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. |
d363f833 | 27 | * @rtp_work: Work queue for invoking callbacks. |
ce9b1c66 PM |
28 | * @barrier_q_head: RCU callback for barrier operation. |
29 | * @cpu: CPU number corresponding to this entry. | |
30 | * @rtpp: Pointer to the rcu_tasks structure. | |
cafafd67 PM |
31 | */ |
32 | struct rcu_tasks_percpu { | |
9b073de1 | 33 | struct rcu_segcblist cblist; |
381a4f3b | 34 | raw_spinlock_t __private lock; |
d363f833 | 35 | struct work_struct rtp_work; |
ce9b1c66 | 36 | struct rcu_head barrier_q_head; |
d363f833 PM |
37 | int cpu; |
38 | struct rcu_tasks *rtpp; | |
cafafd67 PM |
39 | }; |
40 | ||
41 | /** | |
42 | * struct rcu_tasks - Definition for a Tasks-RCU-like mechanism. | |
a616aec9 | 43 | * @cbs_wq: Wait queue allowing new callback to get kthread's attention. |
cafafd67 | 44 | * @cbs_gbl_lock: Lock protecting callback list. |
07e10515 | 45 | * @kthread_ptr: This flavor's grace-period/callback-invocation kthread. |
5873b8a9 | 46 | * @gp_func: This flavor's grace-period-wait function. |
af051ca4 | 47 | * @gp_state: Grace period's most recent state transition (debugging). |
4fe192df | 48 | * @gp_sleep: Per-grace-period sleep to prevent CPU-bound looping. |
2393a613 | 49 | * @init_fract: Initial backoff sleep interval. |
af051ca4 PM |
50 | * @gp_jiffies: Time of last @gp_state transition. |
51 | * @gp_start: Most recent grace-period start in jiffies. | |
b14fb4fb | 52 | * @tasks_gp_seq: Number of grace periods completed since boot. |
238dbce3 | 53 | * @n_ipis: Number of IPIs sent to encourage grace periods to end. |
7e0669c3 | 54 | * @n_ipis_fails: Number of IPI-send failures. |
e4fe5dd6 PM |
55 | * @pregp_func: This flavor's pre-grace-period function (optional). |
56 | * @pertask_func: This flavor's per-task scan function (optional). | |
57 | * @postscan_func: This flavor's post-task scan function (optional). | |
85b86994 | 58 | * @holdouts_func: This flavor's holdout-list scan function (optional). |
e4fe5dd6 | 59 | * @postgp_func: This flavor's post-grace-period function (optional). |
5873b8a9 | 60 | * @call_func: This flavor's call_rcu()-equivalent function. |
cafafd67 | 61 | * @rtpcpu: This flavor's rcu_tasks_percpu structure. |
7a30871b | 62 | * @percpu_enqueue_shift: Shift down CPU ID this much when enqueuing callbacks. |
8dd593fd | 63 | * @percpu_enqueue_lim: Number of per-CPU callback queues in use. |
ce9b1c66 PM |
64 | * @barrier_q_mutex: Serialize barrier operations. |
65 | * @barrier_q_count: Number of queues being waited on. | |
66 | * @barrier_q_completion: Barrier wait/wakeup mechanism. | |
67 | * @barrier_q_seq: Sequence number for barrier operations. | |
c97d12a6 PM |
68 | * @name: This flavor's textual name. |
69 | * @kname: This flavor's kthread name. | |
07e10515 PM |
70 | */ |
71 | struct rcu_tasks { | |
07e10515 | 72 | struct wait_queue_head cbs_wq; |
cafafd67 | 73 | raw_spinlock_t cbs_gbl_lock; |
af051ca4 | 74 | int gp_state; |
4fe192df | 75 | int gp_sleep; |
2393a613 | 76 | int init_fract; |
af051ca4 | 77 | unsigned long gp_jiffies; |
88092d0c | 78 | unsigned long gp_start; |
b14fb4fb | 79 | unsigned long tasks_gp_seq; |
238dbce3 | 80 | unsigned long n_ipis; |
7e0669c3 | 81 | unsigned long n_ipis_fails; |
07e10515 | 82 | struct task_struct *kthread_ptr; |
5873b8a9 | 83 | rcu_tasks_gp_func_t gp_func; |
e4fe5dd6 PM |
84 | pregp_func_t pregp_func; |
85 | pertask_func_t pertask_func; | |
86 | postscan_func_t postscan_func; | |
87 | holdouts_func_t holdouts_func; | |
88 | postgp_func_t postgp_func; | |
5873b8a9 | 89 | call_rcu_func_t call_func; |
cafafd67 | 90 | struct rcu_tasks_percpu __percpu *rtpcpu; |
7a30871b | 91 | int percpu_enqueue_shift; |
8dd593fd | 92 | int percpu_enqueue_lim; |
ce9b1c66 PM |
93 | struct mutex barrier_q_mutex; |
94 | atomic_t barrier_q_count; | |
95 | struct completion barrier_q_completion; | |
96 | unsigned long barrier_q_seq; | |
c97d12a6 PM |
97 | char *name; |
98 | char *kname; | |
07e10515 PM |
99 | }; |
100 | ||
cafafd67 PM |
101 | #define DEFINE_RCU_TASKS(rt_name, gp, call, n) \ |
102 | static DEFINE_PER_CPU(struct rcu_tasks_percpu, rt_name ## __percpu) = { \ | |
381a4f3b | 103 | .lock = __RAW_SPIN_LOCK_UNLOCKED(rt_name ## __percpu.cbs_pcpu_lock), \ |
cafafd67 PM |
104 | }; \ |
105 | static struct rcu_tasks rt_name = \ | |
106 | { \ | |
107 | .cbs_wq = __WAIT_QUEUE_HEAD_INITIALIZER(rt_name.cbs_wq), \ | |
108 | .cbs_gbl_lock = __RAW_SPIN_LOCK_UNLOCKED(rt_name.cbs_gbl_lock), \ | |
109 | .gp_func = gp, \ | |
110 | .call_func = call, \ | |
111 | .rtpcpu = &rt_name ## __percpu, \ | |
112 | .name = n, \ | |
7a30871b | 113 | .percpu_enqueue_shift = ilog2(CONFIG_NR_CPUS), \ |
8dd593fd | 114 | .percpu_enqueue_lim = 1, \ |
ce9b1c66 PM |
115 | .barrier_q_mutex = __MUTEX_INITIALIZER(rt_name.barrier_q_mutex), \ |
116 | .barrier_q_seq = (0UL - 50UL) << RCU_SEQ_CTR_SHIFT, \ | |
cafafd67 | 117 | .kname = #rt_name, \ |
07e10515 PM |
118 | } |
119 | ||
eacd6f04 PM |
120 | /* Track exiting tasks in order to allow them to be waited for. */ |
121 | DEFINE_STATIC_SRCU(tasks_rcu_exit_srcu); | |
122 | ||
b0afa0f0 | 123 | /* Avoid IPIing CPUs early in the grace period. */ |
574de876 | 124 | #define RCU_TASK_IPI_DELAY (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB) ? HZ / 2 : 0) |
b0afa0f0 PM |
125 | static int rcu_task_ipi_delay __read_mostly = RCU_TASK_IPI_DELAY; |
126 | module_param(rcu_task_ipi_delay, int, 0644); | |
127 | ||
eacd6f04 PM |
128 | /* Control stall timeouts. Disable with <= 0, otherwise jiffies till stall. */ |
129 | #define RCU_TASK_STALL_TIMEOUT (HZ * 60 * 10) | |
130 | static int rcu_task_stall_timeout __read_mostly = RCU_TASK_STALL_TIMEOUT; | |
131 | module_param(rcu_task_stall_timeout, int, 0644); | |
132 | ||
8610b656 PM |
133 | static int rcu_task_enqueue_lim __read_mostly = -1; |
134 | module_param(rcu_task_enqueue_lim, int, 0444); | |
135 | ||
af051ca4 PM |
136 | /* RCU tasks grace-period state for debugging. */ |
137 | #define RTGS_INIT 0 | |
138 | #define RTGS_WAIT_WAIT_CBS 1 | |
139 | #define RTGS_WAIT_GP 2 | |
140 | #define RTGS_PRE_WAIT_GP 3 | |
141 | #define RTGS_SCAN_TASKLIST 4 | |
142 | #define RTGS_POST_SCAN_TASKLIST 5 | |
143 | #define RTGS_WAIT_SCAN_HOLDOUTS 6 | |
144 | #define RTGS_SCAN_HOLDOUTS 7 | |
145 | #define RTGS_POST_GP 8 | |
146 | #define RTGS_WAIT_READERS 9 | |
147 | #define RTGS_INVOKE_CBS 10 | |
148 | #define RTGS_WAIT_CBS 11 | |
8344496e | 149 | #ifndef CONFIG_TINY_RCU |
af051ca4 PM |
150 | static const char * const rcu_tasks_gp_state_names[] = { |
151 | "RTGS_INIT", | |
152 | "RTGS_WAIT_WAIT_CBS", | |
153 | "RTGS_WAIT_GP", | |
154 | "RTGS_PRE_WAIT_GP", | |
155 | "RTGS_SCAN_TASKLIST", | |
156 | "RTGS_POST_SCAN_TASKLIST", | |
157 | "RTGS_WAIT_SCAN_HOLDOUTS", | |
158 | "RTGS_SCAN_HOLDOUTS", | |
159 | "RTGS_POST_GP", | |
160 | "RTGS_WAIT_READERS", | |
161 | "RTGS_INVOKE_CBS", | |
162 | "RTGS_WAIT_CBS", | |
163 | }; | |
8344496e | 164 | #endif /* #ifndef CONFIG_TINY_RCU */ |
af051ca4 | 165 | |
5873b8a9 PM |
166 | //////////////////////////////////////////////////////////////////////// |
167 | // | |
168 | // Generic code. | |
169 | ||
d363f833 PM |
170 | static void rcu_tasks_invoke_cbs_wq(struct work_struct *wp); |
171 | ||
af051ca4 PM |
172 | /* Record grace-period phase and time. */ |
173 | static void set_tasks_gp_state(struct rcu_tasks *rtp, int newstate) | |
174 | { | |
175 | rtp->gp_state = newstate; | |
176 | rtp->gp_jiffies = jiffies; | |
177 | } | |
178 | ||
8344496e | 179 | #ifndef CONFIG_TINY_RCU |
af051ca4 PM |
180 | /* Return state name. */ |
181 | static const char *tasks_gp_state_getname(struct rcu_tasks *rtp) | |
182 | { | |
183 | int i = data_race(rtp->gp_state); // Let KCSAN detect update races | |
184 | int j = READ_ONCE(i); // Prevent the compiler from reading twice | |
185 | ||
186 | if (j >= ARRAY_SIZE(rcu_tasks_gp_state_names)) | |
187 | return "???"; | |
188 | return rcu_tasks_gp_state_names[j]; | |
189 | } | |
8344496e | 190 | #endif /* #ifndef CONFIG_TINY_RCU */ |
af051ca4 | 191 | |
cafafd67 PM |
192 | // Initialize per-CPU callback lists for the specified flavor of |
193 | // Tasks RCU. | |
194 | static void cblist_init_generic(struct rcu_tasks *rtp) | |
195 | { | |
196 | int cpu; | |
197 | unsigned long flags; | |
8610b656 | 198 | int lim; |
cafafd67 PM |
199 | |
200 | raw_spin_lock_irqsave(&rtp->cbs_gbl_lock, flags); | |
8610b656 PM |
201 | if (rcu_task_enqueue_lim < 0) |
202 | rcu_task_enqueue_lim = nr_cpu_ids; | |
203 | else if (rcu_task_enqueue_lim == 0) | |
204 | rcu_task_enqueue_lim = 1; | |
205 | lim = rcu_task_enqueue_lim; | |
206 | ||
207 | if (lim > nr_cpu_ids) | |
208 | lim = nr_cpu_ids; | |
209 | WRITE_ONCE(rtp->percpu_enqueue_shift, ilog2(nr_cpu_ids / lim)); | |
210 | smp_store_release(&rtp->percpu_enqueue_lim, lim); | |
cafafd67 PM |
211 | for_each_possible_cpu(cpu) { |
212 | struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu); | |
213 | ||
214 | WARN_ON_ONCE(!rtpcp); | |
215 | if (cpu) | |
381a4f3b PM |
216 | raw_spin_lock_init(&ACCESS_PRIVATE(rtpcp, lock)); |
217 | raw_spin_lock_rcu_node(rtpcp); // irqs already disabled. | |
9b073de1 PM |
218 | if (rcu_segcblist_empty(&rtpcp->cblist)) |
219 | rcu_segcblist_init(&rtpcp->cblist); | |
d363f833 PM |
220 | INIT_WORK(&rtpcp->rtp_work, rcu_tasks_invoke_cbs_wq); |
221 | rtpcp->cpu = cpu; | |
222 | rtpcp->rtpp = rtp; | |
