]> git.proxmox.com Git - mirror_ubuntu-hirsute-kernel.git/commitdiff
Merge branch 'torture.2014.02.23a' into HEAD
authorPaul E. McKenney <paulmck@linux.vnet.ibm.com>
Wed, 26 Feb 2014 14:38:59 +0000 (06:38 -0800)
committerPaul E. McKenney <paulmck@linux.vnet.ibm.com>
Wed, 26 Feb 2014 14:38:59 +0000 (06:38 -0800)
torture.2014.02.23a: locktorture addition and rcutorture changes

1  2 
kernel/rcu/rcutorture.c

index 0000000000000000000000000000000000000000,da6c38d909f15fbb0f312813113134ef8d840269..f59d48597ddeff6cbfea56df85d672a46980fdcd
mode 000000,100644..100644
--- /dev/null
@@@ -1,0 -1,1582 +1,1582 @@@
 - * along with this program; if not, write to the Free Software
 - * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
+ /*
+  * Read-Copy Update module-based torture test facility
+  *
+  * This program is free software; you can redistribute it and/or modify
+  * it under the terms of the GNU General Public License as published by
+  * the Free Software Foundation; either version 2 of the License, or
+  * (at your option) any later version.
+  *
+  * This program is distributed in the hope that it will be useful,
+  * but WITHOUT ANY WARRANTY; without even the implied warranty of
+  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+  * GNU General Public License for more details.
+  *
+  * You should have received a copy of the GNU General Public License
++ * along with this program; if not, you can access it online at
++ * http://www.gnu.org/licenses/gpl-2.0.html.
+  *
+  * Copyright (C) IBM Corporation, 2005, 2006
+  *
+  * Authors: Paul E. McKenney <paulmck@us.ibm.com>
+  *      Josh Triplett <josh@freedesktop.org>
+  *
+  * See also:  Documentation/RCU/torture.txt
+  */
+ #include <linux/types.h>
+ #include <linux/kernel.h>
+ #include <linux/init.h>
+ #include <linux/module.h>
+ #include <linux/kthread.h>
+ #include <linux/err.h>
+ #include <linux/spinlock.h>
+ #include <linux/smp.h>
+ #include <linux/rcupdate.h>
+ #include <linux/interrupt.h>
+ #include <linux/sched.h>
+ #include <linux/atomic.h>
+ #include <linux/bitops.h>
+ #include <linux/completion.h>
+ #include <linux/moduleparam.h>
+ #include <linux/percpu.h>
+ #include <linux/notifier.h>
+ #include <linux/reboot.h>
+ #include <linux/freezer.h>
+ #include <linux/cpu.h>
+ #include <linux/delay.h>
+ #include <linux/stat.h>
+ #include <linux/srcu.h>
+ #include <linux/slab.h>
+ #include <linux/trace_clock.h>
+ #include <asm/byteorder.h>
+ #include <linux/torture.h>
+ MODULE_LICENSE("GPL");
+ MODULE_AUTHOR("Paul E. McKenney <paulmck@us.ibm.com> and Josh Triplett <josh@freedesktop.org>");
+ torture_param(int, fqs_duration, 0,
+             "Duration of fqs bursts (us), 0 to disable");
+ torture_param(int, fqs_holdoff, 0, "Holdoff time within fqs bursts (us)");
+ torture_param(int, fqs_stutter, 3, "Wait time between fqs bursts (s)");
+ torture_param(bool, gp_exp, false, "Use expedited GP wait primitives");
+ torture_param(bool, gp_normal, false,
+            "Use normal (non-expedited) GP wait primitives");
+ torture_param(int, irqreader, 1, "Allow RCU readers from irq handlers");
+ torture_param(int, n_barrier_cbs, 0,
+            "# of callbacks/kthreads for barrier testing");
+ torture_param(int, nfakewriters, 4, "Number of RCU fake writer threads");
+ torture_param(int, nreaders, -1, "Number of RCU reader threads");
+ torture_param(int, object_debug, 0,
+            "Enable debug-object double call_rcu() testing");
+ torture_param(int, onoff_holdoff, 0, "Time after boot before CPU hotplugs (s)");
+ torture_param(int, onoff_interval, 0,
+            "Time between CPU hotplugs (s), 0=disable");
+ torture_param(int, shuffle_interval, 3, "Number of seconds between shuffles");
+ torture_param(int, shutdown_secs, 0, "Shutdown time (s), <= zero to disable.");
+ torture_param(int, stall_cpu, 0, "Stall duration (s), zero to disable.");
+ torture_param(int, stall_cpu_holdoff, 10,
+            "Time to wait before starting stall (s).");
+ torture_param(int, stat_interval, 60,
+            "Number of seconds between stats printk()s");
+ torture_param(int, stutter, 5, "Number of seconds to run/halt test");
+ torture_param(int, test_boost, 1, "Test RCU prio boost: 0=no, 1=maybe, 2=yes.");
+ torture_param(int, test_boost_duration, 4,
+            "Duration of each boost test, seconds.");
+ torture_param(int, test_boost_interval, 7,
+            "Interval between boost tests, seconds.");
+ torture_param(bool, test_no_idle_hz, true,
+            "Test support for tickless idle CPUs");
+ torture_param(bool, verbose, true,
+            "Enable verbose debugging printk()s");
+ static char *torture_type = "rcu";
+ module_param(torture_type, charp, 0444);
+ MODULE_PARM_DESC(torture_type, "Type of RCU to torture (rcu, rcu_bh, ...)");
+ static int nrealreaders;
+ static struct task_struct *writer_task;
+ static struct task_struct **fakewriter_tasks;
+ static struct task_struct **reader_tasks;
+ static struct task_struct *stats_task;
+ static struct task_struct *fqs_task;
+ static struct task_struct *boost_tasks[NR_CPUS];
+ static struct task_struct *stall_task;
+ static struct task_struct **barrier_cbs_tasks;
+ static struct task_struct *barrier_task;
+ #define RCU_TORTURE_PIPE_LEN 10
+ struct rcu_torture {
+       struct rcu_head rtort_rcu;
+       int rtort_pipe_count;
+       struct list_head rtort_free;
+       int rtort_mbtest;
+ };
+ static LIST_HEAD(rcu_torture_freelist);
+ static struct rcu_torture __rcu *rcu_torture_current;
+ static unsigned long rcu_torture_current_version;
+ static struct rcu_torture rcu_tortures[10 * RCU_TORTURE_PIPE_LEN];
+ static DEFINE_SPINLOCK(rcu_torture_lock);
+ static DEFINE_PER_CPU(long [RCU_TORTURE_PIPE_LEN + 1],
+                     rcu_torture_count) = { 0 };
+ static DEFINE_PER_CPU(long [RCU_TORTURE_PIPE_LEN + 1],
+                     rcu_torture_batch) = { 0 };
+ static atomic_t rcu_torture_wcount[RCU_TORTURE_PIPE_LEN + 1];
+ static atomic_t n_rcu_torture_alloc;
+ static atomic_t n_rcu_torture_alloc_fail;
+ static atomic_t n_rcu_torture_free;
+ static atomic_t n_rcu_torture_mberror;
+ static atomic_t n_rcu_torture_error;
+ static long n_rcu_torture_barrier_error;
+ static long n_rcu_torture_boost_ktrerror;
+ static long n_rcu_torture_boost_rterror;
+ static long n_rcu_torture_boost_failure;
+ static long n_rcu_torture_boosts;
+ static long n_rcu_torture_timers;
+ static long n_barrier_attempts;
+ static long n_barrier_successes;
+ static struct list_head rcu_torture_removed;
+ #if defined(MODULE) || defined(CONFIG_RCU_TORTURE_TEST_RUNNABLE)
+ #define RCUTORTURE_RUNNABLE_INIT 1
+ #else
+ #define RCUTORTURE_RUNNABLE_INIT 0
+ #endif
+ int rcutorture_runnable = RCUTORTURE_RUNNABLE_INIT;
+ module_param(rcutorture_runnable, int, 0444);
+ MODULE_PARM_DESC(rcutorture_runnable, "Start rcutorture at boot");
+ #if defined(CONFIG_RCU_BOOST) && !defined(CONFIG_HOTPLUG_CPU)
+ #define rcu_can_boost() 1
+ #else /* #if defined(CONFIG_RCU_BOOST) && !defined(CONFIG_HOTPLUG_CPU) */
+ #define rcu_can_boost() 0
+ #endif /* #else #if defined(CONFIG_RCU_BOOST) && !defined(CONFIG_HOTPLUG_CPU) */
+ #ifdef CONFIG_RCU_TRACE
+ static u64 notrace rcu_trace_clock_local(void)
+ {
+       u64 ts = trace_clock_local();
+       unsigned long __maybe_unused ts_rem = do_div(ts, NSEC_PER_USEC);
+       return ts;
+ }
+ #else /* #ifdef CONFIG_RCU_TRACE */
+ static u64 notrace rcu_trace_clock_local(void)
+ {
+       return 0ULL;
+ }
+ #endif /* #else #ifdef CONFIG_RCU_TRACE */
+ static unsigned long boost_starttime; /* jiffies of next boost test start. */
+ DEFINE_MUTEX(boost_mutex);            /* protect setting boost_starttime */
+                                       /*  and boost task create/destroy. */
+ static atomic_t barrier_cbs_count;    /* Barrier callbacks registered. */
+ static bool barrier_phase;            /* Test phase. */
+ static atomic_t barrier_cbs_invoked;  /* Barrier callbacks invoked. */
+ static wait_queue_head_t *barrier_cbs_wq; /* Coordinate barrier testing. */
+ static DECLARE_WAIT_QUEUE_HEAD(barrier_wq);
+ /*
+  * Allocate an element from the rcu_tortures pool.
