1 /* SPDX-License-Identifier: GPL-2.0+ */
3 * Read-Copy Update mechanism for mutual exclusion (tree-based version)
4 * Internal non-public definitions.
6 * Copyright IBM Corporation, 2008
8 * Author: Ingo Molnar <mingo@elte.hu>
9 * Paul E. McKenney <paulmck@linux.ibm.com>
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>
22 #include "rcu_segcblist.h"
24 /* Communicate arguments to a workqueue handler. */
27 #ifdef CONFIG_RCU_EXP_KTHREAD
28 struct kthread_work rew_work;
30 struct work_struct rew_work;
31 #endif /* CONFIG_RCU_EXP_KTHREAD */
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
43 * Definition for node within the RCU grace-period-detection hierarchy.
46 raw_spinlock_t __private lock; /* Root rcu_node's lock protects */
47 /* some rcu_state fields as well as */
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 */
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 */
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 */
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 */
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;
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;
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.
147 #define leaf_node_cpu_bit(rnp, cpu) (BIT((cpu) - (rnp)->grplo))
150 * Union to allow "aggregate OR" operation on the need for a quiescent
151 * state by the normal and expedited grace periods.
158 u16 s; /* Set of bits, aggregate OR here. */
161 /* Per-CPU data for read-copy update. */
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. */
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 */
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. */
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. */
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 */
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. */
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,
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 */
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 */
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;
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(). */
269 /* Values for nocb_defer_wakeup field in struct rcu_data. */
270 #define RCU_NOCB_WAKE_NOT 0
271 #define RCU_NOCB_WAKE_BYPASS 1
272 #define RCU_NOCB_WAKE 2
273 #define RCU_NOCB_WAKE_FORCE 3
275 #define RCU_JIFFIES_TILL_FORCE_QS (1 + (HZ > 250) + (HZ > 500))
276 /* For jiffies_till_first_fqs and */
277 /* and jiffies_till_next_fqs. */
279 #define RCU_JIFFIES_FQS_DIV 256 /* Very large systems need more */
280 /* delay between bouts of */
281 /* quiescent-state forcing. */
283 #define RCU_STALL_RAT_DELAY 2 /* Allow other CPUs time to take */
284 /* at least one scheduling clock */
285 /* irq before ratting on them. */
287 #define rcu_wait(cond) \
290 set_current_state(TASK_INTERRUPTIBLE); \
295 __set_current_state(TASK_RUNNING); \
299 * RCU global state, including node hierarchy. This hierarchy is
300 * represented in "heap" form in a dense array. The root (first level)
301 * of the hierarchy is in ->node[0] (referenced by ->level[0]), the second
302 * level in ->node[1] through ->node[m] (->node[1] referenced by ->level[1]),
303 * and the third level in ->node[m+1] and following (->node[m+1] referenced
304 * by ->level[2]). The number of levels is determined by the number of
305 * CPUs and by CONFIG_RCU_FANOUT. Small systems will have a "hierarchy"
306 * consisting of a single rcu_node.
