xref: /illumos-gate/usr/src/uts/common/os/cpu.c (revision ef2504f26d1ea5859db9838255bb63f488f1b050)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright 2009 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 /*
27  * Architecture-independent CPU control functions.
28  */
29 
30 #include <sys/types.h>
31 #include <sys/param.h>
32 #include <sys/var.h>
33 #include <sys/thread.h>
34 #include <sys/cpuvar.h>
35 #include <sys/cpu_event.h>
36 #include <sys/kstat.h>
37 #include <sys/uadmin.h>
38 #include <sys/systm.h>
39 #include <sys/errno.h>
40 #include <sys/cmn_err.h>
41 #include <sys/procset.h>
42 #include <sys/processor.h>
43 #include <sys/debug.h>
44 #include <sys/cpupart.h>
45 #include <sys/lgrp.h>
46 #include <sys/pset.h>
47 #include <sys/pghw.h>
48 #include <sys/kmem.h>
49 #include <sys/kmem_impl.h>	/* to set per-cpu kmem_cache offset */
50 #include <sys/atomic.h>
51 #include <sys/callb.h>
52 #include <sys/vtrace.h>
53 #include <sys/cyclic.h>
54 #include <sys/bitmap.h>
55 #include <sys/nvpair.h>
56 #include <sys/pool_pset.h>
57 #include <sys/msacct.h>
58 #include <sys/time.h>
59 #include <sys/archsystm.h>
60 #include <sys/sdt.h>
61 #if defined(__x86) || defined(__amd64)
62 #include <sys/x86_archext.h>
63 #endif
64 #include <sys/callo.h>
65 
66 extern int	mp_cpu_start(cpu_t *);
67 extern int	mp_cpu_stop(cpu_t *);
68 extern int	mp_cpu_poweron(cpu_t *);
69 extern int	mp_cpu_poweroff(cpu_t *);
70 extern int	mp_cpu_configure(int);
71 extern int	mp_cpu_unconfigure(int);
72 extern void	mp_cpu_faulted_enter(cpu_t *);
73 extern void	mp_cpu_faulted_exit(cpu_t *);
74 
75 extern int cmp_cpu_to_chip(processorid_t cpuid);
76 #ifdef __sparcv9
77 extern char *cpu_fru_fmri(cpu_t *cp);
78 #endif
79 
80 static void cpu_add_active_internal(cpu_t *cp);
81 static void cpu_remove_active(cpu_t *cp);
82 static void cpu_info_kstat_create(cpu_t *cp);
83 static void cpu_info_kstat_destroy(cpu_t *cp);
84 static void cpu_stats_kstat_create(cpu_t *cp);
85 static void cpu_stats_kstat_destroy(cpu_t *cp);
86 
87 static int cpu_sys_stats_ks_update(kstat_t *ksp, int rw);
88 static int cpu_vm_stats_ks_update(kstat_t *ksp, int rw);
89 static int cpu_stat_ks_update(kstat_t *ksp, int rw);
90 static int cpu_state_change_hooks(int, cpu_setup_t, cpu_setup_t);
91 
92 /*
93  * cpu_lock protects ncpus, ncpus_online, cpu_flag, cpu_list, cpu_active,
94  * and dispatch queue reallocations.  The lock ordering with respect to
95  * related locks is:
96  *
97  *	cpu_lock --> thread_free_lock  --->  p_lock  --->  thread_lock()
98  *
99  * Warning:  Certain sections of code do not use the cpu_lock when
100  * traversing the cpu_list (e.g. mutex_vector_enter(), clock()).  Since
101  * all cpus are paused during modifications to this list, a solution
102  * to protect the list is too either disable kernel preemption while
103  * walking the list, *or* recheck the cpu_next pointer at each
104  * iteration in the loop.  Note that in no cases can any cached
105  * copies of the cpu pointers be kept as they may become invalid.
106  */
107 kmutex_t	cpu_lock;
108 cpu_t		*cpu_list;		/* list of all CPUs */
109 cpu_t		*clock_cpu_list;	/* used by clock to walk CPUs */
110 cpu_t		*cpu_active;		/* list of active CPUs */
111 static cpuset_t	cpu_available;		/* set of available CPUs */
112 cpuset_t	cpu_seqid_inuse;	/* which cpu_seqids are in use */
113 
114 cpu_t		**cpu_seq;		/* ptrs to CPUs, indexed by seq_id */
115 
116 /*
117  * max_ncpus keeps the max cpus the system can have. Initially
118  * it's NCPU, but since most archs scan the devtree for cpus
119  * fairly early on during boot, the real max can be known before
120  * ncpus is set (useful for early NCPU based allocations).
121  */
122 int max_ncpus = NCPU;
123 /*
124  * platforms that set max_ncpus to maxiumum number of cpus that can be
125  * dynamically added will set boot_max_ncpus to the number of cpus found
126  * at device tree scan time during boot.
127  */
128 int boot_max_ncpus = -1;
129 int boot_ncpus = -1;
130 /*
131  * Maximum possible CPU id.  This can never be >= NCPU since NCPU is
132  * used to size arrays that are indexed by CPU id.
133  */
134 processorid_t max_cpuid = NCPU - 1;
135 
136 int ncpus = 1;
137 int ncpus_online = 1;
138 
139 /*
140  * CPU that we're trying to offline.  Protected by cpu_lock.
141  */
142 cpu_t *cpu_inmotion;
143 
144 /*
145  * Can be raised to suppress further weakbinding, which are instead
146  * satisfied by disabling preemption.  Must be raised/lowered under cpu_lock,
147  * while individual thread weakbinding synchronization is done under thread
148  * lock.
149  */
150 int weakbindingbarrier;
151 
152 /*
153  * Variables used in pause_cpus().
154  */
155 static volatile char safe_list[NCPU];
156 
157 static struct _cpu_pause_info {
158 	int		cp_spl;		/* spl saved in pause_cpus() */
159 	volatile int	cp_go;		/* Go signal sent after all ready */
160 	int		cp_count;	/* # of CPUs to pause */
161 	ksema_t		cp_sem;		/* synch pause_cpus & cpu_pause */
162 	kthread_id_t	cp_paused;
163 } cpu_pause_info;
164 
165 static kmutex_t pause_free_mutex;
166 static kcondvar_t pause_free_cv;
167 
168 void *(*cpu_pause_func)(void *) = NULL;
169 
170 
171 static struct cpu_sys_stats_ks_data {
172 	kstat_named_t cpu_ticks_idle;
173 	kstat_named_t cpu_ticks_user;
174 	kstat_named_t cpu_ticks_kernel;
175 	kstat_named_t cpu_ticks_wait;
176 	kstat_named_t cpu_nsec_idle;
177 	kstat_named_t cpu_nsec_user;
178 	kstat_named_t cpu_nsec_kernel;
179 	kstat_named_t cpu_nsec_intr;
180 	kstat_named_t cpu_load_intr;
181 	kstat_named_t wait_ticks_io;
182 	kstat_named_t bread;
183 	kstat_named_t bwrite;
184 	kstat_named_t lread;
185 	kstat_named_t lwrite;
186 	kstat_named_t phread;
187 	kstat_named_t phwrite;
188 	kstat_named_t pswitch;
189 	kstat_named_t trap;
190 	kstat_named_t intr;
191 	kstat_named_t syscall;
192 	kstat_named_t sysread;
193 	kstat_named_t syswrite;
194 	kstat_named_t sysfork;
195 	kstat_named_t sysvfork;
196 	kstat_named_t sysexec;
197 	kstat_named_t readch;
198 	kstat_named_t writech;
199 	kstat_named_t rcvint;
200 	kstat_named_t xmtint;
201 	kstat_named_t mdmint;
202 	kstat_named_t rawch;
203 	kstat_named_t canch;
204 	kstat_named_t outch;
205 	kstat_named_t msg;
206 	kstat_named_t sema;
207 	kstat_named_t namei;
208 	kstat_named_t ufsiget;
209 	kstat_named_t ufsdirblk;
210 	kstat_named_t ufsipage;
211 	kstat_named_t ufsinopage;
212 	kstat_named_t procovf;
213 	kstat_named_t intrthread;
214 	kstat_named_t intrblk;
215 	kstat_named_t intrunpin;
216 	kstat_named_t idlethread;
217 	kstat_named_t inv_swtch;
218 	kstat_named_t nthreads;
219 	kstat_named_t cpumigrate;
220 	kstat_named_t xcalls;
221 	kstat_named_t mutex_adenters;
222 	kstat_named_t rw_rdfails;
223 	kstat_named_t rw_wrfails;
224 	kstat_named_t modload;
225 	kstat_named_t modunload;
226 	kstat_named_t bawrite;
227 	kstat_named_t iowait;
228 } cpu_sys_stats_ks_data_template = {
229 	{ "cpu_ticks_idle", 	KSTAT_DATA_UINT64 },
230 	{ "cpu_ticks_user", 	KSTAT_DATA_UINT64 },
231 	{ "cpu_ticks_kernel", 	KSTAT_DATA_UINT64 },
232 	{ "cpu_ticks_wait", 	KSTAT_DATA_UINT64 },
233 	{ "cpu_nsec_idle",	KSTAT_DATA_UINT64 },
234 	{ "cpu_nsec_user",	KSTAT_DATA_UINT64 },
235 	{ "cpu_nsec_kernel",	KSTAT_DATA_UINT64 },
236 	{ "cpu_nsec_intr",	KSTAT_DATA_UINT64 },
237 	{ "cpu_load_intr",	KSTAT_DATA_UINT64 },
238 	{ "wait_ticks_io", 	KSTAT_DATA_UINT64 },
239 	{ "bread", 		KSTAT_DATA_UINT64 },
240 	{ "bwrite", 		KSTAT_DATA_UINT64 },
241 	{ "lread", 		KSTAT_DATA_UINT64 },
242 	{ "lwrite", 		KSTAT_DATA_UINT64 },
243 	{ "phread", 		KSTAT_DATA_UINT64 },
244 	{ "phwrite", 		KSTAT_DATA_UINT64 },
245 	{ "pswitch", 		KSTAT_DATA_UINT64 },
246 	{ "trap", 		KSTAT_DATA_UINT64 },
247 	{ "intr", 		KSTAT_DATA_UINT64 },
248 	{ "syscall", 		KSTAT_DATA_UINT64 },
249 	{ "sysread", 		KSTAT_DATA_UINT64 },
250 	{ "syswrite", 		KSTAT_DATA_UINT64 },
251 	{ "sysfork", 		KSTAT_DATA_UINT64 },
252 	{ "sysvfork", 		KSTAT_DATA_UINT64 },
253 	{ "sysexec", 		KSTAT_DATA_UINT64 },
254 	{ "readch", 		KSTAT_DATA_UINT64 },
255 	{ "writech", 		KSTAT_DATA_UINT64 },
256 	{ "rcvint", 		KSTAT_DATA_UINT64 },
257 	{ "xmtint", 		KSTAT_DATA_UINT64 },
258 	{ "mdmint", 		KSTAT_DATA_UINT64 },
259 	{ "rawch", 		KSTAT_DATA_UINT64 },
260 	{ "canch", 		KSTAT_DATA_UINT64 },
261 	{ "outch", 		KSTAT_DATA_UINT64 },
262 	{ "msg", 		KSTAT_DATA_UINT64 },
263 	{ "sema", 		KSTAT_DATA_UINT64 },
264 	{ "namei", 		KSTAT_DATA_UINT64 },
265 	{ "ufsiget", 		KSTAT_DATA_UINT64 },
266 	{ "ufsdirblk", 		KSTAT_DATA_UINT64 },
267 	{ "ufsipage", 		KSTAT_DATA_UINT64 },
268 	{ "ufsinopage", 	KSTAT_DATA_UINT64 },
269 	{ "procovf", 		KSTAT_DATA_UINT64 },
270 	{ "intrthread", 	KSTAT_DATA_UINT64 },
271 	{ "intrblk", 		KSTAT_DATA_UINT64 },
272 	{ "intrunpin",		KSTAT_DATA_UINT64 },
273 	{ "idlethread", 	KSTAT_DATA_UINT64 },
274 	{ "inv_swtch", 		KSTAT_DATA_UINT64 },
275 	{ "nthreads", 		KSTAT_DATA_UINT64 },
276 	{ "cpumigrate", 	KSTAT_DATA_UINT64 },
277 	{ "xcalls", 		KSTAT_DATA_UINT64 },
278 	{ "mutex_adenters", 	KSTAT_DATA_UINT64 },
279 	{ "rw_rdfails", 	KSTAT_DATA_UINT64 },
280 	{ "rw_wrfails", 	KSTAT_DATA_UINT64 },
281 	{ "modload", 		KSTAT_DATA_UINT64 },
282 	{ "modunload", 		KSTAT_DATA_UINT64 },
283 	{ "bawrite", 		KSTAT_DATA_UINT64 },
284 	{ "iowait",		KSTAT_DATA_UINT64 },
285 };
286 
287 static struct cpu_vm_stats_ks_data {
288 	kstat_named_t pgrec;
289 	kstat_named_t pgfrec;
290 	kstat_named_t pgin;
291 	kstat_named_t pgpgin;
292 	kstat_named_t pgout;
293 	kstat_named_t pgpgout;
294 	kstat_named_t swapin;
295 	kstat_named_t pgswapin;
296 	kstat_named_t swapout;
297 	kstat_named_t pgswapout;
298 	kstat_named_t zfod;
299 	kstat_named_t dfree;
300 	kstat_named_t scan;
301 	kstat_named_t rev;
302 	kstat_named_t hat_fault;
303 	kstat_named_t as_fault;
304 	kstat_named_t maj_fault;
305 	kstat_named_t cow_fault;
306 	kstat_named_t prot_fault;
307 	kstat_named_t softlock;
308 	kstat_named_t kernel_asflt;
309 	kstat_named_t pgrrun;
310 	kstat_named_t execpgin;
311 	kstat_named_t execpgout;
312 	kstat_named_t execfree;
313 	kstat_named_t anonpgin;
314 	kstat_named_t anonpgout;
315 	kstat_named_t anonfree;
316 	kstat_named_t fspgin;
317 	kstat_named_t fspgout;
318 	kstat_named_t fsfree;
319 } cpu_vm_stats_ks_data_template = {
320 	{ "pgrec",		KSTAT_DATA_UINT64 },
321 	{ "pgfrec",		KSTAT_DATA_UINT64 },
322 	{ "pgin",		KSTAT_DATA_UINT64 },
323 	{ "pgpgin",		KSTAT_DATA_UINT64 },
324 	{ "pgout",		KSTAT_DATA_UINT64 },
325 	{ "pgpgout",		KSTAT_DATA_UINT64 },
326 	{ "swapin",		KSTAT_DATA_UINT64 },
327 	{ "pgswapin",		KSTAT_DATA_UINT64 },
328 	{ "swapout",		KSTAT_DATA_UINT64 },
329 	{ "pgswapout",		KSTAT_DATA_UINT64 },
330 	{ "zfod",		KSTAT_DATA_UINT64 },
331 	{ "dfree",		KSTAT_DATA_UINT64 },
332 	{ "scan",		KSTAT_DATA_UINT64 },
333 	{ "rev",		KSTAT_DATA_UINT64 },
334 	{ "hat_fault",		KSTAT_DATA_UINT64 },
335 	{ "as_fault",		KSTAT_DATA_UINT64 },
336 	{ "maj_fault",		KSTAT_DATA_UINT64 },
337 	{ "cow_fault",		KSTAT_DATA_UINT64 },
338 	{ "prot_fault",		KSTAT_DATA_UINT64 },
339 	{ "softlock",		KSTAT_DATA_UINT64 },
340 	{ "kernel_asflt",	KSTAT_DATA_UINT64 },
341 	{ "pgrrun",		KSTAT_DATA_UINT64 },
342 	{ "execpgin",		KSTAT_DATA_UINT64 },
343 	{ "execpgout",		KSTAT_DATA_UINT64 },
344 	{ "execfree",		KSTAT_DATA_UINT64 },
345 	{ "anonpgin",		KSTAT_DATA_UINT64 },
346 	{ "anonpgout",		KSTAT_DATA_UINT64 },
347 	{ "anonfree",		KSTAT_DATA_UINT64 },
348 	{ "fspgin",		KSTAT_DATA_UINT64 },
349 	{ "fspgout",		KSTAT_DATA_UINT64 },
350 	{ "fsfree",		KSTAT_DATA_UINT64 },
351 };
352 
353 /*
354  * Force the specified thread to migrate to the appropriate processor.
355  * Called with thread lock held, returns with it dropped.
356  */
357 static void
358 force_thread_migrate(kthread_id_t tp)
359 {
360 	ASSERT(THREAD_LOCK_HELD(tp));
361 	if (tp == curthread) {
362 		THREAD_TRANSITION(tp);
363 		CL_SETRUN(tp);
364 		thread_unlock_nopreempt(tp);
365 		swtch();
366 	} else {
367 		if (tp->t_state == TS_ONPROC) {
368 			cpu_surrender(tp);
369 		} else if (tp->t_state == TS_RUN) {
370 			(void) dispdeq(tp);
371 			setbackdq(tp);
372 		}
373 		thread_unlock(tp);
374 	}
375 }
376 
377 /*
378  * Set affinity for a specified CPU.
379  * A reference count is incremented and the affinity is held until the
380  * reference count is decremented to zero by thread_affinity_clear().
381  * This is so regions of code requiring affinity can be nested.
382  * Caller needs to ensure that cpu_id remains valid, which can be
383  * done by holding cpu_lock across this call, unless the caller
384  * specifies CPU_CURRENT in which case the cpu_lock will be acquired
385  * by thread_affinity_set and CPU->cpu_id will be the target CPU.
386  */
387 void
388 thread_affinity_set(kthread_id_t t, int cpu_id)
389 {
390 	cpu_t		*cp;
391 	int		c;
392 
393 	ASSERT(!(t == curthread && t->t_weakbound_cpu != NULL));
394 
395 	if ((c = cpu_id) == CPU_CURRENT) {
396 		mutex_enter(&cpu_lock);
397 		cpu_id = CPU->cpu_id;
398 	}
399 	/*
400 	 * We should be asserting that cpu_lock is held here, but
401 	 * the NCA code doesn't acquire it.  The following assert
402 	 * should be uncommented when the NCA code is fixed.
