Back to home page

OSCL-LXR
 
 

     

0001 /* SPDX-License-Identifier: GPL-2.0+ */
0002 /*
0003  * Task-based RCU implementations.
0004  *
0005  * Copyright (C) 2020 Paul E. McKenney
0006  */
0007 
0008 #ifdef CONFIG_TASKS_RCU_GENERIC
0009 #include "rcu_segcblist.h"
0010 
0011 ////////////////////////////////////////////////////////////////////////
0012 //
0013 // Generic data structures.
0014 
0015 struct rcu_tasks;
0016 typedef void (*rcu_tasks_gp_func_t)(struct rcu_tasks *rtp);
0017 typedef void (*pregp_func_t)(struct list_head *hop);
0018 typedef void (*pertask_func_t)(struct task_struct *t, struct list_head *hop);
0019 typedef void (*postscan_func_t)(struct list_head *hop);
0020 typedef void (*holdouts_func_t)(struct list_head *hop, bool ndrpt, bool *frptp);
0021 typedef void (*postgp_func_t)(struct rcu_tasks *rtp);
0022 
0023 /**
0024  * struct rcu_tasks_percpu - Per-CPU component of definition for a Tasks-RCU-like mechanism.
0025  * @cblist: Callback list.
0026  * @lock: Lock protecting per-CPU callback list.
0027  * @rtp_jiffies: Jiffies counter value for statistics.
0028  * @rtp_n_lock_retries: Rough lock-contention statistic.
0029  * @rtp_work: Work queue for invoking callbacks.
0030  * @rtp_irq_work: IRQ work queue for deferred wakeups.
0031  * @barrier_q_head: RCU callback for barrier operation.
0032  * @rtp_blkd_tasks: List of tasks blocked as readers.
0033  * @cpu: CPU number corresponding to this entry.
0034  * @rtpp: Pointer to the rcu_tasks structure.
0035  */
0036 struct rcu_tasks_percpu {
0037     struct rcu_segcblist cblist;
0038     raw_spinlock_t __private lock;
0039     unsigned long rtp_jiffies;
0040     unsigned long rtp_n_lock_retries;
0041     struct work_struct rtp_work;
0042     struct irq_work rtp_irq_work;
0043     struct rcu_head barrier_q_head;
0044     struct list_head rtp_blkd_tasks;
0045     int cpu;
0046     struct rcu_tasks *rtpp;
0047 };
0048 
0049 /**
0050  * struct rcu_tasks - Definition for a Tasks-RCU-like mechanism.
0051  * @cbs_wait: RCU wait allowing a new callback to get kthread's attention.
0052  * @cbs_gbl_lock: Lock protecting callback list.
0053  * @tasks_gp_mutex: Mutex protecting grace period, needed during mid-boot dead zone.
0054  * @kthread_ptr: This flavor's grace-period/callback-invocation kthread.
0055  * @gp_func: This flavor's grace-period-wait function.
0056  * @gp_state: Grace period's most recent state transition (debugging).
0057  * @gp_sleep: Per-grace-period sleep to prevent CPU-bound looping.
0058  * @init_fract: Initial backoff sleep interval.
0059  * @gp_jiffies: Time of last @gp_state transition.
0060  * @gp_start: Most recent grace-period start in jiffies.
0061  * @tasks_gp_seq: Number of grace periods completed since boot.
0062  * @n_ipis: Number of IPIs sent to encourage grace periods to end.
0063  * @n_ipis_fails: Number of IPI-send failures.
0064  * @pregp_func: This flavor's pre-grace-period function (optional).
0065  * @pertask_func: This flavor's per-task scan function (optional).
0066  * @postscan_func: This flavor's post-task scan function (optional).
0067  * @holdouts_func: This flavor's holdout-list scan function (optional).
0068  * @postgp_func: This flavor's post-grace-period function (optional).
0069  * @call_func: This flavor's call_rcu()-equivalent function.
0070  * @rtpcpu: This flavor's rcu_tasks_percpu structure.
0071  * @percpu_enqueue_shift: Shift down CPU ID this much when enqueuing callbacks.
0072  * @percpu_enqueue_lim: Number of per-CPU callback queues in use for enqueuing.
0073  * @percpu_dequeue_lim: Number of per-CPU callback queues in use for dequeuing.
0074  * @percpu_dequeue_gpseq: RCU grace-period number to propagate enqueue limit to dequeuers.
0075  * @barrier_q_mutex: Serialize barrier operations.
0076  * @barrier_q_count: Number of queues being waited on.
0077  * @barrier_q_completion: Barrier wait/wakeup mechanism.
0078  * @barrier_q_seq: Sequence number for barrier operations.
0079  * @name: This flavor's textual name.
0080  * @kname: This flavor's kthread name.
0081  */
0082 struct rcu_tasks {
0083     struct rcuwait cbs_wait;
0084     raw_spinlock_t cbs_gbl_lock;
0085     struct mutex tasks_gp_mutex;
0086     int gp_state;
0087     int gp_sleep;
0088     int init_fract;
0089     unsigned long gp_jiffies;
0090     unsigned long gp_start;
0091     unsigned long tasks_gp_seq;
0092     unsigned long n_ipis;
0093     unsigned long n_ipis_fails;
0094     struct task_struct *kthread_ptr;
0095     rcu_tasks_gp_func_t gp_func;
0096     pregp_func_t pregp_func;
0097     pertask_func_t pertask_func;
0098     postscan_func_t postscan_func;
0099     holdouts_func_t holdouts_func;
0100     postgp_func_t postgp_func;
0101     call_rcu_func_t call_func;
0102     struct rcu_tasks_percpu __percpu *rtpcpu;
0103     int percpu_enqueue_shift;
0104     int percpu_enqueue_lim;
0105     int percpu_dequeue_lim;
0106     unsigned long percpu_dequeue_gpseq;
0107     struct mutex barrier_q_mutex;
0108     atomic_t barrier_q_count;
0109     struct completion barrier_q_completion;
0110     unsigned long barrier_q_seq;
0111     char *name;
0112     char *kname;
0113 };
0114 
0115 static void call_rcu_tasks_iw_wakeup(struct irq_work *iwp);
0116 
0117 #define DEFINE_RCU_TASKS(rt_name, gp, call, n)                      \
0118 static DEFINE_PER_CPU(struct rcu_tasks_percpu, rt_name ## __percpu) = {         \
0119     .lock = __RAW_SPIN_LOCK_UNLOCKED(rt_name ## __percpu.cbs_pcpu_lock),        \
0120     .rtp_irq_work = IRQ_WORK_INIT_HARD(call_rcu_tasks_iw_wakeup),           \
0121 };                                          \
0122 static struct rcu_tasks rt_name =                           \
0123 {                                           \
0124     .cbs_wait = __RCUWAIT_INITIALIZER(rt_name.wait),                \
0125     .cbs_gbl_lock = __RAW_SPIN_LOCK_UNLOCKED(rt_name.cbs_gbl_lock),         \
0126     .tasks_gp_mutex = __MUTEX_INITIALIZER(rt_name.tasks_gp_mutex),          \
0127     .gp_func = gp,                                  \
0128     .call_func = call,                              \
0129     .rtpcpu = &rt_name ## __percpu,                         \
0130     .name = n,                                  \
0131     .percpu_enqueue_shift = order_base_2(CONFIG_NR_CPUS),               \
0132     .percpu_enqueue_lim = 1,                            \
0133     .percpu_dequeue_lim = 1,                            \
0134     .barrier_q_mutex = __MUTEX_INITIALIZER(rt_name.barrier_q_mutex),        \
0135     .barrier_q_seq = (0UL - 50UL) << RCU_SEQ_CTR_SHIFT,             \
0136     .kname = #rt_name,                              \
0137 }
0138 
0139 /* Track exiting tasks in order to allow them to be waited for. */
0140 DEFINE_STATIC_SRCU(tasks_rcu_exit_srcu);
0141 
0142 /* Avoid IPIing CPUs early in the grace period. */
0143 #define RCU_TASK_IPI_DELAY (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB) ? HZ / 2 : 0)
0144 static int rcu_task_ipi_delay __read_mostly = RCU_TASK_IPI_DELAY;
0145 module_param(rcu_task_ipi_delay, int, 0644);
0146 
0147 /* Control stall timeouts.  Disable with <= 0, otherwise jiffies till stall. */
0148 #define RCU_TASK_BOOT_STALL_TIMEOUT (HZ * 30)
0149 #define RCU_TASK_STALL_TIMEOUT (HZ * 60 * 10)
0150 static int rcu_task_stall_timeout __read_mostly = RCU_TASK_STALL_TIMEOUT;
0151 module_param(rcu_task_stall_timeout, int, 0644);
0152 #define RCU_TASK_STALL_INFO (HZ * 10)
0153 static int rcu_task_stall_info __read_mostly = RCU_TASK_STALL_INFO;
0154 module_param(rcu_task_stall_info, int, 0644);
0155 static int rcu_task_stall_info_mult __read_mostly = 3;
0156 module_param(rcu_task_stall_info_mult, int, 0444);
0157 
0158 static int rcu_task_enqueue_lim __read_mostly = -1;
0159 module_param(rcu_task_enqueue_lim, int, 0444);
0160 
0161 static bool rcu_task_cb_adjust;
0162 static int rcu_task_contend_lim __read_mostly = 100;
0163 module_param(rcu_task_contend_lim, int, 0444);
0164 static int rcu_task_collapse_lim __read_mostly = 10;
0165 module_param(rcu_task_collapse_lim, int, 0444);
0166 
0167 /* RCU tasks grace-period state for debugging. */
0168 #define RTGS_INIT        0
0169 #define RTGS_WAIT_WAIT_CBS   1
0170 #define RTGS_WAIT_GP         2
0171 #define RTGS_PRE_WAIT_GP     3
0172 #define RTGS_SCAN_TASKLIST   4
0173 #define RTGS_POST_SCAN_TASKLIST  5
0174 #define RTGS_WAIT_SCAN_HOLDOUTS  6
0175 #define RTGS_SCAN_HOLDOUTS   7
0176 #define RTGS_POST_GP         8
0177 #define RTGS_WAIT_READERS    9
0178 #define RTGS_INVOKE_CBS     10
0179 #define RTGS_WAIT_CBS       11
0180 #ifndef CONFIG_TINY_RCU
0181 static const char * const rcu_tasks_gp_state_names[] = {
0182     "RTGS_INIT",
0183     "RTGS_WAIT_WAIT_CBS",
0184     "RTGS_WAIT_GP",
0185     "RTGS_PRE_WAIT_GP",
0186     "RTGS_SCAN_TASKLIST",
0187     "RTGS_POST_SCAN_TASKLIST",
0188     "RTGS_WAIT_SCAN_HOLDOUTS",
0189     "RTGS_SCAN_HOLDOUTS",
0190     "RTGS_POST_GP",
0191     "RTGS_WAIT_READERS",
0192     "RTGS_INVOKE_CBS",
0193     "RTGS_WAIT_CBS",
0194 };
0195 #endif /* #ifndef CONFIG_TINY_RCU */
0196 
0197 ////////////////////////////////////////////////////////////////////////
0198 //
0199 // Generic code.
