xref: /linux/kernel/rcu/tree.h (revision b7019ac550eb3916f34d79db583e9b7ea2524afa)
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/stop_machine.h>
20 #include <linux/rcu_node_tree.h>
21 
22 #include "rcu_segcblist.h"
23 
24 /* Communicate arguments to a workqueue handler. */
25 struct rcu_exp_work {
26 	unsigned long rew_s;
27 	struct work_struct rew_work;
28 };
29 
30 /* RCU's kthread states for tracing. */
31 #define RCU_KTHREAD_STOPPED  0
32 #define RCU_KTHREAD_RUNNING  1
33 #define RCU_KTHREAD_WAITING  2
34 #define RCU_KTHREAD_OFFCPU   3
35 #define RCU_KTHREAD_YIELDING 4
36 #define RCU_KTHREAD_MAX      4
37 
38 /*
39  * Definition for node within the RCU grace-period-detection hierarchy.
40  */
41 struct rcu_node {
42 	raw_spinlock_t __private lock;	/* Root rcu_node's lock protects */
43 					/*  some rcu_state fields as well as */
44 					/*  following. */
45 	unsigned long gp_seq;	/* Track rsp->rcu_gp_seq. */
46 	unsigned long gp_seq_needed; /* Track furthest future GP request. */
47 	unsigned long completedqs; /* All QSes done for this node. */
48 	unsigned long qsmask;	/* CPUs or groups that need to switch in */
49 				/*  order for current grace period to proceed.*/
50 				/*  In leaf rcu_node, each bit corresponds to */
51 				/*  an rcu_data structure, otherwise, each */
52 				/*  bit corresponds to a child rcu_node */
53 				/*  structure. */
54 	unsigned long rcu_gp_init_mask;	/* Mask of offline CPUs at GP init. */
55 	unsigned long qsmaskinit;
56 				/* Per-GP initial value for qsmask. */
57 				/*  Initialized from ->qsmaskinitnext at the */
58 				/*  beginning of each grace period. */
59 	unsigned long qsmaskinitnext;
60 				/* Online CPUs for next grace period. */
61 	unsigned long expmask;	/* CPUs or groups that need to check in */
62 				/*  to allow the current expedited GP */
63 				/*  to complete. */
64 	unsigned long expmaskinit;
65 				/* Per-GP initial values for expmask. */
66 				/*  Initialized from ->expmaskinitnext at the */
67 				/*  beginning of each expedited GP. */
68 	unsigned long expmaskinitnext;
69 				/* Online CPUs for next expedited GP. */
70 				/*  Any CPU that has ever been online will */
71 				/*  have its bit set. */
72 	unsigned long ffmask;	/* Fully functional CPUs. */
73 	unsigned long grpmask;	/* Mask to apply to parent qsmask. */
74 				/*  Only one bit will be set in this mask. */
75 	int	grplo;		/* lowest-numbered CPU or group here. */
76 	int	grphi;		/* highest-numbered CPU or group here. */
77 	u8	grpnum;		/* CPU/group number for next level up. */
78 	u8	level;		/* root is at level 0. */
79 	bool	wait_blkd_tasks;/* Necessary to wait for blocked tasks to */
80 				/*  exit RCU read-side critical sections */
81 				/*  before propagating offline up the */
82 				/*  rcu_node tree? */
83 	struct rcu_node *parent;
84 	struct list_head blkd_tasks;
85 				/* Tasks blocked in RCU read-side critical */
86 				/*  section.  Tasks are placed at the head */
87 				/*  of this list and age towards the tail. */
88 	struct list_head *gp_tasks;
89 				/* Pointer to the first task blocking the */
90 				/*  current grace period, or NULL if there */
91 				/*  is no such task. */
92 	struct list_head *exp_tasks;
