credit2: avoid vCPUs to ever reach lower credits than idle

There have been report of stalls of guest vCPUs, when Credit2 was used.
It seemed like these vCPUs were not getting scheduled for very long
time, even under light load conditions (e.g., during dom0 boot).

Investigations led to the discovery that --although rarely-- it can
happen that a vCPU manages to run for very long timeslices. In Credit2,
this means that, when runtime accounting happens, the vCPU will lose a
large quantity of credits. This in turn may lead to the vCPU having less
credits than the idle vCPUs (-2^30). At this point, the scheduler will
pick the idle vCPU, instead of the ready to run vCPU, for a few
"epochs", which often times is enough for the guest kernel to think the
vCPU is not responding and crashing.

An example of this situation is shown here. In fact, we can see d0v1
sitting in the runqueue while all the CPUs are idle, as it has
-1254238270 credits, which is smaller than -2^30 = −1073741824:

    (XEN) Runqueue 0:
    (XEN)   ncpus              = 28
    (XEN)   cpus               = 0-27
    (XEN)   max_weight         = 256
    (XEN)   pick_bias          = 22
    (XEN)   instload           = 1
    (XEN)   aveload            = 293391 (~111%)
    (XEN)   idlers: 00,00000000,00000000,00000000,00000000,00000000,0fffffff
    (XEN)   tickled: 00,00000000,00000000,00000000,00000000,00000000,00000000
    (XEN)   fully idle cores: 00,00000000,00000000,00000000,00000000,00000000,0fffffff
    [...]
    (XEN) Runqueue 0:
    (XEN) CPU[00] runq=0, sibling=00,..., core=00,...
    (XEN) CPU[01] runq=0, sibling=00,..., core=00,...
    [...]
    (XEN) CPU[26] runq=0, sibling=00,..., core=00,...
    (XEN) CPU[27] runq=0, sibling=00,..., core=00,...
    (XEN) RUNQ:
    (XEN)     0: [0.1] flags=0 cpu=5 credit=-1254238270 [w=256] load=262144 (~100%)

We certainly don't want, under any circumstance, this to happen.
Let's, therefore, define a minimum amount of credits a vCPU can have.
During accounting, we make sure that, for however long the vCPU has
run, it will never get to have less than such minimum amount of
credits. Then, we set the credits of the idle vCPU to an even
smaller value.

NOTE: investigations have been done about _how_ it is possible for a
vCPU to execute for so much time that its credits becomes so low. While
still not completely clear, there are evidence that:
- it only happens very rarely,
- it appears to be both machine and workload specific,
- it does not look to be a Credit2 (e.g., as it happens when
  running with Credit1 as well) issue, or a scheduler issue.

This patch makes Credit2 more robust to events like this, whatever
the cause is, and should hence be backported (as far as possible).

Reported-by: Glen <glenbarney@gmail.com>
Reported-by: Tomas Mozes <hydrapolic@gmail.com>
Signed-off-by: Dario Faggioli <dfaggioli@suse.com>
Reviewed-by: George Dunlap <george.dunlap@citrix.com>