+  */
+ static struct rcu_torture *
+ rcu_torture_alloc(void)
+ {
+       struct list_head *p;
+       spin_lock_bh(&rcu_torture_lock);
+       if (list_empty(&rcu_torture_freelist)) {
+               atomic_inc(&n_rcu_torture_alloc_fail);
+               spin_unlock_bh(&rcu_torture_lock);
+               return NULL;
+       }
+       atomic_inc(&n_rcu_torture_alloc);
+       p = rcu_torture_freelist.next;
+       list_del_init(p);
+       spin_unlock_bh(&rcu_torture_lock);
+       return container_of(p, struct rcu_torture, rtort_free);
+ }
+ /*
+  * Free an element to the rcu_tortures pool.
+  */
+ static void
+ rcu_torture_free(struct rcu_torture *p)
+ {
+       atomic_inc(&n_rcu_torture_free);
+       spin_lock_bh(&rcu_torture_lock);
+       list_add_tail(&p->rtort_free, &rcu_torture_freelist);
+       spin_unlock_bh(&rcu_torture_lock);
+ }
+ /*
+  * Operations vector for selecting different types of tests.
+  */
+ struct rcu_torture_ops {
+       void (*init)(void);
+       int (*readlock)(void);
+       void (*read_delay)(struct torture_random_state *rrsp);
+       void (*readunlock)(int idx);
+       int (*completed)(void);
+       void (*deferred_free)(struct rcu_torture *p);
+       void (*sync)(void);
+       void (*exp_sync)(void);
+       void (*call)(struct rcu_head *head, void (*func)(struct rcu_head *rcu));
+       void (*cb_barrier)(void);
+       void (*fqs)(void);
+       void (*stats)(char *page);
+       int irq_capable;
+       int can_boost;
+       const char *name;
+ };
+ static struct rcu_torture_ops *cur_ops;
+ /*
+  * Definitions for rcu torture testing.
+  */
+ static int rcu_torture_read_lock(void) __acquires(RCU)
+ {
+       rcu_read_lock();
+       return 0;
+ }
+ static void rcu_read_delay(struct torture_random_state *rrsp)
+ {
+       const unsigned long shortdelay_us = 200;
+       const unsigned long longdelay_ms = 50;
+       /* We want a short delay sometimes to make a reader delay the grace
+        * period, and we want a long delay occasionally to trigger
+        * force_quiescent_state. */
+       if (!(torture_random(rrsp) % (nrealreaders * 2000 * longdelay_ms)))
+               mdelay(longdelay_ms);
+       if (!(torture_random(rrsp) % (nrealreaders * 2 * shortdelay_us)))
+               udelay(shortdelay_us);
+ #ifdef CONFIG_PREEMPT
+       if (!preempt_count() &&
+           !(torture_random(rrsp) % (nrealreaders * 20000)))
+               preempt_schedule();  /* No QS if preempt_disable() in effect */
+ #endif
+ }
+ static void rcu_torture_read_unlock(int idx) __releases(RCU)
+ {
+       rcu_read_unlock();
+ }
+ static int rcu_torture_completed(void)
+ {
+       return rcu_batches_completed();
+ }
+ static void
+ rcu_torture_cb(struct rcu_head *p)
+ {
+       int i;
+       struct rcu_torture *rp = container_of(p, struct rcu_torture, rtort_rcu);
+       if (torture_must_stop_irq()) {
+               /* Test is ending, just drop callbacks on the floor. */
+               /* The next initialization will pick up the pieces. */
+               return;
+       }
+       i = rp->rtort_pipe_count;
+       if (i > RCU_TORTURE_PIPE_LEN)
+               i = RCU_TORTURE_PIPE_LEN;
+       atomic_inc(&rcu_torture_wcount[i]);
+       if (++rp->rtort_pipe_count >= RCU_TORTURE_PIPE_LEN) {
+               rp->rtort_mbtest = 0;
+               rcu_torture_free(rp);
+       } else {
+               cur_ops->deferred_free(rp);
+       }
+ }
+ static int rcu_no_completed(void)
+ {
+       return 0;
+ }
+ static void rcu_torture_deferred_free(struct rcu_torture *p)
+ {
+       call_rcu(&p->rtort_rcu, rcu_torture_cb);
+ }
+ static void rcu_sync_torture_init(void)
+ {
+       INIT_LIST_HEAD(&rcu_torture_removed);
+ }
+ static struct rcu_torture_ops rcu_ops = {
+       .init           = rcu_sync_torture_init,
+       .readlock       = rcu_torture_read_lock,
+       .read_delay     = rcu_read_delay,
+       .readunlock     = rcu_torture_read_unlock,
+       .completed      = rcu_torture_completed,
+       .deferred_free  = rcu_torture_deferred_free,
+       .sync           = synchronize_rcu,
+       .exp_sync       = synchronize_rcu_expedited,
+       .call           = call_rcu,
+       .cb_barrier     = rcu_barrier,
+       .fqs            = rcu_force_quiescent_state,
+       .stats          = NULL,
+       .irq_capable    = 1,
+       .can_boost      = rcu_can_boost(),
+       .name           = "rcu"
+ };
+ /*
+  * Definitions for rcu_bh torture testing.
+  */
+ static int rcu_bh_torture_read_lock(void) __acquires(RCU_BH)
+ {
+       rcu_read_lock_bh();
+       return 0;
+ }
+ static void rcu_bh_torture_read_unlock(int idx) __releases(RCU_BH)
+ {
+       rcu_read_unlock_bh();
+ }
+ static int rcu_bh_torture_completed(void)
+ {
+       return rcu_batches_completed_bh();
+ }
+ static void rcu_bh_torture_deferred_free(struct rcu_torture *p)
+ {
+       call_rcu_bh(&p->rtort_rcu, rcu_torture_cb);
+ }
+ static struct rcu_torture_ops rcu_bh_ops = {
+       .init           = rcu_sync_torture_init,
+       .readlock       = rcu_bh_torture_read_lock,
+       .read_delay     = rcu_read_delay,  /* just reuse rcu's version. */
+       .readunlock     = rcu_bh_torture_read_unlock,
+       .completed      = rcu_bh_torture_completed,
+       .deferred_free  = rcu_bh_torture_deferred_free,
+       .sync           = synchronize_rcu_bh,
+       .exp_sync       = synchronize_rcu_bh_expedited,
+       .call           = call_rcu_bh,
+       .cb_barrier     = rcu_barrier_bh,
+       .fqs            = rcu_bh_force_quiescent_state,
+       .stats          = NULL,
+       .irq_capable    = 1,
+       .name           = "rcu_bh"
+ };
+ /*
+  * Don't even think about trying any of these in real life!!!
