rcu: Fix synchronization for rcu_process_gp_end() uses of ->completed counter
authorPaul E. McKenney <paulmck@linux.vnet.ibm.com>
Mon, 2 Nov 2009 21:52:28 +0000 (13:52 -0800)
committerIngo Molnar <mingo@elte.hu>
Tue, 10 Nov 2009 03:11:54 +0000 (04:11 +0100)
Impose a clear locking design on the rcu_process_gp_end()
function's use of the ->completed counter.  This is done by
creating a ->completed field in the rcu_node structure, which
can safely be accessed under the protection of that structure's
lock.  Performance and scalability are maintained by using a
form of double-checked locking, so that rcu_process_gp_end()
only acquires the leaf rcu_node structure's ->lock if a grace
period has recently ended.

This fix reduces rcutorture failure rate by at least two orders
of magnitude under heavy stress with force_quiescent_state()
being invoked artificially often.  Without this fix,
unsynchronized access to the ->completed field can cause
rcu_process_gp_end() to advance callbacks whose grace period has
not yet expired.  (Bad idea!)

Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Cc: laijs@cn.fujitsu.com
Cc: dipankar@in.ibm.com
Cc: mathieu.desnoyers@polymtl.ca
Cc: josh@joshtriplett.org
Cc: dvhltc@us.ibm.com
Cc: niv@us.ibm.com
Cc: peterz@infradead.org
Cc: rostedt@goodmis.org
Cc: Valdis.Kletnieks@vt.edu
Cc: dhowells@redhat.com
Cc: <stable@kernel.org> # .32.x
LKML-Reference: <12571987494069-git-send-email->
Signed-off-by: Ingo Molnar <mingo@elte.hu>
kernel/rcutree.c
kernel/rcutree.h

index 26249abf24dcb59ae144215aa3366438b548ead3..9e068d1121533cf81cecdbeea0654f0bed415e2e 100644 (file)
@@ -569,6 +569,76 @@ check_for_new_grace_period(struct rcu_state *rsp, struct rcu_data *rdp)
        return ret;
 }
 
+/*
+ * Advance this CPU's callbacks, but only if the current grace period
+ * has ended.  This may be called only from the CPU to whom the rdp
+ * belongs.  In addition, the corresponding leaf rcu_node structure's
+ * ->lock must be held by the caller, with irqs disabled.
+ */
+static void
+__rcu_process_gp_end(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
+{
+       /* Did another grace period end? */
+       if (rdp->completed != rnp->completed) {
+
+               /* Advance callbacks.  No harm if list empty. */
+               rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[RCU_WAIT_TAIL];
+               rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_READY_TAIL];
+               rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
+
+               /* Remember that we saw this grace-period completion. */
+               rdp->completed = rnp->completed;
+       }
+}
+
+/*
+ * Advance this CPU's callbacks, but only if the current grace period
+ * has ended.  This may be called only from the CPU to whom the rdp
+ * belongs.
+ */
+static void
+rcu_process_gp_end(struct rcu_state *rsp, struct rcu_data *rdp)
+{
+       unsigned long flags;
+       struct rcu_node *rnp;
+
+       local_irq_save(flags);
+       rnp = rdp->mynode;
+       if (rdp->completed == ACCESS_ONCE(rnp->completed) || /* outside lock. */
+           !spin_trylock(&rnp->lock)) { /* irqs already off, retry later. */
+               local_irq_restore(flags);
+               return;
+       }
+       __rcu_process_gp_end(rsp, rnp, rdp);
+       spin_unlock_irqrestore(&rnp->lock, flags);
+}
+
+/*
+ * Do per-CPU grace-period initialization for running CPU.  The caller
+ * must hold the lock of the leaf rcu_node structure corresponding to
+ * this CPU.
+ */
+static void
+rcu_start_gp_per_cpu(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
+{
+       /* Prior grace period ended, so advance callbacks for current CPU. */
+       __rcu_process_gp_end(rsp, rnp, rdp);
+
+       /*
+        * Because this CPU just now started the new grace period, we know
+        * that all of its callbacks will be covered by this upcoming grace
+        * period, even the ones that were registered arbitrarily recently.
+        * Therefore, advance all outstanding callbacks to RCU_WAIT_TAIL.
+        *
+        * Other CPUs cannot be sure exactly when the grace period started.
+        * Therefore, their recently registered callbacks must pass through
+        * an additional RCU_NEXT_READY stage, so that they will be handled
+        * by the next RCU grace period.
+        */
+       rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
+       rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
+}
+
 /*
  * Start a new RCU grace period if warranted, re-initializing the hierarchy
  * in preparation for detecting the next grace period.  The caller must hold
@@ -596,26 +666,14 @@ rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
        dyntick_record_completed(rsp, rsp->completed - 1);
        note_new_gpnum(rsp, rdp);
 
