2 * Read-Copy Update mechanism for mutual exclusion (tree-based version)
3 * Internal non-public definitions.
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License as published by
7 * the Free Software Foundation; either version 2 of the License, or
8 * (at your option) any later version.
10 * This program is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 * GNU General Public License for more details.
15 * You should have received a copy of the GNU General Public License
16 * along with this program; if not, write to the Free Software
17 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
19 * Copyright IBM Corporation, 2008
21 * Author: Ingo Molnar <mingo@elte.hu>
22 * Paul E. McKenney <paulmck@linux.vnet.ibm.com>
25 #include <linux/cache.h>
26 #include <linux/spinlock.h>
27 #include <linux/threads.h>
28 #include <linux/cpumask.h>
29 #include <linux/seqlock.h>
32 * Define shape of hierarchy based on NR_CPUS, CONFIG_RCU_FANOUT, and
33 * CONFIG_RCU_FANOUT_LEAF.
34 * In theory, it should be possible to add more levels straightforwardly.
35 * In practice, this did work well going from three levels to four.
36 * Of course, your mileage may vary.
38 #define MAX_RCU_LVLS 4
39 #define RCU_FANOUT_1 (CONFIG_RCU_FANOUT_LEAF)
40 #define RCU_FANOUT_2 (RCU_FANOUT_1 * CONFIG_RCU_FANOUT)
41 #define RCU_FANOUT_3 (RCU_FANOUT_2 * CONFIG_RCU_FANOUT)
42 #define RCU_FANOUT_4 (RCU_FANOUT_3 * CONFIG_RCU_FANOUT)
44 #if NR_CPUS <= RCU_FANOUT_1
45 # define NUM_RCU_LVLS 1
46 # define NUM_RCU_LVL_0 1
47 # define NUM_RCU_LVL_1 (NR_CPUS)
48 # define NUM_RCU_LVL_2 0
49 # define NUM_RCU_LVL_3 0
50 # define NUM_RCU_LVL_4 0
51 #elif NR_CPUS <= RCU_FANOUT_2
52 # define NUM_RCU_LVLS 2
53 # define NUM_RCU_LVL_0 1
54 # define NUM_RCU_LVL_1 DIV_ROUND_UP(NR_CPUS, RCU_FANOUT_1)
55 # define NUM_RCU_LVL_2 (NR_CPUS)
56 # define NUM_RCU_LVL_3 0
57 # define NUM_RCU_LVL_4 0
58 #elif NR_CPUS <= RCU_FANOUT_3
59 # define NUM_RCU_LVLS 3
60 # define NUM_RCU_LVL_0 1
61 # define NUM_RCU_LVL_1 DIV_ROUND_UP(NR_CPUS, RCU_FANOUT_2)
62 # define NUM_RCU_LVL_2 DIV_ROUND_UP(NR_CPUS, RCU_FANOUT_1)
63 # define NUM_RCU_LVL_3 (NR_CPUS)
64 # define NUM_RCU_LVL_4 0
65 #elif NR_CPUS <= RCU_FANOUT_4
66 # define NUM_RCU_LVLS 4
67 # define NUM_RCU_LVL_0 1
68 # define NUM_RCU_LVL_1 DIV_ROUND_UP(NR_CPUS, RCU_FANOUT_3)
69 # define NUM_RCU_LVL_2 DIV_ROUND_UP(NR_CPUS, RCU_FANOUT_2)
70 # define NUM_RCU_LVL_3 DIV_ROUND_UP(NR_CPUS, RCU_FANOUT_1)
71 # define NUM_RCU_LVL_4 (NR_CPUS)
73 # error "CONFIG_RCU_FANOUT insufficient for NR_CPUS"
74 #endif /* #if (NR_CPUS) <= RCU_FANOUT_1 */
76 #define RCU_SUM (NUM_RCU_LVL_0 + NUM_RCU_LVL_1 + NUM_RCU_LVL_2 + NUM_RCU_LVL_3 + NUM_RCU_LVL_4)
77 #define NUM_RCU_NODES (RCU_SUM - NR_CPUS)
80 * Dynticks per-CPU state.
