| 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, 2008 |
| 19 | * |
| 20 | * Authors: Dipankar Sarma <dipankar@in.ibm.com> |
| 21 | * Manfred Spraul <manfred@colorfullife.com> |
| 22 | * Paul E. McKenney <paulmck@linux.vnet.ibm.com> Hierarchical version |
| 23 | * |
| 24 | * Based on the original work by Paul McKenney <paulmck@us.ibm.com> |
| 25 | * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen. |
| 26 | * |
| 27 | * For detailed explanation of Read-Copy Update mechanism see - |
| 28 | * Documentation/RCU |
| 29 | */ |
| 30 | #include <linux/types.h> |
| 31 | #include <linux/kernel.h> |
| 32 | #include <linux/init.h> |
| 33 | #include <linux/spinlock.h> |
| 34 | #include <linux/smp.h> |
| 35 | #include <linux/rcupdate.h> |
| 36 | #include <linux/interrupt.h> |
| 37 | #include <linux/sched.h> |
| 38 | #include <linux/nmi.h> |
| 39 | #include <asm/atomic.h> |
| 40 | #include <linux/bitops.h> |
| 41 | #include <linux/module.h> |
| 42 | #include <linux/completion.h> |
| 43 | #include <linux/moduleparam.h> |
| 44 | #include <linux/percpu.h> |
| 45 | #include <linux/notifier.h> |
| 46 | #include <linux/cpu.h> |
| 47 | #include <linux/mutex.h> |
| 48 | #include <linux/time.h> |
| 49 | |
| 50 | #include "rcutree.h" |
| 51 | |
| 52 | /* Data structures. */ |
| 53 | |
| 54 | #define RCU_STATE_INITIALIZER(name) { \ |
| 55 | .level = { &name.node[0] }, \ |
| 56 | .levelcnt = { \ |
| 57 | NUM_RCU_LVL_0, /* root of hierarchy. */ \ |
| 58 | NUM_RCU_LVL_1, \ |
| 59 | NUM_RCU_LVL_2, \ |
| 60 | NUM_RCU_LVL_3, /* == MAX_RCU_LVLS */ \ |
| 61 | }, \ |
| 62 | .signaled = RCU_SIGNAL_INIT, \ |
| 63 | .gpnum = -300, \ |
| 64 | .completed = -300, \ |
| 65 | .onofflock = __SPIN_LOCK_UNLOCKED(&name.onofflock), \ |
| 66 | .fqslock = __SPIN_LOCK_UNLOCKED(&name.fqslock), \ |
| 67 | .n_force_qs = 0, \ |
| 68 | .n_force_qs_ngp = 0, \ |
| 69 | } |
| 70 | |
| 71 | struct rcu_state rcu_sched_state = RCU_STATE_INITIALIZER(rcu_sched_state); |
| 72 | DEFINE_PER_CPU(struct rcu_data, rcu_sched_data); |
| 73 | |
| 74 | struct rcu_state rcu_bh_state = RCU_STATE_INITIALIZER(rcu_bh_state); |
| 75 | DEFINE_PER_CPU(struct rcu_data, rcu_bh_data); |
| 76 | |
| 77 | |
| 78 | /* |
| 79 | * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s |
| 80 | * permit this function to be invoked without holding the root rcu_node |
| 81 | * structure's ->lock, but of course results can be subject to change. |
| 82 | */ |
| 83 | static int rcu_gp_in_progress(struct rcu_state *rsp) |
| 84 | { |
| 85 | return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum); |
| 86 | } |
| 87 | |
| 88 | /* |
| 89 | * Note a quiescent state. Because we do not need to know |
| 90 | * how many quiescent states passed, just if there was at least |
| 91 | * one since the start of the grace period, this just sets a flag. |
| 92 | */ |
| 93 | void rcu_sched_qs(int cpu) |
| 94 | { |
| 95 | struct rcu_data *rdp; |
| 96 | |
| 97 | rdp = &per_cpu(rcu_sched_data, cpu); |
| 98 | rdp->passed_quiesc_completed = rdp->completed; |
| 99 | barrier(); |
| 100 | rdp->passed_quiesc = 1; |
| 101 | rcu_preempt_note_context_switch(cpu); |
| 102 | } |
| 103 | |
| 104 | void rcu_bh_qs(int cpu) |
| 105 | { |
| 106 | struct rcu_data *rdp; |
| 107 | |
| 108 | rdp = &per_cpu(rcu_bh_data, cpu); |
| 109 | rdp->passed_quiesc_completed = rdp->completed; |
| 110 | barrier(); |
| 111 | rdp->passed_quiesc = 1; |
| 112 | } |
| 113 | |
| 114 | #ifdef CONFIG_NO_HZ |
| 115 | DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = { |
| 116 | .dynticks_nesting = 1, |
| 117 | .dynticks = 1, |
| 118 | }; |
| 119 | #endif /* #ifdef CONFIG_NO_HZ */ |
| 120 | |
| 121 | static int blimit = 10; /* Maximum callbacks per softirq. */ |
| 122 | static int qhimark = 10000; /* If this many pending, ignore blimit. */ |
| 123 | static int qlowmark = 100; /* Once only this many pending, use blimit. */ |
| 124 | |
| 125 | module_param(blimit, int, 0); |
| 126 | module_param(qhimark, int, 0); |
| 127 | module_param(qlowmark, int, 0); |
| 128 | |
| 129 | static void force_quiescent_state(struct rcu_state *rsp, int relaxed); |
| 130 | static int rcu_pending(int cpu); |
| 131 | |
| 132 | /* |
| 133 | * Return the number of RCU-sched batches processed thus far for debug & stats. |
| 134 | */ |
| 135 | long rcu_batches_completed_sched(void) |
| 136 | { |
| 137 | return rcu_sched_state.completed; |
| 138 | } |
| 139 | EXPORT_SYMBOL_GPL(rcu_batches_completed_sched); |
| 140 | |
| 141 | /* |
| 142 | * Return the number of RCU BH batches processed thus far for debug & stats. |
| 143 | */ |
| 144 | long rcu_batches_completed_bh(void) |
| 145 | { |
| 146 | return rcu_bh_state.completed; |
| 147 | } |
| 148 | EXPORT_SYMBOL_GPL(rcu_batches_completed_bh); |
| 149 | |
| 150 | /* |
| 151 | * Does the CPU have callbacks ready to be invoked? |
| 152 | */ |
| 153 | static int |
| 154 | cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp) |
| 155 | { |
| 156 | return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL]; |
| 157 | } |
| 158 | |
| 159 | /* |
| 160 | * Does the current CPU require a yet-as-unscheduled grace period? |
| 161 | */ |
| 162 | static int |
| 163 | cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp) |
| 164 | { |
| 165 | return *rdp->nxttail[RCU_DONE_TAIL] && !rcu_gp_in_progress(rsp); |
| 166 | } |
| 167 | |
| 168 | /* |
| 169 | * Return the root node of the specified rcu_state structure. |
| 170 | */ |
| 171 | static struct rcu_node *rcu_get_root(struct rcu_state *rsp) |
| 172 | { |
| 173 | return &rsp->node[0]; |
| 174 | } |
| 175 | |
| 176 | #ifdef CONFIG_SMP |
| 177 | |
| 178 | /* |
| 179 | * If the specified CPU is offline, tell the caller that it is in |
| 180 | * a quiescent state. Otherwise, whack it with a reschedule IPI. |
| 181 | * Grace periods can end up waiting on an offline CPU when that |
| 182 | * CPU is in the process of coming online -- it will be added to the |
| 183 | * rcu_node bitmasks before it actually makes it online. The same thing |
| 184 | * can happen while a CPU is in the process of coming online. Because this |
| 185 | * race is quite rare, we check for it after detecting that the grace |
| 186 | * period has been delayed rather than checking each and every CPU |
| 187 | * each and every time we start a new grace period. |
| 188 | */ |
| 189 | static int rcu_implicit_offline_qs(struct rcu_data *rdp) |
| 190 | { |
| 191 | /* |
| 192 | * If the CPU is offline, it is in a quiescent state. We can |
| 193 | * trust its state not to change because interrupts are disabled. |
| 194 | */ |
| 195 | if (cpu_is_offline(rdp->cpu)) { |
| 196 | rdp->offline_fqs++; |
| 197 | return 1; |
| 198 | } |
| 199 | |
| 200 | /* If preemptable RCU, no point in sending reschedule IPI. */ |
| 201 | if (rdp->preemptable) |
| 202 | return 0; |
| 203 | |
| 204 | /* The CPU is online, so send it a reschedule IPI. */ |
| 205 | if (rdp->cpu != smp_processor_id()) |
| 206 | smp_send_reschedule(rdp->cpu); |
| 207 | else |
| 208 | set_need_resched(); |
| 209 | rdp->resched_ipi++; |
| 210 | return 0; |
| 211 | } |
| 212 | |
| 213 | #endif /* #ifdef CONFIG_SMP */ |
| 214 | |
| 215 | #ifdef CONFIG_NO_HZ |
| 216 | |
| 217 | /** |
| 218 | * rcu_enter_nohz - inform RCU that current CPU is entering nohz |
| 219 | * |
| 220 | * Enter nohz mode, in other words, -leave- the mode in which RCU |
| 221 | * read-side critical sections can occur. (Though RCU read-side |
| 222 | * critical sections can occur in irq handlers in nohz mode, a possibility |
| 223 | * handled by rcu_irq_enter() and rcu_irq_exit()). |
| 224 | */ |
| 225 | void rcu_enter_nohz(void) |
| 226 | { |
| 227 | unsigned long flags; |
| 228 | struct rcu_dynticks *rdtp; |
| 229 | |
| 230 | smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */ |
| 231 | local_irq_save(flags); |
| 232 | rdtp = &__get_cpu_var(rcu_dynticks); |
| 233 | rdtp->dynticks++; |
| 234 | rdtp->dynticks_nesting--; |
| 235 | WARN_ON_ONCE(rdtp->dynticks & 0x1); |
| 236 | local_irq_restore(flags); |
| 237 | } |
