Commit | Line | Data |
---|---|---|
eacd6f04 PM |
1 | /* SPDX-License-Identifier: GPL-2.0+ */ |
2 | /* | |
3 | * Task-based RCU implementations. | |
4 | * | |
5 | * Copyright (C) 2020 Paul E. McKenney | |
6 | */ | |
7 | ||
8fd8ca38 | 8 | #ifdef CONFIG_TASKS_RCU_GENERIC |
9b073de1 | 9 | #include "rcu_segcblist.h" |
5873b8a9 PM |
10 | |
11 | //////////////////////////////////////////////////////////////////////// | |
12 | // | |
13 | // Generic data structures. | |
14 | ||
15 | struct rcu_tasks; | |
16 | typedef void (*rcu_tasks_gp_func_t)(struct rcu_tasks *rtp); | |
e4fe5dd6 PM |
17 | typedef void (*pregp_func_t)(void); |
18 | typedef void (*pertask_func_t)(struct task_struct *t, struct list_head *hop); | |
9796e1ae | 19 | typedef void (*postscan_func_t)(struct list_head *hop); |
e4fe5dd6 | 20 | typedef void (*holdouts_func_t)(struct list_head *hop, bool ndrpt, bool *frptp); |
af051ca4 | 21 | typedef void (*postgp_func_t)(struct rcu_tasks *rtp); |
eacd6f04 | 22 | |
07e10515 | 23 | /** |
cafafd67 | 24 | * struct rcu_tasks_percpu - Per-CPU component of definition for a Tasks-RCU-like mechanism. |
9b073de1 | 25 | * @cblist: Callback list. |
381a4f3b | 26 | * @lock: Lock protecting per-CPU callback list. |
7d13d30b PM |
27 | * @rtp_jiffies: Jiffies counter value for statistics. |
28 | * @rtp_n_lock_retries: Rough lock-contention statistic. | |
d363f833 | 29 | * @rtp_work: Work queue for invoking callbacks. |
3063b33a | 30 | * @rtp_irq_work: IRQ work queue for deferred wakeups. |
ce9b1c66 PM |
31 | * @barrier_q_head: RCU callback for barrier operation. |
32 | * @cpu: CPU number corresponding to this entry. | |
33 | * @rtpp: Pointer to the rcu_tasks structure. | |
cafafd67 PM |
34 | */ |
35 | struct rcu_tasks_percpu { | |
9b073de1 | 36 | struct rcu_segcblist cblist; |
381a4f3b | 37 | raw_spinlock_t __private lock; |
7d13d30b PM |
38 | unsigned long rtp_jiffies; |
39 | unsigned long rtp_n_lock_retries; | |
d363f833 | 40 | struct work_struct rtp_work; |
3063b33a | 41 | struct irq_work rtp_irq_work; |
ce9b1c66 | 42 | struct rcu_head barrier_q_head; |
d363f833 PM |
43 | int cpu; |
44 | struct rcu_tasks *rtpp; | |
cafafd67 PM |
45 | }; |
46 | ||
47 | /** | |
48 | * struct rcu_tasks - Definition for a Tasks-RCU-like mechanism. | |
88db792b | 49 | * @cbs_wait: RCU wait allowing a new callback to get kthread's attention. |
cafafd67 | 50 | * @cbs_gbl_lock: Lock protecting callback list. |
d96225fd | 51 | * @tasks_gp_mutex: Mutex protecting grace period, needed during mid-boot dead zone. |
07e10515 | 52 | * @kthread_ptr: This flavor's grace-period/callback-invocation kthread. |
5873b8a9 | 53 | * @gp_func: This flavor's grace-period-wait function. |
af051ca4 | 54 | * @gp_state: Grace period's most recent state transition (debugging). |
4fe192df | 55 | * @gp_sleep: Per-grace-period sleep to prevent CPU-bound looping. |
2393a613 | 56 | * @init_fract: Initial backoff sleep interval. |
af051ca4 PM |
57 | * @gp_jiffies: Time of last @gp_state transition. |
58 | * @gp_start: Most recent grace-period start in jiffies. | |
b14fb4fb | 59 | * @tasks_gp_seq: Number of grace periods completed since boot. |
238dbce3 | 60 | * @n_ipis: Number of IPIs sent to encourage grace periods to end. |
7e0669c3 | 61 | * @n_ipis_fails: Number of IPI-send failures. |
e4fe5dd6 PM |
62 | * @pregp_func: This flavor's pre-grace-period function (optional). |
63 | * @pertask_func: This flavor's per-task scan function (optional). | |
64 | * @postscan_func: This flavor's post-task scan function (optional). | |
85b86994 | 65 | * @holdouts_func: This flavor's holdout-list scan function (optional). |
e4fe5dd6 | 66 | * @postgp_func: This flavor's post-grace-period function (optional). |
5873b8a9 | 67 | * @call_func: This flavor's call_rcu()-equivalent function. |
cafafd67 | 68 | * @rtpcpu: This flavor's rcu_tasks_percpu structure. |
7a30871b | 69 | * @percpu_enqueue_shift: Shift down CPU ID this much when enqueuing callbacks. |
2cee0789 PM |
70 | * @percpu_enqueue_lim: Number of per-CPU callback queues in use for enqueuing. |
71 | * @percpu_dequeue_lim: Number of per-CPU callback queues in use for dequeuing. | |
fd796e41 | 72 | * @percpu_dequeue_gpseq: RCU grace-period number to propagate enqueue limit to dequeuers. |
ce9b1c66 PM |
73 | * @barrier_q_mutex: Serialize barrier operations. |
74 | * @barrier_q_count: Number of queues being waited on. | |
75 | * @barrier_q_completion: Barrier wait/wakeup mechanism. | |
76 | * @barrier_q_seq: Sequence number for barrier operations. | |
c97d12a6 PM |
77 | * @name: This flavor's textual name. |
78 | * @kname: This flavor's kthread name. | |
07e10515 PM |
79 | */ |
80 | struct rcu_tasks { | |
88db792b | 81 | struct rcuwait cbs_wait; |
cafafd67 | 82 | raw_spinlock_t cbs_gbl_lock; |
d96225fd | 83 | struct mutex tasks_gp_mutex; |
af051ca4 | 84 | int gp_state; |
4fe192df | 85 | int gp_sleep; |
2393a613 | 86 | int init_fract; |
af051ca4 | 87 | unsigned long gp_jiffies; |
88092d0c | 88 | unsigned long gp_start; |
b14fb4fb | 89 | unsigned long tasks_gp_seq; |
238dbce3 | 90 | unsigned long n_ipis; |
7e0669c3 | 91 | unsigned long n_ipis_fails; |
07e10515 | 92 | struct task_struct *kthread_ptr; |
5873b8a9 | 93 | rcu_tasks_gp_func_t gp_func; |
e4fe5dd6 PM |
94 | pregp_func_t pregp_func; |
95 | pertask_func_t pertask_func; | |
96 | postscan_func_t postscan_func; | |
97 | holdouts_func_t holdouts_func; | |
98 | postgp_func_t postgp_func; | |
5873b8a9 | 99 | call_rcu_func_t call_func; |
cafafd67 | 100 | struct rcu_tasks_percpu __percpu *rtpcpu; |
7a30871b | 101 | int percpu_enqueue_shift; |
8dd593fd | 102 | int percpu_enqueue_lim; |
2cee0789 | 103 | int percpu_dequeue_lim; |
fd796e41 | 104 | unsigned long percpu_dequeue_gpseq; |
ce9b1c66 PM |
105 | struct mutex barrier_q_mutex; |
106 | atomic_t barrier_q_count; | |
107 | struct completion barrier_q_completion; | |
108 | unsigned long barrier_q_seq; | |
c97d12a6 PM |
109 | char *name; |
110 | char *kname; | |
07e10515 PM |
111 | }; |
112 | ||
3063b33a PM |
113 | static void call_rcu_tasks_iw_wakeup(struct irq_work *iwp); |
114 | ||
cafafd67 PM |
115 | #define DEFINE_RCU_TASKS(rt_name, gp, call, n) \ |
116 | static DEFINE_PER_CPU(struct rcu_tasks_percpu, rt_name ## __percpu) = { \ | |
381a4f3b | 117 | .lock = __RAW_SPIN_LOCK_UNLOCKED(rt_name ## __percpu.cbs_pcpu_lock), \ |
88db792b | 118 | .rtp_irq_work = IRQ_WORK_INIT_HARD(call_rcu_tasks_iw_wakeup), \ |
cafafd67 PM |
119 | }; \ |
120 | static struct rcu_tasks rt_name = \ | |
121 | { \ | |
88db792b | 122 | .cbs_wait = __RCUWAIT_INITIALIZER(rt_name.wait), \ |
cafafd67 | 123 | .cbs_gbl_lock = __RAW_SPIN_LOCK_UNLOCKED(rt_name.cbs_gbl_lock), \ |
d96225fd | 124 | .tasks_gp_mutex = __MUTEX_INITIALIZER(rt_name.tasks_gp_mutex), \ |
cafafd67 PM |
125 | .gp_func = gp, \ |
126 | .call_func = call, \ | |
127 | .rtpcpu = &rt_name ## __percpu, \ | |
128 | .name = n, \ | |
2bcd18e0 | 129 | .percpu_enqueue_shift = order_base_2(CONFIG_NR_CPUS), \ |
8dd593fd | 130 | .percpu_enqueue_lim = 1, \ |
2cee0789 | 131 | .percpu_dequeue_lim = 1, \ |
ce9b1c66 PM |
132 | .barrier_q_mutex = __MUTEX_INITIALIZER(rt_name.barrier_q_mutex), \ |
133 | .barrier_q_seq = (0UL - 50UL) << RCU_SEQ_CTR_SHIFT, \ | |
cafafd67 | 134 | .kname = #rt_name, \ |
07e10515 PM |
135 | } |
136 | ||
eacd6f04 PM |
137 | /* Track exiting tasks in order to allow them to be waited for. */ |
138 | DEFINE_STATIC_SRCU(tasks_rcu_exit_srcu); | |
139 | ||
b0afa0f0 | 140 | /* Avoid IPIing CPUs early in the grace period. */ |
574de876 | 141 | #define RCU_TASK_IPI_DELAY (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB) ? HZ / 2 : 0) |
b0afa0f0 PM |
142 | static int rcu_task_ipi_delay __read_mostly = RCU_TASK_IPI_DELAY; |
143 | module_param(rcu_task_ipi_delay, int, 0644); | |
144 | ||
eacd6f04 PM |
145 | /* Control stall timeouts. Disable with <= 0, otherwise jiffies till stall. */ |
146 | #define RCU_TASK_STALL_TIMEOUT (HZ * 60 * 10) | |
147 | static int rcu_task_stall_timeout __read_mostly = RCU_TASK_STALL_TIMEOUT; | |
148 | module_param(rcu_task_stall_timeout, int, 0644); | |
f2539003 PM |
149 | #define RCU_TASK_STALL_INFO (HZ * 10) |
150 | static int rcu_task_stall_info __read_mostly = RCU_TASK_STALL_INFO; | |
151 | module_param(rcu_task_stall_info, int, 0644); | |
152 | static int rcu_task_stall_info_mult __read_mostly = 3; | |
153 | module_param(rcu_task_stall_info_mult, int, 0444); | |
eacd6f04 | 154 | |
8610b656 PM |
155 | static int rcu_task_enqueue_lim __read_mostly = -1; |
156 | module_param(rcu_task_enqueue_lim, int, 0444); | |
157 | ||
ab97152f PM |
158 | static bool rcu_task_cb_adjust; |
159 | static int rcu_task_contend_lim __read_mostly = 100; | |
160 | module_param(rcu_task_contend_lim, int, 0444); | |
fd796e41 PM |
161 | static int rcu_task_collapse_lim __read_mostly = 10; |
162 | module_param(rcu_task_collapse_lim, int, 0444); | |
ab97152f | 163 | |
af051ca4 PM |
164 | /* RCU tasks grace-period state for debugging. */ |
165 | #define RTGS_INIT 0 | |
166 | #define RTGS_WAIT_WAIT_CBS 1 | |
167 | #define RTGS_WAIT_GP 2 | |
168 | #define RTGS_PRE_WAIT_GP 3 | |
169 | #define RTGS_SCAN_TASKLIST 4 | |
170 | #define RTGS_POST_SCAN_TASKLIST 5 | |
171 | #define RTGS_WAIT_SCAN_HOLDOUTS 6 | |
172 | #define RTGS_SCAN_HOLDOUTS 7 | |
173 | #define RTGS_POST_GP 8 | |
174 | #define RTGS_WAIT_READERS 9 | |
175 | #define RTGS_INVOKE_CBS 10 | |
176 | #define RTGS_WAIT_CBS 11 | |
8344496e | 177 | #ifndef CONFIG_TINY_RCU |
af051ca4 PM |
178 | static const char * const rcu_tasks_gp_state_names[] = { |
179 | "RTGS_INIT", | |
180 | "RTGS_WAIT_WAIT_CBS", | |
181 | "RTGS_WAIT_GP", | |
182 | "RTGS_PRE_WAIT_GP", | |
183 | "RTGS_SCAN_TASKLIST", | |
184 | "RTGS_POST_SCAN_TASKLIST", | |
185 | "RTGS_WAIT_SCAN_HOLDOUTS", | |
186 | "RTGS_SCAN_HOLDOUTS", | |
187 | "RTGS_POST_GP", | |
188 | "RTGS_WAIT_READERS", | |
189 | "RTGS_INVOKE_CBS", | |
190 | "RTGS_WAIT_CBS", | |
191 | }; | |
8344496e | 192 | #endif /* #ifndef CONFIG_TINY_RCU */ |
af051ca4 | 193 | |
5873b8a9 PM |
194 | //////////////////////////////////////////////////////////////////////// |
195 | // | |
196 | // Generic code. | |
197 | ||
d363f833 PM |
198 | static void rcu_tasks_invoke_cbs_wq(struct work_struct *wp); |
199 | ||
af051ca4 PM |
200 | /* Record grace-period phase and time. */ |
201 | static void set_tasks_gp_state(struct rcu_tasks *rtp, int newstate) | |
202 | { | |
203 | rtp->gp_state = newstate; | |
204 | rtp->gp_jiffies = jiffies; | |
205 | } | |
206 | ||
8344496e | 207 | #ifndef CONFIG_TINY_RCU |
af051ca4 PM |
208 | /* Return state name. */ |