0200 
0201 static void rcu_tasks_invoke_cbs_wq(struct work_struct *wp);
0202 
0203 /* Record grace-period phase and time. */
0204 static void set_tasks_gp_state(struct rcu_tasks *rtp, int newstate)
0205 {
0206     rtp->gp_state = newstate;
0207     rtp->gp_jiffies = jiffies;
0208 }
0209 
0210 #ifndef CONFIG_TINY_RCU
0211 /* Return state name. */
0212 static const char *tasks_gp_state_getname(struct rcu_tasks *rtp)
0213 {
0214     int i = data_race(rtp->gp_state); // Let KCSAN detect update races
0215     int j = READ_ONCE(i); // Prevent the compiler from reading twice
0216 
0217     if (j >= ARRAY_SIZE(rcu_tasks_gp_state_names))
0218         return "???";
0219     return rcu_tasks_gp_state_names[j];
0220 }
0221 #endif /* #ifndef CONFIG_TINY_RCU */
0222 
0223 // Initialize per-CPU callback lists for the specified flavor of
0224 // Tasks RCU.
0225 static void cblist_init_generic(struct rcu_tasks *rtp)
0226 {
0227     int cpu;
0228     unsigned long flags;
0229     int lim;
0230     int shift;
0231 
0232     raw_spin_lock_irqsave(&rtp->cbs_gbl_lock, flags);
0233     if (rcu_task_enqueue_lim < 0) {
0234         rcu_task_enqueue_lim = 1;
0235         rcu_task_cb_adjust = true;
0236         pr_info("%s: Setting adjustable number of callback queues.\n", __func__);
0237     } else if (rcu_task_enqueue_lim == 0) {
0238         rcu_task_enqueue_lim = 1;
0239     }
0240     lim = rcu_task_enqueue_lim;
0241 
0242     if (lim > nr_cpu_ids)
0243         lim = nr_cpu_ids;
0244     shift = ilog2(nr_cpu_ids / lim);
0245     if (((nr_cpu_ids - 1) >> shift) >= lim)
0246         shift++;
0247     WRITE_ONCE(rtp->percpu_enqueue_shift, shift);
0248     WRITE_ONCE(rtp->percpu_dequeue_lim, lim);
0249     smp_store_release(&rtp->percpu_enqueue_lim, lim);
0250     for_each_possible_cpu(cpu) {
0251         struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
0252 
0253         WARN_ON_ONCE(!rtpcp);
0254         if (cpu)
0255             raw_spin_lock_init(&ACCESS_PRIVATE(rtpcp, lock));
0256         raw_spin_lock_rcu_node(rtpcp); // irqs already disabled.
0257         if (rcu_segcblist_empty(&rtpcp->cblist))
0258             rcu_segcblist_init(&rtpcp->cblist);
0259         INIT_WORK(&rtpcp->rtp_work, rcu_tasks_invoke_cbs_wq);
0260         rtpcp->cpu = cpu;
0261         rtpcp->rtpp = rtp;
0262         if (!rtpcp->rtp_blkd_tasks.next)
0263             INIT_LIST_HEAD(&rtpcp->rtp_blkd_tasks);
0264         raw_spin_unlock_rcu_node(rtpcp); // irqs remain disabled.
0265     }
0266     raw_spin_unlock_irqrestore(&rtp->cbs_gbl_lock, flags);
0267     pr_info("%s: Setting shift to %d and lim to %d.\n", __func__, data_race(rtp->percpu_enqueue_shift), data_race(rtp->percpu_enqueue_lim));
0268 }
0269 
0270 // IRQ-work handler that does deferred wakeup for call_rcu_tasks_generic().
0271 static void call_rcu_tasks_iw_wakeup(struct irq_work *iwp)
0272 {
0273     struct rcu_tasks *rtp;
0274     struct rcu_tasks_percpu *rtpcp = container_of(iwp, struct rcu_tasks_percpu, rtp_irq_work);
0275 
0276     rtp = rtpcp->rtpp;
0277     rcuwait_wake_up(&rtp->cbs_wait);
0278 }
0279 
0280 // Enqueue a callback for the specified flavor of Tasks RCU.
0281 static void call_rcu_tasks_generic(struct rcu_head *rhp, rcu_callback_t func,
0282                    struct rcu_tasks *rtp)
0283 {
0284     int chosen_cpu;
0285     unsigned long flags;
0286     int ideal_cpu;
0287     unsigned long j;
0288     bool needadjust = false;
0289     bool needwake;
0290     struct rcu_tasks_percpu *rtpcp;
0291 
0292     rhp->next = NULL;
0293     rhp->func = func;
0294     local_irq_save(flags);
0295     rcu_read_lock();
0296     ideal_cpu = smp_processor_id() >> READ_ONCE(rtp->percpu_enqueue_shift);
0297     chosen_cpu = cpumask_next(ideal_cpu - 1, cpu_possible_mask);
0298     rtpcp = per_cpu_ptr(rtp->rtpcpu, chosen_cpu);
0299     if (!raw_spin_trylock_rcu_node(rtpcp)) { // irqs already disabled.
0300         raw_spin_lock_rcu_node(rtpcp); // irqs already disabled.
0301         j = jiffies;
0302         if (rtpcp->rtp_jiffies != j) {
0303             rtpcp->rtp_jiffies = j;
0304             rtpcp->rtp_n_lock_retries = 0;
0305         }
0306         if (rcu_task_cb_adjust && ++rtpcp->rtp_n_lock_retries > rcu_task_contend_lim &&
0307             READ_ONCE(rtp->percpu_enqueue_lim) != nr_cpu_ids)
0308             needadjust = true;  // Defer adjustment to avoid deadlock.
0309     }
0310     if (!rcu_segcblist_is_enabled(&rtpcp->cblist)) {
0311         raw_spin_unlock_rcu_node(rtpcp); // irqs remain disabled.
0312         cblist_init_generic(rtp);
0313         raw_spin_lock_rcu_node(rtpcp); // irqs already disabled.
0314     }
0315     needwake = rcu_segcblist_empty(&rtpcp->cblist);
0316     rcu_segcblist_enqueue(&rtpcp->cblist, rhp);
0317     raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
0318     if (unlikely(needadjust)) {
0319         raw_spin_lock_irqsave(&rtp->cbs_gbl_lock, flags);
0320         if (rtp->percpu_enqueue_lim != nr_cpu_ids) {
0321             WRITE_ONCE(rtp->percpu_enqueue_shift, 0);
0322             WRITE_ONCE(rtp->percpu_dequeue_lim, nr_cpu_ids);
0323             smp_store_release(&rtp->percpu_enqueue_lim, nr_cpu_ids);
0324             pr_info("Switching %s to per-CPU callback queuing.\n", rtp->name);
0325         }
0326         raw_spin_unlock_irqrestore(&rtp->cbs_gbl_lock, flags);
0327     }
0328     rcu_read_unlock();
0329     /* We can't create the thread unless interrupts are enabled. */
0330     if (needwake && READ_ONCE(rtp->kthread_ptr))
0331         irq_work_queue(&rtpcp->rtp_irq_work);
0332 }
0333 
0334 // RCU callback function for rcu_barrier_tasks_generic().
0335 static void rcu_barrier_tasks_generic_cb(struct rcu_head *rhp)
0336 {
0337     struct rcu_tasks *rtp;
0338     struct rcu_tasks_percpu *rtpcp;
0339 
0340     rtpcp = container_of(rhp, struct rcu_tasks_percpu, barrier_q_head);
0341     rtp = rtpcp->rtpp;
0342     if (atomic_dec_and_test(&rtp->barrier_q_count))
0343         complete(&rtp->barrier_q_completion);
0344 }
0345 
0346 // Wait for all in-flight callbacks for the specified RCU Tasks flavor.
0347 // Operates in a manner similar to rcu_barrier().
0348 static void rcu_barrier_tasks_generic(struct rcu_tasks *rtp)
0349 {
0350     int cpu;
0351     unsigned long flags;
0352     struct rcu_tasks_percpu *rtpcp;
0353     unsigned long s = rcu_seq_snap(&rtp->barrier_q_seq);
0354 
0355     mutex_lock(&rtp->barrier_q_mutex);
0356     if (rcu_seq_done(&rtp->barrier_q_seq, s)) {
0357         smp_mb();
0358         mutex_unlock(&rtp->barrier_q_mutex);
0359         return;
0360     }
0361     rcu_seq_start(&rtp->barrier_q_seq);
0362     init_completion(&rtp->barrier_q_completion);
0363     atomic_set(&rtp->barrier_q_count, 2);
0364     for_each_possible_cpu(cpu) {
0365         if (cpu >= smp_load_acquire(&rtp->percpu_dequeue_lim))
0366             break;
0367         rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
0368         rtpcp->barrier_q_head.func = rcu_barrier_tasks_generic_cb;
0369         raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
0370         if (rcu_segcblist_entrain(&rtpcp->cblist, &rtpcp->barrier_q_head))
0371             atomic_inc(&rtp->barrier_q_count);
0372         raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
0373     }
0374     if (atomic_sub_and_test(2, &rtp->barrier_q_count))
0375         complete(&rtp->barrier_q_completion);
0376     wait_for_completion(&rtp->barrier_q_completion);
0377     rcu_seq_end(&rtp->barrier_q_seq);
0378     mutex_unlock(&rtp->barrier_q_mutex);
0379 }
0380 
0381 // Advance callbacks and indicate whether either a grace period or
0382 // callback invocation is needed.
0383 static int rcu_tasks_need_gpcb(struct rcu_tasks *rtp)
0384 {
0385     int cpu;
0386     unsigned long flags;
0387     long n;
0388     long ncbs = 0;
0389     long ncbsnz = 0;
0390     int needgpcb = 0;
0391 
0392     for (cpu = 0; cpu < smp_load_acquire(&rtp->percpu_dequeue_lim); cpu++) {
0393         struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
0394 
0395         /* Advance and accelerate any new callbacks. */
0396         if (!rcu_segcblist_n_cbs(&rtpcp->cblist))
0397             continue;
0398         raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
0399         // Should we shrink down to a single callback queue?
0400         n = rcu_segcblist_n_cbs(&rtpcp->cblist);
0401         if (n) {
0402             ncbs += n;
0403             if (cpu > 0)
0404                 ncbsnz += n;
0405         }
0406         rcu_segcblist_advance(&rtpcp->cblist, rcu_seq_current(&rtp->tasks_gp_seq));
0407         (void)rcu_segcblist_accelerate(&rtpcp->cblist, rcu_seq_snap(&rtp->tasks_gp_seq));
0408         if (rcu_segcblist_pend_cbs(&rtpcp->cblist))
0409             needgpcb |= 0x3;
0410         if (!rcu_segcblist_empty(&rtpcp->cblist))
0411             needgpcb |= 0x1;
0412         raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
0413     }
0414 
0415     // Shrink down to a single callback queue if appropriate.
0416     // This is done in two stages: (1) If there are no more than
0417     // rcu_task_collapse_lim callbacks on CPU 0 and none on any other
0418     // CPU, limit enqueueing to CPU 0.  (2) After an RCU grace period,
0419     // if there has not been an increase in callbacks, limit dequeuing
0420     // to CPU 0.  Note the matching RCU read-side critical section in
0421     // call_rcu_tasks_generic().
