| 1 | /* |
| 2 | * Sleepable Read-Copy Update mechanism for mutual exclusion. |
| 3 | * |
| 4 | * This program is free software; you can redistribute it and/or modify |
| 5 | * it under the terms of the GNU General Public License as published by |
| 6 | * the Free Software Foundation; either version 2 of the License, or |
| 7 | * (at your option) any later version. |
| 8 | * |
| 9 | * This program is distributed in the hope that it will be useful, |
| 10 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 11 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 12 | * GNU General Public License for more details. |
| 13 | * |
| 14 | * You should have received a copy of the GNU General Public License |
| 15 | * along with this program; if not, write to the Free Software |
| 16 | * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA. |
| 17 | * |
| 18 | * Copyright (C) IBM Corporation, 2006 |
| 19 | * Copyright (C) Fujitsu, 2012 |
| 20 | * |
| 21 | * Author: Paul McKenney <paulmck@us.ibm.com> |
| 22 | * Lai Jiangshan <laijs@cn.fujitsu.com> |
| 23 | * |
| 24 | * For detailed explanation of Read-Copy Update mechanism see - |
| 25 | * Documentation/RCU/ *.txt |
| 26 | * |
| 27 | */ |
| 28 | |
| 29 | #include <linux/export.h> |
| 30 | #include <linux/mutex.h> |
| 31 | #include <linux/percpu.h> |
| 32 | #include <linux/preempt.h> |
| 33 | #include <linux/rcupdate.h> |
| 34 | #include <linux/sched.h> |
| 35 | #include <linux/smp.h> |
| 36 | #include <linux/delay.h> |
| 37 | #include <linux/srcu.h> |
| 38 | |
| 39 | /* |
| 40 | * Initialize an rcu_batch structure to empty. |
| 41 | */ |
| 42 | static inline void rcu_batch_init(struct rcu_batch *b) |
| 43 | { |
| 44 | b->head = NULL; |
| 45 | b->tail = &b->head; |
| 46 | } |
| 47 | |
| 48 | /* |
| 49 | * Enqueue a callback onto the tail of the specified rcu_batch structure. |
| 50 | */ |
| 51 | static inline void rcu_batch_queue(struct rcu_batch *b, struct rcu_head *head) |
| 52 | { |
| 53 | *b->tail = head; |
| 54 | b->tail = &head->next; |
| 55 | } |
| 56 | |
| 57 | /* |
| 58 | * Is the specified rcu_batch structure empty? |
| 59 | */ |
| 60 | static inline bool rcu_batch_empty(struct rcu_batch *b) |
| 61 | { |
| 62 | return b->tail == &b->head; |
| 63 | } |
| 64 | |
| 65 | /* |
| 66 | * Remove the callback at the head of the specified rcu_batch structure |
| 67 | * and return a pointer to it, or return NULL if the structure is empty. |
| 68 | */ |
| 69 | static inline struct rcu_head *rcu_batch_dequeue(struct rcu_batch *b) |
| 70 | { |
| 71 | struct rcu_head *head; |
| 72 | |
| 73 | if (rcu_batch_empty(b)) |
| 74 | return NULL; |
| 75 | |
| 76 | head = b->head; |
| 77 | b->head = head->next; |
| 78 | if (b->tail == &head->next) |
| 79 | rcu_batch_init(b); |
| 80 | |
| 81 | return head; |
| 82 | } |
| 83 | |
| 84 | /* |
| 85 | * Move all callbacks from the rcu_batch structure specified by "from" to |
| 86 | * the structure specified by "to". |
| 87 | */ |
| 88 | static inline void rcu_batch_move(struct rcu_batch *to, struct rcu_batch *from) |
| 89 | { |
| 90 | if (!rcu_batch_empty(from)) { |
| 91 | *to->tail = from->head; |
| 92 | to->tail = from->tail; |
| 93 | rcu_batch_init(from); |
| 94 | } |
| 95 | } |