diff --git a/tools/xentrace/formats b/tools/xentrace/formats
index d6e7e3f80089c0abe7cd51b706bc8e4248432423..8f126f65f1243dc9034bf01ba7aaf818e7cca908 100644 (file)
--- a/tools/xentrace/formats
+++ b/tools/xentrace/formats
@@ -55,7 +55,7 @@
 0x00022204  CPU%(cpu)d  %(tsc)d (+%(reltsc)8d)  csched2:credit_add
 0x00022205  CPU%(cpu)d  %(tsc)d (+%(reltsc)8d)  csched2:tickle_check   [ dom:vcpu = 0x%(1)08x, credit = %(2)d, score = %(3)d ]
 0x00022206  CPU%(cpu)d  %(tsc)d (+%(reltsc)8d)  csched2:tickle         [ cpu = %(1)d ]
-0x00022207  CPU%(cpu)d  %(tsc)d (+%(reltsc)8d)  csched2:credit_reset   [ dom:vcpu = 0x%(1)08x, cr_start = %(2)d, cr_end = %(3)d, mult = %(4)d ]
+0x00022207  CPU%(cpu)d  %(tsc)d (+%(reltsc)8d)  csched2:credit_reset   [ dom:vcpu = 0x%(1)08x, cr_start = %(2)d, cr_end = %(3)d ]
 0x00022208  CPU%(cpu)d  %(tsc)d (+%(reltsc)8d)  csched2:sched_tasklet
 0x00022209  CPU%(cpu)d  %(tsc)d (+%(reltsc)8d)  csched2:update_load
 0x0002220a  CPU%(cpu)d  %(tsc)d (+%(reltsc)8d)  csched2:runq_assign    [ dom:vcpu = 0x%(1)08x, rq_id = %(2)d ]

diff --git a/tools/xentrace/xenalyze.c b/tools/xentrace/xenalyze.c
index aa894673ad0697bbc12d3c4b992693d75f4b03e5..d3c8368e9d779e0371029b5ca677d550f0ae830f 100644 (file)
--- a/tools/xentrace/xenalyze.c
+++ b/tools/xentrace/xenalyze.c
@@ -7718,13 +7718,12 @@ void sched_process(struct pcpu_info *p)
                 struct {
                     unsigned int vcpuid:16, domid:16;
                     int credit_start, credit_end;
-                    unsigned int multiplier;
                 } *r = (typeof(r))ri->d;
 
                 printf(" %s csched2:reset_credits d%uv%u, "
-                       "credit_start = %d, credit_end = %d, mult = %u\n",
+                       "credit_start = %d, credit_end = %d\n",
                        ri->dump_header, r->domid, r->vcpuid,
-                       r->credit_start, r->credit_end, r->multiplier);
+                       r->credit_start, r->credit_end);
             }
             break;
         case TRC_SCHED_CLASS_EVT(CSCHED2, 8):  /* SCHED_TASKLET    */

diff --git a/xen/common/sched/credit2.c b/xen/common/sched/credit2.c
index c7241944a87294e8ee1deba920c50d2f358953c2..cbf9ce2b97dd41d7c3776dddf8dfad1dcc745662 100644 (file)
--- a/xen/common/sched/credit2.c
+++ b/xen/common/sched/credit2.c
@@ -229,12 +229,22 @@
  * before a reset event.
  */
 #define CSCHED2_CREDIT_INIT          MILLISECS(10)
+/*
+ * Minimum amount of credits VMs can have. Ideally, no VM would get
+ * close to this (unless a vCPU manages to execute for really long
+ * time uninterrupted). In case it happens, it makes no sense to
+ * track even deeper undershoots.
+ *
+ * NOTE: If making this smaller than -CSCHED2_CREDIT_INIT, adjust
+ * reset_credit() accordingly.
+ */
+#define CSCHED2_CREDIT_MIN           (-CSCHED2_CREDIT_INIT)
 /*
  * Amount of credit the idle units have. It never changes, as idle
  * units does not consume credits, and it must be lower than whatever
  * amount of credit 'regular' unit would end up with.
  */
-#define CSCHED2_IDLE_CREDIT          (-(1U<<30))
+#define CSCHED2_IDLE_CREDIT          (CSCHED2_CREDIT_MIN-1)
 /*
  * Carryover: How much "extra" credit may be carried over after
  * a reset.
@@ -781,10 +791,15 @@ static int get_fallback_cpu(struct csched2_unit *svc)
 static void t2c_update(const struct csched2_runqueue_data *rqd, s_time_t time,
                           struct csched2_unit *svc)
 {
-    uint64_t val = time * rqd->max_weight + svc->residual;
+    int64_t val = time * rqd->max_weight + svc->residual;
 