83 long long dynticks_nesting; /* Track irq/process nesting level. */
84 /* Process level is worth LLONG_MAX/2. */
85 int dynticks_nmi_nesting; /* Track NMI nesting level. */
86 atomic_t dynticks; /* Even value for idle, else odd. */
87 #ifdef CONFIG_RCU_FAST_NO_HZ
88 int dyntick_drain; /* Prepare-for-idle state variable. */
89 unsigned long dyntick_holdoff;
90 /* No retries for the jiffy of failure. */
91 struct timer_list idle_gp_timer;
92 /* Wake up CPU sleeping with callbacks. */
93 unsigned long idle_gp_timer_expires;
94 /* When to wake up CPU (for repost). */
95 bool idle_first_pass; /* First pass of attempt to go idle? */
96 unsigned long nonlazy_posted;
97 /* # times non-lazy CBs posted to CPU. */
98 unsigned long nonlazy_posted_snap;
99 /* idle-period nonlazy_posted snapshot. */
100 #endif /* #ifdef CONFIG_RCU_FAST_NO_HZ */
103 /* RCU's kthread states for tracing. */
104 #define RCU_KTHREAD_STOPPED 0
105 #define RCU_KTHREAD_RUNNING 1
106 #define RCU_KTHREAD_WAITING 2
107 #define RCU_KTHREAD_OFFCPU 3
108 #define RCU_KTHREAD_YIELDING 4
109 #define RCU_KTHREAD_MAX 4
112 * Definition for node within the RCU grace-period-detection hierarchy.
115 raw_spinlock_t lock; /* Root rcu_node's lock protects some */
116 /* rcu_state fields as well as following. */
117 unsigned long gpnum; /* Current grace period for this node. */
118 /* This will either be equal to or one */
119 /* behind the root rcu_node's gpnum. */
120 unsigned long completed; /* Last GP completed for this node. */
121 /* This will either be equal to or one */
122 /* behind the root rcu_node's gpnum. */
123 unsigned long qsmask; /* CPUs or groups that need to switch in */
124 /* order for current grace period to proceed.*/
125 /* In leaf rcu_node, each bit corresponds to */
126 /* an rcu_data structure, otherwise, each */
127 /* bit corresponds to a child rcu_node */
129 unsigned long expmask; /* Groups that have ->blkd_tasks */
130 /* elements that need to drain to allow the */
131 /* current expedited grace period to */
132 /* complete (only for TREE_PREEMPT_RCU). */
133 atomic_t wakemask; /* CPUs whose kthread needs to be awakened. */
134 /* Since this has meaning only for leaf */
135 /* rcu_node structures, 32 bits suffices. */
136 unsigned long qsmaskinit;
137 /* Per-GP initial value for qsmask & expmask. */
138 unsigned long grpmask; /* Mask to apply to parent qsmask. */
139 /* Only one bit will be set in this mask. */
140 int grplo; /* lowest-numbered CPU or group here. */
141 int grphi; /* highest-numbered CPU or group here. */
142 u8 grpnum; /* CPU/group number for next level up. */
143 u8 level; /* root is at level 0. */
144 struct rcu_node *parent;
145 struct list_head blkd_tasks;
146 /* Tasks blocked in RCU read-side critical */
147 /* section. Tasks are placed at the head */
148 /* of this list and age towards the tail. */
149 struct list_head *gp_tasks;
150 /* Pointer to the first task blocking the */
151 /* current grace period, or NULL if there */
152 /* is no such task. */
153 struct list_head *exp_tasks;
154 /* Pointer to the first task blocking the */
155 /* current expedited grace period, or NULL */
156 /* if there is no such task. If there */
157 /* is no current expedited grace period, */
158 /* then there can cannot be any such task. */
159 #ifdef CONFIG_RCU_BOOST
160 struct list_head *boost_tasks;
161 /* Pointer to first task that needs to be */
162 /* priority boosted, or NULL if no priority */
163 /* boosting is needed for this rcu_node */
164 /* structure. If there are no tasks */
165 /* queued on this rcu_node structure that */
166 /* are blocking the current grace period, */
167 /* there can be no such task. */
168 unsigned long boost_time;
169 /* When to start boosting (jiffies). */
170 struct task_struct *boost_kthread_task;