0422     if (rcu_task_cb_adjust && ncbs <= rcu_task_collapse_lim) {
0423         raw_spin_lock_irqsave(&rtp->cbs_gbl_lock, flags);
0424         if (rtp->percpu_enqueue_lim > 1) {
0425             WRITE_ONCE(rtp->percpu_enqueue_shift, order_base_2(nr_cpu_ids));
0426             smp_store_release(&rtp->percpu_enqueue_lim, 1);
0427             rtp->percpu_dequeue_gpseq = get_state_synchronize_rcu();
0428             pr_info("Starting switch %s to CPU-0 callback queuing.\n", rtp->name);
0429         }
0430         raw_spin_unlock_irqrestore(&rtp->cbs_gbl_lock, flags);
0431     }
0432     if (rcu_task_cb_adjust && !ncbsnz &&
0433         poll_state_synchronize_rcu(rtp->percpu_dequeue_gpseq)) {
0434         raw_spin_lock_irqsave(&rtp->cbs_gbl_lock, flags);
0435         if (rtp->percpu_enqueue_lim < rtp->percpu_dequeue_lim) {
0436             WRITE_ONCE(rtp->percpu_dequeue_lim, 1);
0437             pr_info("Completing switch %s to CPU-0 callback queuing.\n", rtp->name);
0438         }
0439         for (cpu = rtp->percpu_dequeue_lim; cpu < nr_cpu_ids; cpu++) {
0440             struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
0441 
0442             WARN_ON_ONCE(rcu_segcblist_n_cbs(&rtpcp->cblist));
0443         }
0444         raw_spin_unlock_irqrestore(&rtp->cbs_gbl_lock, flags);
0445     }
0446 
0447     return needgpcb;
0448 }
0449 
0450 // Advance callbacks and invoke any that are ready.
0451 static void rcu_tasks_invoke_cbs(struct rcu_tasks *rtp, struct rcu_tasks_percpu *rtpcp)
0452 {
0453     int cpu;
0454     int cpunext;
0455     unsigned long flags;
0456     int len;
0457     struct rcu_head *rhp;
0458     struct rcu_cblist rcl = RCU_CBLIST_INITIALIZER(rcl);
0459     struct rcu_tasks_percpu *rtpcp_next;
0460 
0461     cpu = rtpcp->cpu;
0462     cpunext = cpu * 2 + 1;
0463     if (cpunext < smp_load_acquire(&rtp->percpu_dequeue_lim)) {
0464         rtpcp_next = per_cpu_ptr(rtp->rtpcpu, cpunext);
0465         queue_work_on(cpunext, system_wq, &rtpcp_next->rtp_work);
0466         cpunext++;
0467         if (cpunext < smp_load_acquire(&rtp->percpu_dequeue_lim)) {
0468             rtpcp_next = per_cpu_ptr(rtp->rtpcpu, cpunext);
0469             queue_work_on(cpunext, system_wq, &rtpcp_next->rtp_work);
0470         }
0471     }
0472 
0473     if (rcu_segcblist_empty(&rtpcp->cblist) || !cpu_possible(cpu))
0474         return;
0475     raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
0476     rcu_segcblist_advance(&rtpcp->cblist, rcu_seq_current(&rtp->tasks_gp_seq));
0477     rcu_segcblist_extract_done_cbs(&rtpcp->cblist, &rcl);
0478     raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
0479     len = rcl.len;
0480     for (rhp = rcu_cblist_dequeue(&rcl); rhp; rhp = rcu_cblist_dequeue(&rcl)) {
0481         local_bh_disable();
0482         rhp->func(rhp);
0483         local_bh_enable();
0484         cond_resched();
0485     }
0486     raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
0487     rcu_segcblist_add_len(&rtpcp->cblist, -len);
0488     (void)rcu_segcblist_accelerate(&rtpcp->cblist, rcu_seq_snap(&rtp->tasks_gp_seq));
0489     raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
0490 }
0491 
0492 // Workqueue flood to advance callbacks and invoke any that are ready.
0493 static void rcu_tasks_invoke_cbs_wq(struct work_struct *wp)
0494 {
0495     struct rcu_tasks *rtp;
0496     struct rcu_tasks_percpu *rtpcp = container_of(wp, struct rcu_tasks_percpu, rtp_work);
0497 
0498     rtp = rtpcp->rtpp;
0499     rcu_tasks_invoke_cbs(rtp, rtpcp);
0500 }
0501 
0502 // Wait for one grace period.
0503 static void rcu_tasks_one_gp(struct rcu_tasks *rtp, bool midboot)
0504 {
0505     int needgpcb;
0506 
0507     mutex_lock(&rtp->tasks_gp_mutex);
0508 
0509     // If there were none, wait a bit and start over.
0510     if (unlikely(midboot)) {
0511         needgpcb = 0x2;
0512     } else {
0513         set_tasks_gp_state(rtp, RTGS_WAIT_CBS);
0514         rcuwait_wait_event(&rtp->cbs_wait,
0515                    (needgpcb = rcu_tasks_need_gpcb(rtp)),
0516                    TASK_IDLE);
0517     }
0518 
0519     if (needgpcb & 0x2) {
0520         // Wait for one grace period.
0521         set_tasks_gp_state(rtp, RTGS_WAIT_GP);
0522         rtp->gp_start = jiffies;
0523         rcu_seq_start(&rtp->tasks_gp_seq);
0524         rtp->gp_func(rtp);
0525         rcu_seq_end(&rtp->tasks_gp_seq);
0526     }
0527 
0528     // Invoke callbacks.
0529     set_tasks_gp_state(rtp, RTGS_INVOKE_CBS);
0530     rcu_tasks_invoke_cbs(rtp, per_cpu_ptr(rtp->rtpcpu, 0));
0531     mutex_unlock(&rtp->tasks_gp_mutex);
0532 }
0533 
0534 // RCU-tasks kthread that detects grace periods and invokes callbacks.
0535 static int __noreturn rcu_tasks_kthread(void *arg)
0536 {
0537     struct rcu_tasks *rtp = arg;
0538 
0539     /* Run on housekeeping CPUs by default.  Sysadm can move if desired. */
0540     housekeeping_affine(current, HK_TYPE_RCU);
0541     WRITE_ONCE(rtp->kthread_ptr, current); // Let GPs start!
0542 
0543     /*
0544      * Each pass through the following loop makes one check for
0545      * newly arrived callbacks, and, if there are some, waits for
0546      * one RCU-tasks grace period and then invokes the callbacks.
0547      * This loop is terminated by the system going down.  ;-)
0548      */
0549     for (;;) {
0550         // Wait for one grace period and invoke any callbacks
0551         // that are ready.
0552         rcu_tasks_one_gp(rtp, false);
0553 
0554         // Paranoid sleep to keep this from entering a tight loop.
0555         schedule_timeout_idle(rtp->gp_sleep);
0556     }
0557 }
0558 
0559 // Wait for a grace period for the specified flavor of Tasks RCU.
0560 static void synchronize_rcu_tasks_generic(struct rcu_tasks *rtp)
0561 {
0562     /* Complain if the scheduler has not started.  */
0563     RCU_LOCKDEP_WARN(rcu_scheduler_active == RCU_SCHEDULER_INACTIVE,
0564              "synchronize_rcu_tasks called too soon");
0565 
0566     // If the grace-period kthread is running, use it.
0567     if (READ_ONCE(rtp->kthread_ptr)) {
0568         wait_rcu_gp(rtp->call_func);
0569         return;
0570     }
0571     rcu_tasks_one_gp(rtp, true);
0572 }
0573 
0574 /* Spawn RCU-tasks grace-period kthread. */
0575 static void __init rcu_spawn_tasks_kthread_generic(struct rcu_tasks *rtp)
0576 {
0577     struct task_struct *t;
0578 
0579     t = kthread_run(rcu_tasks_kthread, rtp, "%s_kthread", rtp->kname);
0580     if (WARN_ONCE(IS_ERR(t), "%s: Could not start %s grace-period kthread, OOM is now expected behavior\n", __func__, rtp->name))
0581         return;
0582     smp_mb(); /* Ensure others see full kthread. */
0583 }
0584 
0585 #ifndef CONFIG_TINY_RCU
0586 
0587 /*
0588  * Print any non-default Tasks RCU settings.
0589  */
0590 static void __init rcu_tasks_bootup_oddness(void)
0591 {
0592 #if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU)
0593     int rtsimc;
0594 
0595     if (rcu_task_stall_timeout != RCU_TASK_STALL_TIMEOUT)
0596         pr_info("\tTasks-RCU CPU stall warnings timeout set to %d (rcu_task_stall_timeout).\n", rcu_task_stall_timeout);
0597     rtsimc = clamp(rcu_task_stall_info_mult, 1, 10);
0598     if (rtsimc != rcu_task_stall_info_mult) {
0599         pr_info("\tTasks-RCU CPU stall info multiplier clamped to %d (rcu_task_stall_info_mult).\n", rtsimc);
0600         rcu_task_stall_info_mult = rtsimc;
0601     }
0602 #endif /* #ifdef CONFIG_TASKS_RCU */
0603 #ifdef CONFIG_TASKS_RCU
0604     pr_info("\tTrampoline variant of Tasks RCU enabled.\n");
0605 #endif /* #ifdef CONFIG_TASKS_RCU */
0606 #ifdef CONFIG_TASKS_RUDE_RCU
0607     pr_info("\tRude variant of Tasks RCU enabled.\n");
0608 #endif /* #ifdef CONFIG_TASKS_RUDE_RCU */
0609 #ifdef CONFIG_TASKS_TRACE_RCU
0610     pr_info("\tTracing variant of Tasks RCU enabled.\n");
0611 #endif /* #ifdef CONFIG_TASKS_TRACE_RCU */
0612 }
0613 
0614 #endif /* #ifndef CONFIG_TINY_RCU */
0615 
0616 #ifndef CONFIG_TINY_RCU
0617 /* Dump out rcutorture-relevant state common to all RCU-tasks flavors. */
0618 static void show_rcu_tasks_generic_gp_kthread(struct rcu_tasks *rtp, char *s)
0619 {
0620     int cpu;
0621     bool havecbs = false;
0622 
0623     for_each_possible_cpu(cpu) {
0624         struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
0625 
0626         if (!data_race(rcu_segcblist_empty(&rtpcp->cblist))) {
0627             havecbs = true;
0628             break;
0629         }
0630     }
0631     pr_info("%s: %s(%d) since %lu g:%lu i:%lu/%lu %c%c %s\n",
0632         rtp->kname,
0633         tasks_gp_state_getname(rtp), data_race(rtp->gp_state),
0634         jiffies - data_race(rtp->gp_jiffies),
0635         data_race(rcu_seq_current(&rtp->tasks_gp_seq)),
0636         data_race(rtp->n_ipis_fails), data_race(rtp->n_ipis),
0637         ".k"[!!data_race(rtp->kthread_ptr)],
0638         ".C"[havecbs],
0639         s);
0640 }
0641 #endif // #ifndef CONFIG_TINY_RCU
0642 
0643 static void exit_tasks_rcu_finish_trace(struct task_struct *t);
0644 
0645 #if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU)
0646 
0647 ////////////////////////////////////////////////////////////////////////
0648 //
0649 // Shared code between task-list-scanning variants of Tasks RCU.
0650 
0651 /* Wait for one RCU-tasks grace period. */
0652 static void rcu_tasks_wait_gp(struct rcu_tasks *rtp)
0653 {
0654     struct task_struct *g;
0655     int fract;
0656     LIST_HEAD(holdouts);
0657     unsigned long j;
0658     unsigned long lastinfo;
0659     unsigned long lastreport;
0660     bool reported = false;
0661     int rtsi;
0662     struct task_struct *t;
0663 
0664     set_tasks_gp_state(rtp, RTGS_PRE_WAIT_GP);
0665     rtp->pregp_func(&holdouts);
0666 
0667     /*
0668      * There were callbacks, so we need to wait for an RCU-tasks
0669      * grace period.  Start off by scanning the task list for tasks
0670      * that are not already voluntarily blocked.  Mark these tasks
0671      * and make a list of them in holdouts.
