1 // SPDX-License-Identifier: GPL-2.0+
2 /*
3  * Sleepable Read-Copy Update mechanism for mutual exclusion.
4  *
5  * Copyright (C) IBM Corporation, 2006
6  * Copyright (C) Fujitsu, 2012
7  *
8  * Authors: Paul McKenney <paulmck@linux.ibm.com>
9  *	   Lai Jiangshan <laijs@cn.fujitsu.com>
10  *
11  * For detailed explanation of Read-Copy Update mechanism see -
12  *		Documentation/RCU/ *.txt
13  *
14  */
15 
16 #define pr_fmt(fmt) "rcu: " fmt
17 
18 #include <linux/export.h>
19 #include <linux/mutex.h>
20 #include <linux/percpu.h>
21 #include <linux/preempt.h>
22 #include <linux/rcupdate_wait.h>
23 #include <linux/sched.h>
24 #include <linux/smp.h>
25 #include <linux/delay.h>
26 #include <linux/module.h>
27 #include <linux/slab.h>
28 #include <linux/srcu.h>
29 
30 #include "rcu.h"
31 #include "rcu_segcblist.h"
32 
33 /* Holdoff in nanoseconds for auto-expediting. */
34 #define DEFAULT_SRCU_EXP_HOLDOFF (25 * 1000)
35 static ulong exp_holdoff = DEFAULT_SRCU_EXP_HOLDOFF;
36 module_param(exp_holdoff, ulong, 0444);
37 
38 /* Overflow-check frequency.  N bits roughly says every 2**N grace periods. */
39 static ulong counter_wrap_check = (ULONG_MAX >> 2);
40 module_param(counter_wrap_check, ulong, 0444);
41 
42 /*
43  * Control conversion to SRCU_SIZE_BIG:
44  *    0: Don't convert at all.
45  *    1: Convert at init_srcu_struct() time.
46  *    2: Convert when rcutorture invokes srcu_torture_stats_print().
47  *    3: Decide at boot time based on system shape (default).
48  * 0x1x: Convert when excessive contention encountered.
49  */
50 #define SRCU_SIZING_NONE	0
51 #define SRCU_SIZING_INIT	1
52 #define SRCU_SIZING_TORTURE	2
53 #define SRCU_SIZING_AUTO	3
54 #define SRCU_SIZING_CONTEND	0x10
55 #define SRCU_SIZING_IS(x) ((convert_to_big & ~SRCU_SIZING_CONTEND) == x)
56 #define SRCU_SIZING_IS_NONE() (SRCU_SIZING_IS(SRCU_SIZING_NONE))
57 #define SRCU_SIZING_IS_INIT() (SRCU_SIZING_IS(SRCU_SIZING_INIT))
58 #define SRCU_SIZING_IS_TORTURE() (SRCU_SIZING_IS(SRCU_SIZING_TORTURE))
59 #define SRCU_SIZING_IS_CONTEND() (convert_to_big & SRCU_SIZING_CONTEND)
60 static int convert_to_big = SRCU_SIZING_AUTO;
61 module_param(convert_to_big, int, 0444);
62 
63 /* Number of CPUs to trigger init_srcu_struct()-time transition to big. */
64 static int big_cpu_lim __read_mostly = 128;
65 module_param(big_cpu_lim, int, 0444);
66 
67 /* Contention events per jiffy to initiate transition to big. */
68 static int small_contention_lim __read_mostly = 100;
69 module_param(small_contention_lim, int, 0444);
70 
71 /* Early-boot callback-management, so early that no lock is required! */
72 static LIST_HEAD(srcu_boot_list);
73 static bool __read_mostly srcu_init_done;
74 
75 static void srcu_invoke_callbacks(struct work_struct *work);
76 static void srcu_reschedule(struct srcu_struct *ssp, unsigned long delay);
77 static void process_srcu(struct work_struct *work);
78 static void srcu_delay_timer(struct timer_list *t);
79 
80 /* Wrappers for lock acquisition and release, see raw_spin_lock_rcu_node(). */
81 #define spin_lock_rcu_node(p)							\
82 do {										\
83 	spin_lock(&ACCESS_PRIVATE(p, lock));					\
84 	smp_mb__after_unlock_lock();						\
85 } while (0)
86 
87 #define spin_unlock_rcu_node(p) spin_unlock(&ACCESS_PRIVATE(p, lock))
88 
89 #define spin_lock_irq_rcu_node(p)						\
90 do {										\
91 	spin_lock_irq(&ACCESS_PRIVATE(p, lock));				\
92 	smp_mb__after_unlock_lock();						\
93 } while (0)
94 
95 #define spin_unlock_irq_rcu_node(p)						\
96 	spin_unlock_irq(&ACCESS_PRIVATE(p, lock))
97 
98 #define spin_lock_irqsave_rcu_node(p, flags)					\
99 do {										\
100 	spin_lock_irqsave(&ACCESS_PRIVATE(p, lock), flags);			\
101 	smp_mb__after_unlock_lock();						\
102 } while (0)
103 
104 #define spin_trylock_irqsave_rcu_node(p, flags)					\
105 ({										\
106 	bool ___locked = spin_trylock_irqsave(&ACCESS_PRIVATE(p, lock), flags);	\
107 										\
108 	if (___locked)								\
109 		smp_mb__after_unlock_lock();					\
110 	___locked;								\
111 })
112 
113 #define spin_unlock_irqrestore_rcu_node(p, flags)				\
114 	spin_unlock_irqrestore(&ACCESS_PRIVATE(p, lock), flags)			\
115 
116 /*
117  * Initialize SRCU per-CPU data.  Note that statically allocated
118  * srcu_struct structures might already have srcu_read_lock() and
119  * srcu_read_unlock() running against them.  So if the is_static parameter
120  * is set, don't initialize ->srcu_lock_count[] and ->srcu_unlock_count[].
121  */
init_srcu_struct_data(struct srcu_struct * ssp)122 static void init_srcu_struct_data(struct srcu_struct *ssp)
123 {
124 	int cpu;
125 	struct srcu_data *sdp;
126 
127 	/*
128 	 * Initialize the per-CPU srcu_data array, which feeds into the
129 	 * leaves of the srcu_node tree.
130 	 */
131 	WARN_ON_ONCE(ARRAY_SIZE(sdp->srcu_lock_count) !=
132 		     ARRAY_SIZE(sdp->srcu_unlock_count));
133 	for_each_possible_cpu(cpu) {
134 		sdp = per_cpu_ptr(ssp->sda, cpu);
135 		spin_lock_init(&ACCESS_PRIVATE(sdp, lock));
136 		rcu_segcblist_init(&sdp->srcu_cblist);
137 		sdp->srcu_cblist_invoking = false;
138 		sdp->srcu_gp_seq_needed = ssp->srcu_gp_seq;
139 		sdp->srcu_gp_seq_needed_exp = ssp->srcu_gp_seq;
140 		sdp->mynode = NULL;
141 		sdp->cpu = cpu;
142 		INIT_WORK(&sdp->work, srcu_invoke_callbacks);
143 		timer_setup(&sdp->delay_work, srcu_delay_timer, 0);
144 		sdp->ssp = ssp;
145 	}
146 }
147 
148 /* Invalid seq state, used during snp node initialization */
149 #define SRCU_SNP_INIT_SEQ		0x2
150 
151 /*
152  * Check whether sequence number corresponding to snp node,
153  * is invalid.
154  */
srcu_invl_snp_seq(unsigned long s)155 static inline bool srcu_invl_snp_seq(unsigned long s)
156 {
157 	return rcu_seq_state(s) == SRCU_SNP_INIT_SEQ;
158 }
159 
160 /*
161  * Allocated and initialize SRCU combining tree.  Returns @true if
162  * allocation succeeded and @false otherwise.
163  */
init_srcu_struct_nodes(struct srcu_struct * ssp,gfp_t gfp_flags)164 static bool init_srcu_struct_nodes(struct srcu_struct *ssp, gfp_t gfp_flags)
165 {
166 	int cpu;
167 	int i;
168 	int level = 0;
169 	int levelspread[RCU_NUM_LVLS];
170 	struct srcu_data *sdp;
171 	struct srcu_node *snp;
172 	struct srcu_node *snp_first;
173 
174 	/* Initialize geometry if it has not already been initialized. */
175 	rcu_init_geometry();
176 	ssp->node = kcalloc(rcu_num_nodes, sizeof(*ssp->node), gfp_flags);
177 	if (!ssp->node)
178 		return false;
179 
180 	/* Work out the overall tree geometry. */
181 	ssp->level[0] = &ssp->node[0];
182 	for (i = 1; i < rcu_num_lvls; i++)
183 		ssp->level[i] = ssp->level[i - 1] + num_rcu_lvl[i - 1];
184 	rcu_init_levelspread(levelspread, num_rcu_lvl);
185 
186 	/* Each pass through this loop initializes one srcu_node structure. */
187 	srcu_for_each_node_breadth_first(ssp, snp) {
188 		spin_lock_init(&ACCESS_PRIVATE(snp, lock));
189 		WARN_ON_ONCE(ARRAY_SIZE(snp->srcu_have_cbs) !=
190 			     ARRAY_SIZE(snp->srcu_data_have_cbs));
191 		for (i = 0; i < ARRAY_SIZE(snp->srcu_have_cbs); i++) {
192 			snp->srcu_have_cbs[i] = SRCU_SNP_INIT_SEQ;
193 			snp->srcu_data_have_cbs[i] = 0;
194 		}
195 		snp->srcu_gp_seq_needed_exp = SRCU_SNP_INIT_SEQ;
196 		snp->grplo = -1;
197 		snp->grphi = -1;
198 		if (snp == &ssp->node[0]) {
199 			/* Root node, special case. */
200 			snp->srcu_parent = NULL;
201 			continue;
202 		}
203 
204 		/* Non-root node. */
205 		if (snp == ssp->level[level + 1])
206 			level++;
207 		snp->srcu_parent = ssp->level[level - 1] +
208 				   (snp - ssp->level[level]) /
209 				   levelspread[level - 1];
210 	}
211 
212 	/*
213 	 * Initialize the per-CPU srcu_data array, which feeds into the
214 	 * leaves of the srcu_node tree.
215 	 */
216 	level = rcu_num_lvls - 1;
217 	snp_first = ssp->level[level];
218 	for_each_possible_cpu(cpu) {
219 		sdp = per_cpu_ptr(ssp->sda, cpu);
220 		sdp->mynode = &snp_first[cpu / levelspread[level]];
221 		for (snp = sdp->mynode; snp != NULL; snp = snp->srcu_parent) {
222 			if (snp->grplo < 0)
223 				snp->grplo = cpu;
224 			snp->grphi = cpu;
225 		}
226 		sdp->grpmask = 1 << (cpu - sdp->mynode->grplo);
227 	}
228 	smp_store_release(&ssp->srcu_size_state, SRCU_SIZE_WAIT_BARRIER);
229 	return true;
230 }
231 
232 /*
233  * Initialize non-compile-time initialized fields, including the
234  * associated srcu_node and srcu_data structures.  The is_static parameter
235  * tells us that ->sda has already been wired up to srcu_data.