93 				/* Pointer to the first task blocking the */
94 				/*  current expedited grace period, or NULL */
95 				/*  if there is no such task.  If there */
96 				/*  is no current expedited grace period, */
97 				/*  then there can cannot be any such task. */
98 	struct list_head *boost_tasks;
99 				/* Pointer to first task that needs to be */
100 				/*  priority boosted, or NULL if no priority */
101 				/*  boosting is needed for this rcu_node */
102 				/*  structure.  If there are no tasks */
103 				/*  queued on this rcu_node structure that */
104 				/*  are blocking the current grace period, */
105 				/*  there can be no such task. */
106 	struct rt_mutex boost_mtx;
107 				/* Used only for the priority-boosting */
108 				/*  side effect, not as a lock. */
109 	unsigned long boost_time;
110 				/* When to start boosting (jiffies). */
111 	struct task_struct *boost_kthread_task;
112 				/* kthread that takes care of priority */
113 				/*  boosting for this rcu_node structure. */
114 	unsigned int boost_kthread_status;
115 				/* State of boost_kthread_task for tracing. */
116 #ifdef CONFIG_RCU_NOCB_CPU
117 	struct swait_queue_head nocb_gp_wq[2];
118 				/* Place for rcu_nocb_kthread() to wait GP. */
119 #endif /* #ifdef CONFIG_RCU_NOCB_CPU */
120 	raw_spinlock_t fqslock ____cacheline_internodealigned_in_smp;
121 
122 	spinlock_t exp_lock ____cacheline_internodealigned_in_smp;
123 	unsigned long exp_seq_rq;
124 	wait_queue_head_t exp_wq[4];
125 	struct rcu_exp_work rew;
126 	bool exp_need_flush;	/* Need to flush workitem? */
127 } ____cacheline_internodealigned_in_smp;
128 
129 /*
130  * Bitmasks in an rcu_node cover the interval [grplo, grphi] of CPU IDs, and
131  * are indexed relative to this interval rather than the global CPU ID space.
132  * This generates the bit for a CPU in node-local masks.
133  */
134 #define leaf_node_cpu_bit(rnp, cpu) (BIT((cpu) - (rnp)->grplo))
135 
136 /*
137  * Union to allow "aggregate OR" operation on the need for a quiescent
138  * state by the normal and expedited grace periods.
139  */
140 union rcu_noqs {
141 	struct {
142 		u8 norm;
143 		u8 exp;
144 	} b; /* Bits. */
145 	u16 s; /* Set of bits, aggregate OR here. */
146 };
147 
148 /* Per-CPU data for read-copy update. */
149 struct rcu_data {
150 	/* 1) quiescent-state and grace-period handling : */
151 	unsigned long	gp_seq;		/* Track rsp->rcu_gp_seq counter. */
152 	unsigned long	gp_seq_needed;	/* Track furthest future GP request. */
153 	union rcu_noqs	cpu_no_qs;	/* No QSes yet for this CPU. */
154 	bool		core_needs_qs;	/* Core waits for quiesc state. */
155 	bool		beenonline;	/* CPU online at least once. */
156 	bool		gpwrap;		/* Possible ->gp_seq wrap. */
157 	bool		deferred_qs;	/* This CPU awaiting a deferred QS? */
158 	struct rcu_node *mynode;	/* This CPU's leaf of hierarchy */
159 	unsigned long grpmask;		/* Mask to apply to leaf qsmask. */
160 	unsigned long	ticks_this_gp;	/* The number of scheduling-clock */
161 					/*  ticks this CPU has handled */
162 					/*  during and after the last grace */
163 					/* period it is aware of. */
164 
165 	/* 2) batch handling */
166 	struct rcu_segcblist cblist;	/* Segmented callback list, with */