0672      */
0673     set_tasks_gp_state(rtp, RTGS_SCAN_TASKLIST);
0674     if (rtp->pertask_func) {
0675         rcu_read_lock();
0676         for_each_process_thread(g, t)
0677             rtp->pertask_func(t, &holdouts);
0678         rcu_read_unlock();
0679     }
0680 
0681     set_tasks_gp_state(rtp, RTGS_POST_SCAN_TASKLIST);
0682     rtp->postscan_func(&holdouts);
0683 
0684     /*
0685      * Each pass through the following loop scans the list of holdout
0686      * tasks, removing any that are no longer holdouts.  When the list
0687      * is empty, we are done.
0688      */
0689     lastreport = jiffies;
0690     lastinfo = lastreport;
0691     rtsi = READ_ONCE(rcu_task_stall_info);
0692 
0693     // Start off with initial wait and slowly back off to 1 HZ wait.
0694     fract = rtp->init_fract;
0695 
0696     while (!list_empty(&holdouts)) {
0697         ktime_t exp;
0698         bool firstreport;
0699         bool needreport;
0700         int rtst;
0701 
0702         // Slowly back off waiting for holdouts
0703         set_tasks_gp_state(rtp, RTGS_WAIT_SCAN_HOLDOUTS);
0704         if (!IS_ENABLED(CONFIG_PREEMPT_RT)) {
0705             schedule_timeout_idle(fract);
0706         } else {
0707             exp = jiffies_to_nsecs(fract);
0708             __set_current_state(TASK_IDLE);
0709             schedule_hrtimeout_range(&exp, jiffies_to_nsecs(HZ / 2), HRTIMER_MODE_REL_HARD);
0710         }
0711 
0712         if (fract < HZ)
0713             fract++;
0714 
0715         rtst = READ_ONCE(rcu_task_stall_timeout);
0716         needreport = rtst > 0 && time_after(jiffies, lastreport + rtst);
0717         if (needreport) {
0718             lastreport = jiffies;
0719             reported = true;
0720         }
0721         firstreport = true;
0722         WARN_ON(signal_pending(current));
0723         set_tasks_gp_state(rtp, RTGS_SCAN_HOLDOUTS);
0724         rtp->holdouts_func(&holdouts, needreport, &firstreport);
0725 
0726         // Print pre-stall informational messages if needed.
0727         j = jiffies;
0728         if (rtsi > 0 && !reported && time_after(j, lastinfo + rtsi)) {
0729             lastinfo = j;
0730             rtsi = rtsi * rcu_task_stall_info_mult;
0731             pr_info("%s: %s grace period %lu is %lu jiffies old.\n",
0732                 __func__, rtp->kname, rtp->tasks_gp_seq, j - rtp->gp_start);
0733         }
0734     }
0735 
0736     set_tasks_gp_state(rtp, RTGS_POST_GP);
0737     rtp->postgp_func(rtp);
0738 }
0739 
0740 #endif /* #if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU) */
0741 
0742 #ifdef CONFIG_TASKS_RCU
0743 
0744 ////////////////////////////////////////////////////////////////////////
0745 //
0746 // Simple variant of RCU whose quiescent states are voluntary context
0747 // switch, cond_resched_tasks_rcu_qs(), user-space execution, and idle.
0748 // As such, grace periods can take one good long time.  There are no
0749 // read-side primitives similar to rcu_read_lock() and rcu_read_unlock()
0750 // because this implementation is intended to get the system into a safe
0751 // state for some of the manipulations involved in tracing and the like.
0752 // Finally, this implementation does not support high call_rcu_tasks()
0753 // rates from multiple CPUs.  If this is required, per-CPU callback lists
0754 // will be needed.
0755 //
0756 // The implementation uses rcu_tasks_wait_gp(), which relies on function
0757 // pointers in the rcu_tasks structure.  The rcu_spawn_tasks_kthread()
0758 // function sets these function pointers up so that rcu_tasks_wait_gp()
0759 // invokes these functions in this order:
0760 //
0761 // rcu_tasks_pregp_step():
0762 //  Invokes synchronize_rcu() in order to wait for all in-flight
0763 //  t->on_rq and t->nvcsw transitions to complete.  This works because
0764 //  all such transitions are carried out with interrupts disabled.
0765 // rcu_tasks_pertask(), invoked on every non-idle task:
0766 //  For every runnable non-idle task other than the current one, use
0767 //  get_task_struct() to pin down that task, snapshot that task's
0768 //  number of voluntary context switches, and add that task to the
0769 //  holdout list.
0770 // rcu_tasks_postscan():
0771 //  Invoke synchronize_srcu() to ensure that all tasks that were
0772 //  in the process of exiting (and which thus might not know to
0773 //  synchronize with this RCU Tasks grace period) have completed
0774 //  exiting.
0775 // check_all_holdout_tasks(), repeatedly until holdout list is empty:
0776 //  Scans the holdout list, attempting to identify a quiescent state
0777 //  for each task on the list.  If there is a quiescent state, the
0778 //  corresponding task is removed from the holdout list.
0779 // rcu_tasks_postgp():
0780 //  Invokes synchronize_rcu() in order to ensure that all prior
0781 //  t->on_rq and t->nvcsw transitions are seen by all CPUs and tasks
0782 //  to have happened before the end of this RCU Tasks grace period.
0783 //  Again, this works because all such transitions are carried out
0784 //  with interrupts disabled.
0785 //
0786 // For each exiting task, the exit_tasks_rcu_start() and
0787 // exit_tasks_rcu_finish() functions begin and end, respectively, the SRCU
0788 // read-side critical sections waited for by rcu_tasks_postscan().
0789 //
0790 // Pre-grace-period update-side code is ordered before the grace
0791 // via the raw_spin_lock.*rcu_node().  Pre-grace-period read-side code
0792 // is ordered before the grace period via synchronize_rcu() call in
0793 // rcu_tasks_pregp_step() and by the scheduler's locks and interrupt
0794 // disabling.
0795 
0796 /* Pre-grace-period preparation. */
0797 static void rcu_tasks_pregp_step(struct list_head *hop)
0798 {
0799     /*
0800      * Wait for all pre-existing t->on_rq and t->nvcsw transitions
0801      * to complete.  Invoking synchronize_rcu() suffices because all
0802      * these transitions occur with interrupts disabled.  Without this
0803      * synchronize_rcu(), a read-side critical section that started
0804      * before the grace period might be incorrectly seen as having
0805      * started after the grace period.
0806      *
0807      * This synchronize_rcu() also dispenses with the need for a
0808      * memory barrier on the first store to t->rcu_tasks_holdout,
0809      * as it forces the store to happen after the beginning of the
0810      * grace period.
0811      */
0812     synchronize_rcu();
0813 }
0814 
0815 /* Per-task initial processing. */
0816 static void rcu_tasks_pertask(struct task_struct *t, struct list_head *hop)
0817 {
0818     if (t != current && READ_ONCE(t->on_rq) && !is_idle_task(t)) {
0819         get_task_struct(t);
0820         t->rcu_tasks_nvcsw = READ_ONCE(t->nvcsw);
0821         WRITE_ONCE(t->rcu_tasks_holdout, true);
0822         list_add(&t->rcu_tasks_holdout_list, hop);
0823     }
0824 }
0825 
0826 /* Processing between scanning taskslist and draining the holdout list. */
0827 static void rcu_tasks_postscan(struct list_head *hop)
0828 {
0829     /*
0830      * Wait for tasks that are in the process of exiting.  This
0831      * does only part of the job, ensuring that all tasks that were
0832      * previously exiting reach the point where they have disabled
0833      * preemption, allowing the later synchronize_rcu() to finish
0834      * the job.
0835      */
0836     synchronize_srcu(&tasks_rcu_exit_srcu);
0837 }
0838 
0839 /* See if tasks are still holding out, complain if so. */
0840 static void check_holdout_task(struct task_struct *t,
0841                    bool needreport, bool *firstreport)
0842 {
0843     int cpu;
0844 
0845     if (!READ_ONCE(t->rcu_tasks_holdout) ||
0846         t->rcu_tasks_nvcsw != READ_ONCE(t->nvcsw) ||
0847         !READ_ONCE(t->on_rq) ||
0848         (IS_ENABLED(CONFIG_NO_HZ_FULL) &&
0849          !is_idle_task(t) && t->rcu_tasks_idle_cpu >= 0)) {
0850         WRITE_ONCE(t->rcu_tasks_holdout, false);
0851         list_del_init(&t->rcu_tasks_holdout_list);
0852         put_task_struct(t);
0853         return;
0854     }
0855     rcu_request_urgent_qs_task(t);
0856     if (!needreport)
0857         return;
0858     if (*firstreport) {
0859         pr_err("INFO: rcu_tasks detected stalls on tasks:\n");
0860         *firstreport = false;
0861     }
0862     cpu = task_cpu(t);
0863     pr_alert("%p: %c%c nvcsw: %lu/%lu holdout: %d idle_cpu: %d/%d\n",
0864          t, ".I"[is_idle_task(t)],
0865          "N."[cpu < 0 || !tick_nohz_full_cpu(cpu)],
0866          t->rcu_tasks_nvcsw, t->nvcsw, t->rcu_tasks_holdout,
0867          t->rcu_tasks_idle_cpu, cpu);
0868     sched_show_task(t);
0869 }
0870 
0871 /* Scan the holdout lists for tasks no longer holding out. */
0872 static void check_all_holdout_tasks(struct list_head *hop,
0873                     bool needreport, bool *firstreport)
0874 {
0875     struct task_struct *t, *t1;
0876 
0877     list_for_each_entry_safe(t, t1, hop, rcu_tasks_holdout_list) {
0878         check_holdout_task(t, needreport, firstreport);
0879         cond_resched();
0880     }
0881 }
0882 
0883 /* Finish off the Tasks-RCU grace period. */
0884 static void rcu_tasks_postgp(struct rcu_tasks *rtp)
0885 {
0886     /*
0887      * Because ->on_rq and ->nvcsw are not guaranteed to have a full
0888      * memory barriers prior to them in the schedule() path, memory
0889      * reordering on other CPUs could cause their RCU-tasks read-side
0890      * critical sections to extend past the end of the grace period.
0891      * However, because these ->nvcsw updates are carried out with
0892      * interrupts disabled, we can use synchronize_rcu() to force the
0893      * needed ordering on all such CPUs.
0894      *
0895      * This synchronize_rcu() also confines all ->rcu_tasks_holdout
0896      * accesses to be within the grace period, avoiding the need for
0897      * memory barriers for ->rcu_tasks_holdout accesses.
0898      *
0899      * In addition, this synchronize_rcu() waits for exiting tasks
0900      * to complete their final preempt_disable() region of execution,
0901      * cleaning up after the synchronize_srcu() above.
0902      */
0903     synchronize_rcu();
0904 }
0905 
0906 void call_rcu_tasks(struct rcu_head *rhp, rcu_callback_t func);
0907 DEFINE_RCU_TASKS(rcu_tasks, rcu_tasks_wait_gp, call_rcu_tasks, "RCU Tasks");
0908 
0909 /**
0910  * call_rcu_tasks() - Queue an RCU for invocation task-based grace period
0911  * @rhp: structure to be used for queueing the RCU updates.
0912  * @func: actual callback function to be invoked after the grace period
0913  *
0914  * The callback function will be invoked some time after a full grace
0915  * period elapses, in other words after all currently executing RCU
0916  * read-side critical sections have completed. call_rcu_tasks() assumes
0917  * that the read-side critical sections end at a voluntary context
0918  * switch (not a preemption!), cond_resched_tasks_rcu_qs(), entry into idle,
0919  * or transition to usermode execution.  As such, there are no read-side
0920  * primitives analogous to rcu_read_lock() and rcu_read_unlock() because
0921  * this primitive is intended to determine that all tasks have passed
0922  * through a safe state, not so much for data-structure synchronization.
