kernel/rcupdate.c

   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/rcupdate.h>
  39 #include <linux/interrupt.h>
  40 #include <linux/sched.h>
  41 #include <asm/atomic.h>
  42 #include <linux/bitops.h>
  43 #include <linux/module.h>
  44 #include <linux/completion.h>
  45 #include <linux/moduleparam.h>
  46 #include <linux/percpu.h>
  47 #include <linux/notifier.h>
  48 #include <linux/rcupdate.h>
  49 #include <linux/cpu.h>
  50
  51 /* Definition for rcupdate control block. */
  52 struct rcu_ctrlblk rcu_ctrlblk = {
  53         .cur = -300,
  54         .completed = -300,
  55         .lock = SPIN_LOCK_UNLOCKED,
  56         .cpumask = CPU_MASK_NONE,
  57 };
  58 struct rcu_ctrlblk rcu_bh_ctrlblk = {
  59         .cur = -300,
  60         .completed = -300,
  61         .lock = SPIN_LOCK_UNLOCKED,
  62         .cpumask = CPU_MASK_NONE,
  63 };
  64
  65 DEFINE_PER_CPU(struct rcu_data, rcu_data) = { 0L };
  66 DEFINE_PER_CPU(struct rcu_data, rcu_bh_data) = { 0L };
  67
  68 /* Fake initialization required by compiler */
  69 static DEFINE_PER_CPU(struct tasklet_struct, rcu_tasklet) = {NULL};
  70 static int maxbatch = 10000;
  71
  72 /**
  73  * call_rcu - Queue an RCU callback for invocation after a grace period.
  74  * @head: structure to be used for queueing the RCU updates.
  75  * @func: actual update function to be invoked after the grace period
  76  *
  77  * The update function will be invoked some time after a full grace
  78  * period elapses, in other words after all currently executing RCU
  79  * read-side critical sections have completed.  RCU read-side critical
  80  * sections are delimited by rcu_read_lock() and rcu_read_unlock(),
  81  * and may be nested.
  82  */
  83 void fastcall call_rcu(struct rcu_head *head,
  84                                 void (*func)(struct rcu_head *rcu))
  85 {
  86         unsigned long flags;
  87         struct rcu_data *rdp;
  88
  89         head->func = func;
  90         head->next = NULL;
  91         local_irq_save(flags);
  92         rdp = &__get_cpu_var(rcu_data);
  93         *rdp->nxttail = head;
  94         rdp->nxttail = &head->next;
  95
  96         if (unlikely(++rdp->count > 10000))
  97                 set_need_resched();
  98
  99         local_irq_restore(flags);
 100 }
 101
 102 static atomic_t rcu_barrier_cpu_count;
 103 static struct semaphore rcu_barrier_sema;
 104 static struct completion rcu_barrier_completion;
 105
 106 /**
 107  * call_rcu_bh - Queue an RCU for invocation after a quicker grace period.
 108  * @head: structure to be used for queueing the RCU updates.
 109  * @func: actual update function to be invoked after the grace period
 110  *
 111  * The update function will be invoked some time after a full grace
 112  * period elapses, in other words after all currently executing RCU
 113  * read-side critical sections have completed. call_rcu_bh() assumes
 114  * that the read-side critical sections end on completion of a softirq
 115  * handler. This means that read-side critical sections in process
 116  * context must not be interrupted by softirqs. This interface is to be
 117  * used when most of the read-side critical sections are in softirq context.
 118  * RCU read-side critical sections are delimited by rcu_read_lock() and
 119  * rcu_read_unlock(), * if in interrupt context or rcu_read_lock_bh()
 120  * and rcu_read_unlock_bh(), if in process context. These may be nested.
 121  */
 122 void fastcall call_rcu_bh(struct rcu_head *head,
 123                                 void (*func)(struct rcu_head *rcu))
 124 {
 125         unsigned long flags;
 126         struct rcu_data *rdp;
 127
 128         head->func = func;
 129         head->next = NULL;
 130         local_irq_save(flags);
 131         rdp = &__get_cpu_var(rcu_bh_data);
 132         *rdp->nxttail = head;
 133         rdp->nxttail = &head->next;
 134         rdp->count++;
 135 /*
 136  *  Should we directly call rcu_do_batch() here ?
 137  *  if (unlikely(rdp->count > 10000))
 138  *      rcu_do_batch(rdp);
 139  */
 140         local_irq_restore(flags);
 141 }
 142
 143 /*
 144  * Return the number of RCU batches processed thus far.  Useful
 145  * for debug and statistics.