+  * The names includes "busted", and they really means it!
+  * The only purpose of these functions is to provide a buggy RCU
+  * implementation to make sure that rcutorture correctly emits
+  * buggy-RCU error messages.
+  */
+ static void rcu_busted_torture_deferred_free(struct rcu_torture *p)
+ {
+       /* This is a deliberate bug for testing purposes only! */
+       rcu_torture_cb(&p->rtort_rcu);
+ }
+ static void synchronize_rcu_busted(void)
+ {
+       /* This is a deliberate bug for testing purposes only! */
+ }
+ static void
+ call_rcu_busted(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
+ {
+       /* This is a deliberate bug for testing purposes only! */
+       func(head);
+ }
+ static struct rcu_torture_ops rcu_busted_ops = {
+       .init           = rcu_sync_torture_init,
+       .readlock       = rcu_torture_read_lock,
+       .read_delay     = rcu_read_delay,  /* just reuse rcu's version. */
+       .readunlock     = rcu_torture_read_unlock,
+       .completed      = rcu_no_completed,
+       .deferred_free  = rcu_busted_torture_deferred_free,
+       .sync           = synchronize_rcu_busted,
+       .exp_sync       = synchronize_rcu_busted,
+       .call           = call_rcu_busted,
+       .cb_barrier     = NULL,
+       .fqs            = NULL,
+       .stats          = NULL,
+       .irq_capable    = 1,
+       .name           = "rcu_busted"
+ };
+ /*
+  * Definitions for srcu torture testing.
+  */
+ DEFINE_STATIC_SRCU(srcu_ctl);
+ static int srcu_torture_read_lock(void) __acquires(&srcu_ctl)
+ {
+       return srcu_read_lock(&srcu_ctl);
+ }
+ static void srcu_read_delay(struct torture_random_state *rrsp)
+ {
+       long delay;
+       const long uspertick = 1000000 / HZ;
+       const long longdelay = 10;
+       /* We want there to be long-running readers, but not all the time. */
+       delay = torture_random(rrsp) %
+               (nrealreaders * 2 * longdelay * uspertick);
+       if (!delay)
+               schedule_timeout_interruptible(longdelay);
+       else
+               rcu_read_delay(rrsp);
+ }
+ static void srcu_torture_read_unlock(int idx) __releases(&srcu_ctl)
+ {
+       srcu_read_unlock(&srcu_ctl, idx);
+ }
+ static int srcu_torture_completed(void)
+ {
+       return srcu_batches_completed(&srcu_ctl);
+ }
+ static void srcu_torture_deferred_free(struct rcu_torture *rp)
+ {
+       call_srcu(&srcu_ctl, &rp->rtort_rcu, rcu_torture_cb);
+ }
+ static void srcu_torture_synchronize(void)
+ {
+       synchronize_srcu(&srcu_ctl);
+ }
+ static void srcu_torture_call(struct rcu_head *head,
+                             void (*func)(struct rcu_head *head))
+ {
+       call_srcu(&srcu_ctl, head, func);
+ }
+ static void srcu_torture_barrier(void)
+ {
+       srcu_barrier(&srcu_ctl);
+ }
+ static void srcu_torture_stats(char *page)
+ {
+       int cpu;
+       int idx = srcu_ctl.completed & 0x1;
+       page += sprintf(page, "%s%s per-CPU(idx=%d):",
+                      torture_type, TORTURE_FLAG, idx);
+       for_each_possible_cpu(cpu) {
+               page += sprintf(page, " %d(%lu,%lu)", cpu,
+                              per_cpu_ptr(srcu_ctl.per_cpu_ref, cpu)->c[!idx],
+                              per_cpu_ptr(srcu_ctl.per_cpu_ref, cpu)->c[idx]);
+       }
+       sprintf(page, "\n");
+ }
+ static void srcu_torture_synchronize_expedited(void)
+ {
+       synchronize_srcu_expedited(&srcu_ctl);
+ }
+ static struct rcu_torture_ops srcu_ops = {
+       .init           = rcu_sync_torture_init,
+       .readlock       = srcu_torture_read_lock,
+       .read_delay     = srcu_read_delay,
+       .readunlock     = srcu_torture_read_unlock,
+       .completed      = srcu_torture_completed,
+       .deferred_free  = srcu_torture_deferred_free,
+       .sync           = srcu_torture_synchronize,
+       .exp_sync       = srcu_torture_synchronize_expedited,
+       .call           = srcu_torture_call,
+       .cb_barrier     = srcu_torture_barrier,
+       .stats          = srcu_torture_stats,
+       .name           = "srcu"
+ };
+ /*
+  * Definitions for sched torture testing.
+  */
+ static int sched_torture_read_lock(void)
+ {
+       preempt_disable();
+       return 0;
+ }
+ static void sched_torture_read_unlock(int idx)
+ {
+       preempt_enable();
+ }
+ static void rcu_sched_torture_deferred_free(struct rcu_torture *p)
+ {
+       call_rcu_sched(&p->rtort_rcu, rcu_torture_cb);
+ }
+ static struct rcu_torture_ops sched_ops = {
+       .init           = rcu_sync_torture_init,
+       .readlock       = sched_torture_read_lock,
+       .read_delay     = rcu_read_delay,  /* just reuse rcu's version. */
+       .readunlock     = sched_torture_read_unlock,
+       .completed      = rcu_no_completed,
+       .deferred_free  = rcu_sched_torture_deferred_free,
+       .sync           = synchronize_sched,
+       .exp_sync       = synchronize_sched_expedited,
+       .call           = call_rcu_sched,
+       .cb_barrier     = rcu_barrier_sched,
+       .fqs            = rcu_sched_force_quiescent_state,
+       .stats          = NULL,
+       .irq_capable    = 1,
+       .name           = "sched"
+ };
+ /*
+  * RCU torture priority-boost testing.  Runs one real-time thread per
+  * CPU for moderate bursts, repeatedly registering RCU callbacks and
+  * spinning waiting for them to be invoked.  If a given callback takes
+  * too long to be invoked, we assume that priority inversion has occurred.
+  */
+ struct rcu_boost_inflight {
+       struct rcu_head rcu;
+       int inflight;
+ };
+ static void rcu_torture_boost_cb(struct rcu_head *head)
+ {
+       struct rcu_boost_inflight *rbip =
+               container_of(head, struct rcu_boost_inflight, rcu);
+       smp_mb(); /* Ensure RCU-core accesses precede clearing ->inflight */
+       rbip->inflight = 0;
+ }
+ static int rcu_torture_boost(void *arg)
+ {
+       unsigned long call_rcu_time;
+       unsigned long endtime;
+       unsigned long oldstarttime;
+       struct rcu_boost_inflight rbi = { .inflight = 0 };
+       struct sched_param sp;
+       VERBOSE_TOROUT_STRING("rcu_torture_boost started");
+       /* Set real-time priority. */
+       sp.sched_priority = 1;
+       if (sched_setscheduler(current, SCHED_FIFO, &sp) < 0) {
+               VERBOSE_TOROUT_STRING("rcu_torture_boost RT prio failed!");
+               n_rcu_torture_boost_rterror++;
+       }
+       init_rcu_head_on_stack(&rbi.rcu);
+       /* Each pass through the following loop does one boost-test cycle. */
+       do {
+               /* Wait for the next test interval. */
+               oldstarttime = boost_starttime;
+               while (ULONG_CMP_LT(jiffies, oldstarttime)) {
+                       schedule_timeout_interruptible(oldstarttime - jiffies);
+                       stutter_wait("rcu_torture_boost");
+                       if (torture_must_stop())
+                               goto checkwait;
+               }
+               /* Do one boost-test interval. */
+               endtime = oldstarttime + test_boost_duration * HZ;
+               call_rcu_time = jiffies;
+               while (ULONG_CMP_LT(jiffies, endtime)) {
+                       /* If we don't have a callback in flight, post one. */
+                       if (!rbi.inflight) {
+                               smp_mb(); /* RCU core before ->inflight = 1. */
+                               rbi.inflight = 1;
+                               call_rcu(&rbi.rcu, rcu_torture_boost_cb);
+                               if (jiffies - call_rcu_time >
+                                        test_boost_duration * HZ - HZ / 2) {
+                                       VERBOSE_TOROUT_STRING("rcu_torture_boost boosting failed");
+                                       n_rcu_torture_boost_failure++;
+                               }
+                               call_rcu_time = jiffies;
+                       }
+                       cond_resched();
+                       stutter_wait("rcu_torture_boost");
+                       if (torture_must_stop())
+                               goto checkwait;
+               }
+               /*
+                * Set the start time of the next test interval.
