[PATCH] sched: RCU domains
authorNick Piggin <nickpiggin@yahoo.com.au>
Sat, 25 Jun 2005 21:57:27 +0000 (14:57 -0700)
committerLinus Torvalds <torvalds@ppc970.osdl.org>
Sat, 25 Jun 2005 23:24:44 +0000 (16:24 -0700)
One of the problems with the multilevel balance-on-fork/exec is that it needs
to jump through hoops to satisfy sched-domain's locking semantics (that is,
you may traverse your own domain when not preemptable, and you may traverse
others' domains when holding their runqueue lock).

balance-on-exec had to potentially migrate between more than one CPU before
finding a final CPU to migrate to, and balance-on-fork needed to potentially
take multiple runqueue locks.

So bite the bullet and make sched-domains go completely RCU.  This actually
simplifies the code quite a bit.

From: Ingo Molnar <mingo@elte.hu>

schedstats RCU fix, and a nice comment on for_each_domain, from Ingo.

Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Nick Piggin <nickpiggin@yahoo.com.au>
Acked-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
kernel/sched.c

index ef32389ee7689ed28d2f30b257a62736028237f4..54ce787b6207ff8077b1ab68308e2bedf5fcb9e4 100644 (file)
@@ -260,8 +260,15 @@ struct runqueue {
 
 static DEFINE_PER_CPU(struct runqueue, runqueues);
 
+/*
+ * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
+ * See update_sched_domains: synchronize_kernel for details.
+ *
+ * The domain tree of any CPU may only be accessed from within
+ * preempt-disabled sections.
+ */
 #define for_each_domain(cpu, domain) \
-       for (domain = cpu_rq(cpu)->sd; domain; domain = domain->parent)
+for (domain = rcu_dereference(cpu_rq(cpu)->sd); domain; domain = domain->parent)
 
 #define cpu_rq(cpu)            (&per_cpu(runqueues, (cpu)))
 #define this_rq()              (&__get_cpu_var(runqueues))
@@ -395,6 +402,7 @@ static int show_schedstat(struct seq_file *seq, void *v)
 
 #ifdef CONFIG_SMP
                /* domain-specific stats */
+               preempt_disable();
                for_each_domain(cpu, sd) {
                        enum idle_type itype;
                        char mask_str[NR_CPUS];
@@ -419,6 +427,7 @@ static int show_schedstat(struct seq_file *seq, void *v)
                            sd->sbf_cnt, sd->sbf_balanced, sd->sbf_pushed,
                            sd->ttwu_wake_remote, sd->ttwu_move_affine, sd->ttwu_move_balance);
                }
+               preempt_enable();
 #endif
        }
        return 0;
@@ -824,22 +833,12 @@ inline int task_curr(const task_t *p)
 }
 
 #ifdef CONFIG_SMP
-enum request_type {
-       REQ_MOVE_TASK,
-       REQ_SET_DOMAIN,
-};
-
 typedef struct {
        struct list_head list;
-       enum request_type type;
 
-       /* For REQ_MOVE_TASK */
        task_t *task;
        int dest_cpu;
 
-       /* For REQ_SET_DOMAIN */
-       struct sched_domain *sd;
-
        struct completion done;
 } migration_req_t;
 
@@ -861,7 +860,6 @@ static int migrate_task(task_t *p, int dest_cpu, migration_req_t *req)
        }
 
        init_completion(&req->done);
-       req->type = REQ_MOVE_TASK;
        req->task = p;
        req->dest_cpu = dest_cpu;
        list_add(&req->list, &rq->migration_queue);
@@ -4378,17 +4376,9 @@ static int migration_thread(void * data)
                req = list_entry(head->next, migration_req_t, list);
                list_del_init(head->next);
 
-               if (req->type == REQ_MOVE_TASK) {
-                       spin_unlock(&rq->lock);
-                       __migrate_task(req->task, cpu, req->dest_cpu);
-                       local_irq_enable();
-               } else if (req->type == REQ_SET_DOMAIN) {
-                       rq->sd = req->sd;
-                       spin_unlock_irq(&rq->lock);
-               } else {
-                       spin_unlock_irq(&rq->lock);
-                       WARN_ON(1);
-               }
+               spin_unlock(&rq->lock);
+               __migrate_task(req->task, cpu, req->dest_cpu);
+               local_irq_enable();
 
                complete(&req->done);
        }
@@ -4619,7 +4609,6 @@ static int migration_call(struct notifier_block *nfb, unsigned long action,
                        migration_req_t *req;
                        req = list_entry(rq->migration_queue.next,
                                         migration_req_t, list);
-                       BUG_ON(req->type != REQ_MOVE_TASK);
                        list_del_init(&req->list);
                        complete(&req->done);
                }
@@ -4800,10 +4789,7 @@ static int __devinit sd_parent_degenerate(struct sched_domain *sd,
  */
 void __devinit cpu_attach_domain(struct sched_domain *sd, int cpu)
 {
-       migration_req_t req;
-       unsigned long flags;
        runqueue_t *rq = cpu_rq(cpu);
-       int local = 1;
        struct sched_domain *tmp;
 
        /* Remove the sched domains which do not contribute to scheduling. */
@@ -4820,24 +4806,7 @@ void __devinit cpu_attach_domain(struct sched_domain *sd, int cpu)
 
        sched_domain_debug(sd, cpu);
 
-       spin_lock_irqsave(&rq->lock, flags);
-
-       if (cpu == smp_processor_id() || !cpu_online(cpu)) {
-               rq->sd = sd;
-       } else {
-               init_completion(&req.done);
-               req.type = REQ_SET_DOMAIN;
-               req.sd = sd;
-               list_add(&req.list, &rq->migration_queue);
-               local = 0;
-       }
-
-       spin_unlock_irqrestore(&rq->lock, flags);
-
-       if (!local) {
-               wake_up_process(rq->migration_thread);
-               wait_for_completion(&req.done);
-       }
+       rcu_assign_pointer(rq->sd, sd);
 }
 
 /* cpus with isolated domains */
@@ -5112,6 +5081,7 @@ static int update_sched_domains(struct notifier_block *nfb,
        case CPU_DOWN_PREPARE:
                for_each_online_cpu(i)
                        cpu_attach_domain(NULL, i);
+               synchronize_kernel();
                arch_destroy_sched_domains();
                return NOTIFY_OK;