403 	 *
404 	 * ASSERT(MUTEX_HELD(&cpu_lock));
405 	 */
406 	ASSERT((cpu_id >= 0) && (cpu_id < NCPU));
407 	cp = cpu[cpu_id];
408 	ASSERT(cp != NULL);		/* user must provide a good cpu_id */
409 	/*
410 	 * If there is already a hard affinity requested, and this affinity
411 	 * conflicts with that, panic.
412 	 */
413 	thread_lock(t);
414 	if (t->t_affinitycnt > 0 && t->t_bound_cpu != cp) {
415 		panic("affinity_set: setting %p but already bound to %p",
416 		    (void *)cp, (void *)t->t_bound_cpu);
417 	}
418 	t->t_affinitycnt++;
419 	t->t_bound_cpu = cp;
420 
421 	/*
422 	 * Make sure we're running on the right CPU.
423 	 */
424 	if (cp != t->t_cpu || t != curthread) {
425 		force_thread_migrate(t);	/* drops thread lock */
426 	} else {
427 		thread_unlock(t);
428 	}
429 
430 	if (c == CPU_CURRENT)
431 		mutex_exit(&cpu_lock);
432 }
433 
434 /*
435  *	Wrapper for backward compatibility.
436  */
437 void
438 affinity_set(int cpu_id)
439 {
440 	thread_affinity_set(curthread, cpu_id);
441 }
442 
443 /*
444  * Decrement the affinity reservation count and if it becomes zero,
445  * clear the CPU affinity for the current thread, or set it to the user's
446  * software binding request.
447  */
448 void
449 thread_affinity_clear(kthread_id_t t)
450 {
451 	register processorid_t binding;
452 
453 	thread_lock(t);
454 	if (--t->t_affinitycnt == 0) {
455 		if ((binding = t->t_bind_cpu) == PBIND_NONE) {
456 			/*
457 			 * Adjust disp_max_unbound_pri if necessary.
458 			 */
459 			disp_adjust_unbound_pri(t);
460 			t->t_bound_cpu = NULL;
461 			if (t->t_cpu->cpu_part != t->t_cpupart) {
462 				force_thread_migrate(t);
463 				return;
464 			}
465 		} else {
466 			t->t_bound_cpu = cpu[binding];
467 			/*
468 			 * Make sure the thread is running on the bound CPU.
469 			 */
470 			if (t->t_cpu != t->t_bound_cpu) {
471 				force_thread_migrate(t);
472 				return;		/* already dropped lock */
473 			}
474 		}
475 	}
476 	thread_unlock(t);
477 }
478 
479 /*
480  * Wrapper for backward compatibility.
481  */
482 void
483 affinity_clear(void)
484 {
485 	thread_affinity_clear(curthread);
486 }
487 
488 /*
489  * Weak cpu affinity.  Bind to the "current" cpu for short periods
490  * of time during which the thread must not block (but may be preempted).
491  * Use this instead of kpreempt_disable() when it is only "no migration"
492  * rather than "no preemption" semantics that are required - disabling
493  * preemption holds higher priority threads off of cpu and if the
494  * operation that is protected is more than momentary this is not good
495  * for realtime etc.
496  *
497  * Weakly bound threads will not prevent a cpu from being offlined -
498  * we'll only run them on the cpu to which they are weakly bound but
499  * (because they do not block) we'll always be able to move them on to
500  * another cpu at offline time if we give them just a short moment to
501  * run during which they will unbind.  To give a cpu a chance of offlining,
502  * however, we require a barrier to weak bindings that may be raised for a
503  * given cpu (offline/move code may set this and then wait a short time for
504  * existing weak bindings to drop); the cpu_inmotion pointer is that barrier.
505  *
506  * There are few restrictions on the calling context of thread_nomigrate.
507  * The caller must not hold the thread lock.  Calls may be nested.
508  *
509  * After weakbinding a thread must not perform actions that may block.
510  * In particular it must not call thread_affinity_set; calling that when
511  * already weakbound is nonsensical anyway.
512  *
513  * If curthread is prevented from migrating for other reasons
514  * (kernel preemption disabled; high pil; strongly bound; interrupt thread)
515  * then the weak binding will succeed even if this cpu is the target of an
516  * offline/move request.
517  */
518 void
519 thread_nomigrate(void)
520 {
521 	cpu_t *cp;
522 	kthread_id_t t = curthread;
523 
524 again:
525 	kpreempt_disable();
526 	cp = CPU;
527 
528 	/*
529 	 * A highlevel interrupt must not modify t_nomigrate or
530 	 * t_weakbound_cpu of the thread it has interrupted.  A lowlevel
531 	 * interrupt thread cannot migrate and we can avoid the
532 	 * thread_lock call below by short-circuiting here.  In either
533 	 * case we can just return since no migration is possible and
534 	 * the condition will persist (ie, when we test for these again
535 	 * in thread_allowmigrate they can't have changed).   Migration
536 	 * is also impossible if we're at or above DISP_LEVEL pil.
537 	 */
538 	if (CPU_ON_INTR(cp) || t->t_flag & T_INTR_THREAD ||
539 	    getpil() >= DISP_LEVEL) {
540 		kpreempt_enable();
541 		return;
542 	}
543 
544 	/*
545 	 * We must be consistent with existing weak bindings.  Since we
546 	 * may be interrupted between the increment of t_nomigrate and
547 	 * the store to t_weakbound_cpu below we cannot assume that
548 	 * t_weakbound_cpu will be set if t_nomigrate is.  Note that we
549 	 * cannot assert t_weakbound_cpu == t_bind_cpu since that is not
550 	 * always the case.
551 	 */
552 	if (t->t_nomigrate && t->t_weakbound_cpu && t->t_weakbound_cpu != cp) {
553 		if (!panicstr)
554 			panic("thread_nomigrate: binding to %p but already "
555 			    "bound to %p", (void *)cp,
556 			    (void *)t->t_weakbound_cpu);
557 	}
558 
559 	/*
560 	 * At this point we have preemption disabled and we don't yet hold
561 	 * the thread lock.  So it's possible that somebody else could
562 	 * set t_bind_cpu here and not be able to force us across to the
563 	 * new cpu (since we have preemption disabled).
564 	 */
565 	thread_lock(curthread);
566 
567 	/*
568 	 * If further weak bindings are being (temporarily) suppressed then
569 	 * we'll settle for disabling kernel preemption (which assures
570 	 * no migration provided the thread does not block which it is
571 	 * not allowed to if using thread_nomigrate).  We must remember
572 	 * this disposition so we can take appropriate action in
573 	 * thread_allowmigrate.  If this is a nested call and the
574 	 * thread is already weakbound then fall through as normal.
575 	 * We remember the decision to settle for kpreempt_disable through
576 	 * negative nesting counting in t_nomigrate.  Once a thread has had one
577 	 * weakbinding request satisfied in this way any further (nested)
578 	 * requests will continue to be satisfied in the same way,
579 	 * even if weak bindings have recommenced.
580 	 */
581 	if (t->t_nomigrate < 0 || weakbindingbarrier && t->t_nomigrate == 0) {
582 		--t->t_nomigrate;
583 		thread_unlock(curthread);
584 		return;		/* with kpreempt_disable still active */
585 	}
586 
587 	/*
588 	 * We hold thread_lock so t_bind_cpu cannot change.  We could,
589 	 * however, be running on a different cpu to which we are t_bound_cpu
590 	 * to (as explained above).  If we grant the weak binding request
591 	 * in that case then the dispatcher must favour our weak binding
592 	 * over our strong (in which case, just as when preemption is
593 	 * disabled, we can continue to run on a cpu other than the one to
594 	 * which we are strongbound; the difference in this case is that
595 	 * this thread can be preempted and so can appear on the dispatch
596 	 * queues of a cpu other than the one it is strongbound to).
597 	 *
598 	 * If the cpu we are running on does not appear to be a current
599 	 * offline target (we check cpu_inmotion to determine this - since
600 	 * we don't hold cpu_lock we may not see a recent store to that,
601 	 * so it's possible that we at times can grant a weak binding to a
602 	 * cpu that is an offline target, but that one request will not
603 	 * prevent the offline from succeeding) then we will always grant
604 	 * the weak binding request.  This includes the case above where
605 	 * we grant a weakbinding not commensurate with our strong binding.
606 	 *
607 	 * If our cpu does appear to be an offline target then we're inclined
608 	 * not to grant the weakbinding request just yet - we'd prefer to
609 	 * migrate to another cpu and grant the request there.  The
610 	 * exceptions are those cases where going through preemption code
611 	 * will not result in us changing cpu:
612 	 *
613 	 *	. interrupts have already bypassed this case (see above)
614 	 *	. we are already weakbound to this cpu (dispatcher code will
615 	 *	  always return us to the weakbound cpu)
616 	 *	. preemption was disabled even before we disabled it above
617 	 *	. we are strongbound to this cpu (if we're strongbound to
618 	 *	another and not yet running there the trip through the
619 	 *	dispatcher will move us to the strongbound cpu and we
620 	 *	will grant the weak binding there)
621 	 */
622 	if (cp != cpu_inmotion || t->t_nomigrate > 0 || t->t_preempt > 1 ||
623 	    t->t_bound_cpu == cp) {
624 		/*
625 		 * Don't be tempted to store to t_weakbound_cpu only on
626 		 * the first nested bind request - if we're interrupted
627 		 * after the increment of t_nomigrate and before the
628 		 * store to t_weakbound_cpu and the interrupt calls
629 		 * thread_nomigrate then the assertion in thread_allowmigrate
630 		 * would fail.
631 		 */
632 		t->t_nomigrate++;
633 		t->t_weakbound_cpu = cp;
634 		membar_producer();
635 		thread_unlock(curthread);
636 		/*
637 		 * Now that we have dropped the thread_lock another thread
638 		 * can set our t_weakbound_cpu, and will try to migrate us
639 		 * to the strongbound cpu (which will not be prevented by
640 		 * preemption being disabled since we're about to enable
641 		 * preemption).  We have granted the weakbinding to the current
642 		 * cpu, so again we are in the position that is is is possible
643 		 * that our weak and strong bindings differ.  Again this
644 		 * is catered for by dispatcher code which will favour our
645 		 * weak binding.
646 		 */
647 		kpreempt_enable();
648 	} else {
649 		/*
650 		 * Move to another cpu before granting the request by
651 		 * forcing this thread through preemption code.  When we
652 		 * get to set{front,back}dq called from CL_PREEMPT()
653 		 * cpu_choose() will be used to select a cpu to queue
654 		 * us on - that will see cpu_inmotion and take
655 		 * steps to avoid returning us to this cpu.
656 		 */
657 		cp->cpu_kprunrun = 1;
658 		thread_unlock(curthread);
659 		kpreempt_enable();	/* will call preempt() */
660 		goto again;
661 	}
662 }
663 
664 void
665 thread_allowmigrate(void)
666 {
667 	kthread_id_t t = curthread;
668 
669 	ASSERT(t->t_weakbound_cpu == CPU ||
670 	    (t->t_nomigrate < 0 && t->t_preempt > 0) ||
671 	    CPU_ON_INTR(CPU) || t->t_flag & T_INTR_THREAD ||
672 	    getpil() >= DISP_LEVEL);
673 
674 	if (CPU_ON_INTR(CPU) || (t->t_flag & T_INTR_THREAD) ||
675 	    getpil() >= DISP_LEVEL)
676 		return;
677 
678 	if (t->t_nomigrate < 0) {
679 		/*
680 		 * This thread was granted "weak binding" in the
681 		 * stronger form of kernel preemption disabling.
682 		 * Undo a level of nesting for both t_nomigrate
683 		 * and t_preempt.
684 		 */
685 		++t->t_nomigrate;
686 		kpreempt_enable();
687 	} else if (--t->t_nomigrate == 0) {
688 		/*
689 		 * Time to drop the weak binding.  We need to cater
690 		 * for the case where we're weakbound to a different
691 		 * cpu than that to which we're strongbound (a very
692 		 * temporary arrangement that must only persist until
693 		 * weak binding drops).  We don't acquire thread_lock
694 		 * here so even as this code executes t_bound_cpu
695 		 * may be changing.  So we disable preemption and
696 		 * a) in the case that t_bound_cpu changes while we
697 		 * have preemption disabled kprunrun will be set
698 		 * asynchronously, and b) if before disabling
699 		 * preemption we were already on a different cpu to
700 		 * our t_bound_cpu then we set kprunrun ourselves
701 		 * to force a trip through the dispatcher when
702 		 * preemption is enabled.
703 		 */
704 		kpreempt_disable();
705 		if (t->t_bound_cpu &&
706 		    t->t_weakbound_cpu != t->t_bound_cpu)
707 			CPU->cpu_kprunrun = 1;
708 		t->t_weakbound_cpu = NULL;
709 		membar_producer();
710 		kpreempt_enable();
711 	}
712 }
713 
714 /*
715  * weakbinding_stop can be used to temporarily cause weakbindings made
716  * with thread_nomigrate to be satisfied through the stronger action of
717  * kpreempt_disable.  weakbinding_start recommences normal weakbinding.
718  */
719 
720 void
721 weakbinding_stop(void)
722 {
723 	ASSERT(MUTEX_HELD(&cpu_lock));
724 	weakbindingbarrier = 1;
725 	membar_producer();	/* make visible before subsequent thread_lock */
726 }
727 
728 void
729 weakbinding_start(void)
730 {
731 	ASSERT(MUTEX_HELD(&cpu_lock));
732 	weakbindingbarrier = 0;
733 }
734 
735 void
736 null_xcall(void)
737 {
738 }
739 
740 /*
741  * This routine is called to place the CPUs in a safe place so that
742  * one of them can be taken off line or placed on line.  What we are
743  * trying to do here is prevent a thread from traversing the list
744  * of active CPUs while we are changing it or from getting placed on
745  * the run queue of a CPU that has just gone off line.  We do this by
746  * creating a thread with the highest possible prio for each CPU and
747  * having it call this routine.  The advantage of this method is that
748  * we can eliminate all checks for CPU_ACTIVE in the disp routines.
749  * This makes disp faster at the expense of making p_online() slower
750  * which is a good trade off.
751  */
752 static void
753 cpu_pause(int index)
754 {
755 	int s;
756 	struct _cpu_pause_info *cpi = &cpu_pause_info;
757 	volatile char *safe = &safe_list[index];
758 	long    lindex = index;
759 
760 	ASSERT((curthread->t_bound_cpu != NULL) || (*safe == PAUSE_DIE));
761 
762 	while (*safe != PAUSE_DIE) {
763 		*safe = PAUSE_READY;
764 		membar_enter();		/* make sure stores are flushed */
765 		sema_v(&cpi->cp_sem);	/* signal requesting thread */
766 
767 		/*
768 		 * Wait here until all pause threads are running.  That
769 		 * indicates that it's safe to do the spl.  Until
770 		 * cpu_pause_info.cp_go is set, we don't want to spl
771 		 * because that might block clock interrupts needed
772 		 * to preempt threads on other CPUs.
773 		 */
774 		while (cpi->cp_go == 0)
775 			;
776 		/*
777 		 * Even though we are at the highest disp prio, we need
778 		 * to block out all interrupts below LOCK_LEVEL so that
779 		 * an intr doesn't come in, wake up a thread, and call
780 		 * setbackdq/setfrontdq.
781 		 */
782 		s = splhigh();
783 		/*
784 		 * if cpu_pause_func() has been set then call it using
785 		 * index as the argument, currently only used by
786 		 * cpr_suspend_cpus().  This function is used as the
787 		 * code to execute on the "paused" cpu's when a machine
788 		 * comes out of a sleep state and CPU's were powered off.
789 		 * (could also be used for hotplugging CPU's).
790 		 */
791 		if (cpu_pause_func != NULL)
792 			(*cpu_pause_func)((void *)lindex);
793 
794 		mach_cpu_pause(safe);
795 
796 		splx(s);
797 		/*
798 		 * Waiting is at an end. Switch out of cpu_pause
799 		 * loop and resume useful work.
800 		 */
801 		swtch();
802 	}
803 
804 	mutex_enter(&pause_free_mutex);
805 	*safe = PAUSE_DEAD;
806 	cv_broadcast(&pause_free_cv);
807 	mutex_exit(&pause_free_mutex);
808 }
809 
810 /*
811  * Allow the cpus to start running again.
812  */
813 void
814 start_cpus()
815 {
816 	int i;
817 
818 	ASSERT(MUTEX_HELD(&cpu_lock));
819 	ASSERT(cpu_pause_info.cp_paused);
820 	cpu_pause_info.cp_paused = NULL;
821 	for (i = 0; i < NCPU; i++)
822 		safe_list[i] = PAUSE_IDLE;
823 	membar_enter();			/* make sure stores are flushed */
824 	affinity_clear();
825 	splx(cpu_pause_info.cp_spl);
826 	kpreempt_enable();
827 }
828 
829 /*
830  * Allocate a pause thread for a CPU.
831  */
832 static void
833 cpu_pause_alloc(cpu_t *cp)
834 {
835 	kthread_id_t	t;
836 	long		cpun = cp->cpu_id;
837 
838 	/*
839 	 * Note, v.v_nglobpris will not change value as long as I hold
840 	 * cpu_lock.
841 	 */
842 	t = thread_create(NULL, 0, cpu_pause, (void *)cpun,
843 	    0, &p0, TS_STOPPED, v.v_nglobpris - 1);
844 	thread_lock(t);
845 	t->t_bound_cpu = cp;
846 	t->t_disp_queue = cp->cpu_disp;
847 	t->t_affinitycnt = 1;
848 	t->t_preempt = 1;
849 	thread_unlock(t);
850 	cp->cpu_pause_thread = t;
851 	/*
852 	 * Registering a thread in the callback table is usually done
853 	 * in the initialization code of the thread.  In this
854 	 * case, we do it right after thread creation because the
855 	 * thread itself may never run, and we need to register the
856 	 * fact that it is safe for cpr suspend.
857 	 */
858 	CALLB_CPR_INIT_SAFE(t, "cpu_pause");
859 }
860 
861 /*
862  * Free a pause thread for a CPU.
863  */
864 static void
865 cpu_pause_free(cpu_t *cp)
866 {
867 	kthread_id_t	t;
868 	int		cpun = cp->cpu_id;
869 
870 	ASSERT(MUTEX_HELD(&cpu_lock));
871 	/*
872 	 * We have to get the thread and tell him to die.