| 238 | |
| 239 | /* |
| 240 | * rcu_exit_nohz - inform RCU that current CPU is leaving nohz |
| 241 | * |
| 242 | * Exit nohz mode, in other words, -enter- the mode in which RCU |
| 243 | * read-side critical sections normally occur. |
| 244 | */ |
| 245 | void rcu_exit_nohz(void) |
| 246 | { |
| 247 | unsigned long flags; |
| 248 | struct rcu_dynticks *rdtp; |
| 249 | |
| 250 | local_irq_save(flags); |
| 251 | rdtp = &__get_cpu_var(rcu_dynticks); |
| 252 | rdtp->dynticks++; |
| 253 | rdtp->dynticks_nesting++; |
| 254 | WARN_ON_ONCE(!(rdtp->dynticks & 0x1)); |
| 255 | local_irq_restore(flags); |
| 256 | smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */ |
| 257 | } |
| 258 | |
| 259 | /** |
| 260 | * rcu_nmi_enter - inform RCU of entry to NMI context |
| 261 | * |
| 262 | * If the CPU was idle with dynamic ticks active, and there is no |
| 263 | * irq handler running, this updates rdtp->dynticks_nmi to let the |
| 264 | * RCU grace-period handling know that the CPU is active. |
| 265 | */ |
| 266 | void rcu_nmi_enter(void) |
| 267 | { |
| 268 | struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks); |
| 269 | |
| 270 | if (rdtp->dynticks & 0x1) |
| 271 | return; |
| 272 | rdtp->dynticks_nmi++; |
| 273 | WARN_ON_ONCE(!(rdtp->dynticks_nmi & 0x1)); |
| 274 | smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */ |
| 275 | } |
| 276 | |
| 277 | /** |
| 278 | * rcu_nmi_exit - inform RCU of exit from NMI context |
| 279 | * |
| 280 | * If the CPU was idle with dynamic ticks active, and there is no |
| 281 | * irq handler running, this updates rdtp->dynticks_nmi to let the |
| 282 | * RCU grace-period handling know that the CPU is no longer active. |
| 283 | */ |
| 284 | void rcu_nmi_exit(void) |
| 285 | { |
| 286 | struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks); |
| 287 | |
| 288 | if (rdtp->dynticks & 0x1) |
| 289 | return; |
| 290 | smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */ |
| 291 | rdtp->dynticks_nmi++; |
| 292 | WARN_ON_ONCE(rdtp->dynticks_nmi & 0x1); |
| 293 | } |
| 294 | |
| 295 | /** |
| 296 | * rcu_irq_enter - inform RCU of entry to hard irq context |
| 297 | * |
| 298 | * If the CPU was idle with dynamic ticks active, this updates the |
| 299 | * rdtp->dynticks to let the RCU handling know that the CPU is active. |
| 300 | */ |
| 301 | void rcu_irq_enter(void) |
| 302 | { |
| 303 | struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks); |
| 304 | |
| 305 | if (rdtp->dynticks_nesting++) |
| 306 | return; |
| 307 | rdtp->dynticks++; |
| 308 | WARN_ON_ONCE(!(rdtp->dynticks & 0x1)); |
| 309 | smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */ |
| 310 | } |
| 311 | |
| 312 | /** |
| 313 | * rcu_irq_exit - inform RCU of exit from hard irq context |
| 314 | * |
| 315 | * If the CPU was idle with dynamic ticks active, update the rdp->dynticks |
| 316 | * to put let the RCU handling be aware that the CPU is going back to idle |
| 317 | * with no ticks. |
| 318 | */ |
| 319 | void rcu_irq_exit(void) |
| 320 | { |
| 321 | struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks); |
| 322 | |
| 323 | if (--rdtp->dynticks_nesting) |
| 324 | return; |
| 325 | smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */ |
| 326 | rdtp->dynticks++; |
| 327 | WARN_ON_ONCE(rdtp->dynticks & 0x1); |
| 328 | |
| 329 | /* If the interrupt queued a callback, get out of dyntick mode. */ |
| 330 | if (__get_cpu_var(rcu_sched_data).nxtlist || |
| 331 | __get_cpu_var(rcu_bh_data).nxtlist) |
| 332 | set_need_resched(); |
| 333 | } |
| 334 | |
| 335 | /* |
| 336 | * Record the specified "completed" value, which is later used to validate |
| 337 | * dynticks counter manipulations. Specify "rsp->completed - 1" to |
| 338 | * unconditionally invalidate any future dynticks manipulations (which is |
| 339 | * useful at the beginning of a grace period). |
| 340 | */ |
| 341 | static void dyntick_record_completed(struct rcu_state *rsp, long comp) |
| 342 | { |
| 343 | rsp->dynticks_completed = comp; |
| 344 | } |
| 345 | |
| 346 | #ifdef CONFIG_SMP |
| 347 | |
| 348 | /* |
| 349 | * Recall the previously recorded value of the completion for dynticks. |
| 350 | */ |
| 351 | static long dyntick_recall_completed(struct rcu_state *rsp) |
| 352 | { |
| 353 | return rsp->dynticks_completed; |
| 354 | } |
| 355 | |
| 356 | /* |
| 357 | * Snapshot the specified CPU's dynticks counter so that we can later |
| 358 | * credit them with an implicit quiescent state. Return 1 if this CPU |
| 359 | * is in dynticks idle mode, which is an extended quiescent state. |
| 360 | */ |
| 361 | static int dyntick_save_progress_counter(struct rcu_data *rdp) |
| 362 | { |
| 363 | int ret; |
| 364 | int snap; |
| 365 | int snap_nmi; |
| 366 | |
| 367 | snap = rdp->dynticks->dynticks; |
| 368 | snap_nmi = rdp->dynticks->dynticks_nmi; |
| 369 | smp_mb(); /* Order sampling of snap with end of grace period. */ |
| 370 | rdp->dynticks_snap = snap; |
| 371 | rdp->dynticks_nmi_snap = snap_nmi; |
| 372 | ret = ((snap & 0x1) == 0) && ((snap_nmi & 0x1) == 0); |
| 373 | if (ret) |
| 374 | rdp->dynticks_fqs++; |
| 375 | return ret; |
| 376 | } |
| 377 | |
| 378 | /* |
| 379 | * Return true if the specified CPU has passed through a quiescent |
| 380 | * state by virtue of being in or having passed through an dynticks |
| 381 | * idle state since the last call to dyntick_save_progress_counter() |
| 382 | * for this same CPU. |
| 383 | */ |
| 384 | static int rcu_implicit_dynticks_qs(struct rcu_data *rdp) |
| 385 | { |
| 386 | long curr; |
| 387 | long curr_nmi; |
| 388 | long snap; |
| 389 | long snap_nmi; |
| 390 | |
| 391 | curr = rdp->dynticks->dynticks; |
| 392 | snap = rdp->dynticks_snap; |
| 393 | curr_nmi = rdp->dynticks->dynticks_nmi; |
| 394 | snap_nmi = rdp->dynticks_nmi_snap; |
| 395 | smp_mb(); /* force ordering with cpu entering/leaving dynticks. */ |
| 396 | |
| 397 | /* |
| 398 | * If the CPU passed through or entered a dynticks idle phase with |
| 399 | * no active irq/NMI handlers, then we can safely pretend that the CPU |
| 400 | * already acknowledged the request to pass through a quiescent |
| 401 | * state. Either way, that CPU cannot possibly be in an RCU |
| 402 | * read-side critical section that started before the beginning |
| 403 | * of the current RCU grace period. |
| 404 | */ |
| 405 | if ((curr != snap || (curr & 0x1) == 0) && |
| 406 | (curr_nmi != snap_nmi || (curr_nmi & 0x1) == 0)) { |
| 407 | rdp->dynticks_fqs++; |
| 408 | return 1; |
| 409 | } |
| 410 | |
| 411 | /* Go check for the CPU being offline. */ |
| 412 | return rcu_implicit_offline_qs(rdp); |
| 413 | } |
| 414 | |
| 415 | #endif /* #ifdef CONFIG_SMP */ |
| 416 | |
| 417 | #else /* #ifdef CONFIG_NO_HZ */ |
| 418 | |
| 419 | static void dyntick_record_completed(struct rcu_state *rsp, long comp) |
| 420 | { |
| 421 | } |
| 422 | |
| 423 | #ifdef CONFIG_SMP |
| 424 | |
| 425 | /* |
| 426 | * If there are no dynticks, then the only way that a CPU can passively |
| 427 | * be in a quiescent state is to be offline. Unlike dynticks idle, which |
| 428 | * is a point in time during the prior (already finished) grace period, |
| 429 | * an offline CPU is always in a quiescent state, and thus can be |
| 430 | * unconditionally applied. So just return the current value of completed. |
| 431 | */ |
| 432 | static long dyntick_recall_completed(struct rcu_state *rsp) |
| 433 | { |
| 434 | return rsp->completed; |
| 435 | } |
| 436 | |
| 437 | static int dyntick_save_progress_counter(struct rcu_data *rdp) |
| 438 | { |
| 439 | return 0; |
| 440 | } |
| 441 | |
| 442 | static int rcu_implicit_dynticks_qs(struct rcu_data *rdp) |
| 443 | { |
| 444 | return rcu_implicit_offline_qs(rdp); |
| 445 | } |
| 446 | |
| 447 | #endif /* #ifdef CONFIG_SMP */ |
| 448 | |
| 449 | #endif /* #else #ifdef CONFIG_NO_HZ */ |
| 450 | |
| 451 | #ifdef CONFIG_RCU_CPU_STALL_DETECTOR |
| 452 | |
| 453 | static void record_gp_stall_check_time(struct rcu_state *rsp) |
| 454 | { |