209 | static const char *tasks_gp_state_getname(struct rcu_tasks *rtp) | |
210 | { | |
211 | int i = data_race(rtp->gp_state); // Let KCSAN detect update races | |
212 | int j = READ_ONCE(i); // Prevent the compiler from reading twice | |
213 | ||
214 | if (j >= ARRAY_SIZE(rcu_tasks_gp_state_names)) | |
215 | return "???"; | |
216 | return rcu_tasks_gp_state_names[j]; | |
217 | } | |
8344496e | 218 | #endif /* #ifndef CONFIG_TINY_RCU */ |
af051ca4 | 219 | |
cafafd67 PM |
220 | // Initialize per-CPU callback lists for the specified flavor of |
221 | // Tasks RCU. | |
222 | static void cblist_init_generic(struct rcu_tasks *rtp) | |
223 | { | |
224 | int cpu; | |
225 | unsigned long flags; | |
8610b656 | 226 | int lim; |
da123016 | 227 | int shift; |
cafafd67 PM |
228 | |
229 | raw_spin_lock_irqsave(&rtp->cbs_gbl_lock, flags); | |
ab97152f PM |
230 | if (rcu_task_enqueue_lim < 0) { |
231 | rcu_task_enqueue_lim = 1; | |
232 | rcu_task_cb_adjust = true; | |
233 | pr_info("%s: Setting adjustable number of callback queues.\n", __func__); | |
234 | } else if (rcu_task_enqueue_lim == 0) { | |
8610b656 | 235 | rcu_task_enqueue_lim = 1; |
ab97152f | 236 | } |
8610b656 PM |
237 | lim = rcu_task_enqueue_lim; |
238 | ||
239 | if (lim > nr_cpu_ids) | |
240 | lim = nr_cpu_ids; | |
da123016 PM |
241 | shift = ilog2(nr_cpu_ids / lim); |
242 | if (((nr_cpu_ids - 1) >> shift) >= lim) | |
243 | shift++; | |
244 | WRITE_ONCE(rtp->percpu_enqueue_shift, shift); | |
2cee0789 | 245 | WRITE_ONCE(rtp->percpu_dequeue_lim, lim); |
8610b656 | 246 | smp_store_release(&rtp->percpu_enqueue_lim, lim); |
cafafd67 PM |
247 | for_each_possible_cpu(cpu) { |
248 | struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu); | |
249 | ||
250 | WARN_ON_ONCE(!rtpcp); | |
251 | if (cpu) | |
381a4f3b PM |
252 | raw_spin_lock_init(&ACCESS_PRIVATE(rtpcp, lock)); |
253 | raw_spin_lock_rcu_node(rtpcp); // irqs already disabled. | |
9b073de1 PM |
254 | if (rcu_segcblist_empty(&rtpcp->cblist)) |
255 | rcu_segcblist_init(&rtpcp->cblist); | |
d363f833 PM |
256 | INIT_WORK(&rtpcp->rtp_work, rcu_tasks_invoke_cbs_wq); |
257 | rtpcp->cpu = cpu; | |
258 | rtpcp->rtpp = rtp; | |
381a4f3b | 259 | raw_spin_unlock_rcu_node(rtpcp); // irqs remain disabled. |
cafafd67 PM |
260 | } |
261 | raw_spin_unlock_irqrestore(&rtp->cbs_gbl_lock, flags); | |
8610b656 | 262 | 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)); |
cafafd67 PM |
263 | } |
264 | ||
3063b33a PM |
265 | // IRQ-work handler that does deferred wakeup for call_rcu_tasks_generic(). |
266 | static void call_rcu_tasks_iw_wakeup(struct irq_work *iwp) | |
267 | { | |
268 | struct rcu_tasks *rtp; | |
269 | struct rcu_tasks_percpu *rtpcp = container_of(iwp, struct rcu_tasks_percpu, rtp_irq_work); | |
270 | ||
271 | rtp = rtpcp->rtpp; | |
88db792b | 272 | rcuwait_wake_up(&rtp->cbs_wait); |
3063b33a PM |
273 | } |
274 | ||
5873b8a9 PM |
275 | // Enqueue a callback for the specified flavor of Tasks RCU. |
276 | static void call_rcu_tasks_generic(struct rcu_head *rhp, rcu_callback_t func, | |
277 | struct rcu_tasks *rtp) | |
eacd6f04 | 278 | { |
07d95c34 | 279 | int chosen_cpu; |
eacd6f04 | 280 | unsigned long flags; |
07d95c34 | 281 | int ideal_cpu; |
7d13d30b | 282 | unsigned long j; |
ab97152f | 283 | bool needadjust = false; |
eacd6f04 | 284 | bool needwake; |
cafafd67 | 285 | struct rcu_tasks_percpu *rtpcp; |
eacd6f04 PM |
286 | |
287 | rhp->next = NULL; | |
288 | rhp->func = func; | |
cafafd67 | 289 | local_irq_save(flags); |
fd796e41 | 290 | rcu_read_lock(); |
07d95c34 ED |
291 | ideal_cpu = smp_processor_id() >> READ_ONCE(rtp->percpu_enqueue_shift); |
292 | chosen_cpu = cpumask_next(ideal_cpu - 1, cpu_possible_mask); | |
293 | rtpcp = per_cpu_ptr(rtp->rtpcpu, chosen_cpu); | |
7d13d30b PM |
294 | if (!raw_spin_trylock_rcu_node(rtpcp)) { // irqs already disabled. |
295 | raw_spin_lock_rcu_node(rtpcp); // irqs already disabled. | |
296 | j = jiffies; | |
297 | if (rtpcp->rtp_jiffies != j) { | |
298 | rtpcp->rtp_jiffies = j; | |
299 | rtpcp->rtp_n_lock_retries = 0; | |
300 | } | |
ab97152f PM |
301 | if (rcu_task_cb_adjust && ++rtpcp->rtp_n_lock_retries > rcu_task_contend_lim && |
302 | READ_ONCE(rtp->percpu_enqueue_lim) != nr_cpu_ids) | |
303 | needadjust = true; // Defer adjustment to avoid deadlock. | |
7d13d30b | 304 | } |
9b073de1 | 305 | if (!rcu_segcblist_is_enabled(&rtpcp->cblist)) { |
381a4f3b | 306 | raw_spin_unlock_rcu_node(rtpcp); // irqs remain disabled. |
cafafd67 | 307 | cblist_init_generic(rtp); |
381a4f3b | 308 | raw_spin_lock_rcu_node(rtpcp); // irqs already disabled. |
cafafd67 | 309 | } |
9b073de1 PM |
310 | needwake = rcu_segcblist_empty(&rtpcp->cblist); |
311 | rcu_segcblist_enqueue(&rtpcp->cblist, rhp); | |
381a4f3b | 312 | raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags); |
ab97152f PM |
313 | if (unlikely(needadjust)) { |
314 | raw_spin_lock_irqsave(&rtp->cbs_gbl_lock, flags); | |
315 | if (rtp->percpu_enqueue_lim != nr_cpu_ids) { | |
00a8b4b5 | 316 | WRITE_ONCE(rtp->percpu_enqueue_shift, 0); |
fd796e41 | 317 | WRITE_ONCE(rtp->percpu_dequeue_lim, nr_cpu_ids); |
ab97152f PM |
318 | smp_store_release(&rtp->percpu_enqueue_lim, nr_cpu_ids); |
319 | pr_info("Switching %s to per-CPU callback queuing.\n", rtp->name); | |
320 | } | |
321 | raw_spin_unlock_irqrestore(&rtp->cbs_gbl_lock, flags); | |
322 | } | |
fd796e41 | 323 | rcu_read_unlock(); |
eacd6f04 | 324 | /* We can't create the thread unless interrupts are enabled. */ |
07e10515 | 325 | if (needwake && READ_ONCE(rtp->kthread_ptr)) |
3063b33a | 326 | irq_work_queue(&rtpcp->rtp_irq_work); |
eacd6f04 | 327 | } |
eacd6f04 | 328 | |
ce9b1c66 PM |
329 | // RCU callback function for rcu_barrier_tasks_generic(). |
330 | static void rcu_barrier_tasks_generic_cb(struct rcu_head *rhp) | |
331 | { | |
332 | struct rcu_tasks *rtp; | |
333 | struct rcu_tasks_percpu *rtpcp; | |
334 | ||
335 | rtpcp = container_of(rhp, struct rcu_tasks_percpu, barrier_q_head); | |
336 | rtp = rtpcp->rtpp; | |
337 | if (atomic_dec_and_test(&rtp->barrier_q_count)) | |
338 | complete(&rtp->barrier_q_completion); | |
339 | } | |
340 | ||
341 | // Wait for all in-flight callbacks for the specified RCU Tasks flavor. | |
342 | // Operates in a manner similar to rcu_barrier(). | |
343 | static void rcu_barrier_tasks_generic(struct rcu_tasks *rtp) | |
344 | { | |
345 | int cpu; | |
346 | unsigned long flags; | |
347 | struct rcu_tasks_percpu *rtpcp; | |
348 | unsigned long s = rcu_seq_snap(&rtp->barrier_q_seq); | |
349 | ||
350 | mutex_lock(&rtp->barrier_q_mutex); | |
351 | if (rcu_seq_done(&rtp->barrier_q_seq, s)) { | |
352 | smp_mb(); | |
353 | mutex_unlock(&rtp->barrier_q_mutex); | |
354 | return; | |
355 | } | |
356 | rcu_seq_start(&rtp->barrier_q_seq); | |
357 | init_completion(&rtp->barrier_q_completion); | |
358 | atomic_set(&rtp->barrier_q_count, 2); | |
359 | for_each_possible_cpu(cpu) { | |
2cee0789 | 360 | if (cpu >= smp_load_acquire(&rtp->percpu_dequeue_lim)) |
ce9b1c66 PM |
361 | break; |
362 | rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu); | |
363 | rtpcp->barrier_q_head.func = rcu_barrier_tasks_generic_cb; | |
364 | raw_spin_lock_irqsave_rcu_node(rtpcp, flags); | |
365 | if (rcu_segcblist_entrain(&rtpcp->cblist, &rtpcp->barrier_q_head)) | |
366 | atomic_inc(&rtp->barrier_q_count); | |
367 | raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags); | |
368 | } | |
369 | if (atomic_sub_and_test(2, &rtp->barrier_q_count)) | |
370 | complete(&rtp->barrier_q_completion); | |
371 | wait_for_completion(&rtp->barrier_q_completion); | |
372 | rcu_seq_end(&rtp->barrier_q_seq); | |
373 | mutex_unlock(&rtp->barrier_q_mutex); | |
374 | } | |
375 | ||
4d1114c0 PM |
376 | // Advance callbacks and indicate whether either a grace period or |
377 | // callback invocation is needed. | |
378 | static int rcu_tasks_need_gpcb(struct rcu_tasks *rtp) | |
379 | { | |
380 | int cpu; | |
381 | unsigned long flags; | |
fd796e41 PM |
382 | long n; |
383 | long ncbs = 0; | |
384 | long ncbsnz = 0; | |
4d1114c0 PM |
385 | int needgpcb = 0; |
386 | ||
2cee0789 | 387 | for (cpu = 0; cpu < smp_load_acquire(&rtp->percpu_dequeue_lim); cpu++) { |
4d1114c0 PM |
388 | struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu); |
389 | ||
390 | /* Advance and accelerate any new callbacks. */ | |
fd796e41 | 391 | if (!rcu_segcblist_n_cbs(&rtpcp->cblist)) |
4d1114c0 PM |
392 | continue; |
393 | raw_spin_lock_irqsave_rcu_node(rtpcp, flags); | |
fd796e41 PM |
394 | // Should we shrink down to a single callback queue? |
395 | n = rcu_segcblist_n_cbs(&rtpcp->cblist); | |
396 | if (n) { | |
397 | ncbs += n; | |
398 | if (cpu > 0) | |
399 | ncbsnz += n; | |
400 | } | |
4d1114c0 PM |
401 | rcu_segcblist_advance(&rtpcp->cblist, rcu_seq_current(&rtp->tasks_gp_seq)); |
402 | (void)rcu_segcblist_accelerate(&rtpcp->cblist, rcu_seq_snap(&rtp->tasks_gp_seq)); | |
403 | if (rcu_segcblist_pend_cbs(&rtpcp->cblist)) | |
404 | needgpcb |= 0x3; | |
405 | if (!rcu_segcblist_empty(&rtpcp->cblist)) | |
406 | needgpcb |= 0x1; | |
407 | raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags); | |
408 | } | |
fd796e41 PM |
409 | |
410 | // Shrink down to a single callback queue if appropriate. | |
411 | // This is done in two stages: (1) If there are no more than | |
412 | // rcu_task_collapse_lim callbacks on CPU 0 and none on any other | |
413 | // CPU, limit enqueueing to CPU 0. (2) After an RCU grace period, | |
414 | // if there has not been an increase in callbacks, limit dequeuing | |
415 | // to CPU 0. Note the matching RCU read-side critical section in | |
416 | // call_rcu_tasks_generic(). | |
417 | if (rcu_task_cb_adjust && ncbs <= rcu_task_collapse_lim) { | |
418 | raw_spin_lock_irqsave(&rtp->cbs_gbl_lock, flags); | |
419 | if (rtp->percpu_enqueue_lim > 1) { | |
2bcd18e0 | 420 | WRITE_ONCE(rtp->percpu_enqueue_shift, order_base_2(nr_cpu_ids)); |
fd796e41 PM |
421 | smp_store_release(&rtp->percpu_enqueue_lim, 1); |
422 | rtp->percpu_dequeue_gpseq = get_state_synchronize_rcu(); | |
423 | pr_info("Starting switch %s to CPU-0 callback queuing.\n", rtp->name); | |
424 | } | |
425 | raw_spin_unlock_irqrestore(&rtp->cbs_gbl_lock, flags); | |
426 | } | |
427 | if (rcu_task_cb_adjust && !ncbsnz && | |
428 | poll_state_synchronize_rcu(rtp->percpu_dequeue_gpseq)) { | |
429 | raw_spin_lock_irqsave(&rtp->cbs_gbl_lock, flags); | |
430 | if (rtp->percpu_enqueue_lim < rtp->percpu_dequeue_lim) { | |
431 | WRITE_ONCE(rtp->percpu_dequeue_lim, 1); | |
432 | pr_info("Completing switch %s to CPU-0 callback queuing.\n", rtp->name); | |