309 struct rcu_node node[NUM_RCU_NODES]; /* Hierarchy. */
310 struct rcu_node *level[RCU_NUM_LVLS + 1];
311 /* Hierarchy levels (+1 to */
312 /* shut bogus gcc warning) */
313 int ncpus; /* # CPUs seen so far. */
314 int n_online_cpus; /* # CPUs online for RCU. */
316 /* The following fields are guarded by the root rcu_node's lock. */
318 unsigned long gp_seq ____cacheline_internodealigned_in_smp;
319 /* Grace-period sequence #. */
320 unsigned long gp_max; /* Maximum GP duration in */
322 struct task_struct *gp_kthread; /* Task for grace periods. */
323 struct swait_queue_head gp_wq; /* Where GP task waits. */
324 short gp_flags; /* Commands for GP task. */
325 short gp_state; /* GP kthread sleep state. */
326 unsigned long gp_wake_time; /* Last GP kthread wake. */
327 unsigned long gp_wake_seq; /* ->gp_seq at ^^^. */
328 unsigned long gp_seq_polled; /* GP seq for polled API. */
329 unsigned long gp_seq_polled_snap; /* ->gp_seq_polled at normal GP start. */
330 unsigned long gp_seq_polled_exp_snap; /* ->gp_seq_polled at expedited GP start. */
332 /* End of fields guarded by root rcu_node's lock. */
334 struct mutex barrier_mutex; /* Guards barrier fields. */
335 atomic_t barrier_cpu_count; /* # CPUs waiting on. */
336 struct completion barrier_completion; /* Wake at barrier end. */
337 unsigned long barrier_sequence; /* ++ at start and end of */
339 /* End of fields guarded by barrier_mutex. */
341 raw_spinlock_t barrier_lock; /* Protects ->barrier_seq_snap. */
343 struct mutex exp_mutex; /* Serialize expedited GP. */
344 struct mutex exp_wake_mutex; /* Serialize wakeup. */
345 unsigned long expedited_sequence; /* Take a ticket. */
346 atomic_t expedited_need_qs; /* # CPUs left to check in. */
347 struct swait_queue_head expedited_wq; /* Wait for check-ins. */
348 int ncpus_snap; /* # CPUs seen last time. */
349 u8 cbovld; /* Callback overload now? */
350 u8 cbovldnext; /* ^ ^ next time? */
352 unsigned long jiffies_force_qs; /* Time at which to invoke */
353 /* force_quiescent_state(). */
354 unsigned long jiffies_kick_kthreads; /* Time at which to kick */
355 /* kthreads, if configured. */
356 unsigned long n_force_qs; /* Number of calls to */
357 /* force_quiescent_state(). */
358 unsigned long gp_start; /* Time at which GP started, */
359 /* but in jiffies. */
360 unsigned long gp_end; /* Time last GP ended, again */
362 unsigned long gp_activity; /* Time of last GP kthread */
363 /* activity in jiffies. */
364 unsigned long gp_req_activity; /* Time of last GP request */
366 unsigned long jiffies_stall; /* Time at which to check */
367 /* for CPU stalls. */
368 unsigned long jiffies_resched; /* Time at which to resched */
369 /* a reluctant CPU. */
370 unsigned long n_force_qs_gpstart; /* Snapshot of n_force_qs at */
372 const char *name; /* Name of structure. */
373 char abbr; /* Abbreviated name. */
375 arch_spinlock_t ofl_lock ____cacheline_internodealigned_in_smp;
376 /* Synchronize offline with */
377 /* GP pre-initialization. */
378 int nocb_is_setup; /* nocb is setup from boot */
381 /* Values for rcu_state structure's gp_flags field. */
382 #define RCU_GP_FLAG_INIT 0x1 /* Need grace-period initialization. */
383 #define RCU_GP_FLAG_FQS 0x2 /* Need grace-period quiescent-state forcing. */
384 #define RCU_GP_FLAG_OVLD 0x4 /* Experiencing callback overload. */
386 /* Values for rcu_state structure's gp_state field. */
387 #define RCU_GP_IDLE 0 /* Initial state and no GP in progress. */
388 #define RCU_GP_WAIT_GPS 1 /* Wait for grace-period start. */
389 #define RCU_GP_DONE_GPS 2 /* Wait done for grace-period start. */
390 #define RCU_GP_ONOFF 3 /* Grace-period initialization hotplug. */
391 #define RCU_GP_INIT 4 /* Grace-period initialization. */
392 #define RCU_GP_WAIT_FQS 5 /* Wait for force-quiescent-state time. */
393 #define RCU_GP_DOING_FQS 6 /* Wait done for force-quiescent-state time. */
394 #define RCU_GP_CLEANUP 7 /* Grace-period cleanup started. */
395 #define RCU_GP_CLEANED 8 /* Grace-period cleanup complete. */
398 * In order to export the rcu_state name to the tracing tools, it
399 * needs to be added in the __tracepoint_string section.
400 * This requires defining a separate variable tp_<sname>_varname
401 * that points to the string being used, and this will allow
402 * the tracing userspace tools to be able to decipher the string
403 * address to the matching string.
