This time constant plays a major role in how long a UE has to wait
maximally before it might be scheduled. For instance, if a UE had a lot
of traffic, the average traffic will be high, pushing the PF coefficient
down. If another UE with no traffic in the past becomes active with
heavy traffic, the first UE might not be scheduled for an extended
period of time.
In the AW2S-Amarisoft UE CI test pipeline, we sometimes see that the
maximum number of RLC retransmissions was reached. While this might not
be the (only) cause, scheduling a UE more often might prevent such
situations, and we saw an improved stability.
Different threads (e.g., RRC, DU tasks) might access the scheduler at
the same time as the thread that drivers the scheduler itself
(ru_thread). To avoid data races, introduce this mutex.
Most functions of the scheduler are only used internally and have been
marked static in the previous commit. The remaining ones, in this
commit, either lock the scheduler mutex sched_lock, or a comment has
been added of the assumption of how such function is to be locked (e.g.,
if it might be called from different places, or if it is called by a
function that already locks the scheduler).
To know which functions are used from outside and which are
module-local, make many functions static. This will allow us to protect
only those functions with a mutex for which it is necessary.
This commit fixes a bug caused by global variables in the gNB_MAC_INST
structure used to transfer MAC scheduler's results to the PHY layer.
When using several threads in the PHY layer, it may happen that the
MAC scheduler is called a second time by the second PHY thread for the
next slot before the content of the global variables in the gNB_MAC_INST
structure used to store the result of the MAC scheduler for the current
slot have been consumed by the first thread, leading to unpredictable
results (all of them certainly wrong).
The solution is to introduce a small memory management module for the
interface between the PHY threads and the MAC scheduler. The MAC
scheduler now fills the data structures passed to it, and not a
global variable anymore. The PHY layer manages the structures and
ensures that no data is overwritten before being fully consumed by
the various threads of the PHY layer.
Some care has to be taken because several PHY threads will access the
same structures at the same time for a given slot (in read mode, so
it's okay). A mechanism of reference counting has thus been introduced.
Only one CC is supported. This restriction should not be too hard
to remove if/when needed. (The code before this commit does not seem
to work for more than one CC anyway.)
During the Initial UL RRC Message, the CellGroup is created. Since the
CellGroup specifies the SRB1, the SRB1 should also be created (because
the DL RRC Messages is sent via SRB1). However, before this commit, the
CellGroup is created while having the RLC lock, so we cannot create
SRB1, as it results in a deadlock.
Therefore, we now first create CellGroup, SRB0, and SRB1, before giving
the UE message to the RLC.
- The qsort() function is defined as:
void qsort(void *base, size_t nmemb, size_t size,
int (*compar)(const void *, const void *));
and sorts an array with nmemb elements of size size.
- in the gNB scheduling functions, nmemb and size were swapped
- this was causing a stack overflow ending with a SEGFAULT under certain circumstances
Robert Schmidt