Loading...
1/* SPDX-License-Identifier: GPL-2.0+ */
2/*
3 * Read-Copy Update mechanism for mutual exclusion (tree-based version)
4 * Internal non-public definitions.
5 *
6 * Copyright IBM Corporation, 2008
7 *
8 * Author: Ingo Molnar <mingo@elte.hu>
9 * Paul E. McKenney <paulmck@linux.ibm.com>
10 */
11
12#include <linux/cache.h>
13#include <linux/kthread.h>
14#include <linux/spinlock.h>
15#include <linux/rtmutex.h>
16#include <linux/threads.h>
17#include <linux/cpumask.h>
18#include <linux/seqlock.h>
19#include <linux/swait.h>
20#include <linux/rcu_node_tree.h>
21
22#include "rcu_segcblist.h"
23
24/* Communicate arguments to a workqueue handler. */
25struct rcu_exp_work {
26 unsigned long rew_s;
27#ifdef CONFIG_RCU_EXP_KTHREAD
28 struct kthread_work rew_work;
29#else
30 struct work_struct rew_work;
31#endif /* CONFIG_RCU_EXP_KTHREAD */
32};
33
34/* RCU's kthread states for tracing. */
35#define RCU_KTHREAD_STOPPED 0
36#define RCU_KTHREAD_RUNNING 1
37#define RCU_KTHREAD_WAITING 2
38#define RCU_KTHREAD_OFFCPU 3
39#define RCU_KTHREAD_YIELDING 4
40#define RCU_KTHREAD_MAX 4
41
42/*
43 * Definition for node within the RCU grace-period-detection hierarchy.
44 */
45struct rcu_node {
46 raw_spinlock_t __private lock; /* Root rcu_node's lock protects */
47 /* some rcu_state fields as well as */
48 /* following. */
49 unsigned long gp_seq; /* Track rsp->gp_seq. */
50 unsigned long gp_seq_needed; /* Track furthest future GP request. */
51 unsigned long completedqs; /* All QSes done for this node. */
52 unsigned long qsmask; /* CPUs or groups that need to switch in */
53 /* order for current grace period to proceed.*/
54 /* In leaf rcu_node, each bit corresponds to */
55 /* an rcu_data structure, otherwise, each */
56 /* bit corresponds to a child rcu_node */
57 /* structure. */
58 unsigned long rcu_gp_init_mask; /* Mask of offline CPUs at GP init. */
59 unsigned long qsmaskinit;
60 /* Per-GP initial value for qsmask. */
61 /* Initialized from ->qsmaskinitnext at the */
62 /* beginning of each grace period. */
63 unsigned long qsmaskinitnext;
64 unsigned long expmask; /* CPUs or groups that need to check in */
65 /* to allow the current expedited GP */
66 /* to complete. */
67 unsigned long expmaskinit;
68 /* Per-GP initial values for expmask. */
69 /* Initialized from ->expmaskinitnext at the */
70 /* beginning of each expedited GP. */
71 unsigned long expmaskinitnext;
72 /* Online CPUs for next expedited GP. */
73 /* Any CPU that has ever been online will */
74 /* have its bit set. */
75 unsigned long cbovldmask;
76 /* CPUs experiencing callback overload. */
77 unsigned long ffmask; /* Fully functional CPUs. */
78 unsigned long grpmask; /* Mask to apply to parent qsmask. */
79 /* Only one bit will be set in this mask. */
80 int grplo; /* lowest-numbered CPU here. */
81 int grphi; /* highest-numbered CPU here. */
82 u8 grpnum; /* group number for next level up. */
83 u8 level; /* root is at level 0. */
84 bool wait_blkd_tasks;/* Necessary to wait for blocked tasks to */
85 /* exit RCU read-side critical sections */
86 /* before propagating offline up the */
87 /* rcu_node tree? */
88 struct rcu_node *parent;
89 struct list_head blkd_tasks;
90 /* Tasks blocked in RCU read-side critical */
91 /* section. Tasks are placed at the head */
92 /* of this list and age towards the tail. */
93 struct list_head *gp_tasks;
94 /* Pointer to the first task blocking the */
95 /* current grace period, or NULL if there */
96 /* is no such task. */
97 struct list_head *exp_tasks;
98 /* Pointer to the first task blocking the */
99 /* current expedited grace period, or NULL */
100 /* if there is no such task. If there */
101 /* is no current expedited grace period, */
102 /* then there can cannot be any such task. */
103 struct list_head *boost_tasks;
104 /* Pointer to first task that needs to be */
105 /* priority boosted, or NULL if no priority */
106 /* boosting is needed for this rcu_node */
107 /* structure. If there are no tasks */
108 /* queued on this rcu_node structure that */
109 /* are blocking the current grace period, */
110 /* there can be no such task. */
111 struct rt_mutex boost_mtx;
112 /* Used only for the priority-boosting */
113 /* side effect, not as a lock. */
114 unsigned long boost_time;
115 /* When to start boosting (jiffies). */
116 struct mutex boost_kthread_mutex;
117 /* Exclusion for thread spawning and affinity */
118 /* manipulation. */
119 struct task_struct *boost_kthread_task;
120 /* kthread that takes care of priority */
121 /* boosting for this rcu_node structure. */
122 unsigned int boost_kthread_status;
123 /* State of boost_kthread_task for tracing. */
124 unsigned long n_boosts; /* Number of boosts for this rcu_node structure. */
125#ifdef CONFIG_RCU_NOCB_CPU
126 struct swait_queue_head nocb_gp_wq[2];
127 /* Place for rcu_nocb_kthread() to wait GP. */
128#endif /* #ifdef CONFIG_RCU_NOCB_CPU */
129 raw_spinlock_t fqslock ____cacheline_internodealigned_in_smp;
130
131 spinlock_t exp_lock ____cacheline_internodealigned_in_smp;
132 unsigned long exp_seq_rq;
133 wait_queue_head_t exp_wq[4];
134 struct rcu_exp_work rew;
135 bool exp_need_flush; /* Need to flush workitem? */
136 raw_spinlock_t exp_poll_lock;
137 /* Lock and data for polled expedited grace periods. */
138 unsigned long exp_seq_poll_rq;
139 struct work_struct exp_poll_wq;
140} ____cacheline_internodealigned_in_smp;
141
142/*
143 * Bitmasks in an rcu_node cover the interval [grplo, grphi] of CPU IDs, and
144 * are indexed relative to this interval rather than the global CPU ID space.