| 455 | rsp->gp_start = jiffies; |
| 456 | rsp->jiffies_stall = jiffies + RCU_SECONDS_TILL_STALL_CHECK; |
| 457 | } |
| 458 | |
| 459 | static void print_other_cpu_stall(struct rcu_state *rsp) |
| 460 | { |
| 461 | int cpu; |
| 462 | long delta; |
| 463 | unsigned long flags; |
| 464 | struct rcu_node *rnp = rcu_get_root(rsp); |
| 465 | |
| 466 | /* Only let one CPU complain about others per time interval. */ |
| 467 | |
| 468 | spin_lock_irqsave(&rnp->lock, flags); |
| 469 | delta = jiffies - rsp->jiffies_stall; |
| 470 | if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) { |
| 471 | spin_unlock_irqrestore(&rnp->lock, flags); |
| 472 | return; |
| 473 | } |
| 474 | rsp->jiffies_stall = jiffies + RCU_SECONDS_TILL_STALL_RECHECK; |
| 475 | |
| 476 | /* |
| 477 | * Now rat on any tasks that got kicked up to the root rcu_node |
| 478 | * due to CPU offlining. |
| 479 | */ |
| 480 | rcu_print_task_stall(rnp); |
| 481 | spin_unlock_irqrestore(&rnp->lock, flags); |
| 482 | |
| 483 | /* OK, time to rat on our buddy... */ |
| 484 | |
| 485 | printk(KERN_ERR "INFO: RCU detected CPU stalls:"); |
| 486 | rcu_for_each_leaf_node(rsp, rnp) { |
| 487 | rcu_print_task_stall(rnp); |
| 488 | if (rnp->qsmask == 0) |
| 489 | continue; |
| 490 | for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++) |
| 491 | if (rnp->qsmask & (1UL << cpu)) |
| 492 | printk(" %d", rnp->grplo + cpu); |
| 493 | } |
| 494 | printk(" (detected by %d, t=%ld jiffies)\n", |
| 495 | smp_processor_id(), (long)(jiffies - rsp->gp_start)); |
| 496 | trigger_all_cpu_backtrace(); |
| 497 | |
| 498 | force_quiescent_state(rsp, 0); /* Kick them all. */ |
| 499 | } |
| 500 | |
| 501 | static void print_cpu_stall(struct rcu_state *rsp) |
| 502 | { |
| 503 | unsigned long flags; |
| 504 | struct rcu_node *rnp = rcu_get_root(rsp); |
| 505 | |
| 506 | printk(KERN_ERR "INFO: RCU detected CPU %d stall (t=%lu jiffies)\n", |
| 507 | smp_processor_id(), jiffies - rsp->gp_start); |
| 508 | trigger_all_cpu_backtrace(); |
| 509 | |
| 510 | spin_lock_irqsave(&rnp->lock, flags); |
| 511 | if ((long)(jiffies - rsp->jiffies_stall) >= 0) |
| 512 | rsp->jiffies_stall = |
| 513 | jiffies + RCU_SECONDS_TILL_STALL_RECHECK; |
| 514 | spin_unlock_irqrestore(&rnp->lock, flags); |
| 515 | |
| 516 | set_need_resched(); /* kick ourselves to get things going. */ |
| 517 | } |
| 518 | |
| 519 | static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp) |
| 520 | { |
| 521 | long delta; |
| 522 | struct rcu_node *rnp; |
| 523 | |
| 524 | delta = jiffies - rsp->jiffies_stall; |
| 525 | rnp = rdp->mynode; |
| 526 | if ((rnp->qsmask & rdp->grpmask) && delta >= 0) { |
| 527 | |
| 528 | /* We haven't checked in, so go dump stack. */ |
| 529 | print_cpu_stall(rsp); |
| 530 | |
| 531 | } else if (rcu_gp_in_progress(rsp) && delta >= RCU_STALL_RAT_DELAY) { |
| 532 | |
| 533 | /* They had two time units to dump stack, so complain. */ |
| 534 | print_other_cpu_stall(rsp); |
| 535 | } |
| 536 | } |
| 537 | |
| 538 | #else /* #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */ |
| 539 | |
| 540 | static void record_gp_stall_check_time(struct rcu_state *rsp) |
| 541 | { |
| 542 | } |
| 543 | |
| 544 | static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp) |
| 545 | { |
| 546 | } |
| 547 | |
| 548 | #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */ |
| 549 | |
| 550 | /* |
| 551 | * Update CPU-local rcu_data state to record the newly noticed grace period. |
| 552 | * This is used both when we started the grace period and when we notice |
| 553 | * that someone else started the grace period. |
| 554 | */ |
| 555 | static void note_new_gpnum(struct rcu_state *rsp, struct rcu_data *rdp) |
| 556 | { |
| 557 | rdp->qs_pending = 1; |
| 558 | rdp->passed_quiesc = 0; |
| 559 | rdp->gpnum = rsp->gpnum; |
| 560 | } |
| 561 | |
| 562 | /* |
| 563 | * Did someone else start a new RCU grace period start since we last |
| 564 | * checked? Update local state appropriately if so. Must be called |
| 565 | * on the CPU corresponding to rdp. |
| 566 | */ |
| 567 | static int |
| 568 | check_for_new_grace_period(struct rcu_state *rsp, struct rcu_data *rdp) |
| 569 | { |
| 570 | unsigned long flags; |
| 571 | int ret = 0; |
| 572 | |
| 573 | local_irq_save(flags); |
| 574 | if (rdp->gpnum != rsp->gpnum) { |
| 575 | note_new_gpnum(rsp, rdp); |
| 576 | ret = 1; |
| 577 | } |
| 578 | local_irq_restore(flags); |
| 579 | return ret; |
| 580 | } |
| 581 | |
| 582 | /* |
| 583 | * Start a new RCU grace period if warranted, re-initializing the hierarchy |
| 584 | * in preparation for detecting the next grace period. The caller must hold |
| 585 | * the root node's ->lock, which is released before return. Hard irqs must |
| 586 | * be disabled. |
| 587 | */ |
| 588 | static void |
| 589 | rcu_start_gp(struct rcu_state *rsp, unsigned long flags) |
| 590 | __releases(rcu_get_root(rsp)->lock) |
| 591 | { |
| 592 | struct rcu_data *rdp = rsp->rda[smp_processor_id()]; |
| 593 | struct rcu_node *rnp = rcu_get_root(rsp); |
| 594 | |
| 595 | if (!cpu_needs_another_gp(rsp, rdp)) { |
| 596 | spin_unlock_irqrestore(&rnp->lock, flags); |
| 597 | return; |
| 598 | } |
| 599 | |
| 600 | /* Advance to a new grace period and initialize state. */ |
| 601 | rsp->gpnum++; |
| 602 | WARN_ON_ONCE(rsp->signaled == RCU_GP_INIT); |
| 603 | rsp->signaled = RCU_GP_INIT; /* Hold off force_quiescent_state. */ |
| 604 | rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS; |
| 605 | record_gp_stall_check_time(rsp); |
| 606 | dyntick_record_completed(rsp, rsp->completed - 1); |
| 607 | note_new_gpnum(rsp, rdp); |
| 608 | |
| 609 | /* |
| 610 | * Because this CPU just now started the new grace period, we know |
| 611 | * that all of its callbacks will be covered by this upcoming grace |
| 612 | * period, even the ones that were registered arbitrarily recently. |
| 613 | * Therefore, advance all outstanding callbacks to RCU_WAIT_TAIL. |
| 614 | * |
| 615 | * Other CPUs cannot be sure exactly when the grace period started. |
| 616 | * Therefore, their recently registered callbacks must pass through |
| 617 | * an additional RCU_NEXT_READY stage, so that they will be handled |
| 618 | * by the next RCU grace period. |
| 619 | */ |
| 620 | rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL]; |
| 621 | rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_TAIL]; |
| 622 | |
| 623 | /* Special-case the common single-level case. */ |
| 624 | if (NUM_RCU_NODES == 1) { |
| 625 | rcu_preempt_check_blocked_tasks(rnp); |
| 626 | rnp->qsmask = rnp->qsmaskinit; |
| 627 | rnp->gpnum = rsp->gpnum; |
| 628 | rsp->signaled = RCU_SIGNAL_INIT; /* force_quiescent_state OK. */ |
| 629 | spin_unlock_irqrestore(&rnp->lock, flags); |
| 630 | return; |
| 631 | } |
| 632 | |
| 633 | spin_unlock(&rnp->lock); /* leave irqs disabled. */ |
| 634 | |
| 635 | |
| 636 | /* Exclude any concurrent CPU-hotplug operations. */ |
| 637 | spin_lock(&rsp->onofflock); /* irqs already disabled. */ |
| 638 | |
| 639 | /* |
| 640 | * Set the quiescent-state-needed bits in all the rcu_node |
| 641 | * structures for all currently online CPUs in breadth-first |
| 642 | * order, starting from the root rcu_node structure. This |
| 643 | * operation relies on the layout of the hierarchy within the |
| 644 | * rsp->node[] array. Note that other CPUs will access only |
| 645 | * the leaves of the hierarchy, which still indicate that no |
| 646 | * grace period is in progress, at least until the corresponding |
| 647 | * leaf node has been initialized. In addition, we have excluded |
| 648 | * CPU-hotplug operations. |
| 649 | * |
| 650 | * Note that the grace period cannot complete until we finish |
| 651 | * the initialization process, as there will be at least one |
| 652 | * qsmask bit set in the root node until that time, namely the |
| 653 | * one corresponding to this CPU, due to the fact that we have |
| 654 | * irqs disabled. |
| 655 | */ |
| 656 | rcu_for_each_node_breadth_first(rsp, rnp) { |