236  */
init_srcu_struct_fields(struct srcu_struct * ssp,bool is_static)237 static int init_srcu_struct_fields(struct srcu_struct *ssp, bool is_static)
238 {
239 	ssp->srcu_size_state = SRCU_SIZE_SMALL;
240 	ssp->node = NULL;
241 	mutex_init(&ssp->srcu_cb_mutex);
242 	mutex_init(&ssp->srcu_gp_mutex);
243 	ssp->srcu_idx = 0;
244 	ssp->srcu_gp_seq = 0;
245 	ssp->srcu_barrier_seq = 0;
246 	mutex_init(&ssp->srcu_barrier_mutex);
247 	atomic_set(&ssp->srcu_barrier_cpu_cnt, 0);
248 	INIT_DELAYED_WORK(&ssp->work, process_srcu);
249 	ssp->sda_is_static = is_static;
250 	if (!is_static)
251 		ssp->sda = alloc_percpu(struct srcu_data);
252 	if (!ssp->sda)
253 		return -ENOMEM;
254 	init_srcu_struct_data(ssp);
255 	ssp->srcu_gp_seq_needed_exp = 0;
256 	ssp->srcu_last_gp_end = ktime_get_mono_fast_ns();
257 	if (READ_ONCE(ssp->srcu_size_state) == SRCU_SIZE_SMALL && SRCU_SIZING_IS_INIT()) {
258 		if (!init_srcu_struct_nodes(ssp, GFP_ATOMIC)) {
259 			if (!ssp->sda_is_static) {
260 				free_percpu(ssp->sda);
261 				ssp->sda = NULL;
262 				return -ENOMEM;
263 			}
264 		} else {
265 			WRITE_ONCE(ssp->srcu_size_state, SRCU_SIZE_BIG);
266 		}
267 	}
268 	smp_store_release(&ssp->srcu_gp_seq_needed, 0); /* Init done. */
269 	return 0;
270 }
271 
272 #ifdef CONFIG_DEBUG_LOCK_ALLOC
273 
__init_srcu_struct(struct srcu_struct * ssp,const char * name,struct lock_class_key * key)274 int __init_srcu_struct(struct srcu_struct *ssp, const char *name,
275 		       struct lock_class_key *key)
276 {
277 	/* Don't re-initialize a lock while it is held. */
278 	debug_check_no_locks_freed((void *)ssp, sizeof(*ssp));
279 	lockdep_init_map(&ssp->dep_map, name, key, 0);
280 	spin_lock_init(&ACCESS_PRIVATE(ssp, lock));
281 	return init_srcu_struct_fields(ssp, false);
282 }
283 EXPORT_SYMBOL_GPL(__init_srcu_struct);
284 
285 #else /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */
286 
287 /**
288  * init_srcu_struct - initialize a sleep-RCU structure
289  * @ssp: structure to initialize.
290  *
291  * Must invoke this on a given srcu_struct before passing that srcu_struct
292  * to any other function.  Each srcu_struct represents a separate domain
293  * of SRCU protection.
294  */
init_srcu_struct(struct srcu_struct * ssp)295 int init_srcu_struct(struct srcu_struct *ssp)
296 {
297 	spin_lock_init(&ACCESS_PRIVATE(ssp, lock));
298 	return init_srcu_struct_fields(ssp, false);
299 }
300 EXPORT_SYMBOL_GPL(init_srcu_struct);
301 
302 #endif /* #else #ifdef CONFIG_DEBUG_LOCK_ALLOC */
303 
304 /*
305  * Initiate a transition to SRCU_SIZE_BIG with lock held.
306  */
__srcu_transition_to_big(struct srcu_struct * ssp)307 static void __srcu_transition_to_big(struct srcu_struct *ssp)
308 {
309 	lockdep_assert_held(&ACCESS_PRIVATE(ssp, lock));
310 	smp_store_release(&ssp->srcu_size_state, SRCU_SIZE_ALLOC);
311 }
312 
313 /*
314  * Initiate an idempotent transition to SRCU_SIZE_BIG.
315  */
srcu_transition_to_big(struct srcu_struct * ssp)316 static void srcu_transition_to_big(struct srcu_struct *ssp)
317 {
318 	unsigned long flags;
319 
320 	/* Double-checked locking on ->srcu_size-state. */
321 	if (smp_load_acquire(&ssp->srcu_size_state) != SRCU_SIZE_SMALL)
322 		return;
323 	spin_lock_irqsave_rcu_node(ssp, flags);
324 	if (smp_load_acquire(&ssp->srcu_size_state) != SRCU_SIZE_SMALL) {
325 		spin_unlock_irqrestore_rcu_node(ssp, flags);
326 		return;
327 	}
328 	__srcu_transition_to_big(ssp);
329 	spin_unlock_irqrestore_rcu_node(ssp, flags);
330 }
331 
332 /*
333  * Check to see if the just-encountered contention event justifies
334  * a transition to SRCU_SIZE_BIG.
335  */
spin_lock_irqsave_check_contention(struct srcu_struct * ssp)336 static void spin_lock_irqsave_check_contention(struct srcu_struct *ssp)
337 {
338 	unsigned long j;
339 
340 	if (!SRCU_SIZING_IS_CONTEND() || ssp->srcu_size_state)
341 		return;
342 	j = jiffies;
343 	if (ssp->srcu_size_jiffies != j) {
344 		ssp->srcu_size_jiffies = j;
345 		ssp->srcu_n_lock_retries = 0;
346 	}
347 	if (++ssp->srcu_n_lock_retries <= small_contention_lim)
348 		return;
349 	__srcu_transition_to_big(ssp);
350 }
351 
352 /*
353  * Acquire the specified srcu_data structure's ->lock, but check for
354  * excessive contention, which results in initiation of a transition
355  * to SRCU_SIZE_BIG.  But only if the srcutree.convert_to_big module
356  * parameter permits this.
357  */
spin_lock_irqsave_sdp_contention(struct srcu_data * sdp,unsigned long * flags)358 static void spin_lock_irqsave_sdp_contention(struct srcu_data *sdp, unsigned long *flags)
359 {
360 	struct srcu_struct *ssp = sdp->ssp;
361 
362 	if (spin_trylock_irqsave_rcu_node(sdp, *flags))
363 		return;
364 	spin_lock_irqsave_rcu_node(ssp, *flags);
365 	spin_lock_irqsave_check_contention(ssp);
366 	spin_unlock_irqrestore_rcu_node(ssp, *flags);
367 	spin_lock_irqsave_rcu_node(sdp, *flags);
368 }
369 
370 /*
371  * Acquire the specified srcu_struct structure's ->lock, but check for
372  * excessive contention, which results in initiation of a transition
373  * to SRCU_SIZE_BIG.  But only if the srcutree.convert_to_big module
374  * parameter permits this.
375  */
spin_lock_irqsave_ssp_contention(struct srcu_struct * ssp,unsigned long * flags)376 static void spin_lock_irqsave_ssp_contention(struct srcu_struct *ssp, unsigned long *flags)
377 {
378 	if (spin_trylock_irqsave_rcu_node(ssp, *flags))
379 		return;
380 	spin_lock_irqsave_rcu_node(ssp, *flags);
381 	spin_lock_irqsave_check_contention(ssp);
382 }
383 
384 /*
385  * First-use initialization of statically allocated srcu_struct
386  * structure.  Wiring up the combining tree is more than can be
387  * done with compile-time initialization, so this check is added
388  * to each update-side SRCU primitive.  Use ssp->lock, which -is-
389  * compile-time initialized, to resolve races involving multiple
390  * CPUs trying to garner first-use privileges.
391  */
check_init_srcu_struct(struct srcu_struct * ssp)392 static void check_init_srcu_struct(struct srcu_struct *ssp)
393 {
394 	unsigned long flags;
395 
396 	/* The smp_load_acquire() pairs with the smp_store_release(). */
397 	if (!rcu_seq_state(smp_load_acquire(&ssp->srcu_gp_seq_needed))) /*^^^*/
398 		return; /* Already initialized. */
399 	spin_lock_irqsave_rcu_node(ssp, flags);
400 	if (!rcu_seq_state(ssp->srcu_gp_seq_needed)) {
401 		spin_unlock_irqrestore_rcu_node(ssp, flags);
402 		return;
403 	}
404 	init_srcu_struct_fields(ssp, true);
405 	spin_unlock_irqrestore_rcu_node(ssp, flags);
406 }
407 
408 /*
409  * Returns approximate total of the readers' ->srcu_lock_count[] values
410  * for the rank of per-CPU counters specified by idx.
411  */
srcu_readers_lock_idx(struct srcu_struct * ssp,int idx)412 static unsigned long srcu_readers_lock_idx(struct srcu_struct *ssp, int idx)
413 {
414 	int cpu;
415 	unsigned long sum = 0;
416 
417 	for_each_possible_cpu(cpu) {
418 		struct srcu_data *cpuc = per_cpu_ptr(ssp->sda, cpu);
419 
420 		sum += READ_ONCE(cpuc->srcu_lock_count[idx]);
421 	}
422 	return sum;
423 }
424 
425 /*
426  * Returns approximate total of the readers' ->srcu_unlock_count[] values
427  * for the rank of per-CPU counters specified by idx.
428  */
srcu_readers_unlock_idx(struct srcu_struct * ssp,int idx)429 static unsigned long srcu_readers_unlock_idx(struct srcu_struct *ssp, int idx)
430 {
431 	int cpu;
432 	unsigned long sum = 0;
433 
434 	for_each_possible_cpu(cpu) {
435 		struct srcu_data *cpuc = per_cpu_ptr(ssp->sda, cpu);
436 
437 		sum += READ_ONCE(cpuc->srcu_unlock_count[idx]);
438 	}
439 	return sum;
440 }
441 
442 /*
443  * Return true if the number of pre-existing readers is determined to
444  * be zero.
445  */
srcu_readers_active_idx_check(struct srcu_struct * ssp,int idx)446 static bool srcu_readers_active_idx_check(struct srcu_struct *ssp, int idx)
447 {
448 	unsigned long unlocks;
449 
450 	unlocks = srcu_readers_unlock_idx(ssp, idx);
451 
452 	/*
453 	 * Make sure that a lock is always counted if the corresponding
454 	 * unlock is counted. Needs to be a smp_mb() as the read side may
455 	 * contain a read from a variable that is written to before the
456 	 * synchronize_srcu() in the write side. In this case smp_mb()s
457 	 * A and B act like the store buffering pattern.
458 	 *
459 	 * This smp_mb() also pairs with smp_mb() C to prevent accesses
460 	 * after the synchronize_srcu() from being executed before the
461 	 * grace period ends.
462 	 */
463 	smp_mb(); /* A */
464 
465 	/*
466 	 * If the locks are the same as the unlocks, then there must have
467 	 * been no readers on this index at some time in between. This does
468 	 * not mean that there are no more readers, as one could have read
469 	 * the current index but not have incremented the lock counter yet.
470 	 *
471 	 * So suppose that the updater is preempted here for so long
472 	 * that more than ULONG_MAX non-nested readers come and go in
473 	 * the meantime.  It turns out that this cannot result in overflow
474 	 * because if a reader modifies its unlock count after we read it
475 	 * above, then that reader's next load of ->srcu_idx is guaranteed
476 	 * to get the new value, which will cause it to operate on the
477 	 * other bank of counters, where it cannot contribute to the
478 	 * overflow of these counters.  This means that there is a maximum
479 	 * of 2*NR_CPUS increments, which cannot overflow given current
480 	 * systems, especially not on 64-bit systems.
481 	 *
482 	 * OK, how about nesting?  This does impose a limit on nesting
483 	 * of floor(ULONG_MAX/NR_CPUS/2), which should be sufficient,
484 	 * especially on 64-bit systems.
485 	 */
486 	return srcu_readers_lock_idx(ssp, idx) == unlocks;
487 }
488 
489 /**
490  * srcu_readers_active - returns true if there are readers. and false
491  *                       otherwise
492  * @ssp: which srcu_struct to count active readers (holding srcu_read_lock).
