Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1 | /* |
| 2 | * 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, 2001 |
| 19 | * |
| 20 | * Authors: Dipankar Sarma <dipankar@in.ibm.com> |
| 21 | * Manfred Spraul <manfred@colorfullife.com> |
| 22 | * |
| 23 | * Based on the original work by Paul McKenney <paulmck@us.ibm.com> |
| 24 | * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen. |
| 25 | * Papers: |
| 26 | * http://www.rdrop.com/users/paulmck/paper/rclockpdcsproof.pdf |
| 27 | * http://lse.sourceforge.net/locking/rclock_OLS.2001.05.01c.sc.pdf (OLS2001) |
| 28 | * |
| 29 | * For detailed explanation of Read-Copy Update mechanism see - |
| 30 | * http://lse.sourceforge.net/locking/rcupdate.html |
| 31 | * |
| 32 | */ |
| 33 | #include <linux/types.h> |
| 34 | #include <linux/kernel.h> |
| 35 | #include <linux/init.h> |
| 36 | #include <linux/spinlock.h> |
| 37 | #include <linux/smp.h> |
| 38 | #include <linux/interrupt.h> |
| 39 | #include <linux/sched.h> |
| 40 | #include <asm/atomic.h> |
| 41 | #include <linux/bitops.h> |
| 42 | #include <linux/module.h> |
| 43 | #include <linux/completion.h> |
| 44 | #include <linux/moduleparam.h> |
| 45 | #include <linux/percpu.h> |
| 46 | #include <linux/notifier.h> |
| 47 | #include <linux/rcupdate.h> |
| 48 | #include <linux/cpu.h> |
| 49 | |
| 50 | /* Definition for rcupdate control block. */ |
| 51 | struct rcu_ctrlblk rcu_ctrlblk = |
| 52 | { .cur = -300, .completed = -300 }; |
| 53 | struct rcu_ctrlblk rcu_bh_ctrlblk = |
| 54 | { .cur = -300, .completed = -300 }; |
| 55 | |
| 56 | /* Bookkeeping of the progress of the grace period */ |
| 57 | struct rcu_state { |
| 58 | spinlock_t lock; /* Guard this struct and writes to rcu_ctrlblk */ |
| 59 | cpumask_t cpumask; /* CPUs that need to switch in order */ |
| 60 | /* for current batch to proceed. */ |
| 61 | }; |
| 62 | |
| 63 | static struct rcu_state rcu_state ____cacheline_maxaligned_in_smp = |
| 64 | {.lock = SPIN_LOCK_UNLOCKED, .cpumask = CPU_MASK_NONE }; |
| 65 | static struct rcu_state rcu_bh_state ____cacheline_maxaligned_in_smp = |
| 66 | {.lock = SPIN_LOCK_UNLOCKED, .cpumask = CPU_MASK_NONE }; |
| 67 | |
| 68 | DEFINE_PER_CPU(struct rcu_data, rcu_data) = { 0L }; |
| 69 | DEFINE_PER_CPU(struct rcu_data, rcu_bh_data) = { 0L }; |
| 70 | |
| 71 | /* Fake initialization required by compiler */ |
| 72 | static DEFINE_PER_CPU(struct tasklet_struct, rcu_tasklet) = {NULL}; |
| 73 | static int maxbatch = 10; |
| 74 | |
| 75 | /** |
| 76 | * call_rcu - Queue an RCU callback for invocation after a grace period. |
| 77 | * @head: structure to be used for queueing the RCU updates. |
| 78 | * @func: actual update function to be invoked after the grace period |
| 79 | * |
| 80 | * The update function will be invoked some time after a full grace |
| 81 | * period elapses, in other words after all currently executing RCU |
| 82 | * read-side critical sections have completed. RCU read-side critical |
| 83 | * sections are delimited by rcu_read_lock() and rcu_read_unlock(), |
| 84 | * and may be nested. |