873 	 */
874 	if ((t = cp->cpu_pause_thread) == NULL) {
875 		ASSERT(safe_list[cpun] == PAUSE_IDLE);
876 		return;
877 	}
878 	thread_lock(t);
879 	t->t_cpu = CPU;		/* disp gets upset if last cpu is quiesced. */
880 	t->t_bound_cpu = NULL;	/* Must un-bind; cpu may not be running. */
881 	t->t_pri = v.v_nglobpris - 1;
882 	ASSERT(safe_list[cpun] == PAUSE_IDLE);
883 	safe_list[cpun] = PAUSE_DIE;
884 	THREAD_TRANSITION(t);
885 	setbackdq(t);
886 	thread_unlock_nopreempt(t);
887 
888 	/*
889 	 * If we don't wait for the thread to actually die, it may try to
890 	 * run on the wrong cpu as part of an actual call to pause_cpus().
891 	 */
892 	mutex_enter(&pause_free_mutex);
893 	while (safe_list[cpun] != PAUSE_DEAD) {
894 		cv_wait(&pause_free_cv, &pause_free_mutex);
895 	}
896 	mutex_exit(&pause_free_mutex);
897 	safe_list[cpun] = PAUSE_IDLE;
898 
899 	cp->cpu_pause_thread = NULL;
900 }
901 
902 /*
903  * Initialize basic structures for pausing CPUs.
904  */
905 void
906 cpu_pause_init()
907 {
908 	sema_init(&cpu_pause_info.cp_sem, 0, NULL, SEMA_DEFAULT, NULL);
909 	/*
910 	 * Create initial CPU pause thread.
911 	 */
912 	cpu_pause_alloc(CPU);
913 }
914 
915 /*
916  * Start the threads used to pause another CPU.
917  */
918 static int
919 cpu_pause_start(processorid_t cpu_id)
920 {
921 	int	i;
922 	int	cpu_count = 0;
923 
924 	for (i = 0; i < NCPU; i++) {
925 		cpu_t		*cp;
926 		kthread_id_t	t;
927 
928 		cp = cpu[i];
929 		if (!CPU_IN_SET(cpu_available, i) || (i == cpu_id)) {
930 			safe_list[i] = PAUSE_WAIT;
931 			continue;
932 		}
933 
934 		/*
935 		 * Skip CPU if it is quiesced or not yet started.
936 		 */
937 		if ((cp->cpu_flags & (CPU_QUIESCED | CPU_READY)) != CPU_READY) {
938 			safe_list[i] = PAUSE_WAIT;
939 			continue;
940 		}
941 
942 		/*
943 		 * Start this CPU's pause thread.
944 		 */
945 		t = cp->cpu_pause_thread;
946 		thread_lock(t);
947 		/*
948 		 * Reset the priority, since nglobpris may have
949 		 * changed since the thread was created, if someone
950 		 * has loaded the RT (or some other) scheduling
951 		 * class.
952 		 */
953 		t->t_pri = v.v_nglobpris - 1;
954 		THREAD_TRANSITION(t);
955 		setbackdq(t);
956 		thread_unlock_nopreempt(t);
957 		++cpu_count;
958 	}
959 	return (cpu_count);
960 }
961 
962 
963 /*
964  * Pause all of the CPUs except the one we are on by creating a high
965  * priority thread bound to those CPUs.
966  *
967  * Note that one must be extremely careful regarding code
968  * executed while CPUs are paused.  Since a CPU may be paused
969  * while a thread scheduling on that CPU is holding an adaptive
970  * lock, code executed with CPUs paused must not acquire adaptive
971  * (or low-level spin) locks.  Also, such code must not block,
972  * since the thread that is supposed to initiate the wakeup may
973  * never run.
974  *
975  * With a few exceptions, the restrictions on code executed with CPUs
976  * paused match those for code executed at high-level interrupt
977  * context.
978  */
979 void
980 pause_cpus(cpu_t *off_cp)
981 {
982 	processorid_t	cpu_id;
983 	int		i;
984 	struct _cpu_pause_info	*cpi = &cpu_pause_info;
985 
986 	ASSERT(MUTEX_HELD(&cpu_lock));
987 	ASSERT(cpi->cp_paused == NULL);
988 	cpi->cp_count = 0;
989 	cpi->cp_go = 0;
990 	for (i = 0; i < NCPU; i++)
991 		safe_list[i] = PAUSE_IDLE;
992 	kpreempt_disable();
993 
994 	/*
995 	 * If running on the cpu that is going offline, get off it.
996 	 * This is so that it won't be necessary to rechoose a CPU
997 	 * when done.
998 	 */
999 	if (CPU == off_cp)
1000 		cpu_id = off_cp->cpu_next_part->cpu_id;
1001 	else
1002 		cpu_id = CPU->cpu_id;
1003 	affinity_set(cpu_id);
1004 
1005 	/*
1006 	 * Start the pause threads and record how many were started
1007 	 */
1008 	cpi->cp_count = cpu_pause_start(cpu_id);
1009 
1010 	/*
1011 	 * Now wait for all CPUs to be running the pause thread.
1012 	 */
1013 	while (cpi->cp_count > 0) {
1014 		/*
1015 		 * Spin reading the count without grabbing the disp
1016 		 * lock to make sure we don't prevent the pause
1017 		 * threads from getting the lock.
1018 		 */
1019 		while (sema_held(&cpi->cp_sem))
1020 			;
1021 		if (sema_tryp(&cpi->cp_sem))
1022 			--cpi->cp_count;
1023 	}
1024 	cpi->cp_go = 1;			/* all have reached cpu_pause */
1025 
1026 	/*
1027 	 * Now wait for all CPUs to spl. (Transition from PAUSE_READY
1028 	 * to PAUSE_WAIT.)
1029 	 */
1030 	for (i = 0; i < NCPU; i++) {
1031 		while (safe_list[i] != PAUSE_WAIT)
1032 			;
1033 	}
1034 	cpi->cp_spl = splhigh();	/* block dispatcher on this CPU */
1035 	cpi->cp_paused = curthread;
1036 }
1037 
1038 /*
1039  * Check whether the current thread has CPUs paused
1040  */
1041 int
1042 cpus_paused(void)
1043 {
1044 	if (cpu_pause_info.cp_paused != NULL) {
1045 		ASSERT(cpu_pause_info.cp_paused == curthread);
1046 		return (1);
1047 	}
1048 	return (0);
1049 }
1050 
1051 static cpu_t *
1052 cpu_get_all(processorid_t cpun)
1053 {
1054 	ASSERT(MUTEX_HELD(&cpu_lock));
1055 
1056 	if (cpun >= NCPU || cpun < 0 || !CPU_IN_SET(cpu_available, cpun))
1057 		return (NULL);
1058 	return (cpu[cpun]);
1059 }
1060 
1061 /*
1062  * Check whether cpun is a valid processor id and whether it should be
1063  * visible from the current zone. If it is, return a pointer to the
1064  * associated CPU structure.
1065  */
1066 cpu_t *
1067 cpu_get(processorid_t cpun)
1068 {
1069 	cpu_t *c;
1070 
1071 	ASSERT(MUTEX_HELD(&cpu_lock));
1072 	c = cpu_get_all(cpun);
1073 	if (c != NULL && !INGLOBALZONE(curproc) && pool_pset_enabled() &&
1074 	    zone_pset_get(curproc->p_zone) != cpupart_query_cpu(c))
1075 		return (NULL);
1076 	return (c);
1077 }
1078 
1079 /*
1080  * The following functions should be used to check CPU states in the kernel.
1081  * They should be invoked with cpu_lock held.  Kernel subsystems interested
1082  * in CPU states should *not* use cpu_get_state() and various P_ONLINE/etc
1083  * states.  Those are for user-land (and system call) use only.
1084  */
1085 
1086 /*
1087  * Determine whether the CPU is online and handling interrupts.
1088  */
1089 int
1090 cpu_is_online(cpu_t *cpu)
1091 {
1092 	ASSERT(MUTEX_HELD(&cpu_lock));
1093 	return (cpu_flagged_online(cpu->cpu_flags));
1094 }
1095 
1096 /*
1097  * Determine whether the CPU is offline (this includes spare and faulted).
1098  */
1099 int
1100 cpu_is_offline(cpu_t *cpu)
1101 {
1102 	ASSERT(MUTEX_HELD(&cpu_lock));
1103 	return (cpu_flagged_offline(cpu->cpu_flags));
1104 }
1105 
1106 /*
1107  * Determine whether the CPU is powered off.
1108  */
1109 int
1110 cpu_is_poweredoff(cpu_t *cpu)
1111 {
1112 	ASSERT(MUTEX_HELD(&cpu_lock));
1113 	return (cpu_flagged_poweredoff(cpu->cpu_flags));
1114 }
1115 
1116 /*
1117  * Determine whether the CPU is handling interrupts.
1118  */
1119 int
1120 cpu_is_nointr(cpu_t *cpu)
1121 {
1122 	ASSERT(MUTEX_HELD(&cpu_lock));
1123 	return (cpu_flagged_nointr(cpu->cpu_flags));
1124 }
1125 
1126 /*
1127  * Determine whether the CPU is active (scheduling threads).
1128  */
1129 int
1130 cpu_is_active(cpu_t *cpu)
1131 {
1132 	ASSERT(MUTEX_HELD(&cpu_lock));
1133 	return (cpu_flagged_active(cpu->cpu_flags));
1134 }
1135 
1136 /*
1137  * Same as above, but these require cpu_flags instead of cpu_t pointers.
1138  */
1139 int
1140 cpu_flagged_online(cpu_flag_t cpu_flags)
1141 {
1142 	return (cpu_flagged_active(cpu_flags) &&
1143 	    (cpu_flags & CPU_ENABLE));
1144 }
1145 
1146 int
1147 cpu_flagged_offline(cpu_flag_t cpu_flags)
1148 {
1149 	return (((cpu_flags & CPU_POWEROFF) == 0) &&
1150 	    ((cpu_flags & (CPU_READY | CPU_OFFLINE)) != CPU_READY));
1151 }
1152 
1153 int
1154 cpu_flagged_poweredoff(cpu_flag_t cpu_flags)
1155 {
1156 	return ((cpu_flags & CPU_POWEROFF) == CPU_POWEROFF);
1157 }
1158 
1159 int
1160 cpu_flagged_nointr(cpu_flag_t cpu_flags)
1161 {
1162 	return (cpu_flagged_active(cpu_flags) &&
1163 	    (cpu_flags & CPU_ENABLE) == 0);
1164 }
1165 
1166 int
1167 cpu_flagged_active(cpu_flag_t cpu_flags)
1168 {
1169 	return (((cpu_flags & (CPU_POWEROFF | CPU_FAULTED | CPU_SPARE)) == 0) &&
1170 	    ((cpu_flags & (CPU_READY | CPU_OFFLINE)) == CPU_READY));
1171 }
1172 
1173 /*
1174  * Bring the indicated CPU online.
1175  */
1176 int
1177 cpu_online(cpu_t *cp)
1178 {
1179 	int	error = 0;
1180 
1181 	/*
1182 	 * Handle on-line request.
1183 	 *	This code must put the new CPU on the active list before
1184 	 *	starting it because it will not be paused, and will start
1185 	 * 	using the active list immediately.  The real start occurs
1186 	 *	when the CPU_QUIESCED flag is turned off.
1187 	 */
1188 
1189 	ASSERT(MUTEX_HELD(&cpu_lock));
1190 
1191 	/*
1192 	 * Put all the cpus into a known safe place.
1193 	 * No mutexes can be entered while CPUs are paused.
1194 	 */
1195 	error = mp_cpu_start(cp);	/* arch-dep hook */
1196 	if (error == 0) {
1197 		pg_cpupart_in(cp, cp->cpu_part);
1198 		pause_cpus(NULL);
1199 		cpu_add_active_internal(cp);
1200 		if (cp->cpu_flags & CPU_FAULTED) {
1201 			cp->cpu_flags &= ~CPU_FAULTED;
1202 			mp_cpu_faulted_exit(cp);
1203 		}
1204 		cp->cpu_flags &= ~(CPU_QUIESCED | CPU_OFFLINE | CPU_FROZEN |
1205 		    CPU_SPARE);
1206 		start_cpus();
1207 		cpu_stats_kstat_create(cp);
1208 		cpu_create_intrstat(cp);
1209 		lgrp_kstat_create(cp);
1210 		cpu_state_change_notify(cp->cpu_id, CPU_ON);
1211 		cpu_intr_enable(cp);	/* arch-dep hook */
1212 		cpu_set_state(cp);
1213 		cyclic_online(cp);
1214 		/*
1215 		 * This has to be called only after cyclic_online(). This
1216 		 * function uses cyclics.
1217 		 */
1218 		callout_cpu_online(cp);
1219 		poke_cpu(cp->cpu_id);
1220 	}
1221 
1222 	return (error);
1223 }
1224 
1225 /*
1226  * Take the indicated CPU offline.
1227  */
1228 int
1229 cpu_offline(cpu_t *cp, int flags)
1230 {
1231 	cpupart_t *pp;
1232 	int	error = 0;
1233 	cpu_t	*ncp;
1234 	int	intr_enable;
1235 	int	cyclic_off = 0;
1236 	int	callout_off = 0;
1237 	int	loop_count;
1238 	int	no_quiesce = 0;
1239 	int	(*bound_func)(struct cpu *, int);
1240 	kthread_t *t;
1241 	lpl_t	*cpu_lpl;
1242 	proc_t	*p;
1243 	int	lgrp_diff_lpl;
1244 	boolean_t unbind_all_threads = (flags & CPU_FORCED) != 0;
1245 
1246 	ASSERT(MUTEX_HELD(&cpu_lock));
1247 
1248 	/*
1249 	 * If we're going from faulted or spare to offline, just
1250 	 * clear these flags and update CPU state.
1251 	 */
1252 	if (cp->cpu_flags & (CPU_FAULTED | CPU_SPARE)) {
1253 		if (cp->cpu_flags & CPU_FAULTED) {
1254 			cp->cpu_flags &= ~CPU_FAULTED;
1255 			mp_cpu_faulted_exit(cp);
1256 		}
1257 		cp->cpu_flags &= ~CPU_SPARE;
1258 		cpu_set_state(cp);
1259 		return (0);
1260 	}
1261 
1262 	/*
1263 	 * Handle off-line request.
1264 	 */
1265 	pp = cp->cpu_part;
1266 	/*
1267 	 * Don't offline last online CPU in partition
1268 	 */
1269 	if (ncpus_online <= 1 || pp->cp_ncpus <= 1 || cpu_intr_count(cp) < 2)
1270 		return (EBUSY);
1271 	/*
1272 	 * Unbind all soft-bound threads bound to our CPU and hard bound threads
1273 	 * if we were asked to.
1274 	 */
1275 	error = cpu_unbind(cp->cpu_id, unbind_all_threads);
1276 	if (error != 0)
1277 		return (error);
1278 	/*
1279 	 * We shouldn't be bound to this CPU ourselves.
1280 	 */
1281 	if (curthread->t_bound_cpu == cp)
1282 		return (EBUSY);
1283 
1284 	/*
1285 	 * Tell interested parties that this CPU is going offline.
1286 	 */
1287 	cpu_state_change_notify(cp->cpu_id, CPU_OFF);
1288 
1289 	/*
1290 	 * Tell the PG subsystem that the CPU is leaving the partition
1291 	 */
1292 	pg_cpupart_out(cp, pp);
1293 
1294 	/*
1295 	 * Take the CPU out of interrupt participation so we won't find
1296 	 * bound kernel threads.  If the architecture cannot completely
1297 	 * shut off interrupts on the CPU, don't quiesce it, but don't
1298 	 * run anything but interrupt thread... this is indicated by
1299 	 * the CPU_OFFLINE flag being on but the CPU_QUIESCE flag being
1300 	 * off.
1301 	 */
1302 	intr_enable = cp->cpu_flags & CPU_ENABLE;
1303 	if (intr_enable)
1304 		no_quiesce = cpu_intr_disable(cp);
1305 
1306 	/*
1307 	 * Record that we are aiming to offline this cpu.  This acts as
1308 	 * a barrier to further weak binding requests in thread_nomigrate
1309 	 * and also causes cpu_choose, disp_lowpri_cpu and setfrontdq to
1310 	 * lean away from this cpu.  Further strong bindings are already
1311 	 * avoided since we hold cpu_lock.  Since threads that are set
1312 	 * runnable around now and others coming off the target cpu are
1313 	 * directed away from the target, existing strong and weak bindings
1314 	 * (especially the latter) to the target cpu stand maximum chance of
1315 	 * being able to unbind during the short delay loop below (if other
1316 	 * unbound threads compete they may not see cpu in time to unbind
1317 	 * even if they would do so immediately.
1318 	 */
1319 	cpu_inmotion = cp;
1320 	membar_enter();
1321 
1322 	/*
1323 	 * Check for kernel threads (strong or weak) bound to that CPU.
1324 	 * Strongly bound threads may not unbind, and we'll have to return
1325 	 * EBUSY.  Weakly bound threads should always disappear - we've
1326 	 * stopped more weak binding with cpu_inmotion and existing
1327 	 * bindings will drain imminently (they may not block).  Nonetheless
1328 	 * we will wait for a fixed period for all bound threads to disappear.
1329 	 * Inactive interrupt threads are OK (they'll be in TS_FREE
1330 	 * state).  If test finds some bound threads, wait a few ticks
1331 	 * to give short-lived threads (such as interrupts) chance to
1332 	 * complete.  Note that if no_quiesce is set, i.e. this cpu
1333 	 * is required to service interrupts, then we take the route
1334 	 * that permits interrupt threads to be active (or bypassed).
1335 	 */
1336 	bound_func = no_quiesce ? disp_bound_threads : disp_bound_anythreads;
1337 
1338 again:	for (loop_count = 0; (*bound_func)(cp, 0); loop_count++) {
1339 		if (loop_count >= 5) {
1340 			error = EBUSY;	/* some threads still bound */
1341 			break;
1342 		}
1343 
1344 		/*
1345 		 * If some threads were assigned, give them
1346 		 * a chance to complete or move.