145 * This generates the bit for a CPU in node-local masks.
146 */
147#define leaf_node_cpu_bit(rnp, cpu) (BIT((cpu) - (rnp)->grplo))
148
149/*
150 * Union to allow "aggregate OR" operation on the need for a quiescent
151 * state by the normal and expedited grace periods.
152 */
153union rcu_noqs {
154 struct {
155 u8 norm;
156 u8 exp;
157 } b; /* Bits. */
158 u16 s; /* Set of bits, aggregate OR here. */
159};
160
161/* Per-CPU data for read-copy update. */
162struct rcu_data {
163 /* 1) quiescent-state and grace-period handling : */
164 unsigned long gp_seq; /* Track rsp->gp_seq counter. */
165 unsigned long gp_seq_needed; /* Track furthest future GP request. */
166 union rcu_noqs cpu_no_qs; /* No QSes yet for this CPU. */
167 bool core_needs_qs; /* Core waits for quiescent state. */
168 bool beenonline; /* CPU online at least once. */
169 bool gpwrap; /* Possible ->gp_seq wrap. */
170 bool cpu_started; /* RCU watching this onlining CPU. */
171 struct rcu_node *mynode; /* This CPU's leaf of hierarchy */
172 unsigned long grpmask; /* Mask to apply to leaf qsmask. */
173 unsigned long ticks_this_gp; /* The number of scheduling-clock */
174 /* ticks this CPU has handled */
175 /* during and after the last grace */
176 /* period it is aware of. */
177 struct irq_work defer_qs_iw; /* Obtain later scheduler attention. */
178 bool defer_qs_iw_pending; /* Scheduler attention pending? */
179 struct work_struct strict_work; /* Schedule readers for strict GPs. */
180
181 /* 2) batch handling */
182 struct rcu_segcblist cblist; /* Segmented callback list, with */
183 /* different callbacks waiting for */
184 /* different grace periods. */
185 long qlen_last_fqs_check;
186 /* qlen at last check for QS forcing */
187 unsigned long n_cbs_invoked; /* # callbacks invoked since boot. */
188 unsigned long n_force_qs_snap;
189 /* did other CPU force QS recently? */
190 long blimit; /* Upper limit on a processed batch */
191
192 /* 3) dynticks interface. */
193 int dynticks_snap; /* Per-GP tracking for dynticks. */
194 bool rcu_need_heavy_qs; /* GP old, so heavy quiescent state! */
195 bool rcu_urgent_qs; /* GP old need light quiescent state. */
196 bool rcu_forced_tick; /* Forced tick to provide QS. */
197 bool rcu_forced_tick_exp; /* ... provide QS to expedited GP. */
198
199 /* 4) rcu_barrier(), OOM callbacks, and expediting. */
200 unsigned long barrier_seq_snap; /* Snap of rcu_state.barrier_sequence. */
201 struct rcu_head barrier_head;
202 int exp_dynticks_snap; /* Double-check need for IPI. */
203
204 /* 5) Callback offloading. */
205#ifdef CONFIG_RCU_NOCB_CPU
206 struct swait_queue_head nocb_cb_wq; /* For nocb kthreads to sleep on. */
207 struct swait_queue_head nocb_state_wq; /* For offloading state changes */
208 struct task_struct *nocb_gp_kthread;
209 raw_spinlock_t nocb_lock; /* Guard following pair of fields. */
210 atomic_t nocb_lock_contended; /* Contention experienced. */
211 int nocb_defer_wakeup; /* Defer wakeup of nocb_kthread. */
212 struct timer_list nocb_timer; /* Enforce finite deferral. */
213 unsigned long nocb_gp_adv_time; /* Last call_rcu() CB adv (jiffies). */
214 struct mutex nocb_gp_kthread_mutex; /* Exclusion for nocb gp kthread */
215 /* spawning */
216
217 /* The following fields are used by call_rcu, hence own cacheline. */
218 raw_spinlock_t nocb_bypass_lock ____cacheline_internodealigned_in_smp;
219 struct rcu_cblist nocb_bypass; /* Lock-contention-bypass CB list. */
220 unsigned long nocb_bypass_first; /* Time (jiffies) of first enqueue. */
221 unsigned long nocb_nobypass_last; /* Last ->cblist enqueue (jiffies). */
222 int nocb_nobypass_count; /* # ->cblist enqueues at ^^^ time. */
223
224 /* The following fields are used by GP kthread, hence own cacheline. */
225 raw_spinlock_t nocb_gp_lock ____cacheline_internodealigned_in_smp;
226 u8 nocb_gp_sleep; /* Is the nocb GP thread asleep? */
227 u8 nocb_gp_bypass; /* Found a bypass on last scan? */
228 u8 nocb_gp_gp; /* GP to wait for on last scan? */
229 unsigned long nocb_gp_seq; /* If so, ->gp_seq to wait for. */
230 unsigned long nocb_gp_loops; /* # passes through wait code. */
231 struct swait_queue_head nocb_gp_wq; /* For nocb kthreads to sleep on. */
232 bool nocb_cb_sleep; /* Is the nocb CB thread asleep? */
233 struct task_struct *nocb_cb_kthread;
234 struct list_head nocb_head_rdp; /*
235 * Head of rcu_data list in wakeup chain,
236 * if rdp_gp.
