rcu: Move quiescent-state forcing into kthread
authorPaul E. McKenney <paulmck@linux.vnet.ibm.com>
Sat, 23 Jun 2012 00:06:26 +0000 (17:06 -0700)
committerPaul E. McKenney <paulmck@linux.vnet.ibm.com>
Sun, 23 Sep 2012 14:41:54 +0000 (07:41 -0700)
As the first step towards allowing quiescent-state forcing to be
preemptible, this commit moves RCU quiescent-state forcing into the
same kthread that is now used to initialize and clean up after grace
periods.  This is yet another step towards keeping scheduling
latency down to a dull roar.

Updated to change from raw_spin_lock_irqsave() to raw_spin_lock_irq()
and to remove the now-unused rcu_state structure fields as suggested by
Peter Zijlstra.

Reported-by: Mike Galbraith <mgalbraith@suse.de>
Reported-by: Dimitri Sivanich <sivanich@sgi.com>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
kernel/rcutree.c
kernel/rcutree.h
kernel/rcutree_plugin.h

index 340a5f54b6aff814c0edfbea4aa30af0fd15cda7..6182686de4a671513b1a3039943b05c3621f2a12 100644 (file)
@@ -72,7 +72,6 @@ static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
        .orphan_nxttail = &sname##_state.orphan_nxtlist, \
        .orphan_donetail = &sname##_state.orphan_donelist, \
        .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
-       .fqslock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.fqslock), \
        .name = #sname, \
 }
 
@@ -226,7 +225,8 @@ int rcu_cpu_stall_timeout __read_mostly = CONFIG_RCU_CPU_STALL_TIMEOUT;
 module_param(rcu_cpu_stall_suppress, int, 0644);
 module_param(rcu_cpu_stall_timeout, int, 0644);
 
-static void force_quiescent_state(struct rcu_state *rsp, int relaxed);
+static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *));
+static void force_quiescent_state(struct rcu_state *rsp);
 static int rcu_pending(int cpu);
 
 /*
@@ -252,7 +252,7 @@ EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
  */
 void rcu_bh_force_quiescent_state(void)
 {
-       force_quiescent_state(&rcu_bh_state, 0);
+       force_quiescent_state(&rcu_bh_state);
 }
 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
 
@@ -286,7 +286,7 @@ EXPORT_SYMBOL_GPL(rcutorture_record_progress);
  */
 void rcu_sched_force_quiescent_state(void)
 {
-       force_quiescent_state(&rcu_sched_state, 0);
+       force_quiescent_state(&rcu_sched_state);
 }
 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
 
@@ -784,11 +784,11 @@ static void print_other_cpu_stall(struct rcu_state *rsp)
        else if (!trigger_all_cpu_backtrace())
                dump_stack();
 
-       /* If so configured, complain about tasks blocking the grace period. */
+       /* Complain about tasks blocking the grace period. */
 
        rcu_print_detail_task_stall(rsp);
 
-       force_quiescent_state(rsp, 0);  /* Kick them all. */
+       force_quiescent_state(rsp);  /* Kick them all. */
 }
 
 static void print_cpu_stall(struct rcu_state *rsp)
@@ -1036,7 +1036,7 @@ static int rcu_gp_init(struct rcu_state *rsp)
        struct rcu_node *rnp = rcu_get_root(rsp);
 
        raw_spin_lock_irq(&rnp->lock);
-       rsp->gp_flags = 0;
+       rsp->gp_flags = 0; /* Clear all flags: New grace period. */
 
        if (rcu_gp_in_progress(rsp)) {
                /* Grace period already in progress, don't start another.  */
@@ -1044,22 +1044,9 @@ static int rcu_gp_init(struct rcu_state *rsp)
                return 0;
        }
 
-       if (rsp->fqs_active) {
-               /*
-                * We need a grace period, but force_quiescent_state()
-                * is running.  Tell it to start one on our behalf.
-                */
-               rsp->fqs_need_gp = 1;
-               raw_spin_unlock_irq(&rnp->lock);
-               return 0;
-       }
-
        /* Advance to a new grace period and initialize state. */
        rsp->gpnum++;
        trace_rcu_grace_period(rsp->name, rsp->gpnum, "start");
-       WARN_ON_ONCE(rsp->fqs_state == RCU_GP_INIT);
-       rsp->fqs_state = RCU_GP_INIT; /* Stop force_quiescent_state. */
-       rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
        record_gp_stall_check_time(rsp);
        raw_spin_unlock_irq(&rnp->lock);
 
@@ -1096,19 +1083,40 @@ static int rcu_gp_init(struct rcu_state *rsp)
                cond_resched();
        }
 