167 					/* different callbacks waiting for */
168 					/* different grace periods. */
169 	long		qlen_last_fqs_check;
170 					/* qlen at last check for QS forcing */
171 	unsigned long	n_force_qs_snap;
172 					/* did other CPU force QS recently? */
173 	long		blimit;		/* Upper limit on a processed batch */
174 
175 	/* 3) dynticks interface. */
176 	int dynticks_snap;		/* Per-GP tracking for dynticks. */
177 	long dynticks_nesting;		/* Track process nesting level. */
178 	long dynticks_nmi_nesting;	/* Track irq/NMI nesting level. */
179 	atomic_t dynticks;		/* Even value for idle, else odd. */
180 	bool rcu_need_heavy_qs;		/* GP old, so heavy quiescent state! */
181 	bool rcu_urgent_qs;		/* GP old need light quiescent state. */
182 #ifdef CONFIG_RCU_FAST_NO_HZ
183 	bool all_lazy;			/* All CPU's CBs lazy at idle start? */
184 	unsigned long last_accelerate;	/* Last jiffy CBs were accelerated. */
185 	unsigned long last_advance_all;	/* Last jiffy CBs were all advanced. */
186 	int tick_nohz_enabled_snap;	/* Previously seen value from sysfs. */
187 #endif /* #ifdef CONFIG_RCU_FAST_NO_HZ */
188 
189 	/* 4) rcu_barrier(), OOM callbacks, and expediting. */
190 	struct rcu_head barrier_head;
191 	int exp_dynticks_snap;		/* Double-check need for IPI. */
192 
193 	/* 5) Callback offloading. */
194 #ifdef CONFIG_RCU_NOCB_CPU
195 	struct rcu_head *nocb_head;	/* CBs waiting for kthread. */
196 	struct rcu_head **nocb_tail;
197 	atomic_long_t nocb_q_count;	/* # CBs waiting for nocb */
198 	atomic_long_t nocb_q_count_lazy; /*  invocation (all stages). */
199 	struct rcu_head *nocb_follower_head; /* CBs ready to invoke. */
200 	struct rcu_head **nocb_follower_tail;
201 	struct swait_queue_head nocb_wq; /* For nocb kthreads to sleep on. */
202 	struct task_struct *nocb_kthread;
203 	raw_spinlock_t nocb_lock;	/* Guard following pair of fields. */
204 	int nocb_defer_wakeup;		/* Defer wakeup of nocb_kthread. */
205 	struct timer_list nocb_timer;	/* Enforce finite deferral. */
206 
207 	/* The following fields are used by the leader, hence own cacheline. */
208 	struct rcu_head *nocb_gp_head ____cacheline_internodealigned_in_smp;
209 					/* CBs waiting for GP. */
210 	struct rcu_head **nocb_gp_tail;
211 	bool nocb_leader_sleep;		/* Is the nocb leader thread asleep? */
212 	struct rcu_data *nocb_next_follower;
213 					/* Next follower in wakeup chain. */
214 
215 	/* The following fields are used by the follower, hence new cachline. */
216 	struct rcu_data *nocb_leader ____cacheline_internodealigned_in_smp;
217 					/* Leader CPU takes GP-end wakeups. */
218 #endif /* #ifdef CONFIG_RCU_NOCB_CPU */
219 
220 	/* 6) RCU priority boosting. */
221 	struct task_struct *rcu_cpu_kthread_task;
222 					/* rcuc per-CPU kthread or NULL. */
223 	unsigned int rcu_cpu_kthread_status;
224 	char rcu_cpu_has_work;
225 
226 	/* 7) Diagnostic data, including RCU CPU stall warnings. */
227 	unsigned int softirq_snap;	/* Snapshot of softirq activity. */
228 	/* ->rcu_iw* fields protected by leaf rcu_node ->lock. */
229 	struct irq_work rcu_iw;		/* Check for non-irq activity. */
230 	bool rcu_iw_pending;		/* Is ->rcu_iw pending? */
231 	unsigned long rcu_iw_gp_seq;	/* ->gp_seq associated with ->rcu_iw. */