0923  *
0924  * See the description of call_rcu() for more detailed information on
0925  * memory ordering guarantees.
0926  */
0927 void call_rcu_tasks(struct rcu_head *rhp, rcu_callback_t func)
0928 {
0929     call_rcu_tasks_generic(rhp, func, &rcu_tasks);
0930 }
0931 EXPORT_SYMBOL_GPL(call_rcu_tasks);
0932 
0933 /**
0934  * synchronize_rcu_tasks - wait until an rcu-tasks grace period has elapsed.
0935  *
0936  * Control will return to the caller some time after a full rcu-tasks
0937  * grace period has elapsed, in other words after all currently
0938  * executing rcu-tasks read-side critical sections have elapsed.  These
0939  * read-side critical sections are delimited by calls to schedule(),
0940  * cond_resched_tasks_rcu_qs(), idle execution, userspace execution, calls
0941  * to synchronize_rcu_tasks(), and (in theory, anyway) cond_resched().
0942  *
0943  * This is a very specialized primitive, intended only for a few uses in
0944  * tracing and other situations requiring manipulation of function
0945  * preambles and profiling hooks.  The synchronize_rcu_tasks() function
0946  * is not (yet) intended for heavy use from multiple CPUs.
0947  *
0948  * See the description of synchronize_rcu() for more detailed information
0949  * on memory ordering guarantees.
0950  */
0951 void synchronize_rcu_tasks(void)
0952 {
0953     synchronize_rcu_tasks_generic(&rcu_tasks);
0954 }
0955 EXPORT_SYMBOL_GPL(synchronize_rcu_tasks);
0956 
0957 /**
0958  * rcu_barrier_tasks - Wait for in-flight call_rcu_tasks() callbacks.
0959  *
0960  * Although the current implementation is guaranteed to wait, it is not
0961  * obligated to, for example, if there are no pending callbacks.
0962  */
0963 void rcu_barrier_tasks(void)
0964 {
0965     rcu_barrier_tasks_generic(&rcu_tasks);
0966 }
0967 EXPORT_SYMBOL_GPL(rcu_barrier_tasks);
0968 
0969 static int __init rcu_spawn_tasks_kthread(void)
0970 {
0971     cblist_init_generic(&rcu_tasks);
0972     rcu_tasks.gp_sleep = HZ / 10;
0973     rcu_tasks.init_fract = HZ / 10;
0974     rcu_tasks.pregp_func = rcu_tasks_pregp_step;
0975     rcu_tasks.pertask_func = rcu_tasks_pertask;
0976     rcu_tasks.postscan_func = rcu_tasks_postscan;
0977     rcu_tasks.holdouts_func = check_all_holdout_tasks;
0978     rcu_tasks.postgp_func = rcu_tasks_postgp;
0979     rcu_spawn_tasks_kthread_generic(&rcu_tasks);
0980     return 0;
0981 }
0982 
0983 #if !defined(CONFIG_TINY_RCU)
0984 void show_rcu_tasks_classic_gp_kthread(void)
0985 {
0986     show_rcu_tasks_generic_gp_kthread(&rcu_tasks, "");
0987 }
0988 EXPORT_SYMBOL_GPL(show_rcu_tasks_classic_gp_kthread);
0989 #endif // !defined(CONFIG_TINY_RCU)
0990 
0991 /* Do the srcu_read_lock() for the above synchronize_srcu().  */
0992 void exit_tasks_rcu_start(void) __acquires(&tasks_rcu_exit_srcu)
0993 {
0994     preempt_disable();
0995     current->rcu_tasks_idx = __srcu_read_lock(&tasks_rcu_exit_srcu);
0996     preempt_enable();
0997 }
0998 
0999 /* Do the srcu_read_unlock() for the above synchronize_srcu().  */
1000 void exit_tasks_rcu_finish(void) __releases(&tasks_rcu_exit_srcu)
1001 {
1002     struct task_struct *t = current;
1003 
1004     preempt_disable();
1005     __srcu_read_unlock(&tasks_rcu_exit_srcu, t->rcu_tasks_idx);
1006     preempt_enable();
1007     exit_tasks_rcu_finish_trace(t);
1008 }
1009 
1010 #else /* #ifdef CONFIG_TASKS_RCU */
1011 void exit_tasks_rcu_start(void) { }
1012 void exit_tasks_rcu_finish(void) { exit_tasks_rcu_finish_trace(current); }
1013 #endif /* #else #ifdef CONFIG_TASKS_RCU */
1014 
1015 #ifdef CONFIG_TASKS_RUDE_RCU
1016 
1017 ////////////////////////////////////////////////////////////////////////
1018 //
1019 // "Rude" variant of Tasks RCU, inspired by Steve Rostedt's trick of
1020 // passing an empty function to schedule_on_each_cpu().  This approach
1021 // provides an asynchronous call_rcu_tasks_rude() API and batching of
1022 // concurrent calls to the synchronous synchronize_rcu_tasks_rude() API.
1023 // This invokes schedule_on_each_cpu() in order to send IPIs far and wide
1024 // and induces otherwise unnecessary context switches on all online CPUs,
1025 // whether idle or not.
1026 //
1027 // Callback handling is provided by the rcu_tasks_kthread() function.
1028 //
1029 // Ordering is provided by the scheduler's context-switch code.
1030 
1031 // Empty function to allow workqueues to force a context switch.
1032 static void rcu_tasks_be_rude(struct work_struct *work)
1033 {
1034 }
1035 
1036 // Wait for one rude RCU-tasks grace period.
1037 static void rcu_tasks_rude_wait_gp(struct rcu_tasks *rtp)
1038 {
1039     if (num_online_cpus() <= 1)
1040         return; // Fastpath for only one CPU.
1041 
1042     rtp->n_ipis += cpumask_weight(cpu_online_mask);
1043     schedule_on_each_cpu(rcu_tasks_be_rude);
1044 }
1045 
1046 void call_rcu_tasks_rude(struct rcu_head *rhp, rcu_callback_t func);
1047 DEFINE_RCU_TASKS(rcu_tasks_rude, rcu_tasks_rude_wait_gp, call_rcu_tasks_rude,
1048          "RCU Tasks Rude");
1049 
1050 /**
1051  * call_rcu_tasks_rude() - Queue a callback rude task-based grace period
1052  * @rhp: structure to be used for queueing the RCU updates.
1053  * @func: actual callback function to be invoked after the grace period
1054  *
1055  * The callback function will be invoked some time after a full grace
1056  * period elapses, in other words after all currently executing RCU
1057  * read-side critical sections have completed. call_rcu_tasks_rude()
1058  * assumes that the read-side critical sections end at context switch,
1059  * cond_resched_tasks_rcu_qs(), or transition to usermode execution (as
1060  * usermode execution is schedulable). As such, there are no read-side
1061  * primitives analogous to rcu_read_lock() and rcu_read_unlock() because
1062  * this primitive is intended to determine that all tasks have passed
1063  * through a safe state, not so much for data-structure synchronization.
1064  *
1065  * See the description of call_rcu() for more detailed information on
1066  * memory ordering guarantees.
1067  */
1068 void call_rcu_tasks_rude(struct rcu_head *rhp, rcu_callback_t func)
1069 {
1070     call_rcu_tasks_generic(rhp, func, &rcu_tasks_rude);
1071 }
1072 EXPORT_SYMBOL_GPL(call_rcu_tasks_rude);
1073 
1074 /**
1075  * synchronize_rcu_tasks_rude - wait for a rude rcu-tasks grace period
1076  *
1077  * Control will return to the caller some time after a rude rcu-tasks
1078  * grace period has elapsed, in other words after all currently
1079  * executing rcu-tasks read-side critical sections have elapsed.  These
1080  * read-side critical sections are delimited by calls to schedule(),
1081  * cond_resched_tasks_rcu_qs(), userspace execution (which is a schedulable
1082  * context), and (in theory, anyway) cond_resched().
1083  *
1084  * This is a very specialized primitive, intended only for a few uses in
1085  * tracing and other situations requiring manipulation of function preambles
1086  * and profiling hooks.  The synchronize_rcu_tasks_rude() function is not
1087  * (yet) intended for heavy use from multiple CPUs.
1088  *
1089  * See the description of synchronize_rcu() for more detailed information
1090  * on memory ordering guarantees.
1091  */
1092 void synchronize_rcu_tasks_rude(void)
1093 {
1094     synchronize_rcu_tasks_generic(&rcu_tasks_rude);
1095 }
1096 EXPORT_SYMBOL_GPL(synchronize_rcu_tasks_rude);
1097 
1098 /**
1099  * rcu_barrier_tasks_rude - Wait for in-flight call_rcu_tasks_rude() callbacks.
1100  *
1101  * Although the current implementation is guaranteed to wait, it is not
1102  * obligated to, for example, if there are no pending callbacks.
1103  */
1104 void rcu_barrier_tasks_rude(void)
1105 {
1106     rcu_barrier_tasks_generic(&rcu_tasks_rude);
1107 }
1108 EXPORT_SYMBOL_GPL(rcu_barrier_tasks_rude);
1109 
1110 static int __init rcu_spawn_tasks_rude_kthread(void)
1111 {
1112     cblist_init_generic(&rcu_tasks_rude);
1113     rcu_tasks_rude.gp_sleep = HZ / 10;
1114     rcu_spawn_tasks_kthread_generic(&rcu_tasks_rude);
1115     return 0;
1116 }
1117 
1118 #if !defined(CONFIG_TINY_RCU)
1119 void show_rcu_tasks_rude_gp_kthread(void)
1120 {
1121     show_rcu_tasks_generic_gp_kthread(&rcu_tasks_rude, "");
1122 }
1123 EXPORT_SYMBOL_GPL(show_rcu_tasks_rude_gp_kthread);
1124 #endif // !defined(CONFIG_TINY_RCU)
1125 #endif /* #ifdef CONFIG_TASKS_RUDE_RCU */
1126 
1127 ////////////////////////////////////////////////////////////////////////
1128 //
1129 // Tracing variant of Tasks RCU.  This variant is designed to be used
1130 // to protect tracing hooks, including those of BPF.  This variant
1131 // therefore:
1132 //
1133 // 1.   Has explicit read-side markers to allow finite grace periods
1134 //  in the face of in-kernel loops for PREEMPT=n builds.
1135 //
1136 // 2.   Protects code in the idle loop, exception entry/exit, and
1137 //  CPU-hotplug code paths, similar to the capabilities of SRCU.
1138 //
1139 // 3.   Avoids expensive read-side instructions, having overhead similar
1140 //  to that of Preemptible RCU.
1141 //
1142 // There are of course downsides.  For example, the grace-period code
1143 // can send IPIs to CPUs, even when those CPUs are in the idle loop or
1144 // in nohz_full userspace.  If needed, these downsides can be at least
1145 // partially remedied.
1146 //
1147 // Perhaps most important, this variant of RCU does not affect the vanilla
1148 // flavors, rcu_preempt and rcu_sched.  The fact that RCU Tasks Trace
1149 // readers can operate from idle, offline, and exception entry/exit in no
1150 // way allows rcu_preempt and rcu_sched readers to also do so.