381a4f3b | 223 | raw_spin_unlock_rcu_node(rtpcp); // irqs remain disabled. |
cafafd67 PM |
224 | } |
225 | raw_spin_unlock_irqrestore(&rtp->cbs_gbl_lock, flags); | |
8610b656 | 226 | 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 |
227 | } |
228 | ||
5873b8a9 PM |
229 | // Enqueue a callback for the specified flavor of Tasks RCU. |
230 | static void call_rcu_tasks_generic(struct rcu_head *rhp, rcu_callback_t func, | |
231 | struct rcu_tasks *rtp) | |
eacd6f04 PM |
232 | { |
233 | unsigned long flags; | |
234 | bool needwake; | |
cafafd67 | 235 | struct rcu_tasks_percpu *rtpcp; |
eacd6f04 PM |
236 | |
237 | rhp->next = NULL; | |
238 | rhp->func = func; | |
cafafd67 | 239 | local_irq_save(flags); |
7a30871b PM |
240 | rtpcp = per_cpu_ptr(rtp->rtpcpu, |
241 | smp_processor_id() >> READ_ONCE(rtp->percpu_enqueue_shift)); | |
381a4f3b | 242 | raw_spin_lock_rcu_node(rtpcp); // irqs already disabled. |
9b073de1 | 243 | if (!rcu_segcblist_is_enabled(&rtpcp->cblist)) { |
381a4f3b | 244 | raw_spin_unlock_rcu_node(rtpcp); // irqs remain disabled. |
cafafd67 | 245 | cblist_init_generic(rtp); |
381a4f3b | 246 | raw_spin_lock_rcu_node(rtpcp); // irqs already disabled. |
cafafd67 | 247 | } |
9b073de1 PM |
248 | needwake = rcu_segcblist_empty(&rtpcp->cblist); |
249 | rcu_segcblist_enqueue(&rtpcp->cblist, rhp); | |
381a4f3b | 250 | raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags); |
eacd6f04 | 251 | /* We can't create the thread unless interrupts are enabled. */ |
07e10515 PM |
252 | if (needwake && READ_ONCE(rtp->kthread_ptr)) |
253 | wake_up(&rtp->cbs_wq); | |
eacd6f04 | 254 | } |
eacd6f04 | 255 | |
5873b8a9 PM |
256 | // Wait for a grace period for the specified flavor of Tasks RCU. |
257 | static void synchronize_rcu_tasks_generic(struct rcu_tasks *rtp) | |
eacd6f04 PM |
258 | { |
259 | /* Complain if the scheduler has not started. */ | |
260 | RCU_LOCKDEP_WARN(rcu_scheduler_active == RCU_SCHEDULER_INACTIVE, | |
261 | "synchronize_rcu_tasks called too soon"); | |
262 | ||
263 | /* Wait for the grace period. */ | |
5873b8a9 | 264 | wait_rcu_gp(rtp->call_func); |
eacd6f04 PM |
265 | } |
266 | ||
ce9b1c66 PM |
267 | // RCU callback function for rcu_barrier_tasks_generic(). |
268 | static void rcu_barrier_tasks_generic_cb(struct rcu_head *rhp) | |
269 | { | |
270 | struct rcu_tasks *rtp; | |
271 | struct rcu_tasks_percpu *rtpcp; | |
272 | ||
273 | rtpcp = container_of(rhp, struct rcu_tasks_percpu, barrier_q_head); | |
274 | rtp = rtpcp->rtpp; | |
275 | if (atomic_dec_and_test(&rtp->barrier_q_count)) | |
276 | complete(&rtp->barrier_q_completion); | |
277 | } | |
278 | ||
279 | // Wait for all in-flight callbacks for the specified RCU Tasks flavor. | |
280 | // Operates in a manner similar to rcu_barrier(). | |
281 | static void rcu_barrier_tasks_generic(struct rcu_tasks *rtp) | |
282 | { | |
283 | int cpu; | |
284 | unsigned long flags; | |
285 | struct rcu_tasks_percpu *rtpcp; | |
286 | unsigned long s = rcu_seq_snap(&rtp->barrier_q_seq); | |
287 | ||
288 | mutex_lock(&rtp->barrier_q_mutex); | |
289 | if (rcu_seq_done(&rtp->barrier_q_seq, s)) { | |
290 | smp_mb(); | |
291 | mutex_unlock(&rtp->barrier_q_mutex); | |
292 | return; | |
293 | } | |
294 | rcu_seq_start(&rtp->barrier_q_seq); | |
295 | init_completion(&rtp->barrier_q_completion); | |
296 | atomic_set(&rtp->barrier_q_count, 2); | |
297 | for_each_possible_cpu(cpu) { | |
298 | if (cpu >= smp_load_acquire(&rtp->percpu_enqueue_lim)) | |
299 | break; | |
300 | rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu); | |
301 | rtpcp->barrier_q_head.func = rcu_barrier_tasks_generic_cb; | |
302 | raw_spin_lock_irqsave_rcu_node(rtpcp, flags); | |
303 | if (rcu_segcblist_entrain(&rtpcp->cblist, &rtpcp->barrier_q_head)) | |
304 | atomic_inc(&rtp->barrier_q_count); | |
305 | raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags); | |
306 | } | |
307 | if (atomic_sub_and_test(2, &rtp->barrier_q_count)) | |
308 | complete(&rtp->barrier_q_completion); | |
309 | wait_for_completion(&rtp->barrier_q_completion); | |
310 | rcu_seq_end(&rtp->barrier_q_seq); | |
311 | mutex_unlock(&rtp->barrier_q_mutex); | |
312 | } | |
313 | ||
4d1114c0 PM |
314 | // Advance callbacks and indicate whether either a grace period or |
315 | // callback invocation is needed. | |
316 | static int rcu_tasks_need_gpcb(struct rcu_tasks *rtp) | |
317 | { | |
318 | int cpu; | |
319 | unsigned long flags; | |
320 | int needgpcb = 0; | |
321 | ||
8610b656 | 322 | for (cpu = 0; cpu < smp_load_acquire(&rtp->percpu_enqueue_lim); cpu++) { |
4d1114c0 PM |
323 | struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu); |
324 | ||
325 | /* Advance and accelerate any new callbacks. */ | |
326 | if (rcu_segcblist_empty(&rtpcp->cblist)) | |
327 | continue; | |
328 | raw_spin_lock_irqsave_rcu_node(rtpcp, flags); | |
329 | rcu_segcblist_advance(&rtpcp->cblist, rcu_seq_current(&rtp->tasks_gp_seq)); | |
330 | (void)rcu_segcblist_accelerate(&rtpcp->cblist, rcu_seq_snap(&rtp->tasks_gp_seq)); | |
331 | if (rcu_segcblist_pend_cbs(&rtpcp->cblist)) | |
332 | needgpcb |= 0x3; | |
333 | if (!rcu_segcblist_empty(&rtpcp->cblist)) | |
334 | needgpcb |= 0x1; | |
335 | raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags); | |
336 | } | |
337 | return needgpcb; | |
338 | } | |
339 | ||
57881863 | 340 | // Advance callbacks and invoke any that are ready. |
d363f833 | 341 | static void rcu_tasks_invoke_cbs(struct rcu_tasks *rtp, struct rcu_tasks_percpu *rtpcp) |
eacd6f04 | 342 | { |
57881863 | 343 | int cpu; |
d363f833 | 344 | int cpunext; |
eacd6f04 | 345 | unsigned long flags; |
9b073de1 | 346 | int len; |
9b073de1 | 347 | struct rcu_head *rhp; |
d363f833 PM |
348 | struct rcu_cblist rcl = RCU_CBLIST_INITIALIZER(rcl); |
349 | struct rcu_tasks_percpu *rtpcp_next; | |
350 | ||
351 | cpu = rtpcp->cpu; | |
352 | cpunext = cpu * 2 + 1; | |
8610b656 | 353 | if (cpunext < smp_load_acquire(&rtp->percpu_enqueue_lim)) { |
d363f833 PM |
354 | rtpcp_next = per_cpu_ptr(rtp->rtpcpu, cpunext); |
355 | queue_work_on(cpunext, system_wq, &rtpcp_next->rtp_work); | |
356 | cpunext++; | |
8610b656 | 357 | if (cpunext < smp_load_acquire(&rtp->percpu_enqueue_lim)) { |
d363f833 PM |
358 | rtpcp_next = per_cpu_ptr(rtp->rtpcpu, cpunext); |
359 | queue_work_on(cpunext, system_wq, &rtpcp_next->rtp_work); | |
57881863 | 360 | } |
57881863 | 361 | } |
d363f833 PM |
362 | |
363 | if (rcu_segcblist_empty(&rtpcp->cblist)) | |
364 | return; | |
365 | raw_spin_lock_irqsave_rcu_node(rtpcp, flags); | |
366 | rcu_segcblist_advance(&rtpcp->cblist, rcu_seq_current(&rtp->tasks_gp_seq)); | |
367 | rcu_segcblist_extract_done_cbs(&rtpcp->cblist, &rcl); | |
368 | raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags); | |
369 | len = rcl.len; | |
370 | for (rhp = rcu_cblist_dequeue(&rcl); rhp; rhp = rcu_cblist_dequeue(&rcl)) { | |
371 | local_bh_disable(); | |
372 | rhp->func(rhp); | |
373 | local_bh_enable(); | |
374 | cond_resched(); | |
375 | } | |
376 | raw_spin_lock_irqsave_rcu_node(rtpcp, flags); | |
377 | rcu_segcblist_add_len(&rtpcp->cblist, -len); | |
378 | (void)rcu_segcblist_accelerate(&rtpcp->cblist, rcu_seq_snap(&rtp->tasks_gp_seq)); | |
379 | raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags); | |
380 | } | |
381 | ||
382 | // Workqueue flood to advance callbacks and invoke any that are ready. | |
383 | static void rcu_tasks_invoke_cbs_wq(struct work_struct *wp) | |
384 | { | |
385 | struct rcu_tasks *rtp; | |
386 | struct rcu_tasks_percpu *rtpcp = container_of(wp, struct rcu_tasks_percpu, rtp_work); | |
387 | ||
388 | rtp = rtpcp->rtpp; | |
389 | rcu_tasks_invoke_cbs(rtp, rtpcp); | |
57881863 PM |
390 | } |
391 | ||
392 | /* RCU-tasks kthread that detects grace periods and invokes callbacks. */ | |
393 | static int __noreturn rcu_tasks_kthread(void *arg) | |
394 | { | |
395 | int needgpcb; | |
07e10515 | 396 | struct rcu_tasks *rtp = arg; |
eacd6f04 PM |
397 | |
398 | /* Run on housekeeping CPUs by default. Sysadm can move if desired. */ | |
399 | housekeeping_affine(current, HK_FLAG_RCU); | |
07e10515 | 400 | WRITE_ONCE(rtp->kthread_ptr, current); // Let GPs start! |
eacd6f04 PM |
401 | |
402 | /* | |
403 | * Each pass through the following loop makes one check for | |
404 | * newly arrived callbacks, and, if there are some, waits for | |
405 | * one RCU-tasks grace period and then invokes the callbacks. | |
406 | * This loop is terminated by the system going down. ;-) | |
407 | */ | |
408 | for (;;) { | |
0db7c32a | 409 | set_tasks_gp_state(rtp, RTGS_WAIT_CBS); |
eacd6f04 | 410 | |
eacd6f04 | 411 | /* If there were none, wait a bit and start over. */ |
4d1114c0 | 412 | wait_event_idle(rtp->cbs_wq, (needgpcb = rcu_tasks_need_gpcb(rtp))); |
eacd6f04 | 413 | |
4d1114c0 PM |
414 | if (needgpcb & 0x2) { |
415 | // Wait for one grace period. | |
416 | set_tasks_gp_state(rtp, RTGS_WAIT_GP); | |
417 | rtp->gp_start = jiffies; | |
418 | rcu_seq_start(&rtp->tasks_gp_seq); | |
419 | rtp->gp_func(rtp); | |
420 | rcu_seq_end(&rtp->tasks_gp_seq); | |
421 | } | |
eacd6f04 | 422 | |
57881863 | 423 | /* Invoke callbacks. */ |
af051ca4 | 424 | set_tasks_gp_state(rtp, RTGS_INVOKE_CBS); |
d363f833 | 425 | rcu_tasks_invoke_cbs(rtp, per_cpu_ptr(rtp->rtpcpu, 0)); |
57881863 | 426 | |
eacd6f04 | 427 | /* Paranoid sleep to keep this from entering a tight loop */ |
4fe192df | 428 | schedule_timeout_idle(rtp->gp_sleep); |