232 	unsigned long rcu_ofl_gp_seq;	/* ->gp_seq at last offline. */
233 	short rcu_ofl_gp_flags;		/* ->gp_flags at last offline. */
234 	unsigned long rcu_onl_gp_seq;	/* ->gp_seq at last online. */
235 	short rcu_onl_gp_flags;		/* ->gp_flags at last online. */
236 	unsigned long last_fqs_resched;	/* Time of last rcu_resched(). */
237 
238 	int cpu;
239 };
240 
241 /* Values for nocb_defer_wakeup field in struct rcu_data. */
242 #define RCU_NOCB_WAKE_NOT	0
243 #define RCU_NOCB_WAKE		1
244 #define RCU_NOCB_WAKE_FORCE	2
245 
246 #define RCU_JIFFIES_TILL_FORCE_QS (1 + (HZ > 250) + (HZ > 500))
247 					/* For jiffies_till_first_fqs and */
248 					/*  and jiffies_till_next_fqs. */
249 
250 #define RCU_JIFFIES_FQS_DIV	256	/* Very large systems need more */
251 					/*  delay between bouts of */
252 					/*  quiescent-state forcing. */
253 
254 #define RCU_STALL_RAT_DELAY	2	/* Allow other CPUs time to take */
255 					/*  at least one scheduling clock */
256 					/*  irq before ratting on them. */
257 
258 #define rcu_wait(cond)							\
259 do {									\
260 	for (;;) {							\
261 		set_current_state(TASK_INTERRUPTIBLE);			\
262 		if (cond)						\
263 			break;						\
264 		schedule();						\
265 	}								\
266 	__set_current_state(TASK_RUNNING);				\
267 } while (0)
268 
269 /*
270  * RCU global state, including node hierarchy.  This hierarchy is
271  * represented in "heap" form in a dense array.  The root (first level)
272  * of the hierarchy is in ->node[0] (referenced by ->level[0]), the second
273  * level in ->node[1] through ->node[m] (->node[1] referenced by ->level[1]),
274  * and the third level in ->node[m+1] and following (->node[m+1] referenced
275  * by ->level[2]).  The number of levels is determined by the number of
276  * CPUs and by CONFIG_RCU_FANOUT.  Small systems will have a "hierarchy"
277  * consisting of a single rcu_node.
278  */
279 struct rcu_state {
280 	struct rcu_node node[NUM_RCU_NODES];	/* Hierarchy. */
281 	struct rcu_node *level[RCU_NUM_LVLS + 1];
282 						/* Hierarchy levels (+1 to */
283 						/*  shut bogus gcc warning) */
284 	int ncpus;				/* # CPUs seen so far. */
285 
286 	/* The following fields are guarded by the root rcu_node's lock. */
287 
288 	u8	boost ____cacheline_internodealigned_in_smp;
289 						/* Subject to priority boost. */
290 	unsigned long gp_seq;			/* Grace-period sequence #. */
291 	struct task_struct *gp_kthread;		/* Task for grace periods. */
292 	struct swait_queue_head gp_wq;		/* Where GP task waits. */
293 	short gp_flags;				/* Commands for GP task. */
294 	short gp_state;				/* GP kthread sleep state. */
295 	unsigned long gp_wake_time;		/* Last GP kthread wake. */
296 	unsigned long gp_wake_seq;		/* ->gp_seq at ^^^. */
297 
298 	/* End of fields guarded by root rcu_node's lock. */
299 
300 	struct mutex barrier_mutex;		/* Guards barrier fields. */
301 	atomic_t barrier_cpu_count;		/* # CPUs waiting on. */
302 	struct completion barrier_completion;	/* Wake at barrier end. */
303 	unsigned long barrier_sequence;		/* ++ at start and end of */
304 						/*  rcu_barrier(). */
305 	/* End of fields guarded by barrier_mutex. */
306 
307 	struct mutex exp_mutex;			/* Serialize expedited GP. */
