kernel/rcuclassic.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 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  *              Documentation/RCU
  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/cpu.h>
  49 #include <linux/mutex.h>
  50
  51 #ifdef CONFIG_DEBUG_LOCK_ALLOC
  52 static struct lock_class_key rcu_lock_key;
  53 struct lockdep_map rcu_lock_map =
  54         STATIC_LOCKDEP_MAP_INIT("rcu_read_lock", &rcu_lock_key);
  55 EXPORT_SYMBOL_GPL(rcu_lock_map);
  56 #endif
  57
  58
  59 /* Definition for rcupdate control block. */
  60 static struct rcu_ctrlblk rcu_ctrlblk = {
  61         .cur = -300,
  62         .completed = -300,
  63         .lock = __SPIN_LOCK_UNLOCKED(&rcu_ctrlblk.lock),
  64         .cpumask = CPU_MASK_NONE,
  65 };
  66 static struct rcu_ctrlblk rcu_bh_ctrlblk = {
  67         .cur = -300,
  68         .completed = -300,
  69         .lock = __SPIN_LOCK_UNLOCKED(&rcu_bh_ctrlblk.lock),
  70         .cpumask = CPU_MASK_NONE,
  71 };
  72
  73 DEFINE_PER_CPU(struct rcu_data, rcu_data) = { 0L };
  74 DEFINE_PER_CPU(struct rcu_data, rcu_bh_data) = { 0L };
  75
  76 static int blimit = 10;
  77 static int qhimark = 10000;
  78 static int qlowmark = 100;
  79
  80 #ifdef CONFIG_SMP
  81 static void force_quiescent_state(struct rcu_data *rdp,
  82                         struct rcu_ctrlblk *rcp)
  83 {
  84         int cpu;
  85         cpumask_t cpumask;
  86         set_need_resched();
  87         if (unlikely(!rcp->signaled)) {
  88                 rcp->signaled = 1;
  89                 /*
  90                  * Don't send IPI to itself. With irqs disabled,
  91                  * rdp->cpu is the current cpu.
  92                  */
  93                 cpumask = rcp->cpumask;
  94                 cpu_clear(rdp->cpu, cpumask);
  95                 for_each_cpu_mask(cpu, cpumask)
  96                         smp_send_reschedule(cpu);
  97         }
  98 }
  99 #else
 100 static inline void force_quiescent_state(struct rcu_data *rdp,
 101                         struct rcu_ctrlblk *rcp)
 102 {
 103         set_need_resched();
 104 }
 105 #endif
 106
 107 /**
 108  * call_rcu - Queue an RCU callback for invocation after a grace period.
 109  * @head: structure to be used for queueing the RCU updates.
 110  * @func: actual update function to be invoked after the grace period
 111  *
 112  * The update function will be invoked some time after a full grace
 113  * period elapses, in other words after all currently executing RCU
 114  * read-side critical sections have completed.  RCU read-side critical
 115  * sections are delimited by rcu_read_lock() and rcu_read_unlock(),
 116  * and may be nested.
 117  */
 118 void call_rcu(struct rcu_head *head,
 119                                 void (*func)(struct rcu_head *rcu))
 120 {
 121         unsigned long flags;
 122         struct rcu_data *rdp;
 123
 124         head->func = func;
 125         head->next = NULL;
 126         local_irq_save(flags);
 127         rdp = &__get_cpu_var(rcu_data);
 128         *rdp->nxttail = head;
 129         rdp->nxttail = &head->next;
 130         if (unlikely(++rdp->qlen > qhimark)) {
 131                 rdp->blimit = INT_MAX;
 132                 force_quiescent_state(rdp, &rcu_ctrlblk);
 133         }
 134         local_irq_restore(flags);
 135 }
 136 EXPORT_SYMBOL_GPL(call_rcu);
 137
 138 /**
 139  * call_rcu_bh - Queue an RCU for invocation after a quicker grace period.
 140  * @head: structure to be used for queueing the RCU updates.
 141  * @func: actual update function to be invoked after the grace period
 142  *
 143  * The update function will be invoked some time after a full grace
 144  * period elapses, in other words after all currently executing RCU
 145  * read-side critical sections have completed. call_rcu_bh() assumes
 146  * that the read-side critical sections end on completion of a softirq
 147  * handler. This means that read-side critical sections in process
 148  * context must not be interrupted by softirqs. This interface is to be
 149  * used when most of the read-side critical sections are in softirq context.