     svc->residual = do_div(val, svc->weight);
-    svc->credit -= val;
+    /* Getting to lower credit than CSCHED2_CREDIT_MIN makes no sense. */
+    val = svc->credit - val;
+    if ( unlikely(val < CSCHED2_CREDIT_MIN) )
+        svc->credit = CSCHED2_CREDIT_MIN;
+    else
+        svc->credit = val;
 }
 
 static s_time_t c2t(const struct csched2_runqueue_data *rqd, s_time_t credit,
@@ -1616,28 +1631,25 @@ static void reset_credit(int cpu, s_time_t now, struct csched2_unit *snext)
 {
     struct csched2_runqueue_data *rqd = c2rqd(cpu);
     struct list_head *iter;
-    int m;
+    int reset = CSCHED2_CREDIT_INIT;
 
     /*
      * Under normal circumstances, snext->credit should never be less
      * than -CSCHED2_MIN_TIMER.  However, under some circumstances, an
      * unit with low credits may be allowed to run long enough that
-     * its credits are actually less than -CSCHED2_CREDIT_INIT.
+     * its credits are actually much lower than that.
      * (Instances have been observed, for example, where an unit with
      * 200us of credit was allowed to run for 11ms, giving it -10.8ms
      * of credit.  Thus it was still negative even after the reset.)
      *
      * If this is the case for snext, we simply want to keep moving
-     * everyone up until it is in the black again.  This fair because
-     * none of the other units want to run at the moment.
-     *
-     * Rather than looping, however, we just calculate a multiplier,
-     * avoiding an integer division and multiplication in the common
-     * case.
+     * everyone up until it is in the black again. This means that,
+     * since CSCHED2_CREDIT_MIN is -CSCHED2_CREDIT_INIT, we need to
+     * actually add 2*CSCHED2_CREDIT_INIT.
      */
-    m = 1;
-    if ( snext->credit < -CSCHED2_CREDIT_INIT )
-        m += (-snext->credit) / CSCHED2_CREDIT_INIT;
+    ASSERT(snext->credit >= CSCHED2_CREDIT_MIN);
+    if ( unlikely(snext->credit == CSCHED2_CREDIT_MIN) )
+        reset += CSCHED2_CREDIT_INIT;
 
     list_for_each( iter, &rqd->svc )
     {
@@ -1668,15 +1680,7 @@ static void reset_credit(int cpu, s_time_t now, struct csched2_unit *snext)
         }
 
         start_credit = svc->credit;
-
-        /*
-         * Add INIT * m, avoiding integer multiplication in the common case.
-         */
-        if ( likely(m==1) )
-            svc->credit += CSCHED2_CREDIT_INIT;
-        else
-            svc->credit += m * CSCHED2_CREDIT_INIT;
-
+        svc->credit += reset;
         /* "Clip" credits to max carryover */
         if ( svc->credit > CSCHED2_CREDIT_INIT + CSCHED2_CARRYOVER_MAX )
             svc->credit = CSCHED2_CREDIT_INIT + CSCHED2_CARRYOVER_MAX;
@@ -1688,19 +1692,18 @@ static void reset_credit(int cpu, s_time_t now, struct csched2_unit *snext)
             struct {
                 unsigned unit:16, dom:16;
                 int credit_start, credit_end;
-                unsigned multiplier;
             } d;
             d.dom = svc->unit->domain->domain_id;
             d.unit = svc->unit->unit_id;
             d.credit_start = start_credit;
             d.credit_end = svc->credit;
-            d.multiplier = m;
             __trace_var(TRC_CSCHED2_CREDIT_RESET, 1,
                         sizeof(d),
                         (unsigned char *)&d);
         }
     }
 
+    ASSERT(snext->credit > 0);
     SCHED_STAT_CRANK(credit_reset);
 
     /* No need to resort runqueue, as everyone's order should be the same. */