| 96 | |
| 97 | /* single-thread state-machine */ |
| 98 | static void process_srcu(struct work_struct *work); |
| 99 | |
| 100 | static int init_srcu_struct_fields(struct srcu_struct *sp) |
| 101 | { |
| 102 | sp->completed = 0; |
| 103 | spin_lock_init(&sp->queue_lock); |
| 104 | sp->running = false; |
| 105 | rcu_batch_init(&sp->batch_queue); |
| 106 | rcu_batch_init(&sp->batch_check0); |
| 107 | rcu_batch_init(&sp->batch_check1); |
| 108 | rcu_batch_init(&sp->batch_done); |
| 109 | INIT_DELAYED_WORK(&sp->work, process_srcu); |
| 110 | sp->per_cpu_ref = alloc_percpu(struct srcu_struct_array); |
| 111 | return sp->per_cpu_ref ? 0 : -ENOMEM; |
| 112 | } |
| 113 | |
| 114 | #ifdef CONFIG_DEBUG_LOCK_ALLOC |
| 115 | |
| 116 | int __init_srcu_struct(struct srcu_struct *sp, const char *name, |
| 117 | struct lock_class_key *key) |
| 118 | { |
| 119 | /* Don't re-initialize a lock while it is held. */ |
| 120 | debug_check_no_locks_freed((void *)sp, sizeof(*sp)); |
| 121 | lockdep_init_map(&sp->dep_map, name, key, 0); |
| 122 | return init_srcu_struct_fields(sp); |
| 123 | } |
| 124 | EXPORT_SYMBOL_GPL(__init_srcu_struct); |
| 125 | |
| 126 | #else /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */ |
| 127 | |
| 128 | /** |
| 129 | * init_srcu_struct - initialize a sleep-RCU structure |
| 130 | * @sp: structure to initialize. |
| 131 | * |
| 132 | * Must invoke this on a given srcu_struct before passing that srcu_struct |
| 133 | * to any other function. Each srcu_struct represents a separate domain |
| 134 | * of SRCU protection. |
| 135 | */ |
| 136 | int init_srcu_struct(struct srcu_struct *sp) |
| 137 | { |
| 138 | return init_srcu_struct_fields(sp); |
| 139 | } |
| 140 | EXPORT_SYMBOL_GPL(init_srcu_struct); |
| 141 | |
| 142 | #endif /* #else #ifdef CONFIG_DEBUG_LOCK_ALLOC */ |
| 143 | |
| 144 | /* |
| 145 | * Returns approximate total of the readers' ->seq[] values for the |
| 146 | * rank of per-CPU counters specified by idx. |
| 147 | */ |
| 148 | static unsigned long srcu_readers_seq_idx(struct srcu_struct *sp, int idx) |
| 149 | { |
| 150 | int cpu; |
| 151 | unsigned long sum = 0; |
| 152 | unsigned long t; |
| 153 | |
| 154 | for_each_possible_cpu(cpu) { |
| 155 | t = ACCESS_ONCE(per_cpu_ptr(sp->per_cpu_ref, cpu)->seq[idx]); |
| 156 | sum += t; |
| 157 | } |
| 158 | return sum; |
| 159 | } |
| 160 | |
| 161 | /* |
| 162 | * Returns approximate number of readers active on the specified rank |
| 163 | * of the per-CPU ->c[] counters. |
| 164 | */ |
| 165 | static unsigned long srcu_readers_active_idx(struct srcu_struct *sp, int idx) |
| 166 | { |
| 167 | int cpu; |
| 168 | unsigned long sum = 0; |
| 169 | unsigned long t; |
| 170 | |
| 171 | for_each_possible_cpu(cpu) { |
| 172 | t = ACCESS_ONCE(per_cpu_ptr(sp->per_cpu_ref, cpu)->c[idx]); |
| 173 | sum += t; |
| 174 | } |
| 175 | return sum; |
| 176 | } |
| 177 | |
| 178 | /* |
| 179 | * Return true if the number of pre-existing readers is determined to |
| 180 | * be stably zero. An example unstable zero can occur if the call |
| 181 | * to srcu_readers_active_idx() misses an __srcu_read_lock() increment, |
| 182 | * but due to task migration, sees the corresponding __srcu_read_unlock() |