433 | } | |
4cf0585c PM |
434 | for (cpu = rtp->percpu_dequeue_lim; cpu < nr_cpu_ids; cpu++) { |
435 | struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu); | |
436 | ||
437 | WARN_ON_ONCE(rcu_segcblist_n_cbs(&rtpcp->cblist)); | |
438 | } | |
fd796e41 PM |
439 | raw_spin_unlock_irqrestore(&rtp->cbs_gbl_lock, flags); |
440 | } | |
441 | ||
4d1114c0 PM |
442 | return needgpcb; |
443 | } | |
444 | ||
57881863 | 445 | // Advance callbacks and invoke any that are ready. |
d363f833 | 446 | static void rcu_tasks_invoke_cbs(struct rcu_tasks *rtp, struct rcu_tasks_percpu *rtpcp) |
eacd6f04 | 447 | { |
57881863 | 448 | int cpu; |
d363f833 | 449 | int cpunext; |
eacd6f04 | 450 | unsigned long flags; |
9b073de1 | 451 | int len; |
9b073de1 | 452 | struct rcu_head *rhp; |
d363f833 PM |
453 | struct rcu_cblist rcl = RCU_CBLIST_INITIALIZER(rcl); |
454 | struct rcu_tasks_percpu *rtpcp_next; | |
455 | ||
456 | cpu = rtpcp->cpu; | |
457 | cpunext = cpu * 2 + 1; | |
2cee0789 | 458 | if (cpunext < smp_load_acquire(&rtp->percpu_dequeue_lim)) { |
d363f833 PM |
459 | rtpcp_next = per_cpu_ptr(rtp->rtpcpu, cpunext); |
460 | queue_work_on(cpunext, system_wq, &rtpcp_next->rtp_work); | |
461 | cpunext++; | |
2cee0789 | 462 | if (cpunext < smp_load_acquire(&rtp->percpu_dequeue_lim)) { |
d363f833 PM |
463 | rtpcp_next = per_cpu_ptr(rtp->rtpcpu, cpunext); |
464 | queue_work_on(cpunext, system_wq, &rtpcp_next->rtp_work); | |
57881863 | 465 | } |
57881863 | 466 | } |
d363f833 | 467 | |
ab2756ea | 468 | if (rcu_segcblist_empty(&rtpcp->cblist) || !cpu_possible(cpu)) |
d363f833 PM |
469 | return; |
470 | raw_spin_lock_irqsave_rcu_node(rtpcp, flags); | |
471 | rcu_segcblist_advance(&rtpcp->cblist, rcu_seq_current(&rtp->tasks_gp_seq)); | |
472 | rcu_segcblist_extract_done_cbs(&rtpcp->cblist, &rcl); | |
473 | raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags); | |
474 | len = rcl.len; | |
475 | for (rhp = rcu_cblist_dequeue(&rcl); rhp; rhp = rcu_cblist_dequeue(&rcl)) { | |
476 | local_bh_disable(); | |
477 | rhp->func(rhp); | |
478 | local_bh_enable(); | |
479 | cond_resched(); | |
480 | } | |
481 | raw_spin_lock_irqsave_rcu_node(rtpcp, flags); | |
482 | rcu_segcblist_add_len(&rtpcp->cblist, -len); | |
483 | (void)rcu_segcblist_accelerate(&rtpcp->cblist, rcu_seq_snap(&rtp->tasks_gp_seq)); | |
484 | raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags); | |
485 | } | |
486 | ||
487 | // Workqueue flood to advance callbacks and invoke any that are ready. | |
488 | static void rcu_tasks_invoke_cbs_wq(struct work_struct *wp) | |
489 | { | |
490 | struct rcu_tasks *rtp; | |
491 | struct rcu_tasks_percpu *rtpcp = container_of(wp, struct rcu_tasks_percpu, rtp_work); | |
492 | ||
493 | rtp = rtpcp->rtpp; | |
494 | rcu_tasks_invoke_cbs(rtp, rtpcp); | |
57881863 PM |
495 | } |
496 | ||
d96225fd | 497 | // Wait for one grace period. |
4a8cc433 | 498 | static void rcu_tasks_one_gp(struct rcu_tasks *rtp, bool midboot) |
57881863 PM |
499 | { |
500 | int needgpcb; | |
d96225fd PM |
501 | |
502 | mutex_lock(&rtp->tasks_gp_mutex); | |
d96225fd PM |
503 | |
504 | // If there were none, wait a bit and start over. | |
4a8cc433 PM |
505 | if (unlikely(midboot)) { |
506 | needgpcb = 0x2; | |
507 | } else { | |
508 | set_tasks_gp_state(rtp, RTGS_WAIT_CBS); | |
509 | rcuwait_wait_event(&rtp->cbs_wait, | |
510 | (needgpcb = rcu_tasks_need_gpcb(rtp)), | |
511 | TASK_IDLE); | |
512 | } | |
d96225fd PM |
513 | |
514 | if (needgpcb & 0x2) { | |
515 | // Wait for one grace period. | |
516 | set_tasks_gp_state(rtp, RTGS_WAIT_GP); | |
517 | rtp->gp_start = jiffies; | |
518 | rcu_seq_start(&rtp->tasks_gp_seq); | |
519 | rtp->gp_func(rtp); | |
520 | rcu_seq_end(&rtp->tasks_gp_seq); | |
521 | } | |
522 | ||
523 | // Invoke callbacks. | |
524 | set_tasks_gp_state(rtp, RTGS_INVOKE_CBS); | |
525 | rcu_tasks_invoke_cbs(rtp, per_cpu_ptr(rtp->rtpcpu, 0)); | |
526 | mutex_unlock(&rtp->tasks_gp_mutex); | |
527 | } | |
528 | ||
529 | // RCU-tasks kthread that detects grace periods and invokes callbacks. | |
530 | static int __noreturn rcu_tasks_kthread(void *arg) | |
531 | { | |
07e10515 | 532 | struct rcu_tasks *rtp = arg; |
eacd6f04 PM |
533 | |
534 | /* Run on housekeeping CPUs by default. Sysadm can move if desired. */ | |
04d4e665 | 535 | housekeeping_affine(current, HK_TYPE_RCU); |
07e10515 | 536 | WRITE_ONCE(rtp->kthread_ptr, current); // Let GPs start! |
eacd6f04 PM |
537 | |
538 | /* | |
539 | * Each pass through the following loop makes one check for | |
540 | * newly arrived callbacks, and, if there are some, waits for | |
541 | * one RCU-tasks grace period and then invokes the callbacks. | |
542 | * This loop is terminated by the system going down. ;-) | |
543 | */ | |
544 | for (;;) { | |
d96225fd PM |
545 | // Wait for one grace period and invoke any callbacks |
546 | // that are ready. | |
4a8cc433 | 547 | rcu_tasks_one_gp(rtp, false); |
57881863 | 548 | |
d96225fd | 549 | // Paranoid sleep to keep this from entering a tight loop. |
4fe192df | 550 | schedule_timeout_idle(rtp->gp_sleep); |
eacd6f04 PM |
551 | } |
552 | } | |
553 | ||
68cb4720 PM |
554 | // Wait for a grace period for the specified flavor of Tasks RCU. |
555 | static void synchronize_rcu_tasks_generic(struct rcu_tasks *rtp) | |
556 | { | |
557 | /* Complain if the scheduler has not started. */ | |
558 | RCU_LOCKDEP_WARN(rcu_scheduler_active == RCU_SCHEDULER_INACTIVE, | |
559 | "synchronize_rcu_tasks called too soon"); | |
560 | ||
4a8cc433 PM |
561 | // If the grace-period kthread is running, use it. |
562 | if (READ_ONCE(rtp->kthread_ptr)) { | |
563 | wait_rcu_gp(rtp->call_func); | |
564 | return; | |
565 | } | |
566 | rcu_tasks_one_gp(rtp, true); | |
68cb4720 PM |
567 | } |
568 | ||
1b04fa99 | 569 | /* Spawn RCU-tasks grace-period kthread. */ |
5873b8a9 | 570 | static void __init rcu_spawn_tasks_kthread_generic(struct rcu_tasks *rtp) |
eacd6f04 PM |
571 | { |
572 | struct task_struct *t; | |
573 | ||
c97d12a6 PM |
574 | t = kthread_run(rcu_tasks_kthread, rtp, "%s_kthread", rtp->kname); |
575 | if (WARN_ONCE(IS_ERR(t), "%s: Could not start %s grace-period kthread, OOM is now expected behavior\n", __func__, rtp->name)) | |
5873b8a9 | 576 | return; |
eacd6f04 | 577 | smp_mb(); /* Ensure others see full kthread. */ |
eacd6f04 | 578 | } |
eacd6f04 | 579 | |
eacd6f04 PM |
580 | #ifndef CONFIG_TINY_RCU |
581 | ||
582 | /* | |
583 | * Print any non-default Tasks RCU settings. | |
584 | */ | |
585 | static void __init rcu_tasks_bootup_oddness(void) | |
586 | { | |
d5f177d3 | 587 | #if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU) |
f2539003 PM |
588 | int rtsimc; |
589 | ||
eacd6f04 PM |
590 | if (rcu_task_stall_timeout != RCU_TASK_STALL_TIMEOUT) |
591 | pr_info("\tTasks-RCU CPU stall warnings timeout set to %d (rcu_task_stall_timeout).\n", rcu_task_stall_timeout); | |
f2539003 PM |
592 | rtsimc = clamp(rcu_task_stall_info_mult, 1, 10); |
593 | if (rtsimc != rcu_task_stall_info_mult) { | |
594 | pr_info("\tTasks-RCU CPU stall info multiplier clamped to %d (rcu_task_stall_info_mult).\n", rtsimc); | |
595 | rcu_task_stall_info_mult = rtsimc; | |
596 | } | |
d5f177d3 PM |
597 | #endif /* #ifdef CONFIG_TASKS_RCU */ |
598 | #ifdef CONFIG_TASKS_RCU | |
599 | pr_info("\tTrampoline variant of Tasks RCU enabled.\n"); | |
eacd6f04 | 600 | #endif /* #ifdef CONFIG_TASKS_RCU */ |
c84aad76 PM |
601 | #ifdef CONFIG_TASKS_RUDE_RCU |
602 | pr_info("\tRude variant of Tasks RCU enabled.\n"); | |
603 | #endif /* #ifdef CONFIG_TASKS_RUDE_RCU */ | |
d5f177d3 PM |
604 | #ifdef CONFIG_TASKS_TRACE_RCU |
605 | pr_info("\tTracing variant of Tasks RCU enabled.\n"); | |
606 | #endif /* #ifdef CONFIG_TASKS_TRACE_RCU */ | |
eacd6f04 PM |
607 | } |
608 | ||
609 | #endif /* #ifndef CONFIG_TINY_RCU */ | |
5873b8a9 | 610 | |
8344496e | 611 | #ifndef CONFIG_TINY_RCU |
e21408ce PM |
612 | /* Dump out rcutorture-relevant state common to all RCU-tasks flavors. */ |
613 | static void show_rcu_tasks_generic_gp_kthread(struct rcu_tasks *rtp, char *s) | |
614 | { | |
10b3742f PM |
615 | int cpu; |
616 | bool havecbs = false; | |
617 | ||
618 | for_each_possible_cpu(cpu) { | |
619 | struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu); | |
620 | ||
621 | if (!data_race(rcu_segcblist_empty(&rtpcp->cblist))) { | |
622 | havecbs = true; | |
623 | break; | |
624 | } | |
625 | } | |
7e0669c3 | 626 | pr_info("%s: %s(%d) since %lu g:%lu i:%lu/%lu %c%c %s\n", |
e21408ce | 627 | rtp->kname, |
7e0669c3 | 628 | tasks_gp_state_getname(rtp), data_race(rtp->gp_state), |
af051ca4 | 629 | jiffies - data_race(rtp->gp_jiffies), |
b14fb4fb | 630 | data_race(rcu_seq_current(&rtp->tasks_gp_seq)), |
7e0669c3 | 631 | data_race(rtp->n_ipis_fails), data_race(rtp->n_ipis), |
e21408ce | 632 | ".k"[!!data_race(rtp->kthread_ptr)], |
10b3742f | 633 | ".C"[havecbs], |
e21408ce PM |
634 | s); |
635 | } | |
27c0f144 | 636 | #endif // #ifndef CONFIG_TINY_RCU |
e21408ce | 637 | |
25246fc8 PM |
638 | static void exit_tasks_rcu_finish_trace(struct task_struct *t); |
639 | ||
640 | #if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU) | |
5873b8a9 | 641 | |
d01aa263 PM |
642 | //////////////////////////////////////////////////////////////////////// |
643 | // | |
644 | // Shared code between task-list-scanning variants of Tasks RCU. | |
645 | ||
646 | /* Wait for one RCU-tasks grace period. */ | |
647 | static void rcu_tasks_wait_gp(struct rcu_tasks *rtp) | |
648 | { | |
f2539003 | 649 | struct task_struct *g; |
d01aa263 | 650 | int fract; |
f2539003 PM |
651 | LIST_HEAD(holdouts); |
652 | unsigned long j; | |
653 | unsigned long lastinfo; | |
654 | unsigned long lastreport; | |
655 | bool reported = false; | |
656 | int rtsi; | |
657 | struct task_struct *t; | |
d01aa263 | 658 | |
af051ca4 | 659 | set_tasks_gp_state(rtp, RTGS_PRE_WAIT_GP); |
d01aa263 PM |
660 | rtp->pregp_func(); |
661 | ||
662 | /* | |
663 | * There were callbacks, so we need to wait for an RCU-tasks | |
664 | * grace period. Start off by scanning the task list for tasks | |
665 | * that are not already voluntarily blocked. Mark these tasks | |
666 | * and make a list of them in holdouts. | |
667 | */ | |
af051ca4 | 668 | set_tasks_gp_state(rtp, RTGS_SCAN_TASKLIST); |
d01aa263 PM |
669 | rcu_read_lock(); |
670 | for_each_process_thread(g, t) | |
671 | rtp->pertask_func(t, &holdouts); | |
672 | rcu_read_unlock(); | |
673 | ||
af051ca4 | 674 | set_tasks_gp_state(rtp, RTGS_POST_SCAN_TASKLIST); |
9796e1ae | 675 | rtp->postscan_func(&holdouts); |
d01aa263 PM |
676 | |
677 | /* | |