493  *
494  * Note that this is not an atomic primitive, and can therefore suffer
495  * severe errors when invoked on an active srcu_struct.  That said, it
496  * can be useful as an error check at cleanup time.
497  */
srcu_readers_active(struct srcu_struct * ssp)498 static bool srcu_readers_active(struct srcu_struct *ssp)
499 {
500 	int cpu;
501 	unsigned long sum = 0;
502 
503 	for_each_possible_cpu(cpu) {
504 		struct srcu_data *cpuc = per_cpu_ptr(ssp->sda, cpu);
505 
506 		sum += READ_ONCE(cpuc->srcu_lock_count[0]);
507 		sum += READ_ONCE(cpuc->srcu_lock_count[1]);
508 		sum -= READ_ONCE(cpuc->srcu_unlock_count[0]);
509 		sum -= READ_ONCE(cpuc->srcu_unlock_count[1]);
510 	}
511 	return sum;
512 }
513 
514 /*
515  * We use an adaptive strategy for synchronize_srcu() and especially for
516  * synchronize_srcu_expedited().  We spin for a fixed time period
517  * (defined below, boot time configurable) to allow SRCU readers to exit
518  * their read-side critical sections.  If there are still some readers
519  * after one jiffy, we repeatedly block for one jiffy time periods.
520  * The blocking time is increased as the grace-period age increases,
521  * with max blocking time capped at 10 jiffies.
522  */
523 #define SRCU_DEFAULT_RETRY_CHECK_DELAY		5
524 
525 static ulong srcu_retry_check_delay = SRCU_DEFAULT_RETRY_CHECK_DELAY;
526 module_param(srcu_retry_check_delay, ulong, 0444);
527 
528 #define SRCU_INTERVAL		1		// Base delay if no expedited GPs pending.
529 #define SRCU_MAX_INTERVAL	10		// Maximum incremental delay from slow readers.
530 
531 #define SRCU_DEFAULT_MAX_NODELAY_PHASE_LO	3UL	// Lowmark on default per-GP-phase
532 							// no-delay instances.
533 #define SRCU_DEFAULT_MAX_NODELAY_PHASE_HI	1000UL	// Highmark on default per-GP-phase
534 							// no-delay instances.
535 
536 #define SRCU_UL_CLAMP_LO(val, low)	((val) > (low) ? (val) : (low))
537 #define SRCU_UL_CLAMP_HI(val, high)	((val) < (high) ? (val) : (high))
538 #define SRCU_UL_CLAMP(val, low, high)	SRCU_UL_CLAMP_HI(SRCU_UL_CLAMP_LO((val), (low)), (high))
539 // per-GP-phase no-delay instances adjusted to allow non-sleeping poll upto
540 // one jiffies time duration. Mult by 2 is done to factor in the srcu_get_delay()
541 // called from process_srcu().
542 #define SRCU_DEFAULT_MAX_NODELAY_PHASE_ADJUSTED	\
543 	(2UL * USEC_PER_SEC / HZ / SRCU_DEFAULT_RETRY_CHECK_DELAY)
544 
545 // Maximum per-GP-phase consecutive no-delay instances.
546 #define SRCU_DEFAULT_MAX_NODELAY_PHASE	\
547 	SRCU_UL_CLAMP(SRCU_DEFAULT_MAX_NODELAY_PHASE_ADJUSTED,	\
548 		      SRCU_DEFAULT_MAX_NODELAY_PHASE_LO,	\
549 		      SRCU_DEFAULT_MAX_NODELAY_PHASE_HI)
550 
551 static ulong srcu_max_nodelay_phase = SRCU_DEFAULT_MAX_NODELAY_PHASE;
552 module_param(srcu_max_nodelay_phase, ulong, 0444);
553 
554 // Maximum consecutive no-delay instances.
555 #define SRCU_DEFAULT_MAX_NODELAY	(SRCU_DEFAULT_MAX_NODELAY_PHASE > 100 ?	\
556 					 SRCU_DEFAULT_MAX_NODELAY_PHASE : 100)
557 
558 static ulong srcu_max_nodelay = SRCU_DEFAULT_MAX_NODELAY;
559 module_param(srcu_max_nodelay, ulong, 0444);
560 
561 /*
562  * Return grace-period delay, zero if there are expedited grace
563  * periods pending, SRCU_INTERVAL otherwise.
564  */
srcu_get_delay(struct srcu_struct * ssp)565 static unsigned long srcu_get_delay(struct srcu_struct *ssp)
566 {
567 	unsigned long gpstart;
568 	unsigned long j;
569 	unsigned long jbase = SRCU_INTERVAL;
570 
571 	if (ULONG_CMP_LT(READ_ONCE(ssp->srcu_gp_seq), READ_ONCE(ssp->srcu_gp_seq_needed_exp)))
572 		jbase = 0;
573 	if (rcu_seq_state(READ_ONCE(ssp->srcu_gp_seq))) {
574 		j = jiffies - 1;
575 		gpstart = READ_ONCE(ssp->srcu_gp_start);
576 		if (time_after(j, gpstart))
577 			jbase += j - gpstart;
578 		if (!jbase) {
579 			WRITE_ONCE(ssp->srcu_n_exp_nodelay, READ_ONCE(ssp->srcu_n_exp_nodelay) + 1);
580 			if (READ_ONCE(ssp->srcu_n_exp_nodelay) > srcu_max_nodelay_phase)
581 				jbase = 1;
582 		}
583 	}
584 	return jbase > SRCU_MAX_INTERVAL ? SRCU_MAX_INTERVAL : jbase;
585 }
586 
587 /**
588  * cleanup_srcu_struct - deconstruct a sleep-RCU structure
589  * @ssp: structure to clean up.
590  *
591  * Must invoke this after you are finished using a given srcu_struct that
592  * was initialized via init_srcu_struct(), else you leak memory.
593  */
cleanup_srcu_struct(struct srcu_struct * ssp)594 void cleanup_srcu_struct(struct srcu_struct *ssp)
595 {
596 	int cpu;
597 
598 	if (WARN_ON(!srcu_get_delay(ssp)))
599 		return; /* Just leak it! */
600 	if (WARN_ON(srcu_readers_active(ssp)))
601 		return; /* Just leak it! */
602 	flush_delayed_work(&ssp->work);
603 	for_each_possible_cpu(cpu) {
604 		struct srcu_data *sdp = per_cpu_ptr(ssp->sda, cpu);
605 
606 		del_timer_sync(&sdp->delay_work);
607 		flush_work(&sdp->work);
608 		if (WARN_ON(rcu_segcblist_n_cbs(&sdp->srcu_cblist)))
609 			return; /* Forgot srcu_barrier(), so just leak it! */
610 	}
611 	if (WARN_ON(rcu_seq_state(READ_ONCE(ssp->srcu_gp_seq)) != SRCU_STATE_IDLE) ||
612 	    WARN_ON(rcu_seq_current(&ssp->srcu_gp_seq) != ssp->srcu_gp_seq_needed) ||
613 	    WARN_ON(srcu_readers_active(ssp))) {
614 		pr_info("%s: Active srcu_struct %p read state: %d gp state: %lu/%lu\n",
615 			__func__, ssp, rcu_seq_state(READ_ONCE(ssp->srcu_gp_seq)),
616 			rcu_seq_current(&ssp->srcu_gp_seq), ssp->srcu_gp_seq_needed);
617 		return; /* Caller forgot to stop doing call_srcu()? */
618 	}
619 	if (!ssp->sda_is_static) {
620 		free_percpu(ssp->sda);
621 		ssp->sda = NULL;
622 	}
623 	kfree(ssp->node);
624 	ssp->node = NULL;
625 	ssp->srcu_size_state = SRCU_SIZE_SMALL;
626 }
627 EXPORT_SYMBOL_GPL(cleanup_srcu_struct);
628 
629 /*
630  * Counts the new reader in the appropriate per-CPU element of the
631  * srcu_struct.
632  * Returns an index that must be passed to the matching srcu_read_unlock().
633  */
__srcu_read_lock(struct srcu_struct * ssp)634 int __srcu_read_lock(struct srcu_struct *ssp)
635 {
636 	int idx;
637 
638 	idx = READ_ONCE(ssp->srcu_idx) & 0x1;
639 	this_cpu_inc(ssp->sda->srcu_lock_count[idx]);
640 	smp_mb(); /* B */  /* Avoid leaking the critical section. */
641 	return idx;
642 }
643 EXPORT_SYMBOL_GPL(__srcu_read_lock);
644 
645 /*
646  * Removes the count for the old reader from the appropriate per-CPU
647  * element of the srcu_struct.  Note that this may well be a different
648  * CPU than that which was incremented by the corresponding srcu_read_lock().
649  */
__srcu_read_unlock(struct srcu_struct * ssp,int idx)650 void __srcu_read_unlock(struct srcu_struct *ssp, int idx)
651 {
652 	smp_mb(); /* C */  /* Avoid leaking the critical section. */
653 	this_cpu_inc(ssp->sda->srcu_unlock_count[idx]);
654 }
655 EXPORT_SYMBOL_GPL(__srcu_read_unlock);
656 
657 /*
658  * Start an SRCU grace period.
659  */
srcu_gp_start(struct srcu_struct * ssp)660 static void srcu_gp_start(struct srcu_struct *ssp)
661 {
662 	struct srcu_data *sdp;
663 	int state;
664 
665 	if (smp_load_acquire(&ssp->srcu_size_state) < SRCU_SIZE_WAIT_BARRIER)
666 		sdp = per_cpu_ptr(ssp->sda, 0);
667 	else
668 		sdp = this_cpu_ptr(ssp->sda);
669 	lockdep_assert_held(&ACCESS_PRIVATE(ssp, lock));
670 	WARN_ON_ONCE(ULONG_CMP_GE(ssp->srcu_gp_seq, ssp->srcu_gp_seq_needed));
671 	spin_lock_rcu_node(sdp);  /* Interrupts already disabled. */
672 	rcu_segcblist_advance(&sdp->srcu_cblist,
673 			      rcu_seq_current(&ssp->srcu_gp_seq));
674 	(void)rcu_segcblist_accelerate(&sdp->srcu_cblist,
675 				       rcu_seq_snap(&ssp->srcu_gp_seq));
676 	spin_unlock_rcu_node(sdp);  /* Interrupts remain disabled. */
677 	WRITE_ONCE(ssp->srcu_gp_start, jiffies);
678 	WRITE_ONCE(ssp->srcu_n_exp_nodelay, 0);
679 	smp_mb(); /* Order prior store to ->srcu_gp_seq_needed vs. GP start. */
680 	rcu_seq_start(&ssp->srcu_gp_seq);
681 	state = rcu_seq_state(ssp->srcu_gp_seq);
682 	WARN_ON_ONCE(state != SRCU_STATE_SCAN1);
683 }
684 
685 
srcu_delay_timer(struct timer_list * t)686 static void srcu_delay_timer(struct timer_list *t)
687 {
688 	struct srcu_data *sdp = container_of(t, struct srcu_data, delay_work);
689 
690 	queue_work_on(sdp->cpu, rcu_gp_wq, &sdp->work);
691 }
692 
srcu_queue_delayed_work_on(struct srcu_data * sdp,unsigned long delay)693 static void srcu_queue_delayed_work_on(struct srcu_data *sdp,
694 				       unsigned long delay)
695 {
696 	if (!delay) {
697 		queue_work_on(sdp->cpu, rcu_gp_wq, &sdp->work);
698 		return;
699 	}
700 
701 	timer_reduce(&sdp->delay_work, jiffies + delay);
702 }
703 
704 /*
705  * Schedule callback invocation for the specified srcu_data structure,
706  * if possible, on the corresponding CPU.