+                * Yes, this is vulnerable to long delays, but such
+                * delays simply cause a false negative for the next
+                * interval.  Besides, we are running at RT priority,
+                * so delays should be relatively rare.
+                */
+               while (oldstarttime == boost_starttime &&
+                      !kthread_should_stop()) {
+                       if (mutex_trylock(&boost_mutex)) {
+                               boost_starttime = jiffies +
+                                                 test_boost_interval * HZ;
+                               n_rcu_torture_boosts++;
+                               mutex_unlock(&boost_mutex);
+                               break;
+                       }
+                       schedule_timeout_uninterruptible(1);
+               }
+               /* Go do the stutter. */
+ checkwait:    stutter_wait("rcu_torture_boost");
+       } while (!torture_must_stop());
+       /* Clean up and exit. */
+       while (!kthread_should_stop() || rbi.inflight) {
+               torture_shutdown_absorb("rcu_torture_boost");
+               schedule_timeout_uninterruptible(1);
+       }
+       smp_mb(); /* order accesses to ->inflight before stack-frame death. */
+       destroy_rcu_head_on_stack(&rbi.rcu);
+       torture_kthread_stopping("rcu_torture_boost");
+       return 0;
+ }
+ /*
+  * RCU torture force-quiescent-state kthread.  Repeatedly induces
+  * bursts of calls to force_quiescent_state(), increasing the probability
+  * of occurrence of some important types of race conditions.
+  */
+ static int
+ rcu_torture_fqs(void *arg)
+ {
+       unsigned long fqs_resume_time;
+       int fqs_burst_remaining;
+       VERBOSE_TOROUT_STRING("rcu_torture_fqs task started");
+       do {
+               fqs_resume_time = jiffies + fqs_stutter * HZ;
+               while (ULONG_CMP_LT(jiffies, fqs_resume_time) &&
+                      !kthread_should_stop()) {
+                       schedule_timeout_interruptible(1);
+               }
+               fqs_burst_remaining = fqs_duration;
+               while (fqs_burst_remaining > 0 &&
+                      !kthread_should_stop()) {
+                       cur_ops->fqs();
+                       udelay(fqs_holdoff);
+                       fqs_burst_remaining -= fqs_holdoff;
+               }
+               stutter_wait("rcu_torture_fqs");
+       } while (!torture_must_stop());
+       torture_kthread_stopping("rcu_torture_fqs");
+       return 0;
+ }
+ /*
+  * RCU torture writer kthread.  Repeatedly substitutes a new structure
+  * for that pointed to by rcu_torture_current, freeing the old structure
+  * after a series of grace periods (the "pipeline").
+  */
+ static int
+ rcu_torture_writer(void *arg)
+ {
+       bool exp;
+       int i;
+       struct rcu_torture *rp;
+       struct rcu_torture *rp1;
+       struct rcu_torture *old_rp;
+       static DEFINE_TORTURE_RANDOM(rand);
+       VERBOSE_TOROUT_STRING("rcu_torture_writer task started");
+       set_user_nice(current, 19);
+       do {
+               schedule_timeout_uninterruptible(1);
+               rp = rcu_torture_alloc();
+               if (rp == NULL)
+                       continue;
+               rp->rtort_pipe_count = 0;
+               udelay(torture_random(&rand) & 0x3ff);
+               old_rp = rcu_dereference_check(rcu_torture_current,
+                                              current == writer_task);
+               rp->rtort_mbtest = 1;
+               rcu_assign_pointer(rcu_torture_current, rp);
+               smp_wmb(); /* Mods to old_rp must follow rcu_assign_pointer() */
+               if (old_rp) {
+                       i = old_rp->rtort_pipe_count;
+                       if (i > RCU_TORTURE_PIPE_LEN)
+                               i = RCU_TORTURE_PIPE_LEN;
+                       atomic_inc(&rcu_torture_wcount[i]);
+                       old_rp->rtort_pipe_count++;
+                       if (gp_normal == gp_exp)
+                               exp = !!(torture_random(&rand) & 0x80);
+                       else
+                               exp = gp_exp;
+                       if (!exp) {
+                               cur_ops->deferred_free(old_rp);
+                       } else {
+                               cur_ops->exp_sync();
+                               list_add(&old_rp->rtort_free,
+                                        &rcu_torture_removed);
+                               list_for_each_entry_safe(rp, rp1,
+                                                        &rcu_torture_removed,
+                                                        rtort_free) {
+                                       i = rp->rtort_pipe_count;
+                                       if (i > RCU_TORTURE_PIPE_LEN)
+                                               i = RCU_TORTURE_PIPE_LEN;
+                                       atomic_inc(&rcu_torture_wcount[i]);
+                                       if (++rp->rtort_pipe_count >=
+                                           RCU_TORTURE_PIPE_LEN) {
+                                               rp->rtort_mbtest = 0;
+                                               list_del(&rp->rtort_free);
+                                               rcu_torture_free(rp);
+                                       }
+                                }
+                       }
+               }
+               rcutorture_record_progress(++rcu_torture_current_version);
+               stutter_wait("rcu_torture_writer");
+       } while (!torture_must_stop());
+       torture_kthread_stopping("rcu_torture_writer");
+       return 0;
+ }
+ /*
+  * RCU torture fake writer kthread.  Repeatedly calls sync, with a random
+  * delay between calls.
+  */
+ static int
+ rcu_torture_fakewriter(void *arg)
+ {
+       DEFINE_TORTURE_RANDOM(rand);
+       VERBOSE_TOROUT_STRING("rcu_torture_fakewriter task started");
+       set_user_nice(current, 19);
+       do {
+               schedule_timeout_uninterruptible(1 + torture_random(&rand)%10);
+               udelay(torture_random(&rand) & 0x3ff);
+               if (cur_ops->cb_barrier != NULL &&
+                   torture_random(&rand) % (nfakewriters * 8) == 0) {
+                       cur_ops->cb_barrier();
+               } else if (gp_normal == gp_exp) {
+                       if (torture_random(&rand) & 0x80)
+                               cur_ops->sync();
+                       else
+                               cur_ops->exp_sync();
+               } else if (gp_normal) {
+                       cur_ops->sync();
+               } else {
+                       cur_ops->exp_sync();
+               }
+               stutter_wait("rcu_torture_fakewriter");
+       } while (!torture_must_stop());
+       torture_kthread_stopping("rcu_torture_fakewriter");
+       return 0;
+ }
+ void rcutorture_trace_dump(void)
+ {
+       static atomic_t beenhere = ATOMIC_INIT(0);
+       if (atomic_read(&beenhere))
+               return;
+       if (atomic_xchg(&beenhere, 1) != 0)
+               return;
+       ftrace_dump(DUMP_ALL);
+ }
+ /*
+  * RCU torture reader from timer handler.  Dereferences rcu_torture_current,
+  * incrementing the corresponding element of the pipeline array.  The
+  * counter in the element should never be greater than 1, otherwise, the
+  * RCU implementation is broken.