 146  */
 147 long rcu_batches_completed(void)
 148 {
 149         return rcu_ctrlblk.completed;
 150 }
 151
 152 static void rcu_barrier_callback(struct rcu_head *notused)
 153 {
 154         if (atomic_dec_and_test(&rcu_barrier_cpu_count))
 155                 complete(&rcu_barrier_completion);
 156 }
 157
 158 /*
 159  * Called with preemption disabled, and from cross-cpu IRQ context.
 160  */
 161 static void rcu_barrier_func(void *notused)
 162 {
 163         int cpu = smp_processor_id();
 164         struct rcu_data *rdp = &per_cpu(rcu_data, cpu);
 165         struct rcu_head *head;
 166
 167         head = &rdp->barrier;
 168         atomic_inc(&rcu_barrier_cpu_count);
 169         call_rcu(head, rcu_barrier_callback);
 170 }
 171
 172 /**
 173  * rcu_barrier - Wait until all the in-flight RCUs are complete.
 174  */
 175 void rcu_barrier(void)
 176 {
 177         BUG_ON(in_interrupt());
 178         /* Take cpucontrol semaphore to protect against CPU hotplug */
 179         down(&rcu_barrier_sema);
 180         init_completion(&rcu_barrier_completion);
 181         atomic_set(&rcu_barrier_cpu_count, 0);
 182         on_each_cpu(rcu_barrier_func, NULL, 0, 1);
 183         wait_for_completion(&rcu_barrier_completion);
 184         up(&rcu_barrier_sema);
 185 }
 186 EXPORT_SYMBOL_GPL(rcu_barrier);
 187
 188 /*
 189  * Invoke the completed RCU callbacks. They are expected to be in
 190  * a per-cpu list.
 191  */
 192 static void rcu_do_batch(struct rcu_data *rdp)
 193 {
 194         struct rcu_head *next, *list;
 195         int count = 0;
 196
 197         list = rdp->donelist;
 198         while (list) {
 199                 next = rdp->donelist = list->next;
 200                 list->func(list);
 201                 list = next;
 202                 rdp->count--;
 203                 if (++count >= maxbatch)
 204                         break;
 205         }
 206         if (!rdp->donelist)
 207                 rdp->donetail = &rdp->donelist;
 208         else
 209                 tasklet_schedule(&per_cpu(rcu_tasklet, rdp->cpu));
 210 }
 211
 212 /*
 213  * Grace period handling:
 214  * The grace period handling consists out of two steps:
 215  * - A new grace period is started.
 216  *   This is done by rcu_start_batch. The start is not broadcasted to
 217  *   all cpus, they must pick this up by comparing rcp->cur with
 218  *   rdp->quiescbatch. All cpus are recorded  in the
 219  *   rcu_ctrlblk.cpumask bitmap.
 220  * - All cpus must go through a quiescent state.
 221  *   Since the start of the grace period is not broadcasted, at least two
 222  *   calls to rcu_check_quiescent_state are required:
 223  *   The first call just notices that a new grace period is running. The
 224  *   following calls check if there was a quiescent state since the beginning
 225  *   of the grace period. If so, it updates rcu_ctrlblk.cpumask. If
 226  *   the bitmap is empty, then the grace period is completed.
 227  *   rcu_check_quiescent_state calls rcu_start_batch(0) to start the next grace
 228  *   period (if necessary).
 229  */
 230 /*
 231  * Register a new batch of callbacks, and start it up if there is currently no
 232  * active batch and the batch to be registered has not already occurred.
 233  * Caller must hold rcu_ctrlblk.lock.
 234  */
 235 static void rcu_start_batch(struct rcu_ctrlblk *rcp)
 236 {
 237         if (rcp->next_pending &&
 238                         rcp->completed == rcp->cur) {
 239                 rcp->next_pending = 0;
 240                 /*
 241                  * next_pending == 0 must be visible in
 242                  * __rcu_process_callbacks() before it can see new value of cur.
 243                  */
 244                 smp_wmb();
 245                 rcp->cur++;
 246
 247                 /*
 248                  * Accessing nohz_cpu_mask before incrementing rcp->cur needs a
 249                  * Barrier  Otherwise it can cause tickless idle CPUs to be
 250                  * included in rcp->cpumask, which will extend graceperiods
 251                  * unnecessarily.
