[PATCH v3 2/5] sched/deadline: Fix reclaim inaccuracy with SMP

[Vineeth Pillai]

In a multi-processor system, bandwidth usage is divided equally to
all cpus. This causes issues with reclaiming free bandwidth on a cpu.
"Uextra" is same on all cpus in a root domain and running_bw would be
different based on the reserved bandwidth of tasks running on the cpu.
This causes disproportionate reclaiming - task with lesser bandwidth
reclaims less even if its the only task running on that cpu.

Following is a small test with three tasks with reservations (8,10)
(1,10) and (1, 100). These three tasks run on different cpus. But
since the reclamation logic calculates available bandwidth as a factor
of globally available bandwidth, tasks with lesser bandwidth reclaims
only little compared to higher bandwidth even if cpu has free and
available bandwidth to be reclaimed.

TID[730]: RECLAIM=1, (r=8ms, d=10ms, p=10ms), Util: 95.05
TID[731]: RECLAIM=1, (r=1ms, d=10ms, p=10ms), Util: 31.34
TID[732]: RECLAIM=1, (r=1ms, d=100ms, p=100ms), Util: 3.16

Fix: use the available bandwidth on each cpu to calculate reclaimable
bandwidth. Admission control takes care of total bandwidth and hence
using the available bandwidth on a specific cpu would not break the
deadline guarentees.

With this fix, the above test behaves as follows:
TID[586]: RECLAIM=1, (r=1ms, d=100ms, p=100ms), Util: 95.24
TID[585]: RECLAIM=1, (r=1ms, d=10ms, p=10ms), Util: 95.01
TID[584]: RECLAIM=1, (r=8ms, d=10ms, p=10ms), Util: 95.01

Signed-off-by: Vineeth Pillai (Google) <vineeth@xxxxxxxxxxxxxxx>
---
 kernel/sched/deadline.c | 22 +++++++---------------
 1 file changed, 7 insertions(+), 15 deletions(-)

diff --git a/kernel/sched/deadline.c b/kernel/sched/deadline.c
index 91451c1c7e52..85902c4c484b 100644
--- a/kernel/sched/deadline.c
+++ b/kernel/sched/deadline.c
@@ -1272,7 +1272,7 @@ int dl_runtime_exceeded(struct sched_dl_entity *dl_se)
  *	Umax:		Max usable bandwidth for DL. Currently
  *			= sched_rt_runtime_us / sched_rt_period_us
  *	Uextra:		Extra bandwidth not reserved:
- *			= Umax - \Sum(u_i / #cpus in the root domain)
+ *			= Umax - this_bw
  *	u_i:		Bandwidth of an admitted dl task in the
  *			root domain.
  *
@@ -1286,22 +1286,14 @@ int dl_runtime_exceeded(struct sched_dl_entity *dl_se)
  */
 static u64 grub_reclaim(u64 delta, struct rq *rq, struct sched_dl_entity *dl_se)
 {
-	u64 u_act;
-	u64 u_inact = rq->dl.this_bw - rq->dl.running_bw; /* Utot - Uact */
-
 	/*
-	 * Instead of computing max{u, (rq->dl.max_bw - u_inact - u_extra)},
-	 * we compare u_inact + rq->dl.extra_bw with
-	 * rq->dl.max_bw - u, because u_inact + rq->dl.extra_bw can be larger
-	 * than rq->dl.max_bw (so, rq->dl.max_bw - u_inact - rq->dl.extra_bw
-	 * would be negative leading to wrong results)
+	 * max{u, Umax - Uinact - Uextra}
+	 * = max{u, max_bw - (this_bw - running_bw) + (this_bw - running_bw)}
+	 * = max{u, running_bw} = running_bw
+	 * So dq = -(max{u, Umax - Uinact - Uextra} / Umax) dt
+	 *       = -(running_bw / max_bw) dt
 	 */
-	if (u_inact + rq->dl.extra_bw > rq->dl.max_bw - dl_se->dl_bw)
-		u_act = dl_se->dl_bw;
-	else
-		u_act = rq->dl.max_bw - u_inact - rq->dl.extra_bw;
-
-	return div64_u64(delta * u_act, rq->dl.max_bw);
+	return div64_u64(delta * rq->dl.running_bw, rq->dl.max_bw);
 }
 
 /*
-- 
2.40.1