237 */
238 struct list_head nocb_entry_rdp; /* rcu_data node in wakeup chain. */
239 struct rcu_data *nocb_toggling_rdp; /* rdp queued for (de-)offloading */
240
241 /* The following fields are used by CB kthread, hence new cacheline. */
242 struct rcu_data *nocb_gp_rdp ____cacheline_internodealigned_in_smp;
243 /* GP rdp takes GP-end wakeups. */
244#endif /* #ifdef CONFIG_RCU_NOCB_CPU */
245
246 /* 6) RCU priority boosting. */
247 struct task_struct *rcu_cpu_kthread_task;
248 /* rcuc per-CPU kthread or NULL. */
249 unsigned int rcu_cpu_kthread_status;
250 char rcu_cpu_has_work;
251 unsigned long rcuc_activity;
252
253 /* 7) Diagnostic data, including RCU CPU stall warnings. */
254 unsigned int softirq_snap; /* Snapshot of softirq activity. */
255 /* ->rcu_iw* fields protected by leaf rcu_node ->lock. */
256 struct irq_work rcu_iw; /* Check for non-irq activity. */
257 bool rcu_iw_pending; /* Is ->rcu_iw pending? */
258 unsigned long rcu_iw_gp_seq; /* ->gp_seq associated with ->rcu_iw. */
259 unsigned long rcu_ofl_gp_seq; /* ->gp_seq at last offline. */
260 short rcu_ofl_gp_flags; /* ->gp_flags at last offline. */
261 unsigned long rcu_onl_gp_seq; /* ->gp_seq at last online. */
262 short rcu_onl_gp_flags; /* ->gp_flags at last online. */
263 unsigned long last_fqs_resched; /* Time of last rcu_resched(). */
264 unsigned long last_sched_clock; /* Jiffies of last rcu_sched_clock_irq(). */
265
266 long lazy_len; /* Length of buffered lazy callbacks. */
267 int cpu;
268};
269
270/* Values for nocb_defer_wakeup field in struct rcu_data. */
271#define RCU_NOCB_WAKE_NOT 0
272#define RCU_NOCB_WAKE_BYPASS 1
273#define RCU_NOCB_WAKE_LAZY 2
274#define RCU_NOCB_WAKE 3
275#define RCU_NOCB_WAKE_FORCE 4
276
277#define RCU_JIFFIES_TILL_FORCE_QS (1 + (HZ > 250) + (HZ > 500))
278 /* For jiffies_till_first_fqs and */
279 /* and jiffies_till_next_fqs. */
280
281#define RCU_JIFFIES_FQS_DIV 256 /* Very large systems need more */
282 /* delay between bouts of */
283 /* quiescent-state forcing. */
284
285#define RCU_STALL_RAT_DELAY 2 /* Allow other CPUs time to take */
286 /* at least one scheduling clock */
287 /* irq before ratting on them. */
288
289#define rcu_wait(cond) \
290do { \
291 for (;;) { \
292 set_current_state(TASK_INTERRUPTIBLE); \
293 if (cond) \
294 break; \
295 schedule(); \
296 } \
297 __set_current_state(TASK_RUNNING); \
298} while (0)
299
300/*
301 * RCU global state, including node hierarchy. This hierarchy is
302 * represented in "heap" form in a dense array. The root (first level)
303 * of the hierarchy is in ->node[0] (referenced by ->level[0]), the second
304 * level in ->node[1] through ->node[m] (->node[1] referenced by ->level[1]),
305 * and the third level in ->node[m+1] and following (->node[m+1] referenced
306 * by ->level[2]). The number of levels is determined by the number of
307 * CPUs and by CONFIG_RCU_FANOUT. Small systems will have a "hierarchy"
308 * consisting of a single rcu_node.
309 */
310struct rcu_state {
311 struct rcu_node node[NUM_RCU_NODES]; /* Hierarchy. */
312 struct rcu_node *level[RCU_NUM_LVLS + 1];
313 /* Hierarchy levels (+1 to */
314 /* shut bogus gcc warning) */
315 int ncpus; /* # CPUs seen so far. */
316 int n_online_cpus; /* # CPUs online for RCU. */
317
318 /* The following fields are guarded by the root rcu_node's lock. */
319
320 unsigned long gp_seq ____cacheline_internodealigned_in_smp;
321 /* Grace-period sequence #. */
322 unsigned long gp_max; /* Maximum GP duration in */
323 /* jiffies. */
324 struct task_struct *gp_kthread; /* Task for grace periods. */
325 struct swait_queue_head gp_wq; /* Where GP task waits. */
326 short gp_flags; /* Commands for GP task. */
327 short gp_state; /* GP kthread sleep state. */
328 unsigned long gp_wake_time; /* Last GP kthread wake. */
329 unsigned long gp_wake_seq; /* ->gp_seq at ^^^. */
330 unsigned long gp_seq_polled; /* GP seq for polled API. */
331 unsigned long gp_seq_polled_snap; /* ->gp_seq_polled at normal GP start. */
332 unsigned long gp_seq_polled_exp_snap; /* ->gp_seq_polled at expedited GP start. */
333
334 /* End of fields guarded by root rcu_node's lock. */
335
336 struct mutex barrier_mutex; /* Guards barrier fields. */
337 atomic_t barrier_cpu_count; /* # CPUs waiting on. */
338 struct completion barrier_completion; /* Wake at barrier end. */
339 unsigned long barrier_sequence; /* ++ at start and end of */
340 /* rcu_barrier(). */
341 /* End of fields guarded by barrier_mutex. */
342
343 raw_spinlock_t barrier_lock; /* Protects ->barrier_seq_snap. */
344
345 struct mutex exp_mutex; /* Serialize expedited GP. */
346 struct mutex exp_wake_mutex; /* Serialize wakeup. */
347 unsigned long expedited_sequence; /* Take a ticket. */
348 atomic_t expedited_need_qs; /* # CPUs left to check in. */
349 struct swait_queue_head expedited_wq; /* Wait for check-ins. */
350 int ncpus_snap; /* # CPUs seen last time. */
351 u8 cbovld; /* Callback overload now? */
352 u8 cbovldnext; /* ^ ^ next time? */
353
354 unsigned long jiffies_force_qs; /* Time at which to invoke */
355 /* force_quiescent_state(). */
356 unsigned long jiffies_kick_kthreads; /* Time at which to kick */
357 /* kthreads, if configured. */
358 unsigned long n_force_qs; /* Number of calls to */
359 /* force_quiescent_state(). */
