Fix: call_rcu_thread() affinity failure

Make call_rcu_thread() affine itself more persistently

Currently, URCU simply fails if a call_rcu_thread() fails to affine
itself. This is problematic when execution is constrained by cgroup
and hotunplugged CPUs. This commit therefore makes call_rcu_thread()
retry setting its affinity every 256 grace periods, but only if it
detects that it migrated to a different CPU. Since sched_getcpu() is
cheap on many architectures, this check is less costly than going
through a system call.

Reported-by: Michael Jeanson <mjeanson@efficios.com>
Suggested-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Acked-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

diff --git a/urcu-call-rcu-impl.h b/urcu-call-rcu-impl.h
index 51ee91ffd1ef87e0e8a8e76ff00b64d6dd24c660..e70789a5906d6dec6f707025ec3db5895ea34b0f 100644 (file)
--- a/urcu-call-rcu-impl.h
+++ b/urcu-call-rcu-impl.h
@@ -43,6 +43,9 @@
 #include "urcu/tls-compat.h"
 #include "urcu-die.h"
 
+#define SET_AFFINITY_CHECK_PERIOD              (1U << 8)       /* 256 */
+#define SET_AFFINITY_CHECK_PERIOD_MASK         (SET_AFFINITY_CHECK_PERIOD - 1)
+
 /* Data structure that identifies a call_rcu thread. */
 
 struct call_rcu_data {
@@ -52,6 +55,7 @@ struct call_rcu_data {
        unsigned long qlen; /* maintained for debugging. */
        pthread_t tid;
        int cpu_affinity;
+       unsigned long gp_count;
        struct cds_list_head list;
 } __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
 
@@ -182,22 +186,42 @@ static void call_rcu_unlock(pthread_mutex_t *pmp)
                urcu_die(ret);
 }
 
+/*
+ * Periodically retry setting CPU affinity if we migrate.
+ * Losing affinity can be caused by CPU hotunplug/hotplug, or by
+ * cpuset(7).
+ */
 #if HAVE_SCHED_SETAFFINITY
 static
 int set_thread_cpu_affinity(struct call_rcu_data *crdp)
 {
        cpu_set_t mask;
+       int ret;
 
        if (crdp->cpu_affinity < 0)
                return 0;
+       if (++crdp->gp_count & SET_AFFINITY_CHECK_PERIOD_MASK)
+               return 0;
+       if (urcu_sched_getcpu() == crdp->cpu_affinity)
+               return 0;
 
        CPU_ZERO(&mask);
        CPU_SET(crdp->cpu_affinity, &mask);
 #if SCHED_SETAFFINITY_ARGS == 2
-       return sched_setaffinity(0, &mask);
+       ret = sched_setaffinity(0, &mask);
 #else
-       return sched_setaffinity(0, sizeof(mask), &mask);
+       ret = sched_setaffinity(0, sizeof(mask), &mask);
 #endif
+       /*
+        * EINVAL is fine: can be caused by hotunplugged CPUs, or by
+        * cpuset(7). This is why we should always retry if we detect
+        * migration.
+        */
+       if (ret && errno == EINVAL) {
+               ret = 0;
+               errno = 0;
+       }
+       return ret;
 }
 #else
 static
@@ -237,10 +261,8 @@ static void *call_rcu_thread(void *arg)
        struct call_rcu_data *crdp = (struct call_rcu_data *)arg;
        struct rcu_head *rhp;
        int rt = !!(uatomic_read(&crdp->flags) & URCU_CALL_RCU_RT);
-       int ret;
 
-       ret = set_thread_cpu_affinity(crdp);
-       if (ret)
+       if (set_thread_cpu_affinity(crdp))
                urcu_die(errno);
 
        /*
@@ -255,6 +277,9 @@ static void *call_rcu_thread(void *arg)
                cmm_smp_mb();
        }
        for (;;) {
+               if (set_thread_cpu_affinity(crdp))
+                       urcu_die(errno);
+
                if (uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSE) {
                        /*
                         * Pause requested. Become quiescent: remove
@@ -352,6 +377,7 @@ static void call_rcu_data_init(struct call_rcu_data **crdpp,
        crdp->flags = flags;
        cds_list_add(&crdp->list, &call_rcu_data_list);
        crdp->cpu_affinity = cpu_affinity;
+       crdp->gp_count = 0;
        cmm_smp_mb();  /* Structure initialized before pointer is planted. */
        *crdpp = crdp;
        ret = pthread_create(&crdp->tid, NULL, call_rcu_thread, crdp);