171 /* kthread that takes care of priority */
172 /* boosting for this rcu_node structure. */
173 unsigned int boost_kthread_status;
174 /* State of boost_kthread_task for tracing. */
175 unsigned long n_tasks_boosted;
176 /* Total number of tasks boosted. */
177 unsigned long n_exp_boosts;
178 /* Number of tasks boosted for expedited GP. */
179 unsigned long n_normal_boosts;
180 /* Number of tasks boosted for normal GP. */
181 unsigned long n_balk_blkd_tasks;
182 /* Refused to boost: no blocked tasks. */
183 unsigned long n_balk_exp_gp_tasks;
184 /* Refused to boost: nothing blocking GP. */
185 unsigned long n_balk_boost_tasks;
186 /* Refused to boost: already boosting. */
187 unsigned long n_balk_notblocked;
188 /* Refused to boost: RCU RS CS still running. */
189 unsigned long n_balk_notyet;
190 /* Refused to boost: not yet time. */
191 unsigned long n_balk_nos;
192 /* Refused to boost: not sure why, though. */
193 /* This can happen due to race conditions. */
194 #endif /* #ifdef CONFIG_RCU_BOOST */
195 struct task_struct *node_kthread_task;
196 /* kthread that takes care of this rcu_node */
197 /* structure, for example, awakening the */
198 /* per-CPU kthreads as needed. */
199 unsigned int node_kthread_status;
200 /* State of node_kthread_task for tracing. */
201 } ____cacheline_internodealigned_in_smp;
204 * Do a full breadth-first scan of the rcu_node structures for the
205 * specified rcu_state structure.
207 #define rcu_for_each_node_breadth_first(rsp, rnp) \
208 for ((rnp) = &(rsp)->node[0]; \
209 (rnp) < &(rsp)->node[NUM_RCU_NODES]; (rnp)++)
212 * Do a breadth-first scan of the non-leaf rcu_node structures for the
213 * specified rcu_state structure. Note that if there is a singleton
214 * rcu_node tree with but one rcu_node structure, this loop is a no-op.
216 #define rcu_for_each_nonleaf_node_breadth_first(rsp, rnp) \
217 for ((rnp) = &(rsp)->node[0]; \
218 (rnp) < (rsp)->level[NUM_RCU_LVLS - 1]; (rnp)++)
221 * Scan the leaves of the rcu_node hierarchy for the specified rcu_state
222 * structure. Note that if there is a singleton rcu_node tree with but
223 * one rcu_node structure, this loop -will- visit the rcu_node structure.
224 * It is still a leaf node, even if it is also the root node.
226 #define rcu_for_each_leaf_node(rsp, rnp) \
227 for ((rnp) = (rsp)->level[NUM_RCU_LVLS - 1]; \
228 (rnp) < &(rsp)->node[NUM_RCU_NODES]; (rnp)++)
230 /* Index values for nxttail array in struct rcu_data. */
231 #define RCU_DONE_TAIL 0 /* Also RCU_WAIT head. */
232 #define RCU_WAIT_TAIL 1 /* Also RCU_NEXT_READY head. */
233 #define RCU_NEXT_READY_TAIL 2 /* Also RCU_NEXT head. */
234 #define RCU_NEXT_TAIL 3
235 #define RCU_NEXT_SIZE 4
237 /* Per-CPU data for read-copy update. */
239 /* 1) quiescent-state and grace-period handling : */
240 unsigned long completed; /* Track rsp->completed gp number */
241 /* in order to detect GP end. */
242 unsigned long gpnum; /* Highest gp number that this CPU */
243 /* is aware of having started. */
244 unsigned long passed_quiesce_gpnum;
245 /* gpnum at time of quiescent state. */
246 bool passed_quiesce; /* User-mode/idle loop etc. */
247 bool qs_pending; /* Core waits for quiesc state. */
248 bool beenonline; /* CPU online at least once. */
249 bool preemptible; /* Preemptible RCU? */
250 struct rcu_node *mynode; /* This CPU's leaf of hierarchy */
251 unsigned long grpmask; /* Mask to apply to leaf qsmask. */
252 #ifdef CONFIG_RCU_CPU_STALL_INFO
253 unsigned long ticks_this_gp; /* The number of scheduling-clock */
254 /* ticks this CPU has handled */
255 /* during and after the last grace */
256 /* period it is aware of. */
257 #endif /* #ifdef CONFIG_RCU_CPU_STALL_INFO */
259 /* 2) batch handling */
261 * If nxtlist is not NULL, it is partitioned as follows.