| 85 | */ |
| 86 | void fastcall call_rcu(struct rcu_head *head, |
| 87 | void (*func)(struct rcu_head *rcu)) |
| 88 | { |
| 89 | unsigned long flags; |
| 90 | struct rcu_data *rdp; |
| 91 | |
| 92 | head->func = func; |
| 93 | head->next = NULL; |
| 94 | local_irq_save(flags); |
| 95 | rdp = &__get_cpu_var(rcu_data); |
| 96 | *rdp->nxttail = head; |
| 97 | rdp->nxttail = &head->next; |
| 98 | local_irq_restore(flags); |
| 99 | } |
| 100 | |
| 101 | /** |
| 102 | * call_rcu_bh - Queue an RCU for invocation after a quicker grace period. |
| 103 | * @head: structure to be used for queueing the RCU updates. |
| 104 | * @func: actual update function to be invoked after the grace period |
| 105 | * |
| 106 | * The update function will be invoked some time after a full grace |
| 107 | * period elapses, in other words after all currently executing RCU |
| 108 | * read-side critical sections have completed. call_rcu_bh() assumes |
| 109 | * that the read-side critical sections end on completion of a softirq |
| 110 | * handler. This means that read-side critical sections in process |
| 111 | * context must not be interrupted by softirqs. This interface is to be |
| 112 | * used when most of the read-side critical sections are in softirq context. |
| 113 | * RCU read-side critical sections are delimited by rcu_read_lock() and |
| 114 | * rcu_read_unlock(), * if in interrupt context or rcu_read_lock_bh() |
| 115 | * and rcu_read_unlock_bh(), if in process context. These may be nested. |
| 116 | */ |
| 117 | void fastcall call_rcu_bh(struct rcu_head *head, |
| 118 | void (*func)(struct rcu_head *rcu)) |
| 119 | { |
| 120 | unsigned long flags; |
| 121 | struct rcu_data *rdp; |
| 122 | |
| 123 | head->func = func; |
| 124 | head->next = NULL; |
| 125 | local_irq_save(flags); |
| 126 | rdp = &__get_cpu_var(rcu_bh_data); |
| 127 | *rdp->nxttail = head; |
| 128 | rdp->nxttail = &head->next; |
| 129 | local_irq_restore(flags); |
| 130 | } |
| 131 | |
| 132 | /* |
| 133 | * Invoke the completed RCU callbacks. They are expected to be in |
| 134 | * a per-cpu list. |
| 135 | */ |
| 136 | static void rcu_do_batch(struct rcu_data *rdp) |
| 137 | { |
| 138 | struct rcu_head *next, *list; |
| 139 | int count = 0; |
| 140 | |
| 141 | list = rdp->donelist; |
| 142 | while (list) { |
| 143 | next = rdp->donelist = list->next; |
| 144 | list->func(list); |
| 145 | list = next; |
| 146 | if (++count >= maxbatch) |
| 147 | break; |
| 148 | } |
| 149 | if (!rdp->donelist) |
| 150 | rdp->donetail = &rdp->donelist; |
| 151 | else |
| 152 | tasklet_schedule(&per_cpu(rcu_tasklet, rdp->cpu)); |
| 153 | } |
| 154 | |
| 155 | /* |
| 156 | * Grace period handling: |
| 157 | * The grace period handling consists out of two steps: |
| 158 | * - A new grace period is started. |
| 159 | * This is done by rcu_start_batch. The start is not broadcasted to |
| 160 | * all cpus, they must pick this up by comparing rcp->cur with |
| 161 | * rdp->quiescbatch. All cpus are recorded in the |
| 162 | * rcu_state.cpumask bitmap. |
