
> truncated.
Thus, output could go beyond end, and we get a stack overflow. Instead,
encapsulate the call to snprintf() checking this condition, and only
advancing output as intended, limiting to 0 if we are at the end of the
buffer. This avoids this error:
==964825==ERROR: AddressSanitizer: stack-buffer-overflow on address 0x7be799f51988 at pc 0x7fe7d66b0d39 bp 0x7be79b825ef0 sp 0x7be79b8256c0
WRITE of size 3 at 0x7be799f51988 thread T24
#0 0x7fe7d66b0d38 in vsnprintf (/lib64/libasan.so.8+0xb0d38) (BuildId: 10b8ccd49f75c21babf1d7abe51bb63589d8471f)
#1 0x7fe7d66b2d44 in snprintf (/lib64/libasan.so.8+0xb2d44) (BuildId: 10b8ccd49f75c21babf1d7abe51bb63589d8471f)
#2 0x0000008fe7c6 in dump_mac_stats /home/richie/w/refactor-dlsch/openair2/LAYER2/NR_MAC_gNB/main.c:183
#3 0x00000092071a in gNB_dlsch_ulsch_scheduler /home/richie/w/refactor-dlsch/openair2/LAYER2/NR_MAC_gNB/gNB_scheduler.c:200
#4 0x0000008f8f78 in run_scheduler_monolithic /home/richie/w/refactor-dlsch/openair2/NR_PHY_INTERFACE/NR_IF_Module.c:399
There is the (global) sched_lock that prevents concurrent access. A
dedicated UE lock does not make sense in that case. Worse, this lock
does not prevent all concurrent UE accesses (e.g., there are many loops
over UEs that are not protected), so it's also wrong.
truncate resets the file size to 0. Without this, e.g., when
disconnecting a UE, the old UE remains at the end and is not
overwritten, which is confusing.
Other CSI measurement fixes
Errors fixed in this MR:
- UE currently sends empty CSI measurement report resulting in invalid
cqi_idx 0, default to MCS 9 at gNB because mismanagement of new L1
measurement structure at MAC UE
- nrofReportedRS not updated after UE capability reception at gNB
- multiple RSRP reports multiplexed / de-multiplexed in wrong order
Refactor SDAP adaptation work for QoS
This MR is an adaptation of the code introduced in !2703.
The goal is a cleanup and functional refactor of the SDAP/PDCP
integration, that is:
- centralize SDAP configuration logic
- simplify entity creation
- improve robustness of role-based SDAP handling
- prepare the codebase for clean support of DRB reconfiguration and
teardown flows
More concretely:
1. New sdap_config_t and get_sdap_Config()
- Introduced an internal struct sdap_config_t to abstract SDAP config
input
- Centralized SDAP config parsing into new get_sdap_Config() function
that translates ASN.1 NR_SDAP_Config into this internal
representation. This is called during RRC triggered
(re)configuration procedures.
- Enable mapping of number of QoS flows > 1
2. Unified role handling with bitmasks
- Added a bitmask-based entity_role field to track SDAP directions:
SDAP_UL_RX, SDAP_UL_TX, SDAP_DL_RX, SDAP_DL_TX: header presence was
determined by simple direction (UE vs gNB) and individual
is_sdap_rx/tx() checks, however nr_sdap_rx_entity behaves
differently depending on the direction of the RX entity (DL/UL),
therefore this change can help enable a more accurate SDAP header
handling across directions and node types.
- Replaces old helpers is_sdap_rx() / is_sdap_tx() with internal
checks against the role bitfield
- QFI-to-DRB mapping is now annotated with direction for both data and
control PDUs
3. Simplified DRB and SDAP setup
- add_drb() in PDCP API now delegates SDAP config processing cleanly
- new_nr_sdap_entity() is used consistently to register a fresh SDAP
context for the UE and PDU session.
- Also, re-use and reconfigure existing SDAP entity if already
exists.