678 | * Each pass through the following loop scans the list of holdout | |
679 | * tasks, removing any that are no longer holdouts. When the list | |
680 | * is empty, we are done. | |
681 | */ | |
682 | lastreport = jiffies; | |
f2539003 PM |
683 | lastinfo = lastreport; |
684 | rtsi = READ_ONCE(rcu_task_stall_info); | |
d01aa263 | 685 | |
2393a613 PM |
686 | // Start off with initial wait and slowly back off to 1 HZ wait. |
687 | fract = rtp->init_fract; | |
d01aa263 | 688 | |
77dc1741 | 689 | while (!list_empty(&holdouts)) { |
777570d9 | 690 | ktime_t exp; |
d01aa263 PM |
691 | bool firstreport; |
692 | bool needreport; | |
693 | int rtst; | |
694 | ||
f2539003 | 695 | // Slowly back off waiting for holdouts |
af051ca4 | 696 | set_tasks_gp_state(rtp, RTGS_WAIT_SCAN_HOLDOUTS); |
bddf7122 PM |
697 | if (!IS_ENABLED(CONFIG_PREEMPT_RT)) { |
698 | schedule_timeout_idle(fract); | |
699 | } else { | |
700 | exp = jiffies_to_nsecs(fract); | |
701 | __set_current_state(TASK_IDLE); | |
702 | schedule_hrtimeout_range(&exp, jiffies_to_nsecs(HZ / 2), HRTIMER_MODE_REL_HARD); | |
703 | } | |
d01aa263 | 704 | |
75dc2da5 PM |
705 | if (fract < HZ) |
706 | fract++; | |
d01aa263 PM |
707 | |
708 | rtst = READ_ONCE(rcu_task_stall_timeout); | |
709 | needreport = rtst > 0 && time_after(jiffies, lastreport + rtst); | |
f2539003 | 710 | if (needreport) { |
d01aa263 | 711 | lastreport = jiffies; |
f2539003 PM |
712 | reported = true; |
713 | } | |
d01aa263 PM |
714 | firstreport = true; |
715 | WARN_ON(signal_pending(current)); | |
af051ca4 | 716 | set_tasks_gp_state(rtp, RTGS_SCAN_HOLDOUTS); |
d01aa263 | 717 | rtp->holdouts_func(&holdouts, needreport, &firstreport); |
f2539003 PM |
718 | |
719 | // Print pre-stall informational messages if needed. | |
720 | j = jiffies; | |
721 | if (rtsi > 0 && !reported && time_after(j, lastinfo + rtsi)) { | |
722 | lastinfo = j; | |
723 | rtsi = rtsi * rcu_task_stall_info_mult; | |
724 | pr_info("%s: %s grace period %lu is %lu jiffies old.\n", | |
725 | __func__, rtp->kname, rtp->tasks_gp_seq, j - rtp->gp_start); | |
726 | } | |
d01aa263 PM |
727 | } |
728 | ||
af051ca4 PM |
729 | set_tasks_gp_state(rtp, RTGS_POST_GP); |
730 | rtp->postgp_func(rtp); | |
d01aa263 PM |
731 | } |
732 | ||
25246fc8 PM |
733 | #endif /* #if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU) */ |
734 | ||
735 | #ifdef CONFIG_TASKS_RCU | |
736 | ||
5873b8a9 PM |
737 | //////////////////////////////////////////////////////////////////////// |
738 | // | |
739 | // Simple variant of RCU whose quiescent states are voluntary context | |
8af9e2c7 | 740 | // switch, cond_resched_tasks_rcu_qs(), user-space execution, and idle. |
5873b8a9 PM |
741 | // As such, grace periods can take one good long time. There are no |
742 | // read-side primitives similar to rcu_read_lock() and rcu_read_unlock() | |
743 | // because this implementation is intended to get the system into a safe | |
744 | // state for some of the manipulations involved in tracing and the like. | |
745 | // Finally, this implementation does not support high call_rcu_tasks() | |
746 | // rates from multiple CPUs. If this is required, per-CPU callback lists | |
747 | // will be needed. | |
06a3ec92 PM |
748 | // |
749 | // The implementation uses rcu_tasks_wait_gp(), which relies on function | |
750 | // pointers in the rcu_tasks structure. The rcu_spawn_tasks_kthread() | |
751 | // function sets these function pointers up so that rcu_tasks_wait_gp() | |
752 | // invokes these functions in this order: | |
753 | // | |
754 | // rcu_tasks_pregp_step(): | |
755 | // Invokes synchronize_rcu() in order to wait for all in-flight | |
756 | // t->on_rq and t->nvcsw transitions to complete. This works because | |
757 | // all such transitions are carried out with interrupts disabled. | |
758 | // rcu_tasks_pertask(), invoked on every non-idle task: | |
759 | // For every runnable non-idle task other than the current one, use | |
760 | // get_task_struct() to pin down that task, snapshot that task's | |
761 | // number of voluntary context switches, and add that task to the | |
762 | // holdout list. | |
763 | // rcu_tasks_postscan(): | |
764 | // Invoke synchronize_srcu() to ensure that all tasks that were | |
765 | // in the process of exiting (and which thus might not know to | |
766 | // synchronize with this RCU Tasks grace period) have completed | |
767 | // exiting. | |
768 | // check_all_holdout_tasks(), repeatedly until holdout list is empty: | |
769 | // Scans the holdout list, attempting to identify a quiescent state | |
770 | // for each task on the list. If there is a quiescent state, the | |
771 | // corresponding task is removed from the holdout list. | |
772 | // rcu_tasks_postgp(): | |
773 | // Invokes synchronize_rcu() in order to ensure that all prior | |
774 | // t->on_rq and t->nvcsw transitions are seen by all CPUs and tasks | |
775 | // to have happened before the end of this RCU Tasks grace period. | |
776 | // Again, this works because all such transitions are carried out | |
777 | // with interrupts disabled. | |
778 | // | |
779 | // For each exiting task, the exit_tasks_rcu_start() and | |
780 | // exit_tasks_rcu_finish() functions begin and end, respectively, the SRCU | |
781 | // read-side critical sections waited for by rcu_tasks_postscan(). | |
782 | // | |
381a4f3b PM |
783 | // Pre-grace-period update-side code is ordered before the grace |
784 | // via the raw_spin_lock.*rcu_node(). Pre-grace-period read-side code | |
785 | // is ordered before the grace period via synchronize_rcu() call in | |
786 | // rcu_tasks_pregp_step() and by the scheduler's locks and interrupt | |
06a3ec92 | 787 | // disabling. |
5873b8a9 | 788 | |
e4fe5dd6 PM |
789 | /* Pre-grace-period preparation. */ |
790 | static void rcu_tasks_pregp_step(void) | |
791 | { | |
792 | /* | |
793 | * Wait for all pre-existing t->on_rq and t->nvcsw transitions | |
794 | * to complete. Invoking synchronize_rcu() suffices because all | |
795 | * these transitions occur with interrupts disabled. Without this | |
796 | * synchronize_rcu(), a read-side critical section that started | |
797 | * before the grace period might be incorrectly seen as having | |
798 | * started after the grace period. | |
799 | * | |
800 | * This synchronize_rcu() also dispenses with the need for a | |
801 | * memory barrier on the first store to t->rcu_tasks_holdout, | |
802 | * as it forces the store to happen after the beginning of the | |
803 | * grace period. | |
804 | */ | |
805 | synchronize_rcu(); | |
806 | } | |
807 | ||
808 | /* Per-task initial processing. */ | |
809 | static void rcu_tasks_pertask(struct task_struct *t, struct list_head *hop) | |
810 | { | |
811 | if (t != current && READ_ONCE(t->on_rq) && !is_idle_task(t)) { | |
812 | get_task_struct(t); | |
813 | t->rcu_tasks_nvcsw = READ_ONCE(t->nvcsw); | |
814 | WRITE_ONCE(t->rcu_tasks_holdout, true); | |
815 | list_add(&t->rcu_tasks_holdout_list, hop); | |
816 | } | |
817 | } | |
818 | ||
819 | /* Processing between scanning taskslist and draining the holdout list. */ | |
04a3c5aa | 820 | static void rcu_tasks_postscan(struct list_head *hop) |
e4fe5dd6 PM |
821 | { |
822 | /* | |
823 | * Wait for tasks that are in the process of exiting. This | |
824 | * does only part of the job, ensuring that all tasks that were | |
825 | * previously exiting reach the point where they have disabled | |
826 | * preemption, allowing the later synchronize_rcu() to finish | |
827 | * the job. | |
828 | */ | |
829 | synchronize_srcu(&tasks_rcu_exit_srcu); | |
830 | } | |
831 | ||
5873b8a9 PM |
832 | /* See if tasks are still holding out, complain if so. */ |
833 | static void check_holdout_task(struct task_struct *t, | |
834 | bool needreport, bool *firstreport) | |
835 | { | |
836 | int cpu; | |
837 | ||
838 | if (!READ_ONCE(t->rcu_tasks_holdout) || | |
839 | t->rcu_tasks_nvcsw != READ_ONCE(t->nvcsw) || | |
840 | !READ_ONCE(t->on_rq) || | |
841 | (IS_ENABLED(CONFIG_NO_HZ_FULL) && | |
842 | !is_idle_task(t) && t->rcu_tasks_idle_cpu >= 0)) { | |
843 | WRITE_ONCE(t->rcu_tasks_holdout, false); | |
844 | list_del_init(&t->rcu_tasks_holdout_list); | |
845 | put_task_struct(t); | |
846 | return; | |
847 | } | |
848 | rcu_request_urgent_qs_task(t); | |
849 | if (!needreport) | |
850 | return; | |
851 | if (*firstreport) { | |
852 | pr_err("INFO: rcu_tasks detected stalls on tasks:\n"); | |
853 | *firstreport = false; | |
854 | } | |
855 | cpu = task_cpu(t); | |
856 | pr_alert("%p: %c%c nvcsw: %lu/%lu holdout: %d idle_cpu: %d/%d\n", | |
857 | t, ".I"[is_idle_task(t)], | |
858 | "N."[cpu < 0 || !tick_nohz_full_cpu(cpu)], | |
859 | t->rcu_tasks_nvcsw, t->nvcsw, t->rcu_tasks_holdout, | |
860 | t->rcu_tasks_idle_cpu, cpu); | |
861 | sched_show_task(t); | |
862 | } | |
863 | ||
e4fe5dd6 PM |
864 | /* Scan the holdout lists for tasks no longer holding out. */ |
865 | static void check_all_holdout_tasks(struct list_head *hop, | |
866 | bool needreport, bool *firstreport) | |
867 | { | |
868 | struct task_struct *t, *t1; | |
869 | ||
870 | list_for_each_entry_safe(t, t1, hop, rcu_tasks_holdout_list) { | |
871 | check_holdout_task(t, needreport, firstreport); | |
872 | cond_resched(); | |
873 | } | |
874 | } | |
875 | ||
876 | /* Finish off the Tasks-RCU grace period. */ | |
af051ca4 | 877 | static void rcu_tasks_postgp(struct rcu_tasks *rtp) |
e4fe5dd6 PM |
878 | { |
879 | /* | |
880 | * Because ->on_rq and ->nvcsw are not guaranteed to have a full | |
881 | * memory barriers prior to them in the schedule() path, memory | |
882 | * reordering on other CPUs could cause their RCU-tasks read-side | |
883 | * critical sections to extend past the end of the grace period. | |
884 | * However, because these ->nvcsw updates are carried out with | |
885 | * interrupts disabled, we can use synchronize_rcu() to force the | |
886 | * needed ordering on all such CPUs. | |
887 | * | |
888 | * This synchronize_rcu() also confines all ->rcu_tasks_holdout | |
889 | * accesses to be within the grace period, avoiding the need for | |
890 | * memory barriers for ->rcu_tasks_holdout accesses. | |
891 | * | |
892 | * In addition, this synchronize_rcu() waits for exiting tasks | |
893 | * to complete their final preempt_disable() region of execution, | |
894 | * cleaning up after the synchronize_srcu() above. | |
895 | */ | |
896 | synchronize_rcu(); | |
897 | } | |
898 | ||
5873b8a9 | 899 | void call_rcu_tasks(struct rcu_head *rhp, rcu_callback_t func); |
c97d12a6 | 900 | DEFINE_RCU_TASKS(rcu_tasks, rcu_tasks_wait_gp, call_rcu_tasks, "RCU Tasks"); |