| 163 | * - All cpus must go through a quiescent state. |
| 164 | * Since the start of the grace period is not broadcasted, at least two |
| 165 | * calls to rcu_check_quiescent_state are required: |
| 166 | * The first call just notices that a new grace period is running. The |
| 167 | * following calls check if there was a quiescent state since the beginning |
| 168 | * of the grace period. If so, it updates rcu_state.cpumask. If |
| 169 | * the bitmap is empty, then the grace period is completed. |
| 170 | * rcu_check_quiescent_state calls rcu_start_batch(0) to start the next grace |
| 171 | * period (if necessary). |
| 172 | */ |
| 173 | /* |
| 174 | * Register a new batch of callbacks, and start it up if there is currently no |
| 175 | * active batch and the batch to be registered has not already occurred. |
| 176 | * Caller must hold rcu_state.lock. |
| 177 | */ |
| 178 | static void rcu_start_batch(struct rcu_ctrlblk *rcp, struct rcu_state *rsp, |
| 179 | int next_pending) |
| 180 | { |
| 181 | if (next_pending) |
| 182 | rcp->next_pending = 1; |
| 183 | |
| 184 | if (rcp->next_pending && |
| 185 | rcp->completed == rcp->cur) { |
| 186 | /* Can't change, since spin lock held. */ |
| 187 | cpus_andnot(rsp->cpumask, cpu_online_map, nohz_cpu_mask); |
| 188 | |
| 189 | rcp->next_pending = 0; |
| 190 | /* next_pending == 0 must be visible in __rcu_process_callbacks() |
| 191 | * before it can see new value of cur. |
| 192 | */ |
| 193 | smp_wmb(); |
| 194 | rcp->cur++; |
| 195 | } |
| 196 | } |
| 197 | |
| 198 | /* |
| 199 | * cpu went through a quiescent state since the beginning of the grace period. |
| 200 | * Clear it from the cpu mask and complete the grace period if it was the last |
| 201 | * cpu. Start another grace period if someone has further entries pending |
| 202 | */ |
| 203 | static void cpu_quiet(int cpu, struct rcu_ctrlblk *rcp, struct rcu_state *rsp) |
| 204 | { |
| 205 | cpu_clear(cpu, rsp->cpumask); |
| 206 | if (cpus_empty(rsp->cpumask)) { |
| 207 | /* batch completed ! */ |
| 208 | rcp->completed = rcp->cur; |
| 209 | rcu_start_batch(rcp, rsp, 0); |
| 210 | } |
| 211 | } |
| 212 | |
| 213 | /* |
| 214 | * Check if the cpu has gone through a quiescent state (say context |
| 215 | * switch). If so and if it already hasn't done so in this RCU |
| 216 | * quiescent cycle, then indicate that it has done so. |
| 217 | */ |
| 218 | static void rcu_check_quiescent_state(struct rcu_ctrlblk *rcp, |
| 219 | struct rcu_state *rsp, struct rcu_data *rdp) |
| 220 | { |
| 221 | if (rdp->quiescbatch != rcp->cur) { |
| 222 | /* start new grace period: */ |
| 223 | rdp->qs_pending = 1; |
| 224 | rdp->passed_quiesc = 0; |
| 225 | rdp->quiescbatch = rcp->cur; |
| 226 | return; |
| 227 | } |
| 228 | |
| 229 | /* Grace period already completed for this cpu? |
| 230 | * qs_pending is checked instead of the actual bitmap to avoid |
| 231 | * cacheline trashing. |
| 232 | */ |
| 233 | if (!rdp->qs_pending) |
| 234 | return; |
| 235 | |
| 236 | /* |
| 237 | * Was there a quiescent state since the beginning of the grace |