707  */
srcu_schedule_cbs_sdp(struct srcu_data * sdp,unsigned long delay)708 static void srcu_schedule_cbs_sdp(struct srcu_data *sdp, unsigned long delay)
709 {
710 	srcu_queue_delayed_work_on(sdp, delay);
711 }
712 
713 /*
714  * Schedule callback invocation for all srcu_data structures associated
715  * with the specified srcu_node structure that have callbacks for the
716  * just-completed grace period, the one corresponding to idx.  If possible,
717  * schedule this invocation on the corresponding CPUs.
718  */
srcu_schedule_cbs_snp(struct srcu_struct * ssp,struct srcu_node * snp,unsigned long mask,unsigned long delay)719 static void srcu_schedule_cbs_snp(struct srcu_struct *ssp, struct srcu_node *snp,
720 				  unsigned long mask, unsigned long delay)
721 {
722 	int cpu;
723 
724 	for (cpu = snp->grplo; cpu <= snp->grphi; cpu++) {
725 		if (!(mask & (1 << (cpu - snp->grplo))))
726 			continue;
727 		srcu_schedule_cbs_sdp(per_cpu_ptr(ssp->sda, cpu), delay);
728 	}
729 }
730 
731 /*
732  * Note the end of an SRCU grace period.  Initiates callback invocation
733  * and starts a new grace period if needed.
734  *
735  * The ->srcu_cb_mutex acquisition does not protect any data, but
736  * instead prevents more than one grace period from starting while we
737  * are initiating callback invocation.  This allows the ->srcu_have_cbs[]
738  * array to have a finite number of elements.
739  */
srcu_gp_end(struct srcu_struct * ssp)740 static void srcu_gp_end(struct srcu_struct *ssp)
741 {
742 	unsigned long cbdelay = 1;
743 	bool cbs;
744 	bool last_lvl;
745 	int cpu;
746 	unsigned long flags;
747 	unsigned long gpseq;
748 	int idx;
749 	unsigned long mask;
750 	struct srcu_data *sdp;
751 	unsigned long sgsne;
752 	struct srcu_node *snp;
753 	int ss_state;
754 
755 	/* Prevent more than one additional grace period. */
756 	mutex_lock(&ssp->srcu_cb_mutex);
757 
758 	/* End the current grace period. */
759 	spin_lock_irq_rcu_node(ssp);
760 	idx = rcu_seq_state(ssp->srcu_gp_seq);
761 	WARN_ON_ONCE(idx != SRCU_STATE_SCAN2);
762 	if (ULONG_CMP_LT(READ_ONCE(ssp->srcu_gp_seq), READ_ONCE(ssp->srcu_gp_seq_needed_exp)))
763 		cbdelay = 0;
764 
765 	WRITE_ONCE(ssp->srcu_last_gp_end, ktime_get_mono_fast_ns());
766 	rcu_seq_end(&ssp->srcu_gp_seq);
767 	gpseq = rcu_seq_current(&ssp->srcu_gp_seq);
768 	if (ULONG_CMP_LT(ssp->srcu_gp_seq_needed_exp, gpseq))
769 		WRITE_ONCE(ssp->srcu_gp_seq_needed_exp, gpseq);
770 	spin_unlock_irq_rcu_node(ssp);
771 	mutex_unlock(&ssp->srcu_gp_mutex);
772 	/* A new grace period can start at this point.  But only one. */
773 
774 	/* Initiate callback invocation as needed. */
775 	ss_state = smp_load_acquire(&ssp->srcu_size_state);
776 	if (ss_state < SRCU_SIZE_WAIT_BARRIER) {
777 		srcu_schedule_cbs_sdp(per_cpu_ptr(ssp->sda, 0), cbdelay);
778 	} else {
779 		idx = rcu_seq_ctr(gpseq) % ARRAY_SIZE(snp->srcu_have_cbs);
780 		srcu_for_each_node_breadth_first(ssp, snp) {
781 			spin_lock_irq_rcu_node(snp);
782 			cbs = false;
783 			last_lvl = snp >= ssp->level[rcu_num_lvls - 1];
784 			if (last_lvl)
785 				cbs = ss_state < SRCU_SIZE_BIG || snp->srcu_have_cbs[idx] == gpseq;
786 			snp->srcu_have_cbs[idx] = gpseq;
787 			rcu_seq_set_state(&snp->srcu_have_cbs[idx], 1);
788 			sgsne = snp->srcu_gp_seq_needed_exp;
789 			if (srcu_invl_snp_seq(sgsne) || ULONG_CMP_LT(sgsne, gpseq))
790 				WRITE_ONCE(snp->srcu_gp_seq_needed_exp, gpseq);
791 			if (ss_state < SRCU_SIZE_BIG)
792 				mask = ~0;
793 			else
794 				mask = snp->srcu_data_have_cbs[idx];
795 			snp->srcu_data_have_cbs[idx] = 0;
796 			spin_unlock_irq_rcu_node(snp);
797 			if (cbs)
798 				srcu_schedule_cbs_snp(ssp, snp, mask, cbdelay);
799 		}
800 	}
801 
802 	/* Occasionally prevent srcu_data counter wrap. */
803 	if (!(gpseq & counter_wrap_check))
804 		for_each_possible_cpu(cpu) {
805 			sdp = per_cpu_ptr(ssp->sda, cpu);
806 			spin_lock_irqsave_rcu_node(sdp, flags);
807 			if (ULONG_CMP_GE(gpseq, sdp->srcu_gp_seq_needed + 100))
808 				sdp->srcu_gp_seq_needed = gpseq;
809 			if (ULONG_CMP_GE(gpseq, sdp->srcu_gp_seq_needed_exp + 100))
810 				sdp->srcu_gp_seq_needed_exp = gpseq;
811 			spin_unlock_irqrestore_rcu_node(sdp, flags);
812 		}
813 
814 	/* Callback initiation done, allow grace periods after next. */
815 	mutex_unlock(&ssp->srcu_cb_mutex);
816 
817 	/* Start a new grace period if needed. */
818 	spin_lock_irq_rcu_node(ssp);
819 	gpseq = rcu_seq_current(&ssp->srcu_gp_seq);
820 	if (!rcu_seq_state(gpseq) &&
821 	    ULONG_CMP_LT(gpseq, ssp->srcu_gp_seq_needed)) {
822 		srcu_gp_start(ssp);
823 		spin_unlock_irq_rcu_node(ssp);
824 		srcu_reschedule(ssp, 0);
825 	} else {
826 		spin_unlock_irq_rcu_node(ssp);
827 	}
828 
829 	/* Transition to big if needed. */
830 	if (ss_state != SRCU_SIZE_SMALL && ss_state != SRCU_SIZE_BIG) {
831 		if (ss_state == SRCU_SIZE_ALLOC)
832 			init_srcu_struct_nodes(ssp, GFP_KERNEL);
833 		else
834 			smp_store_release(&ssp->srcu_size_state, ss_state + 1);
835 	}
836 }
837 
838 /*
839  * Funnel-locking scheme to scalably mediate many concurrent expedited
840  * grace-period requests.  This function is invoked for the first known
841  * expedited request for a grace period that has already been requested,
842  * but without expediting.  To start a completely new grace period,
843  * whether expedited or not, use srcu_funnel_gp_start() instead.
844  */
srcu_funnel_exp_start(struct srcu_struct * ssp,struct srcu_node * snp,unsigned long s)845 static void srcu_funnel_exp_start(struct srcu_struct *ssp, struct srcu_node *snp,
846 				  unsigned long s)
847 {
848 	unsigned long flags;
849 	unsigned long sgsne;
850 
851 	if (snp)
852 		for (; snp != NULL; snp = snp->srcu_parent) {
853 			sgsne = READ_ONCE(snp->srcu_gp_seq_needed_exp);
854 			if (rcu_seq_done(&ssp->srcu_gp_seq, s) ||
855 			    (!srcu_invl_snp_seq(sgsne) && ULONG_CMP_GE(sgsne, s)))
856 				return;
857 			spin_lock_irqsave_rcu_node(snp, flags);
858 			sgsne = snp->srcu_gp_seq_needed_exp;
859 			if (!srcu_invl_snp_seq(sgsne) && ULONG_CMP_GE(sgsne, s)) {
860 				spin_unlock_irqrestore_rcu_node(snp, flags);
861 				return;
862 			}
863 			WRITE_ONCE(snp->srcu_gp_seq_needed_exp, s);
864 			spin_unlock_irqrestore_rcu_node(snp, flags);
865 		}
866 	spin_lock_irqsave_ssp_contention(ssp, &flags);
867 	if (ULONG_CMP_LT(ssp->srcu_gp_seq_needed_exp, s))
868 		WRITE_ONCE(ssp->srcu_gp_seq_needed_exp, s);
869 	spin_unlock_irqrestore_rcu_node(ssp, flags);
870 }
871 
872 /*
873  * Funnel-locking scheme to scalably mediate many concurrent grace-period
874  * requests.  The winner has to do the work of actually starting grace
875  * period s.  Losers must either ensure that their desired grace-period
876  * number is recorded on at least their leaf srcu_node structure, or they
877  * must take steps to invoke their own callbacks.
878  *
879  * Note that this function also does the work of srcu_funnel_exp_start(),
880  * in some cases by directly invoking it.
881  */
srcu_funnel_gp_start(struct srcu_struct * ssp,struct srcu_data * sdp,unsigned long s,bool do_norm)882 static void srcu_funnel_gp_start(struct srcu_struct *ssp, struct srcu_data *sdp,
883 				 unsigned long s, bool do_norm)
884 {
885 	unsigned long flags;
886 	int idx = rcu_seq_ctr(s) % ARRAY_SIZE(sdp->mynode->srcu_have_cbs);
887 	unsigned long sgsne;
888 	struct srcu_node *snp;
889 	struct srcu_node *snp_leaf;
890 	unsigned long snp_seq;
891 
892 	/* Ensure that snp node tree is fully initialized before traversing it */
893 	if (smp_load_acquire(&ssp->srcu_size_state) < SRCU_SIZE_WAIT_BARRIER)
894 		snp_leaf = NULL;
895 	else
896 		snp_leaf = sdp->mynode;
897 
898 	if (snp_leaf)
899 		/* Each pass through the loop does one level of the srcu_node tree. */
900 		for (snp = snp_leaf; snp != NULL; snp = snp->srcu_parent) {
901 			if (rcu_seq_done(&ssp->srcu_gp_seq, s) && snp != snp_leaf)
902 				return; /* GP already done and CBs recorded. */
903 			spin_lock_irqsave_rcu_node(snp, flags);
904 			snp_seq = snp->srcu_have_cbs[idx];
905 			if (!srcu_invl_snp_seq(snp_seq) && ULONG_CMP_GE(snp_seq, s)) {
906 				if (snp == snp_leaf && snp_seq == s)
907 					snp->srcu_data_have_cbs[idx] |= sdp->grpmask;
908 				spin_unlock_irqrestore_rcu_node(snp, flags);
909 				if (snp == snp_leaf && snp_seq != s) {
910 					srcu_schedule_cbs_sdp(sdp, do_norm ? SRCU_INTERVAL : 0);
911 					return;
912 				}
913 				if (!do_norm)
914 					srcu_funnel_exp_start(ssp, snp, s);
915 				return;
916 			}
917 			snp->srcu_have_cbs[idx] = s;
918 			if (snp == snp_leaf)
919 				snp->srcu_data_have_cbs[idx] |= sdp->grpmask;
920 			sgsne = snp->srcu_gp_seq_needed_exp;
921 			if (!do_norm && (srcu_invl_snp_seq(sgsne) || ULONG_CMP_LT(sgsne, s)))
922 				WRITE_ONCE(snp->srcu_gp_seq_needed_exp, s);
923 			spin_unlock_irqrestore_rcu_node(snp, flags);
924 		}
925 
926 	/* Top of tree, must ensure the grace period will be started. */
927 	spin_lock_irqsave_ssp_contention(ssp, &flags);
928 	if (ULONG_CMP_LT(ssp->srcu_gp_seq_needed, s)) {
929 		/*
930 		 * Record need for grace period s.  Pair with load
931 		 * acquire setting up for initialization.