eacd6f04 PM |
429 | } |
430 | } | |
431 | ||
1b04fa99 | 432 | /* Spawn RCU-tasks grace-period kthread. */ |
5873b8a9 | 433 | static void __init rcu_spawn_tasks_kthread_generic(struct rcu_tasks *rtp) |
eacd6f04 PM |
434 | { |
435 | struct task_struct *t; | |
436 | ||
c97d12a6 PM |
437 | t = kthread_run(rcu_tasks_kthread, rtp, "%s_kthread", rtp->kname); |
438 | if (WARN_ONCE(IS_ERR(t), "%s: Could not start %s grace-period kthread, OOM is now expected behavior\n", __func__, rtp->name)) | |
5873b8a9 | 439 | return; |
eacd6f04 | 440 | smp_mb(); /* Ensure others see full kthread. */ |
eacd6f04 | 441 | } |
eacd6f04 | 442 | |
eacd6f04 PM |
443 | #ifndef CONFIG_TINY_RCU |
444 | ||
445 | /* | |
446 | * Print any non-default Tasks RCU settings. | |
447 | */ | |
448 | static void __init rcu_tasks_bootup_oddness(void) | |
449 | { | |
d5f177d3 | 450 | #if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU) |
eacd6f04 PM |
451 | if (rcu_task_stall_timeout != RCU_TASK_STALL_TIMEOUT) |
452 | pr_info("\tTasks-RCU CPU stall warnings timeout set to %d (rcu_task_stall_timeout).\n", rcu_task_stall_timeout); | |
d5f177d3 PM |
453 | #endif /* #ifdef CONFIG_TASKS_RCU */ |
454 | #ifdef CONFIG_TASKS_RCU | |
455 | pr_info("\tTrampoline variant of Tasks RCU enabled.\n"); | |
eacd6f04 | 456 | #endif /* #ifdef CONFIG_TASKS_RCU */ |
c84aad76 PM |
457 | #ifdef CONFIG_TASKS_RUDE_RCU |
458 | pr_info("\tRude variant of Tasks RCU enabled.\n"); | |
459 | #endif /* #ifdef CONFIG_TASKS_RUDE_RCU */ | |
d5f177d3 PM |
460 | #ifdef CONFIG_TASKS_TRACE_RCU |
461 | pr_info("\tTracing variant of Tasks RCU enabled.\n"); | |
462 | #endif /* #ifdef CONFIG_TASKS_TRACE_RCU */ | |
eacd6f04 PM |
463 | } |
464 | ||
465 | #endif /* #ifndef CONFIG_TINY_RCU */ | |
5873b8a9 | 466 | |
8344496e | 467 | #ifndef CONFIG_TINY_RCU |
e21408ce PM |
468 | /* Dump out rcutorture-relevant state common to all RCU-tasks flavors. */ |
469 | static void show_rcu_tasks_generic_gp_kthread(struct rcu_tasks *rtp, char *s) | |
470 | { | |
cafafd67 | 471 | struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, 0); // for_each... |
7e0669c3 | 472 | pr_info("%s: %s(%d) since %lu g:%lu i:%lu/%lu %c%c %s\n", |
e21408ce | 473 | rtp->kname, |
7e0669c3 | 474 | tasks_gp_state_getname(rtp), data_race(rtp->gp_state), |
af051ca4 | 475 | jiffies - data_race(rtp->gp_jiffies), |
b14fb4fb | 476 | data_race(rcu_seq_current(&rtp->tasks_gp_seq)), |
7e0669c3 | 477 | data_race(rtp->n_ipis_fails), data_race(rtp->n_ipis), |
e21408ce | 478 | ".k"[!!data_race(rtp->kthread_ptr)], |
9b073de1 | 479 | ".C"[!data_race(rcu_segcblist_empty(&rtpcp->cblist))], |
e21408ce PM |
480 | s); |
481 | } | |
27c0f144 | 482 | #endif // #ifndef CONFIG_TINY_RCU |
e21408ce | 483 | |
25246fc8 PM |
484 | static void exit_tasks_rcu_finish_trace(struct task_struct *t); |
485 | ||
486 | #if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU) | |
5873b8a9 | 487 | |
d01aa263 PM |
488 | //////////////////////////////////////////////////////////////////////// |
489 | // | |
490 | // Shared code between task-list-scanning variants of Tasks RCU. | |
491 | ||
492 | /* Wait for one RCU-tasks grace period. */ | |
493 | static void rcu_tasks_wait_gp(struct rcu_tasks *rtp) | |
494 | { | |
495 | struct task_struct *g, *t; | |
496 | unsigned long lastreport; | |
497 | LIST_HEAD(holdouts); | |
498 | int fract; | |
499 | ||
af051ca4 | 500 | set_tasks_gp_state(rtp, RTGS_PRE_WAIT_GP); |
d01aa263 PM |
501 | rtp->pregp_func(); |
502 | ||
503 | /* | |
504 | * There were callbacks, so we need to wait for an RCU-tasks | |
505 | * grace period. Start off by scanning the task list for tasks | |
506 | * that are not already voluntarily blocked. Mark these tasks | |
507 | * and make a list of them in holdouts. | |
508 | */ | |
af051ca4 | 509 | set_tasks_gp_state(rtp, RTGS_SCAN_TASKLIST); |
d01aa263 PM |
510 | rcu_read_lock(); |
511 | for_each_process_thread(g, t) | |
512 | rtp->pertask_func(t, &holdouts); | |
513 | rcu_read_unlock(); | |
514 | ||
af051ca4 | 515 | set_tasks_gp_state(rtp, RTGS_POST_SCAN_TASKLIST); |
9796e1ae | 516 | rtp->postscan_func(&holdouts); |
d01aa263 PM |
517 | |
518 | /* | |
519 | * Each pass through the following loop scans the list of holdout | |
520 | * tasks, removing any that are no longer holdouts. When the list | |
521 | * is empty, we are done. | |
522 | */ | |
523 | lastreport = jiffies; | |
524 | ||
2393a613 PM |
525 | // Start off with initial wait and slowly back off to 1 HZ wait. |
526 | fract = rtp->init_fract; | |
d01aa263 | 527 | |
77dc1741 | 528 | while (!list_empty(&holdouts)) { |
d01aa263 PM |
529 | bool firstreport; |
530 | bool needreport; | |
531 | int rtst; | |
532 | ||
d01aa263 | 533 | /* Slowly back off waiting for holdouts */ |
af051ca4 | 534 | set_tasks_gp_state(rtp, RTGS_WAIT_SCAN_HOLDOUTS); |
75dc2da5 | 535 | schedule_timeout_idle(fract); |
d01aa263 | 536 | |
75dc2da5 PM |
537 | if (fract < HZ) |
538 | fract++; | |
d01aa263 PM |
539 | |
540 | rtst = READ_ONCE(rcu_task_stall_timeout); | |
541 | needreport = rtst > 0 && time_after(jiffies, lastreport + rtst); | |
542 | if (needreport) | |
543 | lastreport = jiffies; | |
544 | firstreport = true; | |
545 | WARN_ON(signal_pending(current)); | |
af051ca4 | 546 | set_tasks_gp_state(rtp, RTGS_SCAN_HOLDOUTS); |
d01aa263 PM |
547 | rtp->holdouts_func(&holdouts, needreport, &firstreport); |
548 | } | |
549 | ||
af051ca4 PM |
550 | set_tasks_gp_state(rtp, RTGS_POST_GP); |
551 | rtp->postgp_func(rtp); | |
d01aa263 PM |
552 | } |
553 | ||
25246fc8 PM |
554 | #endif /* #if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU) */ |
555 | ||
556 | #ifdef CONFIG_TASKS_RCU | |
557 | ||
5873b8a9 PM |
558 | //////////////////////////////////////////////////////////////////////// |
559 | // | |
560 | // Simple variant of RCU whose quiescent states are voluntary context | |
8af9e2c7 | 561 | // switch, cond_resched_tasks_rcu_qs(), user-space execution, and idle. |
5873b8a9 PM |
562 | // As such, grace periods can take one good long time. There are no |
563 | // read-side primitives similar to rcu_read_lock() and rcu_read_unlock() | |
564 | // because this implementation is intended to get the system into a safe | |
565 | // state for some of the manipulations involved in tracing and the like. | |
566 | // Finally, this implementation does not support high call_rcu_tasks() | |
567 | // rates from multiple CPUs. If this is required, per-CPU callback lists | |
568 | // will be needed. | |
06a3ec92 PM |
569 | // |
570 | // The implementation uses rcu_tasks_wait_gp(), which relies on function | |
571 | // pointers in the rcu_tasks structure. The rcu_spawn_tasks_kthread() | |
572 | // function sets these function pointers up so that rcu_tasks_wait_gp() | |
573 | // invokes these functions in this order: | |
574 | // | |
575 | // rcu_tasks_pregp_step(): | |
576 | // Invokes synchronize_rcu() in order to wait for all in-flight | |
577 | // t->on_rq and t->nvcsw transitions to complete. This works because | |
578 | // all such transitions are carried out with interrupts disabled. | |
579 | // rcu_tasks_pertask(), invoked on every non-idle task: | |
580 | // For every runnable non-idle task other than the current one, use | |
581 | // get_task_struct() to pin down that task, snapshot that task's | |
582 | // number of voluntary context switches, and add that task to the | |
583 | // holdout list. | |
584 | // rcu_tasks_postscan(): | |
585 | // Invoke synchronize_srcu() to ensure that all tasks that were | |
586 | // in the process of exiting (and which thus might not know to | |
587 | // synchronize with this RCU Tasks grace period) have completed | |
588 | // exiting. | |
589 | // check_all_holdout_tasks(), repeatedly until holdout list is empty: | |
590 | // Scans the holdout list, attempting to identify a quiescent state | |
591 | // for each task on the list. If there is a quiescent state, the | |
592 | // corresponding task is removed from the holdout list. | |
593 | // rcu_tasks_postgp(): | |
594 | // Invokes synchronize_rcu() in order to ensure that all prior | |
595 | // t->on_rq and t->nvcsw transitions are seen by all CPUs and tasks | |
596 | // to have happened before the end of this RCU Tasks grace period. | |
597 | // Again, this works because all such transitions are carried out | |
598 | // with interrupts disabled. | |
599 | // | |
600 | // For each exiting task, the exit_tasks_rcu_start() and | |
601 | // exit_tasks_rcu_finish() functions begin and end, respectively, the SRCU | |
602 | // read-side critical sections waited for by rcu_tasks_postscan(). | |
603 | // | |
381a4f3b PM |
604 | // Pre-grace-period update-side code is ordered before the grace |
605 | // via the raw_spin_lock.*rcu_node(). Pre-grace-period read-side code | |
606 | // is ordered before the grace period via synchronize_rcu() call in | |
607 | // rcu_tasks_pregp_step() and by the scheduler's locks and interrupt | |
06a3ec92 | 608 | // disabling. |
5873b8a9 | 609 | |
e4fe5dd6 PM |
610 | /* Pre-grace-period preparation. */ |
611 | static void rcu_tasks_pregp_step(void) | |
612 | { | |
613 | /* | |
614 | * Wait for all pre-existing t->on_rq and t->nvcsw transitions | |
615 | * to complete. Invoking synchronize_rcu() suffices because all | |
616 | * these transitions occur with interrupts disabled. Without this | |
617 | * synchronize_rcu(), a read-side critical section that started | |