262 * Any of the partitions might be empty, in which case the
263 * pointer to that partition will be equal to the pointer for
264 * the following partition. When the list is empty, all of
265 * the nxttail elements point to the ->nxtlist pointer itself,
266 * which in that case is NULL.
268 * [nxtlist, *nxttail[RCU_DONE_TAIL]):
269 * Entries that batch # <= ->completed
270 * The grace period for these entries has completed, and
271 * the other grace-period-completed entries may be moved
272 * here temporarily in rcu_process_callbacks().
273 * [*nxttail[RCU_DONE_TAIL], *nxttail[RCU_WAIT_TAIL]):
274 * Entries that batch # <= ->completed - 1: waiting for current GP
275 * [*nxttail[RCU_WAIT_TAIL], *nxttail[RCU_NEXT_READY_TAIL]):
276 * Entries known to have arrived before current GP ended
277 * [*nxttail[RCU_NEXT_READY_TAIL], *nxttail[RCU_NEXT_TAIL]):
278 * Entries that might have arrived after current GP ended
279 * Note that the value of *nxttail[RCU_NEXT_TAIL] will
280 * always be NULL, as this is the end of the list.
282 struct rcu_head *nxtlist;
283 struct rcu_head **nxttail[RCU_NEXT_SIZE];
284 long qlen_lazy; /* # of lazy queued callbacks */
285 long qlen; /* # of queued callbacks, incl lazy */
286 long qlen_last_fqs_check;
287 /* qlen at last check for QS forcing */
288 unsigned long n_cbs_invoked; /* count of RCU cbs invoked. */
289 unsigned long n_cbs_orphaned; /* RCU cbs orphaned by dying CPU */
290 unsigned long n_cbs_adopted; /* RCU cbs adopted from dying CPU */
291 unsigned long n_force_qs_snap;
292 /* did other CPU force QS recently? */
293 long blimit; /* Upper limit on a processed batch */
295 /* 3) dynticks interface. */
296 struct rcu_dynticks *dynticks; /* Shared per-CPU dynticks state. */
297 int dynticks_snap; /* Per-GP tracking for dynticks. */
299 /* 4) reasons this CPU needed to be kicked by force_quiescent_state */
300 unsigned long dynticks_fqs; /* Kicked due to dynticks idle. */
301 unsigned long offline_fqs; /* Kicked due to being offline. */
303 /* 5) __rcu_pending() statistics. */
304 unsigned long n_rcu_pending; /* rcu_pending() calls since boot. */
305 unsigned long n_rp_qs_pending;
306 unsigned long n_rp_report_qs;
307 unsigned long n_rp_cb_ready;
308 unsigned long n_rp_cpu_needs_gp;
309 unsigned long n_rp_gp_completed;
310 unsigned long n_rp_gp_started;
311 unsigned long n_rp_need_fqs;
312 unsigned long n_rp_need_nothing;
315 struct rcu_state *rsp;
318 /* Values for fqs_state field in struct rcu_state. */
319 #define RCU_GP_IDLE 0 /* No grace period in progress. */
320 #define RCU_GP_INIT 1 /* Grace period being initialized. */
321 #define RCU_SAVE_DYNTICK 2 /* Need to scan dyntick state. */
322 #define RCU_FORCE_QS 3 /* Need to force quiescent state. */
323 #define RCU_SIGNAL_INIT RCU_SAVE_DYNTICK
325 #define RCU_JIFFIES_TILL_FORCE_QS 3 /* for rsp->jiffies_force_qs */
327 #ifdef CONFIG_PROVE_RCU
328 #define RCU_STALL_DELAY_DELTA (5 * HZ)
330 #define RCU_STALL_DELAY_DELTA 0
332 #define RCU_STALL_RAT_DELAY 2 /* Allow other CPUs time */
333 /* to take at least one */
334 /* scheduling clock irq */
335 /* before ratting on them. */
337 #define rcu_wait(cond) \
340 set_current_state(TASK_INTERRUPTIBLE); \
345 __set_current_state(TASK_RUNNING); \
349 * RCU global state, including node hierarchy. This hierarchy is
350 * represented in "heap" form in a dense array. The root (first level)
351 * of the hierarchy is in ->node[0] (referenced by ->level[0]), the second
352 * level in ->node[1] through ->node[m] (->node[1] referenced by ->level[1]),
353 * and the third level in ->node[m+1] and following (->node[m+1] referenced
354 * by ->level[2]). The number of levels is determined by the number of
355 * CPUs and by CONFIG_RCU_FANOUT. Small systems will have a "hierarchy"
356 * consisting of a single rcu_node.