 150  * RCU read-side critical sections are delimited by rcu_read_lock() and
 151  * rcu_read_unlock(), * if in interrupt context or rcu_read_lock_bh()
 152  * and rcu_read_unlock_bh(), if in process context. These may be nested.
 153  */
 154 void call_rcu_bh(struct rcu_head *head,
 155                                 void (*func)(struct rcu_head *rcu))
 156 {
 157         unsigned long flags;
 158         struct rcu_data *rdp;
 159
 160         head->func = func;
 161         head->next = NULL;
 162         local_irq_save(flags);
 163         rdp = &__get_cpu_var(rcu_bh_data);
 164         *rdp->nxttail = head;
 165         rdp->nxttail = &head->next;
 166
 167         if (unlikely(++rdp->qlen > qhimark)) {
 168                 rdp->blimit = INT_MAX;
 169                 force_quiescent_state(rdp, &rcu_bh_ctrlblk);
 170         }
 171
 172         local_irq_restore(flags);
 173 }
 174 EXPORT_SYMBOL_GPL(call_rcu_bh);
 175
 176 /*
 177  * Return the number of RCU batches processed thus far.  Useful
 178  * for debug and statistics.
 179  */
 180 long rcu_batches_completed(void)
 181 {
 182         return rcu_ctrlblk.completed;
 183 }
 184 EXPORT_SYMBOL_GPL(rcu_batches_completed);
 185
 186 /*
 187  * Return the number of RCU batches processed thus far.  Useful
 188  * for debug and statistics.
 189  */
 190 long rcu_batches_completed_bh(void)
 191 {
 192         return rcu_bh_ctrlblk.completed;
 193 }
 194 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
 195
 196 /* Raises the softirq for processing rcu_callbacks. */
 197 static inline void raise_rcu_softirq(void)
 198 {
 199         raise_softirq(RCU_SOFTIRQ);
 200         /*
 201          * The smp_mb() here is required to ensure that this cpu's
 202          * __rcu_process_callbacks() reads the most recently updated
 203          * value of rcu->cur.
 204          */
 205         smp_mb();
 206 }
 207
 208 /*
 209  * Invoke the completed RCU callbacks. They are expected to be in
 210  * a per-cpu list.
 211  */
 212 static void rcu_do_batch(struct rcu_data *rdp)
 213 {
 214         struct rcu_head *next, *list;
 215         int count = 0;
 216
 217         list = rdp->donelist;
 218         while (list) {
 219                 next = list->next;
 220                 prefetch(next);
 221                 list->func(list);
 222                 list = next;
 223                 if (++count >= rdp->blimit)
 224                         break;
 225         }
 226         rdp->donelist = list;
 227
 228         local_irq_disable();
 229         rdp->qlen -= count;
 230         local_irq_enable();
 231         if (rdp->blimit == INT_MAX && rdp->qlen <= qlowmark)
 232                 rdp->blimit = blimit;
 233
 234         if (!rdp->donelist)
 235                 rdp->donetail = &rdp->donelist;
 236         else
 237                 raise_rcu_softirq();
 238 }
 239
 240 /*
 241  * Grace period handling:
 242  * The grace period handling consists out of two steps:
 243  * - A new grace period is started.
 244  *   This is done by rcu_start_batch. The start is not broadcasted to
 245  *   all cpus, they must pick this up by comparing rcp->cur with
 246  *   rdp->quiescbatch. All cpus are recorded  in the
 247  *   rcu_ctrlblk.cpumask bitmap.
 248  * - All cpus must go through a quiescent state.
 249  *   Since the start of the grace period is not broadcasted, at least two
 250  *   calls to rcu_check_quiescent_state are required:
 251  *   The first call just notices that a new grace period is running. The
 252  *   following calls check if there was a quiescent state since the beginning
 253  *   of the grace period. If so, it updates rcu_ctrlblk.cpumask. If
 254  *   the bitmap is empty, then the grace period is completed.
 255  *   rcu_check_quiescent_state calls rcu_start_batch(0) to start the next grace
 256  *   period (if necessary).
 257  */
 258 /*
 259  * Register a new batch of callbacks, and start it up if there is currently no
 260  * active batch and the batch to be registered has not already occurred.