| 183 | * decrement. This can happen because srcu_readers_active_idx() takes |
| 184 | * time to sum the array, and might in fact be interrupted or preempted |
| 185 | * partway through the summation. |
| 186 | */ |
| 187 | static bool srcu_readers_active_idx_check(struct srcu_struct *sp, int idx) |
| 188 | { |
| 189 | unsigned long seq; |
| 190 | |
| 191 | seq = srcu_readers_seq_idx(sp, idx); |
| 192 | |
| 193 | /* |
| 194 | * The following smp_mb() A pairs with the smp_mb() B located in |
| 195 | * __srcu_read_lock(). This pairing ensures that if an |
| 196 | * __srcu_read_lock() increments its counter after the summation |
| 197 | * in srcu_readers_active_idx(), then the corresponding SRCU read-side |
| 198 | * critical section will see any changes made prior to the start |
| 199 | * of the current SRCU grace period. |
| 200 | * |
| 201 | * Also, if the above call to srcu_readers_seq_idx() saw the |
| 202 | * increment of ->seq[], then the call to srcu_readers_active_idx() |
| 203 | * must see the increment of ->c[]. |
| 204 | */ |
| 205 | smp_mb(); /* A */ |
| 206 | |
| 207 | /* |
| 208 | * Note that srcu_readers_active_idx() can incorrectly return |
| 209 | * zero even though there is a pre-existing reader throughout. |
| 210 | * To see this, suppose that task A is in a very long SRCU |
| 211 | * read-side critical section that started on CPU 0, and that |
| 212 | * no other reader exists, so that the sum of the counters |
| 213 | * is equal to one. Then suppose that task B starts executing |
| 214 | * srcu_readers_active_idx(), summing up to CPU 1, and then that |
| 215 | * task C starts reading on CPU 0, so that its increment is not |
| 216 | * summed, but finishes reading on CPU 2, so that its decrement |
| 217 | * -is- summed. Then when task B completes its sum, it will |
| 218 | * incorrectly get zero, despite the fact that task A has been |
| 219 | * in its SRCU read-side critical section the whole time. |
| 220 | * |
| 221 | * We therefore do a validation step should srcu_readers_active_idx() |
| 222 | * return zero. |
| 223 | */ |
| 224 | if (srcu_readers_active_idx(sp, idx) != 0) |
| 225 | return false; |
| 226 | |
| 227 | /* |
| 228 | * The remainder of this function is the validation step. |
| 229 | * The following smp_mb() D pairs with the smp_mb() C in |
| 230 | * __srcu_read_unlock(). If the __srcu_read_unlock() was seen |
| 231 | * by srcu_readers_active_idx() above, then any destructive |
| 232 | * operation performed after the grace period will happen after |
| 233 | * the corresponding SRCU read-side critical section. |
| 234 | * |
| 235 | * Note that there can be at most NR_CPUS worth of readers using |
| 236 | * the old index, which is not enough to overflow even a 32-bit |
| 237 | * integer. (Yes, this does mean that systems having more than |
| 238 | * a billion or so CPUs need to be 64-bit systems.) Therefore, |
| 239 | * the sum of the ->seq[] counters cannot possibly overflow. |
| 240 | * Therefore, the only way that the return values of the two |
| 241 | * calls to srcu_readers_seq_idx() can be equal is if there were |
| 242 | * no increments of the corresponding rank of ->seq[] counts |
| 243 | * in the interim. But the missed-increment scenario laid out |