| 238 | * period? If no, then exit and wait for the next call. |
| 239 | */ |
| 240 | if (!rdp->passed_quiesc) |
| 241 | return; |
| 242 | rdp->qs_pending = 0; |
| 243 | |
| 244 | spin_lock(&rsp->lock); |
| 245 | /* |
| 246 | * rdp->quiescbatch/rcp->cur and the cpu bitmap can come out of sync |
| 247 | * during cpu startup. Ignore the quiescent state. |
| 248 | */ |
| 249 | if (likely(rdp->quiescbatch == rcp->cur)) |
| 250 | cpu_quiet(rdp->cpu, rcp, rsp); |
| 251 | |
| 252 | spin_unlock(&rsp->lock); |
| 253 | } |
| 254 | |
| 255 | |
| 256 | #ifdef CONFIG_HOTPLUG_CPU |
| 257 | |
| 258 | /* warning! helper for rcu_offline_cpu. do not use elsewhere without reviewing |
| 259 | * locking requirements, the list it's pulling from has to belong to a cpu |
| 260 | * which is dead and hence not processing interrupts. |
| 261 | */ |
| 262 | static void rcu_move_batch(struct rcu_data *this_rdp, struct rcu_head *list, |
| 263 | struct rcu_head **tail) |
| 264 | { |
| 265 | local_irq_disable(); |
| 266 | *this_rdp->nxttail = list; |
| 267 | if (list) |
| 268 | this_rdp->nxttail = tail; |
| 269 | local_irq_enable(); |
| 270 | } |
| 271 | |
| 272 | static void __rcu_offline_cpu(struct rcu_data *this_rdp, |
| 273 | struct rcu_ctrlblk *rcp, struct rcu_state *rsp, struct rcu_data *rdp) |
| 274 | { |
| 275 | /* if the cpu going offline owns the grace period |
| 276 | * we can block indefinitely waiting for it, so flush |
| 277 | * it here |
| 278 | */ |
| 279 | spin_lock_bh(&rsp->lock); |
| 280 | if (rcp->cur != rcp->completed) |
| 281 | cpu_quiet(rdp->cpu, rcp, rsp); |
| 282 | spin_unlock_bh(&rsp->lock); |
| 283 | rcu_move_batch(this_rdp, rdp->curlist, rdp->curtail); |
| 284 | rcu_move_batch(this_rdp, rdp->nxtlist, rdp->nxttail); |
| 285 | |
| 286 | } |
| 287 | static void rcu_offline_cpu(int cpu) |
| 288 | { |
| 289 | struct rcu_data *this_rdp = &get_cpu_var(rcu_data); |
| 290 | struct rcu_data *this_bh_rdp = &get_cpu_var(rcu_bh_data); |
| 291 | |
| 292 | __rcu_offline_cpu(this_rdp, &rcu_ctrlblk, &rcu_state, |
| 293 | &per_cpu(rcu_data, cpu)); |
| 294 | __rcu_offline_cpu(this_bh_rdp, &rcu_bh_ctrlblk, &rcu_bh_state, |
| 295 | &per_cpu(rcu_bh_data, cpu)); |
| 296 | put_cpu_var(rcu_data); |
| 297 | put_cpu_var(rcu_bh_data); |
| 298 | tasklet_kill_immediate(&per_cpu(rcu_tasklet, cpu), cpu); |
| 299 | } |
| 300 | |
| 301 | #else |
| 302 | |
| 303 | static void rcu_offline_cpu(int cpu) |
| 304 | { |
| 305 | } |
| 306 | |
| 307 | #endif |
| 308 | |
| 309 | /* |
| 310 | * This does the RCU processing work from tasklet context. |
| 311 | */ |
| 312 | static void __rcu_process_callbacks(struct rcu_ctrlblk *rcp, |
| 313 | struct rcu_state *rsp, struct rcu_data *rdp) |
| 314 | { |
| 315 | if (rdp->curlist && !rcu_batch_before(rcp->completed, rdp->batch)) { |
| 316 | *rdp->donetail = rdp->curlist; |
| 317 | rdp->donetail = rdp->curtail; |
| 318 | rdp->curlist = NULL; |
| 319 | rdp->curtail = &rdp->curlist; |
| 320 | } |
| 321 | |
| 322 | local_irq_disable(); |
| 323 | if (rdp->nxtlist && !rdp->curlist) { |
| 324 | rdp->curlist = rdp->nxtlist; |