+  */
+ static void rcu_torture_timer(unsigned long unused)
+ {
+       int idx;
+       int completed;
+       int completed_end;
+       static DEFINE_TORTURE_RANDOM(rand);
+       static DEFINE_SPINLOCK(rand_lock);
+       struct rcu_torture *p;
+       int pipe_count;
+       unsigned long long ts;
+       idx = cur_ops->readlock();
+       completed = cur_ops->completed();
+       ts = rcu_trace_clock_local();
+       p = rcu_dereference_check(rcu_torture_current,
+                                 rcu_read_lock_bh_held() ||
+                                 rcu_read_lock_sched_held() ||
+                                 srcu_read_lock_held(&srcu_ctl));
+       if (p == NULL) {
+               /* Leave because rcu_torture_writer is not yet underway */
+               cur_ops->readunlock(idx);
+               return;
+       }
+       if (p->rtort_mbtest == 0)
+               atomic_inc(&n_rcu_torture_mberror);
+       spin_lock(&rand_lock);
+       cur_ops->read_delay(&rand);
+       n_rcu_torture_timers++;
+       spin_unlock(&rand_lock);
+       preempt_disable();
+       pipe_count = p->rtort_pipe_count;
+       if (pipe_count > RCU_TORTURE_PIPE_LEN) {
+               /* Should not happen, but... */
+               pipe_count = RCU_TORTURE_PIPE_LEN;
+       }
+       completed_end = cur_ops->completed();
+       if (pipe_count > 1) {
+               do_trace_rcu_torture_read(cur_ops->name, &p->rtort_rcu, ts,
+                                         completed, completed_end);
+               rcutorture_trace_dump();
+       }
+       __this_cpu_inc(rcu_torture_count[pipe_count]);
+       completed = completed_end - completed;
+       if (completed > RCU_TORTURE_PIPE_LEN) {
+               /* Should not happen, but... */
+               completed = RCU_TORTURE_PIPE_LEN;
+       }
+       __this_cpu_inc(rcu_torture_batch[completed]);
+       preempt_enable();
+       cur_ops->readunlock(idx);
+ }
+ /*
+  * RCU torture reader kthread.  Repeatedly dereferences rcu_torture_current,
+  * incrementing the corresponding element of the pipeline array.  The
+  * counter in the element should never be greater than 1, otherwise, the
+  * RCU implementation is broken.
+  */
+ static int
+ rcu_torture_reader(void *arg)
+ {
+       int completed;
+       int completed_end;
+       int idx;
+       DEFINE_TORTURE_RANDOM(rand);
+       struct rcu_torture *p;
+       int pipe_count;
+       struct timer_list t;
+       unsigned long long ts;
+       VERBOSE_TOROUT_STRING("rcu_torture_reader task started");
+       set_user_nice(current, 19);
+       if (irqreader && cur_ops->irq_capable)
+               setup_timer_on_stack(&t, rcu_torture_timer, 0);
+       do {
+               if (irqreader && cur_ops->irq_capable) {
+                       if (!timer_pending(&t))
+                               mod_timer(&t, jiffies + 1);
+               }
+               idx = cur_ops->readlock();
+               completed = cur_ops->completed();
+               ts = rcu_trace_clock_local();
+               p = rcu_dereference_check(rcu_torture_current,
+                                         rcu_read_lock_bh_held() ||
+                                         rcu_read_lock_sched_held() ||
+                                         srcu_read_lock_held(&srcu_ctl));
+               if (p == NULL) {
+                       /* Wait for rcu_torture_writer to get underway */
+                       cur_ops->readunlock(idx);
+                       schedule_timeout_interruptible(HZ);
+                       continue;
+               }
+               if (p->rtort_mbtest == 0)
+                       atomic_inc(&n_rcu_torture_mberror);
+               cur_ops->read_delay(&rand);
+               preempt_disable();
+               pipe_count = p->rtort_pipe_count;
+               if (pipe_count > RCU_TORTURE_PIPE_LEN) {
+                       /* Should not happen, but... */
+                       pipe_count = RCU_TORTURE_PIPE_LEN;
+               }
+               completed_end = cur_ops->completed();
+               if (pipe_count > 1) {
+                       do_trace_rcu_torture_read(cur_ops->name, &p->rtort_rcu,
+                                                 ts, completed, completed_end);
+                       rcutorture_trace_dump();
+               }
+               __this_cpu_inc(rcu_torture_count[pipe_count]);
+               completed = completed_end - completed;
+               if (completed > RCU_TORTURE_PIPE_LEN) {
+                       /* Should not happen, but... */
+                       completed = RCU_TORTURE_PIPE_LEN;
+               }
+               __this_cpu_inc(rcu_torture_batch[completed]);
+               preempt_enable();
+               cur_ops->readunlock(idx);
+               schedule();
+               stutter_wait("rcu_torture_reader");
+       } while (!torture_must_stop());
+       if (irqreader && cur_ops->irq_capable)
+               del_timer_sync(&t);
+       torture_kthread_stopping("rcu_torture_reader");
+       return 0;
+ }
+ /*
+  * Create an RCU-torture statistics message in the specified buffer.
+  */
+ static void
+ rcu_torture_printk(char *page)
+ {
+       int cpu;
+       int i;
+       long pipesummary[RCU_TORTURE_PIPE_LEN + 1] = { 0 };
+       long batchsummary[RCU_TORTURE_PIPE_LEN + 1] = { 0 };
+       for_each_possible_cpu(cpu) {
+               for (i = 0; i < RCU_TORTURE_PIPE_LEN + 1; i++) {
+                       pipesummary[i] += per_cpu(rcu_torture_count, cpu)[i];
+                       batchsummary[i] += per_cpu(rcu_torture_batch, cpu)[i];
+               }
+       }
+       for (i = RCU_TORTURE_PIPE_LEN - 1; i >= 0; i--) {
+               if (pipesummary[i] != 0)
+                       break;
+       }
+       page += sprintf(page, "%s%s ", torture_type, TORTURE_FLAG);
+       page += sprintf(page,
+                      "rtc: %p ver: %lu tfle: %d rta: %d rtaf: %d rtf: %d ",
+                      rcu_torture_current,
+                      rcu_torture_current_version,
+                      list_empty(&rcu_torture_freelist),
+                      atomic_read(&n_rcu_torture_alloc),
+                      atomic_read(&n_rcu_torture_alloc_fail),
+                      atomic_read(&n_rcu_torture_free));
+       page += sprintf(page, "rtmbe: %d rtbke: %ld rtbre: %ld ",
+                      atomic_read(&n_rcu_torture_mberror),
+                      n_rcu_torture_boost_ktrerror,
+                      n_rcu_torture_boost_rterror);
+       page += sprintf(page, "rtbf: %ld rtb: %ld nt: %ld ",
+                      n_rcu_torture_boost_failure,
+                      n_rcu_torture_boosts,
+                      n_rcu_torture_timers);
+       page = torture_onoff_stats(page);
+       page += sprintf(page, "barrier: %ld/%ld:%ld",
+                      n_barrier_successes,
+                      n_barrier_attempts,
+                      n_rcu_torture_barrier_error);
+       page += sprintf(page, "\n%s%s ", torture_type, TORTURE_FLAG);
+       if (atomic_read(&n_rcu_torture_mberror) != 0 ||
+           n_rcu_torture_barrier_error != 0 ||
+           n_rcu_torture_boost_ktrerror != 0 ||
+           n_rcu_torture_boost_rterror != 0 ||
+           n_rcu_torture_boost_failure != 0 ||
+           i > 1) {
+               page += sprintf(page, "!!! ");
+               atomic_inc(&n_rcu_torture_error);
+               WARN_ON_ONCE(1);
+       }
+       page += sprintf(page, "Reader Pipe: ");
+       for (i = 0; i < RCU_TORTURE_PIPE_LEN + 1; i++)
+               page += sprintf(page, " %ld", pipesummary[i]);
+       page += sprintf(page, "\n%s%s ", torture_type, TORTURE_FLAG);
+       page += sprintf(page, "Reader Batch: ");
+       for (i = 0; i < RCU_TORTURE_PIPE_LEN + 1; i++)
+               page += sprintf(page, " %ld", batchsummary[i]);
+       page += sprintf(page, "\n%s%s ", torture_type, TORTURE_FLAG);
+       page += sprintf(page, "Free-Block Circulation: ");
+       for (i = 0; i < RCU_TORTURE_PIPE_LEN + 1; i++) {
+               page += sprintf(page, " %d",
+                              atomic_read(&rcu_torture_wcount[i]));
+       }
+       page += sprintf(page, "\n");
+       if (cur_ops->stats)
+               cur_ops->stats(page);
+ }
+ /*
+  * Print torture statistics.  Caller must ensure that there is only
+  * one call to this function at a given time!!!  This is normally
+  * accomplished by relying on the module system to only have one copy
+  * of the module loaded, and then by giving the rcu_torture_stats
+  * kthread full control (or the init/cleanup functions when rcu_torture_stats
+  * thread is not running).