 261  * Caller must hold rcu_ctrlblk.lock.
 262  */
 263 static void rcu_start_batch(struct rcu_ctrlblk *rcp)
 264 {
 265         if (rcp->next_pending &&
 266                         rcp->completed == rcp->cur) {
 267                 rcp->next_pending = 0;
 268                 /*
 269                  * next_pending == 0 must be visible in
 270                  * __rcu_process_callbacks() before it can see new value of cur.
 271                  */
 272                 smp_wmb();
 273                 rcp->cur++;
 274
 275                 /*
 276                  * Accessing nohz_cpu_mask before incrementing rcp->cur needs a
 277                  * Barrier  Otherwise it can cause tickless idle CPUs to be
 278                  * included in rcp->cpumask, which will extend graceperiods
 279                  * unnecessarily.
 280                  */
 281                 smp_mb();
 282                 cpus_andnot(rcp->cpumask, cpu_online_map, nohz_cpu_mask);
 283
 284                 rcp->signaled = 0;
 285         }
 286 }
 287
 288 /*
 289  * cpu went through a quiescent state since the beginning of the grace period.
 290  * Clear it from the cpu mask and complete the grace period if it was the last
 291  * cpu. Start another grace period if someone has further entries pending
 292  */
 293 static void cpu_quiet(int cpu, struct rcu_ctrlblk *rcp)
 294 {
 295         cpu_clear(cpu, rcp->cpumask);
 296         if (cpus_empty(rcp->cpumask)) {
 297                 /* batch completed ! */
 298                 rcp->completed = rcp->cur;
 299                 rcu_start_batch(rcp);
 300         }
 301 }
 302
 303 /*
 304  * Check if the cpu has gone through a quiescent state (say context
 305  * switch). If so and if it already hasn't done so in this RCU
 306  * quiescent cycle, then indicate that it has done so.
 307  */
 308 static void rcu_check_quiescent_state(struct rcu_ctrlblk *rcp,
 309                                         struct rcu_data *rdp)
 310 {
 311         if (rdp->quiescbatch != rcp->cur) {
 312                 /* start new grace period: */
 313                 rdp->qs_pending = 1;
 314                 rdp->passed_quiesc = 0;
 315                 rdp->quiescbatch = rcp->cur;
 316                 return;
 317         }
 318
 319         /* Grace period already completed for this cpu?
 320          * qs_pending is checked instead of the actual bitmap to avoid
 321          * cacheline trashing.
 322          */
 323         if (!rdp->qs_pending)
 324                 return;
 325
 326         /*
 327          * Was there a quiescent state since the beginning of the grace
 328          * period? If no, then exit and wait for the next call.
 329          */
 330         if (!rdp->passed_quiesc)
 331                 return;
 332         rdp->qs_pending = 0;
 333
 334         spin_lock(&rcp->lock);
 335         /*
 336          * rdp->quiescbatch/rcp->cur and the cpu bitmap can come out of sync
 337          * during cpu startup. Ignore the quiescent state.
 338          */
 339         if (likely(rdp->quiescbatch == rcp->cur))
 340                 cpu_quiet(rdp->cpu, rcp);
 341
 342         spin_unlock(&rcp->lock);
 343 }
 344
 345
 346 #ifdef CONFIG_HOTPLUG_CPU
 347
 348 /* warning! helper for rcu_offline_cpu. do not use elsewhere without reviewing
 349  * locking requirements, the list it's pulling from has to belong to a cpu
 350  * which is dead and hence not processing interrupts.