5873b8a9 PM |
901 | |
902 | /** | |
903 | * call_rcu_tasks() - Queue an RCU for invocation task-based grace period | |
904 | * @rhp: structure to be used for queueing the RCU updates. | |
905 | * @func: actual callback function to be invoked after the grace period | |
906 | * | |
907 | * The callback function will be invoked some time after a full grace | |
908 | * period elapses, in other words after all currently executing RCU | |
909 | * read-side critical sections have completed. call_rcu_tasks() assumes | |
910 | * that the read-side critical sections end at a voluntary context | |
8af9e2c7 | 911 | * switch (not a preemption!), cond_resched_tasks_rcu_qs(), entry into idle, |
5873b8a9 PM |
912 | * or transition to usermode execution. As such, there are no read-side |
913 | * primitives analogous to rcu_read_lock() and rcu_read_unlock() because | |
914 | * this primitive is intended to determine that all tasks have passed | |
a616aec9 | 915 | * through a safe state, not so much for data-structure synchronization. |
5873b8a9 PM |
916 | * |
917 | * See the description of call_rcu() for more detailed information on | |
918 | * memory ordering guarantees. | |
919 | */ | |
920 | void call_rcu_tasks(struct rcu_head *rhp, rcu_callback_t func) | |
921 | { | |
922 | call_rcu_tasks_generic(rhp, func, &rcu_tasks); | |
923 | } | |
924 | EXPORT_SYMBOL_GPL(call_rcu_tasks); | |
925 | ||
926 | /** | |
927 | * synchronize_rcu_tasks - wait until an rcu-tasks grace period has elapsed. | |
928 | * | |
929 | * Control will return to the caller some time after a full rcu-tasks | |
930 | * grace period has elapsed, in other words after all currently | |
931 | * executing rcu-tasks read-side critical sections have elapsed. These | |
932 | * read-side critical sections are delimited by calls to schedule(), | |
933 | * cond_resched_tasks_rcu_qs(), idle execution, userspace execution, calls | |
934 | * to synchronize_rcu_tasks(), and (in theory, anyway) cond_resched(). | |
935 | * | |
936 | * This is a very specialized primitive, intended only for a few uses in | |
937 | * tracing and other situations requiring manipulation of function | |
938 | * preambles and profiling hooks. The synchronize_rcu_tasks() function | |
939 | * is not (yet) intended for heavy use from multiple CPUs. | |
940 | * | |
941 | * See the description of synchronize_rcu() for more detailed information | |
942 | * on memory ordering guarantees. | |
943 | */ | |
944 | void synchronize_rcu_tasks(void) | |
945 | { | |
946 | synchronize_rcu_tasks_generic(&rcu_tasks); | |
947 | } | |
948 | EXPORT_SYMBOL_GPL(synchronize_rcu_tasks); | |
949 | ||
950 | /** | |
951 | * rcu_barrier_tasks - Wait for in-flight call_rcu_tasks() callbacks. | |
952 | * | |
953 | * Although the current implementation is guaranteed to wait, it is not | |
954 | * obligated to, for example, if there are no pending callbacks. | |
955 | */ | |
956 | void rcu_barrier_tasks(void) | |
957 | { | |
ce9b1c66 | 958 | rcu_barrier_tasks_generic(&rcu_tasks); |
5873b8a9 PM |
959 | } |
960 | EXPORT_SYMBOL_GPL(rcu_barrier_tasks); | |
961 | ||
962 | static int __init rcu_spawn_tasks_kthread(void) | |
963 | { | |
cafafd67 | 964 | cblist_init_generic(&rcu_tasks); |
4fe192df | 965 | rcu_tasks.gp_sleep = HZ / 10; |
75dc2da5 | 966 | rcu_tasks.init_fract = HZ / 10; |
e4fe5dd6 PM |
967 | rcu_tasks.pregp_func = rcu_tasks_pregp_step; |
968 | rcu_tasks.pertask_func = rcu_tasks_pertask; | |
969 | rcu_tasks.postscan_func = rcu_tasks_postscan; | |
970 | rcu_tasks.holdouts_func = check_all_holdout_tasks; | |
971 | rcu_tasks.postgp_func = rcu_tasks_postgp; | |
5873b8a9 PM |
972 | rcu_spawn_tasks_kthread_generic(&rcu_tasks); |
973 | return 0; | |
974 | } | |
5873b8a9 | 975 | |
27c0f144 PM |
976 | #if !defined(CONFIG_TINY_RCU) |
977 | void show_rcu_tasks_classic_gp_kthread(void) | |
e21408ce PM |
978 | { |
979 | show_rcu_tasks_generic_gp_kthread(&rcu_tasks, ""); | |
980 | } | |
27c0f144 PM |
981 | EXPORT_SYMBOL_GPL(show_rcu_tasks_classic_gp_kthread); |
982 | #endif // !defined(CONFIG_TINY_RCU) | |
e21408ce | 983 | |
25246fc8 PM |
984 | /* Do the srcu_read_lock() for the above synchronize_srcu(). */ |
985 | void exit_tasks_rcu_start(void) __acquires(&tasks_rcu_exit_srcu) | |
986 | { | |
987 | preempt_disable(); | |
988 | current->rcu_tasks_idx = __srcu_read_lock(&tasks_rcu_exit_srcu); | |
989 | preempt_enable(); | |
990 | } | |
991 | ||
992 | /* Do the srcu_read_unlock() for the above synchronize_srcu(). */ | |
993 | void exit_tasks_rcu_finish(void) __releases(&tasks_rcu_exit_srcu) | |
994 | { | |
995 | struct task_struct *t = current; | |
996 | ||
997 | preempt_disable(); | |
998 | __srcu_read_unlock(&tasks_rcu_exit_srcu, t->rcu_tasks_idx); | |
999 | preempt_enable(); | |
1000 | exit_tasks_rcu_finish_trace(t); | |
1001 | } | |
1002 | ||
e21408ce | 1003 | #else /* #ifdef CONFIG_TASKS_RCU */ |
25246fc8 PM |
1004 | void exit_tasks_rcu_start(void) { } |
1005 | void exit_tasks_rcu_finish(void) { exit_tasks_rcu_finish_trace(current); } | |
e21408ce | 1006 | #endif /* #else #ifdef CONFIG_TASKS_RCU */ |
c84aad76 PM |
1007 | |
1008 | #ifdef CONFIG_TASKS_RUDE_RCU | |
1009 | ||
1010 | //////////////////////////////////////////////////////////////////////// | |
1011 | // | |
1012 | // "Rude" variant of Tasks RCU, inspired by Steve Rostedt's trick of | |
1013 | // passing an empty function to schedule_on_each_cpu(). This approach | |
e4be1f44 PM |
1014 | // provides an asynchronous call_rcu_tasks_rude() API and batching of |
1015 | // concurrent calls to the synchronous synchronize_rcu_tasks_rude() API. | |
9fc98e31 PM |
1016 | // This invokes schedule_on_each_cpu() in order to send IPIs far and wide |
1017 | // and induces otherwise unnecessary context switches on all online CPUs, | |
1018 | // whether idle or not. | |
1019 | // | |
1020 | // Callback handling is provided by the rcu_tasks_kthread() function. | |
1021 | // | |
1022 | // Ordering is provided by the scheduler's context-switch code. | |
c84aad76 PM |
1023 | |
1024 | // Empty function to allow workqueues to force a context switch. | |
1025 | static void rcu_tasks_be_rude(struct work_struct *work) | |
1026 | { | |
1027 | } | |
1028 | ||
1029 | // Wait for one rude RCU-tasks grace period. | |
1030 | static void rcu_tasks_rude_wait_gp(struct rcu_tasks *rtp) | |
1031 | { | |
f75fd4b9 PS |
1032 | if (num_online_cpus() <= 1) |
1033 | return; // Fastpath for only one CPU. | |
1034 | ||
238dbce3 | 1035 | rtp->n_ipis += cpumask_weight(cpu_online_mask); |
c84aad76 PM |
1036 | schedule_on_each_cpu(rcu_tasks_be_rude); |
1037 | } | |
1038 | ||
1039 | void call_rcu_tasks_rude(struct rcu_head *rhp, rcu_callback_t func); | |
c97d12a6 PM |
1040 | DEFINE_RCU_TASKS(rcu_tasks_rude, rcu_tasks_rude_wait_gp, call_rcu_tasks_rude, |
1041 | "RCU Tasks Rude"); | |
c84aad76 PM |
1042 | |
1043 | /** | |
1044 | * call_rcu_tasks_rude() - Queue a callback rude task-based grace period | |
1045 | * @rhp: structure to be used for queueing the RCU updates. | |
1046 | * @func: actual callback function to be invoked after the grace period | |
1047 | * | |
1048 | * The callback function will be invoked some time after a full grace | |
1049 | * period elapses, in other words after all currently executing RCU | |
1050 | * read-side critical sections have completed. call_rcu_tasks_rude() | |
1051 | * assumes that the read-side critical sections end at context switch, | |
8af9e2c7 | 1052 | * cond_resched_tasks_rcu_qs(), or transition to usermode execution (as |
a6517e9c NU |
1053 | * usermode execution is schedulable). As such, there are no read-side |
1054 | * primitives analogous to rcu_read_lock() and rcu_read_unlock() because | |
1055 | * this primitive is intended to determine that all tasks have passed | |
1056 | * through a safe state, not so much for data-structure synchronization. | |
c84aad76 PM |
1057 | * |
1058 | * See the description of call_rcu() for more detailed information on | |
1059 | * memory ordering guarantees. | |
1060 | */ | |
1061 | void call_rcu_tasks_rude(struct rcu_head *rhp, rcu_callback_t func) | |
1062 | { | |
1063 | call_rcu_tasks_generic(rhp, func, &rcu_tasks_rude); | |
1064 | } | |
1065 | EXPORT_SYMBOL_GPL(call_rcu_tasks_rude); | |
1066 | ||
1067 | /** | |
1068 | * synchronize_rcu_tasks_rude - wait for a rude rcu-tasks grace period | |
1069 | * | |
1070 | * Control will return to the caller some time after a rude rcu-tasks | |
1071 | * grace period has elapsed, in other words after all currently | |
1072 | * executing rcu-tasks read-side critical sections have elapsed. These | |
1073 | * read-side critical sections are delimited by calls to schedule(), | |
a6517e9c NU |
1074 | * cond_resched_tasks_rcu_qs(), userspace execution (which is a schedulable |
1075 | * context), and (in theory, anyway) cond_resched(). | |
c84aad76 PM |
1076 | * |
1077 | * This is a very specialized primitive, intended only for a few uses in | |
1078 | * tracing and other situations requiring manipulation of function preambles | |
1079 | * and profiling hooks. The synchronize_rcu_tasks_rude() function is not | |
1080 | * (yet) intended for heavy use from multiple CPUs. | |
1081 | * | |
1082 | * See the description of synchronize_rcu() for more detailed information | |
1083 | * on memory ordering guarantees. | |
1084 | */ | |
1085 | void synchronize_rcu_tasks_rude(void) | |
1086 | { | |
1087 | synchronize_rcu_tasks_generic(&rcu_tasks_rude); | |
1088 | } | |
1089 | EXPORT_SYMBOL_GPL(synchronize_rcu_tasks_rude); | |
1090 | ||
1091 | /** | |
1092 | * rcu_barrier_tasks_rude - Wait for in-flight call_rcu_tasks_rude() callbacks. | |
1093 | * | |
1094 | * Although the current implementation is guaranteed to wait, it is not | |
1095 | * obligated to, for example, if there are no pending callbacks. | |
1096 | */ | |
1097 | void rcu_barrier_tasks_rude(void) | |
1098 | { | |
ce9b1c66 | 1099 | rcu_barrier_tasks_generic(&rcu_tasks_rude); |
c84aad76 PM |
1100 | } |
1101 | EXPORT_SYMBOL_GPL(rcu_barrier_tasks_rude); | |
1102 | ||
1103 | static int __init rcu_spawn_tasks_rude_kthread(void) | |
1104 | { | |
cafafd67 | 1105 | cblist_init_generic(&rcu_tasks_rude); |
4fe192df | 1106 | rcu_tasks_rude.gp_sleep = HZ / 10; |
c84aad76 PM |
1107 | rcu_spawn_tasks_kthread_generic(&rcu_tasks_rude); |
1108 | return 0; | |
1109 | } | |
c84aad76 | 1110 | |
27c0f144 PM |
1111 | #if !defined(CONFIG_TINY_RCU) |
1112 | void show_rcu_tasks_rude_gp_kthread(void) | |
e21408ce PM |
1113 | { |
1114 | show_rcu_tasks_generic_gp_kthread(&rcu_tasks_rude, ""); | |