618 | * before the grace period might be incorrectly seen as having | |
619 | * started after the grace period. | |
620 | * | |
621 | * This synchronize_rcu() also dispenses with the need for a | |
622 | * memory barrier on the first store to t->rcu_tasks_holdout, | |
623 | * as it forces the store to happen after the beginning of the | |
624 | * grace period. | |
625 | */ | |
626 | synchronize_rcu(); | |
627 | } | |
628 | ||
629 | /* Per-task initial processing. */ | |
630 | static void rcu_tasks_pertask(struct task_struct *t, struct list_head *hop) | |
631 | { | |
632 | if (t != current && READ_ONCE(t->on_rq) && !is_idle_task(t)) { | |
633 | get_task_struct(t); | |
634 | t->rcu_tasks_nvcsw = READ_ONCE(t->nvcsw); | |
635 | WRITE_ONCE(t->rcu_tasks_holdout, true); | |
636 | list_add(&t->rcu_tasks_holdout_list, hop); | |
637 | } | |
638 | } | |
639 | ||
640 | /* Processing between scanning taskslist and draining the holdout list. */ | |
04a3c5aa | 641 | static void rcu_tasks_postscan(struct list_head *hop) |
e4fe5dd6 PM |
642 | { |
643 | /* | |
644 | * Wait for tasks that are in the process of exiting. This | |
645 | * does only part of the job, ensuring that all tasks that were | |
646 | * previously exiting reach the point where they have disabled | |
647 | * preemption, allowing the later synchronize_rcu() to finish | |
648 | * the job. | |
649 | */ | |
650 | synchronize_srcu(&tasks_rcu_exit_srcu); | |
651 | } | |
652 | ||
5873b8a9 PM |
653 | /* See if tasks are still holding out, complain if so. */ |
654 | static void check_holdout_task(struct task_struct *t, | |
655 | bool needreport, bool *firstreport) | |
656 | { | |
657 | int cpu; | |
658 | ||
659 | if (!READ_ONCE(t->rcu_tasks_holdout) || | |
660 | t->rcu_tasks_nvcsw != READ_ONCE(t->nvcsw) || | |
661 | !READ_ONCE(t->on_rq) || | |
662 | (IS_ENABLED(CONFIG_NO_HZ_FULL) && | |
663 | !is_idle_task(t) && t->rcu_tasks_idle_cpu >= 0)) { | |
664 | WRITE_ONCE(t->rcu_tasks_holdout, false); | |
665 | list_del_init(&t->rcu_tasks_holdout_list); | |
666 | put_task_struct(t); | |
667 | return; | |
668 | } | |
669 | rcu_request_urgent_qs_task(t); | |
670 | if (!needreport) | |
671 | return; | |
672 | if (*firstreport) { | |
673 | pr_err("INFO: rcu_tasks detected stalls on tasks:\n"); | |
674 | *firstreport = false; | |
675 | } | |
676 | cpu = task_cpu(t); | |
677 | pr_alert("%p: %c%c nvcsw: %lu/%lu holdout: %d idle_cpu: %d/%d\n", | |
678 | t, ".I"[is_idle_task(t)], | |
679 | "N."[cpu < 0 || !tick_nohz_full_cpu(cpu)], | |
680 | t->rcu_tasks_nvcsw, t->nvcsw, t->rcu_tasks_holdout, | |
681 | t->rcu_tasks_idle_cpu, cpu); | |
682 | sched_show_task(t); | |
683 | } | |
684 | ||
e4fe5dd6 PM |
685 | /* Scan the holdout lists for tasks no longer holding out. */ |
686 | static void check_all_holdout_tasks(struct list_head *hop, | |
687 | bool needreport, bool *firstreport) | |
688 | { | |
689 | struct task_struct *t, *t1; | |
690 | ||
691 | list_for_each_entry_safe(t, t1, hop, rcu_tasks_holdout_list) { | |
692 | check_holdout_task(t, needreport, firstreport); | |
693 | cond_resched(); | |
694 | } | |
695 | } | |
696 | ||
697 | /* Finish off the Tasks-RCU grace period. */ | |
af051ca4 | 698 | static void rcu_tasks_postgp(struct rcu_tasks *rtp) |
e4fe5dd6 PM |
699 | { |
700 | /* | |
701 | * Because ->on_rq and ->nvcsw are not guaranteed to have a full | |
702 | * memory barriers prior to them in the schedule() path, memory | |
703 | * reordering on other CPUs could cause their RCU-tasks read-side | |
704 | * critical sections to extend past the end of the grace period. | |
705 | * However, because these ->nvcsw updates are carried out with | |
706 | * interrupts disabled, we can use synchronize_rcu() to force the | |
707 | * needed ordering on all such CPUs. | |
708 | * | |
709 | * This synchronize_rcu() also confines all ->rcu_tasks_holdout | |
710 | * accesses to be within the grace period, avoiding the need for | |
711 | * memory barriers for ->rcu_tasks_holdout accesses. | |
712 | * | |
713 | * In addition, this synchronize_rcu() waits for exiting tasks | |
714 | * to complete their final preempt_disable() region of execution, | |
715 | * cleaning up after the synchronize_srcu() above. | |
716 | */ | |
717 | synchronize_rcu(); | |
718 | } | |
719 | ||
5873b8a9 | 720 | void call_rcu_tasks(struct rcu_head *rhp, rcu_callback_t func); |
c97d12a6 | 721 | DEFINE_RCU_TASKS(rcu_tasks, rcu_tasks_wait_gp, call_rcu_tasks, "RCU Tasks"); |
5873b8a9 PM |
722 | |
723 | /** | |
724 | * call_rcu_tasks() - Queue an RCU for invocation task-based grace period | |
725 | * @rhp: structure to be used for queueing the RCU updates. | |
726 | * @func: actual callback function to be invoked after the grace period | |
727 | * | |
728 | * The callback function will be invoked some time after a full grace | |
729 | * period elapses, in other words after all currently executing RCU | |
730 | * read-side critical sections have completed. call_rcu_tasks() assumes | |
731 | * that the read-side critical sections end at a voluntary context | |
8af9e2c7 | 732 | * switch (not a preemption!), cond_resched_tasks_rcu_qs(), entry into idle, |
5873b8a9 PM |
733 | * or transition to usermode execution. As such, there are no read-side |
734 | * primitives analogous to rcu_read_lock() and rcu_read_unlock() because | |
735 | * this primitive is intended to determine that all tasks have passed | |
a616aec9 | 736 | * through a safe state, not so much for data-structure synchronization. |
5873b8a9 PM |
737 | * |
738 | * See the description of call_rcu() for more detailed information on | |
739 | * memory ordering guarantees. | |
740 | */ | |
741 | void call_rcu_tasks(struct rcu_head *rhp, rcu_callback_t func) | |
742 | { | |
743 | call_rcu_tasks_generic(rhp, func, &rcu_tasks); | |
744 | } | |
745 | EXPORT_SYMBOL_GPL(call_rcu_tasks); | |
746 | ||
747 | /** | |
748 | * synchronize_rcu_tasks - wait until an rcu-tasks grace period has elapsed. | |
749 | * | |
750 | * Control will return to the caller some time after a full rcu-tasks | |
751 | * grace period has elapsed, in other words after all currently | |
752 | * executing rcu-tasks read-side critical sections have elapsed. These | |
753 | * read-side critical sections are delimited by calls to schedule(), | |
754 | * cond_resched_tasks_rcu_qs(), idle execution, userspace execution, calls | |
755 | * to synchronize_rcu_tasks(), and (in theory, anyway) cond_resched(). | |
756 | * | |
757 | * This is a very specialized primitive, intended only for a few uses in | |
758 | * tracing and other situations requiring manipulation of function | |
759 | * preambles and profiling hooks. The synchronize_rcu_tasks() function | |
760 | * is not (yet) intended for heavy use from multiple CPUs. | |
761 | * | |
762 | * See the description of synchronize_rcu() for more detailed information | |
763 | * on memory ordering guarantees. | |
764 | */ | |
765 | void synchronize_rcu_tasks(void) | |
766 | { | |
767 | synchronize_rcu_tasks_generic(&rcu_tasks); | |
768 | } | |
769 | EXPORT_SYMBOL_GPL(synchronize_rcu_tasks); | |
770 | ||
771 | /** | |
772 | * rcu_barrier_tasks - Wait for in-flight call_rcu_tasks() callbacks. | |
773 | * | |
774 | * Although the current implementation is guaranteed to wait, it is not | |
775 | * obligated to, for example, if there are no pending callbacks. | |
776 | */ | |
777 | void rcu_barrier_tasks(void) | |
778 | { | |
ce9b1c66 | 779 | rcu_barrier_tasks_generic(&rcu_tasks); |
5873b8a9 PM |
780 | } |
781 | EXPORT_SYMBOL_GPL(rcu_barrier_tasks); | |
782 | ||
783 | static int __init rcu_spawn_tasks_kthread(void) | |
784 | { | |
cafafd67 | 785 | cblist_init_generic(&rcu_tasks); |
4fe192df | 786 | rcu_tasks.gp_sleep = HZ / 10; |
75dc2da5 | 787 | rcu_tasks.init_fract = HZ / 10; |
e4fe5dd6 PM |
788 | rcu_tasks.pregp_func = rcu_tasks_pregp_step; |
789 | rcu_tasks.pertask_func = rcu_tasks_pertask; | |
790 | rcu_tasks.postscan_func = rcu_tasks_postscan; | |
791 | rcu_tasks.holdouts_func = check_all_holdout_tasks; | |
792 | rcu_tasks.postgp_func = rcu_tasks_postgp; | |
5873b8a9 PM |
793 | rcu_spawn_tasks_kthread_generic(&rcu_tasks); |
794 | return 0; | |
795 | } | |
5873b8a9 | 796 | |
27c0f144 PM |
797 | #if !defined(CONFIG_TINY_RCU) |
798 | void show_rcu_tasks_classic_gp_kthread(void) | |
e21408ce PM |
799 | { |
800 | show_rcu_tasks_generic_gp_kthread(&rcu_tasks, ""); | |
801 | } | |
27c0f144 PM |
802 | EXPORT_SYMBOL_GPL(show_rcu_tasks_classic_gp_kthread); |
803 | #endif // !defined(CONFIG_TINY_RCU) | |
e21408ce | 804 | |
25246fc8 PM |
805 | /* Do the srcu_read_lock() for the above synchronize_srcu(). */ |
806 | void exit_tasks_rcu_start(void) __acquires(&tasks_rcu_exit_srcu) | |
807 | { | |
808 | preempt_disable(); | |
809 | current->rcu_tasks_idx = __srcu_read_lock(&tasks_rcu_exit_srcu); | |
810 | preempt_enable(); | |
811 | } | |
812 | ||
813 | /* Do the srcu_read_unlock() for the above synchronize_srcu(). */ | |
814 | void exit_tasks_rcu_finish(void) __releases(&tasks_rcu_exit_srcu) | |
815 | { | |
816 | struct task_struct *t = current; | |
817 | ||
818 | preempt_disable(); | |
819 | __srcu_read_unlock(&tasks_rcu_exit_srcu, t->rcu_tasks_idx); | |
820 | preempt_enable(); | |
821 | exit_tasks_rcu_finish_trace(t); | |
822 | } | |
823 | ||
e21408ce | 824 | #else /* #ifdef CONFIG_TASKS_RCU */ |
25246fc8 PM |
