Skip to contentSkip to conversation

xen/sched: migrate timers to correct cpus after suspend

PATCH In review Juergen Gross 3 participants 3 messages Nov 2, 2022 – Nov 4, 2022

lore
Share & tools
lore
#0 lore raw

PATCH v3 1/1 xen/sched: migrate timers to correct cpus after suspend

Today all timers are migrated to cpu 0 when the system is being suspended. They are not migrated back after resuming the system again.

This results (at least) to visible problems with the credit scheduler, as the timer isn't handled on the cpu it was expected to occur, which will result in an ASSERT() triggering. Other more subtle problems, like uninterrupted elongated time slices, are probable. The least effect will be worse performance on cpu 0 resulting from most scheduling related timer interrupts happening there after suspend/resume.

Add migrating the scheduling related timers of a specific cpu from cpu 0 back to its original cpu when that cpu has gone up when resuming the system.

Fixes 0763cd268789 ("xen/sched: don't disable scheduler on cpus during suspend")

---
V2:
- fix smt=0 case (Marek Marczykowski-Górecki)
V3:
- minor comment and commit message adjustments (Jan Beulich)

---
 xen/common/sched/core.c    | 29 +++++++++++++++++++
 xen/common/sched/cpupool.c |  2 ++
 xen/common/sched/credit.c  | 13 +++++++++
 xen/common/sched/private.h | 10 +++++++
 xen/common/sched/rt.c      | 58 ++++++++++++++++++++++++++------------
 5 files changed, 94 insertions(+), 18 deletions(-)
xen/common/sched/core.c +29 -0
index 23fa6845a8..43132ff6e0 100644
@@ -1284,6 +1284,35 @@ static int cpu_disable_scheduler_check(unsigned int cpu)
     return 0;
 }
 
+/*
+ * Called after a cpu has come up again in a suspend/resume cycle.
+ * Migrate all timers for this cpu (they have been migrated to cpu 0 when the
+ * cpu was going down).
+ * Note that only timers related to a physical cpu are migrated, not the ones
+ * related to a vcpu or domain.
+ */
+void sched_migrate_timers(unsigned int cpu)
+{
+    struct sched_resource *sr;
+
+    rcu_read_lock(&sched_res_rculock);
+
+    sr = get_sched_res(cpu);
+
+    /*
+     * Note that on a system with parked cpus (e.g. smt=0 on Intel cpus) this
+     * will be called for the parked cpus, too, so the case for no scheduling
+     * resource being available must be considered.
+     */
+    if ( sr && sr->master_cpu == cpu )
+    {
+        migrate_timer(&sr->s_timer, cpu);
+        sched_move_timers(sr->scheduler, sr);
+    }
+
+    rcu_read_unlock(&sched_res_rculock);
+}
+
 /*
  * In general, this must be called with the scheduler lock held, because the
  * adjust_affinity hook may want to modify the vCPU state. However, when the
xen/common/sched/cpupool.c +2 -0
index b2c6f520c3..bdf6030ab0 100644
@@ -1035,6 +1035,8 @@ static int cf_check cpu_callback(
     case CPU_ONLINE:
         if ( system_state <= SYS_STATE_active )
             rc = cpupool_cpu_add(cpu);
+        else
+            sched_migrate_timers(cpu);
         break;
     case CPU_DOWN_PREPARE:
         /* Suspend/Resume don't change assignments of cpus to cpupools. */
xen/common/sched/credit.c +13 -0
index 47945c2834..f2cd3d9da3 100644
@@ -614,6 +614,18 @@ init_pdata(struct csched_private *prv, struct csched_pcpu *spc, int cpu)
     spc->nr_runnable = 0;
 }
 
+static void cf_check
+csched_move_timers(const struct scheduler *ops, struct sched_resource *sr)
+{
+    struct csched_private *prv = CSCHED_PRIV(ops);
+    struct csched_pcpu *spc = sr->sched_priv;
+
+    if ( sr->master_cpu == prv->master )
+        migrate_timer(&prv->master_ticker, prv->master);
+
+    migrate_timer(&spc->ticker, sr->master_cpu);
+}
+
 /* Change the scheduler of cpu to us (Credit). */
 static spinlock_t *cf_check
 csched_switch_sched(struct scheduler *new_ops, unsigned int cpu,
@@ -2264,6 +2276,7 @@ static const struct scheduler sched_credit_def = {
     .switch_sched   = csched_switch_sched,
     .alloc_domdata  = csched_alloc_domdata,
     .free_domdata   = csched_free_domdata,
+    .move_timers    = csched_move_timers,
 };
 