 351  */
 352 static void rcu_move_batch(struct rcu_data *this_rdp, struct rcu_head *list,
 353                                 struct rcu_head **tail)
 354 {
 355         local_irq_disable();
 356         *this_rdp->nxttail = list;
 357         if (list)
 358                 this_rdp->nxttail = tail;
 359         local_irq_enable();
 360 }
 361
 362 static void __rcu_offline_cpu(struct rcu_data *this_rdp,
 363                                 struct rcu_ctrlblk *rcp, struct rcu_data *rdp)
 364 {
 365         /* if the cpu going offline owns the grace period
 366          * we can block indefinitely waiting for it, so flush
 367          * it here
 368          */
 369         spin_lock_bh(&rcp->lock);
 370         if (rcp->cur != rcp->completed)
 371                 cpu_quiet(rdp->cpu, rcp);
 372         spin_unlock_bh(&rcp->lock);
 373         rcu_move_batch(this_rdp, rdp->donelist, rdp->donetail);
 374         rcu_move_batch(this_rdp, rdp->curlist, rdp->curtail);
 375         rcu_move_batch(this_rdp, rdp->nxtlist, rdp->nxttail);
 376 }
 377
 378 static void rcu_offline_cpu(int cpu)
 379 {
 380         struct rcu_data *this_rdp = &get_cpu_var(rcu_data);
 381         struct rcu_data *this_bh_rdp = &get_cpu_var(rcu_bh_data);
 382
 383         __rcu_offline_cpu(this_rdp, &rcu_ctrlblk,
 384                                         &per_cpu(rcu_data, cpu));
 385         __rcu_offline_cpu(this_bh_rdp, &rcu_bh_ctrlblk,
 386                                         &per_cpu(rcu_bh_data, cpu));
 387         put_cpu_var(rcu_data);
 388         put_cpu_var(rcu_bh_data);
 389 }
 390
 391 #else
 392
 393 static void rcu_offline_cpu(int cpu)
 394 {
 395 }
 396
 397 #endif
 398
 399 /*
 400  * This does the RCU processing work from softirq context.
 401  */
 402 static void __rcu_process_callbacks(struct rcu_ctrlblk *rcp,
 403                                         struct rcu_data *rdp)
 404 {
 405         if (rdp->curlist && !rcu_batch_before(rcp->completed, rdp->batch)) {
 406                 *rdp->donetail = rdp->curlist;
 407                 rdp->donetail = rdp->curtail;
 408                 rdp->curlist = NULL;
 409                 rdp->curtail = &rdp->curlist;
 410         }
 411
 412         if (rdp->nxtlist && !rdp->curlist) {
 413                 local_irq_disable();
 414                 rdp->curlist = rdp->nxtlist;
 415                 rdp->curtail = rdp->nxttail;
 416                 rdp->nxtlist = NULL;
 417                 rdp->nxttail = &rdp->nxtlist;
 418                 local_irq_enable();
 419
 420                 /*
 421                  * start the next batch of callbacks
 422                  */
 423
 424                 /* determine batch number */
 425                 rdp->batch = rcp->cur + 1;
 426                 /* see the comment and corresponding wmb() in
 427                  * the rcu_start_batch()
 428                  */
 429                 smp_rmb();
 430
 431                 if (!rcp->next_pending) {
 432                         /* and start it/schedule start if it's a new batch */
 433                         spin_lock(&rcp->lock);
 434                         rcp->next_pending = 1;
 435                         rcu_start_batch(rcp);
 436                         spin_unlock(&rcp->lock);
 437                 }
 438         }
 439
 440         rcu_check_quiescent_state(rcp, rdp);
 441         if (rdp->donelist)
 442                 rcu_do_batch(rdp);
 443 }
 444
 445 static void rcu_process_callbacks(struct softirq_action *unused)
 446 {
 447         __rcu_process_callbacks(&rcu_ctrlblk, &__get_cpu_var(rcu_data));
 448         __rcu_process_callbacks(&rcu_bh_ctrlblk, &__get_cpu_var(rcu_bh_data));
 449 }
 450
 451 static int __rcu_pending(struct rcu_ctrlblk *rcp, struct rcu_data *rdp)
 452 {
 453         /* This cpu has pending rcu entries and the grace period
 454          * for them has completed.
 455          */
 456         if (rdp->curlist && !rcu_batch_before(rcp->completed, rdp->batch))
 457                 return 1;
 458
 459         /* This cpu has no pending entries, but there are new entries */
 460         if (!rdp->curlist && rdp->nxtlist)
 461                 return 1;
 462
 463         /* This cpu has finished callbacks to invoke */
 464         if (rdp->donelist)
 465                 return 1;
 466
 467         /* The rcu core waits for a quiescent state from the cpu */
 468         if (rdp->quiescbatch != rcp->cur || rdp->qs_pending)
 469                 return 1;
 470
 471         /* nothing to do */
 472         return 0;
 473 }
 474
 475 /*
 476  * Check to see if there is any immediate RCU-related work to be done
 477  * by the current CPU, returning 1 if so.  This function is part of the
 478  * RCU implementation; it is -not- an exported member of the RCU API.