1347 		 *
1348 		 * This assumes that the clock_thread is not bound
1349 		 * to any CPU, because the clock_thread is needed to
1350 		 * do the delay(hz/100).
1351 		 *
1352 		 * Note: we still hold the cpu_lock while waiting for
1353 		 * the next clock tick.  This is OK since it isn't
1354 		 * needed for anything else except processor_bind(2),
1355 		 * and system initialization.  If we drop the lock,
1356 		 * we would risk another p_online disabling the last
1357 		 * processor.
1358 		 */
1359 		delay(hz/100);
1360 	}
1361 
1362 	if (error == 0 && callout_off == 0) {
1363 		callout_cpu_offline(cp);
1364 		callout_off = 1;
1365 	}
1366 
1367 	if (error == 0 && cyclic_off == 0) {
1368 		if (!cyclic_offline(cp)) {
1369 			/*
1370 			 * We must have bound cyclics...
1371 			 */
1372 			error = EBUSY;
1373 			goto out;
1374 		}
1375 		cyclic_off = 1;
1376 	}
1377 
1378 	/*
1379 	 * Call mp_cpu_stop() to perform any special operations
1380 	 * needed for this machine architecture to offline a CPU.
1381 	 */
1382 	if (error == 0)
1383 		error = mp_cpu_stop(cp);	/* arch-dep hook */
1384 
1385 	/*
1386 	 * If that all worked, take the CPU offline and decrement
1387 	 * ncpus_online.
1388 	 */
1389 	if (error == 0) {
1390 		/*
1391 		 * Put all the cpus into a known safe place.
1392 		 * No mutexes can be entered while CPUs are paused.
1393 		 */
1394 		pause_cpus(cp);
1395 		/*
1396 		 * Repeat the operation, if necessary, to make sure that
1397 		 * all outstanding low-level interrupts run to completion
1398 		 * before we set the CPU_QUIESCED flag.  It's also possible
1399 		 * that a thread has weak bound to the cpu despite our raising
1400 		 * cpu_inmotion above since it may have loaded that
1401 		 * value before the barrier became visible (this would have
1402 		 * to be the thread that was on the target cpu at the time
1403 		 * we raised the barrier).
1404 		 */
1405 		if ((!no_quiesce && cp->cpu_intr_actv != 0) ||
1406 		    (*bound_func)(cp, 1)) {
1407 			start_cpus();
1408 			(void) mp_cpu_start(cp);
1409 			goto again;
1410 		}
1411 		ncp = cp->cpu_next_part;
1412 		cpu_lpl = cp->cpu_lpl;
1413 		ASSERT(cpu_lpl != NULL);
1414 
1415 		/*
1416 		 * Remove the CPU from the list of active CPUs.
1417 		 */
1418 		cpu_remove_active(cp);
1419 
1420 		/*
1421 		 * Walk the active process list and look for threads
1422 		 * whose home lgroup needs to be updated, or
1423 		 * the last CPU they run on is the one being offlined now.
1424 		 */
1425 
1426 		ASSERT(curthread->t_cpu != cp);
1427 		for (p = practive; p != NULL; p = p->p_next) {
1428 
1429 			t = p->p_tlist;
1430 
1431 			if (t == NULL)
1432 				continue;
1433 
1434 			lgrp_diff_lpl = 0;
1435 
1436 			do {
1437 				ASSERT(t->t_lpl != NULL);
1438 				/*
1439 				 * Taking last CPU in lpl offline
1440 				 * Rehome thread if it is in this lpl
1441 				 * Otherwise, update the count of how many
1442 				 * threads are in this CPU's lgroup but have
1443 				 * a different lpl.
1444 				 */
1445 
1446 				if (cpu_lpl->lpl_ncpu == 0) {
1447 					if (t->t_lpl == cpu_lpl)
1448 						lgrp_move_thread(t,
1449 						    lgrp_choose(t,
1450 						    t->t_cpupart), 0);
1451 					else if (t->t_lpl->lpl_lgrpid ==
1452 					    cpu_lpl->lpl_lgrpid)
1453 						lgrp_diff_lpl++;
1454 				}
1455 				ASSERT(t->t_lpl->lpl_ncpu > 0);
1456 
1457 				/*
1458 				 * Update CPU last ran on if it was this CPU
1459 				 */
1460 				if (t->t_cpu == cp && t->t_bound_cpu != cp)
1461 					t->t_cpu = disp_lowpri_cpu(ncp,
1462 					    t->t_lpl, t->t_pri, NULL);
1463 				ASSERT(t->t_cpu != cp || t->t_bound_cpu == cp ||
1464 				    t->t_weakbound_cpu == cp);
1465 
1466 				t = t->t_forw;
1467 			} while (t != p->p_tlist);
1468 
1469 			/*
1470 			 * Didn't find any threads in the same lgroup as this
1471 			 * CPU with a different lpl, so remove the lgroup from
1472 			 * the process lgroup bitmask.
1473 			 */
1474 
1475 			if (lgrp_diff_lpl == 0)
1476 				klgrpset_del(p->p_lgrpset, cpu_lpl->lpl_lgrpid);
1477 		}
1478 
1479 		/*
1480 		 * Walk thread list looking for threads that need to be
1481 		 * rehomed, since there are some threads that are not in
1482 		 * their process's p_tlist.
1483 		 */
1484 
1485 		t = curthread;
1486 		do {
1487 			ASSERT(t != NULL && t->t_lpl != NULL);
1488 
1489 			/*
1490 			 * Rehome threads with same lpl as this CPU when this
1491 			 * is the last CPU in the lpl.
1492 			 */
1493 
1494 			if ((cpu_lpl->lpl_ncpu == 0) && (t->t_lpl == cpu_lpl))
1495 				lgrp_move_thread(t,
1496 				    lgrp_choose(t, t->t_cpupart), 1);
1497 
1498 			ASSERT(t->t_lpl->lpl_ncpu > 0);
1499 
1500 			/*
1501 			 * Update CPU last ran on if it was this CPU
1502 			 */
1503 
1504 			if (t->t_cpu == cp && t->t_bound_cpu != cp) {
1505 				t->t_cpu = disp_lowpri_cpu(ncp,
1506 				    t->t_lpl, t->t_pri, NULL);
1507 			}
1508 			ASSERT(t->t_cpu != cp || t->t_bound_cpu == cp ||
1509 			    t->t_weakbound_cpu == cp);
1510 			t = t->t_next;
1511 
1512 		} while (t != curthread);
1513 		ASSERT((cp->cpu_flags & (CPU_FAULTED | CPU_SPARE)) == 0);
1514 		cp->cpu_flags |= CPU_OFFLINE;
1515 		disp_cpu_inactive(cp);
1516 		if (!no_quiesce)
1517 			cp->cpu_flags |= CPU_QUIESCED;
1518 		ncpus_online--;
1519 		cpu_set_state(cp);
1520 		cpu_inmotion = NULL;
1521 		start_cpus();
1522 		cpu_stats_kstat_destroy(cp);
1523 		cpu_delete_intrstat(cp);
1524 		lgrp_kstat_destroy(cp);
1525 	}
1526 
1527 out:
1528 	cpu_inmotion = NULL;
1529 
1530 	/*
1531 	 * If we failed, re-enable interrupts.
1532 	 * Do this even if cpu_intr_disable returned an error, because
1533 	 * it may have partially disabled interrupts.
1534 	 */
1535 	if (error && intr_enable)
1536 		cpu_intr_enable(cp);
1537 
1538 	/*
1539 	 * If we failed, but managed to offline the cyclic subsystem on this
1540 	 * CPU, bring it back online.
1541 	 */
1542 	if (error && cyclic_off)
1543 		cyclic_online(cp);
1544 
1545 	/*
1546 	 * If we failed, but managed to offline callouts on this CPU,
1547 	 * bring it back online.
1548 	 */
1549 	if (error && callout_off)
1550 		callout_cpu_online(cp);
1551 
1552 	/*
1553 	 * If we failed, tell the PG subsystem that the CPU is back
1554 	 */
1555 	pg_cpupart_in(cp, pp);
1556 
1557 	/*
1558 	 * If we failed, we need to notify everyone that this CPU is back on.
1559 	 */
1560 	if (error != 0)
1561 		cpu_state_change_notify(cp->cpu_id, CPU_ON);
1562 
1563 	return (error);
1564 }
1565 
1566 /*
1567  * Mark the indicated CPU as faulted, taking it offline.
1568  */
1569 int
1570 cpu_faulted(cpu_t *cp, int flags)
1571 {
1572 	int	error = 0;
1573 
1574 	ASSERT(MUTEX_HELD(&cpu_lock));
1575 	ASSERT(!cpu_is_poweredoff(cp));
1576 
1577 	if (cpu_is_offline(cp)) {
1578 		cp->cpu_flags &= ~CPU_SPARE;
1579 		cp->cpu_flags |= CPU_FAULTED;
1580 		mp_cpu_faulted_enter(cp);
1581 		cpu_set_state(cp);
1582 		return (0);
1583 	}
1584 
1585 	if ((error = cpu_offline(cp, flags)) == 0) {
1586 		cp->cpu_flags |= CPU_FAULTED;
1587 		mp_cpu_faulted_enter(cp);
1588 		cpu_set_state(cp);
1589 	}
1590 
1591 	return (error);
1592 }
1593 
1594 /*
1595  * Mark the indicated CPU as a spare, taking it offline.
1596  */
1597 int
1598 cpu_spare(cpu_t *cp, int flags)
1599 {
1600 	int	error = 0;
1601 
1602 	ASSERT(MUTEX_HELD(&cpu_lock));
1603 	ASSERT(!cpu_is_poweredoff(cp));
1604 
1605 	if (cpu_is_offline(cp)) {
1606 		if (cp->cpu_flags & CPU_FAULTED) {
1607 			cp->cpu_flags &= ~CPU_FAULTED;
1608 			mp_cpu_faulted_exit(cp);
1609 		}
1610 		cp->cpu_flags |= CPU_SPARE;
1611 		cpu_set_state(cp);
1612 		return (0);
1613 	}
1614 
1615 	if ((error = cpu_offline(cp, flags)) == 0) {
1616 		cp->cpu_flags |= CPU_SPARE;
1617 		cpu_set_state(cp);
1618 	}
1619 
1620 	return (error);
1621 }
1622 
1623 /*
1624  * Take the indicated CPU from poweroff to offline.
1625  */
1626 int
1627 cpu_poweron(cpu_t *cp)
1628 {
1629 	int	error = ENOTSUP;
1630 
1631 	ASSERT(MUTEX_HELD(&cpu_lock));
1632 	ASSERT(cpu_is_poweredoff(cp));
1633 
1634 	error = mp_cpu_poweron(cp);	/* arch-dep hook */
1635 	if (error == 0)
1636 		cpu_set_state(cp);
1637 
1638 	return (error);
1639 }
1640 
1641 /*
1642  * Take the indicated CPU from any inactive state to powered off.
1643  */
1644 int
1645 cpu_poweroff(cpu_t *cp)
1646 {
1647 	int	error = ENOTSUP;
1648 
1649 	ASSERT(MUTEX_HELD(&cpu_lock));
1650 	ASSERT(cpu_is_offline(cp));
1651 
1652 	if (!(cp->cpu_flags & CPU_QUIESCED))
1653 		return (EBUSY);		/* not completely idle */
1654 
1655 	error = mp_cpu_poweroff(cp);	/* arch-dep hook */
1656 	if (error == 0)
1657 		cpu_set_state(cp);
1658 
1659 	return (error);
1660 }
1661 
1662 /*
1663  * Initialize the Sequential CPU id lookup table
1664  */
1665 void
1666 cpu_seq_tbl_init()
1667 {
1668 	cpu_t	**tbl;
1669 
1670 	tbl = kmem_zalloc(sizeof (struct cpu *) * max_ncpus, KM_SLEEP);
1671 	tbl[0] = CPU;
1672 
1673 	cpu_seq = tbl;
1674 }
1675 
1676 /*
1677  * Initialize the CPU lists for the first CPU.
1678  */
1679 void
1680 cpu_list_init(cpu_t *cp)
1681 {
1682 	cp->cpu_next = cp;
1683 	cp->cpu_prev = cp;
1684 	cpu_list = cp;
1685 	clock_cpu_list = cp;
1686 
1687 	cp->cpu_next_onln = cp;
1688 	cp->cpu_prev_onln = cp;
1689 	cpu_active = cp;
1690 
1691 	cp->cpu_seqid = 0;
1692 	CPUSET_ADD(cpu_seqid_inuse, 0);
1693 
1694 	/*
1695 	 * Bootstrap cpu_seq using cpu_list
1696 	 * The cpu_seq[] table will be dynamically allocated
1697 	 * when kmem later becomes available (but before going MP)
1698 	 */
1699 	cpu_seq = &cpu_list;
1700 
1701 	cp->cpu_cache_offset = KMEM_CPU_CACHE_OFFSET(cp->cpu_seqid);
1702 	cp_default.cp_cpulist = cp;
1703 	cp_default.cp_ncpus = 1;
1704 	cp->cpu_next_part = cp;
1705 	cp->cpu_prev_part = cp;
1706 	cp->cpu_part = &cp_default;
1707 
1708 	CPUSET_ADD(cpu_available, cp->cpu_id);
1709 }
1710 
1711 /*
1712  * Insert a CPU into the list of available CPUs.
1713  */
1714 void
1715 cpu_add_unit(cpu_t *cp)
1716 {
1717 	int seqid;
1718 
1719 	ASSERT(MUTEX_HELD(&cpu_lock));
1720 	ASSERT(cpu_list != NULL);	/* list started in cpu_list_init */
1721 
1722 	lgrp_config(LGRP_CONFIG_CPU_ADD, (uintptr_t)cp, 0);
1723 
1724 	/*
1725 	 * Note: most users of the cpu_list will grab the
1726 	 * cpu_lock to insure that it isn't modified.  However,
1727 	 * certain users can't or won't do that.  To allow this
1728 	 * we pause the other cpus.  Users who walk the list
1729 	 * without cpu_lock, must disable kernel preemption
1730 	 * to insure that the list isn't modified underneath
1731 	 * them.  Also, any cached pointers to cpu structures
1732 	 * must be revalidated by checking to see if the
1733 	 * cpu_next pointer points to itself.  This check must
1734 	 * be done with the cpu_lock held or kernel preemption
1735 	 * disabled.  This check relies upon the fact that
1736 	 * old cpu structures are not free'ed or cleared after
1737 	 * then are removed from the cpu_list.
1738 	 *
1739 	 * Note that the clock code walks the cpu list dereferencing
1740 	 * the cpu_part pointer, so we need to initialize it before
1741 	 * adding the cpu to the list.
1742 	 */
1743 	cp->cpu_part = &cp_default;
1744 	(void) pause_cpus(NULL);
1745 	cp->cpu_next = cpu_list;
1746 	cp->cpu_prev = cpu_list->cpu_prev;
1747 	cpu_list->cpu_prev->cpu_next = cp;
1748 	cpu_list->cpu_prev = cp;
1749 	start_cpus();
1750 
1751 	for (seqid = 0; CPU_IN_SET(cpu_seqid_inuse, seqid); seqid++)
1752 		continue;
1753 	CPUSET_ADD(cpu_seqid_inuse, seqid);
1754 	cp->cpu_seqid = seqid;
1755 	ASSERT(ncpus < max_ncpus);
1756 	ncpus++;
1757 	cp->cpu_cache_offset = KMEM_CPU_CACHE_OFFSET(cp->cpu_seqid);
1758 	cpu[cp->cpu_id] = cp;
1759 	CPUSET_ADD(cpu_available, cp->cpu_id);
1760 	cpu_seq[cp->cpu_seqid] = cp;
1761 
1762 	/*
1763 	 * allocate a pause thread for this CPU.
1764 	 */
1765 	cpu_pause_alloc(cp);
1766 
1767 	/*
1768 	 * So that new CPUs won't have NULL prev_onln and next_onln pointers,
1769 	 * link them into a list of just that CPU.
1770 	 * This is so that disp_lowpri_cpu will work for thread_create in
1771 	 * pause_cpus() when called from the startup thread in a new CPU.
1772 	 */
1773 	cp->cpu_next_onln = cp;
1774 	cp->cpu_prev_onln = cp;
1775 	cpu_info_kstat_create(cp);
1776 	cp->cpu_next_part = cp;
1777 	cp->cpu_prev_part = cp;
1778 
1779 	init_cpu_mstate(cp, CMS_SYSTEM);
1780 
1781 	pool_pset_mod = gethrtime();
1782 }
1783 
1784 /*
1785  * Do the opposite of cpu_add_unit().
1786  */
1787 void
1788 cpu_del_unit(int cpuid)
1789 {
1790 	struct cpu	*cp, *cpnext;
1791 
1792 	ASSERT(MUTEX_HELD(&cpu_lock));
1793 	cp = cpu[cpuid];
1794 	ASSERT(cp != NULL);
1795 
1796 	ASSERT(cp->cpu_next_onln == cp);
1797 	ASSERT(cp->cpu_prev_onln == cp);
1798 	ASSERT(cp->cpu_next_part == cp);
1799 	ASSERT(cp->cpu_prev_part == cp);
1800 
1801 	/*
1802 	 * Tear down the CPU's physical ID cache, and update any
1803 	 * processor groups
1804 	 */
1805 	pg_cpu_fini(cp);
1806 	pghw_physid_destroy(cp);
1807 
1808 	/*
1809 	 * Destroy kstat stuff.
1810 	 */
1811 	cpu_info_kstat_destroy(cp);
1812 	term_cpu_mstate(cp);
1813 	/*
1814 	 * Free up pause thread.
1815 	 */
1816 	cpu_pause_free(cp);
1817 	CPUSET_DEL(cpu_available, cp->cpu_id);
1818 	cpu[cp->cpu_id] = NULL;
1819 	cpu_seq[cp->cpu_seqid] = NULL;
1820 
1821 	/*
1822 	 * The clock thread and mutex_vector_enter cannot hold the
1823 	 * cpu_lock while traversing the cpu list, therefore we pause
1824 	 * all other threads by pausing the other cpus. These, and any
1825 	 * other routines holding cpu pointers while possibly sleeping
1826 	 * must be sure to call kpreempt_disable before processing the
1827 	 * list and be sure to check that the cpu has not been deleted
1828 	 * after any sleeps (check cp->cpu_next != NULL). We guarantee
1829 	 * to keep the deleted cpu structure around.