825 | void exit_tasks_rcu_start(void) { } |
826 | void exit_tasks_rcu_finish(void) { exit_tasks_rcu_finish_trace(current); } | |
e21408ce | 827 | #endif /* #else #ifdef CONFIG_TASKS_RCU */ |
c84aad76 PM |
828 | |
829 | #ifdef CONFIG_TASKS_RUDE_RCU | |
830 | ||
831 | //////////////////////////////////////////////////////////////////////// | |
832 | // | |
833 | // "Rude" variant of Tasks RCU, inspired by Steve Rostedt's trick of | |
834 | // passing an empty function to schedule_on_each_cpu(). This approach | |
e4be1f44 PM |
835 | // provides an asynchronous call_rcu_tasks_rude() API and batching of |
836 | // concurrent calls to the synchronous synchronize_rcu_tasks_rude() API. | |
9fc98e31 PM |
837 | // This invokes schedule_on_each_cpu() in order to send IPIs far and wide |
838 | // and induces otherwise unnecessary context switches on all online CPUs, | |
839 | // whether idle or not. | |
840 | // | |
841 | // Callback handling is provided by the rcu_tasks_kthread() function. | |
842 | // | |
843 | // Ordering is provided by the scheduler's context-switch code. | |
c84aad76 PM |
844 | |
845 | // Empty function to allow workqueues to force a context switch. | |
846 | static void rcu_tasks_be_rude(struct work_struct *work) | |
847 | { | |
848 | } | |
849 | ||
850 | // Wait for one rude RCU-tasks grace period. | |
851 | static void rcu_tasks_rude_wait_gp(struct rcu_tasks *rtp) | |
852 | { | |
238dbce3 | 853 | rtp->n_ipis += cpumask_weight(cpu_online_mask); |
c84aad76 PM |
854 | schedule_on_each_cpu(rcu_tasks_be_rude); |
855 | } | |
856 | ||
857 | void call_rcu_tasks_rude(struct rcu_head *rhp, rcu_callback_t func); | |
c97d12a6 PM |
858 | DEFINE_RCU_TASKS(rcu_tasks_rude, rcu_tasks_rude_wait_gp, call_rcu_tasks_rude, |
859 | "RCU Tasks Rude"); | |
c84aad76 PM |
860 | |
861 | /** | |
862 | * call_rcu_tasks_rude() - Queue a callback rude task-based grace period | |
863 | * @rhp: structure to be used for queueing the RCU updates. | |
864 | * @func: actual callback function to be invoked after the grace period | |
865 | * | |
866 | * The callback function will be invoked some time after a full grace | |
867 | * period elapses, in other words after all currently executing RCU | |
868 | * read-side critical sections have completed. call_rcu_tasks_rude() | |
869 | * assumes that the read-side critical sections end at context switch, | |
8af9e2c7 | 870 | * cond_resched_tasks_rcu_qs(), or transition to usermode execution (as |
a6517e9c NU |
871 | * usermode execution is schedulable). As such, there are no read-side |
872 | * primitives analogous to rcu_read_lock() and rcu_read_unlock() because | |
873 | * this primitive is intended to determine that all tasks have passed | |
874 | * through a safe state, not so much for data-structure synchronization. | |
c84aad76 PM |
875 | * |
876 | * See the description of call_rcu() for more detailed information on | |
877 | * memory ordering guarantees. | |
878 | */ | |
879 | void call_rcu_tasks_rude(struct rcu_head *rhp, rcu_callback_t func) | |
880 | { | |
881 | call_rcu_tasks_generic(rhp, func, &rcu_tasks_rude); | |
882 | } | |
883 | EXPORT_SYMBOL_GPL(call_rcu_tasks_rude); | |
884 | ||
885 | /** | |
886 | * synchronize_rcu_tasks_rude - wait for a rude rcu-tasks grace period | |
887 | * | |
888 | * Control will return to the caller some time after a rude rcu-tasks | |
889 | * grace period has elapsed, in other words after all currently | |
890 | * executing rcu-tasks read-side critical sections have elapsed. These | |
891 | * read-side critical sections are delimited by calls to schedule(), | |
a6517e9c NU |
892 | * cond_resched_tasks_rcu_qs(), userspace execution (which is a schedulable |
893 | * context), and (in theory, anyway) cond_resched(). | |
c84aad76 PM |
894 | * |
895 | * This is a very specialized primitive, intended only for a few uses in | |
896 | * tracing and other situations requiring manipulation of function preambles | |
897 | * and profiling hooks. The synchronize_rcu_tasks_rude() function is not | |
898 | * (yet) intended for heavy use from multiple CPUs. | |
899 | * | |
900 | * See the description of synchronize_rcu() for more detailed information | |
901 | * on memory ordering guarantees. | |
902 | */ | |
903 | void synchronize_rcu_tasks_rude(void) | |
904 | { | |
905 | synchronize_rcu_tasks_generic(&rcu_tasks_rude); | |
906 | } | |
907 | EXPORT_SYMBOL_GPL(synchronize_rcu_tasks_rude); | |
908 | ||
909 | /** | |
910 | * rcu_barrier_tasks_rude - Wait for in-flight call_rcu_tasks_rude() callbacks. | |
911 | * | |
912 | * Although the current implementation is guaranteed to wait, it is not | |
913 | * obligated to, for example, if there are no pending callbacks. | |
914 | */ | |
915 | void rcu_barrier_tasks_rude(void) | |
916 | { | |
ce9b1c66 | 917 | rcu_barrier_tasks_generic(&rcu_tasks_rude); |
c84aad76 PM |
918 | } |
919 | EXPORT_SYMBOL_GPL(rcu_barrier_tasks_rude); | |
920 | ||
921 | static int __init rcu_spawn_tasks_rude_kthread(void) | |
922 | { | |
cafafd67 | 923 | cblist_init_generic(&rcu_tasks_rude); |
4fe192df | 924 | rcu_tasks_rude.gp_sleep = HZ / 10; |
c84aad76 PM |
925 | rcu_spawn_tasks_kthread_generic(&rcu_tasks_rude); |
926 | return 0; | |
927 | } | |
c84aad76 | 928 | |
27c0f144 PM |
929 | #if !defined(CONFIG_TINY_RCU) |
930 | void show_rcu_tasks_rude_gp_kthread(void) | |
e21408ce PM |
931 | { |
932 | show_rcu_tasks_generic_gp_kthread(&rcu_tasks_rude, ""); | |
933 | } | |
27c0f144 PM |
934 | EXPORT_SYMBOL_GPL(show_rcu_tasks_rude_gp_kthread); |
935 | #endif // !defined(CONFIG_TINY_RCU) | |
936 | #endif /* #ifdef CONFIG_TASKS_RUDE_RCU */ | |
d5f177d3 PM |
937 | |
938 | //////////////////////////////////////////////////////////////////////// | |
939 | // | |
940 | // Tracing variant of Tasks RCU. This variant is designed to be used | |
941 | // to protect tracing hooks, including those of BPF. This variant | |
942 | // therefore: | |
943 | // | |
944 | // 1. Has explicit read-side markers to allow finite grace periods | |
945 | // in the face of in-kernel loops for PREEMPT=n builds. | |
946 | // | |
947 | // 2. Protects code in the idle loop, exception entry/exit, and | |
948 | // CPU-hotplug code paths, similar to the capabilities of SRCU. | |
949 | // | |
c4f113ac | 950 | // 3. Avoids expensive read-side instructions, having overhead similar |
d5f177d3 PM |
951 | // to that of Preemptible RCU. |
952 | // | |
953 | // There are of course downsides. The grace-period code can send IPIs to | |
954 | // CPUs, even when those CPUs are in the idle loop or in nohz_full userspace. | |
955 | // It is necessary to scan the full tasklist, much as for Tasks RCU. There | |
956 | // is a single callback queue guarded by a single lock, again, much as for | |
957 | // Tasks RCU. If needed, these downsides can be at least partially remedied. | |
958 | // | |
959 | // Perhaps most important, this variant of RCU does not affect the vanilla | |
960 | // flavors, rcu_preempt and rcu_sched. The fact that RCU Tasks Trace | |
961 | // readers can operate from idle, offline, and exception entry/exit in no | |
962 | // way allows rcu_preempt and rcu_sched readers to also do so. | |
a434dd10 PM |
963 | // |
964 | // The implementation uses rcu_tasks_wait_gp(), which relies on function | |
965 | // pointers in the rcu_tasks structure. The rcu_spawn_tasks_trace_kthread() | |
966 | // function sets these function pointers up so that rcu_tasks_wait_gp() | |
967 | // invokes these functions in this order: | |
968 | // | |
969 | // rcu_tasks_trace_pregp_step(): | |
970 | // Initialize the count of readers and block CPU-hotplug operations. | |
971 | // rcu_tasks_trace_pertask(), invoked on every non-idle task: | |
972 | // Initialize per-task state and attempt to identify an immediate | |
973 | // quiescent state for that task, or, failing that, attempt to | |
974 | // set that task's .need_qs flag so that task's next outermost | |
975 | // rcu_read_unlock_trace() will report the quiescent state (in which | |
976 | // case the count of readers is incremented). If both attempts fail, | |
45f4b4a2 PM |
977 | // the task is added to a "holdout" list. Note that IPIs are used |
978 | // to invoke trc_read_check_handler() in the context of running tasks | |
979 | // in order to avoid ordering overhead on common-case shared-variable | |
980 | // accessses. | |
a434dd10 PM |
981 | // rcu_tasks_trace_postscan(): |
982 | // Initialize state and attempt to identify an immediate quiescent | |
983 | // state as above (but only for idle tasks), unblock CPU-hotplug | |
984 | // operations, and wait for an RCU grace period to avoid races with | |
985 | // tasks that are in the process of exiting. | |
986 | // check_all_holdout_tasks_trace(), repeatedly until holdout list is empty: | |
987 | // Scans the holdout list, attempting to identify a quiescent state | |
988 | // for each task on the list. If there is a quiescent state, the | |
989 | // corresponding task is removed from the holdout list. | |
990 | // rcu_tasks_trace_postgp(): | |
991 | // Wait for the count of readers do drop to zero, reporting any stalls. | |
992 | // Also execute full memory barriers to maintain ordering with code | |
993 | // executing after the grace period. | |
994 | // | |
995 | // The exit_tasks_rcu_finish_trace() synchronizes with exiting tasks. | |
996 | // | |