1151 //
1152 // The implementation uses rcu_tasks_wait_gp(), which relies on function
1153 // pointers in the rcu_tasks structure.  The rcu_spawn_tasks_trace_kthread()
1154 // function sets these function pointers up so that rcu_tasks_wait_gp()
1155 // invokes these functions in this order:
1156 //
1157 // rcu_tasks_trace_pregp_step():
1158 //  Disables CPU hotplug, adds all currently executing tasks to the
1159 //  holdout list, then checks the state of all tasks that blocked
1160 //  or were preempted within their current RCU Tasks Trace read-side
1161 //  critical section, adding them to the holdout list if appropriate.
1162 //  Finally, this function re-enables CPU hotplug.
1163 // The ->pertask_func() pointer is NULL, so there is no per-task processing.
1164 // rcu_tasks_trace_postscan():
1165 //  Invokes synchronize_rcu() to wait for late-stage exiting tasks
1166 //  to finish exiting.
1167 // check_all_holdout_tasks_trace(), repeatedly until holdout list is empty:
1168 //  Scans the holdout list, attempting to identify a quiescent state
1169 //  for each task on the list.  If there is a quiescent state, the
1170 //  corresponding task is removed from the holdout list.  Once this
1171 //  list is empty, the grace period has completed.
1172 // rcu_tasks_trace_postgp():
1173 //  Provides the needed full memory barrier and does debug checks.
1174 //
1175 // The exit_tasks_rcu_finish_trace() synchronizes with exiting tasks.
1176 //
1177 // Pre-grace-period update-side code is ordered before the grace period
1178 // via the ->cbs_lock and barriers in rcu_tasks_kthread().  Pre-grace-period
1179 // read-side code is ordered before the grace period by atomic operations
1180 // on .b.need_qs flag of each task involved in this process, or by scheduler
1181 // context-switch ordering (for locked-down non-running readers).
1182 
1183 // The lockdep state must be outside of #ifdef to be useful.
1184 #ifdef CONFIG_DEBUG_LOCK_ALLOC
1185 static struct lock_class_key rcu_lock_trace_key;
1186 struct lockdep_map rcu_trace_lock_map =
1187     STATIC_LOCKDEP_MAP_INIT("rcu_read_lock_trace", &rcu_lock_trace_key);
1188 EXPORT_SYMBOL_GPL(rcu_trace_lock_map);
1189 #endif /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */
1190 
1191 #ifdef CONFIG_TASKS_TRACE_RCU
1192 
1193 // Record outstanding IPIs to each CPU.  No point in sending two...
1194 static DEFINE_PER_CPU(bool, trc_ipi_to_cpu);
1195 
1196 // The number of detections of task quiescent state relying on
1197 // heavyweight readers executing explicit memory barriers.
1198 static unsigned long n_heavy_reader_attempts;
1199 static unsigned long n_heavy_reader_updates;
1200 static unsigned long n_heavy_reader_ofl_updates;
1201 static unsigned long n_trc_holdouts;
1202 
1203 void call_rcu_tasks_trace(struct rcu_head *rhp, rcu_callback_t func);
1204 DEFINE_RCU_TASKS(rcu_tasks_trace, rcu_tasks_wait_gp, call_rcu_tasks_trace,
1205          "RCU Tasks Trace");
1206 
1207 /* Load from ->trc_reader_special.b.need_qs with proper ordering. */
1208 static u8 rcu_ld_need_qs(struct task_struct *t)
1209 {
1210     smp_mb(); // Enforce full grace-period ordering.
1211     return smp_load_acquire(&t->trc_reader_special.b.need_qs);
1212 }
1213 
1214 /* Store to ->trc_reader_special.b.need_qs with proper ordering. */
1215 static void rcu_st_need_qs(struct task_struct *t, u8 v)
1216 {
1217     smp_store_release(&t->trc_reader_special.b.need_qs, v);
1218     smp_mb(); // Enforce full grace-period ordering.
1219 }
1220 
1221 /*
1222  * Do a cmpxchg() on ->trc_reader_special.b.need_qs, allowing for
1223  * the four-byte operand-size restriction of some platforms.
1224  * Returns the old value, which is often ignored.
1225  */
1226 u8 rcu_trc_cmpxchg_need_qs(struct task_struct *t, u8 old, u8 new)
1227 {
1228     union rcu_special ret;
1229     union rcu_special trs_old = READ_ONCE(t->trc_reader_special);
1230     union rcu_special trs_new = trs_old;
1231 
1232     if (trs_old.b.need_qs != old)
1233         return trs_old.b.need_qs;
1234     trs_new.b.need_qs = new;
1235     ret.s = cmpxchg(&t->trc_reader_special.s, trs_old.s, trs_new.s);
1236     return ret.b.need_qs;
1237 }
1238 EXPORT_SYMBOL_GPL(rcu_trc_cmpxchg_need_qs);
1239 
1240 /*
1241  * If we are the last reader, signal the grace-period kthread.
1242  * Also remove from the per-CPU list of blocked tasks.
1243  */
1244 void rcu_read_unlock_trace_special(struct task_struct *t)
1245 {
1246     unsigned long flags;
1247     struct rcu_tasks_percpu *rtpcp;
1248     union rcu_special trs;
1249 
1250     // Open-coded full-word version of rcu_ld_need_qs().
1251     smp_mb(); // Enforce full grace-period ordering.
1252     trs = smp_load_acquire(&t->trc_reader_special);
1253 
1254     if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB) && t->trc_reader_special.b.need_mb)
1255         smp_mb(); // Pairs with update-side barriers.
1256     // Update .need_qs before ->trc_reader_nesting for irq/NMI handlers.
1257     if (trs.b.need_qs == (TRC_NEED_QS_CHECKED | TRC_NEED_QS)) {
1258         u8 result = rcu_trc_cmpxchg_need_qs(t, TRC_NEED_QS_CHECKED | TRC_NEED_QS,
1259                                TRC_NEED_QS_CHECKED);
1260 
1261         WARN_ONCE(result != trs.b.need_qs, "%s: result = %d", __func__, result);
1262     }
1263     if (trs.b.blocked) {
1264         rtpcp = per_cpu_ptr(rcu_tasks_trace.rtpcpu, t->trc_blkd_cpu);
1265         raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
1266         list_del_init(&t->trc_blkd_node);
1267         WRITE_ONCE(t->trc_reader_special.b.blocked, false);
1268         raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
1269     }
1270     WRITE_ONCE(t->trc_reader_nesting, 0);
1271 }
1272 EXPORT_SYMBOL_GPL(rcu_read_unlock_trace_special);
1273 
1274 /* Add a newly blocked reader task to its CPU's list. */
1275 void rcu_tasks_trace_qs_blkd(struct task_struct *t)
1276 {
1277     unsigned long flags;
1278     struct rcu_tasks_percpu *rtpcp;
1279 
1280     local_irq_save(flags);
1281     rtpcp = this_cpu_ptr(rcu_tasks_trace.rtpcpu);
1282     raw_spin_lock_rcu_node(rtpcp); // irqs already disabled
1283     t->trc_blkd_cpu = smp_processor_id();
1284     if (!rtpcp->rtp_blkd_tasks.next)
1285         INIT_LIST_HEAD(&rtpcp->rtp_blkd_tasks);
1286     list_add(&t->trc_blkd_node, &rtpcp->rtp_blkd_tasks);
1287     WRITE_ONCE(t->trc_reader_special.b.blocked, true);
1288     raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
1289 }
1290 EXPORT_SYMBOL_GPL(rcu_tasks_trace_qs_blkd);
1291 
1292 /* Add a task to the holdout list, if it is not already on the list. */
1293 static void trc_add_holdout(struct task_struct *t, struct list_head *bhp)
1294 {
1295     if (list_empty(&t->trc_holdout_list)) {
1296         get_task_struct(t);
1297         list_add(&t->trc_holdout_list, bhp);
1298         n_trc_holdouts++;
1299     }
1300 }
1301 
1302 /* Remove a task from the holdout list, if it is in fact present. */
1303 static void trc_del_holdout(struct task_struct *t)
1304 {
1305     if (!list_empty(&t->trc_holdout_list)) {
1306         list_del_init(&t->trc_holdout_list);
1307         put_task_struct(t);
1308         n_trc_holdouts--;
1309     }
1310 }
1311 
1312 /* IPI handler to check task state. */
1313 static void trc_read_check_handler(void *t_in)
1314 {
1315     int nesting;
1316     struct task_struct *t = current;
1317     struct task_struct *texp = t_in;
1318 
1319     // If the task is no longer running on this CPU, leave.
1320     if (unlikely(texp != t))
1321         goto reset_ipi; // Already on holdout list, so will check later.
1322 
1323     // If the task is not in a read-side critical section, and
1324     // if this is the last reader, awaken the grace-period kthread.
1325     nesting = READ_ONCE(t->trc_reader_nesting);
1326     if (likely(!nesting)) {
1327         rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS_CHECKED);
1328         goto reset_ipi;
1329     }
1330     // If we are racing with an rcu_read_unlock_trace(), try again later.
1331     if (unlikely(nesting < 0))
1332         goto reset_ipi;
1333 
1334     // Get here if the task is in a read-side critical section.
1335     // Set its state so that it will update state for the grace-period
1336     // kthread upon exit from that critical section.
1337     rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS | TRC_NEED_QS_CHECKED);
1338 
1339 reset_ipi:
1340     // Allow future IPIs to be sent on CPU and for task.
1341     // Also order this IPI handler against any later manipulations of
1342     // the intended task.
1343     smp_store_release(per_cpu_ptr(&trc_ipi_to_cpu, smp_processor_id()), false); // ^^^
1344     smp_store_release(&texp->trc_ipi_to_cpu, -1); // ^^^
1345 }
1346 
1347 /* Callback function for scheduler to check locked-down task.  */
1348 static int trc_inspect_reader(struct task_struct *t, void *bhp_in)
1349 {
1350     struct list_head *bhp = bhp_in;
1351     int cpu = task_cpu(t);
1352     int nesting;
1353     bool ofl = cpu_is_offline(cpu);
1354 
1355     if (task_curr(t) && !ofl) {
1356         // If no chance of heavyweight readers, do it the hard way.
1357         if (!IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB))
1358             return -EINVAL;
1359 
1360         // If heavyweight readers are enabled on the remote task,
1361         // we can inspect its state despite its currently running.
1362         // However, we cannot safely change its state.
1363         n_heavy_reader_attempts++;
1364         // Check for "running" idle tasks on offline CPUs.
1365         if (!rcu_dynticks_zero_in_eqs(cpu, &t->trc_reader_nesting))
1366             return -EINVAL; // No quiescent state, do it the hard way.
1367         n_heavy_reader_updates++;
1368         nesting = 0;
1369     } else {
1370         // The task is not running, so C-language access is safe.
1371         nesting = t->trc_reader_nesting;
1372         WARN_ON_ONCE(ofl && task_curr(t) && !is_idle_task(t));
1373         if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB) && ofl)
1374             n_heavy_reader_ofl_updates++;
1375     }
1376 
1377     // If not exiting a read-side critical section, mark as checked
1378     // so that the grace-period kthread will remove it from the
1379     // holdout list.
1380     if (!nesting) {
1381         rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS_CHECKED);
1382         return 0;  // In QS, so done.
1383     }
1384     if (nesting < 0)
1385         return -EINVAL; // Reader transitioning, try again later.
1386 
1387     // The task is in a read-side critical section, so set up its
1388     // state so that it will update state upon exit from that critical
1389     // section.