| 244 | * above includes an increment of the ->seq[] counter by |
| 245 | * the corresponding __srcu_read_lock(). Therefore, if this |
| 246 | * scenario occurs, the return values from the two calls to |
| 247 | * srcu_readers_seq_idx() will differ, and thus the validation |
| 248 | * step below suffices. |
| 249 | */ |
| 250 | smp_mb(); /* D */ |
| 251 | |
| 252 | return srcu_readers_seq_idx(sp, idx) == seq; |
| 253 | } |
| 254 | |
| 255 | /** |
| 256 | * srcu_readers_active - returns approximate number of readers. |
| 257 | * @sp: which srcu_struct to count active readers (holding srcu_read_lock). |
| 258 | * |
| 259 | * Note that this is not an atomic primitive, and can therefore suffer |
| 260 | * severe errors when invoked on an active srcu_struct. That said, it |
| 261 | * can be useful as an error check at cleanup time. |
| 262 | */ |
| 263 | static int srcu_readers_active(struct srcu_struct *sp) |
| 264 | { |
| 265 | int cpu; |
| 266 | unsigned long sum = 0; |
| 267 | |
| 268 | for_each_possible_cpu(cpu) { |
| 269 | sum += ACCESS_ONCE(per_cpu_ptr(sp->per_cpu_ref, cpu)->c[0]); |
| 270 | sum += ACCESS_ONCE(per_cpu_ptr(sp->per_cpu_ref, cpu)->c[1]); |
| 271 | } |
| 272 | return sum; |
| 273 | } |
| 274 | |
| 275 | /** |
| 276 | * cleanup_srcu_struct - deconstruct a sleep-RCU structure |
| 277 | * @sp: structure to clean up. |
| 278 | * |
| 279 | * Must invoke this after you are finished using a given srcu_struct that |
| 280 | * was initialized via init_srcu_struct(), else you leak memory. |
| 281 | */ |
| 282 | void cleanup_srcu_struct(struct srcu_struct *sp) |
| 283 | { |
| 284 | int sum; |
| 285 | |
| 286 | sum = srcu_readers_active(sp); |
| 287 | WARN_ON(sum); /* Leakage unless caller handles error. */ |
| 288 | if (sum != 0) |
| 289 | return; |
| 290 | free_percpu(sp->per_cpu_ref); |
| 291 | sp->per_cpu_ref = NULL; |
| 292 | } |
| 293 | EXPORT_SYMBOL_GPL(cleanup_srcu_struct); |
| 294 | |
| 295 | /* |
| 296 | * Counts the new reader in the appropriate per-CPU element of the |
| 297 | * srcu_struct. Must be called from process context. |
| 298 | * Returns an index that must be passed to the matching srcu_read_unlock(). |
| 299 | */ |
| 300 | int __srcu_read_lock(struct srcu_struct *sp) |
| 301 | { |
| 302 | int idx; |
| 303 | |
| 304 | preempt_disable(); |
| 305 | idx = rcu_dereference_index_check(sp->completed, |
| 306 | rcu_read_lock_sched_held()) & 0x1; |
| 307 | ACCESS_ONCE(this_cpu_ptr(sp->per_cpu_ref)->c[idx]) += 1; |
| 308 | smp_mb(); /* B */ /* Avoid leaking the critical section. */ |
| 309 | ACCESS_ONCE(this_cpu_ptr(sp->per_cpu_ref)->seq[idx]) += 1; |
| 310 | preempt_enable(); |
| 311 | return idx; |
| 312 | } |
| 313 | EXPORT_SYMBOL_GPL(__srcu_read_lock); |
| 314 | |
| 315 | /* |
| 316 | * Removes the count for the old reader from the appropriate per-CPU |
| 317 | * element of the srcu_struct. Note that this may well be a different |
| 318 | * CPU than that which was incremented by the corresponding srcu_read_lock(). |
| 319 | * Must be called from process context. |
| 320 | */ |
| 321 | void __srcu_read_unlock(struct srcu_struct *sp, int idx) |
| 322 | { |
| 323 | preempt_disable(); |