| 657 | spin_lock(&rnp->lock); /* irqs already disabled. */ |
| 658 | rcu_preempt_check_blocked_tasks(rnp); |
| 659 | rnp->qsmask = rnp->qsmaskinit; |
| 660 | rnp->gpnum = rsp->gpnum; |
| 661 | spin_unlock(&rnp->lock); /* irqs already disabled. */ |
| 662 | } |
| 663 | |
| 664 | rsp->signaled = RCU_SIGNAL_INIT; /* force_quiescent_state now OK. */ |
| 665 | spin_unlock_irqrestore(&rsp->onofflock, flags); |
| 666 | } |
| 667 | |
| 668 | /* |
| 669 | * Advance this CPU's callbacks, but only if the current grace period |
| 670 | * has ended. This may be called only from the CPU to whom the rdp |
| 671 | * belongs. |
| 672 | */ |
| 673 | static void |
| 674 | rcu_process_gp_end(struct rcu_state *rsp, struct rcu_data *rdp) |
| 675 | { |
| 676 | long completed_snap; |
| 677 | unsigned long flags; |
| 678 | |
| 679 | local_irq_save(flags); |
| 680 | completed_snap = ACCESS_ONCE(rsp->completed); /* outside of lock. */ |
| 681 | |
| 682 | /* Did another grace period end? */ |
| 683 | if (rdp->completed != completed_snap) { |
| 684 | |
| 685 | /* Advance callbacks. No harm if list empty. */ |
| 686 | rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[RCU_WAIT_TAIL]; |
| 687 | rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_READY_TAIL]; |
| 688 | rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL]; |
| 689 | |
| 690 | /* Remember that we saw this grace-period completion. */ |
| 691 | rdp->completed = completed_snap; |
| 692 | } |
| 693 | local_irq_restore(flags); |
| 694 | } |
| 695 | |
| 696 | /* |
| 697 | * Clean up after the prior grace period and let rcu_start_gp() start up |
| 698 | * the next grace period if one is needed. Note that the caller must |
| 699 | * hold rnp->lock, as required by rcu_start_gp(), which will release it. |
| 700 | */ |
| 701 | static void cpu_quiet_msk_finish(struct rcu_state *rsp, unsigned long flags) |
| 702 | __releases(rcu_get_root(rsp)->lock) |
| 703 | { |
| 704 | WARN_ON_ONCE(!rcu_gp_in_progress(rsp)); |
| 705 | rsp->completed = rsp->gpnum; |
| 706 | rcu_process_gp_end(rsp, rsp->rda[smp_processor_id()]); |
| 707 | rcu_start_gp(rsp, flags); /* releases root node's rnp->lock. */ |
| 708 | } |
| 709 | |
| 710 | /* |
| 711 | * Similar to cpu_quiet(), for which it is a helper function. Allows |
| 712 | * a group of CPUs to be quieted at one go, though all the CPUs in the |
| 713 | * group must be represented by the same leaf rcu_node structure. |
| 714 | * That structure's lock must be held upon entry, and it is released |
| 715 | * before return. |
| 716 | */ |
| 717 | static void |
| 718 | cpu_quiet_msk(unsigned long mask, struct rcu_state *rsp, struct rcu_node *rnp, |
| 719 | unsigned long flags) |
| 720 | __releases(rnp->lock) |
| 721 | { |
| 722 | struct rcu_node *rnp_c; |
| 723 | |
| 724 | /* Walk up the rcu_node hierarchy. */ |
| 725 | for (;;) { |
| 726 | if (!(rnp->qsmask & mask)) { |
| 727 | |
| 728 | /* Our bit has already been cleared, so done. */ |
| 729 | spin_unlock_irqrestore(&rnp->lock, flags); |
| 730 | return; |
| 731 | } |
| 732 | rnp->qsmask &= ~mask; |
| 733 | if (rnp->qsmask != 0 || rcu_preempted_readers(rnp)) { |
| 734 | |
| 735 | /* Other bits still set at this level, so done. */ |
| 736 | spin_unlock_irqrestore(&rnp->lock, flags); |
| 737 | return; |
| 738 | } |
| 739 | mask = rnp->grpmask; |
| 740 | if (rnp->parent == NULL) { |
| 741 | |
| 742 | /* No more levels. Exit loop holding root lock. */ |
| 743 | |
| 744 | break; |
| 745 | } |
| 746 | spin_unlock_irqrestore(&rnp->lock, flags); |
| 747 | rnp_c = rnp; |
| 748 | rnp = rnp->parent; |
| 749 | spin_lock_irqsave(&rnp->lock, flags); |
| 750 | WARN_ON_ONCE(rnp_c->qsmask); |
| 751 | } |
| 752 | |
| 753 | /* |
| 754 | * Get here if we are the last CPU to pass through a quiescent |
| 755 | * state for this grace period. Invoke cpu_quiet_msk_finish() |
| 756 | * to clean up and start the next grace period if one is needed. |
| 757 | */ |
| 758 | cpu_quiet_msk_finish(rsp, flags); /* releases rnp->lock. */ |
| 759 | } |
| 760 | |
| 761 | /* |
| 762 | * Record a quiescent state for the specified CPU, which must either be |
| 763 | * the current CPU. The lastcomp argument is used to make sure we are |
| 764 | * still in the grace period of interest. We don't want to end the current |
| 765 | * grace period based on quiescent states detected in an earlier grace |
| 766 | * period! |
| 767 | */ |
| 768 | static void |
| 769 | cpu_quiet(int cpu, struct rcu_state *rsp, struct rcu_data *rdp, long lastcomp) |
| 770 | { |
| 771 | unsigned long flags; |
| 772 | unsigned long mask; |
| 773 | struct rcu_node *rnp; |
| 774 | |
| 775 | rnp = rdp->mynode; |
| 776 | spin_lock_irqsave(&rnp->lock, flags); |
| 777 | if (lastcomp != ACCESS_ONCE(rsp->completed)) { |
| 778 | |
| 779 | /* |
| 780 | * Someone beat us to it for this grace period, so leave. |
| 781 | * The race with GP start is resolved by the fact that we |
| 782 | * hold the leaf rcu_node lock, so that the per-CPU bits |
| 783 | * cannot yet be initialized -- so we would simply find our |
| 784 | * CPU's bit already cleared in cpu_quiet_msk() if this race |
| 785 | * occurred. |
| 786 | */ |
| 787 | rdp->passed_quiesc = 0; /* try again later! */ |
| 788 | spin_unlock_irqrestore(&rnp->lock, flags); |
| 789 | return; |
| 790 | } |
| 791 | mask = rdp->grpmask; |
| 792 | if ((rnp->qsmask & mask) == 0) { |
| 793 | spin_unlock_irqrestore(&rnp->lock, flags); |
| 794 | } else { |
| 795 | rdp->qs_pending = 0; |
| 796 | |
| 797 | /* |
| 798 | * This GP can't end until cpu checks in, so all of our |
| 799 | * callbacks can be processed during the next GP. |
| 800 | */ |
| 801 | rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL]; |
| 802 | |
| 803 | cpu_quiet_msk(mask, rsp, rnp, flags); /* releases rnp->lock */ |
| 804 | } |
| 805 | } |
| 806 | |
| 807 | /* |
| 808 | * Check to see if there is a new grace period of which this CPU |
| 809 | * is not yet aware, and if so, set up local rcu_data state for it. |
| 810 | * Otherwise, see if this CPU has just passed through its first |
| 811 | * quiescent state for this grace period, and record that fact if so. |
| 812 | */ |
| 813 | static void |
| 814 | rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp) |
| 815 | { |
| 816 | /* If there is now a new grace period, record and return. */ |
| 817 | if (check_for_new_grace_period(rsp, rdp)) |
| 818 | return; |
| 819 | |
| 820 | /* |
| 821 | * Does this CPU still need to do its part for current grace period? |
| 822 | * If no, return and let the other CPUs do their part as well. |
| 823 | */ |
| 824 | if (!rdp->qs_pending) |
| 825 | return; |
| 826 | |
| 827 | /* |
| 828 | * Was there a quiescent state since the beginning of the grace |
| 829 | * period? If no, then exit and wait for the next call. |
| 830 | */ |
| 831 | if (!rdp->passed_quiesc) |
| 832 | return; |
| 833 | |
| 834 | /* Tell RCU we are done (but cpu_quiet() will be the judge of that). */ |
| 835 | cpu_quiet(rdp->cpu, rsp, rdp, rdp->passed_quiesc_completed); |
| 836 | } |
| 837 | |
| 838 | #ifdef CONFIG_HOTPLUG_CPU |
| 839 | |
| 840 | /* |
| 841 | * Remove the outgoing CPU from the bitmasks in the rcu_node hierarchy |
| 842 | * and move all callbacks from the outgoing CPU to the current one. |
| 843 | */ |
| 844 | static void __rcu_offline_cpu(int cpu, struct rcu_state *rsp) |
| 845 | { |
| 846 | int i; |
| 847 | unsigned long flags; |
| 848 | long lastcomp; |
| 849 | unsigned long mask; |
| 850 | struct rcu_data *rdp = rsp->rda[cpu]; |
| 851 | struct rcu_data *rdp_me; |
| 852 | struct rcu_node *rnp; |
| 853 | |
| 854 | /* Exclude any attempts to start a new grace period. */ |
| 855 | spin_lock_irqsave(&rsp->onofflock, flags); |
| 856 | |
| 857 | /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */ |
| 858 | rnp = rdp->mynode; /* this is the outgoing CPU's rnp. */ |
| 859 | mask = rdp->grpmask; /* rnp->grplo is constant. */ |
| 860 | do { |
| 861 | spin_lock(&rnp->lock); /* irqs already disabled. */ |