1390     if (!rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS | TRC_NEED_QS_CHECKED))
1391         trc_add_holdout(t, bhp);
1392     return 0;
1393 }
1394 
1395 /* Attempt to extract the state for the specified task. */
1396 static void trc_wait_for_one_reader(struct task_struct *t,
1397                     struct list_head *bhp)
1398 {
1399     int cpu;
1400 
1401     // If a previous IPI is still in flight, let it complete.
1402     if (smp_load_acquire(&t->trc_ipi_to_cpu) != -1) // Order IPI
1403         return;
1404 
1405     // The current task had better be in a quiescent state.
1406     if (t == current) {
1407         rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS_CHECKED);
1408         WARN_ON_ONCE(READ_ONCE(t->trc_reader_nesting));
1409         return;
1410     }
1411 
1412     // Attempt to nail down the task for inspection.
1413     get_task_struct(t);
1414     if (!task_call_func(t, trc_inspect_reader, bhp)) {
1415         put_task_struct(t);
1416         return;
1417     }
1418     put_task_struct(t);
1419 
1420     // If this task is not yet on the holdout list, then we are in
1421     // an RCU read-side critical section.  Otherwise, the invocation of
1422     // trc_add_holdout() that added it to the list did the necessary
1423     // get_task_struct().  Either way, the task cannot be freed out
1424     // from under this code.
1425 
1426     // If currently running, send an IPI, either way, add to list.
1427     trc_add_holdout(t, bhp);
1428     if (task_curr(t) &&
1429         time_after(jiffies + 1, rcu_tasks_trace.gp_start + rcu_task_ipi_delay)) {
1430         // The task is currently running, so try IPIing it.
1431         cpu = task_cpu(t);
1432 
1433         // If there is already an IPI outstanding, let it happen.
1434         if (per_cpu(trc_ipi_to_cpu, cpu) || t->trc_ipi_to_cpu >= 0)
1435             return;
1436 
1437         per_cpu(trc_ipi_to_cpu, cpu) = true;
1438         t->trc_ipi_to_cpu = cpu;
1439         rcu_tasks_trace.n_ipis++;
1440         if (smp_call_function_single(cpu, trc_read_check_handler, t, 0)) {
1441             // Just in case there is some other reason for
1442             // failure than the target CPU being offline.
1443             WARN_ONCE(1, "%s():  smp_call_function_single() failed for CPU: %d\n",
1444                   __func__, cpu);
1445             rcu_tasks_trace.n_ipis_fails++;
1446             per_cpu(trc_ipi_to_cpu, cpu) = false;
1447             t->trc_ipi_to_cpu = -1;
1448         }
1449     }
1450 }
1451 
1452 /*
1453  * Initialize for first-round processing for the specified task.
1454  * Return false if task is NULL or already taken care of, true otherwise.
1455  */
1456 static bool rcu_tasks_trace_pertask_prep(struct task_struct *t, bool notself)
1457 {
1458     // During early boot when there is only the one boot CPU, there
1459     // is no idle task for the other CPUs.  Also, the grace-period
1460     // kthread is always in a quiescent state.  In addition, just return
1461     // if this task is already on the list.
1462     if (unlikely(t == NULL) || (t == current && notself) || !list_empty(&t->trc_holdout_list))
1463         return false;
1464 
1465     rcu_st_need_qs(t, 0);
1466     t->trc_ipi_to_cpu = -1;
1467     return true;
1468 }
1469 
1470 /* Do first-round processing for the specified task. */
1471 static void rcu_tasks_trace_pertask(struct task_struct *t, struct list_head *hop)
1472 {
1473     if (rcu_tasks_trace_pertask_prep(t, true))
1474         trc_wait_for_one_reader(t, hop);
1475 }
1476 
1477 /* Initialize for a new RCU-tasks-trace grace period. */
1478 static void rcu_tasks_trace_pregp_step(struct list_head *hop)
1479 {
1480     LIST_HEAD(blkd_tasks);
1481     int cpu;
1482     unsigned long flags;
1483     struct rcu_tasks_percpu *rtpcp;
1484     struct task_struct *t;
1485 
1486     // There shouldn't be any old IPIs, but...
1487     for_each_possible_cpu(cpu)
1488         WARN_ON_ONCE(per_cpu(trc_ipi_to_cpu, cpu));
1489 
1490     // Disable CPU hotplug across the CPU scan for the benefit of
1491     // any IPIs that might be needed.  This also waits for all readers
1492     // in CPU-hotplug code paths.
1493     cpus_read_lock();
1494 
1495     // These rcu_tasks_trace_pertask_prep() calls are serialized to
1496     // allow safe access to the hop list.
1497     for_each_online_cpu(cpu) {
1498         rcu_read_lock();
1499         t = cpu_curr_snapshot(cpu);
1500         if (rcu_tasks_trace_pertask_prep(t, true))
1501             trc_add_holdout(t, hop);
1502         rcu_read_unlock();
1503     }
1504 
1505     // Only after all running tasks have been accounted for is it
1506     // safe to take care of the tasks that have blocked within their
1507     // current RCU tasks trace read-side critical section.
1508     for_each_possible_cpu(cpu) {
1509         rtpcp = per_cpu_ptr(rcu_tasks_trace.rtpcpu, cpu);
1510         raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
1511         list_splice_init(&rtpcp->rtp_blkd_tasks, &blkd_tasks);
1512         while (!list_empty(&blkd_tasks)) {
1513             rcu_read_lock();
1514             t = list_first_entry(&blkd_tasks, struct task_struct, trc_blkd_node);
1515             list_del_init(&t->trc_blkd_node);
1516             list_add(&t->trc_blkd_node, &rtpcp->rtp_blkd_tasks);
1517             raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
1518             rcu_tasks_trace_pertask(t, hop);
1519             rcu_read_unlock();
1520             raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
1521         }
1522         raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
1523     }
1524 
1525     // Re-enable CPU hotplug now that the holdout list is populated.
1526     cpus_read_unlock();
1527 }
1528 
1529 /*
1530  * Do intermediate processing between task and holdout scans.
1531  */
1532 static void rcu_tasks_trace_postscan(struct list_head *hop)
1533 {
1534     // Wait for late-stage exiting tasks to finish exiting.
1535     // These might have passed the call to exit_tasks_rcu_finish().
1536     synchronize_rcu();
1537     // Any tasks that exit after this point will set
1538     // TRC_NEED_QS_CHECKED in ->trc_reader_special.b.need_qs.
1539 }
1540 
1541 /* Communicate task state back to the RCU tasks trace stall warning request. */
1542 struct trc_stall_chk_rdr {
1543     int nesting;
1544     int ipi_to_cpu;
1545     u8 needqs;
1546 };
1547 
1548 static int trc_check_slow_task(struct task_struct *t, void *arg)
1549 {
1550     struct trc_stall_chk_rdr *trc_rdrp = arg;
1551 
1552     if (task_curr(t) && cpu_online(task_cpu(t)))
1553         return false; // It is running, so decline to inspect it.
1554     trc_rdrp->nesting = READ_ONCE(t->trc_reader_nesting);
1555     trc_rdrp->ipi_to_cpu = READ_ONCE(t->trc_ipi_to_cpu);
1556     trc_rdrp->needqs = rcu_ld_need_qs(t);
1557     return true;
1558 }
1559 
1560 /* Show the state of a task stalling the current RCU tasks trace GP. */
1561 static void show_stalled_task_trace(struct task_struct *t, bool *firstreport)
1562 {
1563     int cpu;
1564     struct trc_stall_chk_rdr trc_rdr;
1565     bool is_idle_tsk = is_idle_task(t);
1566 
1567     if (*firstreport) {
1568         pr_err("INFO: rcu_tasks_trace detected stalls on tasks:\n");
1569         *firstreport = false;
1570     }
1571     cpu = task_cpu(t);
1572     if (!task_call_func(t, trc_check_slow_task, &trc_rdr))
1573         pr_alert("P%d: %c%c\n",
1574              t->pid,
1575              ".I"[t->trc_ipi_to_cpu >= 0],
1576              ".i"[is_idle_tsk]);
1577     else
1578         pr_alert("P%d: %c%c%c%c nesting: %d%c%c cpu: %d%s\n",
1579              t->pid,
1580              ".I"[trc_rdr.ipi_to_cpu >= 0],
1581              ".i"[is_idle_tsk],
1582              ".N"[cpu >= 0 && tick_nohz_full_cpu(cpu)],
1583              ".B"[!!data_race(t->trc_reader_special.b.blocked)],
1584              trc_rdr.nesting,
1585              " !CN"[trc_rdr.needqs & 0x3],
1586              " ?"[trc_rdr.needqs > 0x3],
1587              cpu, cpu_online(cpu) ? "" : "(offline)");
1588     sched_show_task(t);
1589 }
1590 
1591 /* List stalled IPIs for RCU tasks trace. */
1592 static void show_stalled_ipi_trace(void)
1593 {
1594     int cpu;
1595 
1596     for_each_possible_cpu(cpu)
1597         if (per_cpu(trc_ipi_to_cpu, cpu))
1598             pr_alert("\tIPI outstanding to CPU %d\n", cpu);
1599 }
1600 
1601 /* Do one scan of the holdout list. */
1602 static void check_all_holdout_tasks_trace(struct list_head *hop,
1603                       bool needreport, bool *firstreport)
1604 {
1605     struct task_struct *g, *t;
1606 
1607     // Disable CPU hotplug across the holdout list scan for IPIs.
1608     cpus_read_lock();
1609 
1610     list_for_each_entry_safe(t, g, hop, trc_holdout_list) {
1611         // If safe and needed, try to check the current task.
1612         if (READ_ONCE(t->trc_ipi_to_cpu) == -1 &&
1613             !(rcu_ld_need_qs(t) & TRC_NEED_QS_CHECKED))
1614             trc_wait_for_one_reader(t, hop);
1615 
1616         // If check succeeded, remove this task from the list.
1617         if (smp_load_acquire(&t->trc_ipi_to_cpu) == -1 &&
1618             rcu_ld_need_qs(t) == TRC_NEED_QS_CHECKED)
1619             trc_del_holdout(t);
1620         else if (needreport)
1621             show_stalled_task_trace(t, firstreport);
1622     }
1623 
1624     // Re-enable CPU hotplug now that the holdout list scan has completed.
1625     cpus_read_unlock();
1626 
1627     if (needreport) {
1628         if (*firstreport)
1629             pr_err("INFO: rcu_tasks_trace detected stalls? (Late IPI?)\n");
1630         show_stalled_ipi_trace();
1631     }
1632 }
1633 
1634 static void rcu_tasks_trace_empty_fn(void *unused)
1635 {
1636 }
1637 
1638 /* Wait for grace period to complete and provide ordering. */
1639 static void rcu_tasks_trace_postgp(struct rcu_tasks *rtp)
1640 {
1641     int cpu;
1642 
1643     // Wait for any lingering IPI handlers to complete.  Note that
1644     // if a CPU has gone offline or transitioned to userspace in the
1645     // meantime, all IPI handlers should have been drained beforehand.
1646     // Yes, this assumes that CPUs process IPIs in order.  If that ever
1647     // changes, there will need to be a recheck and/or timed wait.
1648     for_each_online_cpu(cpu)
1649         if (WARN_ON_ONCE(smp_load_acquire(per_cpu_ptr(&trc_ipi_to_cpu, cpu))))
1650             smp_call_function_single(cpu, rcu_tasks_trace_empty_fn, NULL, 1);
1651 
1652     smp_mb(); // Caller's code must be ordered after wakeup.
1653           // Pairs with pretty much every ordering primitive.