1115 | } | |
27c0f144 PM |
1116 | EXPORT_SYMBOL_GPL(show_rcu_tasks_rude_gp_kthread); |
1117 | #endif // !defined(CONFIG_TINY_RCU) | |
1118 | #endif /* #ifdef CONFIG_TASKS_RUDE_RCU */ | |
d5f177d3 PM |
1119 | |
1120 | //////////////////////////////////////////////////////////////////////// | |
1121 | // | |
1122 | // Tracing variant of Tasks RCU. This variant is designed to be used | |
1123 | // to protect tracing hooks, including those of BPF. This variant | |
1124 | // therefore: | |
1125 | // | |
1126 | // 1. Has explicit read-side markers to allow finite grace periods | |
1127 | // in the face of in-kernel loops for PREEMPT=n builds. | |
1128 | // | |
1129 | // 2. Protects code in the idle loop, exception entry/exit, and | |
1130 | // CPU-hotplug code paths, similar to the capabilities of SRCU. | |
1131 | // | |
c4f113ac | 1132 | // 3. Avoids expensive read-side instructions, having overhead similar |
d5f177d3 PM |
1133 | // to that of Preemptible RCU. |
1134 | // | |
1135 | // There are of course downsides. The grace-period code can send IPIs to | |
1136 | // CPUs, even when those CPUs are in the idle loop or in nohz_full userspace. | |
1137 | // It is necessary to scan the full tasklist, much as for Tasks RCU. There | |
1138 | // is a single callback queue guarded by a single lock, again, much as for | |
1139 | // Tasks RCU. If needed, these downsides can be at least partially remedied. | |
1140 | // | |
1141 | // Perhaps most important, this variant of RCU does not affect the vanilla | |
1142 | // flavors, rcu_preempt and rcu_sched. The fact that RCU Tasks Trace | |
1143 | // readers can operate from idle, offline, and exception entry/exit in no | |
1144 | // way allows rcu_preempt and rcu_sched readers to also do so. | |
a434dd10 PM |
1145 | // |
1146 | // The implementation uses rcu_tasks_wait_gp(), which relies on function | |
1147 | // pointers in the rcu_tasks structure. The rcu_spawn_tasks_trace_kthread() | |
1148 | // function sets these function pointers up so that rcu_tasks_wait_gp() | |
1149 | // invokes these functions in this order: | |
1150 | // | |
1151 | // rcu_tasks_trace_pregp_step(): | |
1152 | // Initialize the count of readers and block CPU-hotplug operations. | |
1153 | // rcu_tasks_trace_pertask(), invoked on every non-idle task: | |
1154 | // Initialize per-task state and attempt to identify an immediate | |
1155 | // quiescent state for that task, or, failing that, attempt to | |
1156 | // set that task's .need_qs flag so that task's next outermost | |
1157 | // rcu_read_unlock_trace() will report the quiescent state (in which | |
1158 | // case the count of readers is incremented). If both attempts fail, | |
45f4b4a2 PM |
1159 | // the task is added to a "holdout" list. Note that IPIs are used |
1160 | // to invoke trc_read_check_handler() in the context of running tasks | |
1161 | // in order to avoid ordering overhead on common-case shared-variable | |
1162 | // accessses. | |
a434dd10 PM |
1163 | // rcu_tasks_trace_postscan(): |
1164 | // Initialize state and attempt to identify an immediate quiescent | |
1165 | // state as above (but only for idle tasks), unblock CPU-hotplug | |
1166 | // operations, and wait for an RCU grace period to avoid races with | |
1167 | // tasks that are in the process of exiting. | |
1168 | // check_all_holdout_tasks_trace(), repeatedly until holdout list is empty: | |
1169 | // Scans the holdout list, attempting to identify a quiescent state | |
1170 | // for each task on the list. If there is a quiescent state, the | |
1171 | // corresponding task is removed from the holdout list. | |
1172 | // rcu_tasks_trace_postgp(): | |
1173 | // Wait for the count of readers do drop to zero, reporting any stalls. | |
1174 | // Also execute full memory barriers to maintain ordering with code | |
1175 | // executing after the grace period. | |
1176 | // | |
1177 | // The exit_tasks_rcu_finish_trace() synchronizes with exiting tasks. | |
1178 | // | |
1179 | // Pre-grace-period update-side code is ordered before the grace | |
1180 | // period via the ->cbs_lock and barriers in rcu_tasks_kthread(). | |
1181 | // Pre-grace-period read-side code is ordered before the grace period by | |
1182 | // atomic_dec_and_test() of the count of readers (for IPIed readers) and by | |
1183 | // scheduler context-switch ordering (for locked-down non-running readers). | |
d5f177d3 PM |
1184 | |
1185 | // The lockdep state must be outside of #ifdef to be useful. | |
1186 | #ifdef CONFIG_DEBUG_LOCK_ALLOC | |
1187 | static struct lock_class_key rcu_lock_trace_key; | |
1188 | struct lockdep_map rcu_trace_lock_map = | |
1189 | STATIC_LOCKDEP_MAP_INIT("rcu_read_lock_trace", &rcu_lock_trace_key); | |
1190 | EXPORT_SYMBOL_GPL(rcu_trace_lock_map); | |
1191 | #endif /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */ | |
1192 | ||
1193 | #ifdef CONFIG_TASKS_TRACE_RCU | |
1194 | ||
d5f177d3 PM |
1195 | // Record outstanding IPIs to each CPU. No point in sending two... |
1196 | static DEFINE_PER_CPU(bool, trc_ipi_to_cpu); | |
1197 | ||
40471509 PM |
1198 | // The number of detections of task quiescent state relying on |
1199 | // heavyweight readers executing explicit memory barriers. | |
6731da9e PM |
1200 | static unsigned long n_heavy_reader_attempts; |
1201 | static unsigned long n_heavy_reader_updates; | |
1202 | static unsigned long n_heavy_reader_ofl_updates; | |
40471509 | 1203 | |
b0afa0f0 PM |
1204 | void call_rcu_tasks_trace(struct rcu_head *rhp, rcu_callback_t func); |
1205 | DEFINE_RCU_TASKS(rcu_tasks_trace, rcu_tasks_wait_gp, call_rcu_tasks_trace, | |
1206 | "RCU Tasks Trace"); | |
1207 | ||
3847b645 PM |
1208 | /* Load from ->trc_reader_special.b.need_qs with proper ordering. */ |
1209 | static u8 rcu_ld_need_qs(struct task_struct *t) | |
1210 | { | |
1211 | smp_mb(); // Enforce full grace-period ordering. | |
1212 | return smp_load_acquire(&t->trc_reader_special.b.need_qs); | |
1213 | } | |
1214 | ||
1215 | /* Store to ->trc_reader_special.b.need_qs with proper ordering. */ | |
1216 | static void rcu_st_need_qs(struct task_struct *t, u8 v) | |
1217 | { | |
1218 | smp_store_release(&t->trc_reader_special.b.need_qs, v); | |
1219 | smp_mb(); // Enforce full grace-period ordering. | |
1220 | } | |
1221 | ||
1222 | /* | |
1223 | * Do a cmpxchg() on ->trc_reader_special.b.need_qs, allowing for | |
1224 | * the four-byte operand-size restriction of some platforms. | |
1225 | * Returns the old value, which is often ignored. | |
1226 | */ | |
1227 | u8 rcu_trc_cmpxchg_need_qs(struct task_struct *t, u8 old, u8 new) | |
1228 | { | |
1229 | union rcu_special ret; | |
1230 | union rcu_special trs_old = READ_ONCE(t->trc_reader_special); | |
1231 | union rcu_special trs_new = trs_old; | |
1232 | ||
1233 | if (trs_old.b.need_qs != old) | |
1234 | return trs_old.b.need_qs; | |
1235 | trs_new.b.need_qs = new; | |
1236 | ret.s = cmpxchg(&t->trc_reader_special.s, trs_old.s, trs_new.s); | |
1237 | return ret.b.need_qs; | |
1238 | } | |
1239 | EXPORT_SYMBOL_GPL(rcu_trc_cmpxchg_need_qs); | |
1240 | ||
d5f177d3 | 1241 | /* If we are the last reader, wake up the grace-period kthread. */ |
a5c071cc | 1242 | void rcu_read_unlock_trace_special(struct task_struct *t) |
d5f177d3 | 1243 | { |
3847b645 | 1244 | int nqs = (rcu_ld_need_qs(t) == (TRC_NEED_QS_CHECKED | TRC_NEED_QS)); |
276c4104 | 1245 | |
3847b645 | 1246 | if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB) && t->trc_reader_special.b.need_mb) |
276c4104 PM |
1247 | smp_mb(); // Pairs with update-side barriers. |
1248 | // Update .need_qs before ->trc_reader_nesting for irq/NMI handlers. | |
3847b645 PM |
1249 | if (nqs) { |
1250 | u8 result = rcu_trc_cmpxchg_need_qs(t, TRC_NEED_QS_CHECKED | TRC_NEED_QS, | |
1251 | TRC_NEED_QS_CHECKED); | |
1252 | ||
1253 | WARN_ONCE(result != (TRC_NEED_QS_CHECKED | TRC_NEED_QS), | |
1254 | "%s: result = %d", __func__, result); | |
1255 | } | |
a5c071cc | 1256 | WRITE_ONCE(t->trc_reader_nesting, 0); |
d5f177d3 PM |
1257 | } |
1258 | EXPORT_SYMBOL_GPL(rcu_read_unlock_trace_special); | |
1259 | ||
1260 | /* Add a task to the holdout list, if it is not already on the list. */ | |
1261 | static void trc_add_holdout(struct task_struct *t, struct list_head *bhp) | |
1262 | { | |
1263 | if (list_empty(&t->trc_holdout_list)) { | |
1264 | get_task_struct(t); | |
1265 | list_add(&t->trc_holdout_list, bhp); | |
1266 | } | |
1267 | } | |
1268 | ||
1269 | /* Remove a task from the holdout list, if it is in fact present. */ | |
1270 | static void trc_del_holdout(struct task_struct *t) | |
1271 | { | |
1272 | if (!list_empty(&t->trc_holdout_list)) { | |
1273 | list_del_init(&t->trc_holdout_list); | |
1274 | put_task_struct(t); | |
1275 | } | |
1276 | } | |
1277 | ||
1278 | /* IPI handler to check task state. */ | |
1279 | static void trc_read_check_handler(void *t_in) | |
1280 | { | |
9ff86b4c | 1281 | int nesting; |
d5f177d3 PM |
1282 | struct task_struct *t = current; |
1283 | struct task_struct *texp = t_in; | |
1284 | ||
1285 | // If the task is no longer running on this CPU, leave. | |
3847b645 | 1286 | if (unlikely(texp != t)) |
d5f177d3 | 1287 | goto reset_ipi; // Already on holdout list, so will check later. |
d5f177d3 PM |
1288 | |
1289 | // If the task is not in a read-side critical section, and | |
1290 | // if this is the last reader, awaken the grace-period kthread. | |
9ff86b4c PM |
1291 | nesting = READ_ONCE(t->trc_reader_nesting); |
1292 | if (likely(!nesting)) { | |
3847b645 | 1293 | rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS_CHECKED); |
d5f177d3 PM |
1294 | goto reset_ipi; |
1295 | } | |
ba3a86e4 | 1296 | // If we are racing with an rcu_read_unlock_trace(), try again later. |
9ff86b4c | 1297 | if (unlikely(nesting < 0)) |
ba3a86e4 | 1298 | goto reset_ipi; |
d5f177d3 PM |
1299 | |
1300 | // Get here if the task is in a read-side critical section. Set | |
1301 | // its state so that it will awaken the grace-period kthread upon | |
1302 | // exit from that critical section. | |
55061126 | 1303 | rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS | TRC_NEED_QS_CHECKED); |
d5f177d3 PM |
1304 | |
1305 | reset_ipi: | |
1306 | // Allow future IPIs to be sent on CPU and for task. | |
1307 | // Also order this IPI handler against any later manipulations of | |
1308 | // the intended task. | |
8211e922 | 1309 | smp_store_release(per_cpu_ptr(&trc_ipi_to_cpu, smp_processor_id()), false); // ^^^ |
d5f177d3 PM |
1310 | smp_store_release(&texp->trc_ipi_to_cpu, -1); // ^^^ |
1311 | } | |
1312 | ||