| 862 | rnp->qsmaskinit &= ~mask; |
| 863 | if (rnp->qsmaskinit != 0) { |
| 864 | spin_unlock(&rnp->lock); /* irqs remain disabled. */ |
| 865 | break; |
| 866 | } |
| 867 | rcu_preempt_offline_tasks(rsp, rnp, rdp); |
| 868 | mask = rnp->grpmask; |
| 869 | spin_unlock(&rnp->lock); /* irqs remain disabled. */ |
| 870 | rnp = rnp->parent; |
| 871 | } while (rnp != NULL); |
| 872 | lastcomp = rsp->completed; |
| 873 | |
| 874 | spin_unlock(&rsp->onofflock); /* irqs remain disabled. */ |
| 875 | |
| 876 | /* |
| 877 | * Move callbacks from the outgoing CPU to the running CPU. |
| 878 | * Note that the outgoing CPU is now quiescent, so it is now |
| 879 | * (uncharacteristically) safe to access its rcu_data structure. |
| 880 | * Note also that we must carefully retain the order of the |
| 881 | * outgoing CPU's callbacks in order for rcu_barrier() to work |
| 882 | * correctly. Finally, note that we start all the callbacks |
| 883 | * afresh, even those that have passed through a grace period |
| 884 | * and are therefore ready to invoke. The theory is that hotplug |
| 885 | * events are rare, and that if they are frequent enough to |
| 886 | * indefinitely delay callbacks, you have far worse things to |
| 887 | * be worrying about. |
| 888 | */ |
| 889 | if (rdp->nxtlist != NULL) { |
| 890 | rdp_me = rsp->rda[smp_processor_id()]; |
| 891 | *rdp_me->nxttail[RCU_NEXT_TAIL] = rdp->nxtlist; |
| 892 | rdp_me->nxttail[RCU_NEXT_TAIL] = rdp->nxttail[RCU_NEXT_TAIL]; |
| 893 | rdp->nxtlist = NULL; |
| 894 | for (i = 0; i < RCU_NEXT_SIZE; i++) |
| 895 | rdp->nxttail[i] = &rdp->nxtlist; |
| 896 | rdp_me->qlen += rdp->qlen; |
| 897 | rdp->qlen = 0; |
| 898 | } |
| 899 | local_irq_restore(flags); |
| 900 | } |
| 901 | |
| 902 | /* |
| 903 | * Remove the specified CPU from the RCU hierarchy and move any pending |
| 904 | * callbacks that it might have to the current CPU. This code assumes |
| 905 | * that at least one CPU in the system will remain running at all times. |
| 906 | * Any attempt to offline -all- CPUs is likely to strand RCU callbacks. |
| 907 | */ |
| 908 | static void rcu_offline_cpu(int cpu) |
| 909 | { |
| 910 | __rcu_offline_cpu(cpu, &rcu_sched_state); |
| 911 | __rcu_offline_cpu(cpu, &rcu_bh_state); |
| 912 | rcu_preempt_offline_cpu(cpu); |
| 913 | } |
| 914 | |
| 915 | #else /* #ifdef CONFIG_HOTPLUG_CPU */ |
| 916 | |
| 917 | static void rcu_offline_cpu(int cpu) |
| 918 | { |
| 919 | } |
| 920 | |
| 921 | #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */ |
| 922 | |
| 923 | /* |
| 924 | * Invoke any RCU callbacks that have made it to the end of their grace |
| 925 | * period. Thottle as specified by rdp->blimit. |
| 926 | */ |
| 927 | static void rcu_do_batch(struct rcu_data *rdp) |
| 928 | { |
| 929 | unsigned long flags; |
| 930 | struct rcu_head *next, *list, **tail; |
| 931 | int count; |
| 932 | |
| 933 | /* If no callbacks are ready, just return.*/ |
| 934 | if (!cpu_has_callbacks_ready_to_invoke(rdp)) |
| 935 | return; |
| 936 | |
| 937 | /* |
| 938 | * Extract the list of ready callbacks, disabling to prevent |
| 939 | * races with call_rcu() from interrupt handlers. |
| 940 | */ |
| 941 | local_irq_save(flags); |
| 942 | list = rdp->nxtlist; |
| 943 | rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL]; |
| 944 | *rdp->nxttail[RCU_DONE_TAIL] = NULL; |
| 945 | tail = rdp->nxttail[RCU_DONE_TAIL]; |
| 946 | for (count = RCU_NEXT_SIZE - 1; count >= 0; count--) |
| 947 | if (rdp->nxttail[count] == rdp->nxttail[RCU_DONE_TAIL]) |
| 948 | rdp->nxttail[count] = &rdp->nxtlist; |
| 949 | local_irq_restore(flags); |
| 950 | |
| 951 | /* Invoke callbacks. */ |
| 952 | count = 0; |
| 953 | while (list) { |
| 954 | next = list->next; |
| 955 | prefetch(next); |
| 956 | list->func(list); |
| 957 | list = next; |
| 958 | if (++count >= rdp->blimit) |
| 959 | break; |
| 960 | } |
| 961 | |
| 962 | local_irq_save(flags); |
| 963 | |
| 964 | /* Update count, and requeue any remaining callbacks. */ |
| 965 | rdp->qlen -= count; |
| 966 | if (list != NULL) { |
| 967 | *tail = rdp->nxtlist; |
| 968 | rdp->nxtlist = list; |
| 969 | for (count = 0; count < RCU_NEXT_SIZE; count++) |
| 970 | if (&rdp->nxtlist == rdp->nxttail[count]) |
| 971 | rdp->nxttail[count] = tail; |
| 972 | else |
| 973 | break; |
| 974 | } |
| 975 | |
| 976 | /* Reinstate batch limit if we have worked down the excess. */ |
| 977 | if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark) |
| 978 | rdp->blimit = blimit; |
| 979 | |
| 980 | local_irq_restore(flags); |
| 981 | |
| 982 | /* Re-raise the RCU softirq if there are callbacks remaining. */ |
| 983 | if (cpu_has_callbacks_ready_to_invoke(rdp)) |
| 984 | raise_softirq(RCU_SOFTIRQ); |
| 985 | } |
| 986 | |
| 987 | /* |
| 988 | * Check to see if this CPU is in a non-context-switch quiescent state |
| 989 | * (user mode or idle loop for rcu, non-softirq execution for rcu_bh). |
| 990 | * Also schedule the RCU softirq handler. |
| 991 | * |
| 992 | * This function must be called with hardirqs disabled. It is normally |
| 993 | * invoked from the scheduling-clock interrupt. If rcu_pending returns |
| 994 | * false, there is no point in invoking rcu_check_callbacks(). |
| 995 | */ |
| 996 | void rcu_check_callbacks(int cpu, int user) |
| 997 | { |
| 998 | if (!rcu_pending(cpu)) |
| 999 | return; /* if nothing for RCU to do. */ |
| 1000 | if (user || |
| 1001 | (idle_cpu(cpu) && rcu_scheduler_active && |
| 1002 | !in_softirq() && hardirq_count() <= (1 << HARDIRQ_SHIFT))) { |
| 1003 | |
| 1004 | /* |
| 1005 | * Get here if this CPU took its interrupt from user |
| 1006 | * mode or from the idle loop, and if this is not a |
| 1007 | * nested interrupt. In this case, the CPU is in |
| 1008 | * a quiescent state, so note it. |
| 1009 | * |
| 1010 | * No memory barrier is required here because both |
| 1011 | * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local |
| 1012 | * variables that other CPUs neither access nor modify, |
| 1013 | * at least not while the corresponding CPU is online. |
| 1014 | */ |
| 1015 | |
| 1016 | rcu_sched_qs(cpu); |
| 1017 | rcu_bh_qs(cpu); |
| 1018 | |
| 1019 | } else if (!in_softirq()) { |
| 1020 | |
| 1021 | /* |
| 1022 | * Get here if this CPU did not take its interrupt from |
| 1023 | * softirq, in other words, if it is not interrupting |
| 1024 | * a rcu_bh read-side critical section. This is an _bh |
| 1025 | * critical section, so note it. |
| 1026 | */ |
| 1027 | |
| 1028 | rcu_bh_qs(cpu); |
| 1029 | } |
| 1030 | rcu_preempt_check_callbacks(cpu); |
| 1031 | raise_softirq(RCU_SOFTIRQ); |
| 1032 | } |
| 1033 | |
| 1034 | #ifdef CONFIG_SMP |
| 1035 | |
| 1036 | /* |
| 1037 | * Scan the leaf rcu_node structures, processing dyntick state for any that |
| 1038 | * have not yet encountered a quiescent state, using the function specified. |
| 1039 | * Returns 1 if the current grace period ends while scanning (possibly |
| 1040 | * because we made it end). |
| 1041 | */ |
| 1042 | static int rcu_process_dyntick(struct rcu_state *rsp, long lastcomp, |
| 1043 | int (*f)(struct rcu_data *)) |
| 1044 | { |
| 1045 | unsigned long bit; |
| 1046 | int cpu; |
| 1047 | unsigned long flags; |
| 1048 | unsigned long mask; |
| 1049 | struct rcu_node *rnp; |
| 1050 | |
| 1051 | rcu_for_each_leaf_node(rsp, rnp) { |
| 1052 | mask = 0; |
| 1053 | spin_lock_irqsave(&rnp->lock, flags); |
| 1054 | if (rsp->completed != lastcomp) { |
| 1055 | spin_unlock_irqrestore(&rnp->lock, flags); |
| 1056 | return 1; |
| 1057 | } |
| 1058 | if (rnp->qsmask == 0) { |
| 1059 | spin_unlock_irqrestore(&rnp->lock, flags); |
| 1060 | continue; |
| 1061 | } |
| 1062 | cpu = rnp->grplo; |
| 1063 | bit = 1; |
| 1064 | for (; cpu <= rnp->grphi; cpu++, bit <<= 1) { |
| 1065 | if ((rnp->qsmask & bit) != 0 && f(rsp->rda[cpu])) |
| 1066 | mask |= bit; |
| 1067 | } |
| 1068 | if (mask != 0 && rsp->completed == lastcomp) { |
| 1069 | |
| 1070 | /* cpu_quiet_msk() releases rnp->lock. */ |
| 1071 | cpu_quiet_msk(mask, rsp, rnp, flags); |