- logging has been added to trace mapped QFIs from RRC and resulting
table entries
4. Improved QFI-to-DRB management
- Refactored nr_sdap_qfi2drb_map_update() to register QFIs based on
parsed config
- Ensures control PDUs are correctly routed even after reconfiguration
- Extended nr_sdap_ue_qfi2drb_config: perform UL QoS flow to DRB
mapping configuration for a SDAP entity has already been established
according to TS 37.324, 5.3 QoS flow to DRB Mapping, 5.3.1
Configuration Procedures. Handle both UL QoS Flows mapping rules to
add and to remove. QFIs will now be properly unmapped if specified
via mappedQoS_FlowsToRelease, this might prevent stale QFI mappings
from being used incorrectly.
- nr_sdap_qfi2drb() and sdap_map_ctrl_pdu() now strictly fallback to
default_drb only if a QFI mapping is missing. Missing default DRB
now triggers LOG_E() (previously silent or misrouted to DRB 0).
5. Other cleanup
- Remove dependencies from legacy rb_id_t
Improve handling of selection of RSRP/SINR report via configuration file
In current develop, the selection of SSB or CRI RSRP depends on the
do_CSIRS flag being activated. That's not wanted. CRI report for now it
is never handled, while we need SSB RSRP report for beam switching
regardless of CSI-RS configuration.
This commit refactors the SDAP entity QoS flow to DRB mapping functionality
to improve code organization and enhance maintainability.
Key changes:
- Split nr_sdap_qfi2drb_map_update into focused functions:
* nr_sdap_add_qos_flows_to_drb: handles QFI addition to DRB
* nr_sdap_rm_qos_flows_from_drb: handles QFI removal from DRB
* nr_sdap_ue_control_pdu_config: handles UE control PDU generation
- Consolidate add/modify logic into reusable nr_sdap_addmod_entity function
- Remove entity existence checking and updating from new_nr_sdap_entity:
this function should only add SDAP entities
- Improve function separation of concerns: do SDAP entity add or update
separately and improve modularity by splitting operations in self
contained functions
- At SDAP creation, do only add of QFI to DRB mapping
- rename new_nr_sdap_entity to nr_sdap_add_entity for function naming consinstency
The TX/RX entity should have different behavior according to
5.2.1, 5.2.2, 3GPP TS 37.324, depending on whether it is UL and DL.
The gNB can play a role both as UL RX entity and DL TX entity, while
the UE can play a role both as UL TX entity and DL RX entity.
The goal of this commit is to extend the information stored in the
SDAP entity as the current code is just marking the entity as TX/RX.
The commit introduces a bitmap to store the role of the
entity, which can be DL RX, DL TX, UL RX and UL TX. Based on this
the TX SDAP entity knows whether a DL/UL PDU should be build, and
the RX SDAP entity knows whether a DL/UL PDU should be processed.
See also Figure 4.2.2-1.
SDAP headers are present only if enable_sdap option is used.
* use get_sdap_Config to process sdap_Config
* move SDAP config logic to the scope where it is relevant
* add SDAP entity direcly in RRC, not from PDCP
* replaced nr_pdcp_add_drbs with single add_drb call: it was no longer
efficient to keep it in the new logic. The addMod list can be looped directly in RRC.
* add specific function to add bearers from addMod list in E1, NSA, and UE
UE symbol based PDCCH receiver
This is a partial rebase/rewrite of !2895
The goal of this MR is to check how we can merge this.
The eventual result is stated in !2895. The way forward is to modify UE
RX procedures to work symbol by symbol by extracting the symbol loop out
of RX procedures. After all channels are modified this way, the main
thread can be modified to work symbol by symbol.
At this point only PDCCH was modified.
1. Refactor PDCCH PHY functions to take rxdataF of one OFDM symbol and
produce LLR for each search space. The modified functions take
rxdataF in a different format than that produced by nr_slot_fep(). So
for now a memcpy is used to change the array structure. Once
nr_slot_fep() is modified, this memcpy will no longer be needed and
will be removed.
2. Modify DCI functions to decode DCI in the last PDCCH symbol and
inidcate to MAC once.
3. Fix PDCCH monitoring for start symbol > 0.
4. Fix multiple PDCCH monitoring within slot.
Co-authored-by: Sakthivel Velumani <mail@sakthi.me>
Robert Schmidt