1830 	 *
1831 	 * Note that this MUST be done AFTER cpu_available
1832 	 * has been updated so that we don't waste time
1833 	 * trying to pause the cpu we're trying to delete.
1834 	 */
1835 	(void) pause_cpus(NULL);
1836 
1837 	cpnext = cp->cpu_next;
1838 	cp->cpu_prev->cpu_next = cp->cpu_next;
1839 	cp->cpu_next->cpu_prev = cp->cpu_prev;
1840 	if (cp == cpu_list)
1841 		cpu_list = cpnext;
1842 
1843 	/*
1844 	 * Signals that the cpu has been deleted (see above).
1845 	 */
1846 	cp->cpu_next = NULL;
1847 	cp->cpu_prev = NULL;
1848 
1849 	start_cpus();
1850 
1851 	CPUSET_DEL(cpu_seqid_inuse, cp->cpu_seqid);
1852 	ncpus--;
1853 	lgrp_config(LGRP_CONFIG_CPU_DEL, (uintptr_t)cp, 0);
1854 
1855 	pool_pset_mod = gethrtime();
1856 }
1857 
1858 /*
1859  * Add a CPU to the list of active CPUs.
1860  *	This routine must not get any locks, because other CPUs are paused.
1861  */
1862 static void
1863 cpu_add_active_internal(cpu_t *cp)
1864 {
1865 	cpupart_t	*pp = cp->cpu_part;
1866 
1867 	ASSERT(MUTEX_HELD(&cpu_lock));
1868 	ASSERT(cpu_list != NULL);	/* list started in cpu_list_init */
1869 
1870 	ncpus_online++;
1871 	cpu_set_state(cp);
1872 	cp->cpu_next_onln = cpu_active;
1873 	cp->cpu_prev_onln = cpu_active->cpu_prev_onln;
1874 	cpu_active->cpu_prev_onln->cpu_next_onln = cp;
1875 	cpu_active->cpu_prev_onln = cp;
1876 
1877 	if (pp->cp_cpulist) {
1878 		cp->cpu_next_part = pp->cp_cpulist;
1879 		cp->cpu_prev_part = pp->cp_cpulist->cpu_prev_part;
1880 		pp->cp_cpulist->cpu_prev_part->cpu_next_part = cp;
1881 		pp->cp_cpulist->cpu_prev_part = cp;
1882 	} else {
1883 		ASSERT(pp->cp_ncpus == 0);
1884 		pp->cp_cpulist = cp->cpu_next_part = cp->cpu_prev_part = cp;
1885 	}
1886 	pp->cp_ncpus++;
1887 	if (pp->cp_ncpus == 1) {
1888 		cp_numparts_nonempty++;
1889 		ASSERT(cp_numparts_nonempty != 0);
1890 	}
1891 
1892 	pg_cpu_active(cp);
1893 	lgrp_config(LGRP_CONFIG_CPU_ONLINE, (uintptr_t)cp, 0);
1894 
1895 	bzero(&cp->cpu_loadavg, sizeof (cp->cpu_loadavg));
1896 }
1897 
1898 /*
1899  * Add a CPU to the list of active CPUs.
1900  *	This is called from machine-dependent layers when a new CPU is started.
1901  */
1902 void
1903 cpu_add_active(cpu_t *cp)
1904 {
1905 	pg_cpupart_in(cp, cp->cpu_part);
1906 
1907 	pause_cpus(NULL);
1908 	cpu_add_active_internal(cp);
1909 	start_cpus();
1910 
1911 	cpu_stats_kstat_create(cp);
1912 	cpu_create_intrstat(cp);
1913 	lgrp_kstat_create(cp);
1914 	cpu_state_change_notify(cp->cpu_id, CPU_INIT);
1915 }
1916 
1917 
1918 /*
1919  * Remove a CPU from the list of active CPUs.
1920  *	This routine must not get any locks, because other CPUs are paused.
1921  */
1922 /* ARGSUSED */
1923 static void
1924 cpu_remove_active(cpu_t *cp)
1925 {
1926 	cpupart_t	*pp = cp->cpu_part;
1927 
1928 	ASSERT(MUTEX_HELD(&cpu_lock));
1929 	ASSERT(cp->cpu_next_onln != cp);	/* not the last one */
1930 	ASSERT(cp->cpu_prev_onln != cp);	/* not the last one */
1931 
1932 	pg_cpu_inactive(cp);
1933 
1934 	lgrp_config(LGRP_CONFIG_CPU_OFFLINE, (uintptr_t)cp, 0);
1935 
1936 	if (cp == clock_cpu_list)
1937 		clock_cpu_list = cp->cpu_next_onln;
1938 
1939 	cp->cpu_prev_onln->cpu_next_onln = cp->cpu_next_onln;
1940 	cp->cpu_next_onln->cpu_prev_onln = cp->cpu_prev_onln;
1941 	if (cpu_active == cp) {
1942 		cpu_active = cp->cpu_next_onln;
1943 	}
1944 	cp->cpu_next_onln = cp;
1945 	cp->cpu_prev_onln = cp;
1946 
1947 	cp->cpu_prev_part->cpu_next_part = cp->cpu_next_part;
1948 	cp->cpu_next_part->cpu_prev_part = cp->cpu_prev_part;
1949 	if (pp->cp_cpulist == cp) {
1950 		pp->cp_cpulist = cp->cpu_next_part;
1951 		ASSERT(pp->cp_cpulist != cp);
1952 	}
1953 	cp->cpu_next_part = cp;
1954 	cp->cpu_prev_part = cp;
1955 	pp->cp_ncpus--;
1956 	if (pp->cp_ncpus == 0) {
1957 		cp_numparts_nonempty--;
1958 		ASSERT(cp_numparts_nonempty != 0);
1959 	}
1960 }
1961 
1962 /*
1963  * Routine used to setup a newly inserted CPU in preparation for starting
1964  * it running code.
1965  */
1966 int
1967 cpu_configure(int cpuid)
1968 {
1969 	int retval = 0;
1970 
1971 	ASSERT(MUTEX_HELD(&cpu_lock));
1972 
1973 	/*
1974 	 * Some structures are statically allocated based upon
1975 	 * the maximum number of cpus the system supports.  Do not
1976 	 * try to add anything beyond this limit.
1977 	 */
1978 	if (cpuid < 0 || cpuid >= NCPU) {
1979 		return (EINVAL);
1980 	}
1981 
1982 	if ((cpu[cpuid] != NULL) && (cpu[cpuid]->cpu_flags != 0)) {
1983 		return (EALREADY);
1984 	}
1985 
1986 	if ((retval = mp_cpu_configure(cpuid)) != 0) {
1987 		return (retval);
1988 	}
1989 
1990 	cpu[cpuid]->cpu_flags = CPU_QUIESCED | CPU_OFFLINE | CPU_POWEROFF;
1991 	cpu_set_state(cpu[cpuid]);
1992 	retval = cpu_state_change_hooks(cpuid, CPU_CONFIG, CPU_UNCONFIG);
1993 	if (retval != 0)
1994 		(void) mp_cpu_unconfigure(cpuid);
1995 
1996 	return (retval);
1997 }
1998 
1999 /*
2000  * Routine used to cleanup a CPU that has been powered off.  This will
2001  * destroy all per-cpu information related to this cpu.
2002  */
2003 int
2004 cpu_unconfigure(int cpuid)
2005 {
2006 	int error;
2007 
2008 	ASSERT(MUTEX_HELD(&cpu_lock));
2009 
2010 	if (cpu[cpuid] == NULL) {
2011 		return (ENODEV);
2012 	}
2013 
2014 	if (cpu[cpuid]->cpu_flags == 0) {
2015 		return (EALREADY);
2016 	}
2017 
2018 	if ((cpu[cpuid]->cpu_flags & CPU_POWEROFF) == 0) {
2019 		return (EBUSY);
2020 	}
2021 
2022 	if (cpu[cpuid]->cpu_props != NULL) {
2023 		(void) nvlist_free(cpu[cpuid]->cpu_props);
2024 		cpu[cpuid]->cpu_props = NULL;
2025 	}
2026 
2027 	error = cpu_state_change_hooks(cpuid, CPU_UNCONFIG, CPU_CONFIG);
2028 
2029 	if (error != 0)
2030 		return (error);
2031 
2032 	return (mp_cpu_unconfigure(cpuid));
2033 }
2034 
2035 /*
2036  * Routines for registering and de-registering cpu_setup callback functions.
2037  *
2038  * Caller's context
2039  *	These routines must not be called from a driver's attach(9E) or
2040  *	detach(9E) entry point.
2041  *
2042  * NOTE: CPU callbacks should not block. They are called with cpu_lock held.
2043  */
2044 
2045 /*
2046  * Ideally, these would be dynamically allocated and put into a linked
2047  * list; however that is not feasible because the registration routine
2048  * has to be available before the kmem allocator is working (in fact,
2049  * it is called by the kmem allocator init code).  In any case, there
2050  * are quite a few extra entries for future users.
2051  */
2052 #define	NCPU_SETUPS	20
2053 
2054 struct cpu_setup {
2055 	cpu_setup_func_t *func;
2056 	void *arg;
2057 } cpu_setups[NCPU_SETUPS];
2058 
2059 void
2060 register_cpu_setup_func(cpu_setup_func_t *func, void *arg)
2061 {
2062 	int i;
2063 
2064 	ASSERT(MUTEX_HELD(&cpu_lock));
2065 
2066 	for (i = 0; i < NCPU_SETUPS; i++)
2067 		if (cpu_setups[i].func == NULL)
2068 			break;
2069 	if (i >= NCPU_SETUPS)
2070 		cmn_err(CE_PANIC, "Ran out of cpu_setup callback entries");
2071 
2072 	cpu_setups[i].func = func;
2073 	cpu_setups[i].arg = arg;
2074 }
2075 
2076 void
2077 unregister_cpu_setup_func(cpu_setup_func_t *func, void *arg)
2078 {
2079 	int i;
2080 
2081 	ASSERT(MUTEX_HELD(&cpu_lock));
2082 
2083 	for (i = 0; i < NCPU_SETUPS; i++)
2084 		if ((cpu_setups[i].func == func) &&
2085 		    (cpu_setups[i].arg == arg))
2086 			break;
2087 	if (i >= NCPU_SETUPS)
2088 		cmn_err(CE_PANIC, "Could not find cpu_setup callback to "
2089 		    "deregister");
2090 
2091 	cpu_setups[i].func = NULL;
2092 	cpu_setups[i].arg = 0;
2093 }
2094 
2095 /*
2096  * Call any state change hooks for this CPU, ignore any errors.
2097  */
2098 void
2099 cpu_state_change_notify(int id, cpu_setup_t what)
2100 {
2101 	int i;
2102 
2103 	ASSERT(MUTEX_HELD(&cpu_lock));
2104 
2105 	for (i = 0; i < NCPU_SETUPS; i++) {
2106 		if (cpu_setups[i].func != NULL) {
2107 			cpu_setups[i].func(what, id, cpu_setups[i].arg);
2108 		}
2109 	}
2110 }
2111 
2112 /*
2113  * Call any state change hooks for this CPU, undo it if error found.
2114  */
2115 static int
2116 cpu_state_change_hooks(int id, cpu_setup_t what, cpu_setup_t undo)
2117 {
2118 	int i;
2119 	int retval = 0;
2120 
2121 	ASSERT(MUTEX_HELD(&cpu_lock));
2122 
2123 	for (i = 0; i < NCPU_SETUPS; i++) {
2124 		if (cpu_setups[i].func != NULL) {
2125 			retval = cpu_setups[i].func(what, id,
2126 			    cpu_setups[i].arg);
2127 			if (retval) {
2128 				for (i--; i >= 0; i--) {
2129 					if (cpu_setups[i].func != NULL)
2130 						cpu_setups[i].func(undo,
2131 						    id, cpu_setups[i].arg);
2132 				}
2133 				break;
2134 			}
2135 		}
2136 	}
2137 	return (retval);
2138 }
2139 
2140 /*
2141  * Export information about this CPU via the kstat mechanism.
2142  */
2143 static struct {
2144 	kstat_named_t ci_state;
2145 	kstat_named_t ci_state_begin;
2146 	kstat_named_t ci_cpu_type;
2147 	kstat_named_t ci_fpu_type;
2148 	kstat_named_t ci_clock_MHz;
2149 	kstat_named_t ci_chip_id;
2150 	kstat_named_t ci_implementation;
2151 	kstat_named_t ci_brandstr;
2152 	kstat_named_t ci_core_id;
2153 	kstat_named_t ci_curr_clock_Hz;
2154 	kstat_named_t ci_supp_freq_Hz;
2155 #if defined(__sparcv9)
2156 	kstat_named_t ci_device_ID;
2157 	kstat_named_t ci_cpu_fru;
2158 #endif
2159 #if defined(__x86)
2160 	kstat_named_t ci_vendorstr;
2161 	kstat_named_t ci_family;
2162 	kstat_named_t ci_model;
2163 	kstat_named_t ci_step;
2164 	kstat_named_t ci_clogid;
2165 	kstat_named_t ci_pkg_core_id;
2166 	kstat_named_t ci_ncpuperchip;
2167 	kstat_named_t ci_ncoreperchip;
2168 	kstat_named_t ci_max_cstates;
2169 	kstat_named_t ci_curr_cstate;
2170 	kstat_named_t ci_sktstr;
2171 #endif
2172 } cpu_info_template = {
2173 	{ "state",			KSTAT_DATA_CHAR },
2174 	{ "state_begin",		KSTAT_DATA_LONG },
2175 	{ "cpu_type",			KSTAT_DATA_CHAR },
2176 	{ "fpu_type",			KSTAT_DATA_CHAR },
2177 	{ "clock_MHz",			KSTAT_DATA_LONG },
2178 	{ "chip_id",			KSTAT_DATA_LONG },
2179 	{ "implementation",		KSTAT_DATA_STRING },
2180 	{ "brand",			KSTAT_DATA_STRING },
2181 	{ "core_id",			KSTAT_DATA_LONG },
2182 	{ "current_clock_Hz",		KSTAT_DATA_UINT64 },
2183 	{ "supported_frequencies_Hz",	KSTAT_DATA_STRING },
2184 #if defined(__sparcv9)
2185 	{ "device_ID",			KSTAT_DATA_UINT64 },
2186 	{ "cpu_fru",			KSTAT_DATA_STRING },
2187 #endif
2188 #if defined(__x86)
2189 	{ "vendor_id",			KSTAT_DATA_STRING },
2190 	{ "family",			KSTAT_DATA_INT32 },
2191 	{ "model",			KSTAT_DATA_INT32 },
2192 	{ "stepping",			KSTAT_DATA_INT32 },
2193 	{ "clog_id",			KSTAT_DATA_INT32 },
2194 	{ "pkg_core_id",		KSTAT_DATA_LONG },
2195 	{ "ncpu_per_chip",		KSTAT_DATA_INT32 },
2196 	{ "ncore_per_chip",		KSTAT_DATA_INT32 },
2197 	{ "supported_max_cstates",	KSTAT_DATA_INT32 },
2198 	{ "current_cstate",		KSTAT_DATA_INT32 },
2199 	{ "socket_type",		KSTAT_DATA_STRING },
2200 #endif
2201 };
2202 
2203 static kmutex_t cpu_info_template_lock;
2204 
2205 static int
2206 cpu_info_kstat_update(kstat_t *ksp, int rw)
2207 {
2208 	cpu_t	*cp = ksp->ks_private;
2209 	const char *pi_state;
2210 
2211 	if (rw == KSTAT_WRITE)
2212 		return (EACCES);
2213 
2214 	switch (cp->cpu_type_info.pi_state) {
2215 	case P_ONLINE:
2216 		pi_state = PS_ONLINE;
2217 		break;
2218 	case P_POWEROFF:
2219 		pi_state = PS_POWEROFF;
2220 		break;
2221 	case P_NOINTR:
2222 		pi_state = PS_NOINTR;
2223 		break;
2224 	case P_FAULTED:
2225 		pi_state = PS_FAULTED;
2226 		break;
2227 	case P_SPARE:
2228 		pi_state = PS_SPARE;
2229 		break;
2230 	case P_OFFLINE:
2231 		pi_state = PS_OFFLINE;
2232 		break;
2233 	default:
2234 		pi_state = "unknown";
2235 	}
2236 	(void) strcpy(cpu_info_template.ci_state.value.c, pi_state);
2237 	cpu_info_template.ci_state_begin.value.l = cp->cpu_state_begin;
2238 	(void) strncpy(cpu_info_template.ci_cpu_type.value.c,
2239 	    cp->cpu_type_info.pi_processor_type, 15);
2240 	(void) strncpy(cpu_info_template.ci_fpu_type.value.c,
2241 	    cp->cpu_type_info.pi_fputypes, 15);
2242 	cpu_info_template.ci_clock_MHz.value.l = cp->cpu_type_info.pi_clock;
2243 	cpu_info_template.ci_chip_id.value.l =
2244 	    pg_plat_hw_instance_id(cp, PGHW_CHIP);
2245 	kstat_named_setstr(&cpu_info_template.ci_implementation,
2246 	    cp->cpu_idstr);
2247 	kstat_named_setstr(&cpu_info_template.ci_brandstr, cp->cpu_brandstr);
2248 	cpu_info_template.ci_core_id.value.l = pg_plat_get_core_id(cp);
2249 	cpu_info_template.ci_curr_clock_Hz.value.ui64 =
2250 	    cp->cpu_curr_clock;
2251 	kstat_named_setstr(&cpu_info_template.ci_supp_freq_Hz,
2252 	    cp->cpu_supp_freqs);
2253 #if defined(__sparcv9)
2254 	cpu_info_template.ci_device_ID.value.ui64 =
2255 	    cpunodes[cp->cpu_id].device_id;
2256 	kstat_named_setstr(&cpu_info_template.ci_cpu_fru, cpu_fru_fmri(cp));
2257 #endif
2258 #if defined(__x86)
2259 	kstat_named_setstr(&cpu_info_template.ci_vendorstr,
2260 	    cpuid_getvendorstr(cp));
2261 	cpu_info_template.ci_family.value.l = cpuid_getfamily(cp);
2262 	cpu_info_template.ci_model.value.l = cpuid_getmodel(cp);
2263 	cpu_info_template.ci_step.value.l = cpuid_getstep(cp);
2264 	cpu_info_template.ci_clogid.value.l = cpuid_get_clogid(cp);
2265 	cpu_info_template.ci_ncpuperchip.value.l = cpuid_get_ncpu_per_chip(cp);
2266 	cpu_info_template.ci_ncoreperchip.value.l =
2267 	    cpuid_get_ncore_per_chip(cp);
2268 	cpu_info_template.ci_pkg_core_id.value.l = cpuid_get_pkgcoreid(cp);
2269 	cpu_info_template.ci_max_cstates.value.l = cp->cpu_m.max_cstates;
2270 	cpu_info_template.ci_curr_cstate.value.l = cpu_idle_get_cpu_state(cp);
2271 	kstat_named_setstr(&cpu_info_template.ci_sktstr,
2272 	    cpuid_getsocketstr(cp));
2273 #endif
2274 
2275 	return (0);
2276 }
2277 
2278 static void
2279 cpu_info_kstat_create(cpu_t *cp)
2280 {
2281 	zoneid_t zoneid;
2282 
2283 	ASSERT(MUTEX_HELD(&cpu_lock));
2284 
2285 	if (pool_pset_enabled())
2286 		zoneid = GLOBAL_ZONEID;
2287 	else
2288 		zoneid = ALL_ZONES;
2289 	if ((cp->cpu_info_kstat = kstat_create_zone("cpu_info", cp->cpu_id,
2290 	    NULL, "misc", KSTAT_TYPE_NAMED,
2291 	    sizeof (cpu_info_template) / sizeof (kstat_named_t),
2292 	    KSTAT_FLAG_VIRTUAL, zoneid)) != NULL) {
2293 		cp->cpu_info_kstat->ks_data_size += 2 * CPU_IDSTRLEN;
2294 #if defined(__sparcv9)
2295 		cp->cpu_info_kstat->ks_data_size +=
2296 		    strlen(cpu_fru_fmri(cp)) + 1;
2297 #endif
2298 #if defined(__x86)
2299 		cp->cpu_info_kstat->ks_data_size += X86_VENDOR_STRLEN;
2300 #endif
2301 		if (cp->cpu_supp_freqs != NULL)
2302 			cp->cpu_info_kstat->ks_data_size +=
2303 			    strlen(cp->cpu_supp_freqs) + 1;
2304 		cp->cpu_info_kstat->ks_lock = &cpu_info_template_lock;
2305 		cp->cpu_info_kstat->ks_data = &cpu_info_template;
2306 		cp->cpu_info_kstat->ks_private = cp;
2307 		cp->cpu_info_kstat->ks_update = cpu_info_kstat_update;
2308 		kstat_install(cp->cpu_info_kstat);
2309 	}
2310 }
2311 
2312 static void
2313 cpu_info_kstat_destroy(cpu_t *cp)
2314 {
2315 	ASSERT(MUTEX_HELD(&cpu_lock));
2316 
2317 	kstat_delete(cp->cpu_info_kstat);
2318 	cp->cpu_info_kstat = NULL;
2319 }
2320 
2321 /*
2322  * Create and install kstats for the boot CPU.