359 struct rcu_node node[NUM_RCU_NODES]; /* Hierarchy. */
360 struct rcu_node *level[NUM_RCU_LVLS]; /* Hierarchy levels. */
361 u32 levelcnt[MAX_RCU_LVLS + 1]; /* # nodes in each level. */
362 u8 levelspread[NUM_RCU_LVLS]; /* kids/node in each level. */
363 struct rcu_data __percpu *rda; /* pointer of percu rcu_data. */
365 /* The following fields are guarded by the root rcu_node's lock. */
367 u8 fqs_state ____cacheline_internodealigned_in_smp;
368 /* Force QS state. */
369 u8 fqs_active; /* force_quiescent_state() */
371 u8 fqs_need_gp; /* A CPU was prevented from */
372 /* starting a new grace */
374 /* force_quiescent_state() */
376 u8 boost; /* Subject to priority boost. */
377 unsigned long gpnum; /* Current gp number. */
378 unsigned long completed; /* # of last completed gp. */
380 /* End of fields guarded by root rcu_node's lock. */
382 raw_spinlock_t onofflock; /* exclude on/offline and */
383 /* starting new GP. */
384 struct rcu_head *orphan_nxtlist; /* Orphaned callbacks that */
385 /* need a grace period. */
386 struct rcu_head **orphan_nxttail; /* Tail of above. */
387 struct rcu_head *orphan_donelist; /* Orphaned callbacks that */
388 /* are ready to invoke. */
389 struct rcu_head **orphan_donetail; /* Tail of above. */
390 long qlen_lazy; /* Number of lazy callbacks. */
391 long qlen; /* Total number of callbacks. */
392 struct task_struct *rcu_barrier_in_progress;
393 /* Task doing rcu_barrier(), */
394 /* or NULL if no barrier. */
395 raw_spinlock_t fqslock; /* Only one task forcing */
396 /* quiescent states. */
397 unsigned long jiffies_force_qs; /* Time at which to invoke */
398 /* force_quiescent_state(). */
399 unsigned long n_force_qs; /* Number of calls to */
400 /* force_quiescent_state(). */
401 unsigned long n_force_qs_lh; /* ~Number of calls leaving */
402 /* due to lock unavailable. */
403 unsigned long n_force_qs_ngp; /* Number of calls leaving */
404 /* due to no GP active. */
405 unsigned long gp_start; /* Time at which GP started, */
406 /* but in jiffies. */
407 unsigned long jiffies_stall; /* Time at which to check */
408 /* for CPU stalls. */
409 unsigned long gp_max; /* Maximum GP duration in */
411 char *name; /* Name of structure. */
414 /* Return values for rcu_preempt_offline_tasks(). */
416 #define RCU_OFL_TASKS_NORM_GP 0x1 /* Tasks blocking normal */
417 /* GP were moved to root. */
418 #define RCU_OFL_TASKS_EXP_GP 0x2 /* Tasks blocking expedited */
419 /* GP were moved to root. */
422 * RCU implementation internal declarations:
424 extern struct rcu_state rcu_sched_state;
425 DECLARE_PER_CPU(struct rcu_data, rcu_sched_data);
427 extern struct rcu_state rcu_bh_state;
428 DECLARE_PER_CPU(struct rcu_data, rcu_bh_data);
430 #ifdef CONFIG_TREE_PREEMPT_RCU
431 extern struct rcu_state rcu_preempt_state;
432 DECLARE_PER_CPU(struct rcu_data, rcu_preempt_data);
433 #endif /* #ifdef CONFIG_TREE_PREEMPT_RCU */
435 #ifdef CONFIG_RCU_BOOST
436 DECLARE_PER_CPU(unsigned int, rcu_cpu_kthread_status);
437 DECLARE_PER_CPU(int, rcu_cpu_kthread_cpu);