 252                  */
 253                 smp_mb();
 254                 cpus_andnot(rcp->cpumask, cpu_online_map, nohz_cpu_mask);
 255
 256         }
 257 }
 258
 259 /*
 260  * cpu went through a quiescent state since the beginning of the grace period.
 261  * Clear it from the cpu mask and complete the grace period if it was the last
 262  * cpu. Start another grace period if someone has further entries pending
 263  */
 264 static void cpu_quiet(int cpu, struct rcu_ctrlblk *rcp)
 265 {
 266         cpu_clear(cpu, rcp->cpumask);
 267         if (cpus_empty(rcp->cpumask)) {
 268                 /* batch completed ! */
 269                 rcp->completed = rcp->cur;
 270                 rcu_start_batch(rcp);
 271         }
 272 }
 273
 274 /*
 275  * Check if the cpu has gone through a quiescent state (say context
 276  * switch). If so and if it already hasn't done so in this RCU
 277  * quiescent cycle, then indicate that it has done so.
 278  */
 279 static void rcu_check_quiescent_state(struct rcu_ctrlblk *rcp,
 280                                         struct rcu_data *rdp)
 281 {
 282         if (rdp->quiescbatch != rcp->cur) {
 283                 /* start new grace period: */
 284                 rdp->qs_pending = 1;
 285                 rdp->passed_quiesc = 0;
 286                 rdp->quiescbatch = rcp->cur;
 287                 return;
 288         }
 289
 290         /* Grace period already completed for this cpu?
 291          * qs_pending is checked instead of the actual bitmap to avoid
 292          * cacheline trashing.
 293          */
 294         if (!rdp->qs_pending)
 295                 return;
 296
 297         /*
 298          * Was there a quiescent state since the beginning of the grace
 299          * period? If no, then exit and wait for the next call.
 300          */
 301         if (!rdp->passed_quiesc)
 302                 return;
 303         rdp->qs_pending = 0;
 304
 305         spin_lock(&rcp->lock);
 306         /*
 307          * rdp->quiescbatch/rcp->cur and the cpu bitmap can come out of sync
 308          * during cpu startup. Ignore the quiescent state.
 309          */
 310         if (likely(rdp->quiescbatch == rcp->cur))
 311                 cpu_quiet(rdp->cpu, rcp);
 312
 313         spin_unlock(&rcp->lock);
 314 }
 315
 316
 317 #ifdef CONFIG_HOTPLUG_CPU
 318
 319 /* warning! helper for rcu_offline_cpu. do not use elsewhere without reviewing
 320  * locking requirements, the list it's pulling from has to belong to a cpu
 321  * which is dead and hence not processing interrupts.
 322  */
 323 static void rcu_move_batch(struct rcu_data *this_rdp, struct rcu_head *list,
 324                                 struct rcu_head **tail)
 325 {
 326         local_irq_disable();
 327         *this_rdp->nxttail = list;
 328         if (list)
 329                 this_rdp->nxttail = tail;
 330         local_irq_enable();
 331 }
 332
 333 static void __rcu_offline_cpu(struct rcu_data *this_rdp,
 334                                 struct rcu_ctrlblk *rcp, struct rcu_data *rdp)
 335 {
 336         /* if the cpu going offline owns the grace period
 337          * we can block indefinitely waiting for it, so flush
 338          * it here
 339          */
 340         spin_lock_bh(&rcp->lock);
 341         if (rcp->cur != rcp->completed)
 342                 cpu_quiet(rdp->cpu, rcp);
 343         spin_unlock_bh(&rcp->lock);
 344         rcu_move_batch(this_rdp, rdp->curlist, rdp->curtail);
 345         rcu_move_batch(this_rdp, rdp->nxtlist, rdp->nxttail);
 346         rcu_move_batch(this_rdp, rdp->donelist, rdp->donetail);
 347 }
 348
 349 static void rcu_offline_cpu(int cpu)
 350 {
 351         struct rcu_data *this_rdp = &get_cpu_var(rcu_data);
 352         struct rcu_data *this_bh_rdp = &get_cpu_var(rcu_bh_data);
 353
 354         __rcu_offline_cpu(this_rdp, &rcu_ctrlblk,
 355                                         &per_cpu(rcu_data, cpu));
 356         __rcu_offline_cpu(this_bh_rdp, &rcu_bh_ctrlblk,
 357                                         &per_cpu(rcu_bh_data, cpu));
 358         put_cpu_var(rcu_data);
 359         put_cpu_var(rcu_bh_data);
 360         tasklet_kill_immediate(&per_cpu(rcu_tasklet, cpu), cpu);
 361 }
 362
 363 #else
 364
 365 static void rcu_offline_cpu(int cpu)
 366 {
 367 }
 368
 369 #endif
 370
 371 /*
 372  * This does the RCU processing work from tasklet context.