932 		 */
933 		smp_store_release(&ssp->srcu_gp_seq_needed, s); /*^^^*/
934 	}
935 	if (!do_norm && ULONG_CMP_LT(ssp->srcu_gp_seq_needed_exp, s))
936 		WRITE_ONCE(ssp->srcu_gp_seq_needed_exp, s);
937 
938 	/* If grace period not already done and none in progress, start it. */
939 	if (!rcu_seq_done(&ssp->srcu_gp_seq, s) &&
940 	    rcu_seq_state(ssp->srcu_gp_seq) == SRCU_STATE_IDLE) {
941 		WARN_ON_ONCE(ULONG_CMP_GE(ssp->srcu_gp_seq, ssp->srcu_gp_seq_needed));
942 		srcu_gp_start(ssp);
943 
944 		// And how can that list_add() in the "else" clause
945 		// possibly be safe for concurrent execution?  Well,
946 		// it isn't.  And it does not have to be.  After all, it
947 		// can only be executed during early boot when there is only
948 		// the one boot CPU running with interrupts still disabled.
949 		if (likely(srcu_init_done))
950 			queue_delayed_work(rcu_gp_wq, &ssp->work,
951 					   !!srcu_get_delay(ssp));
952 		else if (list_empty(&ssp->work.work.entry))
953 			list_add(&ssp->work.work.entry, &srcu_boot_list);
954 	}
955 	spin_unlock_irqrestore_rcu_node(ssp, flags);
956 }
957 
958 /*
959  * Wait until all readers counted by array index idx complete, but
960  * loop an additional time if there is an expedited grace period pending.
961  * The caller must ensure that ->srcu_idx is not changed while checking.
962  */
try_check_zero(struct srcu_struct * ssp,int idx,int trycount)963 static bool try_check_zero(struct srcu_struct *ssp, int idx, int trycount)
964 {
965 	unsigned long curdelay;
966 
967 	curdelay = !srcu_get_delay(ssp);
968 
969 	for (;;) {
970 		if (srcu_readers_active_idx_check(ssp, idx))
971 			return true;
972 		if ((--trycount + curdelay) <= 0)
973 			return false;
974 		udelay(srcu_retry_check_delay);
975 	}
976 }
977 
978 /*
979  * Increment the ->srcu_idx counter so that future SRCU readers will
980  * use the other rank of the ->srcu_(un)lock_count[] arrays.  This allows
981  * us to wait for pre-existing readers in a starvation-free manner.
982  */
srcu_flip(struct srcu_struct * ssp)983 static void srcu_flip(struct srcu_struct *ssp)
984 {
985 	/*
986 	 * Ensure that if this updater saw a given reader's increment
987 	 * from __srcu_read_lock(), that reader was using an old value
988 	 * of ->srcu_idx.  Also ensure that if a given reader sees the
989 	 * new value of ->srcu_idx, this updater's earlier scans cannot
990 	 * have seen that reader's increments (which is OK, because this
991 	 * grace period need not wait on that reader).
992 	 */
993 	smp_mb(); /* E */  /* Pairs with B and C. */
994 
995 	WRITE_ONCE(ssp->srcu_idx, ssp->srcu_idx + 1);
996 
997 	/*
998 	 * Ensure that if the updater misses an __srcu_read_unlock()
999 	 * increment, that task's next __srcu_read_lock() will see the
1000 	 * above counter update.  Note that both this memory barrier
1001 	 * and the one in srcu_readers_active_idx_check() provide the
1002 	 * guarantee for __srcu_read_lock().
1003 	 */
1004 	smp_mb(); /* D */  /* Pairs with C. */
1005 }
1006 
1007 /*
1008  * If SRCU is likely idle, return true, otherwise return false.
1009  *
1010  * Note that it is OK for several current from-idle requests for a new
1011  * grace period from idle to specify expediting because they will all end
1012  * up requesting the same grace period anyhow.  So no loss.
1013  *
1014  * Note also that if any CPU (including the current one) is still invoking
1015  * callbacks, this function will nevertheless say "idle".  This is not
1016  * ideal, but the overhead of checking all CPUs' callback lists is even
1017  * less ideal, especially on large systems.  Furthermore, the wakeup
1018  * can happen before the callback is fully removed, so we have no choice
1019  * but to accept this type of error.
1020  *
1021  * This function is also subject to counter-wrap errors, but let's face
1022  * it, if this function was preempted for enough time for the counters
1023  * to wrap, it really doesn't matter whether or not we expedite the grace
1024  * period.  The extra overhead of a needlessly expedited grace period is
1025  * negligible when amortized over that time period, and the extra latency
1026  * of a needlessly non-expedited grace period is similarly negligible.
1027  */
srcu_might_be_idle(struct srcu_struct * ssp)1028 static bool srcu_might_be_idle(struct srcu_struct *ssp)
1029 {
1030 	unsigned long curseq;
1031 	unsigned long flags;
1032 	struct srcu_data *sdp;
1033 	unsigned long t;
1034 	unsigned long tlast;
1035 
1036 	check_init_srcu_struct(ssp);
1037 	/* If the local srcu_data structure has callbacks, not idle.  */
1038 	sdp = raw_cpu_ptr(ssp->sda);
1039 	spin_lock_irqsave_rcu_node(sdp, flags);
1040 	if (rcu_segcblist_pend_cbs(&sdp->srcu_cblist)) {
1041 		spin_unlock_irqrestore_rcu_node(sdp, flags);
1042 		return false; /* Callbacks already present, so not idle. */
1043 	}
1044 	spin_unlock_irqrestore_rcu_node(sdp, flags);
1045 
1046 	/*
1047 	 * No local callbacks, so probabilistically probe global state.
1048 	 * Exact information would require acquiring locks, which would
1049 	 * kill scalability, hence the probabilistic nature of the probe.
1050 	 */
1051 
1052 	/* First, see if enough time has passed since the last GP. */
1053 	t = ktime_get_mono_fast_ns();
1054 	tlast = READ_ONCE(ssp->srcu_last_gp_end);
1055 	if (exp_holdoff == 0 ||
1056 	    time_in_range_open(t, tlast, tlast + exp_holdoff))
1057 		return false; /* Too soon after last GP. */
1058 
1059 	/* Next, check for probable idleness. */
1060 	curseq = rcu_seq_current(&ssp->srcu_gp_seq);
1061 	smp_mb(); /* Order ->srcu_gp_seq with ->srcu_gp_seq_needed. */
1062 	if (ULONG_CMP_LT(curseq, READ_ONCE(ssp->srcu_gp_seq_needed)))
1063 		return false; /* Grace period in progress, so not idle. */
1064 	smp_mb(); /* Order ->srcu_gp_seq with prior access. */
1065 	if (curseq != rcu_seq_current(&ssp->srcu_gp_seq))
1066 		return false; /* GP # changed, so not idle. */
1067 	return true; /* With reasonable probability, idle! */
1068 }
1069 
1070 /*
1071  * SRCU callback function to leak a callback.
1072  */
srcu_leak_callback(struct rcu_head * rhp)1073 static void srcu_leak_callback(struct rcu_head *rhp)
1074 {
1075 }
1076 
1077 /*
1078  * Start an SRCU grace period, and also queue the callback if non-NULL.
1079  */
srcu_gp_start_if_needed(struct srcu_struct * ssp,struct rcu_head * rhp,bool do_norm)1080 static unsigned long srcu_gp_start_if_needed(struct srcu_struct *ssp,
1081 					     struct rcu_head *rhp, bool do_norm)
1082 {
1083 	unsigned long flags;
1084 	int idx;
1085 	bool needexp = false;
1086 	bool needgp = false;
1087 	unsigned long s;
1088 	struct srcu_data *sdp;
1089 	struct srcu_node *sdp_mynode;
1090 	int ss_state;
1091 
1092 	check_init_srcu_struct(ssp);
1093 	idx = srcu_read_lock(ssp);
1094 	ss_state = smp_load_acquire(&ssp->srcu_size_state);
1095 	if (ss_state < SRCU_SIZE_WAIT_CALL)
1096 		sdp = per_cpu_ptr(ssp->sda, 0);
1097 	else
1098 		sdp = raw_cpu_ptr(ssp->sda);
1099 	spin_lock_irqsave_sdp_contention(sdp, &flags);
1100 	if (rhp)
1101 		rcu_segcblist_enqueue(&sdp->srcu_cblist, rhp);
1102 	rcu_segcblist_advance(&sdp->srcu_cblist,
1103 			      rcu_seq_current(&ssp->srcu_gp_seq));
1104 	s = rcu_seq_snap(&ssp->srcu_gp_seq);
1105 	(void)rcu_segcblist_accelerate(&sdp->srcu_cblist, s);
1106 	if (ULONG_CMP_LT(sdp->srcu_gp_seq_needed, s)) {
1107 		sdp->srcu_gp_seq_needed = s;
1108 		needgp = true;
1109 	}
1110 	if (!do_norm && ULONG_CMP_LT(sdp->srcu_gp_seq_needed_exp, s)) {
1111 		sdp->srcu_gp_seq_needed_exp = s;
1112 		needexp = true;
1113 	}
1114 	spin_unlock_irqrestore_rcu_node(sdp, flags);
1115 
1116 	/* Ensure that snp node tree is fully initialized before traversing it */
1117 	if (ss_state < SRCU_SIZE_WAIT_BARRIER)
1118 		sdp_mynode = NULL;
1119 	else
1120 		sdp_mynode = sdp->mynode;
1121 
1122 	if (needgp)
1123 		srcu_funnel_gp_start(ssp, sdp, s, do_norm);
1124 	else if (needexp)
1125 		srcu_funnel_exp_start(ssp, sdp_mynode, s);
1126 	srcu_read_unlock(ssp, idx);
1127 	return s;
1128 }
1129 
1130 /*
1131  * Enqueue an SRCU callback on the srcu_data structure associated with
1132  * the current CPU and the specified srcu_struct structure, initiating
1133  * grace-period processing if it is not already running.
1134  *
1135  * Note that all CPUs must agree that the grace period extended beyond
1136  * all pre-existing SRCU read-side critical section.  On systems with
1137  * more than one CPU, this means that when "func()" is invoked, each CPU
1138  * is guaranteed to have executed a full memory barrier since the end of
1139  * its last corresponding SRCU read-side critical section whose beginning
1140  * preceded the call to call_srcu().  It also means that each CPU executing
1141  * an SRCU read-side critical section that continues beyond the start of
1142  * "func()" must have executed a memory barrier after the call_srcu()
1143  * but before the beginning of that SRCU read-side critical section.