1313 | /* Callback function for scheduler to check locked-down task. */ | |
3847b645 | 1314 | static int trc_inspect_reader(struct task_struct *t, void *bhp_in) |
d5f177d3 | 1315 | { |
3847b645 | 1316 | struct list_head *bhp = bhp_in; |
7d0c9c50 | 1317 | int cpu = task_cpu(t); |
18f08e75 | 1318 | int nesting; |
7e3b70e0 | 1319 | bool ofl = cpu_is_offline(cpu); |
7d0c9c50 | 1320 | |
897ba84d | 1321 | if (task_curr(t) && !ofl) { |
7d0c9c50 | 1322 | // If no chance of heavyweight readers, do it the hard way. |
897ba84d | 1323 | if (!IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB)) |
9b3c4ab3 | 1324 | return -EINVAL; |
7d0c9c50 PM |
1325 | |
1326 | // If heavyweight readers are enabled on the remote task, | |
1327 | // we can inspect its state despite its currently running. | |
1328 | // However, we cannot safely change its state. | |
40471509 | 1329 | n_heavy_reader_attempts++; |
897ba84d PM |
1330 | // Check for "running" idle tasks on offline CPUs. |
1331 | if (!rcu_dynticks_zero_in_eqs(cpu, &t->trc_reader_nesting)) | |
9b3c4ab3 | 1332 | return -EINVAL; // No quiescent state, do it the hard way. |
40471509 | 1333 | n_heavy_reader_updates++; |
18f08e75 | 1334 | nesting = 0; |
7d0c9c50 | 1335 | } else { |
bdb0cca0 | 1336 | // The task is not running, so C-language access is safe. |
18f08e75 | 1337 | nesting = t->trc_reader_nesting; |
897ba84d PM |
1338 | WARN_ON_ONCE(ofl && task_curr(t) && !is_idle_task(t)); |
1339 | if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB) && ofl) | |
1340 | n_heavy_reader_ofl_updates++; | |
7d0c9c50 | 1341 | } |
d5f177d3 | 1342 | |
18f08e75 PM |
1343 | // If not exiting a read-side critical section, mark as checked |
1344 | // so that the grace-period kthread will remove it from the | |
1345 | // holdout list. | |
0968e892 PM |
1346 | if (!nesting) { |
1347 | rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS_CHECKED); | |
1348 | return 0; // In QS, so done. | |
3847b645 | 1349 | } |
0968e892 PM |
1350 | if (nesting < 0) |
1351 | return -EINVAL; // QS transitioning, try again later. | |
7d0c9c50 PM |
1352 | |
1353 | // The task is in a read-side critical section, so set up its | |
0968e892 PM |
1354 | // state so that it will update state upon exit from that critical |
1355 | // section. | |
55061126 | 1356 | if (!rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS | TRC_NEED_QS_CHECKED)) |
3847b645 | 1357 | trc_add_holdout(t, bhp); |
9b3c4ab3 | 1358 | return 0; |
d5f177d3 PM |
1359 | } |
1360 | ||
1361 | /* Attempt to extract the state for the specified task. */ | |
1362 | static void trc_wait_for_one_reader(struct task_struct *t, | |
1363 | struct list_head *bhp) | |
1364 | { | |
1365 | int cpu; | |
1366 | ||
1367 | // If a previous IPI is still in flight, let it complete. | |
1368 | if (smp_load_acquire(&t->trc_ipi_to_cpu) != -1) // Order IPI | |
1369 | return; | |
1370 | ||
1371 | // The current task had better be in a quiescent state. | |
1372 | if (t == current) { | |
3847b645 | 1373 | rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS_CHECKED); |
bdb0cca0 | 1374 | WARN_ON_ONCE(READ_ONCE(t->trc_reader_nesting)); |
d5f177d3 PM |
1375 | return; |
1376 | } | |
1377 | ||
1378 | // Attempt to nail down the task for inspection. | |
1379 | get_task_struct(t); | |
3847b645 | 1380 | if (!task_call_func(t, trc_inspect_reader, bhp)) { |
d5f177d3 PM |
1381 | put_task_struct(t); |
1382 | return; | |
1383 | } | |
1384 | put_task_struct(t); | |
1385 | ||
45f4b4a2 PM |
1386 | // If this task is not yet on the holdout list, then we are in |
1387 | // an RCU read-side critical section. Otherwise, the invocation of | |
d0a85858 | 1388 | // trc_add_holdout() that added it to the list did the necessary |
45f4b4a2 PM |
1389 | // get_task_struct(). Either way, the task cannot be freed out |
1390 | // from under this code. | |
1391 | ||
d5f177d3 PM |
1392 | // If currently running, send an IPI, either way, add to list. |
1393 | trc_add_holdout(t, bhp); | |
574de876 PM |
1394 | if (task_curr(t) && |
1395 | time_after(jiffies + 1, rcu_tasks_trace.gp_start + rcu_task_ipi_delay)) { | |
d5f177d3 PM |
1396 | // The task is currently running, so try IPIing it. |
1397 | cpu = task_cpu(t); | |
1398 | ||
1399 | // If there is already an IPI outstanding, let it happen. | |
1400 | if (per_cpu(trc_ipi_to_cpu, cpu) || t->trc_ipi_to_cpu >= 0) | |
1401 | return; | |
1402 | ||
d5f177d3 PM |
1403 | per_cpu(trc_ipi_to_cpu, cpu) = true; |
1404 | t->trc_ipi_to_cpu = cpu; | |
238dbce3 | 1405 | rcu_tasks_trace.n_ipis++; |
96017bf9 | 1406 | if (smp_call_function_single(cpu, trc_read_check_handler, t, 0)) { |
d5f177d3 PM |
1407 | // Just in case there is some other reason for |
1408 | // failure than the target CPU being offline. | |
46aa886c NU |
1409 | WARN_ONCE(1, "%s(): smp_call_function_single() failed for CPU: %d\n", |
1410 | __func__, cpu); | |
7e0669c3 | 1411 | rcu_tasks_trace.n_ipis_fails++; |
d5f177d3 | 1412 | per_cpu(trc_ipi_to_cpu, cpu) = false; |
46aa886c | 1413 | t->trc_ipi_to_cpu = -1; |
d5f177d3 PM |
1414 | } |
1415 | } | |
1416 | } | |
1417 | ||
1418 | /* Initialize for a new RCU-tasks-trace grace period. */ | |
1419 | static void rcu_tasks_trace_pregp_step(void) | |
1420 | { | |
1421 | int cpu; | |
1422 | ||
d5f177d3 PM |
1423 | // There shouldn't be any old IPIs, but... |
1424 | for_each_possible_cpu(cpu) | |
1425 | WARN_ON_ONCE(per_cpu(trc_ipi_to_cpu, cpu)); | |
81b4a7bc PM |
1426 | |
1427 | // Disable CPU hotplug across the tasklist scan. | |
1428 | // This also waits for all readers in CPU-hotplug code paths. | |
1429 | cpus_read_lock(); | |
d5f177d3 PM |
1430 | } |
1431 | ||
1432 | /* Do first-round processing for the specified task. */ | |
1433 | static void rcu_tasks_trace_pertask(struct task_struct *t, | |
1434 | struct list_head *hop) | |
1435 | { | |
1b04fa99 | 1436 | // During early boot when there is only the one boot CPU, there |
5d4c90d7 PM |
1437 | // is no idle task for the other CPUs. Also, the grace-period |
1438 | // kthread is always in a quiescent state. Either way, just return. | |
1439 | if (unlikely(t == NULL) || t == current) | |
1b04fa99 URS |
1440 | return; |
1441 | ||
3847b645 | 1442 | rcu_st_need_qs(t, 0); |
d5f177d3 PM |
1443 | t->trc_ipi_to_cpu = -1; |
1444 | trc_wait_for_one_reader(t, hop); | |
1445 | } | |
1446 | ||
9796e1ae PM |
1447 | /* |
1448 | * Do intermediate processing between task and holdout scans and | |
1449 | * pick up the idle tasks. | |
1450 | */ | |
1451 | static void rcu_tasks_trace_postscan(struct list_head *hop) | |
d5f177d3 | 1452 | { |
9796e1ae PM |
1453 | int cpu; |
1454 | ||
5c9a9ca4 | 1455 | for_each_online_cpu(cpu) |
9796e1ae PM |
1456 | rcu_tasks_trace_pertask(idle_task(cpu), hop); |
1457 | ||
81b4a7bc PM |
1458 | // Re-enable CPU hotplug now that the tasklist scan has completed. |
1459 | cpus_read_unlock(); | |
1460 | ||
d5f177d3 PM |
1461 | // Wait for late-stage exiting tasks to finish exiting. |
1462 | // These might have passed the call to exit_tasks_rcu_finish(). | |
1463 | synchronize_rcu(); | |
3847b645 PM |
1464 | // Any tasks that exit after this point will set |
1465 | // TRC_NEED_QS_CHECKED in ->trc_reader_special.b.need_qs. | |
d5f177d3 PM |
1466 | } |
1467 | ||
65b629e7 NU |
1468 | /* Communicate task state back to the RCU tasks trace stall warning request. */ |
1469 | struct trc_stall_chk_rdr { | |
1470 | int nesting; | |
1471 | int ipi_to_cpu; | |
1472 | u8 needqs; | |
1473 | }; | |
1474 | ||
1475 | static int trc_check_slow_task(struct task_struct *t, void *arg) | |
1476 | { | |
1477 | struct trc_stall_chk_rdr *trc_rdrp = arg; | |
1478 | ||
1479 | if (task_curr(t)) | |
1480 | return false; // It is running, so decline to inspect it. | |
1481 | trc_rdrp->nesting = READ_ONCE(t->trc_reader_nesting); | |
1482 | trc_rdrp->ipi_to_cpu = READ_ONCE(t->trc_ipi_to_cpu); | |
3847b645 | 1483 | trc_rdrp->needqs = rcu_ld_need_qs(t); |
65b629e7 NU |
1484 | return true; |
1485 | } | |
1486 | ||
4593e772 PM |
1487 | /* Show the state of a task stalling the current RCU tasks trace GP. */ |
1488 | static void show_stalled_task_trace(struct task_struct *t, bool *firstreport) | |
1489 | { | |
1490 | int cpu; | |
65b629e7 NU |
1491 | struct trc_stall_chk_rdr trc_rdr; |
1492 | bool is_idle_tsk = is_idle_task(t); | |
4593e772 PM |
1493 | |
1494 | if (*firstreport) { | |
1495 | pr_err("INFO: rcu_tasks_trace detected stalls on tasks:\n"); | |
1496 | *firstreport = false; | |
1497 | } | |
4593e772 | 1498 | cpu = task_cpu(t); |
65b629e7 | 1499 | if (!task_call_func(t, trc_check_slow_task, &trc_rdr)) |
9f3eb5fb | 1500 | pr_alert("P%d: %c%c\n", |
65b629e7 | 1501 | t->pid, |
9f3eb5fb | 1502 | ".I"[t->trc_ipi_to_cpu >= 0], |
65b629e7 NU |
1503 | ".i"[is_idle_tsk]); |
1504 | else | |
1505 | pr_alert("P%d: %c%c%c nesting: %d%c cpu: %d\n", | |
1506 | t->pid, | |
1507 | ".I"[trc_rdr.ipi_to_cpu >= 0], | |
1508 | ".i"[is_idle_tsk], | |
1509 | ".N"[cpu >= 0 && tick_nohz_full_cpu(cpu)], | |
1510 | trc_rdr.nesting, | |
1511 | " N"[!!trc_rdr.needqs], | |
1512 | cpu); | |
4593e772 PM |
1513 | sched_show_task(t); |
1514 | } | |
1515 | ||
1516 | /* List stalled IPIs for RCU tasks trace. */ | |
1517 | static void show_stalled_ipi_trace(void) | |
1518 | { | |
1519 | int cpu; | |
1520 | ||
1521 | for_each_possible_cpu(cpu) | |
1522 | if (per_cpu(trc_ipi_to_cpu, cpu)) | |
1523 | pr_alert("\tIPI outstanding to CPU %d\n", cpu); | |
1524 | } | |
1525 | ||
d5f177d3 PM |
1526 | /* Do one scan of the holdout list. */ |
1527 | static void check_all_holdout_tasks_trace(struct list_head *hop, | |
4593e772 | 1528 | bool needreport, bool *firstreport) |
d5f177d3 PM |
1529 | { |
1530 | struct task_struct *g, *t; | |
1531 | ||
81b4a7bc PM |
1532 | // Disable CPU hotplug across the holdout list scan. |
1533 | cpus_read_lock(); | |
1534 | ||
d5f177d3 PM |
1535 | list_for_each_entry_safe(t, g, hop, trc_holdout_list) { |
1536 | // If safe and needed, try to check the current task. | |
1537 | if (READ_ONCE(t->trc_ipi_to_cpu) == -1 && | |
3847b645 | 1538 | !(rcu_ld_need_qs(t) & TRC_NEED_QS_CHECKED)) |
d5f177d3 PM |
1539 | trc_wait_for_one_reader(t, hop); |
1540 | ||
1541 | // If check succeeded, remove this task from the list. | |
f5dbc594 | 1542 | if (smp_load_acquire(&t->trc_ipi_to_cpu) == -1 && |
3847b645 | 1543 | rcu_ld_need_qs(t) == TRC_NEED_QS_CHECKED) |
d5f177d3 | 1544 | trc_del_holdout(t); |
4593e772 PM |
1545 | else if (needreport) |
1546 | show_stalled_task_trace(t, firstreport); | |
1547 | } | |
81b4a7bc PM |
1548 | |
1549 | // Re-enable CPU hotplug now that the holdout list scan has completed. | |