| 325 | rdp->curtail = rdp->nxttail; |
| 326 | rdp->nxtlist = NULL; |
| 327 | rdp->nxttail = &rdp->nxtlist; |
| 328 | local_irq_enable(); |
| 329 | |
| 330 | /* |
| 331 | * start the next batch of callbacks |
| 332 | */ |
| 333 | |
| 334 | /* determine batch number */ |
| 335 | rdp->batch = rcp->cur + 1; |
| 336 | /* see the comment and corresponding wmb() in |
| 337 | * the rcu_start_batch() |
| 338 | */ |
| 339 | smp_rmb(); |
| 340 | |
| 341 | if (!rcp->next_pending) { |
| 342 | /* and start it/schedule start if it's a new batch */ |
| 343 | spin_lock(&rsp->lock); |
| 344 | rcu_start_batch(rcp, rsp, 1); |
| 345 | spin_unlock(&rsp->lock); |
| 346 | } |
| 347 | } else { |
| 348 | local_irq_enable(); |
| 349 | } |
| 350 | rcu_check_quiescent_state(rcp, rsp, rdp); |
| 351 | if (rdp->donelist) |
| 352 | rcu_do_batch(rdp); |
| 353 | } |
| 354 | |
| 355 | static void rcu_process_callbacks(unsigned long unused) |
| 356 | { |
| 357 | __rcu_process_callbacks(&rcu_ctrlblk, &rcu_state, |
| 358 | &__get_cpu_var(rcu_data)); |
| 359 | __rcu_process_callbacks(&rcu_bh_ctrlblk, &rcu_bh_state, |
| 360 | &__get_cpu_var(rcu_bh_data)); |
| 361 | } |
| 362 | |
| 363 | void rcu_check_callbacks(int cpu, int user) |
| 364 | { |
| 365 | if (user || |
| 366 | (idle_cpu(cpu) && !in_softirq() && |
| 367 | hardirq_count() <= (1 << HARDIRQ_SHIFT))) { |
| 368 | rcu_qsctr_inc(cpu); |
| 369 | rcu_bh_qsctr_inc(cpu); |
| 370 | } else if (!in_softirq()) |
| 371 | rcu_bh_qsctr_inc(cpu); |
| 372 | tasklet_schedule(&per_cpu(rcu_tasklet, cpu)); |
| 373 | } |
| 374 | |
| 375 | static void rcu_init_percpu_data(int cpu, struct rcu_ctrlblk *rcp, |
| 376 | struct rcu_data *rdp) |
| 377 | { |
| 378 | memset(rdp, 0, sizeof(*rdp)); |
| 379 | rdp->curtail = &rdp->curlist; |
| 380 | rdp->nxttail = &rdp->nxtlist; |
| 381 | rdp->donetail = &rdp->donelist; |
| 382 | rdp->quiescbatch = rcp->completed; |
| 383 | rdp->qs_pending = 0; |
| 384 | rdp->cpu = cpu; |
| 385 | } |
| 386 | |
| 387 | static void __devinit rcu_online_cpu(int cpu) |
| 388 | { |
| 389 | struct rcu_data *rdp = &per_cpu(rcu_data, cpu); |
| 390 | struct rcu_data *bh_rdp = &per_cpu(rcu_bh_data, cpu); |
| 391 | |
| 392 | rcu_init_percpu_data(cpu, &rcu_ctrlblk, rdp); |
| 393 | rcu_init_percpu_data(cpu, &rcu_bh_ctrlblk, bh_rdp); |
| 394 | tasklet_init(&per_cpu(rcu_tasklet, cpu), rcu_process_callbacks, 0UL); |
| 395 | } |
| 396 | |
| 397 | static int __devinit rcu_cpu_notify(struct notifier_block *self, |
| 398 | unsigned long action, void *hcpu) |
| 399 | { |
| 400 | long cpu = (long)hcpu; |
| 401 | switch (action) { |
| 402 | case CPU_UP_PREPARE: |
| 403 | rcu_online_cpu(cpu); |
| 404 | break; |
| 405 | case CPU_DEAD: |
| 406 | rcu_offline_cpu(cpu); |
| 407 | break; |
| 408 | default: |
| 409 | break; |
| 410 | } |
| 411 | return NOTIFY_OK; |
| 412 | } |
| 413 | |
| 414 | static struct notifier_block __devinitdata rcu_nb = { |
| 415 | .notifier_call = rcu_cpu_notify, |
| 416 | }; |
| 417 | |
| 418 | /* |
| 419 | * Initializes rcu mechanism. Assumed to be called early. |