997 | // Pre-grace-period update-side code is ordered before the grace | |
998 | // period via the ->cbs_lock and barriers in rcu_tasks_kthread(). | |
999 | // Pre-grace-period read-side code is ordered before the grace period by | |
1000 | // atomic_dec_and_test() of the count of readers (for IPIed readers) and by | |
1001 | // scheduler context-switch ordering (for locked-down non-running readers). | |
d5f177d3 PM |
1002 | |
1003 | // The lockdep state must be outside of #ifdef to be useful. | |
1004 | #ifdef CONFIG_DEBUG_LOCK_ALLOC | |
1005 | static struct lock_class_key rcu_lock_trace_key; | |
1006 | struct lockdep_map rcu_trace_lock_map = | |
1007 | STATIC_LOCKDEP_MAP_INIT("rcu_read_lock_trace", &rcu_lock_trace_key); | |
1008 | EXPORT_SYMBOL_GPL(rcu_trace_lock_map); | |
1009 | #endif /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */ | |
1010 | ||
1011 | #ifdef CONFIG_TASKS_TRACE_RCU | |
1012 | ||
30d8aa51 PM |
1013 | static atomic_t trc_n_readers_need_end; // Number of waited-for readers. |
1014 | static DECLARE_WAIT_QUEUE_HEAD(trc_wait); // List of holdout tasks. | |
d5f177d3 PM |
1015 | |
1016 | // Record outstanding IPIs to each CPU. No point in sending two... | |
1017 | static DEFINE_PER_CPU(bool, trc_ipi_to_cpu); | |
1018 | ||
40471509 PM |
1019 | // The number of detections of task quiescent state relying on |
1020 | // heavyweight readers executing explicit memory barriers. | |
6731da9e PM |
1021 | static unsigned long n_heavy_reader_attempts; |
1022 | static unsigned long n_heavy_reader_updates; | |
1023 | static unsigned long n_heavy_reader_ofl_updates; | |
40471509 | 1024 | |
b0afa0f0 PM |
1025 | void call_rcu_tasks_trace(struct rcu_head *rhp, rcu_callback_t func); |
1026 | DEFINE_RCU_TASKS(rcu_tasks_trace, rcu_tasks_wait_gp, call_rcu_tasks_trace, | |
1027 | "RCU Tasks Trace"); | |
1028 | ||
b38f57c1 PM |
1029 | /* |
1030 | * This irq_work handler allows rcu_read_unlock_trace() to be invoked | |
1031 | * while the scheduler locks are held. | |
1032 | */ | |
1033 | static void rcu_read_unlock_iw(struct irq_work *iwp) | |
1034 | { | |
1035 | wake_up(&trc_wait); | |
1036 | } | |
1037 | static DEFINE_IRQ_WORK(rcu_tasks_trace_iw, rcu_read_unlock_iw); | |
1038 | ||
d5f177d3 | 1039 | /* If we are the last reader, wake up the grace-period kthread. */ |
a5c071cc | 1040 | void rcu_read_unlock_trace_special(struct task_struct *t) |
d5f177d3 | 1041 | { |
f8ab3fad | 1042 | int nq = READ_ONCE(t->trc_reader_special.b.need_qs); |
276c4104 | 1043 | |
9ae58d7b PM |
1044 | if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB) && |
1045 | t->trc_reader_special.b.need_mb) | |
276c4104 PM |
1046 | smp_mb(); // Pairs with update-side barriers. |
1047 | // Update .need_qs before ->trc_reader_nesting for irq/NMI handlers. | |
1048 | if (nq) | |
1049 | WRITE_ONCE(t->trc_reader_special.b.need_qs, false); | |
a5c071cc | 1050 | WRITE_ONCE(t->trc_reader_nesting, 0); |
276c4104 | 1051 | if (nq && atomic_dec_and_test(&trc_n_readers_need_end)) |
b38f57c1 | 1052 | irq_work_queue(&rcu_tasks_trace_iw); |
d5f177d3 PM |
1053 | } |
1054 | EXPORT_SYMBOL_GPL(rcu_read_unlock_trace_special); | |
1055 | ||
1056 | /* Add a task to the holdout list, if it is not already on the list. */ | |
1057 | static void trc_add_holdout(struct task_struct *t, struct list_head *bhp) | |
1058 | { | |
1059 | if (list_empty(&t->trc_holdout_list)) { | |
1060 | get_task_struct(t); | |
1061 | list_add(&t->trc_holdout_list, bhp); | |
1062 | } | |
1063 | } | |
1064 | ||
1065 | /* Remove a task from the holdout list, if it is in fact present. */ | |
1066 | static void trc_del_holdout(struct task_struct *t) | |
1067 | { | |
1068 | if (!list_empty(&t->trc_holdout_list)) { | |
1069 | list_del_init(&t->trc_holdout_list); | |
1070 | put_task_struct(t); | |
1071 | } | |
1072 | } | |
1073 | ||
1074 | /* IPI handler to check task state. */ | |
1075 | static void trc_read_check_handler(void *t_in) | |
1076 | { | |
1077 | struct task_struct *t = current; | |
1078 | struct task_struct *texp = t_in; | |
1079 | ||
1080 | // If the task is no longer running on this CPU, leave. | |
1081 | if (unlikely(texp != t)) { | |
d5f177d3 PM |
1082 | goto reset_ipi; // Already on holdout list, so will check later. |
1083 | } | |
1084 | ||
1085 | // If the task is not in a read-side critical section, and | |
1086 | // if this is the last reader, awaken the grace-period kthread. | |
bdb0cca0 | 1087 | if (likely(!READ_ONCE(t->trc_reader_nesting))) { |
d5f177d3 PM |
1088 | WRITE_ONCE(t->trc_reader_checked, true); |
1089 | goto reset_ipi; | |
1090 | } | |
ba3a86e4 | 1091 | // If we are racing with an rcu_read_unlock_trace(), try again later. |
96017bf9 | 1092 | if (unlikely(READ_ONCE(t->trc_reader_nesting) < 0)) |
ba3a86e4 | 1093 | goto reset_ipi; |
d5f177d3 PM |
1094 | WRITE_ONCE(t->trc_reader_checked, true); |
1095 | ||
1096 | // Get here if the task is in a read-side critical section. Set | |
1097 | // its state so that it will awaken the grace-period kthread upon | |
1098 | // exit from that critical section. | |
96017bf9 | 1099 | atomic_inc(&trc_n_readers_need_end); // One more to wait on. |
f8ab3fad | 1100 | WARN_ON_ONCE(READ_ONCE(t->trc_reader_special.b.need_qs)); |
276c4104 | 1101 | WRITE_ONCE(t->trc_reader_special.b.need_qs, true); |
d5f177d3 PM |
1102 | |
1103 | reset_ipi: | |
1104 | // Allow future IPIs to be sent on CPU and for task. | |
1105 | // Also order this IPI handler against any later manipulations of | |
1106 | // the intended task. | |
8211e922 | 1107 | smp_store_release(per_cpu_ptr(&trc_ipi_to_cpu, smp_processor_id()), false); // ^^^ |
d5f177d3 PM |
1108 | smp_store_release(&texp->trc_ipi_to_cpu, -1); // ^^^ |
1109 | } | |
1110 | ||
1111 | /* Callback function for scheduler to check locked-down task. */ | |
9b3c4ab3 | 1112 | static int trc_inspect_reader(struct task_struct *t, void *arg) |
d5f177d3 | 1113 | { |
7d0c9c50 | 1114 | int cpu = task_cpu(t); |
18f08e75 | 1115 | int nesting; |
7e3b70e0 | 1116 | bool ofl = cpu_is_offline(cpu); |
7d0c9c50 PM |
1117 | |
1118 | if (task_curr(t)) { | |
30d8aa51 | 1119 | WARN_ON_ONCE(ofl && !is_idle_task(t)); |
7e3b70e0 | 1120 | |
7d0c9c50 | 1121 | // If no chance of heavyweight readers, do it the hard way. |
7e3b70e0 | 1122 | if (!ofl && !IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB)) |
9b3c4ab3 | 1123 | return -EINVAL; |
7d0c9c50 PM |
1124 | |
1125 | // If heavyweight readers are enabled on the remote task, | |
1126 | // we can inspect its state despite its currently running. | |
1127 | // However, we cannot safely change its state. | |
40471509 | 1128 | n_heavy_reader_attempts++; |
7e3b70e0 PM |
1129 | if (!ofl && // Check for "running" idle tasks on offline CPUs. |
1130 | !rcu_dynticks_zero_in_eqs(cpu, &t->trc_reader_nesting)) | |
9b3c4ab3 | 1131 | return -EINVAL; // No quiescent state, do it the hard way. |
40471509 | 1132 | n_heavy_reader_updates++; |
edf3775f PM |
1133 | if (ofl) |
1134 | n_heavy_reader_ofl_updates++; | |
18f08e75 | 1135 | nesting = 0; |
7d0c9c50 | 1136 | } else { |
bdb0cca0 | 1137 | // The task is not running, so C-language access is safe. |
18f08e75 | 1138 | nesting = t->trc_reader_nesting; |
7d0c9c50 | 1139 | } |
d5f177d3 | 1140 | |
18f08e75 PM |
1141 | // If not exiting a read-side critical section, mark as checked |
1142 | // so that the grace-period kthread will remove it from the | |
1143 | // holdout list. | |
1144 | t->trc_reader_checked = nesting >= 0; | |
1145 | if (nesting <= 0) | |
6fedc280 | 1146 | return nesting ? -EINVAL : 0; // If in QS, done, otherwise try again later. |
7d0c9c50 PM |
1147 | |
1148 | // The task is in a read-side critical section, so set up its | |
1149 | // state so that it will awaken the grace-period kthread upon exit | |
1150 | // from that critical section. | |
1151 | atomic_inc(&trc_n_readers_need_end); // One more to wait on. | |
f8ab3fad | 1152 | WARN_ON_ONCE(READ_ONCE(t->trc_reader_special.b.need_qs)); |
7d0c9c50 | 1153 | WRITE_ONCE(t->trc_reader_special.b.need_qs, true); |
9b3c4ab3 | 1154 | return 0; |
d5f177d3 PM |
1155 | } |
1156 | ||
1157 | /* Attempt to extract the state for the specified task. */ | |
1158 | static void trc_wait_for_one_reader(struct task_struct *t, | |
1159 | struct list_head *bhp) | |
1160 | { | |
1161 | int cpu; | |
1162 | ||
1163 | // If a previous IPI is still in flight, let it complete. | |
1164 | if (smp_load_acquire(&t->trc_ipi_to_cpu) != -1) // Order IPI | |
1165 | return; | |
1166 | ||
1167 | // The current task had better be in a quiescent state. | |
1168 | if (t == current) { | |
1169 | t->trc_reader_checked = true; | |
bdb0cca0 | 1170 | WARN_ON_ONCE(READ_ONCE(t->trc_reader_nesting)); |
d5f177d3 PM |
1171 | return; |
1172 | } | |
1173 | ||
1174 | // Attempt to nail down the task for inspection. | |
1175 | get_task_struct(t); | |
9b3c4ab3 | 1176 | if (!task_call_func(t, trc_inspect_reader, NULL)) { |
d5f177d3 PM |
1177 | put_task_struct(t); |
1178 | return; | |
1179 | } | |
1180 | put_task_struct(t); | |
1181 | ||
45f4b4a2 PM |
1182 | // If this task is not yet on the holdout list, then we are in |
1183 | // an RCU read-side critical section. Otherwise, the invocation of | |
d0a85858 | 1184 | // trc_add_holdout() that added it to the list did the necessary |
45f4b4a2 PM |
1185 | // get_task_struct(). Either way, the task cannot be freed out |
1186 | // from under this code. | |
1187 | ||
d5f177d3 PM |