308 	struct mutex exp_wake_mutex;		/* Serialize wakeup. */
309 	unsigned long expedited_sequence;	/* Take a ticket. */
310 	atomic_t expedited_need_qs;		/* # CPUs left to check in. */
311 	struct swait_queue_head expedited_wq;	/* Wait for check-ins. */
312 	int ncpus_snap;				/* # CPUs seen last time. */
313 
314 	unsigned long jiffies_force_qs;		/* Time at which to invoke */
315 						/*  force_quiescent_state(). */
316 	unsigned long jiffies_kick_kthreads;	/* Time at which to kick */
317 						/*  kthreads, if configured. */
318 	unsigned long n_force_qs;		/* Number of calls to */
319 						/*  force_quiescent_state(). */
320 	unsigned long gp_start;			/* Time at which GP started, */
321 						/*  but in jiffies. */
322 	unsigned long gp_end;			/* Time last GP ended, again */
323 						/*  in jiffies. */
324 	unsigned long gp_activity;		/* Time of last GP kthread */
325 						/*  activity in jiffies. */
326 	unsigned long gp_req_activity;		/* Time of last GP request */
327 						/*  in jiffies. */
328 	unsigned long jiffies_stall;		/* Time at which to check */
329 						/*  for CPU stalls. */
330 	unsigned long jiffies_resched;		/* Time at which to resched */
331 						/*  a reluctant CPU. */
332 	unsigned long n_force_qs_gpstart;	/* Snapshot of n_force_qs at */
333 						/*  GP start. */
334 	unsigned long gp_max;			/* Maximum GP duration in */
335 						/*  jiffies. */
336 	const char *name;			/* Name of structure. */
337 	char abbr;				/* Abbreviated name. */
338 
339 	raw_spinlock_t ofl_lock ____cacheline_internodealigned_in_smp;
340 						/* Synchronize offline with */
341 						/*  GP pre-initialization. */
342 };
343 
344 /* Values for rcu_state structure's gp_flags field. */
345 #define RCU_GP_FLAG_INIT 0x1	/* Need grace-period initialization. */
346 #define RCU_GP_FLAG_FQS  0x2	/* Need grace-period quiescent-state forcing. */
347 
348 /* Values for rcu_state structure's gp_state field. */
349 #define RCU_GP_IDLE	 0	/* Initial state and no GP in progress. */
350 #define RCU_GP_WAIT_GPS  1	/* Wait for grace-period start. */
351 #define RCU_GP_DONE_GPS  2	/* Wait done for grace-period start. */
352 #define RCU_GP_ONOFF     3	/* Grace-period initialization hotplug. */
353 #define RCU_GP_INIT      4	/* Grace-period initialization. */
354 #define RCU_GP_WAIT_FQS  5	/* Wait for force-quiescent-state time. */
355 #define RCU_GP_DOING_FQS 6	/* Wait done for force-quiescent-state time. */
356 #define RCU_GP_CLEANUP   7	/* Grace-period cleanup started. */
357 #define RCU_GP_CLEANED   8	/* Grace-period cleanup complete. */
358 
359 static const char * const gp_state_names[] = {
360 	"RCU_GP_IDLE",
361 	"RCU_GP_WAIT_GPS",
362 	"RCU_GP_DONE_GPS",
363 	"RCU_GP_ONOFF",
364 	"RCU_GP_INIT",
365 	"RCU_GP_WAIT_FQS",
366 	"RCU_GP_DOING_FQS",
367 	"RCU_GP_CLEANUP",
368 	"RCU_GP_CLEANED",
369 };
370 
371 /*
372  * In order to export the rcu_state name to the tracing tools, it
373  * needs to be added in the __tracepoint_string section.
374  * This requires defining a separate variable tp_<sname>_varname
375  * that points to the string being used, and this will allow
376  * the tracing userspace tools to be able to decipher the string
377  * address to the matching string.