2323  */
2324 void
2325 cpu_kstat_init(cpu_t *cp)
2326 {
2327 	mutex_enter(&cpu_lock);
2328 	cpu_info_kstat_create(cp);
2329 	cpu_stats_kstat_create(cp);
2330 	cpu_create_intrstat(cp);
2331 	cpu_set_state(cp);
2332 	mutex_exit(&cpu_lock);
2333 }
2334 
2335 /*
2336  * Make visible to the zone that subset of the cpu information that would be
2337  * initialized when a cpu is configured (but still offline).
2338  */
2339 void
2340 cpu_visibility_configure(cpu_t *cp, zone_t *zone)
2341 {
2342 	zoneid_t zoneid = zone ? zone->zone_id : ALL_ZONES;
2343 
2344 	ASSERT(MUTEX_HELD(&cpu_lock));
2345 	ASSERT(pool_pset_enabled());
2346 	ASSERT(cp != NULL);
2347 
2348 	if (zoneid != ALL_ZONES && zoneid != GLOBAL_ZONEID) {
2349 		zone->zone_ncpus++;
2350 		ASSERT(zone->zone_ncpus <= ncpus);
2351 	}
2352 	if (cp->cpu_info_kstat != NULL)
2353 		kstat_zone_add(cp->cpu_info_kstat, zoneid);
2354 }
2355 
2356 /*
2357  * Make visible to the zone that subset of the cpu information that would be
2358  * initialized when a previously configured cpu is onlined.
2359  */
2360 void
2361 cpu_visibility_online(cpu_t *cp, zone_t *zone)
2362 {
2363 	kstat_t *ksp;
2364 	char name[sizeof ("cpu_stat") + 10];	/* enough for 32-bit cpuids */
2365 	zoneid_t zoneid = zone ? zone->zone_id : ALL_ZONES;
2366 	processorid_t cpun;
2367 
2368 	ASSERT(MUTEX_HELD(&cpu_lock));
2369 	ASSERT(pool_pset_enabled());
2370 	ASSERT(cp != NULL);
2371 	ASSERT(cpu_is_active(cp));
2372 
2373 	cpun = cp->cpu_id;
2374 	if (zoneid != ALL_ZONES && zoneid != GLOBAL_ZONEID) {
2375 		zone->zone_ncpus_online++;
2376 		ASSERT(zone->zone_ncpus_online <= ncpus_online);
2377 	}
2378 	(void) snprintf(name, sizeof (name), "cpu_stat%d", cpun);
2379 	if ((ksp = kstat_hold_byname("cpu_stat", cpun, name, ALL_ZONES))
2380 	    != NULL) {
2381 		kstat_zone_add(ksp, zoneid);
2382 		kstat_rele(ksp);
2383 	}
2384 	if ((ksp = kstat_hold_byname("cpu", cpun, "sys", ALL_ZONES)) != NULL) {
2385 		kstat_zone_add(ksp, zoneid);
2386 		kstat_rele(ksp);
2387 	}
2388 	if ((ksp = kstat_hold_byname("cpu", cpun, "vm", ALL_ZONES)) != NULL) {
2389 		kstat_zone_add(ksp, zoneid);
2390 		kstat_rele(ksp);
2391 	}
2392 	if ((ksp = kstat_hold_byname("cpu", cpun, "intrstat", ALL_ZONES)) !=
2393 	    NULL) {
2394 		kstat_zone_add(ksp, zoneid);
2395 		kstat_rele(ksp);
2396 	}
2397 }
2398 
2399 /*
2400  * Update relevant kstats such that cpu is now visible to processes
2401  * executing in specified zone.
2402  */
2403 void
2404 cpu_visibility_add(cpu_t *cp, zone_t *zone)
2405 {
2406 	cpu_visibility_configure(cp, zone);
2407 	if (cpu_is_active(cp))
2408 		cpu_visibility_online(cp, zone);
2409 }
2410 
2411 /*
2412  * Make invisible to the zone that subset of the cpu information that would be
2413  * torn down when a previously offlined cpu is unconfigured.
2414  */
2415 void
2416 cpu_visibility_unconfigure(cpu_t *cp, zone_t *zone)
2417 {
2418 	zoneid_t zoneid = zone ? zone->zone_id : ALL_ZONES;
2419 
2420 	ASSERT(MUTEX_HELD(&cpu_lock));
2421 	ASSERT(pool_pset_enabled());
2422 	ASSERT(cp != NULL);
2423 
2424 	if (zoneid != ALL_ZONES && zoneid != GLOBAL_ZONEID) {
2425 		ASSERT(zone->zone_ncpus != 0);
2426 		zone->zone_ncpus--;
2427 	}
2428 	if (cp->cpu_info_kstat)
2429 		kstat_zone_remove(cp->cpu_info_kstat, zoneid);
2430 }
2431 
2432 /*
2433  * Make invisible to the zone that subset of the cpu information that would be
2434  * torn down when a cpu is offlined (but still configured).
2435  */
2436 void
2437 cpu_visibility_offline(cpu_t *cp, zone_t *zone)
2438 {
2439 	kstat_t *ksp;
2440 	char name[sizeof ("cpu_stat") + 10];	/* enough for 32-bit cpuids */
2441 	zoneid_t zoneid = zone ? zone->zone_id : ALL_ZONES;
2442 	processorid_t cpun;
2443 
2444 	ASSERT(MUTEX_HELD(&cpu_lock));
2445 	ASSERT(pool_pset_enabled());
2446 	ASSERT(cp != NULL);
2447 	ASSERT(cpu_is_active(cp));
2448 
2449 	cpun = cp->cpu_id;
2450 	if (zoneid != ALL_ZONES && zoneid != GLOBAL_ZONEID) {
2451 		ASSERT(zone->zone_ncpus_online != 0);
2452 		zone->zone_ncpus_online--;
2453 	}
2454 
2455 	if ((ksp = kstat_hold_byname("cpu", cpun, "intrstat", ALL_ZONES)) !=
2456 	    NULL) {
2457 		kstat_zone_remove(ksp, zoneid);
2458 		kstat_rele(ksp);
2459 	}
2460 	if ((ksp = kstat_hold_byname("cpu", cpun, "vm", ALL_ZONES)) != NULL) {
2461 		kstat_zone_remove(ksp, zoneid);
2462 		kstat_rele(ksp);
2463 	}
2464 	if ((ksp = kstat_hold_byname("cpu", cpun, "sys", ALL_ZONES)) != NULL) {
2465 		kstat_zone_remove(ksp, zoneid);
2466 		kstat_rele(ksp);
2467 	}
2468 	(void) snprintf(name, sizeof (name), "cpu_stat%d", cpun);
2469 	if ((ksp = kstat_hold_byname("cpu_stat", cpun, name, ALL_ZONES))
2470 	    != NULL) {
2471 		kstat_zone_remove(ksp, zoneid);
2472 		kstat_rele(ksp);
2473 	}
2474 }
2475 
2476 /*
2477  * Update relevant kstats such that cpu is no longer visible to processes
2478  * executing in specified zone.
2479  */
2480 void
2481 cpu_visibility_remove(cpu_t *cp, zone_t *zone)
2482 {
2483 	if (cpu_is_active(cp))
2484 		cpu_visibility_offline(cp, zone);
2485 	cpu_visibility_unconfigure(cp, zone);
2486 }
2487 
2488 /*
2489  * Bind a thread to a CPU as requested.
2490  */
2491 int
2492 cpu_bind_thread(kthread_id_t tp, processorid_t bind, processorid_t *obind,
2493     int *error)
2494 {
2495 	processorid_t	binding;
2496 	cpu_t		*cp = NULL;
2497 
2498 	ASSERT(MUTEX_HELD(&cpu_lock));
2499 	ASSERT(MUTEX_HELD(&ttoproc(tp)->p_lock));
2500 
2501 	thread_lock(tp);
2502 
2503 	/*
2504 	 * Record old binding, but change the obind, which was initialized
2505 	 * to PBIND_NONE, only if this thread has a binding.  This avoids
2506 	 * reporting PBIND_NONE for a process when some LWPs are bound.
2507 	 */
2508 	binding = tp->t_bind_cpu;
2509 	if (binding != PBIND_NONE)
2510 		*obind = binding;	/* record old binding */
2511 
2512 	switch (bind) {
2513 	case PBIND_QUERY:
2514 		/* Just return the old binding */
2515 		thread_unlock(tp);
2516 		return (0);
2517 
2518 	case PBIND_QUERY_TYPE:
2519 		/* Return the binding type */
2520 		*obind = TB_CPU_IS_SOFT(tp) ? PBIND_SOFT : PBIND_HARD;
2521 		thread_unlock(tp);
2522 		return (0);
2523 
2524 	case PBIND_SOFT:
2525 		/*
2526 		 *  Set soft binding for this thread and return the actual
2527 		 *  binding
2528 		 */
2529 		TB_CPU_SOFT_SET(tp);
2530 		thread_unlock(tp);
2531 		return (0);
2532 
2533 	case PBIND_HARD:
2534 		/*
2535 		 *  Set hard binding for this thread and return the actual
2536 		 *  binding
2537 		 */
2538 		TB_CPU_HARD_SET(tp);
2539 		thread_unlock(tp);
2540 		return (0);
2541 
2542 	default:
2543 		break;
2544 	}
2545 
2546 	/*
2547 	 * If this thread/LWP cannot be bound because of permission
2548 	 * problems, just note that and return success so that the
2549 	 * other threads/LWPs will be bound.  This is the way
2550 	 * processor_bind() is defined to work.
2551 	 *
2552 	 * Binding will get EPERM if the thread is of system class
2553 	 * or hasprocperm() fails.
2554 	 */
2555 	if (tp->t_cid == 0 || !hasprocperm(tp->t_cred, CRED())) {
2556 		*error = EPERM;
2557 		thread_unlock(tp);
2558 		return (0);
2559 	}
2560 
2561 	binding = bind;
2562 	if (binding != PBIND_NONE) {
2563 		cp = cpu_get((processorid_t)binding);
2564 		/*
2565 		 * Make sure binding is valid and is in right partition.
2566 		 */
2567 		if (cp == NULL || tp->t_cpupart != cp->cpu_part) {
2568 			*error = EINVAL;
2569 			thread_unlock(tp);
2570 			return (0);
2571 		}
2572 	}
2573 	tp->t_bind_cpu = binding;	/* set new binding */
2574 
2575 	/*
2576 	 * If there is no system-set reason for affinity, set
2577 	 * the t_bound_cpu field to reflect the binding.
2578 	 */
2579 	if (tp->t_affinitycnt == 0) {
2580 		if (binding == PBIND_NONE) {
2581 			/*
2582 			 * We may need to adjust disp_max_unbound_pri
2583 			 * since we're becoming unbound.
2584 			 */
2585 			disp_adjust_unbound_pri(tp);
2586 
2587 			tp->t_bound_cpu = NULL;	/* set new binding */
2588 
2589 			/*
2590 			 * Move thread to lgroup with strongest affinity
2591 			 * after unbinding
2592 			 */
2593 			if (tp->t_lgrp_affinity)
2594 				lgrp_move_thread(tp,
2595 				    lgrp_choose(tp, tp->t_cpupart), 1);
2596 
2597 			if (tp->t_state == TS_ONPROC &&
2598 			    tp->t_cpu->cpu_part != tp->t_cpupart)
2599 				cpu_surrender(tp);
2600 		} else {
2601 			lpl_t	*lpl;
2602 
2603 			tp->t_bound_cpu = cp;
2604 			ASSERT(cp->cpu_lpl != NULL);
2605 
2606 			/*
2607 			 * Set home to lgroup with most affinity containing CPU
2608 			 * that thread is being bound or minimum bounding
2609 			 * lgroup if no affinities set
2610 			 */
2611 			if (tp->t_lgrp_affinity)
2612 				lpl = lgrp_affinity_best(tp, tp->t_cpupart,
2613 				    LGRP_NONE, B_FALSE);
2614 			else
2615 				lpl = cp->cpu_lpl;
2616 
2617 			if (tp->t_lpl != lpl) {
2618 				/* can't grab cpu_lock */
2619 				lgrp_move_thread(tp, lpl, 1);
2620 			}
2621 
2622 			/*
2623 			 * Make the thread switch to the bound CPU.
2624 			 * If the thread is runnable, we need to
2625 			 * requeue it even if t_cpu is already set
2626 			 * to the right CPU, since it may be on a
2627 			 * kpreempt queue and need to move to a local
2628 			 * queue.  We could check t_disp_queue to
2629 			 * avoid unnecessary overhead if it's already
2630 			 * on the right queue, but since this isn't
2631 			 * a performance-critical operation it doesn't
2632 			 * seem worth the extra code and complexity.
2633 			 *
2634 			 * If the thread is weakbound to the cpu then it will
2635 			 * resist the new binding request until the weak
2636 			 * binding drops.  The cpu_surrender or requeueing
2637 			 * below could be skipped in such cases (since it
2638 			 * will have no effect), but that would require
2639 			 * thread_allowmigrate to acquire thread_lock so
2640 			 * we'll take the very occasional hit here instead.
2641 			 */
2642 			if (tp->t_state == TS_ONPROC) {
2643 				cpu_surrender(tp);
2644 			} else if (tp->t_state == TS_RUN) {
2645 				cpu_t *ocp = tp->t_cpu;
2646 
2647 				(void) dispdeq(tp);
2648 				setbackdq(tp);
2649 				/*
2650 				 * Either on the bound CPU's disp queue now,
2651 				 * or swapped out or on the swap queue.
2652 				 */
2653 				ASSERT(tp->t_disp_queue == cp->cpu_disp ||
2654 				    tp->t_weakbound_cpu == ocp ||
2655 				    (tp->t_schedflag & (TS_LOAD | TS_ON_SWAPQ))
2656 				    != TS_LOAD);
2657 			}
2658 		}
2659 	}
2660 
2661 	/*
2662 	 * Our binding has changed; set TP_CHANGEBIND.
2663 	 */
2664 	tp->t_proc_flag |= TP_CHANGEBIND;
2665 	aston(tp);
2666 
2667 	thread_unlock(tp);
2668 
2669 	return (0);
2670 }
2671 
2672 #if CPUSET_WORDS > 1
2673 
2674 /*
2675  * Functions for implementing cpuset operations when a cpuset is more
2676  * than one word.  On platforms where a cpuset is a single word these
2677  * are implemented as macros in cpuvar.h.