| 1072 | continue; |
| 1073 | } |
| 1074 | spin_unlock_irqrestore(&rnp->lock, flags); |
| 1075 | } |
| 1076 | return 0; |
| 1077 | } |
| 1078 | |
| 1079 | /* |
| 1080 | * Force quiescent states on reluctant CPUs, and also detect which |
| 1081 | * CPUs are in dyntick-idle mode. |
| 1082 | */ |
| 1083 | static void force_quiescent_state(struct rcu_state *rsp, int relaxed) |
| 1084 | { |
| 1085 | unsigned long flags; |
| 1086 | long lastcomp; |
| 1087 | struct rcu_node *rnp = rcu_get_root(rsp); |
| 1088 | u8 signaled; |
| 1089 | |
| 1090 | if (!rcu_gp_in_progress(rsp)) |
| 1091 | return; /* No grace period in progress, nothing to force. */ |
| 1092 | if (!spin_trylock_irqsave(&rsp->fqslock, flags)) { |
| 1093 | rsp->n_force_qs_lh++; /* Inexact, can lose counts. Tough! */ |
| 1094 | return; /* Someone else is already on the job. */ |
| 1095 | } |
| 1096 | if (relaxed && |
| 1097 | (long)(rsp->jiffies_force_qs - jiffies) >= 0) |
| 1098 | goto unlock_ret; /* no emergency and done recently. */ |
| 1099 | rsp->n_force_qs++; |
| 1100 | spin_lock(&rnp->lock); |
| 1101 | lastcomp = rsp->completed; |
| 1102 | signaled = rsp->signaled; |
| 1103 | rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS; |
| 1104 | if (lastcomp == rsp->gpnum) { |
| 1105 | rsp->n_force_qs_ngp++; |
| 1106 | spin_unlock(&rnp->lock); |
| 1107 | goto unlock_ret; /* no GP in progress, time updated. */ |
| 1108 | } |
| 1109 | spin_unlock(&rnp->lock); |
| 1110 | switch (signaled) { |
| 1111 | case RCU_GP_INIT: |
| 1112 | |
| 1113 | break; /* grace period still initializing, ignore. */ |
| 1114 | |
| 1115 | case RCU_SAVE_DYNTICK: |
| 1116 | |
| 1117 | if (RCU_SIGNAL_INIT != RCU_SAVE_DYNTICK) |
| 1118 | break; /* So gcc recognizes the dead code. */ |
| 1119 | |
| 1120 | /* Record dyntick-idle state. */ |
| 1121 | if (rcu_process_dyntick(rsp, lastcomp, |
| 1122 | dyntick_save_progress_counter)) |
| 1123 | goto unlock_ret; |
| 1124 | |
| 1125 | /* Update state, record completion counter. */ |
| 1126 | spin_lock(&rnp->lock); |
| 1127 | if (lastcomp == rsp->completed) { |
| 1128 | rsp->signaled = RCU_FORCE_QS; |
| 1129 | dyntick_record_completed(rsp, lastcomp); |
| 1130 | } |
| 1131 | spin_unlock(&rnp->lock); |
| 1132 | break; |
| 1133 | |
| 1134 | case RCU_FORCE_QS: |
| 1135 | |
| 1136 | /* Check dyntick-idle state, send IPI to laggarts. */ |
| 1137 | if (rcu_process_dyntick(rsp, dyntick_recall_completed(rsp), |
| 1138 | rcu_implicit_dynticks_qs)) |
| 1139 | goto unlock_ret; |
| 1140 | |
| 1141 | /* Leave state in case more forcing is required. */ |
| 1142 | |
| 1143 | break; |
| 1144 | } |
| 1145 | unlock_ret: |
| 1146 | spin_unlock_irqrestore(&rsp->fqslock, flags); |
| 1147 | } |
| 1148 | |
| 1149 | #else /* #ifdef CONFIG_SMP */ |
| 1150 | |
| 1151 | static void force_quiescent_state(struct rcu_state *rsp, int relaxed) |
| 1152 | { |
| 1153 | set_need_resched(); |
| 1154 | } |
| 1155 | |
| 1156 | #endif /* #else #ifdef CONFIG_SMP */ |
| 1157 | |
| 1158 | /* |
| 1159 | * This does the RCU processing work from softirq context for the |
| 1160 | * specified rcu_state and rcu_data structures. This may be called |
| 1161 | * only from the CPU to whom the rdp belongs. |
| 1162 | */ |
| 1163 | static void |
| 1164 | __rcu_process_callbacks(struct rcu_state *rsp, struct rcu_data *rdp) |
| 1165 | { |
| 1166 | unsigned long flags; |
| 1167 | |
| 1168 | WARN_ON_ONCE(rdp->beenonline == 0); |
| 1169 | |
| 1170 | /* |
| 1171 | * If an RCU GP has gone long enough, go check for dyntick |
| 1172 | * idle CPUs and, if needed, send resched IPIs. |
| 1173 | */ |
| 1174 | if ((long)(ACCESS_ONCE(rsp->jiffies_force_qs) - jiffies) < 0) |
| 1175 | force_quiescent_state(rsp, 1); |
| 1176 | |
| 1177 | /* |
| 1178 | * Advance callbacks in response to end of earlier grace |
| 1179 | * period that some other CPU ended. |
| 1180 | */ |
| 1181 | rcu_process_gp_end(rsp, rdp); |
| 1182 | |
| 1183 | /* Update RCU state based on any recent quiescent states. */ |
| 1184 | rcu_check_quiescent_state(rsp, rdp); |
| 1185 | |
| 1186 | /* Does this CPU require a not-yet-started grace period? */ |
| 1187 | if (cpu_needs_another_gp(rsp, rdp)) { |
| 1188 | spin_lock_irqsave(&rcu_get_root(rsp)->lock, flags); |
| 1189 | rcu_start_gp(rsp, flags); /* releases above lock */ |
| 1190 | } |
| 1191 | |
| 1192 | /* If there are callbacks ready, invoke them. */ |
| 1193 | rcu_do_batch(rdp); |
| 1194 | } |
| 1195 | |
| 1196 | /* |
| 1197 | * Do softirq processing for the current CPU. |
| 1198 | */ |
| 1199 | static void rcu_process_callbacks(struct softirq_action *unused) |
| 1200 | { |
| 1201 | /* |
| 1202 | * Memory references from any prior RCU read-side critical sections |
| 1203 | * executed by the interrupted code must be seen before any RCU |
| 1204 | * grace-period manipulations below. |
| 1205 | */ |
| 1206 | smp_mb(); /* See above block comment. */ |
| 1207 | |
| 1208 | __rcu_process_callbacks(&rcu_sched_state, |
| 1209 | &__get_cpu_var(rcu_sched_data)); |
| 1210 | __rcu_process_callbacks(&rcu_bh_state, &__get_cpu_var(rcu_bh_data)); |
| 1211 | rcu_preempt_process_callbacks(); |
| 1212 | |
| 1213 | /* |
| 1214 | * Memory references from any later RCU read-side critical sections |
| 1215 | * executed by the interrupted code must be seen after any RCU |
| 1216 | * grace-period manipulations above. |
| 1217 | */ |
| 1218 | smp_mb(); /* See above block comment. */ |
| 1219 | } |
| 1220 | |
| 1221 | static void |
| 1222 | __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu), |
| 1223 | struct rcu_state *rsp) |
| 1224 | { |
| 1225 | unsigned long flags; |
| 1226 | struct rcu_data *rdp; |
| 1227 | |
| 1228 | head->func = func; |
| 1229 | head->next = NULL; |
| 1230 | |
| 1231 | smp_mb(); /* Ensure RCU update seen before callback registry. */ |
| 1232 | |
| 1233 | /* |
| 1234 | * Opportunistically note grace-period endings and beginnings. |
| 1235 | * Note that we might see a beginning right after we see an |
| 1236 | * end, but never vice versa, since this CPU has to pass through |
| 1237 | * a quiescent state betweentimes. |
| 1238 | */ |
| 1239 | local_irq_save(flags); |
| 1240 | rdp = rsp->rda[smp_processor_id()]; |
| 1241 | rcu_process_gp_end(rsp, rdp); |
| 1242 | check_for_new_grace_period(rsp, rdp); |
| 1243 | |
| 1244 | /* Add the callback to our list. */ |
| 1245 | *rdp->nxttail[RCU_NEXT_TAIL] = head; |
| 1246 | rdp->nxttail[RCU_NEXT_TAIL] = &head->next; |
| 1247 | |
| 1248 | /* Start a new grace period if one not already started. */ |
| 1249 | if (!rcu_gp_in_progress(rsp)) { |
| 1250 | unsigned long nestflag; |
| 1251 | struct rcu_node *rnp_root = rcu_get_root(rsp); |
| 1252 | |
| 1253 | spin_lock_irqsave(&rnp_root->lock, nestflag); |
| 1254 | rcu_start_gp(rsp, nestflag); /* releases rnp_root->lock. */ |
| 1255 | } |
| 1256 | |
| 1257 | /* Force the grace period if too many callbacks or too long waiting. */ |
| 1258 | if (unlikely(++rdp->qlen > qhimark)) { |
| 1259 | rdp->blimit = LONG_MAX; |
| 1260 | force_quiescent_state(rsp, 0); |
| 1261 | } else if ((long)(ACCESS_ONCE(rsp->jiffies_force_qs) - jiffies) < 0) |
| 1262 | force_quiescent_state(rsp, 1); |
| 1263 | local_irq_restore(flags); |
| 1264 | } |
| 1265 | |
| 1266 | /* |
| 1267 | * Queue an RCU-sched callback for invocation after a grace period. |
| 1268 | */ |
| 1269 | void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu)) |
| 1270 | { |
| 1271 | __call_rcu(head, func, &rcu_sched_state); |
| 1272 | } |
| 1273 | EXPORT_SYMBOL_GPL(call_rcu_sched); |
| 1274 | |
| 1275 | /* |
| 1276 | * Queue an RCU for invocation after a quicker grace period. |
| 1277 | */ |
| 1278 | void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu)) |
| 1279 | { |
| 1280 | __call_rcu(head, func, &rcu_bh_state); |
| 1281 | } |
| 1282 | EXPORT_SYMBOL_GPL(call_rcu_bh); |
| 1283 | |
| 1284 | /* |
| 1285 | * Check to see if there is any immediate RCU-related work to be done |
| 1286 | * by the current CPU, for the specified type of RCU, returning 1 if so. |