-       rnp = rcu_get_root(rsp);
-       raw_spin_lock_irq(&rnp->lock);
-       /* force_quiescent_state() now OK. */
-       rsp->fqs_state = RCU_SIGNAL_INIT;
-       raw_spin_unlock_irq(&rnp->lock);
        put_online_cpus();
        return 1;
 }
 
+/*
+ * Do one round of quiescent-state forcing.
+ */
+int rcu_gp_fqs(struct rcu_state *rsp, int fqs_state_in)
+{
+       int fqs_state = fqs_state_in;
+       struct rcu_node *rnp = rcu_get_root(rsp);
+
+       rsp->n_force_qs++;
+       if (fqs_state == RCU_SAVE_DYNTICK) {
+               /* Collect dyntick-idle snapshots. */
+               force_qs_rnp(rsp, dyntick_save_progress_counter);
+               fqs_state = RCU_FORCE_QS;
+       } else {
+               /* Handle dyntick-idle and offline CPUs. */
+               force_qs_rnp(rsp, rcu_implicit_dynticks_qs);
+       }
+       /* Clear flag to prevent immediate re-entry. */
+       if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
+               raw_spin_lock_irq(&rnp->lock);
+               rsp->gp_flags &= ~RCU_GP_FLAG_FQS;
+               raw_spin_unlock_irq(&rnp->lock);
+       }
+       return fqs_state;
+}
+
 /*
  * Clean up after the old grace period.
  */
-static int rcu_gp_cleanup(struct rcu_state *rsp)
+static void rcu_gp_cleanup(struct rcu_state *rsp)
 {
        unsigned long gp_duration;
        struct rcu_data *rdp;
@@ -1160,7 +1168,6 @@ static int rcu_gp_cleanup(struct rcu_state *rsp)
        if (cpu_needs_another_gp(rsp, rdp))
                rsp->gp_flags = 1;
        raw_spin_unlock_irq(&rnp->lock);
-       return 1;
 }
 
 /*
@@ -1168,6 +1175,8 @@ static int rcu_gp_cleanup(struct rcu_state *rsp)
  */
 static int __noreturn rcu_gp_kthread(void *arg)
 {
+       int fqs_state;
+       int ret;
        struct rcu_state *rsp = arg;
        struct rcu_node *rnp = rcu_get_root(rsp);
 
@@ -1175,26 +1184,43 @@ static int __noreturn rcu_gp_kthread(void *arg)
 
                /* Handle grace-period start. */
                for (;;) {
-                       wait_event_interruptible(rsp->gp_wq, rsp->gp_flags);
-                       if (rsp->gp_flags && rcu_gp_init(rsp))
+                       wait_event_interruptible(rsp->gp_wq,
+                                                rsp->gp_flags &
+                                                RCU_GP_FLAG_INIT);
+                       if ((rsp->gp_flags & RCU_GP_FLAG_INIT) &&
+                           rcu_gp_init(rsp))
                                break;
                        cond_resched();
                        flush_signals(current);
                }
 
-               /* Handle grace-period end. */
-               rnp = rcu_get_root(rsp);
+               /* Handle quiescent-state forcing. */
+               fqs_state = RCU_SAVE_DYNTICK;
                for (;;) {
-                       wait_event_interruptible(rsp->gp_wq,
-                                                !ACCESS_ONCE(rnp->qsmask) &&
-                                                !rcu_preempt_blocked_readers_cgp(rnp));
+                       rsp->jiffies_force_qs = jiffies +
+                                               RCU_JIFFIES_TILL_FORCE_QS;
+                       ret = wait_event_interruptible_timeout(rsp->gp_wq,
+                                       (rsp->gp_flags & RCU_GP_FLAG_FQS) ||
+                                       (!ACCESS_ONCE(rnp->qsmask) &&
+                                        !rcu_preempt_blocked_readers_cgp(rnp)),
+                                       RCU_JIFFIES_TILL_FORCE_QS);
+                       /* If grace period done, leave loop. */
                        if (!ACCESS_ONCE(rnp->qsmask) &&
-                           !rcu_preempt_blocked_readers_cgp(rnp) &&
-                           rcu_gp_cleanup(rsp))
+                           !rcu_preempt_blocked_readers_cgp(rnp))
                                break;
-                       cond_resched();
-                       flush_signals(current);
+                       /* If time for quiescent-state forcing, do it. */
+                       if (ret == 0 || (rsp->gp_flags & RCU_GP_FLAG_FQS)) {
+                               fqs_state = rcu_gp_fqs(rsp, fqs_state);
+                               cond_resched();
+                       } else {
+                               /* Deal with stray signal. */
+                               cond_resched();
+                               flush_signals(current);
+                       }
                }
+
+               /* Handle grace-period end. */
+               rcu_gp_cleanup(rsp);
        }
 }
 