1550 | cpus_read_unlock(); | |
1551 | ||
4593e772 | 1552 | if (needreport) { |
89401176 | 1553 | if (*firstreport) |
4593e772 PM |
1554 | pr_err("INFO: rcu_tasks_trace detected stalls? (Late IPI?)\n"); |
1555 | show_stalled_ipi_trace(); | |
d5f177d3 PM |
1556 | } |
1557 | } | |
1558 | ||
cbe0d8d9 PM |
1559 | static void rcu_tasks_trace_empty_fn(void *unused) |
1560 | { | |
1561 | } | |
1562 | ||
d5f177d3 | 1563 | /* Wait for grace period to complete and provide ordering. */ |
af051ca4 | 1564 | static void rcu_tasks_trace_postgp(struct rcu_tasks *rtp) |
d5f177d3 | 1565 | { |
cbe0d8d9 | 1566 | int cpu; |
4593e772 | 1567 | |
cbe0d8d9 PM |
1568 | // Wait for any lingering IPI handlers to complete. Note that |
1569 | // if a CPU has gone offline or transitioned to userspace in the | |
1570 | // meantime, all IPI handlers should have been drained beforehand. | |
1571 | // Yes, this assumes that CPUs process IPIs in order. If that ever | |
1572 | // changes, there will need to be a recheck and/or timed wait. | |
1573 | for_each_online_cpu(cpu) | |
f5dbc594 | 1574 | if (WARN_ON_ONCE(smp_load_acquire(per_cpu_ptr(&trc_ipi_to_cpu, cpu)))) |
cbe0d8d9 PM |
1575 | smp_call_function_single(cpu, rcu_tasks_trace_empty_fn, NULL, 1); |
1576 | ||
d5f177d3 | 1577 | smp_mb(); // Caller's code must be ordered after wakeup. |
43766c3e | 1578 | // Pairs with pretty much every ordering primitive. |
d5f177d3 PM |
1579 | } |
1580 | ||
1581 | /* Report any needed quiescent state for this exiting task. */ | |
25246fc8 | 1582 | static void exit_tasks_rcu_finish_trace(struct task_struct *t) |
d5f177d3 | 1583 | { |
3847b645 | 1584 | rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS_CHECKED); |
bdb0cca0 | 1585 | WARN_ON_ONCE(READ_ONCE(t->trc_reader_nesting)); |
3847b645 | 1586 | if (WARN_ON_ONCE(rcu_ld_need_qs(t) & TRC_NEED_QS)) |
a5c071cc | 1587 | rcu_read_unlock_trace_special(t); |
3847b645 PM |
1588 | else |
1589 | WRITE_ONCE(t->trc_reader_nesting, 0); | |
d5f177d3 PM |
1590 | } |
1591 | ||
d5f177d3 PM |
1592 | /** |
1593 | * call_rcu_tasks_trace() - Queue a callback trace task-based grace period | |
1594 | * @rhp: structure to be used for queueing the RCU updates. | |
1595 | * @func: actual callback function to be invoked after the grace period | |
1596 | * | |
ed42c380 NU |
1597 | * The callback function will be invoked some time after a trace rcu-tasks |
1598 | * grace period elapses, in other words after all currently executing | |
1599 | * trace rcu-tasks read-side critical sections have completed. These | |
1600 | * read-side critical sections are delimited by calls to rcu_read_lock_trace() | |
1601 | * and rcu_read_unlock_trace(). | |
d5f177d3 PM |
1602 | * |
1603 | * See the description of call_rcu() for more detailed information on | |
1604 | * memory ordering guarantees. | |
1605 | */ | |
1606 | void call_rcu_tasks_trace(struct rcu_head *rhp, rcu_callback_t func) | |
1607 | { | |
1608 | call_rcu_tasks_generic(rhp, func, &rcu_tasks_trace); | |
1609 | } | |
1610 | EXPORT_SYMBOL_GPL(call_rcu_tasks_trace); | |
1611 | ||
1612 | /** | |
1613 | * synchronize_rcu_tasks_trace - wait for a trace rcu-tasks grace period | |
1614 | * | |
1615 | * Control will return to the caller some time after a trace rcu-tasks | |
c7dcf810 | 1616 | * grace period has elapsed, in other words after all currently executing |
ed42c380 | 1617 | * trace rcu-tasks read-side critical sections have elapsed. These read-side |
c7dcf810 PM |
1618 | * critical sections are delimited by calls to rcu_read_lock_trace() |
1619 | * and rcu_read_unlock_trace(). | |
d5f177d3 PM |
1620 | * |
1621 | * This is a very specialized primitive, intended only for a few uses in | |
1622 | * tracing and other situations requiring manipulation of function preambles | |
1623 | * and profiling hooks. The synchronize_rcu_tasks_trace() function is not | |
1624 | * (yet) intended for heavy use from multiple CPUs. | |
1625 | * | |
1626 | * See the description of synchronize_rcu() for more detailed information | |
1627 | * on memory ordering guarantees. | |
1628 | */ | |
1629 | void synchronize_rcu_tasks_trace(void) | |
1630 | { | |
1631 | RCU_LOCKDEP_WARN(lock_is_held(&rcu_trace_lock_map), "Illegal synchronize_rcu_tasks_trace() in RCU Tasks Trace read-side critical section"); | |
1632 | synchronize_rcu_tasks_generic(&rcu_tasks_trace); | |
1633 | } | |
1634 | EXPORT_SYMBOL_GPL(synchronize_rcu_tasks_trace); | |
1635 | ||
1636 | /** | |
1637 | * rcu_barrier_tasks_trace - Wait for in-flight call_rcu_tasks_trace() callbacks. | |
1638 | * | |
1639 | * Although the current implementation is guaranteed to wait, it is not | |
1640 | * obligated to, for example, if there are no pending callbacks. | |
1641 | */ | |
1642 | void rcu_barrier_tasks_trace(void) | |
1643 | { | |
ce9b1c66 | 1644 | rcu_barrier_tasks_generic(&rcu_tasks_trace); |
d5f177d3 PM |
1645 | } |
1646 | EXPORT_SYMBOL_GPL(rcu_barrier_tasks_trace); | |
1647 | ||
1648 | static int __init rcu_spawn_tasks_trace_kthread(void) | |
1649 | { | |
cafafd67 | 1650 | cblist_init_generic(&rcu_tasks_trace); |
2393a613 | 1651 | if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB)) { |
4fe192df | 1652 | rcu_tasks_trace.gp_sleep = HZ / 10; |
75dc2da5 | 1653 | rcu_tasks_trace.init_fract = HZ / 10; |
2393a613 | 1654 | } else { |
4fe192df PM |
1655 | rcu_tasks_trace.gp_sleep = HZ / 200; |
1656 | if (rcu_tasks_trace.gp_sleep <= 0) | |
1657 | rcu_tasks_trace.gp_sleep = 1; | |
75dc2da5 | 1658 | rcu_tasks_trace.init_fract = HZ / 200; |
2393a613 PM |
1659 | if (rcu_tasks_trace.init_fract <= 0) |
1660 | rcu_tasks_trace.init_fract = 1; | |
1661 | } | |
d5f177d3 PM |
1662 | rcu_tasks_trace.pregp_func = rcu_tasks_trace_pregp_step; |
1663 | rcu_tasks_trace.pertask_func = rcu_tasks_trace_pertask; | |
1664 | rcu_tasks_trace.postscan_func = rcu_tasks_trace_postscan; | |
1665 | rcu_tasks_trace.holdouts_func = check_all_holdout_tasks_trace; | |
1666 | rcu_tasks_trace.postgp_func = rcu_tasks_trace_postgp; | |
1667 | rcu_spawn_tasks_kthread_generic(&rcu_tasks_trace); | |
1668 | return 0; | |
1669 | } | |
d5f177d3 | 1670 | |
27c0f144 PM |
1671 | #if !defined(CONFIG_TINY_RCU) |
1672 | void show_rcu_tasks_trace_gp_kthread(void) | |
e21408ce | 1673 | { |
40471509 | 1674 | char buf[64]; |
e21408ce | 1675 | |
55061126 | 1676 | sprintf(buf, "h:%lu/%lu/%lu", |
edf3775f | 1677 | data_race(n_heavy_reader_ofl_updates), |
40471509 PM |
1678 | data_race(n_heavy_reader_updates), |
1679 | data_race(n_heavy_reader_attempts)); | |
e21408ce PM |
1680 | show_rcu_tasks_generic_gp_kthread(&rcu_tasks_trace, buf); |
1681 | } | |
27c0f144 PM |
1682 | EXPORT_SYMBOL_GPL(show_rcu_tasks_trace_gp_kthread); |
1683 | #endif // !defined(CONFIG_TINY_RCU) | |
e21408ce | 1684 | |
d5f177d3 | 1685 | #else /* #ifdef CONFIG_TASKS_TRACE_RCU */ |
25246fc8 | 1686 | static void exit_tasks_rcu_finish_trace(struct task_struct *t) { } |
d5f177d3 | 1687 | #endif /* #else #ifdef CONFIG_TASKS_TRACE_RCU */ |
8fd8ca38 | 1688 | |
8344496e | 1689 | #ifndef CONFIG_TINY_RCU |
e21408ce PM |
1690 | void show_rcu_tasks_gp_kthreads(void) |
1691 | { | |
1692 | show_rcu_tasks_classic_gp_kthread(); | |
1693 | show_rcu_tasks_rude_gp_kthread(); | |
1694 | show_rcu_tasks_trace_gp_kthread(); | |
1695 | } | |
8344496e | 1696 | #endif /* #ifndef CONFIG_TINY_RCU */ |
e21408ce | 1697 | |
bfba7ed0 URS |
1698 | #ifdef CONFIG_PROVE_RCU |
1699 | struct rcu_tasks_test_desc { | |
1700 | struct rcu_head rh; | |
1701 | const char *name; | |
1702 | bool notrun; | |
1703 | }; | |
1704 | ||
1705 | static struct rcu_tasks_test_desc tests[] = { | |
1706 | { | |
1707 | .name = "call_rcu_tasks()", | |
1708 | /* If not defined, the test is skipped. */ | |
1709 | .notrun = !IS_ENABLED(CONFIG_TASKS_RCU), | |
1710 | }, | |
1711 | { | |
1712 | .name = "call_rcu_tasks_rude()", | |
1713 | /* If not defined, the test is skipped. */ | |
1714 | .notrun = !IS_ENABLED(CONFIG_TASKS_RUDE_RCU), | |
1715 | }, | |
1716 | { | |
1717 | .name = "call_rcu_tasks_trace()", | |
1718 | /* If not defined, the test is skipped. */ | |
1719 | .notrun = !IS_ENABLED(CONFIG_TASKS_TRACE_RCU) | |
1720 | } | |
1721 | }; | |
1722 | ||
1723 | static void test_rcu_tasks_callback(struct rcu_head *rhp) | |
1724 | { | |
1725 | struct rcu_tasks_test_desc *rttd = | |
1726 | container_of(rhp, struct rcu_tasks_test_desc, rh); | |
1727 | ||
1728 | pr_info("Callback from %s invoked.\n", rttd->name); | |
1729 | ||
1730 | rttd->notrun = true; | |
1731 | } | |
1732 | ||
1733 | static void rcu_tasks_initiate_self_tests(void) | |
1734 | { | |
1735 | pr_info("Running RCU-tasks wait API self tests\n"); | |
1736 | #ifdef CONFIG_TASKS_RCU | |
1737 | synchronize_rcu_tasks(); | |
1738 | call_rcu_tasks(&tests[0].rh, test_rcu_tasks_callback); | |
1739 | #endif | |
1740 | ||
1741 | #ifdef CONFIG_TASKS_RUDE_RCU | |
1742 | synchronize_rcu_tasks_rude(); | |
1743 | call_rcu_tasks_rude(&tests[1].rh, test_rcu_tasks_callback); | |
1744 | #endif | |
1745 | ||
1746 | #ifdef CONFIG_TASKS_TRACE_RCU | |
1747 | synchronize_rcu_tasks_trace(); | |
1748 | call_rcu_tasks_trace(&tests[2].rh, test_rcu_tasks_callback); | |
1749 | #endif | |
1750 | } | |
1751 | ||
1752 | static int rcu_tasks_verify_self_tests(void) | |
1753 | { | |
1754 | int ret = 0; | |
1755 | int i; | |
1756 | ||
1757 | for (i = 0; i < ARRAY_SIZE(tests); i++) { | |
1758 | if (!tests[i].notrun) { // still hanging. | |
1759 | pr_err("%s has been failed.\n", tests[i].name); | |
1760 | ret = -1; | |
1761 | } | |
1762 | } | |
1763 | ||
1764 | if (ret) | |
1765 | WARN_ON(1); | |
1766 | ||
1767 | return ret; | |
1768 | } | |
1769 | late_initcall(rcu_tasks_verify_self_tests); | |
1770 | #else /* #ifdef CONFIG_PROVE_RCU */ | |
1771 | static void rcu_tasks_initiate_self_tests(void) { } | |
1772 | #endif /* #else #ifdef CONFIG_PROVE_RCU */ | |
1773 | ||
1b04fa99 URS |
1774 | void __init rcu_init_tasks_generic(void) |
1775 | { | |
1776 | #ifdef CONFIG_TASKS_RCU | |
1777 | rcu_spawn_tasks_kthread(); | |
1778 | #endif | |
1779 | ||
1780 | #ifdef CONFIG_TASKS_RUDE_RCU | |
1781 | rcu_spawn_tasks_rude_kthread(); | |
1782 | #endif | |
1783 | ||
1784 | #ifdef CONFIG_TASKS_TRACE_RCU | |
1785 | rcu_spawn_tasks_trace_kthread(); | |
1786 | #endif | |
bfba7ed0 URS |
1787 | |
1788 | // Run the self-tests. | |
1789 | rcu_tasks_initiate_self_tests(); | |
1b04fa99 URS |
1790 | } |
1791 | ||
8fd8ca38 PM |
1792 | #else /* #ifdef CONFIG_TASKS_RCU_GENERIC */ |
1793 | static inline void rcu_tasks_bootup_oddness(void) {} | |
1794 | #endif /* #else #ifdef CONFIG_TASKS_RCU_GENERIC */ |