+  */
+ static void
+ rcu_torture_stats_print(void)
+ {
+       int size = nr_cpu_ids * 200 + 8192;
+       char *buf;
+       buf = kmalloc(size, GFP_KERNEL);
+       if (!buf) {
+               pr_err("rcu-torture: Out of memory, need: %d", size);
+               return;
+       }
+       rcu_torture_printk(buf);
+       pr_alert("%s", buf);
+       kfree(buf);
+ }
+ /*
+  * Periodically prints torture statistics, if periodic statistics printing
+  * was specified via the stat_interval module parameter.
+  */
+ static int
+ rcu_torture_stats(void *arg)
+ {
+       VERBOSE_TOROUT_STRING("rcu_torture_stats task started");
+       do {
+               schedule_timeout_interruptible(stat_interval * HZ);
+               rcu_torture_stats_print();
+               torture_shutdown_absorb("rcu_torture_stats");
+       } while (!torture_must_stop());
+       torture_kthread_stopping("rcu_torture_stats");
+       return 0;
+ }
+ static inline void
+ rcu_torture_print_module_parms(struct rcu_torture_ops *cur_ops, const char *tag)
+ {
+       pr_alert("%s" TORTURE_FLAG
+                "--- %s: nreaders=%d nfakewriters=%d "
+                "stat_interval=%d verbose=%d test_no_idle_hz=%d "
+                "shuffle_interval=%d stutter=%d irqreader=%d "
+                "fqs_duration=%d fqs_holdoff=%d fqs_stutter=%d "
+                "test_boost=%d/%d test_boost_interval=%d "
+                "test_boost_duration=%d shutdown_secs=%d "
+                "stall_cpu=%d stall_cpu_holdoff=%d "
+                "n_barrier_cbs=%d "
+                "onoff_interval=%d onoff_holdoff=%d\n",
+                torture_type, tag, nrealreaders, nfakewriters,
+                stat_interval, verbose, test_no_idle_hz, shuffle_interval,
+                stutter, irqreader, fqs_duration, fqs_holdoff, fqs_stutter,
+                test_boost, cur_ops->can_boost,
+                test_boost_interval, test_boost_duration, shutdown_secs,
+                stall_cpu, stall_cpu_holdoff,
+                n_barrier_cbs,
+                onoff_interval, onoff_holdoff);
+ }
+ static void rcutorture_booster_cleanup(int cpu)
+ {
+       struct task_struct *t;
+       if (boost_tasks[cpu] == NULL)
+               return;
+       mutex_lock(&boost_mutex);
+       t = boost_tasks[cpu];
+       boost_tasks[cpu] = NULL;
+       mutex_unlock(&boost_mutex);
+       /* This must be outside of the mutex, otherwise deadlock! */
+       torture_stop_kthread(rcu_torture_boost, t);
+ }
+ static int rcutorture_booster_init(int cpu)
+ {
+       int retval;
+       if (boost_tasks[cpu] != NULL)
+               return 0;  /* Already created, nothing more to do. */
+       /* Don't allow time recalculation while creating a new task. */
+       mutex_lock(&boost_mutex);
+       VERBOSE_TOROUT_STRING("Creating rcu_torture_boost task");
+       boost_tasks[cpu] = kthread_create_on_node(rcu_torture_boost, NULL,
+                                                 cpu_to_node(cpu),
+                                                 "rcu_torture_boost");
+       if (IS_ERR(boost_tasks[cpu])) {
+               retval = PTR_ERR(boost_tasks[cpu]);
+               VERBOSE_TOROUT_STRING("rcu_torture_boost task create failed");
+               n_rcu_torture_boost_ktrerror++;
+               boost_tasks[cpu] = NULL;
+               mutex_unlock(&boost_mutex);
+               return retval;
+       }
+       kthread_bind(boost_tasks[cpu], cpu);
+       wake_up_process(boost_tasks[cpu]);
+       mutex_unlock(&boost_mutex);
+       return 0;
+ }
+ /*
+  * CPU-stall kthread.  It waits as specified by stall_cpu_holdoff, then
+  * induces a CPU stall for the time specified by stall_cpu.
+  */
+ static int rcu_torture_stall(void *args)
+ {
+       unsigned long stop_at;
+       VERBOSE_TOROUT_STRING("rcu_torture_stall task started");
+       if (stall_cpu_holdoff > 0) {
+               VERBOSE_TOROUT_STRING("rcu_torture_stall begin holdoff");
+               schedule_timeout_interruptible(stall_cpu_holdoff * HZ);
+               VERBOSE_TOROUT_STRING("rcu_torture_stall end holdoff");
+       }
+       if (!kthread_should_stop()) {
+               stop_at = get_seconds() + stall_cpu;
+               /* RCU CPU stall is expected behavior in following code. */
+               pr_alert("rcu_torture_stall start.\n");
+               rcu_read_lock();
+               preempt_disable();
+               while (ULONG_CMP_LT(get_seconds(), stop_at))
+                       continue;  /* Induce RCU CPU stall warning. */
+               preempt_enable();
+               rcu_read_unlock();
+               pr_alert("rcu_torture_stall end.\n");
+       }
+       torture_shutdown_absorb("rcu_torture_stall");
+       while (!kthread_should_stop())
+               schedule_timeout_interruptible(10 * HZ);
+       return 0;
+ }
+ /* Spawn CPU-stall kthread, if stall_cpu specified. */
+ static int __init rcu_torture_stall_init(void)
+ {
+       if (stall_cpu <= 0)
+               return 0;
+       return torture_create_kthread(rcu_torture_stall, NULL, stall_task);
+ }
+ /* Callback function for RCU barrier testing. */
+ void rcu_torture_barrier_cbf(struct rcu_head *rcu)
+ {
+       atomic_inc(&barrier_cbs_invoked);
+ }
+ /* kthread function to register callbacks used to test RCU barriers. */
+ static int rcu_torture_barrier_cbs(void *arg)
+ {
+       long myid = (long)arg;
+       bool lastphase = 0;
+       bool newphase;
+       struct rcu_head rcu;
+       init_rcu_head_on_stack(&rcu);
+       VERBOSE_TOROUT_STRING("rcu_torture_barrier_cbs task started");
+       set_user_nice(current, 19);
+       do {
+               wait_event(barrier_cbs_wq[myid],
+                          (newphase =
+                           ACCESS_ONCE(barrier_phase)) != lastphase ||
+                          torture_must_stop());
+               lastphase = newphase;
+               smp_mb(); /* ensure barrier_phase load before ->call(). */
+               if (torture_must_stop())
+                       break;
+               cur_ops->call(&rcu, rcu_torture_barrier_cbf);
+               if (atomic_dec_and_test(&barrier_cbs_count))
+                       wake_up(&barrier_wq);
+       } while (!torture_must_stop());
+       cur_ops->cb_barrier();
+       destroy_rcu_head_on_stack(&rcu);
+       torture_kthread_stopping("rcu_torture_barrier_cbs");
+       return 0;
+ }
+ /* kthread function to drive and coordinate RCU barrier testing. */
+ static int rcu_torture_barrier(void *arg)
+ {
+       int i;
+       VERBOSE_TOROUT_STRING("rcu_torture_barrier task starting");
+       do {
+               atomic_set(&barrier_cbs_invoked, 0);
+               atomic_set(&barrier_cbs_count, n_barrier_cbs);
+               smp_mb(); /* Ensure barrier_phase after prior assignments. */