 373  */
 374 static void __rcu_process_callbacks(struct rcu_ctrlblk *rcp,
 375                                         struct rcu_data *rdp)
 376 {
 377         if (rdp->curlist && !rcu_batch_before(rcp->completed, rdp->batch)) {
 378                 *rdp->donetail = rdp->curlist;
 379                 rdp->donetail = rdp->curtail;
 380                 rdp->curlist = NULL;
 381                 rdp->curtail = &rdp->curlist;
 382         }
 383
 384         local_irq_disable();
 385         if (rdp->nxtlist && !rdp->curlist) {
 386                 rdp->curlist = rdp->nxtlist;
 387                 rdp->curtail = rdp->nxttail;
 388                 rdp->nxtlist = NULL;
 389                 rdp->nxttail = &rdp->nxtlist;
 390                 local_irq_enable();
 391
 392                 /*
 393                  * start the next batch of callbacks
 394                  */
 395
 396                 /* determine batch number */
 397                 rdp->batch = rcp->cur + 1;
 398                 /* see the comment and corresponding wmb() in
 399                  * the rcu_start_batch()
 400                  */
 401                 smp_rmb();
 402
 403                 if (!rcp->next_pending) {
 404                         /* and start it/schedule start if it's a new batch */
 405                         spin_lock(&rcp->lock);
 406                         rcp->next_pending = 1;
 407                         rcu_start_batch(rcp);
 408                         spin_unlock(&rcp->lock);
 409                 }
 410         } else {
 411                 local_irq_enable();
 412         }
 413         rcu_check_quiescent_state(rcp, rdp);
 414         if (rdp->donelist)
 415                 rcu_do_batch(rdp);
 416 }
 417
 418 static void rcu_process_callbacks(unsigned long unused)
 419 {
 420         __rcu_process_callbacks(&rcu_ctrlblk, &__get_cpu_var(rcu_data));
 421         __rcu_process_callbacks(&rcu_bh_ctrlblk, &__get_cpu_var(rcu_bh_data));
 422 }
 423
 424 static int __rcu_pending(struct rcu_ctrlblk *rcp, struct rcu_data *rdp)
 425 {
 426         /* This cpu has pending rcu entries and the grace period
 427          * for them has completed.
 428          */
 429         if (rdp->curlist && !rcu_batch_before(rcp->completed, rdp->batch))
 430                 return 1;
 431
 432         /* This cpu has no pending entries, but there are new entries */
 433         if (!rdp->curlist && rdp->nxtlist)
 434                 return 1;
 435
 436         /* This cpu has finished callbacks to invoke */
 437         if (rdp->donelist)
 438                 return 1;
 439
 440         /* The rcu core waits for a quiescent state from the cpu */
 441         if (rdp->quiescbatch != rcp->cur || rdp->qs_pending)
 442                 return 1;
 443
 444         /* nothing to do */
 445         return 0;
 446 }
 447
 448 int rcu_pending(int cpu)
 449 {
 450         return __rcu_pending(&rcu_ctrlblk, &per_cpu(rcu_data, cpu)) ||
 451                 __rcu_pending(&rcu_bh_ctrlblk, &per_cpu(rcu_bh_data, cpu));
 452 }
 453
 454 void rcu_check_callbacks(int cpu, int user)
 455 {
 456         if (user ||
 457             (idle_cpu(cpu) && !in_softirq() &&
 458                                 hardirq_count() <= (1 << HARDIRQ_SHIFT))) {
 459                 rcu_qsctr_inc(cpu);
 460                 rcu_bh_qsctr_inc(cpu);
 461         } else if (!in_softirq())
 462                 rcu_bh_qsctr_inc(cpu);
 463         tasklet_schedule(&per_cpu(rcu_tasklet, cpu));
 464 }
 465
 466 static void rcu_init_percpu_data(int cpu, struct rcu_ctrlblk *rcp,
 467                                                 struct rcu_data *rdp)
 468 {
 469         memset(rdp, 0, sizeof(*rdp));