| 324 | smp_mb(); /* C */ /* Avoid leaking the critical section. */ |
| 325 | ACCESS_ONCE(this_cpu_ptr(sp->per_cpu_ref)->c[idx]) -= 1; |
| 326 | preempt_enable(); |
| 327 | } |
| 328 | EXPORT_SYMBOL_GPL(__srcu_read_unlock); |
| 329 | |
| 330 | /* |
| 331 | * We use an adaptive strategy for synchronize_srcu() and especially for |
| 332 | * synchronize_srcu_expedited(). We spin for a fixed time period |
| 333 | * (defined below) to allow SRCU readers to exit their read-side critical |
| 334 | * sections. If there are still some readers after 10 microseconds, |
| 335 | * we repeatedly block for 1-millisecond time periods. This approach |
| 336 | * has done well in testing, so there is no need for a config parameter. |
| 337 | */ |
| 338 | #define SRCU_RETRY_CHECK_DELAY 5 |
| 339 | #define SYNCHRONIZE_SRCU_TRYCOUNT 2 |
| 340 | #define SYNCHRONIZE_SRCU_EXP_TRYCOUNT 12 |
| 341 | |
| 342 | /* |
| 343 | * @@@ Wait until all pre-existing readers complete. Such readers |
| 344 | * will have used the index specified by "idx". |
| 345 | * the caller should ensures the ->completed is not changed while checking |
| 346 | * and idx = (->completed & 1) ^ 1 |
| 347 | */ |
| 348 | static bool try_check_zero(struct srcu_struct *sp, int idx, int trycount) |
| 349 | { |
| 350 | for (;;) { |
| 351 | if (srcu_readers_active_idx_check(sp, idx)) |
| 352 | return true; |
| 353 | if (--trycount <= 0) |
| 354 | return false; |
| 355 | udelay(SRCU_RETRY_CHECK_DELAY); |
| 356 | } |
| 357 | } |
| 358 | |
| 359 | /* |
| 360 | * Increment the ->completed counter so that future SRCU readers will |
| 361 | * use the other rank of the ->c[] and ->seq[] arrays. This allows |
| 362 | * us to wait for pre-existing readers in a starvation-free manner. |
| 363 | */ |
| 364 | static void srcu_flip(struct srcu_struct *sp) |
| 365 | { |
| 366 | sp->completed++; |
| 367 | } |
| 368 | |
| 369 | /* |
| 370 | * Enqueue an SRCU callback on the specified srcu_struct structure, |
| 371 | * initiating grace-period processing if it is not already running. |
| 372 | */ |
| 373 | void call_srcu(struct srcu_struct *sp, struct rcu_head *head, |
| 374 | void (*func)(struct rcu_head *head)) |
| 375 | { |
| 376 | unsigned long flags; |
| 377 | |
| 378 | head->next = NULL; |
| 379 | head->func = func; |
| 380 | spin_lock_irqsave(&sp->queue_lock, flags); |
| 381 | rcu_batch_queue(&sp->batch_queue, head); |
| 382 | if (!sp->running) { |
| 383 | sp->running = true; |
| 384 | schedule_delayed_work(&sp->work, 0); |
| 385 | } |
| 386 | spin_unlock_irqrestore(&sp->queue_lock, flags); |
| 387 | } |
| 388 | EXPORT_SYMBOL_GPL(call_srcu); |
| 389 | |
| 390 | struct rcu_synchronize { |
| 391 | struct rcu_head head; |
| 392 | struct completion completion; |
| 393 | }; |
| 394 | |
| 395 | /* |
| 396 | * Awaken the corresponding synchronize_srcu() instance now that a |
| 397 | * grace period has elapsed. |
| 398 | */ |
| 399 | static void wakeme_after_rcu(struct rcu_head *head) |
| 400 | { |
| 401 | struct rcu_synchronize *rcu; |
| 402 | |
| 403 | rcu = container_of(head, struct rcu_synchronize, head); |
| 404 | complete(&rcu->completion); |
| 405 | } |
| 406 | |
| 407 | static void srcu_advance_batches(struct srcu_struct *sp, int trycount); |