| 1287 | * The checks are in order of increasing expense: checks that can be |
| 1288 | * carried out against CPU-local state are performed first. However, |
| 1289 | * we must check for CPU stalls first, else we might not get a chance. |
| 1290 | */ |
| 1291 | static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp) |
| 1292 | { |
| 1293 | rdp->n_rcu_pending++; |
| 1294 | |
| 1295 | /* Check for CPU stalls, if enabled. */ |
| 1296 | check_cpu_stall(rsp, rdp); |
| 1297 | |
| 1298 | /* Is the RCU core waiting for a quiescent state from this CPU? */ |
| 1299 | if (rdp->qs_pending) { |
| 1300 | rdp->n_rp_qs_pending++; |
| 1301 | return 1; |
| 1302 | } |
| 1303 | |
| 1304 | /* Does this CPU have callbacks ready to invoke? */ |
| 1305 | if (cpu_has_callbacks_ready_to_invoke(rdp)) { |
| 1306 | rdp->n_rp_cb_ready++; |
| 1307 | return 1; |
| 1308 | } |
| 1309 | |
| 1310 | /* Has RCU gone idle with this CPU needing another grace period? */ |
| 1311 | if (cpu_needs_another_gp(rsp, rdp)) { |
| 1312 | rdp->n_rp_cpu_needs_gp++; |
| 1313 | return 1; |
| 1314 | } |
| 1315 | |
| 1316 | /* Has another RCU grace period completed? */ |
| 1317 | if (ACCESS_ONCE(rsp->completed) != rdp->completed) { /* outside lock */ |
| 1318 | rdp->n_rp_gp_completed++; |
| 1319 | return 1; |
| 1320 | } |
| 1321 | |
| 1322 | /* Has a new RCU grace period started? */ |
| 1323 | if (ACCESS_ONCE(rsp->gpnum) != rdp->gpnum) { /* outside lock */ |
| 1324 | rdp->n_rp_gp_started++; |
| 1325 | return 1; |
| 1326 | } |
| 1327 | |
| 1328 | /* Has an RCU GP gone long enough to send resched IPIs &c? */ |
| 1329 | if (rcu_gp_in_progress(rsp) && |
| 1330 | ((long)(ACCESS_ONCE(rsp->jiffies_force_qs) - jiffies) < 0)) { |
| 1331 | rdp->n_rp_need_fqs++; |
| 1332 | return 1; |
| 1333 | } |
| 1334 | |
| 1335 | /* nothing to do */ |
| 1336 | rdp->n_rp_need_nothing++; |
| 1337 | return 0; |
| 1338 | } |
| 1339 | |
| 1340 | /* |
| 1341 | * Check to see if there is any immediate RCU-related work to be done |
| 1342 | * by the current CPU, returning 1 if so. This function is part of the |
| 1343 | * RCU implementation; it is -not- an exported member of the RCU API. |
| 1344 | */ |
| 1345 | static int rcu_pending(int cpu) |
| 1346 | { |
| 1347 | return __rcu_pending(&rcu_sched_state, &per_cpu(rcu_sched_data, cpu)) || |
| 1348 | __rcu_pending(&rcu_bh_state, &per_cpu(rcu_bh_data, cpu)) || |
| 1349 | rcu_preempt_pending(cpu); |
| 1350 | } |
| 1351 | |
| 1352 | /* |
| 1353 | * Check to see if any future RCU-related work will need to be done |
| 1354 | * by the current CPU, even if none need be done immediately, returning |
| 1355 | * 1 if so. This function is part of the RCU implementation; it is -not- |
| 1356 | * an exported member of the RCU API. |
| 1357 | */ |
| 1358 | int rcu_needs_cpu(int cpu) |
| 1359 | { |
| 1360 | /* RCU callbacks either ready or pending? */ |
| 1361 | return per_cpu(rcu_sched_data, cpu).nxtlist || |
| 1362 | per_cpu(rcu_bh_data, cpu).nxtlist || |
| 1363 | rcu_preempt_needs_cpu(cpu); |
| 1364 | } |
| 1365 | |
| 1366 | static DEFINE_PER_CPU(struct rcu_head, rcu_barrier_head) = {NULL}; |
| 1367 | static atomic_t rcu_barrier_cpu_count; |
| 1368 | static DEFINE_MUTEX(rcu_barrier_mutex); |
| 1369 | static struct completion rcu_barrier_completion; |
| 1370 | static atomic_t rcu_migrate_type_count = ATOMIC_INIT(0); |
| 1371 | static struct rcu_head rcu_migrate_head[3]; |
| 1372 | static DECLARE_WAIT_QUEUE_HEAD(rcu_migrate_wq); |
| 1373 | |
| 1374 | static void rcu_barrier_callback(struct rcu_head *notused) |
| 1375 | { |
| 1376 | if (atomic_dec_and_test(&rcu_barrier_cpu_count)) |
| 1377 | complete(&rcu_barrier_completion); |
| 1378 | } |
| 1379 | |
| 1380 | /* |
| 1381 | * Called with preemption disabled, and from cross-cpu IRQ context. |
| 1382 | */ |
| 1383 | static void rcu_barrier_func(void *type) |
| 1384 | { |
| 1385 | int cpu = smp_processor_id(); |
| 1386 | struct rcu_head *head = &per_cpu(rcu_barrier_head, cpu); |
| 1387 | void (*call_rcu_func)(struct rcu_head *head, |
| 1388 | void (*func)(struct rcu_head *head)); |
| 1389 | |
| 1390 | atomic_inc(&rcu_barrier_cpu_count); |
| 1391 | call_rcu_func = type; |
| 1392 | call_rcu_func(head, rcu_barrier_callback); |
| 1393 | } |
| 1394 | |
| 1395 | static inline void wait_migrated_callbacks(void) |
| 1396 | { |
| 1397 | wait_event(rcu_migrate_wq, !atomic_read(&rcu_migrate_type_count)); |
| 1398 | smp_mb(); /* In case we didn't sleep. */ |
| 1399 | } |
| 1400 | |
| 1401 | /* |
| 1402 | * Orchestrate the specified type of RCU barrier, waiting for all |
| 1403 | * RCU callbacks of the specified type to complete. |
| 1404 | */ |
| 1405 | static void _rcu_barrier(void (*call_rcu_func)(struct rcu_head *head, |
| 1406 | void (*func)(struct rcu_head *head))) |
| 1407 | { |
| 1408 | BUG_ON(in_interrupt()); |
| 1409 | /* Take cpucontrol mutex to protect against CPU hotplug */ |
| 1410 | mutex_lock(&rcu_barrier_mutex); |
| 1411 | init_completion(&rcu_barrier_completion); |
| 1412 | /* |
| 1413 | * Initialize rcu_barrier_cpu_count to 1, then invoke |
| 1414 | * rcu_barrier_func() on each CPU, so that each CPU also has |
| 1415 | * incremented rcu_barrier_cpu_count. Only then is it safe to |
| 1416 | * decrement rcu_barrier_cpu_count -- otherwise the first CPU |
| 1417 | * might complete its grace period before all of the other CPUs |
| 1418 | * did their increment, causing this function to return too |
| 1419 | * early. |
| 1420 | */ |
| 1421 | atomic_set(&rcu_barrier_cpu_count, 1); |
| 1422 | on_each_cpu(rcu_barrier_func, (void *)call_rcu_func, 1); |
| 1423 | if (atomic_dec_and_test(&rcu_barrier_cpu_count)) |
| 1424 | complete(&rcu_barrier_completion); |
| 1425 | wait_for_completion(&rcu_barrier_completion); |
| 1426 | mutex_unlock(&rcu_barrier_mutex); |
| 1427 | wait_migrated_callbacks(); |
| 1428 | } |
| 1429 | |
| 1430 | /** |
| 1431 | * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete. |
| 1432 | */ |
| 1433 | void rcu_barrier(void) |
| 1434 | { |
| 1435 | _rcu_barrier(call_rcu); |
| 1436 | } |
| 1437 | EXPORT_SYMBOL_GPL(rcu_barrier); |
| 1438 | |
| 1439 | /** |
| 1440 | * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete. |
| 1441 | */ |
| 1442 | void rcu_barrier_bh(void) |
| 1443 | { |
| 1444 | _rcu_barrier(call_rcu_bh); |
| 1445 | } |
| 1446 | EXPORT_SYMBOL_GPL(rcu_barrier_bh); |
| 1447 | |
| 1448 | /** |
| 1449 | * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks. |
| 1450 | */ |
| 1451 | void rcu_barrier_sched(void) |
| 1452 | { |
| 1453 | _rcu_barrier(call_rcu_sched); |
| 1454 | } |
| 1455 | EXPORT_SYMBOL_GPL(rcu_barrier_sched); |
| 1456 | |
| 1457 | static void rcu_migrate_callback(struct rcu_head *notused) |
| 1458 | { |
| 1459 | if (atomic_dec_and_test(&rcu_migrate_type_count)) |
| 1460 | wake_up(&rcu_migrate_wq); |
| 1461 | } |
| 1462 | |
| 1463 | /* |
| 1464 | * Do boot-time initialization of a CPU's per-CPU RCU data. |
| 1465 | */ |
| 1466 | static void __init |
| 1467 | rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp) |
| 1468 | { |
| 1469 | unsigned long flags; |
| 1470 | int i; |
| 1471 | struct rcu_data *rdp = rsp->rda[cpu]; |
| 1472 | struct rcu_node *rnp = rcu_get_root(rsp); |
| 1473 | |
| 1474 | /* Set up local state, ensuring consistent view of global state. */ |
| 1475 | spin_lock_irqsave(&rnp->lock, flags); |
| 1476 | rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo); |
| 1477 | rdp->nxtlist = NULL; |
| 1478 | for (i = 0; i < RCU_NEXT_SIZE; i++) |
| 1479 | rdp->nxttail[i] = &rdp->nxtlist; |
| 1480 | rdp->qlen = 0; |
| 1481 | #ifdef CONFIG_NO_HZ |
| 1482 | rdp->dynticks = &per_cpu(rcu_dynticks, cpu); |
| 1483 | #endif /* #ifdef CONFIG_NO_HZ */ |
| 1484 | rdp->cpu = cpu; |
| 1485 | spin_unlock_irqrestore(&rnp->lock, flags); |
| 1486 | } |
| 1487 | |
| 1488 | /* |
| 1489 | * Initialize a CPU's per-CPU RCU data. Note that only one online or |
| 1490 | * offline event can be happening at a given time. Note also that we |
| 1491 | * can accept some slop in the rsp->completed access due to the fact |