405 #ifdef CONFIG_PREEMPT_RCU
407 #define RCU_NAME_RAW "rcu_preempt"
408 #else /* #ifdef CONFIG_PREEMPT_RCU */
410 #define RCU_NAME_RAW "rcu_sched"
411 #endif /* #else #ifdef CONFIG_PREEMPT_RCU */
412 #ifndef CONFIG_TRACING
413 #define RCU_NAME RCU_NAME_RAW
414 #else /* #ifdef CONFIG_TRACING */
415 static char rcu_name[] = RCU_NAME_RAW;
416 static const char *tp_rcu_varname __used __tracepoint_string = rcu_name;
417 #define RCU_NAME rcu_name
418 #endif /* #else #ifdef CONFIG_TRACING */
420 /* Forward declarations for tree_plugin.h */
421 static void rcu_bootup_announce(void);
422 static void rcu_qs(void);
423 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp);
424 #ifdef CONFIG_HOTPLUG_CPU
425 static bool rcu_preempt_has_tasks(struct rcu_node *rnp);
426 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
427 static int rcu_print_task_exp_stall(struct rcu_node *rnp);
428 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp);
429 static void rcu_flavor_sched_clock_irq(int user);
430 static void dump_blkd_tasks(struct rcu_node *rnp, int ncheck);
431 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags);
432 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp);
433 static bool rcu_is_callbacks_kthread(struct rcu_data *rdp);
434 static void rcu_cpu_kthread_setup(unsigned int cpu);
435 static void rcu_spawn_one_boost_kthread(struct rcu_node *rnp);
436 static bool rcu_preempt_has_tasks(struct rcu_node *rnp);
437 static bool rcu_preempt_need_deferred_qs(struct task_struct *t);
438 static void zero_cpu_stall_ticks(struct rcu_data *rdp);
439 static struct swait_queue_head *rcu_nocb_gp_get(struct rcu_node *rnp);
440 static void rcu_nocb_gp_cleanup(struct swait_queue_head *sq);
441 static void rcu_init_one_nocb(struct rcu_node *rnp);
442 static bool rcu_nocb_flush_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
444 static bool rcu_nocb_try_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
445 bool *was_alldone, unsigned long flags);
446 static void __call_rcu_nocb_wake(struct rcu_data *rdp, bool was_empty,
447 unsigned long flags);
448 static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp, int level);
449 static bool do_nocb_deferred_wakeup(struct rcu_data *rdp);
450 static void rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp);
451 static void rcu_spawn_cpu_nocb_kthread(int cpu);
452 static void show_rcu_nocb_state(struct rcu_data *rdp);
453 static void rcu_nocb_lock(struct rcu_data *rdp);
454 static void rcu_nocb_unlock(struct rcu_data *rdp);
455 static void rcu_nocb_unlock_irqrestore(struct rcu_data *rdp,
456 unsigned long flags);
457 static void rcu_lockdep_assert_cblist_protected(struct rcu_data *rdp);
458 #ifdef CONFIG_RCU_NOCB_CPU
459 static void __init rcu_organize_nocb_kthreads(void);
462 * Disable IRQs before checking offloaded state so that local
463 * locking is safe against concurrent de-offloading.
465 #define rcu_nocb_lock_irqsave(rdp, flags) \
467 local_irq_save(flags); \
468 if (rcu_segcblist_is_offloaded(&(rdp)->cblist)) \
469 raw_spin_lock(&(rdp)->nocb_lock); \
471 #else /* #ifdef CONFIG_RCU_NOCB_CPU */
472 #define rcu_nocb_lock_irqsave(rdp, flags) local_irq_save(flags)
473 #endif /* #else #ifdef CONFIG_RCU_NOCB_CPU */
475 static void rcu_bind_gp_kthread(void);
476 static bool rcu_nohz_full_cpu(void);
478 /* Forward declarations for tree_stall.h */
479 static void record_gp_stall_check_time(void);
480 static void rcu_iw_handler(struct irq_work *iwp);
481 static void check_cpu_stall(struct rcu_data *rdp);
482 static void rcu_check_gp_start_stall(struct rcu_node *rnp, struct rcu_data *rdp,
483 const unsigned long gpssdelay);
485 /* Forward declarations for tree_exp.h. */
486 static void sync_rcu_do_polled_gp(struct work_struct *wp);