2678  */
2679 
2680 void
2681 cpuset_all(cpuset_t *s)
2682 {
2683 	int i;
2684 
2685 	for (i = 0; i < CPUSET_WORDS; i++)
2686 		s->cpub[i] = ~0UL;
2687 }
2688 
2689 void
2690 cpuset_all_but(cpuset_t *s, uint_t cpu)
2691 {
2692 	cpuset_all(s);
2693 	CPUSET_DEL(*s, cpu);
2694 }
2695 
2696 void
2697 cpuset_only(cpuset_t *s, uint_t cpu)
2698 {
2699 	CPUSET_ZERO(*s);
2700 	CPUSET_ADD(*s, cpu);
2701 }
2702 
2703 int
2704 cpuset_isnull(cpuset_t *s)
2705 {
2706 	int i;
2707 
2708 	for (i = 0; i < CPUSET_WORDS; i++)
2709 		if (s->cpub[i] != 0)
2710 			return (0);
2711 	return (1);
2712 }
2713 
2714 int
2715 cpuset_cmp(cpuset_t *s1, cpuset_t *s2)
2716 {
2717 	int i;
2718 
2719 	for (i = 0; i < CPUSET_WORDS; i++)
2720 		if (s1->cpub[i] != s2->cpub[i])
2721 			return (0);
2722 	return (1);
2723 }
2724 
2725 uint_t
2726 cpuset_find(cpuset_t *s)
2727 {
2728 
2729 	uint_t	i;
2730 	uint_t	cpu = (uint_t)-1;
2731 
2732 	/*
2733 	 * Find a cpu in the cpuset
2734 	 */
2735 	for (i = 0; i < CPUSET_WORDS; i++) {
2736 		cpu = (uint_t)(lowbit(s->cpub[i]) - 1);
2737 		if (cpu != (uint_t)-1) {
2738 			cpu += i * BT_NBIPUL;
2739 			break;
2740 		}
2741 	}
2742 	return (cpu);
2743 }
2744 
2745 void
2746 cpuset_bounds(cpuset_t *s, uint_t *smallestid, uint_t *largestid)
2747 {
2748 	int	i, j;
2749 	uint_t	bit;
2750 
2751 	/*
2752 	 * First, find the smallest cpu id in the set.
2753 	 */
2754 	for (i = 0; i < CPUSET_WORDS; i++) {
2755 		if (s->cpub[i] != 0) {
2756 			bit = (uint_t)(lowbit(s->cpub[i]) - 1);
2757 			ASSERT(bit != (uint_t)-1);
2758 			*smallestid = bit + (i * BT_NBIPUL);
2759 
2760 			/*
2761 			 * Now find the largest cpu id in
2762 			 * the set and return immediately.
2763 			 * Done in an inner loop to avoid
2764 			 * having to break out of the first
2765 			 * loop.
2766 			 */
2767 			for (j = CPUSET_WORDS - 1; j >= i; j--) {
2768 				if (s->cpub[j] != 0) {
2769 					bit = (uint_t)(highbit(s->cpub[j]) - 1);
2770 					ASSERT(bit != (uint_t)-1);
2771 					*largestid = bit + (j * BT_NBIPUL);
2772 					ASSERT(*largestid >= *smallestid);
2773 					return;
2774 				}
2775 			}
2776 
2777 			/*
2778 			 * If this code is reached, a
2779 			 * smallestid was found, but not a
2780 			 * largestid. The cpuset must have
2781 			 * been changed during the course
2782 			 * of this function call.
2783 			 */
2784 			ASSERT(0);
2785 		}
2786 	}
2787 	*smallestid = *largestid = CPUSET_NOTINSET;
2788 }
2789 
2790 #endif	/* CPUSET_WORDS */
2791 
2792 /*
2793  * Unbind threads bound to specified CPU.
2794  *
2795  * If `unbind_all_threads' is true, unbind all user threads bound to a given
2796  * CPU. Otherwise unbind all soft-bound user threads.
2797  */
2798 int
2799 cpu_unbind(processorid_t cpu, boolean_t unbind_all_threads)
2800 {
2801 	processorid_t obind;
2802 	kthread_t *tp;
2803 	int ret = 0;
2804 	proc_t *pp;
2805 	int err, berr = 0;
2806 
2807 	ASSERT(MUTEX_HELD(&cpu_lock));
2808 
2809 	mutex_enter(&pidlock);
2810 	for (pp = practive; pp != NULL; pp = pp->p_next) {
2811 		mutex_enter(&pp->p_lock);
2812 		tp = pp->p_tlist;
2813 		/*
2814 		 * Skip zombies, kernel processes, and processes in
2815 		 * other zones, if called from a non-global zone.
2816 		 */
2817 		if (tp == NULL || (pp->p_flag & SSYS) ||
2818 		    !HASZONEACCESS(curproc, pp->p_zone->zone_id)) {
2819 			mutex_exit(&pp->p_lock);
2820 			continue;
2821 		}
2822 		do {
2823 			if (tp->t_bind_cpu != cpu)
2824 				continue;
2825 			/*
2826 			 * Skip threads with hard binding when
2827 			 * `unbind_all_threads' is not specified.
2828 			 */
2829 			if (!unbind_all_threads && TB_CPU_IS_HARD(tp))
2830 				continue;
2831 			err = cpu_bind_thread(tp, PBIND_NONE, &obind, &berr);
2832 			if (ret == 0)
2833 				ret = err;
2834 		} while ((tp = tp->t_forw) != pp->p_tlist);
2835 		mutex_exit(&pp->p_lock);
2836 	}
2837 	mutex_exit(&pidlock);
2838 	if (ret == 0)
2839 		ret = berr;
2840 	return (ret);
2841 }
2842 
2843 
2844 /*
2845  * Destroy all remaining bound threads on a cpu.
2846  */
2847 void
2848 cpu_destroy_bound_threads(cpu_t *cp)
2849 {
2850 	extern id_t syscid;
2851 	register kthread_id_t	t, tlist, tnext;
2852 
2853 	/*
2854 	 * Destroy all remaining bound threads on the cpu.  This
2855 	 * should include both the interrupt threads and the idle thread.
2856 	 * This requires some care, since we need to traverse the
2857 	 * thread list with the pidlock mutex locked, but thread_free
2858 	 * also locks the pidlock mutex.  So, we collect the threads
2859 	 * we're going to reap in a list headed by "tlist", then we
2860 	 * unlock the pidlock mutex and traverse the tlist list,
2861 	 * doing thread_free's on the thread's.	 Simple, n'est pas?
2862 	 * Also, this depends on thread_free not mucking with the
2863 	 * t_next and t_prev links of the thread.
2864 	 */
2865 
2866 	if ((t = curthread) != NULL) {
2867 
2868 		tlist = NULL;
2869 		mutex_enter(&pidlock);
2870 		do {
2871 			tnext = t->t_next;
2872 			if (t->t_bound_cpu == cp) {
2873 
2874 				/*
2875 				 * We've found a bound thread, carefully unlink
2876 				 * it out of the thread list, and add it to
2877 				 * our "tlist".	 We "know" we don't have to
2878 				 * worry about unlinking curthread (the thread
2879 				 * that is executing this code).
2880 				 */
2881 				t->t_next->t_prev = t->t_prev;
2882 				t->t_prev->t_next = t->t_next;
2883 				t->t_next = tlist;
2884 				tlist = t;
2885 				ASSERT(t->t_cid == syscid);
2886 				/* wake up anyone blocked in thread_join */
2887 				cv_broadcast(&t->t_joincv);
2888 				/*
2889 				 * t_lwp set by interrupt threads and not
2890 				 * cleared.
2891 				 */
2892 				t->t_lwp = NULL;
2893 				/*
2894 				 * Pause and idle threads always have
2895 				 * t_state set to TS_ONPROC.
2896 				 */
2897 				t->t_state = TS_FREE;
2898 				t->t_prev = NULL;	/* Just in case */
2899 			}
2900 
2901 		} while ((t = tnext) != curthread);
2902 
2903 		mutex_exit(&pidlock);
2904 
2905 		mutex_sync();
2906 		for (t = tlist; t != NULL; t = tnext) {
2907 			tnext = t->t_next;
2908 			thread_free(t);
2909 		}
2910 	}
2911 }
2912 
2913 /*
2914  * Update the cpu_supp_freqs of this cpu. This information is returned
2915  * as part of cpu_info kstats. If the cpu_info_kstat exists already, then
2916  * maintain the kstat data size.
2917  */
2918 void
2919 cpu_set_supp_freqs(cpu_t *cp, const char *freqs)
2920 {
2921 	char clkstr[sizeof ("18446744073709551615") + 1]; /* ui64 MAX */
2922 	const char *lfreqs = clkstr;
2923 	boolean_t kstat_exists = B_FALSE;
2924 	kstat_t *ksp;
2925 	size_t len;
2926 
2927 	/*
2928 	 * A NULL pointer means we only support one speed.
2929 	 */
2930 	if (freqs == NULL)
2931 		(void) snprintf(clkstr, sizeof (clkstr), "%"PRIu64,
2932 		    cp->cpu_curr_clock);
2933 	else
2934 		lfreqs = freqs;
2935 
2936 	/*
2937 	 * Make sure the frequency doesn't change while a snapshot is
2938 	 * going on. Of course, we only need to worry about this if
2939 	 * the kstat exists.
2940 	 */
2941 	if ((ksp = cp->cpu_info_kstat) != NULL) {
2942 		mutex_enter(ksp->ks_lock);
2943 		kstat_exists = B_TRUE;
2944 	}
2945 
2946 	/*
2947 	 * Free any previously allocated string and if the kstat
2948 	 * already exists, then update its data size.
2949 	 */
2950 	if (cp->cpu_supp_freqs != NULL) {
2951 		len = strlen(cp->cpu_supp_freqs) + 1;
2952 		kmem_free(cp->cpu_supp_freqs, len);
2953 		if (kstat_exists)
2954 			ksp->ks_data_size -= len;
2955 	}
2956 
2957 	/*
2958 	 * Allocate the new string and set the pointer.
2959 	 */
2960 	len = strlen(lfreqs) + 1;
2961 	cp->cpu_supp_freqs = kmem_alloc(len, KM_SLEEP);
2962 	(void) strcpy(cp->cpu_supp_freqs, lfreqs);
2963 
2964 	/*
2965 	 * If the kstat already exists then update the data size and
2966 	 * free the lock.
2967 	 */
2968 	if (kstat_exists) {
2969 		ksp->ks_data_size += len;
2970 		mutex_exit(ksp->ks_lock);
2971 	}
2972 }
2973 
2974 /*
2975  * Indicate the current CPU's clock freqency (in Hz).
2976  * The calling context must be such that CPU references are safe.
2977  */
2978 void
2979 cpu_set_curr_clock(uint64_t new_clk)
2980 {
2981 	uint64_t old_clk;
2982 
2983 	old_clk = CPU->cpu_curr_clock;
2984 	CPU->cpu_curr_clock = new_clk;
2985 
2986 	/*
2987 	 * The cpu-change-speed DTrace probe exports the frequency in Hz
2988 	 */
2989 	DTRACE_PROBE3(cpu__change__speed, processorid_t, CPU->cpu_id,
2990 	    uint64_t, old_clk, uint64_t, new_clk);
2991 }
2992 
2993 /*
2994  * processor_info(2) and p_online(2) status support functions
2995  *   The constants returned by the cpu_get_state() and cpu_get_state_str() are
2996  *   for use in communicating processor state information to userland.  Kernel
2997  *   subsystems should only be using the cpu_flags value directly.  Subsystems
2998  *   modifying cpu_flags should record the state change via a call to the
2999  *   cpu_set_state().
3000  */
3001 
3002 /*
3003  * Update the pi_state of this CPU.  This function provides the CPU status for
3004  * the information returned by processor_info(2).
3005  */
3006 void
3007 cpu_set_state(cpu_t *cpu)
3008 {
3009 	ASSERT(MUTEX_HELD(&cpu_lock));
3010 	cpu->cpu_type_info.pi_state = cpu_get_state(cpu);
3011 	cpu->cpu_state_begin = gethrestime_sec();
3012 	pool_cpu_mod = gethrtime();
3013 }
3014 
3015 /*
3016  * Return offline/online/other status for the indicated CPU.  Use only for
3017  * communication with user applications; cpu_flags provides the in-kernel
3018  * interface.
3019  */
3020 int
3021 cpu_get_state(cpu_t *cpu)
3022 {
3023 	ASSERT(MUTEX_HELD(&cpu_lock));
3024 	if (cpu->cpu_flags & CPU_POWEROFF)
3025 		return (P_POWEROFF);
3026 	else if (cpu->cpu_flags & CPU_FAULTED)
3027 		return (P_FAULTED);
3028 	else if (cpu->cpu_flags & CPU_SPARE)
3029 		return (P_SPARE);
3030 	else if ((cpu->cpu_flags & (CPU_READY | CPU_OFFLINE)) != CPU_READY)
3031 		return (P_OFFLINE);
3032 	else if (cpu->cpu_flags & CPU_ENABLE)
3033 		return (P_ONLINE);
3034 	else
3035 		return (P_NOINTR);
3036 }
3037 
3038 /*
3039  * Return processor_info(2) state as a string.
3040  */
3041 const char *
3042 cpu_get_state_str(cpu_t *cpu)
3043 {
3044 	const char *string;
3045 
3046 	switch (cpu_get_state(cpu)) {
3047 	case P_ONLINE:
3048 		string = PS_ONLINE;
3049 		break;
3050 	case P_POWEROFF:
3051 		string = PS_POWEROFF;
3052 		break;
3053 	case P_NOINTR:
3054 		string = PS_NOINTR;
3055 		break;
3056 	case P_SPARE:
3057 		string = PS_SPARE;
3058 		break;
3059 	case P_FAULTED:
3060 		string = PS_FAULTED;
3061 		break;
3062 	case P_OFFLINE:
3063 		string = PS_OFFLINE;
3064 		break;
3065 	default:
3066 		string = "unknown";
3067 		break;
3068 	}
3069 	return (string);
3070 }
3071 
3072 /*
3073  * Export this CPU's statistics (cpu_stat_t and cpu_stats_t) as raw and named
3074  * kstats, respectively.  This is done when a CPU is initialized or placed
3075  * online via p_online(2).
3076  */
3077 static void
3078 cpu_stats_kstat_create(cpu_t *cp)
3079 {
3080 	int 	instance = cp->cpu_id;
3081 	char 	*module = "cpu";
3082 	char 	*class = "misc";
3083 	kstat_t	*ksp;
3084 	zoneid_t zoneid;
3085 
3086 	ASSERT(MUTEX_HELD(&cpu_lock));
3087 
3088 	if (pool_pset_enabled())
3089 		zoneid = GLOBAL_ZONEID;
3090 	else
3091 		zoneid = ALL_ZONES;
3092 	/*
3093 	 * Create named kstats
3094 	 */
3095 #define	CPU_STATS_KS_CREATE(name, tsize, update_func)                    \
3096 	ksp = kstat_create_zone(module, instance, (name), class,         \
3097 	    KSTAT_TYPE_NAMED, (tsize) / sizeof (kstat_named_t), 0,       \
3098 	    zoneid);                                                     \
3099 	if (ksp != NULL) {                                               \
3100 		ksp->ks_private = cp;                                    \
3101 		ksp->ks_update = (update_func);                          \
3102 		kstat_install(ksp);                                      \
3103 	} else                                                           \
3104 		cmn_err(CE_WARN, "cpu: unable to create %s:%d:%s kstat", \
3105 		    module, instance, (name));
3106 
3107 	CPU_STATS_KS_CREATE("sys", sizeof (cpu_sys_stats_ks_data_template),
3108 	    cpu_sys_stats_ks_update);
3109 	CPU_STATS_KS_CREATE("vm", sizeof (cpu_vm_stats_ks_data_template),
3110 	    cpu_vm_stats_ks_update);
3111 
3112 	/*
3113 	 * Export the familiar cpu_stat_t KSTAT_TYPE_RAW kstat.
3114 	 */
3115 	ksp = kstat_create_zone("cpu_stat", cp->cpu_id, NULL,
3116 	    "misc", KSTAT_TYPE_RAW, sizeof (cpu_stat_t), 0, zoneid);
3117 	if (ksp != NULL) {
3118 		ksp->ks_update = cpu_stat_ks_update;
3119 		ksp->ks_private = cp;
3120 		kstat_install(ksp);
3121 	}
3122 }
3123 
3124 static void
3125 cpu_stats_kstat_destroy(cpu_t *cp)
3126 {
3127 	char ks_name[KSTAT_STRLEN];
3128 
3129 	(void) sprintf(ks_name, "cpu_stat%d", cp->cpu_id);
3130 	kstat_delete_byname("cpu_stat", cp->cpu_id, ks_name);
3131 
3132 	kstat_delete_byname("cpu", cp->cpu_id, "sys");
3133 	kstat_delete_byname("cpu", cp->cpu_id, "vm");
3134 }
3135 
3136 static int
3137 cpu_sys_stats_ks_update(kstat_t *ksp, int rw)
3138 {
3139 	cpu_t *cp = (cpu_t *)ksp->ks_private;
3140 	struct cpu_sys_stats_ks_data *csskd;
3141 	cpu_sys_stats_t *css;
3142 	hrtime_t msnsecs[NCMSTATES];
3143 	int	i;
3144 
3145 	if (rw == KSTAT_WRITE)
3146 		return (EACCES);
3147 
3148 	csskd = ksp->ks_data;
3149 	css = &cp->cpu_stats.sys;
3150 
3151 	/*
3152 	 * Read CPU mstate, but compare with the last values we
3153 	 * received to make sure that the returned kstats never
3154 	 * decrease.