1144  * Note that these guarantees include CPUs that are offline, idle, or
1145  * executing in user mode, as well as CPUs that are executing in the kernel.
1146  *
1147  * Furthermore, if CPU A invoked call_srcu() and CPU B invoked the
1148  * resulting SRCU callback function "func()", then both CPU A and CPU
1149  * B are guaranteed to execute a full memory barrier during the time
1150  * interval between the call to call_srcu() and the invocation of "func()".
1151  * This guarantee applies even if CPU A and CPU B are the same CPU (but
1152  * again only if the system has more than one CPU).
1153  *
1154  * Of course, these guarantees apply only for invocations of call_srcu(),
1155  * srcu_read_lock(), and srcu_read_unlock() that are all passed the same
1156  * srcu_struct structure.
1157  */
__call_srcu(struct srcu_struct * ssp,struct rcu_head * rhp,rcu_callback_t func,bool do_norm)1158 static void __call_srcu(struct srcu_struct *ssp, struct rcu_head *rhp,
1159 			rcu_callback_t func, bool do_norm)
1160 {
1161 	if (debug_rcu_head_queue(rhp)) {
1162 		/* Probable double call_srcu(), so leak the callback. */
1163 		WRITE_ONCE(rhp->func, srcu_leak_callback);
1164 		WARN_ONCE(1, "call_srcu(): Leaked duplicate callback\n");
1165 		return;
1166 	}
1167 	rhp->func = func;
1168 	(void)srcu_gp_start_if_needed(ssp, rhp, do_norm);
1169 }
1170 
1171 /**
1172  * call_srcu() - Queue a callback for invocation after an SRCU grace period
1173  * @ssp: srcu_struct in queue the callback
1174  * @rhp: structure to be used for queueing the SRCU callback.
1175  * @func: function to be invoked after the SRCU grace period
1176  *
1177  * The callback function will be invoked some time after a full SRCU
1178  * grace period elapses, in other words after all pre-existing SRCU
1179  * read-side critical sections have completed.  However, the callback
1180  * function might well execute concurrently with other SRCU read-side
1181  * critical sections that started after call_srcu() was invoked.  SRCU
1182  * read-side critical sections are delimited by srcu_read_lock() and
1183  * srcu_read_unlock(), and may be nested.
1184  *
1185  * The callback will be invoked from process context, but must nevertheless
1186  * be fast and must not block.
1187  */
call_srcu(struct srcu_struct * ssp,struct rcu_head * rhp,rcu_callback_t func)1188 void call_srcu(struct srcu_struct *ssp, struct rcu_head *rhp,
1189 	       rcu_callback_t func)
1190 {
1191 	__call_srcu(ssp, rhp, func, true);
1192 }
1193 EXPORT_SYMBOL_GPL(call_srcu);
1194 
1195 /*
1196  * Helper function for synchronize_srcu() and synchronize_srcu_expedited().
1197  */
__synchronize_srcu(struct srcu_struct * ssp,bool do_norm)1198 static void __synchronize_srcu(struct srcu_struct *ssp, bool do_norm)
1199 {
1200 	struct rcu_synchronize rcu;
1201 
1202 	RCU_LOCKDEP_WARN(lockdep_is_held(ssp) ||
1203 			 lock_is_held(&rcu_bh_lock_map) ||
1204 			 lock_is_held(&rcu_lock_map) ||
1205 			 lock_is_held(&rcu_sched_lock_map),
1206 			 "Illegal synchronize_srcu() in same-type SRCU (or in RCU) read-side critical section");
1207 
1208 	if (rcu_scheduler_active == RCU_SCHEDULER_INACTIVE)
1209 		return;
1210 	might_sleep();
1211 	check_init_srcu_struct(ssp);
1212 	init_completion(&rcu.completion);
1213 	init_rcu_head_on_stack(&rcu.head);
1214 	__call_srcu(ssp, &rcu.head, wakeme_after_rcu, do_norm);
1215 	wait_for_completion(&rcu.completion);
1216 	destroy_rcu_head_on_stack(&rcu.head);
1217 
1218 	/*
1219 	 * Make sure that later code is ordered after the SRCU grace
1220 	 * period.  This pairs with the spin_lock_irq_rcu_node()
1221 	 * in srcu_invoke_callbacks().  Unlike Tree RCU, this is needed
1222 	 * because the current CPU might have been totally uninvolved with
1223 	 * (and thus unordered against) that grace period.
1224 	 */
1225 	smp_mb();
1226 }
1227 
1228 /**
1229  * synchronize_srcu_expedited - Brute-force SRCU grace period
1230  * @ssp: srcu_struct with which to synchronize.
1231  *
1232  * Wait for an SRCU grace period to elapse, but be more aggressive about
1233  * spinning rather than blocking when waiting.
1234  *
1235  * Note that synchronize_srcu_expedited() has the same deadlock and
1236  * memory-ordering properties as does synchronize_srcu().
1237  */
synchronize_srcu_expedited(struct srcu_struct * ssp)1238 void synchronize_srcu_expedited(struct srcu_struct *ssp)
1239 {
1240 	__synchronize_srcu(ssp, rcu_gp_is_normal());
1241 }
1242 EXPORT_SYMBOL_GPL(synchronize_srcu_expedited);
1243 
1244 /**
1245  * synchronize_srcu - wait for prior SRCU read-side critical-section completion
1246  * @ssp: srcu_struct with which to synchronize.
1247  *
1248  * Wait for the count to drain to zero of both indexes. To avoid the
1249  * possible starvation of synchronize_srcu(), it waits for the count of
1250  * the index=((->srcu_idx & 1) ^ 1) to drain to zero at first,
1251  * and then flip the srcu_idx and wait for the count of the other index.
1252  *
1253  * Can block; must be called from process context.
1254  *
1255  * Note that it is illegal to call synchronize_srcu() from the corresponding
1256  * SRCU read-side critical section; doing so will result in deadlock.
1257  * However, it is perfectly legal to call synchronize_srcu() on one
1258  * srcu_struct from some other srcu_struct's read-side critical section,
1259  * as long as the resulting graph of srcu_structs is acyclic.
1260  *
1261  * There are memory-ordering constraints implied by synchronize_srcu().
1262  * On systems with more than one CPU, when synchronize_srcu() returns,
1263  * each CPU is guaranteed to have executed a full memory barrier since
1264  * the end of its last corresponding SRCU read-side critical section
1265  * whose beginning preceded the call to synchronize_srcu().  In addition,
1266  * each CPU having an SRCU read-side critical section that extends beyond
1267  * the return from synchronize_srcu() is guaranteed to have executed a
1268  * full memory barrier after the beginning of synchronize_srcu() and before
1269  * the beginning of that SRCU read-side critical section.  Note that these
1270  * guarantees include CPUs that are offline, idle, or executing in user mode,
1271  * as well as CPUs that are executing in the kernel.
1272  *
1273  * Furthermore, if CPU A invoked synchronize_srcu(), which returned
1274  * to its caller on CPU B, then both CPU A and CPU B are guaranteed
1275  * to have executed a full memory barrier during the execution of
1276  * synchronize_srcu().  This guarantee applies even if CPU A and CPU B
1277  * are the same CPU, but again only if the system has more than one CPU.
1278  *
1279  * Of course, these memory-ordering guarantees apply only when
1280  * synchronize_srcu(), srcu_read_lock(), and srcu_read_unlock() are
1281  * passed the same srcu_struct structure.
1282  *
1283  * Implementation of these memory-ordering guarantees is similar to
1284  * that of synchronize_rcu().
1285  *
1286  * If SRCU is likely idle, expedite the first request.  This semantic
1287  * was provided by Classic SRCU, and is relied upon by its users, so TREE
1288  * SRCU must also provide it.  Note that detecting idleness is heuristic
1289  * and subject to both false positives and negatives.
1290  */
synchronize_srcu(struct srcu_struct * ssp)1291 void synchronize_srcu(struct srcu_struct *ssp)
1292 {
1293 	if (srcu_might_be_idle(ssp) || rcu_gp_is_expedited())
1294 		synchronize_srcu_expedited(ssp);
1295 	else
1296 		__synchronize_srcu(ssp, true);
1297 }
1298 EXPORT_SYMBOL_GPL(synchronize_srcu);
1299 
1300 /**
1301  * get_state_synchronize_srcu - Provide an end-of-grace-period cookie
1302  * @ssp: srcu_struct to provide cookie for.
1303  *
1304  * This function returns a cookie that can be passed to
1305  * poll_state_synchronize_srcu(), which will return true if a full grace
1306  * period has elapsed in the meantime.  It is the caller's responsibility
1307  * to make sure that grace period happens, for example, by invoking
1308  * call_srcu() after return from get_state_synchronize_srcu().
1309  */
get_state_synchronize_srcu(struct srcu_struct * ssp)1310 unsigned long get_state_synchronize_srcu(struct srcu_struct *ssp)
1311 {
1312 	// Any prior manipulation of SRCU-protected data must happen
1313 	// before the load from ->srcu_gp_seq.
1314 	smp_mb();
1315 	return rcu_seq_snap(&ssp->srcu_gp_seq);
1316 }
1317 EXPORT_SYMBOL_GPL(get_state_synchronize_srcu);
1318 
1319 /**
1320  * start_poll_synchronize_srcu - Provide cookie and start grace period
1321  * @ssp: srcu_struct to provide cookie for.
1322  *
1323  * This function returns a cookie that can be passed to
1324  * poll_state_synchronize_srcu(), which will return true if a full grace
1325  * period has elapsed in the meantime.  Unlike get_state_synchronize_srcu(),
1326  * this function also ensures that any needed SRCU grace period will be
1327  * started.  This convenience does come at a cost in terms of CPU overhead.
1328  */
start_poll_synchronize_srcu(struct srcu_struct * ssp)1329 unsigned long start_poll_synchronize_srcu(struct srcu_struct *ssp)
1330 {
1331 	return srcu_gp_start_if_needed(ssp, NULL, true);
1332 }
1333 EXPORT_SYMBOL_GPL(start_poll_synchronize_srcu);
1334 
1335 /**
1336  * poll_state_synchronize_srcu - Has cookie's grace period ended?
1337  * @ssp: srcu_struct to provide cookie for.
1338  * @cookie: Return value from get_state_synchronize_srcu() or start_poll_synchronize_srcu().
1339  *
1340  * This function takes the cookie that was returned from either
1341  * get_state_synchronize_srcu() or start_poll_synchronize_srcu(), and
1342  * returns @true if an SRCU grace period elapsed since the time that the
1343  * cookie was created.
1344  *
1345  * Because cookies are finite in size, wrapping/overflow is possible.
1346  * This is more pronounced on 32-bit systems where cookies are 32 bits,
1347  * where in theory wrapping could happen in about 14 hours assuming
1348  * 25-microsecond expedited SRCU grace periods.  However, a more likely
1349  * overflow lower bound is on the order of 24 days in the case of
1350  * one-millisecond SRCU grace periods.  Of course, wrapping in a 64-bit
1351  * system requires geologic timespans, as in more than seven million years
1352  * even for expedited SRCU grace periods.
1353  *
1354  * Wrapping/overflow is much more of an issue for CONFIG_SMP=n systems
1355  * that also have CONFIG_PREEMPTION=n, which selects Tiny SRCU.  This uses
1356  * a 16-bit cookie, which rcutorture routinely wraps in a matter of a
1357  * few minutes.  If this proves to be a problem, this counter will be
1358  * expanded to the same size as for Tree SRCU.