1188 | // If currently running, send an IPI, either way, add to list. |
1189 | trc_add_holdout(t, bhp); | |
574de876 PM |
1190 | if (task_curr(t) && |
1191 | time_after(jiffies + 1, rcu_tasks_trace.gp_start + rcu_task_ipi_delay)) { | |
d5f177d3 PM |
1192 | // The task is currently running, so try IPIing it. |
1193 | cpu = task_cpu(t); | |
1194 | ||
1195 | // If there is already an IPI outstanding, let it happen. | |
1196 | if (per_cpu(trc_ipi_to_cpu, cpu) || t->trc_ipi_to_cpu >= 0) | |
1197 | return; | |
1198 | ||
d5f177d3 PM |
1199 | per_cpu(trc_ipi_to_cpu, cpu) = true; |
1200 | t->trc_ipi_to_cpu = cpu; | |
238dbce3 | 1201 | rcu_tasks_trace.n_ipis++; |
96017bf9 | 1202 | if (smp_call_function_single(cpu, trc_read_check_handler, t, 0)) { |
d5f177d3 PM |
1203 | // Just in case there is some other reason for |
1204 | // failure than the target CPU being offline. | |
46aa886c NU |
1205 | WARN_ONCE(1, "%s(): smp_call_function_single() failed for CPU: %d\n", |
1206 | __func__, cpu); | |
7e0669c3 | 1207 | rcu_tasks_trace.n_ipis_fails++; |
d5f177d3 | 1208 | per_cpu(trc_ipi_to_cpu, cpu) = false; |
46aa886c | 1209 | t->trc_ipi_to_cpu = -1; |
d5f177d3 PM |
1210 | } |
1211 | } | |
1212 | } | |
1213 | ||
1214 | /* Initialize for a new RCU-tasks-trace grace period. */ | |
1215 | static void rcu_tasks_trace_pregp_step(void) | |
1216 | { | |
1217 | int cpu; | |
1218 | ||
d5f177d3 PM |
1219 | // Allow for fast-acting IPIs. |
1220 | atomic_set(&trc_n_readers_need_end, 1); | |
1221 | ||
1222 | // There shouldn't be any old IPIs, but... | |
1223 | for_each_possible_cpu(cpu) | |
1224 | WARN_ON_ONCE(per_cpu(trc_ipi_to_cpu, cpu)); | |
81b4a7bc PM |
1225 | |
1226 | // Disable CPU hotplug across the tasklist scan. | |
1227 | // This also waits for all readers in CPU-hotplug code paths. | |
1228 | cpus_read_lock(); | |
d5f177d3 PM |
1229 | } |
1230 | ||
1231 | /* Do first-round processing for the specified task. */ | |
1232 | static void rcu_tasks_trace_pertask(struct task_struct *t, | |
1233 | struct list_head *hop) | |
1234 | { | |
1b04fa99 URS |
1235 | // During early boot when there is only the one boot CPU, there |
1236 | // is no idle task for the other CPUs. Just return. | |
1237 | if (unlikely(t == NULL)) | |
1238 | return; | |
1239 | ||
276c4104 | 1240 | WRITE_ONCE(t->trc_reader_special.b.need_qs, false); |
43766c3e | 1241 | WRITE_ONCE(t->trc_reader_checked, false); |
d5f177d3 PM |
1242 | t->trc_ipi_to_cpu = -1; |
1243 | trc_wait_for_one_reader(t, hop); | |
1244 | } | |
1245 | ||
9796e1ae PM |
1246 | /* |
1247 | * Do intermediate processing between task and holdout scans and | |
1248 | * pick up the idle tasks. | |
1249 | */ | |
1250 | static void rcu_tasks_trace_postscan(struct list_head *hop) | |
d5f177d3 | 1251 | { |
9796e1ae PM |
1252 | int cpu; |
1253 | ||
1254 | for_each_possible_cpu(cpu) | |
1255 | rcu_tasks_trace_pertask(idle_task(cpu), hop); | |
1256 | ||
81b4a7bc PM |
1257 | // Re-enable CPU hotplug now that the tasklist scan has completed. |
1258 | cpus_read_unlock(); | |
1259 | ||
d5f177d3 PM |
1260 | // Wait for late-stage exiting tasks to finish exiting. |
1261 | // These might have passed the call to exit_tasks_rcu_finish(). | |
1262 | synchronize_rcu(); | |
1263 | // Any tasks that exit after this point will set ->trc_reader_checked. | |
1264 | } | |
1265 | ||
65b629e7 NU |
1266 | /* Communicate task state back to the RCU tasks trace stall warning request. */ |
1267 | struct trc_stall_chk_rdr { | |
1268 | int nesting; | |
1269 | int ipi_to_cpu; | |
1270 | u8 needqs; | |
1271 | }; | |
1272 | ||
1273 | static int trc_check_slow_task(struct task_struct *t, void *arg) | |
1274 | { | |
1275 | struct trc_stall_chk_rdr *trc_rdrp = arg; | |
1276 | ||
1277 | if (task_curr(t)) | |
1278 | return false; // It is running, so decline to inspect it. | |
1279 | trc_rdrp->nesting = READ_ONCE(t->trc_reader_nesting); | |
1280 | trc_rdrp->ipi_to_cpu = READ_ONCE(t->trc_ipi_to_cpu); | |
1281 | trc_rdrp->needqs = READ_ONCE(t->trc_reader_special.b.need_qs); | |
1282 | return true; | |
1283 | } | |
1284 | ||
4593e772 PM |
1285 | /* Show the state of a task stalling the current RCU tasks trace GP. */ |
1286 | static void show_stalled_task_trace(struct task_struct *t, bool *firstreport) | |
1287 | { | |
1288 | int cpu; | |
65b629e7 NU |
1289 | struct trc_stall_chk_rdr trc_rdr; |
1290 | bool is_idle_tsk = is_idle_task(t); | |
4593e772 PM |
1291 | |
1292 | if (*firstreport) { | |
1293 | pr_err("INFO: rcu_tasks_trace detected stalls on tasks:\n"); | |
1294 | *firstreport = false; | |
1295 | } | |
4593e772 | 1296 | cpu = task_cpu(t); |
65b629e7 NU |
1297 | if (!task_call_func(t, trc_check_slow_task, &trc_rdr)) |
1298 | pr_alert("P%d: %c\n", | |
1299 | t->pid, | |
1300 | ".i"[is_idle_tsk]); | |
1301 | else | |
1302 | pr_alert("P%d: %c%c%c nesting: %d%c cpu: %d\n", | |
1303 | t->pid, | |
1304 | ".I"[trc_rdr.ipi_to_cpu >= 0], | |
1305 | ".i"[is_idle_tsk], | |
1306 | ".N"[cpu >= 0 && tick_nohz_full_cpu(cpu)], | |
1307 | trc_rdr.nesting, | |
1308 | " N"[!!trc_rdr.needqs], | |
1309 | cpu); | |
4593e772 PM |
1310 | sched_show_task(t); |
1311 | } | |
1312 | ||
1313 | /* List stalled IPIs for RCU tasks trace. */ | |
1314 | static void show_stalled_ipi_trace(void) | |
1315 | { | |
1316 | int cpu; | |
1317 | ||
1318 | for_each_possible_cpu(cpu) | |
1319 | if (per_cpu(trc_ipi_to_cpu, cpu)) | |
1320 | pr_alert("\tIPI outstanding to CPU %d\n", cpu); | |
1321 | } | |
1322 | ||
d5f177d3 PM |
1323 | /* Do one scan of the holdout list. */ |
1324 | static void check_all_holdout_tasks_trace(struct list_head *hop, | |
4593e772 | 1325 | bool needreport, bool *firstreport) |
d5f177d3 PM |
1326 | { |
1327 | struct task_struct *g, *t; | |
1328 | ||
81b4a7bc PM |
1329 | // Disable CPU hotplug across the holdout list scan. |
1330 | cpus_read_lock(); | |
1331 | ||
d5f177d3 PM |
1332 | list_for_each_entry_safe(t, g, hop, trc_holdout_list) { |
1333 | // If safe and needed, try to check the current task. | |
1334 | if (READ_ONCE(t->trc_ipi_to_cpu) == -1 && | |
1335 | !READ_ONCE(t->trc_reader_checked)) | |
1336 | trc_wait_for_one_reader(t, hop); | |
1337 | ||
1338 | // If check succeeded, remove this task from the list. | |
f5dbc594 PM |
1339 | if (smp_load_acquire(&t->trc_ipi_to_cpu) == -1 && |
1340 | READ_ONCE(t->trc_reader_checked)) | |
d5f177d3 | 1341 | trc_del_holdout(t); |
4593e772 PM |
1342 | else if (needreport) |
1343 | show_stalled_task_trace(t, firstreport); | |
1344 | } | |
81b4a7bc PM |
1345 | |
1346 | // Re-enable CPU hotplug now that the holdout list scan has completed. | |
1347 | cpus_read_unlock(); | |
1348 | ||
4593e772 | 1349 | if (needreport) { |
89401176 | 1350 | if (*firstreport) |
4593e772 PM |
1351 | pr_err("INFO: rcu_tasks_trace detected stalls? (Late IPI?)\n"); |
1352 | show_stalled_ipi_trace(); | |
d5f177d3 PM |
1353 | } |
1354 | } | |
1355 | ||
cbe0d8d9 PM |
1356 | static void rcu_tasks_trace_empty_fn(void *unused) |
1357 | { | |
1358 | } | |
1359 | ||
d5f177d3 | 1360 | /* Wait for grace period to complete and provide ordering. */ |
af051ca4 | 1361 | static void rcu_tasks_trace_postgp(struct rcu_tasks *rtp) |
d5f177d3 | 1362 | { |
cbe0d8d9 | 1363 | int cpu; |
4593e772 PM |
1364 | bool firstreport; |
1365 | struct task_struct *g, *t; | |
1366 | LIST_HEAD(holdouts); | |
1367 | long ret; | |
1368 | ||
cbe0d8d9 PM |
1369 | // Wait for any lingering IPI handlers to complete. Note that |
1370 | // if a CPU has gone offline or transitioned to userspace in the | |
1371 | // meantime, all IPI handlers should have been drained beforehand. | |
1372 | // Yes, this assumes that CPUs process IPIs in order. If that ever | |
1373 | // changes, there will need to be a recheck and/or timed wait. | |
1374 | for_each_online_cpu(cpu) | |
f5dbc594 | 1375 | if (WARN_ON_ONCE(smp_load_acquire(per_cpu_ptr(&trc_ipi_to_cpu, cpu)))) |
cbe0d8d9 PM |
1376 | smp_call_function_single(cpu, rcu_tasks_trace_empty_fn, NULL, 1); |
1377 | ||
d5f177d3 PM |
1378 | // Remove the safety count. |
1379 | smp_mb__before_atomic(); // Order vs. earlier atomics | |
1380 | atomic_dec(&trc_n_readers_need_end); | |
1381 | smp_mb__after_atomic(); // Order vs. later atomics | |
1382 | ||
1383 | // Wait for readers. | |
af051ca4 | 1384 | set_tasks_gp_state(rtp, RTGS_WAIT_READERS); |
4593e772 PM |
1385 | for (;;) { |
1386 | ret = wait_event_idle_exclusive_timeout( | |
1387 | trc_wait, | |
1388 | atomic_read(&trc_n_readers_need_end) == 0, | |
1389 | READ_ONCE(rcu_task_stall_timeout)); | |
1390 | if (ret) | |
1391 | break; // Count reached zero. | |
af051ca4 | 1392 | // Stall warning time, so make a list of the offenders. |
f747c7e1 | 1393 | rcu_read_lock(); |
4593e772 | 1394 | for_each_process_thread(g, t) |
276c4104 | 1395 | if (READ_ONCE(t->trc_reader_special.b.need_qs)) |
4593e772 | 1396 | trc_add_holdout(t, &holdouts); |
f747c7e1 | 1397 | rcu_read_unlock(); |
4593e772 | 1398 | firstreport = true; |
592031cc PM |
1399 | list_for_each_entry_safe(t, g, &holdouts, trc_holdout_list) { |
1400 | if (READ_ONCE(t->trc_reader_special.b.need_qs)) | |
4593e772 | 1401 | show_stalled_task_trace(t, &firstreport); |
592031cc PM |
1402 | trc_del_holdout(t); // Release task_struct reference. |
1403 | } | |
4593e772 PM |
1404 | if (firstreport) |
1405 | pr_err("INFO: rcu_tasks_trace detected stalls? (Counter/taskslist mismatch?)\n"); | |
1406 | show_stalled_ipi_trace(); | |