360 unsigned long gp_start; /* Time at which GP started, */
361 /* but in jiffies. */
362 unsigned long gp_end; /* Time last GP ended, again */
363 /* in jiffies. */
364 unsigned long gp_activity; /* Time of last GP kthread */
365 /* activity in jiffies. */
366 unsigned long gp_req_activity; /* Time of last GP request */
367 /* in jiffies. */
368 unsigned long jiffies_stall; /* Time at which to check */
369 /* for CPU stalls. */
370 unsigned long jiffies_resched; /* Time at which to resched */
371 /* a reluctant CPU. */
372 unsigned long n_force_qs_gpstart; /* Snapshot of n_force_qs at */
373 /* GP start. */
374 const char *name; /* Name of structure. */
375 char abbr; /* Abbreviated name. */
376
377 arch_spinlock_t ofl_lock ____cacheline_internodealigned_in_smp;
378 /* Synchronize offline with */
379 /* GP pre-initialization. */
380 int nocb_is_setup; /* nocb is setup from boot */
381};
382
383/* Values for rcu_state structure's gp_flags field. */
384#define RCU_GP_FLAG_INIT 0x1 /* Need grace-period initialization. */
385#define RCU_GP_FLAG_FQS 0x2 /* Need grace-period quiescent-state forcing. */
386#define RCU_GP_FLAG_OVLD 0x4 /* Experiencing callback overload. */
387
388/* Values for rcu_state structure's gp_state field. */
389#define RCU_GP_IDLE 0 /* Initial state and no GP in progress. */
390#define RCU_GP_WAIT_GPS 1 /* Wait for grace-period start. */
391#define RCU_GP_DONE_GPS 2 /* Wait done for grace-period start. */
392#define RCU_GP_ONOFF 3 /* Grace-period initialization hotplug. */
393#define RCU_GP_INIT 4 /* Grace-period initialization. */
394#define RCU_GP_WAIT_FQS 5 /* Wait for force-quiescent-state time. */
395#define RCU_GP_DOING_FQS 6 /* Wait done for force-quiescent-state time. */
396#define RCU_GP_CLEANUP 7 /* Grace-period cleanup started. */
397#define RCU_GP_CLEANED 8 /* Grace-period cleanup complete. */
398
399/*
400 * In order to export the rcu_state name to the tracing tools, it
401 * needs to be added in the __tracepoint_string section.
402 * This requires defining a separate variable tp_<sname>_varname
403 * that points to the string being used, and this will allow
404 * the tracing userspace tools to be able to decipher the string
405 * address to the matching string.
406 */
407#ifdef CONFIG_PREEMPT_RCU
408#define RCU_ABBR 'p'
409#define RCU_NAME_RAW "rcu_preempt"
410#else /* #ifdef CONFIG_PREEMPT_RCU */
411#define RCU_ABBR 's'
412#define RCU_NAME_RAW "rcu_sched"
413#endif /* #else #ifdef CONFIG_PREEMPT_RCU */
414#ifndef CONFIG_TRACING
415#define RCU_NAME RCU_NAME_RAW
416#else /* #ifdef CONFIG_TRACING */
417static char rcu_name[] = RCU_NAME_RAW;
418static const char *tp_rcu_varname __used __tracepoint_string = rcu_name;
419#define RCU_NAME rcu_name
420#endif /* #else #ifdef CONFIG_TRACING */
421
422/* Forward declarations for tree_plugin.h */
423static void rcu_bootup_announce(void);
424static void rcu_qs(void);
425static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp);
426#ifdef CONFIG_HOTPLUG_CPU
427static bool rcu_preempt_has_tasks(struct rcu_node *rnp);
428#endif /* #ifdef CONFIG_HOTPLUG_CPU */
429static int rcu_print_task_exp_stall(struct rcu_node *rnp);
430static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp);
431static void rcu_flavor_sched_clock_irq(int user);
432static void dump_blkd_tasks(struct rcu_node *rnp, int ncheck);
433static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags);
434static void rcu_preempt_boost_start_gp(struct rcu_node *rnp);
435static bool rcu_is_callbacks_kthread(struct rcu_data *rdp);
436static void rcu_cpu_kthread_setup(unsigned int cpu);
437static void rcu_spawn_one_boost_kthread(struct rcu_node *rnp);
438static bool rcu_preempt_has_tasks(struct rcu_node *rnp);
439static bool rcu_preempt_need_deferred_qs(struct task_struct *t);
440static void zero_cpu_stall_ticks(struct rcu_data *rdp);
441static struct swait_queue_head *rcu_nocb_gp_get(struct rcu_node *rnp);
442static void rcu_nocb_gp_cleanup(struct swait_queue_head *sq);
443static void rcu_init_one_nocb(struct rcu_node *rnp);
444static bool wake_nocb_gp(struct rcu_data *rdp, bool force);
445static bool rcu_nocb_flush_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
446 unsigned long j, bool lazy);
447static bool rcu_nocb_try_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
448 bool *was_alldone, unsigned long flags,
449 bool lazy);
450static void __call_rcu_nocb_wake(struct rcu_data *rdp, bool was_empty,
451 unsigned long flags);
452static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp, int level);
453static bool do_nocb_deferred_wakeup(struct rcu_data *rdp);
454static void rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp);
455static void rcu_spawn_cpu_nocb_kthread(int cpu);
456static void show_rcu_nocb_state(struct rcu_data *rdp);
457static void rcu_nocb_lock(struct rcu_data *rdp);
458static void rcu_nocb_unlock(struct rcu_data *rdp);
459static void rcu_nocb_unlock_irqrestore(struct rcu_data *rdp,
460 unsigned long flags);
461static void rcu_lockdep_assert_cblist_protected(struct rcu_data *rdp);
462#ifdef CONFIG_RCU_NOCB_CPU
463static void __init rcu_organize_nocb_kthreads(void);
464
465/*
466 * Disable IRQs before checking offloaded state so that local
467 * locking is safe against concurrent de-offloading.