+               barrier_phase = !barrier_phase;
+               for (i = 0; i < n_barrier_cbs; i++)
+                       wake_up(&barrier_cbs_wq[i]);
+               wait_event(barrier_wq,
+                          atomic_read(&barrier_cbs_count) == 0 ||
+                          torture_must_stop());
+               if (torture_must_stop())
+                       break;
+               n_barrier_attempts++;
+               cur_ops->cb_barrier(); /* Implies smp_mb() for wait_event(). */
+               if (atomic_read(&barrier_cbs_invoked) != n_barrier_cbs) {
+                       n_rcu_torture_barrier_error++;
+                       WARN_ON_ONCE(1);
+               }
+               n_barrier_successes++;
+               schedule_timeout_interruptible(HZ / 10);
+       } while (!torture_must_stop());
+       torture_kthread_stopping("rcu_torture_barrier");
+       return 0;
+ }
+ /* Initialize RCU barrier testing. */
+ static int rcu_torture_barrier_init(void)
+ {
+       int i;
+       int ret;
+       if (n_barrier_cbs == 0)
+               return 0;
+       if (cur_ops->call == NULL || cur_ops->cb_barrier == NULL) {
+               pr_alert("%s" TORTURE_FLAG
+                        " Call or barrier ops missing for %s,\n",
+                        torture_type, cur_ops->name);
+               pr_alert("%s" TORTURE_FLAG
+                        " RCU barrier testing omitted from run.\n",
+                        torture_type);
+               return 0;
+       }
+       atomic_set(&barrier_cbs_count, 0);
+       atomic_set(&barrier_cbs_invoked, 0);
+       barrier_cbs_tasks =
+               kzalloc(n_barrier_cbs * sizeof(barrier_cbs_tasks[0]),
+                       GFP_KERNEL);
+       barrier_cbs_wq =
+               kzalloc(n_barrier_cbs * sizeof(barrier_cbs_wq[0]),
+                       GFP_KERNEL);
+       if (barrier_cbs_tasks == NULL || !barrier_cbs_wq)
+               return -ENOMEM;
+       for (i = 0; i < n_barrier_cbs; i++) {
+               init_waitqueue_head(&barrier_cbs_wq[i]);
+               ret = torture_create_kthread(rcu_torture_barrier_cbs,
+                                            (void *)(long)i,
+                                            barrier_cbs_tasks[i]);
+               if (ret)
+                       return ret;
+       }
+       return torture_create_kthread(rcu_torture_barrier, NULL, barrier_task);
+ }
+ /* Clean up after RCU barrier testing. */
+ static void rcu_torture_barrier_cleanup(void)
+ {
+       int i;
+       torture_stop_kthread(rcu_torture_barrier, barrier_task);
+       if (barrier_cbs_tasks != NULL) {
+               for (i = 0; i < n_barrier_cbs; i++)
+                       torture_stop_kthread(rcu_torture_barrier_cbs,
+                                            barrier_cbs_tasks[i]);
+               kfree(barrier_cbs_tasks);
+               barrier_cbs_tasks = NULL;
+       }
+       if (barrier_cbs_wq != NULL) {
+               kfree(barrier_cbs_wq);
+               barrier_cbs_wq = NULL;
+       }
+ }
+ static int rcutorture_cpu_notify(struct notifier_block *self,
+                                unsigned long action, void *hcpu)
+ {
+       long cpu = (long)hcpu;
+       switch (action) {
+       case CPU_ONLINE:
+       case CPU_DOWN_FAILED:
+               (void)rcutorture_booster_init(cpu);
+               break;
+       case CPU_DOWN_PREPARE:
+               rcutorture_booster_cleanup(cpu);
+               break;
+       default:
+               break;
+       }
+       return NOTIFY_OK;
+ }
+ static struct notifier_block rcutorture_cpu_nb = {
+       .notifier_call = rcutorture_cpu_notify,
+ };
+ static void
+ rcu_torture_cleanup(void)
+ {
+       int i;
+       rcutorture_record_test_transition();
+       if (torture_cleanup()) {
+               if (cur_ops->cb_barrier != NULL)
+                       cur_ops->cb_barrier();
+               return;
+       }
+       rcu_torture_barrier_cleanup();
+       torture_stop_kthread(rcu_torture_stall, stall_task);
+       torture_stop_kthread(rcu_torture_writer, writer_task);
+       if (reader_tasks) {
+               for (i = 0; i < nrealreaders; i++)
+                       torture_stop_kthread(rcu_torture_reader,
+                                            reader_tasks[i]);
+               kfree(reader_tasks);
+       }
+       rcu_torture_current = NULL;
+       if (fakewriter_tasks) {
+               for (i = 0; i < nfakewriters; i++) {
+                       torture_stop_kthread(rcu_torture_fakewriter,
+                                            fakewriter_tasks[i]);
+               }
+               kfree(fakewriter_tasks);
+               fakewriter_tasks = NULL;
+       }
+       torture_stop_kthread(rcu_torture_stats, stats_task);
+       torture_stop_kthread(rcu_torture_fqs, fqs_task);
+       if ((test_boost == 1 && cur_ops->can_boost) ||
+           test_boost == 2) {
+               unregister_cpu_notifier(&rcutorture_cpu_nb);
+               for_each_possible_cpu(i)
+                       rcutorture_booster_cleanup(i);
+       }
+       /* Wait for all RCU callbacks to fire.  */
+       if (cur_ops->cb_barrier != NULL)
+               cur_ops->cb_barrier();
+       rcu_torture_stats_print();  /* -After- the stats thread is stopped! */
+       if (atomic_read(&n_rcu_torture_error) || n_rcu_torture_barrier_error)
+               rcu_torture_print_module_parms(cur_ops, "End of test: FAILURE");
+       else if (torture_onoff_failures())
+               rcu_torture_print_module_parms(cur_ops,
+                                              "End of test: RCU_HOTPLUG");
+       else
+               rcu_torture_print_module_parms(cur_ops, "End of test: SUCCESS");
+ }
+ #ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD
+ static void rcu_torture_leak_cb(struct rcu_head *rhp)
+ {
+ }
+ static void rcu_torture_err_cb(struct rcu_head *rhp)
+ {
+       /*
+        * This -might- happen due to race conditions, but is unlikely.
+        * The scenario that leads to this happening is that the
+        * first of the pair of duplicate callbacks is queued,
+        * someone else starts a grace period that includes that
+        * callback, then the second of the pair must wait for the
+        * next grace period.  Unlikely, but can happen.  If it
+        * does happen, the debug-objects subsystem won't have splatted.
+        */
+       pr_alert("rcutorture: duplicated callback was invoked.\n");
+ }
+ #endif /* #ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD */
+ /*
+  * Verify that double-free causes debug-objects to complain, but only
+  * if CONFIG_DEBUG_OBJECTS_RCU_HEAD=y.  Otherwise, say that the test
+  * cannot be carried out.