1407 | pr_err("\t%d holdouts\n", atomic_read(&trc_n_readers_need_end)); | |
1408 | } | |
d5f177d3 | 1409 | smp_mb(); // Caller's code must be ordered after wakeup. |
43766c3e | 1410 | // Pairs with pretty much every ordering primitive. |
d5f177d3 PM |
1411 | } |
1412 | ||
1413 | /* Report any needed quiescent state for this exiting task. */ | |
25246fc8 | 1414 | static void exit_tasks_rcu_finish_trace(struct task_struct *t) |
d5f177d3 PM |
1415 | { |
1416 | WRITE_ONCE(t->trc_reader_checked, true); | |
bdb0cca0 | 1417 | WARN_ON_ONCE(READ_ONCE(t->trc_reader_nesting)); |
d5f177d3 | 1418 | WRITE_ONCE(t->trc_reader_nesting, 0); |
276c4104 | 1419 | if (WARN_ON_ONCE(READ_ONCE(t->trc_reader_special.b.need_qs))) |
a5c071cc | 1420 | rcu_read_unlock_trace_special(t); |
d5f177d3 PM |
1421 | } |
1422 | ||
d5f177d3 PM |
1423 | /** |
1424 | * call_rcu_tasks_trace() - Queue a callback trace task-based grace period | |
1425 | * @rhp: structure to be used for queueing the RCU updates. | |
1426 | * @func: actual callback function to be invoked after the grace period | |
1427 | * | |
ed42c380 NU |
1428 | * The callback function will be invoked some time after a trace rcu-tasks |
1429 | * grace period elapses, in other words after all currently executing | |
1430 | * trace rcu-tasks read-side critical sections have completed. These | |
1431 | * read-side critical sections are delimited by calls to rcu_read_lock_trace() | |
1432 | * and rcu_read_unlock_trace(). | |
d5f177d3 PM |
1433 | * |
1434 | * See the description of call_rcu() for more detailed information on | |
1435 | * memory ordering guarantees. | |
1436 | */ | |
1437 | void call_rcu_tasks_trace(struct rcu_head *rhp, rcu_callback_t func) | |
1438 | { | |
1439 | call_rcu_tasks_generic(rhp, func, &rcu_tasks_trace); | |
1440 | } | |
1441 | EXPORT_SYMBOL_GPL(call_rcu_tasks_trace); | |
1442 | ||
1443 | /** | |
1444 | * synchronize_rcu_tasks_trace - wait for a trace rcu-tasks grace period | |
1445 | * | |
1446 | * Control will return to the caller some time after a trace rcu-tasks | |
c7dcf810 | 1447 | * grace period has elapsed, in other words after all currently executing |
ed42c380 | 1448 | * trace rcu-tasks read-side critical sections have elapsed. These read-side |
c7dcf810 PM |
1449 | * critical sections are delimited by calls to rcu_read_lock_trace() |
1450 | * and rcu_read_unlock_trace(). | |
d5f177d3 PM |
1451 | * |
1452 | * This is a very specialized primitive, intended only for a few uses in | |
1453 | * tracing and other situations requiring manipulation of function preambles | |
1454 | * and profiling hooks. The synchronize_rcu_tasks_trace() function is not | |
1455 | * (yet) intended for heavy use from multiple CPUs. | |
1456 | * | |
1457 | * See the description of synchronize_rcu() for more detailed information | |
1458 | * on memory ordering guarantees. | |
1459 | */ | |
1460 | void synchronize_rcu_tasks_trace(void) | |
1461 | { | |
1462 | RCU_LOCKDEP_WARN(lock_is_held(&rcu_trace_lock_map), "Illegal synchronize_rcu_tasks_trace() in RCU Tasks Trace read-side critical section"); | |
1463 | synchronize_rcu_tasks_generic(&rcu_tasks_trace); | |
1464 | } | |
1465 | EXPORT_SYMBOL_GPL(synchronize_rcu_tasks_trace); | |
1466 | ||
1467 | /** | |
1468 | * rcu_barrier_tasks_trace - Wait for in-flight call_rcu_tasks_trace() callbacks. | |
1469 | * | |
1470 | * Although the current implementation is guaranteed to wait, it is not | |
1471 | * obligated to, for example, if there are no pending callbacks. | |
1472 | */ | |
1473 | void rcu_barrier_tasks_trace(void) | |
1474 | { | |
ce9b1c66 | 1475 | rcu_barrier_tasks_generic(&rcu_tasks_trace); |
d5f177d3 PM |
1476 | } |
1477 | EXPORT_SYMBOL_GPL(rcu_barrier_tasks_trace); | |
1478 | ||
1479 | static int __init rcu_spawn_tasks_trace_kthread(void) | |
1480 | { | |
cafafd67 | 1481 | cblist_init_generic(&rcu_tasks_trace); |
2393a613 | 1482 | if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB)) { |
4fe192df | 1483 | rcu_tasks_trace.gp_sleep = HZ / 10; |
75dc2da5 | 1484 | rcu_tasks_trace.init_fract = HZ / 10; |
2393a613 | 1485 | } else { |
4fe192df PM |
1486 | rcu_tasks_trace.gp_sleep = HZ / 200; |
1487 | if (rcu_tasks_trace.gp_sleep <= 0) | |
1488 | rcu_tasks_trace.gp_sleep = 1; | |
75dc2da5 | 1489 | rcu_tasks_trace.init_fract = HZ / 200; |
2393a613 PM |
1490 | if (rcu_tasks_trace.init_fract <= 0) |
1491 | rcu_tasks_trace.init_fract = 1; | |
1492 | } | |
d5f177d3 PM |
1493 | rcu_tasks_trace.pregp_func = rcu_tasks_trace_pregp_step; |
1494 | rcu_tasks_trace.pertask_func = rcu_tasks_trace_pertask; | |
1495 | rcu_tasks_trace.postscan_func = rcu_tasks_trace_postscan; | |
1496 | rcu_tasks_trace.holdouts_func = check_all_holdout_tasks_trace; | |
1497 | rcu_tasks_trace.postgp_func = rcu_tasks_trace_postgp; | |
1498 | rcu_spawn_tasks_kthread_generic(&rcu_tasks_trace); | |
1499 | return 0; | |
1500 | } | |
d5f177d3 | 1501 | |
27c0f144 PM |
1502 | #if !defined(CONFIG_TINY_RCU) |
1503 | void show_rcu_tasks_trace_gp_kthread(void) | |
e21408ce | 1504 | { |
40471509 | 1505 | char buf[64]; |
e21408ce | 1506 | |
edf3775f PM |
1507 | sprintf(buf, "N%d h:%lu/%lu/%lu", atomic_read(&trc_n_readers_need_end), |
1508 | data_race(n_heavy_reader_ofl_updates), | |
40471509 PM |
1509 | data_race(n_heavy_reader_updates), |
1510 | data_race(n_heavy_reader_attempts)); | |
e21408ce PM |
1511 | show_rcu_tasks_generic_gp_kthread(&rcu_tasks_trace, buf); |
1512 | } | |
27c0f144 PM |
1513 | EXPORT_SYMBOL_GPL(show_rcu_tasks_trace_gp_kthread); |
1514 | #endif // !defined(CONFIG_TINY_RCU) | |
e21408ce | 1515 | |
d5f177d3 | 1516 | #else /* #ifdef CONFIG_TASKS_TRACE_RCU */ |
25246fc8 | 1517 | static void exit_tasks_rcu_finish_trace(struct task_struct *t) { } |
d5f177d3 | 1518 | #endif /* #else #ifdef CONFIG_TASKS_TRACE_RCU */ |
8fd8ca38 | 1519 | |
8344496e | 1520 | #ifndef CONFIG_TINY_RCU |
e21408ce PM |
1521 | void show_rcu_tasks_gp_kthreads(void) |
1522 | { | |
1523 | show_rcu_tasks_classic_gp_kthread(); | |
1524 | show_rcu_tasks_rude_gp_kthread(); | |
1525 | show_rcu_tasks_trace_gp_kthread(); | |
1526 | } | |
8344496e | 1527 | #endif /* #ifndef CONFIG_TINY_RCU */ |
e21408ce | 1528 | |
bfba7ed0 URS |
1529 | #ifdef CONFIG_PROVE_RCU |
1530 | struct rcu_tasks_test_desc { | |
1531 | struct rcu_head rh; | |
1532 | const char *name; | |
1533 | bool notrun; | |
1534 | }; | |
1535 | ||
1536 | static struct rcu_tasks_test_desc tests[] = { | |
1537 | { | |
1538 | .name = "call_rcu_tasks()", | |
1539 | /* If not defined, the test is skipped. */ | |
1540 | .notrun = !IS_ENABLED(CONFIG_TASKS_RCU), | |
1541 | }, | |
1542 | { | |
1543 | .name = "call_rcu_tasks_rude()", | |
1544 | /* If not defined, the test is skipped. */ | |
1545 | .notrun = !IS_ENABLED(CONFIG_TASKS_RUDE_RCU), | |
1546 | }, | |
1547 | { | |
1548 | .name = "call_rcu_tasks_trace()", | |
1549 | /* If not defined, the test is skipped. */ | |
1550 | .notrun = !IS_ENABLED(CONFIG_TASKS_TRACE_RCU) | |
1551 | } | |
1552 | }; | |
1553 | ||
1554 | static void test_rcu_tasks_callback(struct rcu_head *rhp) | |
1555 | { | |
1556 | struct rcu_tasks_test_desc *rttd = | |
1557 | container_of(rhp, struct rcu_tasks_test_desc, rh); | |
1558 | ||
1559 | pr_info("Callback from %s invoked.\n", rttd->name); | |
1560 | ||
1561 | rttd->notrun = true; | |
1562 | } | |
1563 | ||
1564 | static void rcu_tasks_initiate_self_tests(void) | |
1565 | { | |
1566 | pr_info("Running RCU-tasks wait API self tests\n"); | |
1567 | #ifdef CONFIG_TASKS_RCU | |
1568 | synchronize_rcu_tasks(); | |
1569 | call_rcu_tasks(&tests[0].rh, test_rcu_tasks_callback); | |
1570 | #endif | |
1571 | ||
1572 | #ifdef CONFIG_TASKS_RUDE_RCU | |
1573 | synchronize_rcu_tasks_rude(); | |
1574 | call_rcu_tasks_rude(&tests[1].rh, test_rcu_tasks_callback); | |
1575 | #endif | |
1576 | ||
1577 | #ifdef CONFIG_TASKS_TRACE_RCU | |
1578 | synchronize_rcu_tasks_trace(); | |
1579 | call_rcu_tasks_trace(&tests[2].rh, test_rcu_tasks_callback); | |
1580 | #endif | |
1581 | } | |
1582 | ||
1583 | static int rcu_tasks_verify_self_tests(void) | |
1584 | { | |
1585 | int ret = 0; | |
1586 | int i; | |
1587 | ||
1588 | for (i = 0; i < ARRAY_SIZE(tests); i++) { | |
1589 | if (!tests[i].notrun) { // still hanging. | |
1590 | pr_err("%s has been failed.\n", tests[i].name); | |
1591 | ret = -1; | |
1592 | } | |
1593 | } | |
1594 | ||
1595 | if (ret) | |
1596 | WARN_ON(1); | |
1597 | ||
1598 | return ret; | |
1599 | } | |
1600 | late_initcall(rcu_tasks_verify_self_tests); | |
1601 | #else /* #ifdef CONFIG_PROVE_RCU */ | |
1602 | static void rcu_tasks_initiate_self_tests(void) { } | |
1603 | #endif /* #else #ifdef CONFIG_PROVE_RCU */ | |
1604 | ||
1b04fa99 URS |
1605 | void __init rcu_init_tasks_generic(void) |
1606 | { | |
1607 | #ifdef CONFIG_TASKS_RCU | |
1608 | rcu_spawn_tasks_kthread(); | |
1609 | #endif | |
1610 | ||
1611 | #ifdef CONFIG_TASKS_RUDE_RCU | |
1612 | rcu_spawn_tasks_rude_kthread(); | |
1613 | #endif | |
1614 | ||
1615 | #ifdef CONFIG_TASKS_TRACE_RCU | |
1616 | rcu_spawn_tasks_trace_kthread(); | |
1617 | #endif | |
bfba7ed0 URS |
1618 | |
1619 | // Run the self-tests. | |
1620 | rcu_tasks_initiate_self_tests(); | |
1b04fa99 URS |
1621 | } |
1622 | ||
8fd8ca38 PM |
1623 | #else /* #ifdef CONFIG_TASKS_RCU_GENERIC */ |
1624 | static inline void rcu_tasks_bootup_oddness(void) {} | |
1625 | #endif /* #else #ifdef CONFIG_TASKS_RCU_GENERIC */ |