468 */
469#define rcu_nocb_lock_irqsave(rdp, flags) \
470do { \
471 local_irq_save(flags); \
472 if (rcu_segcblist_is_offloaded(&(rdp)->cblist)) \
473 raw_spin_lock(&(rdp)->nocb_lock); \
474} while (0)
475#else /* #ifdef CONFIG_RCU_NOCB_CPU */
476#define rcu_nocb_lock_irqsave(rdp, flags) local_irq_save(flags)
477#endif /* #else #ifdef CONFIG_RCU_NOCB_CPU */
478
479static void rcu_bind_gp_kthread(void);
480static bool rcu_nohz_full_cpu(void);
481
482/* Forward declarations for tree_stall.h */
483static void record_gp_stall_check_time(void);
484static void rcu_iw_handler(struct irq_work *iwp);
485static void check_cpu_stall(struct rcu_data *rdp);
486static void rcu_check_gp_start_stall(struct rcu_node *rnp, struct rcu_data *rdp,
487 const unsigned long gpssdelay);
488
489/* Forward declarations for tree_exp.h. */
490static void sync_rcu_do_polled_gp(struct work_struct *wp);
1/* SPDX-License-Identifier: GPL-2.0+ */
2/*
3 * Read-Copy Update mechanism for mutual exclusion (tree-based version)
4 * Internal non-public definitions.
5 *
6 * Copyright IBM Corporation, 2008
7 *
8 * Author: Ingo Molnar <mingo@elte.hu>
9 * Paul E. McKenney <paulmck@linux.ibm.com>
10 */
11
12#include <linux/cache.h>
13#include <linux/spinlock.h>
14#include <linux/rtmutex.h>
15#include <linux/threads.h>
16#include <linux/cpumask.h>
17#include <linux/seqlock.h>
18#include <linux/swait.h>
19#include <linux/rcu_node_tree.h>
20
21#include "rcu_segcblist.h"
22
23/* Communicate arguments to a workqueue handler. */
24struct rcu_exp_work {
25 unsigned long rew_s;
26 struct work_struct rew_work;
27};
28
29/* RCU's kthread states for tracing. */
30#define RCU_KTHREAD_STOPPED 0
31#define RCU_KTHREAD_RUNNING 1
32#define RCU_KTHREAD_WAITING 2
33#define RCU_KTHREAD_OFFCPU 3
34#define RCU_KTHREAD_YIELDING 4
35#define RCU_KTHREAD_MAX 4
36
37/*
38 * Definition for node within the RCU grace-period-detection hierarchy.
39 */
40struct rcu_node {
41 raw_spinlock_t __private lock; /* Root rcu_node's lock protects */
42 /* some rcu_state fields as well as */
43 /* following. */
44 unsigned long gp_seq; /* Track rsp->gp_seq. */
45 unsigned long gp_seq_needed; /* Track furthest future GP request. */
46 unsigned long completedqs; /* All QSes done for this node. */
47 unsigned long qsmask; /* CPUs or groups that need to switch in */
48 /* order for current grace period to proceed.*/
49 /* In leaf rcu_node, each bit corresponds to */
50 /* an rcu_data structure, otherwise, each */
51 /* bit corresponds to a child rcu_node */
52 /* structure. */
53 unsigned long rcu_gp_init_mask; /* Mask of offline CPUs at GP init. */
54 unsigned long qsmaskinit;
55 /* Per-GP initial value for qsmask. */
56 /* Initialized from ->qsmaskinitnext at the */
57 /* beginning of each grace period. */
58 unsigned long qsmaskinitnext;
59 /* Online CPUs for next grace period. */
60 unsigned long expmask; /* CPUs or groups that need to check in */
61 /* to allow the current expedited GP */
62 /* to complete. */
63 unsigned long expmaskinit;
64 /* Per-GP initial values for expmask. */
65 /* Initialized from ->expmaskinitnext at the */
66 /* beginning of each expedited GP. */
67 unsigned long expmaskinitnext;
68 /* Online CPUs for next expedited GP. */
69 /* Any CPU that has ever been online will */
70 /* have its bit set. */
71 unsigned long cbovldmask;
72 /* CPUs experiencing callback overload. */
73 unsigned long ffmask; /* Fully functional CPUs. */
74 unsigned long grpmask; /* Mask to apply to parent qsmask. */
75 /* Only one bit will be set in this mask. */
76 int grplo; /* lowest-numbered CPU here. */
77 int grphi; /* highest-numbered CPU here. */
78 u8 grpnum; /* group number for next level up. */
79 u8 level; /* root is at level 0. */
80 bool wait_blkd_tasks;/* Necessary to wait for blocked tasks to */
81 /* exit RCU read-side critical sections */
82 /* before propagating offline up the */
83 /* rcu_node tree? */
84 struct rcu_node *parent;
85 struct list_head blkd_tasks;
86 /* Tasks blocked in RCU read-side critical */
87 /* section. Tasks are placed at the head */
88 /* of this list and age towards the tail. */
89 struct list_head *gp_tasks;
90 /* Pointer to the first task blocking the */
91 /* current grace period, or NULL if there */
92 /* is no such task. */
93 struct list_head *exp_tasks;
94 /* Pointer to the first task blocking the */
95 /* current expedited grace period, or NULL */
96 /* if there is no such task. If there */
97 /* is no current expedited grace period, */
98 /* then there can cannot be any such task. */
99 struct list_head *boost_tasks;
100 /* Pointer to first task that needs to be */
101 /* priority boosted, or NULL if no priority */
102 /* boosting is needed for this rcu_node */
103 /* structure. If there are no tasks */
104 /* queued on this rcu_node structure that */
105 /* are blocking the current grace period, */
106 /* there can be no such task. */
107 struct rt_mutex boost_mtx;
108 /* Used only for the priority-boosting */
109 /* side effect, not as a lock. */
110 unsigned long boost_time;
111 /* When to start boosting (jiffies). */
112 struct task_struct *boost_kthread_task;
113 /* kthread that takes care of priority */
114 /* boosting for this rcu_node structure. */
115 unsigned int boost_kthread_status;
116 /* State of boost_kthread_task for tracing. */
117#ifdef CONFIG_RCU_NOCB_CPU
118 struct swait_queue_head nocb_gp_wq[2];
119 /* Place for rcu_nocb_kthread() to wait GP. */
120#endif /* #ifdef CONFIG_RCU_NOCB_CPU */
121 raw_spinlock_t fqslock ____cacheline_internodealigned_in_smp;
122
123 spinlock_t exp_lock ____cacheline_internodealigned_in_smp;
124 unsigned long exp_seq_rq;
125 wait_queue_head_t exp_wq[4];
126 struct rcu_exp_work rew;
127 bool exp_need_flush; /* Need to flush workitem? */
128} ____cacheline_internodealigned_in_smp;
129
130/*
131 * Bitmasks in an rcu_node cover the interval [grplo, grphi] of CPU IDs, and
132 * are indexed relative to this interval rather than the global CPU ID space.