 REGISTER_SCHEDULER(sched_credit_def);
xen/common/sched/private.h +10 -0
index 0126a4bb9e..0527a8c70d 100644
@@ -331,6 +331,8 @@ struct scheduler {
                                     struct xen_sysctl_scheduler_op *);
     void         (*dump_settings)  (const struct scheduler *);
     void         (*dump_cpu_state) (const struct scheduler *, int);
+    void         (*move_timers)    (const struct scheduler *,
+                                    struct sched_resource *);
 };
 
 static inline int sched_init(struct scheduler *s)
@@ -485,6 +487,13 @@ static inline int sched_adjust_cpupool(const struct scheduler *s,
     return s->adjust_global ? s->adjust_global(s, op) : 0;
 }
 
+static inline void sched_move_timers(const struct scheduler *s,
+                                     struct sched_resource *sr)
+{
+    if ( s->move_timers )
+        s->move_timers(s, sr);
+}
+
 static inline void sched_unit_pause_nosync(const struct sched_unit *unit)
 {
     struct vcpu *v;
@@ -622,6 +631,7 @@ struct cpu_rm_data *alloc_cpu_rm_data(unsigned int cpu, bool aff_alloc);
 void free_cpu_rm_data(struct cpu_rm_data *mem, unsigned int cpu);
 int schedule_cpu_rm(unsigned int cpu, struct cpu_rm_data *mem);
 int sched_move_domain(struct domain *d, struct cpupool *c);
+void sched_migrate_timers(unsigned int cpu);
 struct cpupool *cpupool_get_by_id(unsigned int poolid);
 void cpupool_put(struct cpupool *pool);
 int cpupool_add_domain(struct domain *d, unsigned int poolid);
xen/common/sched/rt.c +40 -18

77 lines not on this page · raw message

Signature · 1 line
2.35.3
Acked-by on v3 1/1
#1 lore raw
Quoted 19 lines · Juergen Gross

Today all timers are migrated to cpu 0 when the system is being suspended. They are not migrated back after resuming the system again.

This results (at least) to visible problems with the credit scheduler,
as the timer isn't handled on the cpu it was expected to occur, which will result in an ASSERT() triggering. Other more subtle problems, like
uninterrupted elongated time slices, are probable. The least effect will be worse performance on cpu 0 resulting from most scheduling related timer interrupts happening there after suspend/resume.

Add migrating the scheduling related timers of a specific cpu from cpu
0 back to its original cpu when that cpu has gone up when resuming the
system.

Quoted 6 lines
Fixes 0763cd268789 ("xen/sched: don't disable scheduler on cpus

during suspend")

Tested-by Marek Marczykowski-Górecki

<[email protected]>

Regards

Signature · 6 lines
Dario Faggioli, Ph.D
http://about.me/dario.faggioli
Virtualization Software Engineer
SUSE Labs, SUSE https://www.suse.com/
-------------------------------------------------------------------
<<This happens because _I_ choose it to happen!>> (Raistlin Majere)
#2 lore raw

Hi Juergen,

Quoted 20 lines · Juergen Gross
Quoted 19 lines

Today all timers are migrated to cpu 0 when the system is being suspended. They are not migrated back after resuming the system again.

This results (at least) to visible problems with the credit scheduler,
as the timer isn't handled on the cpu it was expected to occur, which will result in an ASSERT() triggering. Other more subtle problems, like
uninterrupted elongated time slices, are probable. The least effect will be worse performance on cpu 0 resulting from most scheduling related timer interrupts happening there after suspend/resume.

Add migrating the scheduling related timers of a specific cpu from cpu
0 back to its original cpu when that cpu has gone up when resuming the
system.

Quoted 7 lines
Quoted 6 lines
Fixes 0763cd268789 ("xen/sched: don't disable scheduler on cpus

during suspend")

Tested-by Marek Marczykowski-Górecki

<[email protected]>

Kind regards,
Henry