+  */
+ static void rcu_test_debug_objects(void)
+ {
+ #ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD
+       struct rcu_head rh1;
+       struct rcu_head rh2;
+       init_rcu_head_on_stack(&rh1);
+       init_rcu_head_on_stack(&rh2);
+       pr_alert("rcutorture: WARN: Duplicate call_rcu() test starting.\n");
+       /* Try to queue the rh2 pair of callbacks for the same grace period. */
+       preempt_disable(); /* Prevent preemption from interrupting test. */
+       rcu_read_lock(); /* Make it impossible to finish a grace period. */
+       call_rcu(&rh1, rcu_torture_leak_cb); /* Start grace period. */
+       local_irq_disable(); /* Make it harder to start a new grace period. */
+       call_rcu(&rh2, rcu_torture_leak_cb);
+       call_rcu(&rh2, rcu_torture_err_cb); /* Duplicate callback. */
+       local_irq_enable();
+       rcu_read_unlock();
+       preempt_enable();
+       /* Wait for them all to get done so we can safely return. */
+       rcu_barrier();
+       pr_alert("rcutorture: WARN: Duplicate call_rcu() test complete.\n");
+       destroy_rcu_head_on_stack(&rh1);
+       destroy_rcu_head_on_stack(&rh2);
+ #else /* #ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD */
+       pr_alert("rcutorture: !CONFIG_DEBUG_OBJECTS_RCU_HEAD, not testing duplicate call_rcu()\n");
+ #endif /* #else #ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD */
+ }
+ static int __init
+ rcu_torture_init(void)
+ {
+       int i;
+       int cpu;
+       int firsterr = 0;
+       static struct rcu_torture_ops *torture_ops[] = {
+               &rcu_ops, &rcu_bh_ops, &rcu_busted_ops, &srcu_ops, &sched_ops,
+       };
+       torture_init_begin(torture_type, verbose, &rcutorture_runnable);
+       /* Process args and tell the world that the torturer is on the job. */
+       for (i = 0; i < ARRAY_SIZE(torture_ops); i++) {
+               cur_ops = torture_ops[i];
+               if (strcmp(torture_type, cur_ops->name) == 0)
+                       break;
+       }
+       if (i == ARRAY_SIZE(torture_ops)) {
+               pr_alert("rcu-torture: invalid torture type: \"%s\"\n",
+                        torture_type);
+               pr_alert("rcu-torture types:");
+               for (i = 0; i < ARRAY_SIZE(torture_ops); i++)
+                       pr_alert(" %s", torture_ops[i]->name);
+               pr_alert("\n");
+               torture_init_end();
+               return -EINVAL;
+       }
+       if (cur_ops->fqs == NULL && fqs_duration != 0) {
+               pr_alert("rcu-torture: ->fqs NULL and non-zero fqs_duration, fqs disabled.\n");
+               fqs_duration = 0;
+       }
+       if (cur_ops->init)
+               cur_ops->init(); /* no "goto unwind" prior to this point!!! */
+       if (nreaders >= 0)
+               nrealreaders = nreaders;
+       else
+               nrealreaders = 2 * num_online_cpus();
+       rcu_torture_print_module_parms(cur_ops, "Start of test");
+       /* Set up the freelist. */
+       INIT_LIST_HEAD(&rcu_torture_freelist);
+       for (i = 0; i < ARRAY_SIZE(rcu_tortures); i++) {
+               rcu_tortures[i].rtort_mbtest = 0;
+               list_add_tail(&rcu_tortures[i].rtort_free,
+                             &rcu_torture_freelist);
+       }
+       /* Initialize the statistics so that each run gets its own numbers. */
+       rcu_torture_current = NULL;
+       rcu_torture_current_version = 0;
+       atomic_set(&n_rcu_torture_alloc, 0);
+       atomic_set(&n_rcu_torture_alloc_fail, 0);
+       atomic_set(&n_rcu_torture_free, 0);
+       atomic_set(&n_rcu_torture_mberror, 0);
+       atomic_set(&n_rcu_torture_error, 0);
+       n_rcu_torture_barrier_error = 0;
+       n_rcu_torture_boost_ktrerror = 0;
+       n_rcu_torture_boost_rterror = 0;
+       n_rcu_torture_boost_failure = 0;
+       n_rcu_torture_boosts = 0;
+       for (i = 0; i < RCU_TORTURE_PIPE_LEN + 1; i++)
+               atomic_set(&rcu_torture_wcount[i], 0);
+       for_each_possible_cpu(cpu) {
+               for (i = 0; i < RCU_TORTURE_PIPE_LEN + 1; i++) {
+                       per_cpu(rcu_torture_count, cpu)[i] = 0;
+                       per_cpu(rcu_torture_batch, cpu)[i] = 0;
+               }
+       }
+       /* Start up the kthreads. */
+       firsterr = torture_create_kthread(rcu_torture_writer, NULL,
+                                         writer_task);
+       if (firsterr)
+               goto unwind;
+       fakewriter_tasks = kzalloc(nfakewriters * sizeof(fakewriter_tasks[0]),
+                                  GFP_KERNEL);
+       if (fakewriter_tasks == NULL) {
+               VERBOSE_TOROUT_ERRSTRING("out of memory");
+               firsterr = -ENOMEM;
+               goto unwind;
+       }
+       for (i = 0; i < nfakewriters; i++) {
+               firsterr = torture_create_kthread(rcu_torture_fakewriter,
+                                                 NULL, fakewriter_tasks[i]);
+               if (firsterr)
+                       goto unwind;
+       }
+       reader_tasks = kzalloc(nrealreaders * sizeof(reader_tasks[0]),
+                              GFP_KERNEL);
+       if (reader_tasks == NULL) {
+               VERBOSE_TOROUT_ERRSTRING("out of memory");
+               firsterr = -ENOMEM;
+               goto unwind;
+       }
+       for (i = 0; i < nrealreaders; i++) {
+               firsterr = torture_create_kthread(rcu_torture_reader, NULL,
+                                                 reader_tasks[i]);
+               if (firsterr)
+                       goto unwind;
+       }
+       if (stat_interval > 0) {
+               firsterr = torture_create_kthread(rcu_torture_stats, NULL,
+                                                 stats_task);
+               if (firsterr)
+                       goto unwind;
+       }
+       if (test_no_idle_hz) {
+               firsterr = torture_shuffle_init(shuffle_interval * HZ);
+               if (firsterr)
+                       goto unwind;
+       }
+       if (stutter < 0)
+               stutter = 0;
+       if (stutter) {
+               firsterr = torture_stutter_init(stutter * HZ);
+               if (firsterr)
+                       goto unwind;
+       }
+       if (fqs_duration < 0)
+               fqs_duration = 0;
+       if (fqs_duration) {
+               /* Create the fqs thread */
+               torture_create_kthread(rcu_torture_fqs, NULL, fqs_task);
+               if (firsterr)
+                       goto unwind;
+       }
+       if (test_boost_interval < 1)
+               test_boost_interval = 1;
+       if (test_boost_duration < 2)
+               test_boost_duration = 2;
+       if ((test_boost == 1 && cur_ops->can_boost) ||
+           test_boost == 2) {
+               boost_starttime = jiffies + test_boost_interval * HZ;
+               register_cpu_notifier(&rcutorture_cpu_nb);
+               for_each_possible_cpu(i) {
+                       if (cpu_is_offline(i))
+                               continue;  /* Heuristic: CPU can go offline. */
+                       firsterr = rcutorture_booster_init(i);
+                       if (firsterr)
+                               goto unwind;
+               }
+       }
+       firsterr = torture_shutdown_init(shutdown_secs, rcu_torture_cleanup);
+       if (firsterr)
+               goto unwind;
+       firsterr = torture_onoff_init(onoff_holdoff * HZ, onoff_interval * HZ);
+       if (firsterr)
+               goto unwind;
+       firsterr = rcu_torture_stall_init();
+       if (firsterr)
+               goto unwind;
+       firsterr = rcu_torture_barrier_init();
+       if (firsterr)
+               goto unwind;
+       if (object_debug)
+               rcu_test_debug_objects();
+       rcutorture_record_test_transition();
+       torture_init_end();
+       return 0;
+ unwind:
+       torture_init_end();
+       rcu_torture_cleanup();
+       return firsterr;
+ }
+ module_init(rcu_torture_init);
+ module_exit(rcu_torture_cleanup);