133 * This generates the bit for a CPU in node-local masks.
134 */
135#define leaf_node_cpu_bit(rnp, cpu) (BIT((cpu) - (rnp)->grplo))
136
137/*
138 * Union to allow "aggregate OR" operation on the need for a quiescent
139 * state by the normal and expedited grace periods.
140 */
141union rcu_noqs {
142 struct {
143 u8 norm;
144 u8 exp;
145 } b; /* Bits. */
146 u16 s; /* Set of bits, aggregate OR here. */
147};
148
149/* Per-CPU data for read-copy update. */
150struct rcu_data {
151 /* 1) quiescent-state and grace-period handling : */
152 unsigned long gp_seq; /* Track rsp->gp_seq counter. */
153 unsigned long gp_seq_needed; /* Track furthest future GP request. */
154 union rcu_noqs cpu_no_qs; /* No QSes yet for this CPU. */
155 bool core_needs_qs; /* Core waits for quiesc state. */
156 bool beenonline; /* CPU online at least once. */
157 bool gpwrap; /* Possible ->gp_seq wrap. */
158 bool exp_deferred_qs; /* This CPU awaiting a deferred QS? */
159 struct rcu_node *mynode; /* This CPU's leaf of hierarchy */
160 unsigned long grpmask; /* Mask to apply to leaf qsmask. */
161 unsigned long ticks_this_gp; /* The number of scheduling-clock */
162 /* ticks this CPU has handled */
163 /* during and after the last grace */
164 /* period it is aware of. */
165 struct irq_work defer_qs_iw; /* Obtain later scheduler attention. */
166 bool defer_qs_iw_pending; /* Scheduler attention pending? */
167
168 /* 2) batch handling */
169 struct rcu_segcblist cblist; /* Segmented callback list, with */
170 /* different callbacks waiting for */
171 /* different grace periods. */
172 long qlen_last_fqs_check;
173 /* qlen at last check for QS forcing */
174 unsigned long n_cbs_invoked; /* # callbacks invoked since boot. */
175 unsigned long n_force_qs_snap;
176 /* did other CPU force QS recently? */
177 long blimit; /* Upper limit on a processed batch */
178
179 /* 3) dynticks interface. */
180 int dynticks_snap; /* Per-GP tracking for dynticks. */
181 long dynticks_nesting; /* Track process nesting level. */
182 long dynticks_nmi_nesting; /* Track irq/NMI nesting level. */
183 atomic_t dynticks; /* Even value for idle, else odd. */
184 bool rcu_need_heavy_qs; /* GP old, so heavy quiescent state! */
185 bool rcu_urgent_qs; /* GP old need light quiescent state. */
186 bool rcu_forced_tick; /* Forced tick to provide QS. */
187 bool rcu_forced_tick_exp; /* ... provide QS to expedited GP. */
188#ifdef CONFIG_RCU_FAST_NO_HZ
189 unsigned long last_accelerate; /* Last jiffy CBs were accelerated. */
190 unsigned long last_advance_all; /* Last jiffy CBs were all advanced. */
191 int tick_nohz_enabled_snap; /* Previously seen value from sysfs. */
192#endif /* #ifdef CONFIG_RCU_FAST_NO_HZ */
193
194 /* 4) rcu_barrier(), OOM callbacks, and expediting. */
195 struct rcu_head barrier_head;
196 int exp_dynticks_snap; /* Double-check need for IPI. */
197
198 /* 5) Callback offloading. */
199#ifdef CONFIG_RCU_NOCB_CPU
200 struct swait_queue_head nocb_cb_wq; /* For nocb kthreads to sleep on. */
201 struct task_struct *nocb_gp_kthread;
202 raw_spinlock_t nocb_lock; /* Guard following pair of fields. */
203 atomic_t nocb_lock_contended; /* Contention experienced. */
204 int nocb_defer_wakeup; /* Defer wakeup of nocb_kthread. */
205 struct timer_list nocb_timer; /* Enforce finite deferral. */
206 unsigned long nocb_gp_adv_time; /* Last call_rcu() CB adv (jiffies). */
207
208 /* The following fields are used by call_rcu, hence own cacheline. */
209 raw_spinlock_t nocb_bypass_lock ____cacheline_internodealigned_in_smp;
210 struct rcu_cblist nocb_bypass; /* Lock-contention-bypass CB list. */
211 unsigned long nocb_bypass_first; /* Time (jiffies) of first enqueue. */
212 unsigned long nocb_nobypass_last; /* Last ->cblist enqueue (jiffies). */
213 int nocb_nobypass_count; /* # ->cblist enqueues at ^^^ time. */
214
215 /* The following fields are used by GP kthread, hence own cacheline. */
216 raw_spinlock_t nocb_gp_lock ____cacheline_internodealigned_in_smp;
217 struct timer_list nocb_bypass_timer; /* Force nocb_bypass flush. */
218 u8 nocb_gp_sleep; /* Is the nocb GP thread asleep? */
219 u8 nocb_gp_bypass; /* Found a bypass on last scan? */
220 u8 nocb_gp_gp; /* GP to wait for on last scan? */
221 unsigned long nocb_gp_seq; /* If so, ->gp_seq to wait for. */
222 unsigned long nocb_gp_loops; /* # passes through wait code. */
223 struct swait_queue_head nocb_gp_wq; /* For nocb kthreads to sleep on. */
224 bool nocb_cb_sleep; /* Is the nocb CB thread asleep? */
225 struct task_struct *nocb_cb_kthread;
226 struct rcu_data *nocb_next_cb_rdp;
227 /* Next rcu_data in wakeup chain. */
228
229 /* The following fields are used by CB kthread, hence new cacheline. */
230 struct rcu_data *nocb_gp_rdp ____cacheline_internodealigned_in_smp;
231 /* GP rdp takes GP-end wakeups. */
232#endif /* #ifdef CONFIG_RCU_NOCB_CPU */
233
234 /* 6) RCU priority boosting. */
235 struct task_struct *rcu_cpu_kthread_task;
236 /* rcuc per-CPU kthread or NULL. */
237 unsigned int rcu_cpu_kthread_status;
238 char rcu_cpu_has_work;
239
240 /* 7) Diagnostic data, including RCU CPU stall warnings. */
241 unsigned int softirq_snap; /* Snapshot of softirq activity. */
242 /* ->rcu_iw* fields protected by leaf rcu_node ->lock. */
243 struct irq_work rcu_iw; /* Check for non-irq activity. */
244 bool rcu_iw_pending; /* Is ->rcu_iw pending? */
245 unsigned long rcu_iw_gp_seq; /* ->gp_seq associated with ->rcu_iw. */
246 unsigned long rcu_ofl_gp_seq; /* ->gp_seq at last offline. */
247 short rcu_ofl_gp_flags; /* ->gp_flags at last offline. */
