sched/credit: avoid priority boost for capped domains when unpark

When unpausing a capped domain, the scheduler currently clears the
CSCHED_FLAG_VCPU_PARKED flag before vcpu_wake(). This, in turn, causes the
vcpu_wake to set CSCHED_PRI_TS_BOOST, resulting in an unfair credit boost. The
comment around the changed lines already states that clearing the flag should
happen AFTER the unpause. This bug was introduced in commit be650750945
"credit1: Use atomic bit operations for the flags structure".

Original patch author credit: Xi Xiong while at Amazon.

Signed-off-by: Eslam Elnikety <elnikety@amazon.com>
Reviewed-by: Leonard Foerster <foersleo@amazon.de>
Reviewed-by: Petre Eftime <epetre@amazon.com>
Acked-by: Dario Faggioli <dfaggioli@suse.com>

diff --git a/xen/common/sched_credit.c b/xen/common/sched_credit.c
index 3abe20def8423117f6e04474106f48d34096b84f..7b7facbace07f03d93269393b9be70b2d84a3faf 100644 (file)
--- a/xen/common/sched_credit.c
+++ b/xen/common/sched_credit.c
@@ -1538,7 +1538,7 @@ csched_acct(void* dummy)
                 svc->pri = CSCHED_PRI_TS_UNDER;
 
                 /* Unpark any capped domains whose credits go positive */
-                if ( test_and_clear_bit(CSCHED_FLAG_VCPU_PARKED, &svc->flags) )
+                if ( test_bit(CSCHED_FLAG_VCPU_PARKED, &svc->flags) )
                 {
                     /*
                      * It's important to unset the flag AFTER the unpause()
@@ -1547,6 +1547,7 @@ csched_acct(void* dummy)
                      */
                     SCHED_STAT_CRANK(vcpu_unpark);
                     vcpu_unpause(svc->vcpu);
+                    clear_bit(CSCHED_FLAG_VCPU_PARKED, &svc->flags);
                 }
 
                 /* Upper bound on credits means VCPU stops earning */