@@ -1226,7 +1252,7 @@ rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
                return;
        }
 
-       rsp->gp_flags = 1;
+       rsp->gp_flags = RCU_GP_FLAG_INIT;
        raw_spin_unlock_irqrestore(&rnp->lock, flags);
        wake_up(&rsp->gp_wq);
 }
@@ -1777,72 +1803,20 @@ static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *))
  * Force quiescent states on reluctant CPUs, and also detect which
  * CPUs are in dyntick-idle mode.
  */
-static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
+static void force_quiescent_state(struct rcu_state *rsp)
 {
        unsigned long flags;
        struct rcu_node *rnp = rcu_get_root(rsp);
 
-       trace_rcu_utilization("Start fqs");
-       if (!rcu_gp_in_progress(rsp)) {
-               trace_rcu_utilization("End fqs");
-               return;  /* No grace period in progress, nothing to force. */
-       }
-       if (!raw_spin_trylock_irqsave(&rsp->fqslock, flags)) {
+       if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS)
+               return;  /* Someone beat us to it. */
+       if (!raw_spin_trylock_irqsave(&rnp->lock, flags)) {
                rsp->n_force_qs_lh++; /* Inexact, can lose counts.  Tough! */
-               trace_rcu_utilization("End fqs");
-               return; /* Someone else is already on the job. */
-       }
-       if (relaxed && ULONG_CMP_GE(rsp->jiffies_force_qs, jiffies))
-               goto unlock_fqs_ret; /* no emergency and done recently. */
-       rsp->n_force_qs++;
-       raw_spin_lock(&rnp->lock);  /* irqs already disabled */
-       rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
-       if(!rcu_gp_in_progress(rsp)) {
-               rsp->n_force_qs_ngp++;
-               raw_spin_unlock(&rnp->lock);  /* irqs remain disabled */
-               goto unlock_fqs_ret;  /* no GP in progress, time updated. */
-       }
-       rsp->fqs_active = 1;
-       switch (rsp->fqs_state) {
-       case RCU_GP_IDLE:
-       case RCU_GP_INIT:
-
-               break; /* grace period idle or initializing, ignore. */
-
-       case RCU_SAVE_DYNTICK:
-
-               raw_spin_unlock(&rnp->lock);  /* irqs remain disabled */
-
-               /* Record dyntick-idle state. */
-               force_qs_rnp(rsp, dyntick_save_progress_counter);
-               raw_spin_lock(&rnp->lock);  /* irqs already disabled */
-               if (rcu_gp_in_progress(rsp))
-                       rsp->fqs_state = RCU_FORCE_QS;
-               break;
-
-       case RCU_FORCE_QS:
-
-               /* Check dyntick-idle state, send IPI to laggarts. */
-               raw_spin_unlock(&rnp->lock);  /* irqs remain disabled */
-               force_qs_rnp(rsp, rcu_implicit_dynticks_qs);
-
-               /* Leave state in case more forcing is required. */
-
-               raw_spin_lock(&rnp->lock);  /* irqs already disabled */
-               break;
-       }
-       rsp->fqs_active = 0;
-       if (rsp->fqs_need_gp) {
-               raw_spin_unlock(&rsp->fqslock); /* irqs remain disabled */
-               rsp->fqs_need_gp = 0;
-               rcu_start_gp(rsp, flags); /* releases rnp->lock */
-               trace_rcu_utilization("End fqs");
                return;
        }
-       raw_spin_unlock(&rnp->lock);  /* irqs remain disabled */
-unlock_fqs_ret:
-       raw_spin_unlock_irqrestore(&rsp->fqslock, flags);
-       trace_rcu_utilization("End fqs");
+       rsp->gp_flags |= RCU_GP_FLAG_FQS;
+       raw_spin_unlock_irqrestore(&rnp->lock, flags);
+       wake_up(&rsp->gp_wq);  /* Memory barrier implied by wake_up() path. */
 }
 
 /*
@@ -1858,13 +1832,6 @@ __rcu_process_callbacks(struct rcu_state *rsp)
 
        WARN_ON_ONCE(rdp->beenonline == 0);
 