248 unsigned long rcu_onl_gp_seq; /* ->gp_seq at last online. */
249 short rcu_onl_gp_flags; /* ->gp_flags at last online. */
250 unsigned long last_fqs_resched; /* Time of last rcu_resched(). */
251
252 int cpu;
253};
254
255/* Values for nocb_defer_wakeup field in struct rcu_data. */
256#define RCU_NOCB_WAKE_NOT 0
257#define RCU_NOCB_WAKE 1
258#define RCU_NOCB_WAKE_FORCE 2
259
260#define RCU_JIFFIES_TILL_FORCE_QS (1 + (HZ > 250) + (HZ > 500))
261 /* For jiffies_till_first_fqs and */
262 /* and jiffies_till_next_fqs. */
263
264#define RCU_JIFFIES_FQS_DIV 256 /* Very large systems need more */
265 /* delay between bouts of */
266 /* quiescent-state forcing. */
267
268#define RCU_STALL_RAT_DELAY 2 /* Allow other CPUs time to take */
269 /* at least one scheduling clock */
270 /* irq before ratting on them. */
271
272#define rcu_wait(cond) \
273do { \
274 for (;;) { \
275 set_current_state(TASK_INTERRUPTIBLE); \
276 if (cond) \
277 break; \
278 schedule(); \
279 } \
280 __set_current_state(TASK_RUNNING); \
281} while (0)
282
283/*
284 * RCU global state, including node hierarchy. This hierarchy is
285 * represented in "heap" form in a dense array. The root (first level)
286 * of the hierarchy is in ->node[0] (referenced by ->level[0]), the second
287 * level in ->node[1] through ->node[m] (->node[1] referenced by ->level[1]),
288 * and the third level in ->node[m+1] and following (->node[m+1] referenced
289 * by ->level[2]). The number of levels is determined by the number of
290 * CPUs and by CONFIG_RCU_FANOUT. Small systems will have a "hierarchy"
291 * consisting of a single rcu_node.
292 */
293struct rcu_state {
294 struct rcu_node node[NUM_RCU_NODES]; /* Hierarchy. */
295 struct rcu_node *level[RCU_NUM_LVLS + 1];
296 /* Hierarchy levels (+1 to */
297 /* shut bogus gcc warning) */
298 int ncpus; /* # CPUs seen so far. */
299
300 /* The following fields are guarded by the root rcu_node's lock. */
301
302 u8 boost ____cacheline_internodealigned_in_smp;
303 /* Subject to priority boost. */
304 unsigned long gp_seq; /* Grace-period sequence #. */
305 unsigned long gp_max; /* Maximum GP duration in */
306 /* jiffies. */
307 struct task_struct *gp_kthread; /* Task for grace periods. */
308 struct swait_queue_head gp_wq; /* Where GP task waits. */
309 short gp_flags; /* Commands for GP task. */
310 short gp_state; /* GP kthread sleep state. */
311 unsigned long gp_wake_time; /* Last GP kthread wake. */
312 unsigned long gp_wake_seq; /* ->gp_seq at ^^^. */
313
314 /* End of fields guarded by root rcu_node's lock. */
315
316 struct mutex barrier_mutex; /* Guards barrier fields. */
317 atomic_t barrier_cpu_count; /* # CPUs waiting on. */
318 struct completion barrier_completion; /* Wake at barrier end. */
319 unsigned long barrier_sequence; /* ++ at start and end of */
320 /* rcu_barrier(). */
321 /* End of fields guarded by barrier_mutex. */
322
323 struct mutex exp_mutex; /* Serialize expedited GP. */
324 struct mutex exp_wake_mutex; /* Serialize wakeup. */
325 unsigned long expedited_sequence; /* Take a ticket. */
326 atomic_t expedited_need_qs; /* # CPUs left to check in. */
327 struct swait_queue_head expedited_wq; /* Wait for check-ins. */
328 int ncpus_snap; /* # CPUs seen last time. */
329 u8 cbovld; /* Callback overload now? */
330 u8 cbovldnext; /* ^ ^ next time? */
331
332 unsigned long jiffies_force_qs; /* Time at which to invoke */
333 /* force_quiescent_state(). */
334 unsigned long jiffies_kick_kthreads; /* Time at which to kick */
335 /* kthreads, if configured. */
336 unsigned long n_force_qs; /* Number of calls to */
337 /* force_quiescent_state(). */
338 unsigned long gp_start; /* Time at which GP started, */
339 /* but in jiffies. */
340 unsigned long gp_end; /* Time last GP ended, again */
341 /* in jiffies. */
342 unsigned long gp_activity; /* Time of last GP kthread */
343 /* activity in jiffies. */
344 unsigned long gp_req_activity; /* Time of last GP request */
345 /* in jiffies. */
346 unsigned long jiffies_stall; /* Time at which to check */
347 /* for CPU stalls. */
348 unsigned long jiffies_resched; /* Time at which to resched */
349 /* a reluctant CPU. */
350 unsigned long n_force_qs_gpstart; /* Snapshot of n_force_qs at */
351 /* GP start. */
352 const char *name; /* Name of structure. */
353 char abbr; /* Abbreviated name. */
354
355 raw_spinlock_t ofl_lock ____cacheline_internodealigned_in_smp;
356 /* Synchronize offline with */
357 /* GP pre-initialization. */
358};
359
360/* Values for rcu_state structure's gp_flags field. */
361#define RCU_GP_FLAG_INIT 0x1 /* Need grace-period initialization. */
362#define RCU_GP_FLAG_FQS 0x2 /* Need grace-period quiescent-state forcing. */
363#define RCU_GP_FLAG_OVLD 0x4 /* Experiencing callback overload. */
364
365/* Values for rcu_state structure's gp_state field. */
366#define RCU_GP_IDLE 0 /* Initial state and no GP in progress. */
367#define RCU_GP_WAIT_GPS 1 /* Wait for grace-period start. */
368#define RCU_GP_DONE_GPS 2 /* Wait done for grace-period start. */
369#define RCU_GP_ONOFF 3 /* Grace-period initialization hotplug. */
370#define RCU_GP_INIT 4 /* Grace-period initialization. */
371#define RCU_GP_WAIT_FQS 5 /* Wait for force-quiescent-state time. */
372#define RCU_GP_DOING_FQS 6 /* Wait done for force-quiescent-state time. */
373#define RCU_GP_CLEANUP 7 /* Grace-period cleanup started. */
374#define RCU_GP_CLEANED 8 /* Grace-period cleanup complete. */
375
376/*
377 * In order to export the rcu_state name to the tracing tools, it
378 * needs to be added in the __tracepoint_string section.