1654 }
1655 
1656 /* Report any needed quiescent state for this exiting task. */
1657 static void exit_tasks_rcu_finish_trace(struct task_struct *t)
1658 {
1659     union rcu_special trs = READ_ONCE(t->trc_reader_special);
1660 
1661     rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS_CHECKED);
1662     WARN_ON_ONCE(READ_ONCE(t->trc_reader_nesting));
1663     if (WARN_ON_ONCE(rcu_ld_need_qs(t) & TRC_NEED_QS || trs.b.blocked))
1664         rcu_read_unlock_trace_special(t);
1665     else
1666         WRITE_ONCE(t->trc_reader_nesting, 0);
1667 }
1668 
1669 /**
1670  * call_rcu_tasks_trace() - Queue a callback trace task-based grace period
1671  * @rhp: structure to be used for queueing the RCU updates.
1672  * @func: actual callback function to be invoked after the grace period
1673  *
1674  * The callback function will be invoked some time after a trace rcu-tasks
1675  * grace period elapses, in other words after all currently executing
1676  * trace rcu-tasks read-side critical sections have completed. These
1677  * read-side critical sections are delimited by calls to rcu_read_lock_trace()
1678  * and rcu_read_unlock_trace().
1679  *
1680  * See the description of call_rcu() for more detailed information on
1681  * memory ordering guarantees.
1682  */
1683 void call_rcu_tasks_trace(struct rcu_head *rhp, rcu_callback_t func)
1684 {
1685     call_rcu_tasks_generic(rhp, func, &rcu_tasks_trace);
1686 }
1687 EXPORT_SYMBOL_GPL(call_rcu_tasks_trace);
1688 
1689 /**
1690  * synchronize_rcu_tasks_trace - wait for a trace rcu-tasks grace period
1691  *
1692  * Control will return to the caller some time after a trace rcu-tasks
1693  * grace period has elapsed, in other words after all currently executing
1694  * trace rcu-tasks read-side critical sections have elapsed. These read-side
1695  * critical sections are delimited by calls to rcu_read_lock_trace()
1696  * and rcu_read_unlock_trace().
1697  *
1698  * This is a very specialized primitive, intended only for a few uses in
1699  * tracing and other situations requiring manipulation of function preambles
1700  * and profiling hooks.  The synchronize_rcu_tasks_trace() function is not
1701  * (yet) intended for heavy use from multiple CPUs.
1702  *
1703  * See the description of synchronize_rcu() for more detailed information
1704  * on memory ordering guarantees.
1705  */
1706 void synchronize_rcu_tasks_trace(void)
1707 {
1708     RCU_LOCKDEP_WARN(lock_is_held(&rcu_trace_lock_map), "Illegal synchronize_rcu_tasks_trace() in RCU Tasks Trace read-side critical section");
1709     synchronize_rcu_tasks_generic(&rcu_tasks_trace);
1710 }
1711 EXPORT_SYMBOL_GPL(synchronize_rcu_tasks_trace);
1712 
1713 /**
1714  * rcu_barrier_tasks_trace - Wait for in-flight call_rcu_tasks_trace() callbacks.
1715  *
1716  * Although the current implementation is guaranteed to wait, it is not
1717  * obligated to, for example, if there are no pending callbacks.
1718  */
1719 void rcu_barrier_tasks_trace(void)
1720 {
1721     rcu_barrier_tasks_generic(&rcu_tasks_trace);
1722 }
1723 EXPORT_SYMBOL_GPL(rcu_barrier_tasks_trace);
1724 
1725 static int __init rcu_spawn_tasks_trace_kthread(void)
1726 {
1727     cblist_init_generic(&rcu_tasks_trace);
1728     if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB)) {
1729         rcu_tasks_trace.gp_sleep = HZ / 10;
1730         rcu_tasks_trace.init_fract = HZ / 10;
1731     } else {
1732         rcu_tasks_trace.gp_sleep = HZ / 200;
1733         if (rcu_tasks_trace.gp_sleep <= 0)
1734             rcu_tasks_trace.gp_sleep = 1;
1735         rcu_tasks_trace.init_fract = HZ / 200;
1736         if (rcu_tasks_trace.init_fract <= 0)
1737             rcu_tasks_trace.init_fract = 1;
1738     }
1739     rcu_tasks_trace.pregp_func = rcu_tasks_trace_pregp_step;
1740     rcu_tasks_trace.postscan_func = rcu_tasks_trace_postscan;
1741     rcu_tasks_trace.holdouts_func = check_all_holdout_tasks_trace;
1742     rcu_tasks_trace.postgp_func = rcu_tasks_trace_postgp;
1743     rcu_spawn_tasks_kthread_generic(&rcu_tasks_trace);
1744     return 0;
1745 }
1746 
1747 #if !defined(CONFIG_TINY_RCU)
1748 void show_rcu_tasks_trace_gp_kthread(void)
1749 {
1750     char buf[64];
1751 
1752     sprintf(buf, "N%lu h:%lu/%lu/%lu",
1753         data_race(n_trc_holdouts),
1754         data_race(n_heavy_reader_ofl_updates),
1755         data_race(n_heavy_reader_updates),
1756         data_race(n_heavy_reader_attempts));
1757     show_rcu_tasks_generic_gp_kthread(&rcu_tasks_trace, buf);
1758 }
1759 EXPORT_SYMBOL_GPL(show_rcu_tasks_trace_gp_kthread);
1760 #endif // !defined(CONFIG_TINY_RCU)
1761 
1762 #else /* #ifdef CONFIG_TASKS_TRACE_RCU */
1763 static void exit_tasks_rcu_finish_trace(struct task_struct *t) { }
1764 #endif /* #else #ifdef CONFIG_TASKS_TRACE_RCU */
1765 
1766 #ifndef CONFIG_TINY_RCU
1767 void show_rcu_tasks_gp_kthreads(void)
1768 {
1769     show_rcu_tasks_classic_gp_kthread();
1770     show_rcu_tasks_rude_gp_kthread();
1771     show_rcu_tasks_trace_gp_kthread();
1772 }
1773 #endif /* #ifndef CONFIG_TINY_RCU */
1774 
1775 #ifdef CONFIG_PROVE_RCU
1776 struct rcu_tasks_test_desc {
1777     struct rcu_head rh;
1778     const char *name;
1779     bool notrun;
1780     unsigned long runstart;
1781 };
1782 
1783 static struct rcu_tasks_test_desc tests[] = {
1784     {
1785         .name = "call_rcu_tasks()",
1786         /* If not defined, the test is skipped. */
1787         .notrun = IS_ENABLED(CONFIG_TASKS_RCU),
1788     },
1789     {
1790         .name = "call_rcu_tasks_rude()",
1791         /* If not defined, the test is skipped. */
1792         .notrun = IS_ENABLED(CONFIG_TASKS_RUDE_RCU),
1793     },
1794     {
1795         .name = "call_rcu_tasks_trace()",
1796         /* If not defined, the test is skipped. */
1797         .notrun = IS_ENABLED(CONFIG_TASKS_TRACE_RCU)
1798     }
1799 };
1800 
1801 static void test_rcu_tasks_callback(struct rcu_head *rhp)
1802 {
1803     struct rcu_tasks_test_desc *rttd =
1804         container_of(rhp, struct rcu_tasks_test_desc, rh);
1805 
1806     pr_info("Callback from %s invoked.\n", rttd->name);
1807 
1808     rttd->notrun = false;
1809 }
1810 
1811 static void rcu_tasks_initiate_self_tests(void)
1812 {
1813     unsigned long j = jiffies;
1814 
1815     pr_info("Running RCU-tasks wait API self tests\n");
1816 #ifdef CONFIG_TASKS_RCU
1817     tests[0].runstart = j;
1818     synchronize_rcu_tasks();
1819     call_rcu_tasks(&tests[0].rh, test_rcu_tasks_callback);
1820 #endif
1821 
1822 #ifdef CONFIG_TASKS_RUDE_RCU
1823     tests[1].runstart = j;
1824     synchronize_rcu_tasks_rude();
1825     call_rcu_tasks_rude(&tests[1].rh, test_rcu_tasks_callback);
1826 #endif
1827 
1828 #ifdef CONFIG_TASKS_TRACE_RCU
1829     tests[2].runstart = j;
1830     synchronize_rcu_tasks_trace();
1831     call_rcu_tasks_trace(&tests[2].rh, test_rcu_tasks_callback);
1832 #endif
1833 }
1834 
1835 /*
1836  * Return:  0 - test passed
1837  *      1 - test failed, but have not timed out yet
1838  *     -1 - test failed and timed out
1839  */
1840 static int rcu_tasks_verify_self_tests(void)
1841 {
1842     int ret = 0;
1843     int i;
1844     unsigned long bst = rcu_task_stall_timeout;
1845 
1846     if (bst <= 0 || bst > RCU_TASK_BOOT_STALL_TIMEOUT)
1847         bst = RCU_TASK_BOOT_STALL_TIMEOUT;
1848     for (i = 0; i < ARRAY_SIZE(tests); i++) {
1849         while (tests[i].notrun) {       // still hanging.
1850             if (time_after(jiffies, tests[i].runstart + bst)) {
1851                 pr_err("%s has failed boot-time tests.\n", tests[i].name);
1852                 ret = -1;
1853                 break;
1854             }
1855             ret = 1;
1856             break;
1857         }
1858     }
1859     WARN_ON(ret < 0);
1860 
1861     return ret;
1862 }
1863 
1864 /*
1865  * Repeat the rcu_tasks_verify_self_tests() call once every second until the
1866  * test passes or has timed out.
1867  */
1868 static struct delayed_work rcu_tasks_verify_work;
1869 static void rcu_tasks_verify_work_fn(struct work_struct *work __maybe_unused)
1870 {
1871     int ret = rcu_tasks_verify_self_tests();
1872 
1873     if (ret <= 0)
1874         return;
1875 
1876     /* Test fails but not timed out yet, reschedule another check */
1877     schedule_delayed_work(&rcu_tasks_verify_work, HZ);
1878 }
1879 
1880 static int rcu_tasks_verify_schedule_work(void)
1881 {
1882     INIT_DELAYED_WORK(&rcu_tasks_verify_work, rcu_tasks_verify_work_fn);
1883     rcu_tasks_verify_work_fn(NULL);
1884     return 0;
1885 }
1886 late_initcall(rcu_tasks_verify_schedule_work);
1887 #else /* #ifdef CONFIG_PROVE_RCU */
1888 static void rcu_tasks_initiate_self_tests(void) { }
1889 #endif /* #else #ifdef CONFIG_PROVE_RCU */
1890 
1891 void __init rcu_init_tasks_generic(void)
1892 {
1893 #ifdef CONFIG_TASKS_RCU
1894     rcu_spawn_tasks_kthread();
1895 #endif
1896 
1897 #ifdef CONFIG_TASKS_RUDE_RCU
1898     rcu_spawn_tasks_rude_kthread();
1899 #endif
1900 
1901 #ifdef CONFIG_TASKS_TRACE_RCU
1902     rcu_spawn_tasks_trace_kthread();
1903 #endif
1904 
1905     // Run the self-tests.
1906     rcu_tasks_initiate_self_tests();
1907 }
1908 
1909 #else /* #ifdef CONFIG_TASKS_RCU_GENERIC */
1910 static inline void rcu_tasks_bootup_oddness(void) {}
1911 #endif /* #else #ifdef CONFIG_TASKS_RCU_GENERIC */
[ Source navigation ] [ Diff markup ] [ Identifier search ] [ general search ]

This page was automatically generated by the 2.1.0 LXR engine.

The LXR team
Valid CSS 2.1! Valid HTML 4.01!