| 408 | static void srcu_reschedule(struct srcu_struct *sp); |
| 409 | |
| 410 | /* |
| 411 | * Helper function for synchronize_srcu() and synchronize_srcu_expedited(). |
| 412 | */ |
| 413 | static void __synchronize_srcu(struct srcu_struct *sp, int trycount) |
| 414 | { |
| 415 | struct rcu_synchronize rcu; |
| 416 | struct rcu_head *head = &rcu.head; |
| 417 | bool done = false; |
| 418 | |
| 419 | rcu_lockdep_assert(!lock_is_held(&sp->dep_map) && |
| 420 | !lock_is_held(&rcu_bh_lock_map) && |
| 421 | !lock_is_held(&rcu_lock_map) && |
| 422 | !lock_is_held(&rcu_sched_lock_map), |
| 423 | "Illegal synchronize_srcu() in same-type SRCU (or RCU) read-side critical section"); |
| 424 | |
| 425 | init_completion(&rcu.completion); |
| 426 | |
| 427 | head->next = NULL; |
| 428 | head->func = wakeme_after_rcu; |
| 429 | spin_lock_irq(&sp->queue_lock); |
| 430 | if (!sp->running) { |
| 431 | /* steal the processing owner */ |
| 432 | sp->running = true; |
| 433 | rcu_batch_queue(&sp->batch_check0, head); |
| 434 | spin_unlock_irq(&sp->queue_lock); |
| 435 | |
| 436 | srcu_advance_batches(sp, trycount); |
| 437 | if (!rcu_batch_empty(&sp->batch_done)) { |
| 438 | BUG_ON(sp->batch_done.head != head); |
| 439 | rcu_batch_dequeue(&sp->batch_done); |
| 440 | done = true; |
| 441 | } |
| 442 | /* give the processing owner to work_struct */ |
| 443 | srcu_reschedule(sp); |
| 444 | } else { |
| 445 | rcu_batch_queue(&sp->batch_queue, head); |
| 446 | spin_unlock_irq(&sp->queue_lock); |
| 447 | } |
| 448 | |
| 449 | if (!done) |
| 450 | wait_for_completion(&rcu.completion); |
| 451 | } |
| 452 | |
| 453 | /** |
| 454 | * synchronize_srcu - wait for prior SRCU read-side critical-section completion |
| 455 | * @sp: srcu_struct with which to synchronize. |
| 456 | * |
| 457 | * Flip the completed counter, and wait for the old count to drain to zero. |
| 458 | * As with classic RCU, the updater must use some separate means of |
| 459 | * synchronizing concurrent updates. Can block; must be called from |
| 460 | * process context. |
| 461 | * |
| 462 | * Note that it is illegal to call synchronize_srcu() from the corresponding |
| 463 | * SRCU read-side critical section; doing so will result in deadlock. |
| 464 | * However, it is perfectly legal to call synchronize_srcu() on one |
| 465 | * srcu_struct from some other srcu_struct's read-side critical section. |
| 466 | */ |
| 467 | void synchronize_srcu(struct srcu_struct *sp) |
| 468 | { |
| 469 | __synchronize_srcu(sp, SYNCHRONIZE_SRCU_TRYCOUNT); |
| 470 | } |
| 471 | EXPORT_SYMBOL_GPL(synchronize_srcu); |
| 472 | |
| 473 | /** |
| 474 | * synchronize_srcu_expedited - Brute-force SRCU grace period |
| 475 | * @sp: srcu_struct with which to synchronize. |
| 476 | * |
| 477 | * Wait for an SRCU grace period to elapse, but be more aggressive about |
| 478 | * spinning rather than blocking when waiting. |
| 479 | * |
| 480 | * Note that it is illegal to call this function while holding any lock |
| 481 | * that is acquired by a CPU-hotplug notifier. It is also illegal to call |
| 482 | * synchronize_srcu_expedited() from the corresponding SRCU read-side |
| 483 | * critical section; doing so will result in deadlock. However, it is |