 479  */
 480 int rcu_pending(int cpu)
 481 {
 482         return __rcu_pending(&rcu_ctrlblk, &per_cpu(rcu_data, cpu)) ||
 483                 __rcu_pending(&rcu_bh_ctrlblk, &per_cpu(rcu_bh_data, cpu));
 484 }
 485
 486 /*
 487  * Check to see if any future RCU-related work will need to be done
 488  * by the current CPU, even if none need be done immediately, returning
 489  * 1 if so.  This function is part of the RCU implementation; it is -not-
 490  * an exported member of the RCU API.
 491  */
 492 int rcu_needs_cpu(int cpu)
 493 {
 494         struct rcu_data *rdp = &per_cpu(rcu_data, cpu);
 495         struct rcu_data *rdp_bh = &per_cpu(rcu_bh_data, cpu);
 496
 497         return (!!rdp->curlist || !!rdp_bh->curlist || rcu_pending(cpu));
 498 }
 499
 500 void rcu_check_callbacks(int cpu, int user)
 501 {
 502         if (user ||
 503             (idle_cpu(cpu) && !in_softirq() &&
 504                                 hardirq_count() <= (1 << HARDIRQ_SHIFT))) {
 505                 rcu_qsctr_inc(cpu);
 506                 rcu_bh_qsctr_inc(cpu);
 507         } else if (!in_softirq())
 508                 rcu_bh_qsctr_inc(cpu);
 509         raise_rcu_softirq();
 510 }
 511
 512 static void rcu_init_percpu_data(int cpu, struct rcu_ctrlblk *rcp,
 513                                                 struct rcu_data *rdp)
 514 {
 515         memset(rdp, 0, sizeof(*rdp));
 516         rdp->curtail = &rdp->curlist;
 517         rdp->nxttail = &rdp->nxtlist;
 518         rdp->donetail = &rdp->donelist;
 519         rdp->quiescbatch = rcp->completed;
 520         rdp->qs_pending = 0;
 521         rdp->cpu = cpu;
 522         rdp->blimit = blimit;
 523 }
 524
 525 static void __cpuinit rcu_online_cpu(int cpu)
 526 {
 527         struct rcu_data *rdp = &per_cpu(rcu_data, cpu);
 528         struct rcu_data *bh_rdp = &per_cpu(rcu_bh_data, cpu);
 529
 530         rcu_init_percpu_data(cpu, &rcu_ctrlblk, rdp);
 531         rcu_init_percpu_data(cpu, &rcu_bh_ctrlblk, bh_rdp);
 532         open_softirq(RCU_SOFTIRQ, rcu_process_callbacks, NULL);
 533 }
 534
 535 static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
 536                                 unsigned long action, void *hcpu)
 537 {
 538         long cpu = (long)hcpu;
 539
 540         switch (action) {
 541         case CPU_UP_PREPARE:
 542         case CPU_UP_PREPARE_FROZEN:
 543                 rcu_online_cpu(cpu);
 544                 break;
 545         case CPU_DEAD:
 546         case CPU_DEAD_FROZEN:
 547                 rcu_offline_cpu(cpu);
 548                 break;
 549         default:
 550                 break;
 551         }
 552         return NOTIFY_OK;
 553 }
 554
 555 static struct notifier_block __cpuinitdata rcu_nb = {
 556         .notifier_call  = rcu_cpu_notify,
 557 };
 558
 559 /*
 560  * Initializes rcu mechanism.  Assumed to be called early.
 561  * That is before local timer(SMP) or jiffie timer (uniproc) is setup.
 562  * Note that rcu_qsctr and friends are implicitly
 563  * initialized due to the choice of ``0'' for RCU_CTR_INVALID.
 564  */
 565 void __init __rcu_init(void)
 566 {
 567         rcu_cpu_notify(&rcu_nb, CPU_UP_PREPARE,
 568                         (void *)(long)smp_processor_id());
 569         /* Register notifier for non-boot CPUs */
 570         register_cpu_notifier(&rcu_nb);
 571 }
 572
 573 module_param(blimit, int, 0);
 574 module_param(qhimark, int, 0);
 575 module_param(qlowmark, int, 0);