fix performance for chan_counts and all_allocated stats

The all_allocated_update_bsc() does inefficient iterating to count
active/inactive lchans, which scales badly for high numbers of TRX
managed by osmo-bsc.

We need to update the all_allocated flags immediately (periodic counting
alone would suffer from undersampling), so, until now, we are calling
this inefficient function every time a channel state changes.

Instead of iterating all channels for any chan state changes anywhere,
keep global state of the current channel counts, and on channel state
change only update those ts, trx, bts counts that actually change.

A desirable side effect: for connection stats and handover decision 2,
we can now also use the globally updated channel counts and save a bunch
of inefficient iterations.

To get accurate channel counts at all times, spread around some
chan_counts_ts_update() calls in pivotal places. It re-counts the given
timeslot and cascades counter changes, iff required.

Just in case I missed some channel accounting, still run an inefficient
iterating count regularly that detects errors, logs them and fixes them.
No real harm done if such error appears. None show in ttcn3 BSC_Tests.
It is fine to do the inefficient iteration once per second; channel
state changes can realistically happen hundreds of times per second.

Related: SYS#5976
Change-Id: I580bfae329aac8d4552723164741536af6512011

1 files changed, 1 insertions, 0 deletions

diff --git a/include/osmocom/bsc/bts.h b/include/osmocom/bsc/bts.h
index fc3ebce12..66cf68f02 100644
--- a/include/osmocom/bsc/bts.h
+++ b/include/osmocom/bsc/bts.h
@@ -637,6 +637,7 @@ struct gsm_bts {
 	/* At what point in the channel allocation sequence to dispatch the Immediate Assignment (Abis optimization) */
 	enum imm_ass_time imm_ass_time;
 
+	struct chan_counts chan_counts;
 	struct all_allocated all_allocated;
 };