| 484 | * perfectly legal to call synchronize_srcu_expedited() on one srcu_struct |
| 485 | * from some other srcu_struct's read-side critical section, as long as |
| 486 | * the resulting graph of srcu_structs is acyclic. |
| 487 | */ |
| 488 | void synchronize_srcu_expedited(struct srcu_struct *sp) |
| 489 | { |
| 490 | __synchronize_srcu(sp, SYNCHRONIZE_SRCU_EXP_TRYCOUNT); |
| 491 | } |
| 492 | EXPORT_SYMBOL_GPL(synchronize_srcu_expedited); |
| 493 | |
| 494 | /** |
| 495 | * srcu_barrier - Wait until all in-flight call_srcu() callbacks complete. |
| 496 | */ |
| 497 | void srcu_barrier(struct srcu_struct *sp) |
| 498 | { |
| 499 | synchronize_srcu(sp); |
| 500 | } |
| 501 | EXPORT_SYMBOL_GPL(srcu_barrier); |
| 502 | |
| 503 | /** |
| 504 | * srcu_batches_completed - return batches completed. |
| 505 | * @sp: srcu_struct on which to report batch completion. |
| 506 | * |
| 507 | * Report the number of batches, correlated with, but not necessarily |
| 508 | * precisely the same as, the number of grace periods that have elapsed. |
| 509 | */ |
| 510 | long srcu_batches_completed(struct srcu_struct *sp) |
| 511 | { |
| 512 | return sp->completed; |
| 513 | } |
| 514 | EXPORT_SYMBOL_GPL(srcu_batches_completed); |
| 515 | |
| 516 | #define SRCU_CALLBACK_BATCH 10 |
| 517 | #define SRCU_INTERVAL 1 |
| 518 | |
| 519 | /* |
| 520 | * Move any new SRCU callbacks to the first stage of the SRCU grace |
| 521 | * period pipeline. |
| 522 | */ |
| 523 | static void srcu_collect_new(struct srcu_struct *sp) |
| 524 | { |
| 525 | if (!rcu_batch_empty(&sp->batch_queue)) { |
| 526 | spin_lock_irq(&sp->queue_lock); |
| 527 | rcu_batch_move(&sp->batch_check0, &sp->batch_queue); |
| 528 | spin_unlock_irq(&sp->queue_lock); |
| 529 | } |
| 530 | } |
| 531 | |
| 532 | /* |
| 533 | * Core SRCU state machine. Advance callbacks from ->batch_check0 to |
| 534 | * ->batch_check1 and then to ->batch_done as readers drain. |
| 535 | */ |
| 536 | static void srcu_advance_batches(struct srcu_struct *sp, int trycount) |
| 537 | { |
| 538 | int idx = 1 ^ (sp->completed & 1); |
| 539 | |
| 540 | /* |
| 541 | * Because readers might be delayed for an extended period after |
| 542 | * fetching ->completed for their index, at any point in time there |
| 543 | * might well be readers using both idx=0 and idx=1. We therefore |
| 544 | * need to wait for readers to clear from both index values before |
| 545 | * invoking a callback. |
| 546 | */ |
| 547 | |
| 548 | if (rcu_batch_empty(&sp->batch_check0) && |
| 549 | rcu_batch_empty(&sp->batch_check1)) |
| 550 | return; /* no callbacks need to be advanced */ |
| 551 | |
| 552 | if (!try_check_zero(sp, idx, trycount)) |
| 553 | return; /* failed to advance, will try after SRCU_INTERVAL */ |
| 554 | |
| 555 | /* |
| 556 | * The callbacks in ->batch_check1 have already done with their |
| 557 | * first zero check and flip back when they were enqueued on |
| 558 | * ->batch_check0 in a previous invocation of srcu_advance_batches(). |
| 559 | * (Presumably try_check_zero() returned false during that |
| 560 | * invocation, leaving the callbacks stranded on ->batch_check1.) |
| 561 | * They are therefore ready to invoke, so move them to ->batch_done. |