3155 	 */
3156 
3157 	get_cpu_mstate(cp, msnsecs);
3158 	if (csskd->cpu_nsec_idle.value.ui64 > msnsecs[CMS_IDLE])
3159 		msnsecs[CMS_IDLE] = csskd->cpu_nsec_idle.value.ui64;
3160 	if (csskd->cpu_nsec_user.value.ui64 > msnsecs[CMS_USER])
3161 		msnsecs[CMS_USER] = csskd->cpu_nsec_user.value.ui64;
3162 	if (csskd->cpu_nsec_kernel.value.ui64 > msnsecs[CMS_SYSTEM])
3163 		msnsecs[CMS_SYSTEM] = csskd->cpu_nsec_kernel.value.ui64;
3164 
3165 	bcopy(&cpu_sys_stats_ks_data_template, ksp->ks_data,
3166 	    sizeof (cpu_sys_stats_ks_data_template));
3167 
3168 	csskd->cpu_ticks_wait.value.ui64 = 0;
3169 	csskd->wait_ticks_io.value.ui64 = 0;
3170 
3171 	csskd->cpu_nsec_idle.value.ui64 = msnsecs[CMS_IDLE];
3172 	csskd->cpu_nsec_user.value.ui64 = msnsecs[CMS_USER];
3173 	csskd->cpu_nsec_kernel.value.ui64 = msnsecs[CMS_SYSTEM];
3174 	csskd->cpu_ticks_idle.value.ui64 =
3175 	    NSEC_TO_TICK(csskd->cpu_nsec_idle.value.ui64);
3176 	csskd->cpu_ticks_user.value.ui64 =
3177 	    NSEC_TO_TICK(csskd->cpu_nsec_user.value.ui64);
3178 	csskd->cpu_ticks_kernel.value.ui64 =
3179 	    NSEC_TO_TICK(csskd->cpu_nsec_kernel.value.ui64);
3180 	csskd->cpu_nsec_intr.value.ui64 = cp->cpu_intrlast;
3181 	csskd->cpu_load_intr.value.ui64 = cp->cpu_intrload;
3182 	csskd->bread.value.ui64 = css->bread;
3183 	csskd->bwrite.value.ui64 = css->bwrite;
3184 	csskd->lread.value.ui64 = css->lread;
3185 	csskd->lwrite.value.ui64 = css->lwrite;
3186 	csskd->phread.value.ui64 = css->phread;
3187 	csskd->phwrite.value.ui64 = css->phwrite;
3188 	csskd->pswitch.value.ui64 = css->pswitch;
3189 	csskd->trap.value.ui64 = css->trap;
3190 	csskd->intr.value.ui64 = 0;
3191 	for (i = 0; i < PIL_MAX; i++)
3192 		csskd->intr.value.ui64 += css->intr[i];
3193 	csskd->syscall.value.ui64 = css->syscall;
3194 	csskd->sysread.value.ui64 = css->sysread;
3195 	csskd->syswrite.value.ui64 = css->syswrite;
3196 	csskd->sysfork.value.ui64 = css->sysfork;
3197 	csskd->sysvfork.value.ui64 = css->sysvfork;
3198 	csskd->sysexec.value.ui64 = css->sysexec;
3199 	csskd->readch.value.ui64 = css->readch;
3200 	csskd->writech.value.ui64 = css->writech;
3201 	csskd->rcvint.value.ui64 = css->rcvint;
3202 	csskd->xmtint.value.ui64 = css->xmtint;
3203 	csskd->mdmint.value.ui64 = css->mdmint;
3204 	csskd->rawch.value.ui64 = css->rawch;
3205 	csskd->canch.value.ui64 = css->canch;
3206 	csskd->outch.value.ui64 = css->outch;
3207 	csskd->msg.value.ui64 = css->msg;
3208 	csskd->sema.value.ui64 = css->sema;
3209 	csskd->namei.value.ui64 = css->namei;
3210 	csskd->ufsiget.value.ui64 = css->ufsiget;
3211 	csskd->ufsdirblk.value.ui64 = css->ufsdirblk;
3212 	csskd->ufsipage.value.ui64 = css->ufsipage;
3213 	csskd->ufsinopage.value.ui64 = css->ufsinopage;
3214 	csskd->procovf.value.ui64 = css->procovf;
3215 	csskd->intrthread.value.ui64 = 0;
3216 	for (i = 0; i < LOCK_LEVEL - 1; i++)
3217 		csskd->intrthread.value.ui64 += css->intr[i];
3218 	csskd->intrblk.value.ui64 = css->intrblk;
3219 	csskd->intrunpin.value.ui64 = css->intrunpin;
3220 	csskd->idlethread.value.ui64 = css->idlethread;
3221 	csskd->inv_swtch.value.ui64 = css->inv_swtch;
3222 	csskd->nthreads.value.ui64 = css->nthreads;
3223 	csskd->cpumigrate.value.ui64 = css->cpumigrate;
3224 	csskd->xcalls.value.ui64 = css->xcalls;
3225 	csskd->mutex_adenters.value.ui64 = css->mutex_adenters;
3226 	csskd->rw_rdfails.value.ui64 = css->rw_rdfails;
3227 	csskd->rw_wrfails.value.ui64 = css->rw_wrfails;
3228 	csskd->modload.value.ui64 = css->modload;
3229 	csskd->modunload.value.ui64 = css->modunload;
3230 	csskd->bawrite.value.ui64 = css->bawrite;
3231 	csskd->iowait.value.ui64 = css->iowait;
3232 
3233 	return (0);
3234 }
3235 
3236 static int
3237 cpu_vm_stats_ks_update(kstat_t *ksp, int rw)
3238 {
3239 	cpu_t *cp = (cpu_t *)ksp->ks_private;
3240 	struct cpu_vm_stats_ks_data *cvskd;
3241 	cpu_vm_stats_t *cvs;
3242 
3243 	if (rw == KSTAT_WRITE)
3244 		return (EACCES);
3245 
3246 	cvs = &cp->cpu_stats.vm;
3247 	cvskd = ksp->ks_data;
3248 
3249 	bcopy(&cpu_vm_stats_ks_data_template, ksp->ks_data,
3250 	    sizeof (cpu_vm_stats_ks_data_template));
3251 	cvskd->pgrec.value.ui64 = cvs->pgrec;
3252 	cvskd->pgfrec.value.ui64 = cvs->pgfrec;
3253 	cvskd->pgin.value.ui64 = cvs->pgin;
3254 	cvskd->pgpgin.value.ui64 = cvs->pgpgin;
3255 	cvskd->pgout.value.ui64 = cvs->pgout;
3256 	cvskd->pgpgout.value.ui64 = cvs->pgpgout;
3257 	cvskd->swapin.value.ui64 = cvs->swapin;
3258 	cvskd->pgswapin.value.ui64 = cvs->pgswapin;
3259 	cvskd->swapout.value.ui64 = cvs->swapout;
3260 	cvskd->pgswapout.value.ui64 = cvs->pgswapout;
3261 	cvskd->zfod.value.ui64 = cvs->zfod;
3262 	cvskd->dfree.value.ui64 = cvs->dfree;
3263 	cvskd->scan.value.ui64 = cvs->scan;
3264 	cvskd->rev.value.ui64 = cvs->rev;
3265 	cvskd->hat_fault.value.ui64 = cvs->hat_fault;
3266 	cvskd->as_fault.value.ui64 = cvs->as_fault;
3267 	cvskd->maj_fault.value.ui64 = cvs->maj_fault;
3268 	cvskd->cow_fault.value.ui64 = cvs->cow_fault;
3269 	cvskd->prot_fault.value.ui64 = cvs->prot_fault;
3270 	cvskd->softlock.value.ui64 = cvs->softlock;
3271 	cvskd->kernel_asflt.value.ui64 = cvs->kernel_asflt;
3272 	cvskd->pgrrun.value.ui64 = cvs->pgrrun;
3273 	cvskd->execpgin.value.ui64 = cvs->execpgin;
3274 	cvskd->execpgout.value.ui64 = cvs->execpgout;
3275 	cvskd->execfree.value.ui64 = cvs->execfree;
3276 	cvskd->anonpgin.value.ui64 = cvs->anonpgin;
3277 	cvskd->anonpgout.value.ui64 = cvs->anonpgout;
3278 	cvskd->anonfree.value.ui64 = cvs->anonfree;
3279 	cvskd->fspgin.value.ui64 = cvs->fspgin;
3280 	cvskd->fspgout.value.ui64 = cvs->fspgout;
3281 	cvskd->fsfree.value.ui64 = cvs->fsfree;
3282 
3283 	return (0);
3284 }
3285 
3286 static int
3287 cpu_stat_ks_update(kstat_t *ksp, int rw)
3288 {
3289 	cpu_stat_t *cso;
3290 	cpu_t *cp;
3291 	int i;
3292 	hrtime_t msnsecs[NCMSTATES];
3293 
3294 	cso = (cpu_stat_t *)ksp->ks_data;
3295 	cp = (cpu_t *)ksp->ks_private;
3296 
3297 	if (rw == KSTAT_WRITE)
3298 		return (EACCES);
3299 
3300 	/*
3301 	 * Read CPU mstate, but compare with the last values we
3302 	 * received to make sure that the returned kstats never
3303 	 * decrease.
3304 	 */
3305 
3306 	get_cpu_mstate(cp, msnsecs);
3307 	msnsecs[CMS_IDLE] = NSEC_TO_TICK(msnsecs[CMS_IDLE]);
3308 	msnsecs[CMS_USER] = NSEC_TO_TICK(msnsecs[CMS_USER]);
3309 	msnsecs[CMS_SYSTEM] = NSEC_TO_TICK(msnsecs[CMS_SYSTEM]);
3310 	if (cso->cpu_sysinfo.cpu[CPU_IDLE] < msnsecs[CMS_IDLE])
3311 		cso->cpu_sysinfo.cpu[CPU_IDLE] = msnsecs[CMS_IDLE];
3312 	if (cso->cpu_sysinfo.cpu[CPU_USER] < msnsecs[CMS_USER])
3313 		cso->cpu_sysinfo.cpu[CPU_USER] = msnsecs[CMS_USER];
3314 	if (cso->cpu_sysinfo.cpu[CPU_KERNEL] < msnsecs[CMS_SYSTEM])
3315 		cso->cpu_sysinfo.cpu[CPU_KERNEL] = msnsecs[CMS_SYSTEM];
3316 	cso->cpu_sysinfo.cpu[CPU_WAIT] 	= 0;
3317 	cso->cpu_sysinfo.wait[W_IO] 	= 0;
3318 	cso->cpu_sysinfo.wait[W_SWAP]	= 0;
3319 	cso->cpu_sysinfo.wait[W_PIO]	= 0;
3320 	cso->cpu_sysinfo.bread 		= CPU_STATS(cp, sys.bread);
3321 	cso->cpu_sysinfo.bwrite 	= CPU_STATS(cp, sys.bwrite);
3322 	cso->cpu_sysinfo.lread 		= CPU_STATS(cp, sys.lread);
3323 	cso->cpu_sysinfo.lwrite 	= CPU_STATS(cp, sys.lwrite);
3324 	cso->cpu_sysinfo.phread 	= CPU_STATS(cp, sys.phread);
3325 	cso->cpu_sysinfo.phwrite 	= CPU_STATS(cp, sys.phwrite);
3326 	cso->cpu_sysinfo.pswitch 	= CPU_STATS(cp, sys.pswitch);
3327 	cso->cpu_sysinfo.trap 		= CPU_STATS(cp, sys.trap);
3328 	cso->cpu_sysinfo.intr		= 0;
3329 	for (i = 0; i < PIL_MAX; i++)
3330 		cso->cpu_sysinfo.intr += CPU_STATS(cp, sys.intr[i]);
3331 	cso->cpu_sysinfo.syscall	= CPU_STATS(cp, sys.syscall);
3332 	cso->cpu_sysinfo.sysread	= CPU_STATS(cp, sys.sysread);
3333 	cso->cpu_sysinfo.syswrite	= CPU_STATS(cp, sys.syswrite);
3334 	cso->cpu_sysinfo.sysfork	= CPU_STATS(cp, sys.sysfork);
3335 	cso->cpu_sysinfo.sysvfork	= CPU_STATS(cp, sys.sysvfork);
3336 	cso->cpu_sysinfo.sysexec	= CPU_STATS(cp, sys.sysexec);
3337 	cso->cpu_sysinfo.readch		= CPU_STATS(cp, sys.readch);
3338 	cso->cpu_sysinfo.writech	= CPU_STATS(cp, sys.writech);
3339 	cso->cpu_sysinfo.rcvint		= CPU_STATS(cp, sys.rcvint);
3340 	cso->cpu_sysinfo.xmtint		= CPU_STATS(cp, sys.xmtint);
3341 	cso->cpu_sysinfo.mdmint		= CPU_STATS(cp, sys.mdmint);
3342 	cso->cpu_sysinfo.rawch		= CPU_STATS(cp, sys.rawch);
3343 	cso->cpu_sysinfo.canch		= CPU_STATS(cp, sys.canch);
3344 	cso->cpu_sysinfo.outch		= CPU_STATS(cp, sys.outch);
3345 	cso->cpu_sysinfo.msg		= CPU_STATS(cp, sys.msg);
3346 	cso->cpu_sysinfo.sema		= CPU_STATS(cp, sys.sema);
3347 	cso->cpu_sysinfo.namei		= CPU_STATS(cp, sys.namei);
3348 	cso->cpu_sysinfo.ufsiget	= CPU_STATS(cp, sys.ufsiget);
3349 	cso->cpu_sysinfo.ufsdirblk	= CPU_STATS(cp, sys.ufsdirblk);
3350 	cso->cpu_sysinfo.ufsipage	= CPU_STATS(cp, sys.ufsipage);
3351 	cso->cpu_sysinfo.ufsinopage	= CPU_STATS(cp, sys.ufsinopage);
3352 	cso->cpu_sysinfo.inodeovf	= 0;
3353 	cso->cpu_sysinfo.fileovf	= 0;
3354 	cso->cpu_sysinfo.procovf	= CPU_STATS(cp, sys.procovf);
3355 	cso->cpu_sysinfo.intrthread	= 0;
3356 	for (i = 0; i < LOCK_LEVEL - 1; i++)
3357 		cso->cpu_sysinfo.intrthread += CPU_STATS(cp, sys.intr[i]);
3358 	cso->cpu_sysinfo.intrblk	= CPU_STATS(cp, sys.intrblk);
3359 	cso->cpu_sysinfo.idlethread	= CPU_STATS(cp, sys.idlethread);
3360 	cso->cpu_sysinfo.inv_swtch	= CPU_STATS(cp, sys.inv_swtch);
3361 	cso->cpu_sysinfo.nthreads	= CPU_STATS(cp, sys.nthreads);
3362 	cso->cpu_sysinfo.cpumigrate	= CPU_STATS(cp, sys.cpumigrate);
3363 	cso->cpu_sysinfo.xcalls		= CPU_STATS(cp, sys.xcalls);
3364 	cso->cpu_sysinfo.mutex_adenters	= CPU_STATS(cp, sys.mutex_adenters);
3365 	cso->cpu_sysinfo.rw_rdfails	= CPU_STATS(cp, sys.rw_rdfails);
3366 	cso->cpu_sysinfo.rw_wrfails	= CPU_STATS(cp, sys.rw_wrfails);
3367 	cso->cpu_sysinfo.modload	= CPU_STATS(cp, sys.modload);
3368 	cso->cpu_sysinfo.modunload	= CPU_STATS(cp, sys.modunload);
3369 	cso->cpu_sysinfo.bawrite	= CPU_STATS(cp, sys.bawrite);
3370 	cso->cpu_sysinfo.rw_enters	= 0;
3371 	cso->cpu_sysinfo.win_uo_cnt	= 0;
3372 	cso->cpu_sysinfo.win_uu_cnt	= 0;
3373 	cso->cpu_sysinfo.win_so_cnt	= 0;
3374 	cso->cpu_sysinfo.win_su_cnt	= 0;
3375 	cso->cpu_sysinfo.win_suo_cnt	= 0;
3376 
3377 	cso->cpu_syswait.iowait		= CPU_STATS(cp, sys.iowait);
3378 	cso->cpu_syswait.swap		= 0;
3379 	cso->cpu_syswait.physio		= 0;
3380 
3381 	cso->cpu_vminfo.pgrec		= CPU_STATS(cp, vm.pgrec);
3382 	cso->cpu_vminfo.pgfrec		= CPU_STATS(cp, vm.pgfrec);
3383 	cso->cpu_vminfo.pgin		= CPU_STATS(cp, vm.pgin);
3384 	cso->cpu_vminfo.pgpgin		= CPU_STATS(cp, vm.pgpgin);
3385 	cso->cpu_vminfo.pgout		= CPU_STATS(cp, vm.pgout);
3386 	cso->cpu_vminfo.pgpgout		= CPU_STATS(cp, vm.pgpgout);
3387 	cso->cpu_vminfo.swapin		= CPU_STATS(cp, vm.swapin);
3388 	cso->cpu_vminfo.pgswapin	= CPU_STATS(cp, vm.pgswapin);
3389 	cso->cpu_vminfo.swapout		= CPU_STATS(cp, vm.swapout);
3390 	cso->cpu_vminfo.pgswapout	= CPU_STATS(cp, vm.pgswapout);
3391 	cso->cpu_vminfo.zfod		= CPU_STATS(cp, vm.zfod);
3392 	cso->cpu_vminfo.dfree		= CPU_STATS(cp, vm.dfree);
3393 	cso->cpu_vminfo.scan		= CPU_STATS(cp, vm.scan);
3394 	cso->cpu_vminfo.rev		= CPU_STATS(cp, vm.rev);
3395 	cso->cpu_vminfo.hat_fault	= CPU_STATS(cp, vm.hat_fault);
3396 	cso->cpu_vminfo.as_fault	= CPU_STATS(cp, vm.as_fault);
3397 	cso->cpu_vminfo.maj_fault	= CPU_STATS(cp, vm.maj_fault);
3398 	cso->cpu_vminfo.cow_fault	= CPU_STATS(cp, vm.cow_fault);
3399 	cso->cpu_vminfo.prot_fault	= CPU_STATS(cp, vm.prot_fault);
3400 	cso->cpu_vminfo.softlock	= CPU_STATS(cp, vm.softlock);
3401 	cso->cpu_vminfo.kernel_asflt	= CPU_STATS(cp, vm.kernel_asflt);
3402 	cso->cpu_vminfo.pgrrun		= CPU_STATS(cp, vm.pgrrun);
3403 	cso->cpu_vminfo.execpgin	= CPU_STATS(cp, vm.execpgin);
3404 	cso->cpu_vminfo.execpgout	= CPU_STATS(cp, vm.execpgout);
3405 	cso->cpu_vminfo.execfree	= CPU_STATS(cp, vm.execfree);
3406 	cso->cpu_vminfo.anonpgin	= CPU_STATS(cp, vm.anonpgin);
3407 	cso->cpu_vminfo.anonpgout	= CPU_STATS(cp, vm.anonpgout);
3408 	cso->cpu_vminfo.anonfree	= CPU_STATS(cp, vm.anonfree);
3409 	cso->cpu_vminfo.fspgin		= CPU_STATS(cp, vm.fspgin);
3410 	cso->cpu_vminfo.fspgout		= CPU_STATS(cp, vm.fspgout);
3411 	cso->cpu_vminfo.fsfree		= CPU_STATS(cp, vm.fsfree);
3412 
3413 	return (0);
3414 }
3415