1359  */
poll_state_synchronize_srcu(struct srcu_struct * ssp,unsigned long cookie)1360 bool poll_state_synchronize_srcu(struct srcu_struct *ssp, unsigned long cookie)
1361 {
1362 	if (!rcu_seq_done(&ssp->srcu_gp_seq, cookie))
1363 		return false;
1364 	// Ensure that the end of the SRCU grace period happens before
1365 	// any subsequent code that the caller might execute.
1366 	smp_mb(); // ^^^
1367 	return true;
1368 }
1369 EXPORT_SYMBOL_GPL(poll_state_synchronize_srcu);
1370 
1371 /*
1372  * Callback function for srcu_barrier() use.
1373  */
srcu_barrier_cb(struct rcu_head * rhp)1374 static void srcu_barrier_cb(struct rcu_head *rhp)
1375 {
1376 	struct srcu_data *sdp;
1377 	struct srcu_struct *ssp;
1378 
1379 	sdp = container_of(rhp, struct srcu_data, srcu_barrier_head);
1380 	ssp = sdp->ssp;
1381 	if (atomic_dec_and_test(&ssp->srcu_barrier_cpu_cnt))
1382 		complete(&ssp->srcu_barrier_completion);
1383 }
1384 
1385 /*
1386  * Enqueue an srcu_barrier() callback on the specified srcu_data
1387  * structure's ->cblist.  but only if that ->cblist already has at least one
1388  * callback enqueued.  Note that if a CPU already has callbacks enqueue,
1389  * it must have already registered the need for a future grace period,
1390  * so all we need do is enqueue a callback that will use the same grace
1391  * period as the last callback already in the queue.
1392  */
srcu_barrier_one_cpu(struct srcu_struct * ssp,struct srcu_data * sdp)1393 static void srcu_barrier_one_cpu(struct srcu_struct *ssp, struct srcu_data *sdp)
1394 {
1395 	spin_lock_irq_rcu_node(sdp);
1396 	atomic_inc(&ssp->srcu_barrier_cpu_cnt);
1397 	sdp->srcu_barrier_head.func = srcu_barrier_cb;
1398 	debug_rcu_head_queue(&sdp->srcu_barrier_head);
1399 	if (!rcu_segcblist_entrain(&sdp->srcu_cblist,
1400 				   &sdp->srcu_barrier_head)) {
1401 		debug_rcu_head_unqueue(&sdp->srcu_barrier_head);
1402 		atomic_dec(&ssp->srcu_barrier_cpu_cnt);
1403 	}
1404 	spin_unlock_irq_rcu_node(sdp);
1405 }
1406 
1407 /**
1408  * srcu_barrier - Wait until all in-flight call_srcu() callbacks complete.
1409  * @ssp: srcu_struct on which to wait for in-flight callbacks.
1410  */
srcu_barrier(struct srcu_struct * ssp)1411 void srcu_barrier(struct srcu_struct *ssp)
1412 {
1413 	int cpu;
1414 	int idx;
1415 	unsigned long s = rcu_seq_snap(&ssp->srcu_barrier_seq);
1416 
1417 	check_init_srcu_struct(ssp);
1418 	mutex_lock(&ssp->srcu_barrier_mutex);
1419 	if (rcu_seq_done(&ssp->srcu_barrier_seq, s)) {
1420 		smp_mb(); /* Force ordering following return. */
1421 		mutex_unlock(&ssp->srcu_barrier_mutex);
1422 		return; /* Someone else did our work for us. */
1423 	}
1424 	rcu_seq_start(&ssp->srcu_barrier_seq);
1425 	init_completion(&ssp->srcu_barrier_completion);
1426 
1427 	/* Initial count prevents reaching zero until all CBs are posted. */
1428 	atomic_set(&ssp->srcu_barrier_cpu_cnt, 1);
1429 
1430 	idx = srcu_read_lock(ssp);
1431 	if (smp_load_acquire(&ssp->srcu_size_state) < SRCU_SIZE_WAIT_BARRIER)
1432 		srcu_barrier_one_cpu(ssp, per_cpu_ptr(ssp->sda, 0));
1433 	else
1434 		for_each_possible_cpu(cpu)
1435 			srcu_barrier_one_cpu(ssp, per_cpu_ptr(ssp->sda, cpu));
1436 	srcu_read_unlock(ssp, idx);
1437 
1438 	/* Remove the initial count, at which point reaching zero can happen. */
1439 	if (atomic_dec_and_test(&ssp->srcu_barrier_cpu_cnt))
1440 		complete(&ssp->srcu_barrier_completion);
1441 	wait_for_completion(&ssp->srcu_barrier_completion);
1442 
1443 	rcu_seq_end(&ssp->srcu_barrier_seq);
1444 	mutex_unlock(&ssp->srcu_barrier_mutex);
1445 }
1446 EXPORT_SYMBOL_GPL(srcu_barrier);
1447 
1448 /**
1449  * srcu_batches_completed - return batches completed.
1450  * @ssp: srcu_struct on which to report batch completion.
1451  *
1452  * Report the number of batches, correlated with, but not necessarily
1453  * precisely the same as, the number of grace periods that have elapsed.
1454  */
srcu_batches_completed(struct srcu_struct * ssp)1455 unsigned long srcu_batches_completed(struct srcu_struct *ssp)
1456 {
1457 	return READ_ONCE(ssp->srcu_idx);
1458 }
1459 EXPORT_SYMBOL_GPL(srcu_batches_completed);
1460 
1461 /*
1462  * Core SRCU state machine.  Push state bits of ->srcu_gp_seq
1463  * to SRCU_STATE_SCAN2, and invoke srcu_gp_end() when scan has
1464  * completed in that state.
1465  */
srcu_advance_state(struct srcu_struct * ssp)1466 static void srcu_advance_state(struct srcu_struct *ssp)
1467 {
1468 	int idx;
1469 
1470 	mutex_lock(&ssp->srcu_gp_mutex);
1471 
1472 	/*
1473 	 * Because readers might be delayed for an extended period after
1474 	 * fetching ->srcu_idx for their index, at any point in time there
1475 	 * might well be readers using both idx=0 and idx=1.  We therefore
1476 	 * need to wait for readers to clear from both index values before
1477 	 * invoking a callback.
1478 	 *
1479 	 * The load-acquire ensures that we see the accesses performed
1480 	 * by the prior grace period.
1481 	 */
1482 	idx = rcu_seq_state(smp_load_acquire(&ssp->srcu_gp_seq)); /* ^^^ */
1483 	if (idx == SRCU_STATE_IDLE) {
1484 		spin_lock_irq_rcu_node(ssp);
1485 		if (ULONG_CMP_GE(ssp->srcu_gp_seq, ssp->srcu_gp_seq_needed)) {
1486 			WARN_ON_ONCE(rcu_seq_state(ssp->srcu_gp_seq));
1487 			spin_unlock_irq_rcu_node(ssp);
1488 			mutex_unlock(&ssp->srcu_gp_mutex);
1489 			return;
1490 		}
1491 		idx = rcu_seq_state(READ_ONCE(ssp->srcu_gp_seq));
1492 		if (idx == SRCU_STATE_IDLE)
1493 			srcu_gp_start(ssp);
1494 		spin_unlock_irq_rcu_node(ssp);
1495 		if (idx != SRCU_STATE_IDLE) {
1496 			mutex_unlock(&ssp->srcu_gp_mutex);
1497 			return; /* Someone else started the grace period. */
1498 		}
1499 	}
1500 
1501 	if (rcu_seq_state(READ_ONCE(ssp->srcu_gp_seq)) == SRCU_STATE_SCAN1) {
1502 		idx = 1 ^ (ssp->srcu_idx & 1);
1503 		if (!try_check_zero(ssp, idx, 1)) {
1504 			mutex_unlock(&ssp->srcu_gp_mutex);
1505 			return; /* readers present, retry later. */
1506 		}
1507 		srcu_flip(ssp);
1508 		spin_lock_irq_rcu_node(ssp);
1509 		rcu_seq_set_state(&ssp->srcu_gp_seq, SRCU_STATE_SCAN2);
1510 		ssp->srcu_n_exp_nodelay = 0;
1511 		spin_unlock_irq_rcu_node(ssp);
1512 	}
1513 
1514 	if (rcu_seq_state(READ_ONCE(ssp->srcu_gp_seq)) == SRCU_STATE_SCAN2) {
1515 
1516 		/*
1517 		 * SRCU read-side critical sections are normally short,
1518 		 * so check at least twice in quick succession after a flip.
1519 		 */
1520 		idx = 1 ^ (ssp->srcu_idx & 1);
1521 		if (!try_check_zero(ssp, idx, 2)) {
1522 			mutex_unlock(&ssp->srcu_gp_mutex);
1523 			return; /* readers present, retry later. */
1524 		}
1525 		ssp->srcu_n_exp_nodelay = 0;
1526 		srcu_gp_end(ssp);  /* Releases ->srcu_gp_mutex. */
1527 	}
1528 }
1529 
1530 /*
1531  * Invoke a limited number of SRCU callbacks that have passed through
1532  * their grace period.  If there are more to do, SRCU will reschedule
1533  * the workqueue.  Note that needed memory barriers have been executed
1534  * in this task's context by srcu_readers_active_idx_check().
1535  */
srcu_invoke_callbacks(struct work_struct * work)1536 static void srcu_invoke_callbacks(struct work_struct *work)
1537 {
1538 	long len;
1539 	bool more;
1540 	struct rcu_cblist ready_cbs;
1541 	struct rcu_head *rhp;
1542 	struct srcu_data *sdp;
1543 	struct srcu_struct *ssp;
1544 
1545 	sdp = container_of(work, struct srcu_data, work);
1546 
1547 	ssp = sdp->ssp;
1548 	rcu_cblist_init(&ready_cbs);
1549 	spin_lock_irq_rcu_node(sdp);
1550 	rcu_segcblist_advance(&sdp->srcu_cblist,
1551 			      rcu_seq_current(&ssp->srcu_gp_seq));
1552 	if (sdp->srcu_cblist_invoking ||
1553 	    !rcu_segcblist_ready_cbs(&sdp->srcu_cblist)) {
1554 		spin_unlock_irq_rcu_node(sdp);
1555 		return;  /* Someone else on the job or nothing to do. */
1556 	}
1557 
1558 	/* We are on the job!  Extract and invoke ready callbacks. */
1559 	sdp->srcu_cblist_invoking = true;
1560 	rcu_segcblist_extract_done_cbs(&sdp->srcu_cblist, &ready_cbs);
1561 	len = ready_cbs.len;
1562 	spin_unlock_irq_rcu_node(sdp);
1563 	rhp = rcu_cblist_dequeue(&ready_cbs);
1564 	for (; rhp != NULL; rhp = rcu_cblist_dequeue(&ready_cbs)) {
1565 		debug_rcu_head_unqueue(rhp);
1566 		local_bh_disable();
1567 		rhp->func(rhp);
1568 		local_bh_enable();
1569 	}
1570 	WARN_ON_ONCE(ready_cbs.len);
1571 
1572 	/*
1573 	 * Update counts, accelerate new callbacks, and if needed,
1574 	 * schedule another round of callback invocation.
1575 	 */
1576 	spin_lock_irq_rcu_node(sdp);
1577 	rcu_segcblist_add_len(&sdp->srcu_cblist, -len);
1578 	(void)rcu_segcblist_accelerate(&sdp->srcu_cblist,
1579 				       rcu_seq_snap(&ssp->srcu_gp_seq));
1580 	sdp->srcu_cblist_invoking = false;
1581 	more = rcu_segcblist_ready_cbs(&sdp->srcu_cblist);
1582 	spin_unlock_irq_rcu_node(sdp);
1583 	if (more)
1584 		srcu_schedule_cbs_sdp(sdp, 0);
1585 }
1586 
1587 /*
1588  * Finished one round of SRCU grace period.  Start another if there are
1589  * more SRCU callbacks queued, otherwise put SRCU into not-running state.