 470         rdp->curtail = &rdp->curlist;
 471         rdp->nxttail = &rdp->nxtlist;
 472         rdp->donetail = &rdp->donelist;
 473         rdp->quiescbatch = rcp->completed;
 474         rdp->qs_pending = 0;
 475         rdp->cpu = cpu;
 476 }
 477
 478 static void __devinit rcu_online_cpu(int cpu)
 479 {
 480         struct rcu_data *rdp = &per_cpu(rcu_data, cpu);
 481         struct rcu_data *bh_rdp = &per_cpu(rcu_bh_data, cpu);
 482
 483         rcu_init_percpu_data(cpu, &rcu_ctrlblk, rdp);
 484         rcu_init_percpu_data(cpu, &rcu_bh_ctrlblk, bh_rdp);
 485         tasklet_init(&per_cpu(rcu_tasklet, cpu), rcu_process_callbacks, 0UL);
 486 }
 487
 488 static int __devinit rcu_cpu_notify(struct notifier_block *self,
 489                                 unsigned long action, void *hcpu)
 490 {
 491         long cpu = (long)hcpu;
 492         switch (action) {
 493         case CPU_UP_PREPARE:
 494                 rcu_online_cpu(cpu);
 495                 break;
 496         case CPU_DEAD:
 497                 rcu_offline_cpu(cpu);
 498                 break;
 499         default:
 500                 break;
 501         }
 502         return NOTIFY_OK;
 503 }
 504
 505 static struct notifier_block __devinitdata rcu_nb = {
 506         .notifier_call  = rcu_cpu_notify,
 507 };
 508
 509 /*
 510  * Initializes rcu mechanism.  Assumed to be called early.
 511  * That is before local timer(SMP) or jiffie timer (uniproc) is setup.
 512  * Note that rcu_qsctr and friends are implicitly
 513  * initialized due to the choice of ``0'' for RCU_CTR_INVALID.
 514  */
 515 void __init rcu_init(void)
 516 {
 517         sema_init(&rcu_barrier_sema, 1);
 518         rcu_cpu_notify(&rcu_nb, CPU_UP_PREPARE,
 519                         (void *)(long)smp_processor_id());
 520         /* Register notifier for non-boot CPUs */
 521         register_cpu_notifier(&rcu_nb);
 522 }
 523
 524 struct rcu_synchronize {
 525         struct rcu_head head;
 526         struct completion completion;
 527 };
 528
 529 /* Because of FASTCALL declaration of complete, we use this wrapper */
 530 static void wakeme_after_rcu(struct rcu_head  *head)
 531 {
 532         struct rcu_synchronize *rcu;
 533
 534         rcu = container_of(head, struct rcu_synchronize, head);
 535         complete(&rcu->completion);
 536 }
 537
 538 /**
 539  * synchronize_rcu - wait until a grace period has elapsed.
 540  *
 541  * Control will return to the caller some time after a full grace
 542  * period has elapsed, in other words after all currently executing RCU
 543  * read-side critical sections have completed.  RCU read-side critical
 544  * sections are delimited by rcu_read_lock() and rcu_read_unlock(),
 545  * and may be nested.
 546  *
 547  * If your read-side code is not protected by rcu_read_lock(), do -not-
 548  * use synchronize_rcu().
 549  */
 550 void synchronize_rcu(void)
 551 {
 552         struct rcu_synchronize rcu;
 553
 554         init_completion(&rcu.completion);
 555         /* Will wake me after RCU finished */
 556         call_rcu(&rcu.head, wakeme_after_rcu);
 557
 558         /* Wait for it */
 559         wait_for_completion(&rcu.completion);
 560 }
 561
 562 /*
 563  * Deprecated, use synchronize_rcu() or synchronize_sched() instead.
 564  */
 565 void synchronize_kernel(void)
 566 {
 567         synchronize_rcu();
 568 }
 569
 570 module_param(maxbatch, int, 0);
 571 EXPORT_SYMBOL_GPL(rcu_batches_completed);
 572 EXPORT_SYMBOL(call_rcu);  /* WARNING: GPL-only in April 2006. */
 573 EXPORT_SYMBOL(call_rcu_bh);  /* WARNING: GPL-only in April 2006. */
 574 EXPORT_SYMBOL_GPL(synchronize_rcu);
 575 EXPORT_SYMBOL(synchronize_kernel);  /* WARNING: GPL-only in April 2006. */