-       /*
-        * If an RCU GP has gone long enough, go check for dyntick
-        * idle CPUs and, if needed, send resched IPIs.
-        */
-       if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
-               force_quiescent_state(rsp, 1);
-
        /*
         * Advance callbacks in response to end of earlier grace
         * period that some other CPU ended.
@@ -1965,12 +1932,11 @@ static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp,
                        rdp->blimit = LONG_MAX;
                        if (rsp->n_force_qs == rdp->n_force_qs_snap &&
                            *rdp->nxttail[RCU_DONE_TAIL] != head)
-                               force_quiescent_state(rsp, 0);
+                               force_quiescent_state(rsp);
                        rdp->n_force_qs_snap = rsp->n_force_qs;
                        rdp->qlen_last_fqs_check = rdp->qlen;
                }
-       } else if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
-               force_quiescent_state(rsp, 1);
+       }
 }
 
 static void
@@ -2251,17 +2217,7 @@ static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
        /* Is the RCU core waiting for a quiescent state from this CPU? */
        if (rcu_scheduler_fully_active &&
            rdp->qs_pending && !rdp->passed_quiesce) {
-
-               /*
-                * If force_quiescent_state() coming soon and this CPU
-                * needs a quiescent state, and this is either RCU-sched
-                * or RCU-bh, force a local reschedule.
-                */
                rdp->n_rp_qs_pending++;
-               if (!rdp->preemptible &&
-                   ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs) - 1,
-                                jiffies))
-                       set_need_resched();
        } else if (rdp->qs_pending && rdp->passed_quiesce) {
                rdp->n_rp_report_qs++;
                return 1;
@@ -2291,13 +2247,6 @@ static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
                return 1;
        }
 
-       /* Has an RCU GP gone long enough to send resched IPIs &c? */
-       if (rcu_gp_in_progress(rsp) &&
-           ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies)) {
-               rdp->n_rp_need_fqs++;
-               return 1;
-       }
-
        /* nothing to do */
        rdp->n_rp_need_nothing++;
        return 0;
index 5d92b80a0a281c58dcc887c377ff84af594714c6..2d04106d15330b6cfaf2855b4955758b936d88ee 100644 (file)
@@ -378,13 +378,6 @@ struct rcu_state {
 
        u8      fqs_state ____cacheline_internodealigned_in_smp;
                                                /* Force QS state. */
-       u8      fqs_active;                     /* force_quiescent_state() */
-                                               /*  is running. */
-       u8      fqs_need_gp;                    /* A CPU was prevented from */
-                                               /*  starting a new grace */
-                                               /*  period because */
-                                               /*  force_quiescent_state() */
-                                               /*  was running. */
        u8      boost;                          /* Subject to priority boost. */
        unsigned long gpnum;                    /* Current gp number. */
        unsigned long completed;                /* # of last completed gp. */
@@ -413,8 +406,6 @@ struct rcu_state {
        struct completion barrier_completion;   /* Wake at barrier end. */
        unsigned long n_barrier_done;           /* ++ at start and end of */
                                                /*  _rcu_barrier(). */
-       raw_spinlock_t fqslock;                 /* Only one task forcing */
-                                               /*  quiescent states. */
        unsigned long jiffies_force_qs;         /* Time at which to invoke */
                                                /*  force_quiescent_state(). */
        unsigned long n_force_qs;               /* Number of calls to */
@@ -433,6 +424,10 @@ struct rcu_state {
        struct list_head flavors;               /* List of RCU flavors. */
 };
 
+/* Values for rcu_state structure's gp_flags field. */
+#define RCU_GP_FLAG_INIT 0x1   /* Need grace-period initialization. */
+#define RCU_GP_FLAG_FQS  0x2   /* Need grace-period quiescent-state forcing. */
+
 extern struct list_head rcu_struct_flavors;
 #define for_each_rcu_flavor(rsp) \
        list_for_each_entry((rsp), &rcu_struct_flavors, flavors)
index 5879636893287d717574cd3e8ddf789c99031732..eb8dcd1bc4b5425b696af6b45da824e2a85a6b17 100644 (file)
@@ -119,7 +119,7 @@ EXPORT_SYMBOL_GPL(rcu_batches_completed);
  */
 void rcu_force_quiescent_state(void)
 {
-       force_quiescent_state(&rcu_preempt_state, 0);
+       force_quiescent_state(&rcu_preempt_state);
 }
 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
 
@@ -2076,16 +2076,16 @@ static void rcu_prepare_for_idle(int cpu)
 #ifdef CONFIG_TREE_PREEMPT_RCU
        if (per_cpu(rcu_preempt_data, cpu).nxtlist) {
                rcu_preempt_qs(cpu);
-               force_quiescent_state(&rcu_preempt_state, 0);
+               force_quiescent_state(&rcu_preempt_state);
        }
 #endif /* #ifdef CONFIG_TREE_PREEMPT_RCU */
        if (per_cpu(rcu_sched_data, cpu).nxtlist) {
                rcu_sched_qs(cpu);
-               force_quiescent_state(&rcu_sched_state, 0);
+               force_quiescent_state(&rcu_sched_state);
        }
        if (per_cpu(rcu_bh_data, cpu).nxtlist) {
                rcu_bh_qs(cpu);
-               force_quiescent_state(&rcu_bh_state, 0);
+               force_quiescent_state(&rcu_bh_state);
        }
 
        /*