[PATCH v3 2/2] xen: credit1: avoid boosting vCPUs being "just" migrated

[Dario Faggioli <dario.faggioli@citrix.com>]

Moving a vCPU to a different pCPU means offlining it and
then waking it up, on the new pCPU. Credit1 grants BOOST
priority to vCPUs that wakes up, with the aim of improving
I/O latency. The net effect of this all is that vCPUs get
boosted when migrating, which shouldn't happen.

For instance, this causes scheduling anomalies and,
potentially, performance problems, as reported here:
  http://lists.xen.org/archives/html/xen-devel/2015-10/msg02851.html

This patch fixes this by noting down (by means of a flag)
the fact that the vCPU is about to undergo a migration.
This way we can tell, later, during a wakeup, whether the
vCPU is migrating or unblocking, and decide whether or
not to apply the boosting.

Note that it is important that atomic-safe bit operations
are used when manipulating vCPUs' flags. Take the chance
and add a comment about this.

Signed-off-by: Dario Faggioli <dario.faggioli@citrix.com>
---
Cc: George Dunlap <george.dunlap@eu.citrix.com>
Cc: Jan Beulich <JBeulich@suse.com>
---
Changes from v2:
 * test_and_clear() is necessary when accessing svc->flags;
 * added a comment about such need at the top, where the flags
   are defined.

Changes from v1:
 * rewritten, following suggestion got during review: there
   are no wakeup flags any longer, and all is done in sched_credit.c
   by setting a flag in csched_cpu_pick() and testing (and
   cleating) it in csched_vcpu_wake().
---
 xen/common/sched_credit.c |   30 ++++++++++++++++++++++++++----
 1 file changed, 26 insertions(+), 4 deletions(-)

diff --git a/xen/common/sched_credit.c b/xen/common/sched_credit.c
index 5708701..597a784 100644
--- a/xen/common/sched_credit.c
+++ b/xen/common/sched_credit.c
@@ -63,9 +63,14 @@
 
 /*
  * Flags
+ *
+ * Note that svc->flags (where these flags live) is protected by an
+ * inconsistent set of locks. Therefore atomic-safe bit operations must
+ * be used for accessing it.
  */
 #define CSCHED_FLAG_VCPU_PARKED    0x0  /* VCPU over capped credits */
 #define CSCHED_FLAG_VCPU_YIELD     0x1  /* VCPU yielding */
+#define CSCHED_FLAG_VCPU_MIGRATING 0x2  /* VCPU may have moved to a new pcpu */
 
 
 /*
@@ -787,6 +792,16 @@ _csched_cpu_pick(const struct scheduler *ops, struct vcpu *vc, bool_t commit)
 static int
 csched_cpu_pick(const struct scheduler *ops, struct vcpu *vc)
 {
+    struct csched_vcpu *svc = CSCHED_VCPU(vc);
+
+    /*
+     * We have been called by vcpu_migrate() (in schedule.c), as part
+     * of the process of seeing if vc can be migrated to another pcpu.
+     * We make a note about this in svc->flags so that later, in
+     * csched_vcpu_wake() (still called from vcpu_migrate()) we won't
+     * get boosted, which we don't deserve as we are "only" migrating.
+     */
+    set_bit(CSCHED_FLAG_VCPU_MIGRATING, &svc->flags);
     return _csched_cpu_pick(ops, vc, 1);
 }
 
@@ -1022,11 +1037,18 @@ csched_vcpu_wake(const struct scheduler *ops, struct vcpu *vc)
      * more CPU resource intensive VCPUs without impacting overall 
      * system fairness.
      *
-     * The one exception is for VCPUs of capped domains unpausing
-     * after earning credits they had overspent. We don't boost
-     * those.
+     * There are two cases, when we don't want to boost:
+     *  - VCPUs that are waking up after a migration, rather than
+     *    after having block;
+     *  - VCPUs of capped domains unpausing after earning credits
+     *    they had overspent.
+     *
+     * Note that checking whether we are "only" migrating must be
+     * done up front, as we do not want the clearing of the bit we
+     * set in csched_cpu_pick() to be short-circuited away.
      */
-    if ( svc->pri == CSCHED_PRI_TS_UNDER &&
+    if ( !test_and_clear_bit(CSCHED_FLAG_VCPU_MIGRATING, &svc->flags)  &&
+         svc->pri == CSCHED_PRI_TS_UNDER &&
          !test_bit(CSCHED_FLAG_VCPU_PARKED, &svc->flags) )
     {
         TRACE_2D(TRC_CSCHED_BOOST_START, vc->domain->domain_id, vc->vcpu_id);