1590  */
srcu_reschedule(struct srcu_struct * ssp,unsigned long delay)1591 static void srcu_reschedule(struct srcu_struct *ssp, unsigned long delay)
1592 {
1593 	bool pushgp = true;
1594 
1595 	spin_lock_irq_rcu_node(ssp);
1596 	if (ULONG_CMP_GE(ssp->srcu_gp_seq, ssp->srcu_gp_seq_needed)) {
1597 		if (!WARN_ON_ONCE(rcu_seq_state(ssp->srcu_gp_seq))) {
1598 			/* All requests fulfilled, time to go idle. */
1599 			pushgp = false;
1600 		}
1601 	} else if (!rcu_seq_state(ssp->srcu_gp_seq)) {
1602 		/* Outstanding request and no GP.  Start one. */
1603 		srcu_gp_start(ssp);
1604 	}
1605 	spin_unlock_irq_rcu_node(ssp);
1606 
1607 	if (pushgp)
1608 		queue_delayed_work(rcu_gp_wq, &ssp->work, delay);
1609 }
1610 
1611 /*
1612  * This is the work-queue function that handles SRCU grace periods.
1613  */
process_srcu(struct work_struct * work)1614 static void process_srcu(struct work_struct *work)
1615 {
1616 	unsigned long curdelay;
1617 	unsigned long j;
1618 	struct srcu_struct *ssp;
1619 
1620 	ssp = container_of(work, struct srcu_struct, work.work);
1621 
1622 	srcu_advance_state(ssp);
1623 	curdelay = srcu_get_delay(ssp);
1624 	if (curdelay) {
1625 		WRITE_ONCE(ssp->reschedule_count, 0);
1626 	} else {
1627 		j = jiffies;
1628 		if (READ_ONCE(ssp->reschedule_jiffies) == j) {
1629 			WRITE_ONCE(ssp->reschedule_count, READ_ONCE(ssp->reschedule_count) + 1);
1630 			if (READ_ONCE(ssp->reschedule_count) > srcu_max_nodelay)
1631 				curdelay = 1;
1632 		} else {
1633 			WRITE_ONCE(ssp->reschedule_count, 1);
1634 			WRITE_ONCE(ssp->reschedule_jiffies, j);
1635 		}
1636 	}
1637 	srcu_reschedule(ssp, curdelay);
1638 }
1639 
srcutorture_get_gp_data(enum rcutorture_type test_type,struct srcu_struct * ssp,int * flags,unsigned long * gp_seq)1640 void srcutorture_get_gp_data(enum rcutorture_type test_type,
1641 			     struct srcu_struct *ssp, int *flags,
1642 			     unsigned long *gp_seq)
1643 {
1644 	if (test_type != SRCU_FLAVOR)
1645 		return;
1646 	*flags = 0;
1647 	*gp_seq = rcu_seq_current(&ssp->srcu_gp_seq);
1648 }
1649 EXPORT_SYMBOL_GPL(srcutorture_get_gp_data);
1650 
1651 static const char * const srcu_size_state_name[] = {
1652 	"SRCU_SIZE_SMALL",
1653 	"SRCU_SIZE_ALLOC",
1654 	"SRCU_SIZE_WAIT_BARRIER",
1655 	"SRCU_SIZE_WAIT_CALL",
1656 	"SRCU_SIZE_WAIT_CBS1",
1657 	"SRCU_SIZE_WAIT_CBS2",
1658 	"SRCU_SIZE_WAIT_CBS3",
1659 	"SRCU_SIZE_WAIT_CBS4",
1660 	"SRCU_SIZE_BIG",
1661 	"SRCU_SIZE_???",
1662 };
1663 
srcu_torture_stats_print(struct srcu_struct * ssp,char * tt,char * tf)1664 void srcu_torture_stats_print(struct srcu_struct *ssp, char *tt, char *tf)
1665 {
1666 	int cpu;
1667 	int idx;
1668 	unsigned long s0 = 0, s1 = 0;
1669 	int ss_state = READ_ONCE(ssp->srcu_size_state);
1670 	int ss_state_idx = ss_state;
1671 
1672 	idx = ssp->srcu_idx & 0x1;
1673 	if (ss_state < 0 || ss_state >= ARRAY_SIZE(srcu_size_state_name))
1674 		ss_state_idx = ARRAY_SIZE(srcu_size_state_name) - 1;
1675 	pr_alert("%s%s Tree SRCU g%ld state %d (%s)",
1676 		 tt, tf, rcu_seq_current(&ssp->srcu_gp_seq), ss_state,
1677 		 srcu_size_state_name[ss_state_idx]);
1678 	if (!ssp->sda) {
1679 		// Called after cleanup_srcu_struct(), perhaps.
1680 		pr_cont(" No per-CPU srcu_data structures (->sda == NULL).\n");
1681 	} else {
1682 		pr_cont(" per-CPU(idx=%d):", idx);
1683 		for_each_possible_cpu(cpu) {
1684 			unsigned long l0, l1;
1685 			unsigned long u0, u1;
1686 			long c0, c1;
1687 			struct srcu_data *sdp;
1688 
1689 			sdp = per_cpu_ptr(ssp->sda, cpu);
1690 			u0 = data_race(sdp->srcu_unlock_count[!idx]);
1691 			u1 = data_race(sdp->srcu_unlock_count[idx]);
1692 
1693 			/*
1694 			 * Make sure that a lock is always counted if the corresponding
1695 			 * unlock is counted.
1696 			 */
1697 			smp_rmb();
1698 
1699 			l0 = data_race(sdp->srcu_lock_count[!idx]);
1700 			l1 = data_race(sdp->srcu_lock_count[idx]);
1701 
1702 			c0 = l0 - u0;
1703 			c1 = l1 - u1;
1704 			pr_cont(" %d(%ld,%ld %c)",
1705 				cpu, c0, c1,
1706 				"C."[rcu_segcblist_empty(&sdp->srcu_cblist)]);
1707 			s0 += c0;
1708 			s1 += c1;
1709 		}
1710 		pr_cont(" T(%ld,%ld)\n", s0, s1);
1711 	}
1712 	if (SRCU_SIZING_IS_TORTURE())
1713 		srcu_transition_to_big(ssp);
1714 }
1715 EXPORT_SYMBOL_GPL(srcu_torture_stats_print);
1716 
srcu_bootup_announce(void)1717 static int __init srcu_bootup_announce(void)
1718 {
1719 	pr_info("Hierarchical SRCU implementation.\n");
1720 	if (exp_holdoff != DEFAULT_SRCU_EXP_HOLDOFF)
1721 		pr_info("\tNon-default auto-expedite holdoff of %lu ns.\n", exp_holdoff);
1722 	if (srcu_retry_check_delay != SRCU_DEFAULT_RETRY_CHECK_DELAY)
1723 		pr_info("\tNon-default retry check delay of %lu us.\n", srcu_retry_check_delay);
1724 	if (srcu_max_nodelay != SRCU_DEFAULT_MAX_NODELAY)
1725 		pr_info("\tNon-default max no-delay of %lu.\n", srcu_max_nodelay);
1726 	pr_info("\tMax phase no-delay instances is %lu.\n", srcu_max_nodelay_phase);
1727 	return 0;
1728 }
1729 early_initcall(srcu_bootup_announce);
1730 
srcu_init(void)1731 void __init srcu_init(void)
1732 {
1733 	struct srcu_struct *ssp;
1734 
1735 	/* Decide on srcu_struct-size strategy. */
1736 	if (SRCU_SIZING_IS(SRCU_SIZING_AUTO)) {
1737 		if (nr_cpu_ids >= big_cpu_lim) {
1738 			convert_to_big = SRCU_SIZING_INIT; // Don't bother waiting for contention.
1739 			pr_info("%s: Setting srcu_struct sizes to big.\n", __func__);
1740 		} else {
1741 			convert_to_big = SRCU_SIZING_NONE | SRCU_SIZING_CONTEND;
1742 			pr_info("%s: Setting srcu_struct sizes based on contention.\n", __func__);
1743 		}
1744 	}
1745 
1746 	/*
1747 	 * Once that is set, call_srcu() can follow the normal path and
1748 	 * queue delayed work. This must follow RCU workqueues creation
1749 	 * and timers initialization.
1750 	 */
1751 	srcu_init_done = true;
1752 	while (!list_empty(&srcu_boot_list)) {
1753 		ssp = list_first_entry(&srcu_boot_list, struct srcu_struct,
1754 				      work.work.entry);
1755 		list_del_init(&ssp->work.work.entry);
1756 		if (SRCU_SIZING_IS(SRCU_SIZING_INIT) && ssp->srcu_size_state == SRCU_SIZE_SMALL)
1757 			ssp->srcu_size_state = SRCU_SIZE_ALLOC;
1758 		queue_work(rcu_gp_wq, &ssp->work.work);
1759 	}
1760 }
1761 
1762 #ifdef CONFIG_MODULES
1763 
1764 /* Initialize any global-scope srcu_struct structures used by this module. */
srcu_module_coming(struct module * mod)1765 static int srcu_module_coming(struct module *mod)
1766 {
1767 	int i;
1768 	struct srcu_struct **sspp = mod->srcu_struct_ptrs;
1769 	int ret;
1770 
1771 	for (i = 0; i < mod->num_srcu_structs; i++) {
1772 		ret = init_srcu_struct(*(sspp++));
1773 		if (WARN_ON_ONCE(ret))
1774 			return ret;
1775 	}
1776 	return 0;
1777 }
1778 
1779 /* Clean up any global-scope srcu_struct structures used by this module. */
srcu_module_going(struct module * mod)1780 static void srcu_module_going(struct module *mod)
1781 {
1782 	int i;
1783 	struct srcu_struct **sspp = mod->srcu_struct_ptrs;
1784 
1785 	for (i = 0; i < mod->num_srcu_structs; i++)
1786 		cleanup_srcu_struct(*(sspp++));
1787 }
1788 
1789 /* Handle one module, either coming or going. */
srcu_module_notify(struct notifier_block * self,unsigned long val,void * data)1790 static int srcu_module_notify(struct notifier_block *self,
1791 			      unsigned long val, void *data)
1792 {
1793 	struct module *mod = data;
1794 	int ret = 0;
1795 
1796 	switch (val) {
1797 	case MODULE_STATE_COMING:
1798 		ret = srcu_module_coming(mod);
1799 		break;
1800 	case MODULE_STATE_GOING:
1801 		srcu_module_going(mod);
1802 		break;
1803 	default:
1804 		break;
1805 	}
1806 	return ret;
1807 }
1808 
1809 static struct notifier_block srcu_module_nb = {
1810 	.notifier_call = srcu_module_notify,
1811 	.priority = 0,
1812 };
1813 
init_srcu_module_notifier(void)1814 static __init int init_srcu_module_notifier(void)
1815 {
1816 	int ret;
1817 
1818 	ret = register_module_notifier(&srcu_module_nb);
1819 	if (ret)
1820 		pr_warn("Failed to register srcu module notifier\n");
1821 	return ret;
1822 }
1823 late_initcall(init_srcu_module_notifier);
1824 
1825 #endif /* #ifdef CONFIG_MODULES */
1826