| 1492 | * that this CPU cannot possibly have any RCU callbacks in flight yet. |
| 1493 | */ |
| 1494 | static void __cpuinit |
| 1495 | rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptable) |
| 1496 | { |
| 1497 | unsigned long flags; |
| 1498 | long lastcomp; |
| 1499 | unsigned long mask; |
| 1500 | struct rcu_data *rdp = rsp->rda[cpu]; |
| 1501 | struct rcu_node *rnp = rcu_get_root(rsp); |
| 1502 | |
| 1503 | /* Set up local state, ensuring consistent view of global state. */ |
| 1504 | spin_lock_irqsave(&rnp->lock, flags); |
| 1505 | lastcomp = rsp->completed; |
| 1506 | rdp->completed = lastcomp; |
| 1507 | rdp->gpnum = lastcomp; |
| 1508 | rdp->passed_quiesc = 0; /* We could be racing with new GP, */ |
| 1509 | rdp->qs_pending = 1; /* so set up to respond to current GP. */ |
| 1510 | rdp->beenonline = 1; /* We have now been online. */ |
| 1511 | rdp->preemptable = preemptable; |
| 1512 | rdp->passed_quiesc_completed = lastcomp - 1; |
| 1513 | rdp->blimit = blimit; |
| 1514 | spin_unlock(&rnp->lock); /* irqs remain disabled. */ |
| 1515 | |
| 1516 | /* |
| 1517 | * A new grace period might start here. If so, we won't be part |
| 1518 | * of it, but that is OK, as we are currently in a quiescent state. |
| 1519 | */ |
| 1520 | |
| 1521 | /* Exclude any attempts to start a new GP on large systems. */ |
| 1522 | spin_lock(&rsp->onofflock); /* irqs already disabled. */ |
| 1523 | |
| 1524 | /* Add CPU to rcu_node bitmasks. */ |
| 1525 | rnp = rdp->mynode; |
| 1526 | mask = rdp->grpmask; |
| 1527 | do { |
| 1528 | /* Exclude any attempts to start a new GP on small systems. */ |
| 1529 | spin_lock(&rnp->lock); /* irqs already disabled. */ |
| 1530 | rnp->qsmaskinit |= mask; |
| 1531 | mask = rnp->grpmask; |
| 1532 | spin_unlock(&rnp->lock); /* irqs already disabled. */ |
| 1533 | rnp = rnp->parent; |
| 1534 | } while (rnp != NULL && !(rnp->qsmaskinit & mask)); |
| 1535 | |
| 1536 | spin_unlock_irqrestore(&rsp->onofflock, flags); |
| 1537 | } |
| 1538 | |
| 1539 | static void __cpuinit rcu_online_cpu(int cpu) |
| 1540 | { |
| 1541 | rcu_init_percpu_data(cpu, &rcu_sched_state, 0); |
| 1542 | rcu_init_percpu_data(cpu, &rcu_bh_state, 0); |
| 1543 | rcu_preempt_init_percpu_data(cpu); |
| 1544 | } |
| 1545 | |
| 1546 | /* |
| 1547 | * Handle CPU online/offline notification events. |
| 1548 | */ |
| 1549 | int __cpuinit rcu_cpu_notify(struct notifier_block *self, |
| 1550 | unsigned long action, void *hcpu) |
| 1551 | { |
| 1552 | long cpu = (long)hcpu; |
| 1553 | |
| 1554 | switch (action) { |
| 1555 | case CPU_UP_PREPARE: |
| 1556 | case CPU_UP_PREPARE_FROZEN: |
| 1557 | rcu_online_cpu(cpu); |
| 1558 | break; |
| 1559 | case CPU_DOWN_PREPARE: |
| 1560 | case CPU_DOWN_PREPARE_FROZEN: |
| 1561 | /* Don't need to wait until next removal operation. */ |
| 1562 | /* rcu_migrate_head is protected by cpu_add_remove_lock */ |
| 1563 | wait_migrated_callbacks(); |
| 1564 | break; |
| 1565 | case CPU_DYING: |
| 1566 | case CPU_DYING_FROZEN: |
| 1567 | /* |
| 1568 | * preempt_disable() in on_each_cpu() prevents stop_machine(), |
| 1569 | * so when "on_each_cpu(rcu_barrier_func, (void *)type, 1);" |
| 1570 | * returns, all online cpus have queued rcu_barrier_func(), |
| 1571 | * and the dead cpu(if it exist) queues rcu_migrate_callback()s. |
| 1572 | * |
| 1573 | * These callbacks ensure _rcu_barrier() waits for all |
| 1574 | * RCU callbacks of the specified type to complete. |
| 1575 | */ |
| 1576 | atomic_set(&rcu_migrate_type_count, 3); |
| 1577 | call_rcu_bh(rcu_migrate_head, rcu_migrate_callback); |
| 1578 | call_rcu_sched(rcu_migrate_head + 1, rcu_migrate_callback); |
| 1579 | call_rcu(rcu_migrate_head + 2, rcu_migrate_callback); |
| 1580 | break; |
| 1581 | case CPU_DEAD: |
| 1582 | case CPU_DEAD_FROZEN: |
| 1583 | case CPU_UP_CANCELED: |
| 1584 | case CPU_UP_CANCELED_FROZEN: |
| 1585 | rcu_offline_cpu(cpu); |
| 1586 | break; |
| 1587 | default: |
| 1588 | break; |
| 1589 | } |
| 1590 | return NOTIFY_OK; |
| 1591 | } |
| 1592 | |
| 1593 | /* |
| 1594 | * Compute the per-level fanout, either using the exact fanout specified |
| 1595 | * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT. |
| 1596 | */ |
| 1597 | #ifdef CONFIG_RCU_FANOUT_EXACT |
| 1598 | static void __init rcu_init_levelspread(struct rcu_state *rsp) |
| 1599 | { |
| 1600 | int i; |
| 1601 | |
| 1602 | for (i = NUM_RCU_LVLS - 1; i >= 0; i--) |
| 1603 | rsp->levelspread[i] = CONFIG_RCU_FANOUT; |
| 1604 | } |
| 1605 | #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */ |
| 1606 | static void __init rcu_init_levelspread(struct rcu_state *rsp) |
| 1607 | { |
| 1608 | int ccur; |
| 1609 | int cprv; |
| 1610 | int i; |
| 1611 | |
| 1612 | cprv = NR_CPUS; |
| 1613 | for (i = NUM_RCU_LVLS - 1; i >= 0; i--) { |
| 1614 | ccur = rsp->levelcnt[i]; |
| 1615 | rsp->levelspread[i] = (cprv + ccur - 1) / ccur; |
| 1616 | cprv = ccur; |
| 1617 | } |
| 1618 | } |
| 1619 | #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */ |
| 1620 | |
| 1621 | /* |
| 1622 | * Helper function for rcu_init() that initializes one rcu_state structure. |
| 1623 | */ |
| 1624 | static void __init rcu_init_one(struct rcu_state *rsp) |
| 1625 | { |
| 1626 | int cpustride = 1; |
| 1627 | int i; |
| 1628 | int j; |
| 1629 | struct rcu_node *rnp; |
| 1630 | |
| 1631 | /* Initialize the level-tracking arrays. */ |
| 1632 | |
| 1633 | for (i = 1; i < NUM_RCU_LVLS; i++) |
| 1634 | rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1]; |
| 1635 | rcu_init_levelspread(rsp); |
| 1636 | |
| 1637 | /* Initialize the elements themselves, starting from the leaves. */ |
| 1638 | |
| 1639 | for (i = NUM_RCU_LVLS - 1; i >= 0; i--) { |
| 1640 | cpustride *= rsp->levelspread[i]; |
| 1641 | rnp = rsp->level[i]; |
| 1642 | for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) { |
| 1643 | spin_lock_init(&rnp->lock); |
| 1644 | rnp->gpnum = 0; |
| 1645 | rnp->qsmask = 0; |
| 1646 | rnp->qsmaskinit = 0; |
| 1647 | rnp->grplo = j * cpustride; |
| 1648 | rnp->grphi = (j + 1) * cpustride - 1; |
| 1649 | if (rnp->grphi >= NR_CPUS) |
| 1650 | rnp->grphi = NR_CPUS - 1; |
| 1651 | if (i == 0) { |
| 1652 | rnp->grpnum = 0; |
| 1653 | rnp->grpmask = 0; |
| 1654 | rnp->parent = NULL; |
| 1655 | } else { |
| 1656 | rnp->grpnum = j % rsp->levelspread[i - 1]; |
| 1657 | rnp->grpmask = 1UL << rnp->grpnum; |
| 1658 | rnp->parent = rsp->level[i - 1] + |
| 1659 | j / rsp->levelspread[i - 1]; |
| 1660 | } |
| 1661 | rnp->level = i; |
| 1662 | INIT_LIST_HEAD(&rnp->blocked_tasks[0]); |
| 1663 | INIT_LIST_HEAD(&rnp->blocked_tasks[1]); |
| 1664 | } |
| 1665 | } |
| 1666 | } |
| 1667 | |
| 1668 | /* |
| 1669 | * Helper macro for __rcu_init() and __rcu_init_preempt(). To be used |
| 1670 | * nowhere else! Assigns leaf node pointers into each CPU's rcu_data |
| 1671 | * structure. |
| 1672 | */ |
| 1673 | #define RCU_INIT_FLAVOR(rsp, rcu_data) \ |
| 1674 | do { \ |
| 1675 | int i; \ |
| 1676 | int j; \ |
| 1677 | struct rcu_node *rnp; \ |
| 1678 | \ |
| 1679 | rcu_init_one(rsp); \ |
| 1680 | rnp = (rsp)->level[NUM_RCU_LVLS - 1]; \ |
| 1681 | j = 0; \ |
| 1682 | for_each_possible_cpu(i) { \ |
| 1683 | if (i > rnp[j].grphi) \ |
| 1684 | j++; \ |
| 1685 | per_cpu(rcu_data, i).mynode = &rnp[j]; \ |
| 1686 | (rsp)->rda[i] = &per_cpu(rcu_data, i); \ |
| 1687 | rcu_boot_init_percpu_data(i, rsp); \ |
| 1688 | } \ |
| 1689 | } while (0) |
| 1690 | |
| 1691 | void __init __rcu_init(void) |
| 1692 | { |
| 1693 | rcu_bootup_announce(); |
| 1694 | #ifdef CONFIG_RCU_CPU_STALL_DETECTOR |
| 1695 | printk(KERN_INFO "RCU-based detection of stalled CPUs is enabled.\n"); |
| 1696 | #endif /* #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */ |
| 1697 | RCU_INIT_FLAVOR(&rcu_sched_state, rcu_sched_data); |
| 1698 | RCU_INIT_FLAVOR(&rcu_bh_state, rcu_bh_data); |
| 1699 | __rcu_init_preempt(); |
| 1700 | open_softirq(RCU_SOFTIRQ, rcu_process_callbacks); |
| 1701 | } |
| 1702 | |
| 1703 | #include "rcutree_plugin.h" |