| 562 | */ |
| 563 | rcu_batch_move(&sp->batch_done, &sp->batch_check1); |
| 564 | |
| 565 | if (rcu_batch_empty(&sp->batch_check0)) |
| 566 | return; /* no callbacks need to be advanced */ |
| 567 | srcu_flip(sp); |
| 568 | |
| 569 | /* |
| 570 | * The callbacks in ->batch_check0 just finished their |
| 571 | * first check zero and flip, so move them to ->batch_check1 |
| 572 | * for future checking on the other idx. |
| 573 | */ |
| 574 | rcu_batch_move(&sp->batch_check1, &sp->batch_check0); |
| 575 | |
| 576 | /* |
| 577 | * SRCU read-side critical sections are normally short, so check |
| 578 | * at least twice in quick succession after a flip. |
| 579 | */ |
| 580 | trycount = trycount < 2 ? 2 : trycount; |
| 581 | if (!try_check_zero(sp, idx^1, trycount)) |
| 582 | return; /* failed to advance, will try after SRCU_INTERVAL */ |
| 583 | |
| 584 | /* |
| 585 | * The callbacks in ->batch_check1 have now waited for all |
| 586 | * pre-existing readers using both idx values. They are therefore |
| 587 | * ready to invoke, so move them to ->batch_done. |
| 588 | */ |
| 589 | rcu_batch_move(&sp->batch_done, &sp->batch_check1); |
| 590 | } |
| 591 | |
| 592 | /* |
| 593 | * Invoke a limited number of SRCU callbacks that have passed through |
| 594 | * their grace period. If there are more to do, SRCU will reschedule |
| 595 | * the workqueue. |
| 596 | */ |
| 597 | static void srcu_invoke_callbacks(struct srcu_struct *sp) |
| 598 | { |
| 599 | int i; |
| 600 | struct rcu_head *head; |
| 601 | |
| 602 | for (i = 0; i < SRCU_CALLBACK_BATCH; i++) { |
| 603 | head = rcu_batch_dequeue(&sp->batch_done); |
| 604 | if (!head) |
| 605 | break; |
| 606 | local_bh_disable(); |
| 607 | head->func(head); |
| 608 | local_bh_enable(); |
| 609 | } |
| 610 | } |
| 611 | |
| 612 | /* |
| 613 | * Finished one round of SRCU grace period. Start another if there are |
| 614 | * more SRCU callbacks queued, otherwise put SRCU into not-running state. |
| 615 | */ |
| 616 | static void srcu_reschedule(struct srcu_struct *sp) |
| 617 | { |
| 618 | bool pending = true; |
| 619 | |
| 620 | if (rcu_batch_empty(&sp->batch_done) && |
| 621 | rcu_batch_empty(&sp->batch_check1) && |
| 622 | rcu_batch_empty(&sp->batch_check0) && |
| 623 | rcu_batch_empty(&sp->batch_queue)) { |
| 624 | spin_lock_irq(&sp->queue_lock); |
| 625 | if (rcu_batch_empty(&sp->batch_done) && |
| 626 | rcu_batch_empty(&sp->batch_check1) && |
| 627 | rcu_batch_empty(&sp->batch_check0) && |
| 628 | rcu_batch_empty(&sp->batch_queue)) { |
| 629 | sp->running = false; |
| 630 | pending = false; |
| 631 | } |
| 632 | spin_unlock_irq(&sp->queue_lock); |
| 633 | } |
| 634 | |
| 635 | if (pending) |
| 636 | schedule_delayed_work(&sp->work, SRCU_INTERVAL); |
| 637 | } |
| 638 | |
| 639 | /* |
| 640 | * This is the work-queue function that handles SRCU grace periods. |
| 641 | */ |
| 642 | static void process_srcu(struct work_struct *work) |
| 643 | { |
| 644 | struct srcu_struct *sp; |
| 645 | |
| 646 | sp = container_of(work, struct srcu_struct, work.work); |
| 647 | |
| 648 | srcu_collect_new(sp); |
| 649 | srcu_advance_batches(sp, 1); |
| 650 | srcu_invoke_callbacks(sp); |
| 651 | srcu_reschedule(sp); |
| 652 | } |