379 * This requires defining a separate variable tp_<sname>_varname
380 * that points to the string being used, and this will allow
381 * the tracing userspace tools to be able to decipher the string
382 * address to the matching string.
383 */
384#ifdef CONFIG_PREEMPT_RCU
385#define RCU_ABBR 'p'
386#define RCU_NAME_RAW "rcu_preempt"
387#else /* #ifdef CONFIG_PREEMPT_RCU */
388#define RCU_ABBR 's'
389#define RCU_NAME_RAW "rcu_sched"
390#endif /* #else #ifdef CONFIG_PREEMPT_RCU */
391#ifndef CONFIG_TRACING
392#define RCU_NAME RCU_NAME_RAW
393#else /* #ifdef CONFIG_TRACING */
394static char rcu_name[] = RCU_NAME_RAW;
395static const char *tp_rcu_varname __used __tracepoint_string = rcu_name;
396#define RCU_NAME rcu_name
397#endif /* #else #ifdef CONFIG_TRACING */
398
399/* Forward declarations for tree_plugin.h */
400static void rcu_bootup_announce(void);
401static void rcu_qs(void);
402static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp);
403#ifdef CONFIG_HOTPLUG_CPU
404static bool rcu_preempt_has_tasks(struct rcu_node *rnp);
405#endif /* #ifdef CONFIG_HOTPLUG_CPU */
406static int rcu_print_task_exp_stall(struct rcu_node *rnp);
407static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp);
408static void rcu_flavor_sched_clock_irq(int user);
409static void dump_blkd_tasks(struct rcu_node *rnp, int ncheck);
410static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags);
411static void rcu_preempt_boost_start_gp(struct rcu_node *rnp);
412static bool rcu_is_callbacks_kthread(void);
413static void rcu_cpu_kthread_setup(unsigned int cpu);
414static void __init rcu_spawn_boost_kthreads(void);
415static void rcu_prepare_kthreads(int cpu);
416static void rcu_cleanup_after_idle(void);
417static void rcu_prepare_for_idle(void);
418static bool rcu_preempt_has_tasks(struct rcu_node *rnp);
419static bool rcu_preempt_need_deferred_qs(struct task_struct *t);
420static void rcu_preempt_deferred_qs(struct task_struct *t);
421static void zero_cpu_stall_ticks(struct rcu_data *rdp);
422static struct swait_queue_head *rcu_nocb_gp_get(struct rcu_node *rnp);
423static void rcu_nocb_gp_cleanup(struct swait_queue_head *sq);
424static void rcu_init_one_nocb(struct rcu_node *rnp);
425static bool rcu_nocb_flush_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
426 unsigned long j);
427static bool rcu_nocb_try_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
428 bool *was_alldone, unsigned long flags);
429static void __call_rcu_nocb_wake(struct rcu_data *rdp, bool was_empty,
430 unsigned long flags);
431static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp);
432static void do_nocb_deferred_wakeup(struct rcu_data *rdp);
433static void rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp);
434static void rcu_spawn_cpu_nocb_kthread(int cpu);
435static void __init rcu_spawn_nocb_kthreads(void);
436static void show_rcu_nocb_state(struct rcu_data *rdp);
437static void rcu_nocb_lock(struct rcu_data *rdp);
438static void rcu_nocb_unlock(struct rcu_data *rdp);
439static void rcu_nocb_unlock_irqrestore(struct rcu_data *rdp,
440 unsigned long flags);
441static void rcu_lockdep_assert_cblist_protected(struct rcu_data *rdp);
442#ifdef CONFIG_RCU_NOCB_CPU
443static void __init rcu_organize_nocb_kthreads(void);
444#define rcu_nocb_lock_irqsave(rdp, flags) \
445do { \
446 if (!rcu_segcblist_is_offloaded(&(rdp)->cblist)) \
447 local_irq_save(flags); \
448 else \
449 raw_spin_lock_irqsave(&(rdp)->nocb_lock, (flags)); \
450} while (0)
451#else /* #ifdef CONFIG_RCU_NOCB_CPU */
452#define rcu_nocb_lock_irqsave(rdp, flags) local_irq_save(flags)
453#endif /* #else #ifdef CONFIG_RCU_NOCB_CPU */
454
455static void rcu_bind_gp_kthread(void);
456static bool rcu_nohz_full_cpu(void);
457static void rcu_dynticks_task_enter(void);
458static void rcu_dynticks_task_exit(void);
459static void rcu_dynticks_task_trace_enter(void);
460static void rcu_dynticks_task_trace_exit(void);
461
462/* Forward declarations for tree_stall.h */
463static void record_gp_stall_check_time(void);
464static void rcu_iw_handler(struct irq_work *iwp);
465static void check_cpu_stall(struct rcu_data *rdp);
466static void rcu_check_gp_start_stall(struct rcu_node *rnp, struct rcu_data *rdp,
467 const unsigned long gpssdelay);