| 420 | * That is before local timer(SMP) or jiffie timer (uniproc) is setup. |
| 421 | * Note that rcu_qsctr and friends are implicitly |
| 422 | * initialized due to the choice of ``0'' for RCU_CTR_INVALID. |
| 423 | */ |
| 424 | void __init rcu_init(void) |
| 425 | { |
| 426 | rcu_cpu_notify(&rcu_nb, CPU_UP_PREPARE, |
| 427 | (void *)(long)smp_processor_id()); |
| 428 | /* Register notifier for non-boot CPUs */ |
| 429 | register_cpu_notifier(&rcu_nb); |
| 430 | } |
| 431 | |
| 432 | struct rcu_synchronize { |
| 433 | struct rcu_head head; |
| 434 | struct completion completion; |
| 435 | }; |
| 436 | |
| 437 | /* Because of FASTCALL declaration of complete, we use this wrapper */ |
| 438 | static void wakeme_after_rcu(struct rcu_head *head) |
| 439 | { |
| 440 | struct rcu_synchronize *rcu; |
| 441 | |
| 442 | rcu = container_of(head, struct rcu_synchronize, head); |
| 443 | complete(&rcu->completion); |
| 444 | } |
| 445 | |
| 446 | /** |
Paul E. McKenney | 9b06e81 | 2005-05-01 08:59:04 -0700 | [diff] [blame] | 447 | * synchronize_rcu - wait until a grace period has elapsed. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 448 | * |
| 449 | * Control will return to the caller some time after a full grace |
| 450 | * period has elapsed, in other words after all currently executing RCU |
| 451 | * read-side critical sections have completed. RCU read-side critical |
| 452 | * sections are delimited by rcu_read_lock() and rcu_read_unlock(), |
| 453 | * and may be nested. |
Paul E. McKenney | 9b06e81 | 2005-05-01 08:59:04 -0700 | [diff] [blame] | 454 | * |
| 455 | * If your read-side code is not protected by rcu_read_lock(), do -not- |
| 456 | * use synchronize_rcu(). |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 457 | */ |
Paul E. McKenney | 9b06e81 | 2005-05-01 08:59:04 -0700 | [diff] [blame] | 458 | void synchronize_rcu(void) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 459 | { |
| 460 | struct rcu_synchronize rcu; |
| 461 | |
| 462 | init_completion(&rcu.completion); |
| 463 | /* Will wake me after RCU finished */ |
| 464 | call_rcu(&rcu.head, wakeme_after_rcu); |
| 465 | |
| 466 | /* Wait for it */ |
| 467 | wait_for_completion(&rcu.completion); |
| 468 | } |
| 469 | |
Paul E. McKenney | 9b06e81 | 2005-05-01 08:59:04 -0700 | [diff] [blame] | 470 | /* |
| 471 | * Deprecated, use synchronize_rcu() or synchronize_sched() instead. |
| 472 | */ |
| 473 | void synchronize_kernel(void) |
| 474 | { |
| 475 | synchronize_rcu(); |
| 476 | } |
| 477 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 478 | module_param(maxbatch, int, 0); |
Paul E. McKenney | 66cf8f1 | 2005-05-01 08:59:03 -0700 | [diff] [blame] | 479 | EXPORT_SYMBOL(call_rcu); /* WARNING: GPL-only in April 2006. */ |
| 480 | EXPORT_SYMBOL(call_rcu_bh); /* WARNING: GPL-only in April 2006. */ |
Paul E. McKenney | 9b06e81 | 2005-05-01 08:59:04 -0700 | [diff] [blame] | 481 | EXPORT_SYMBOL_GPL(synchronize_rcu); |
Paul E. McKenney | 66cf8f1 | 2005-05-01 08:59:03 -0700 | [diff] [blame] | 482 | EXPORT_SYMBOL(synchronize_kernel); /* WARNING: GPL-only in April 2006. */ |