438 DECLARE_PER_CPU(unsigned int, rcu_cpu_kthread_loops);
439 DECLARE_PER_CPU(char, rcu_cpu_has_work);
440 #endif /* #ifdef CONFIG_RCU_BOOST */
442 #ifndef RCU_TREE_NONCORE
444 /* Forward declarations for rcutree_plugin.h */
445 static void rcu_bootup_announce(void);
446 long rcu_batches_completed(void);
447 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp);
448 #ifdef CONFIG_HOTPLUG_CPU
449 static void rcu_report_unblock_qs_rnp(struct rcu_node *rnp,
450 unsigned long flags);
451 static void rcu_stop_cpu_kthread(int cpu);
452 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
453 static void rcu_print_detail_task_stall(struct rcu_state *rsp);
454 static int rcu_print_task_stall(struct rcu_node *rnp);
455 static void rcu_preempt_stall_reset(void);
456 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp);
457 #ifdef CONFIG_HOTPLUG_CPU
458 static int rcu_preempt_offline_tasks(struct rcu_state *rsp,
459 struct rcu_node *rnp,
460 struct rcu_data *rdp);
461 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
462 static void rcu_preempt_cleanup_dead_cpu(int cpu);
463 static void rcu_preempt_check_callbacks(int cpu);
464 static void rcu_preempt_process_callbacks(void);
465 void call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu));
466 #if defined(CONFIG_HOTPLUG_CPU) || defined(CONFIG_TREE_PREEMPT_RCU)
467 static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp,
469 #endif /* #if defined(CONFIG_HOTPLUG_CPU) || defined(CONFIG_TREE_PREEMPT_RCU) */
470 static int rcu_preempt_pending(int cpu);
471 static int rcu_preempt_cpu_has_callbacks(int cpu);
472 static void __cpuinit rcu_preempt_init_percpu_data(int cpu);
473 static void rcu_preempt_cleanup_dying_cpu(void);
474 static void __init __rcu_init_preempt(void);
475 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags);
476 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp);
477 static void invoke_rcu_callbacks_kthread(void);
478 static bool rcu_is_callbacks_kthread(void);
479 #ifdef CONFIG_RCU_BOOST
480 static void rcu_preempt_do_callbacks(void);
481 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp,
483 static int __cpuinit rcu_spawn_one_boost_kthread(struct rcu_state *rsp,
484 struct rcu_node *rnp,
486 static void invoke_rcu_node_kthread(struct rcu_node *rnp);
487 static void rcu_yield(void (*f)(unsigned long), unsigned long arg);
488 #endif /* #ifdef CONFIG_RCU_BOOST */
489 static void rcu_cpu_kthread_setrt(int cpu, int to_rt);
490 static void __cpuinit rcu_prepare_kthreads(int cpu);
491 static void rcu_prepare_for_idle_init(int cpu);
492 static void rcu_cleanup_after_idle(int cpu);
493 static void rcu_prepare_for_idle(int cpu);
494 static void rcu_idle_count_callbacks_posted(void);
495 static void print_cpu_stall_info_begin(void);
496 static void print_cpu_stall_info(struct rcu_state *rsp, int cpu);
497 static void print_cpu_stall_info_end(void);
498 static void zero_cpu_stall_ticks(struct rcu_data *rdp);
499 static void increment_cpu_stall_ticks(void);
501 #endif /* #ifndef RCU_TREE_NONCORE */