378  */
379 #ifdef CONFIG_PREEMPT_RCU
380 #define RCU_ABBR 'p'
381 #define RCU_NAME_RAW "rcu_preempt"
382 #else /* #ifdef CONFIG_PREEMPT_RCU */
383 #define RCU_ABBR 's'
384 #define RCU_NAME_RAW "rcu_sched"
385 #endif /* #else #ifdef CONFIG_PREEMPT_RCU */
386 #ifndef CONFIG_TRACING
387 #define RCU_NAME RCU_NAME_RAW
388 #else /* #ifdef CONFIG_TRACING */
389 static char rcu_name[] = RCU_NAME_RAW;
390 static const char *tp_rcu_varname __used __tracepoint_string = rcu_name;
391 #define RCU_NAME rcu_name
392 #endif /* #else #ifdef CONFIG_TRACING */
393 
394 int rcu_dynticks_snap(struct rcu_data *rdp);
395 
396 /* Forward declarations for tree_plugin.h */
397 static void rcu_bootup_announce(void);
398 static void rcu_qs(void);
399 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp);
400 #ifdef CONFIG_HOTPLUG_CPU
401 static bool rcu_preempt_has_tasks(struct rcu_node *rnp);
402 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
403 static int rcu_print_task_exp_stall(struct rcu_node *rnp);
404 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp);
405 static void rcu_flavor_sched_clock_irq(int user);
406 void call_rcu(struct rcu_head *head, rcu_callback_t func);
407 static void dump_blkd_tasks(struct rcu_node *rnp, int ncheck);
408 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags);
409 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp);
410 static void invoke_rcu_callbacks_kthread(void);
411 static bool rcu_is_callbacks_kthread(void);
412 static void __init rcu_spawn_boost_kthreads(void);
413 static void rcu_prepare_kthreads(int cpu);
414 static void rcu_cleanup_after_idle(void);
415 static void rcu_prepare_for_idle(void);
416 static bool rcu_preempt_has_tasks(struct rcu_node *rnp);
417 static bool rcu_preempt_need_deferred_qs(struct task_struct *t);
418 static void rcu_preempt_deferred_qs(struct task_struct *t);
419 static void zero_cpu_stall_ticks(struct rcu_data *rdp);
420 static bool rcu_nocb_cpu_needs_barrier(int cpu);
421 static struct swait_queue_head *rcu_nocb_gp_get(struct rcu_node *rnp);
422 static void rcu_nocb_gp_cleanup(struct swait_queue_head *sq);
423 static void rcu_init_one_nocb(struct rcu_node *rnp);
424 static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,
425 			    bool lazy, unsigned long flags);
426 static bool rcu_nocb_adopt_orphan_cbs(struct rcu_data *my_rdp,
427 				      struct rcu_data *rdp,
428 				      unsigned long flags);
429 static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp);
430 static void do_nocb_deferred_wakeup(struct rcu_data *rdp);
431 static void rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp);
432 static void rcu_spawn_cpu_nocb_kthread(int cpu);
433 static void __init rcu_spawn_nocb_kthreads(void);
434 #ifdef CONFIG_RCU_NOCB_CPU
435 static void __init rcu_organize_nocb_kthreads(void);
436 #endif /* #ifdef CONFIG_RCU_NOCB_CPU */
437 static bool init_nocb_callback_list(struct rcu_data *rdp);
438 static unsigned long rcu_get_n_cbs_nocb_cpu(struct rcu_data *rdp);
439 static void rcu_bind_gp_kthread(void);
440 static bool rcu_nohz_full_cpu(void);
441 static void rcu_dynticks_task_enter(void);
442 static void rcu_dynticks_task_exit(void);
443 
444 /* Forward declarations for tree_stall.h */
445 static void record_gp_stall_check_time(void);
446 static void rcu_iw_handler(struct irq_work *iwp);
447 static void check_cpu_stall(struct rcu_data *rdp);
448 static void rcu_check_gp_start_stall(struct rcu_node *rnp, struct rcu_data *rdp,
449 				     const unsigned long gpssdelay);
450