-       /*
-        * Because this CPU just now started the new grace period, we know
-        * that all of its callbacks will be covered by this upcoming grace
-        * period, even the ones that were registered arbitrarily recently.
-        * Therefore, advance all outstanding callbacks to RCU_WAIT_TAIL.
-        *
-        * Other CPUs cannot be sure exactly when the grace period started.
-        * Therefore, their recently registered callbacks must pass through
-        * an additional RCU_NEXT_READY stage, so that they will be handled
-        * by the next RCU grace period.
-        */
-       rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
-       rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
-
        /* Special-case the common single-level case. */
        if (NUM_RCU_NODES == 1) {
                rcu_preempt_check_blocked_tasks(rnp);
                rnp->qsmask = rnp->qsmaskinit;
                rnp->gpnum = rsp->gpnum;
+               rnp->completed = rsp->completed;
                rsp->signaled = RCU_SIGNAL_INIT; /* force_quiescent_state OK. */
+               rcu_start_gp_per_cpu(rsp, rnp, rdp);
                spin_unlock_irqrestore(&rnp->lock, flags);
                return;
        }
@@ -648,6 +706,9 @@ rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
                rcu_preempt_check_blocked_tasks(rnp);
                rnp->qsmask = rnp->qsmaskinit;
                rnp->gpnum = rsp->gpnum;
+               rnp->completed = rsp->completed;
+               if (rnp == rdp->mynode)
+                       rcu_start_gp_per_cpu(rsp, rnp, rdp);
                spin_unlock(&rnp->lock);        /* irqs remain disabled. */
        }
 
@@ -658,34 +719,6 @@ rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
        spin_unlock_irqrestore(&rsp->onofflock, flags);
 }
 
-/*
- * Advance this CPU's callbacks, but only if the current grace period
- * has ended.  This may be called only from the CPU to whom the rdp
- * belongs.
- */
-static void
-rcu_process_gp_end(struct rcu_state *rsp, struct rcu_data *rdp)
-{
-       long completed_snap;
-       unsigned long flags;
-
-       local_irq_save(flags);
-       completed_snap = ACCESS_ONCE(rsp->completed);  /* outside of lock. */
-
-       /* Did another grace period end? */
-       if (rdp->completed != completed_snap) {
-
-               /* Advance callbacks.  No harm if list empty. */
-               rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[RCU_WAIT_TAIL];
-               rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_READY_TAIL];
-               rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
-
-               /* Remember that we saw this grace-period completion. */
-               rdp->completed = completed_snap;
-       }
-       local_irq_restore(flags);
-}
-
 /*
  * Clean up after the prior grace period and let rcu_start_gp() start up
  * the next grace period if one is needed.  Note that the caller must
@@ -697,7 +730,6 @@ static void cpu_quiet_msk_finish(struct rcu_state *rsp, unsigned long flags)
        WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
        rsp->completed = rsp->gpnum;
        rsp->signaled = RCU_GP_IDLE;
-       rcu_process_gp_end(rsp, rsp->rda[smp_processor_id()]);
        rcu_start_gp(rsp, flags);  /* releases root node's rnp->lock. */
 }
 
@@ -1539,21 +1571,16 @@ static void __cpuinit
 rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptable)
 {
        unsigned long flags;
-       long lastcomp;
        unsigned long mask;
        struct rcu_data *rdp = rsp->rda[cpu];
        struct rcu_node *rnp = rcu_get_root(rsp);
 
        /* Set up local state, ensuring consistent view of global state. */
        spin_lock_irqsave(&rnp->lock, flags);
-       lastcomp = rsp->completed;
-       rdp->completed = lastcomp;
-       rdp->gpnum = lastcomp;
        rdp->passed_quiesc = 0;  /* We could be racing with new GP, */
        rdp->qs_pending = 1;     /*  so set up to respond to current GP. */
        rdp->beenonline = 1;     /* We have now been online. */
        rdp->preemptable = preemptable;
-       rdp->passed_quiesc_completed = lastcomp - 1;
        rdp->qlen_last_fqs_check = 0;
        rdp->n_force_qs_snap = rsp->n_force_qs;
        rdp->blimit = blimit;
@@ -1575,6 +1602,11 @@ rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptable)
                spin_lock(&rnp->lock);  /* irqs already disabled. */
                rnp->qsmaskinit |= mask;
                mask = rnp->grpmask;
+               if (rnp == rdp->mynode) {
+                       rdp->gpnum = rnp->completed; /* if GP in progress... */
+                       rdp->completed = rnp->completed;
+                       rdp->passed_quiesc_completed = rnp->completed - 1;
+               }
                spin_unlock(&rnp->lock); /* irqs already disabled. */
                rnp = rnp->parent;
        } while (rnp != NULL && !(rnp->qsmaskinit & mask));
index 8a4c1650ad8d9d22785f49c5c8e647b61f4ee006..c1891c3cae63c0333438f8abac310ad94c598a5d 100644 (file)
@@ -84,6 +84,9 @@ struct rcu_node {
        long    gpnum;          /* Current grace period for this node. */
                                /*  This will either be equal to or one */
                                /*  behind the root rcu_node's gpnum. */
+       long    completed;      /* Last grace period completed for this node. */
+                               /*  This will either be equal to or one */
+                               /*  behind the root rcu_node's gpnum. */
        unsigned long qsmask;   /* CPUs or groups that need to switch in */
                                /*  order for current grace period to proceed.*/
                                /*  In leaf rcu_node, each bit corresponds to */