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 (c) 1991, 2010, Oracle and/or its affiliates. All rights reserved.
  23  * Copyright (c) 2012 by Delphix. All rights reserved.
  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  * max_cpu_seqid_ever, and dispatch queue reallocations.  The lock ordering with
  95  * respect to 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 /*
 137  * Maximum cpu_seqid was given. This number can only grow and never shrink. It
 138  * can be used to optimize NCPU loops to avoid going through CPUs which were
 139  * never on-line.
 140  */
 141 processorid_t max_cpu_seqid_ever = 0;
 142 
 143 int ncpus = 1;
 144 int ncpus_online = 1;
 145 
 146 /*
 147  * CPU that we're trying to offline.  Protected by cpu_lock.
 148  */
 149 cpu_t *cpu_inmotion;
 150 
 151 /*
 152  * Can be raised to suppress further weakbinding, which are instead
 153  * satisfied by disabling preemption.  Must be raised/lowered under cpu_lock,
 154  * while individual thread weakbinding synchronization is done under thread
 155  * lock.
 156  */
 157 int weakbindingbarrier;
 158 
 159 /*
 160  * Variables used in pause_cpus().
 161  */
 162 static volatile char safe_list[NCPU];
 163 
 164 static struct _cpu_pause_info {
 165         int             cp_spl;         /* spl saved in pause_cpus() */
 166         volatile int    cp_go;          /* Go signal sent after all ready */
 167         int             cp_count;       /* # of CPUs to pause */
 168         ksema_t         cp_sem;         /* synch pause_cpus & cpu_pause */
 169         kthread_id_t    cp_paused;
 170         void            *(*cp_func)(void *);
 171 } cpu_pause_info;
 172 
 173 static kmutex_t pause_free_mutex;
 174 static kcondvar_t pause_free_cv;
 175 
 176 
 177 static struct cpu_sys_stats_ks_data {
 178         kstat_named_t cpu_ticks_idle;
 179         kstat_named_t cpu_ticks_user;
 180         kstat_named_t cpu_ticks_kernel;
 181         kstat_named_t cpu_ticks_wait;
 182         kstat_named_t cpu_nsec_idle;
 183         kstat_named_t cpu_nsec_user;
 184         kstat_named_t cpu_nsec_kernel;
 185         kstat_named_t cpu_nsec_dtrace;
 186         kstat_named_t cpu_nsec_intr;
 187         kstat_named_t cpu_load_intr;
 188         kstat_named_t wait_ticks_io;
 189         kstat_named_t dtrace_probes;
 190         kstat_named_t bread;
 191         kstat_named_t bwrite;
 192         kstat_named_t lread;
 193         kstat_named_t lwrite;
 194         kstat_named_t phread;
 195         kstat_named_t phwrite;
 196         kstat_named_t pswitch;
 197         kstat_named_t trap;
 198         kstat_named_t intr;
 199         kstat_named_t syscall;
 200         kstat_named_t sysread;
 201         kstat_named_t syswrite;
 202         kstat_named_t sysfork;
 203         kstat_named_t sysvfork;
 204         kstat_named_t sysexec;
 205         kstat_named_t readch;
 206         kstat_named_t writech;
 207         kstat_named_t rcvint;
 208         kstat_named_t xmtint;
 209         kstat_named_t mdmint;
 210         kstat_named_t rawch;
 211         kstat_named_t canch;
 212         kstat_named_t outch;
 213         kstat_named_t msg;
 214         kstat_named_t sema;
 215         kstat_named_t namei;
 216         kstat_named_t ufsiget;
 217         kstat_named_t ufsdirblk;
 218         kstat_named_t ufsipage;
 219         kstat_named_t ufsinopage;
 220         kstat_named_t procovf;
 221         kstat_named_t intrthread;
 222         kstat_named_t intrblk;
 223         kstat_named_t intrunpin;
 224         kstat_named_t idlethread;
 225         kstat_named_t inv_swtch;
 226         kstat_named_t nthreads;
 227         kstat_named_t cpumigrate;
 228         kstat_named_t xcalls;
 229         kstat_named_t mutex_adenters;
 230         kstat_named_t rw_rdfails;
 231         kstat_named_t rw_wrfails;
 232         kstat_named_t modload;
 233         kstat_named_t modunload;
 234         kstat_named_t bawrite;
 235         kstat_named_t iowait;
 236 } cpu_sys_stats_ks_data_template = {
 237         { "cpu_ticks_idle",     KSTAT_DATA_UINT64 },
 238         { "cpu_ticks_user",     KSTAT_DATA_UINT64 },
 239         { "cpu_ticks_kernel",   KSTAT_DATA_UINT64 },
 240         { "cpu_ticks_wait",     KSTAT_DATA_UINT64 },
 241         { "cpu_nsec_idle",      KSTAT_DATA_UINT64 },
 242         { "cpu_nsec_user",      KSTAT_DATA_UINT64 },
 243         { "cpu_nsec_kernel",    KSTAT_DATA_UINT64 },
 244         { "cpu_nsec_dtrace",    KSTAT_DATA_UINT64 },
 245         { "cpu_nsec_intr",      KSTAT_DATA_UINT64 },
 246         { "cpu_load_intr",      KSTAT_DATA_UINT64 },
 247         { "wait_ticks_io",      KSTAT_DATA_UINT64 },
 248         { "dtrace_probes",      KSTAT_DATA_UINT64 },
 249         { "bread",              KSTAT_DATA_UINT64 },
 250         { "bwrite",             KSTAT_DATA_UINT64 },
 251         { "lread",              KSTAT_DATA_UINT64 },
 252         { "lwrite",             KSTAT_DATA_UINT64 },
 253         { "phread",             KSTAT_DATA_UINT64 },
 254         { "phwrite",            KSTAT_DATA_UINT64 },
 255         { "pswitch",            KSTAT_DATA_UINT64 },
 256         { "trap",               KSTAT_DATA_UINT64 },
 257         { "intr",               KSTAT_DATA_UINT64 },
 258         { "syscall",            KSTAT_DATA_UINT64 },
 259         { "sysread",            KSTAT_DATA_UINT64 },
 260         { "syswrite",           KSTAT_DATA_UINT64 },
 261         { "sysfork",            KSTAT_DATA_UINT64 },
 262         { "sysvfork",           KSTAT_DATA_UINT64 },
 263         { "sysexec",            KSTAT_DATA_UINT64 },
 264         { "readch",             KSTAT_DATA_UINT64 },
 265         { "writech",            KSTAT_DATA_UINT64 },
 266         { "rcvint",             KSTAT_DATA_UINT64 },
 267         { "xmtint",             KSTAT_DATA_UINT64 },
 268         { "mdmint",             KSTAT_DATA_UINT64 },
 269         { "rawch",              KSTAT_DATA_UINT64 },
 270         { "canch",              KSTAT_DATA_UINT64 },
 271         { "outch",              KSTAT_DATA_UINT64 },
 272         { "msg",                KSTAT_DATA_UINT64 },
 273         { "sema",               KSTAT_DATA_UINT64 },
 274         { "namei",              KSTAT_DATA_UINT64 },
 275         { "ufsiget",            KSTAT_DATA_UINT64 },
 276         { "ufsdirblk",          KSTAT_DATA_UINT64 },
 277         { "ufsipage",           KSTAT_DATA_UINT64 },
 278         { "ufsinopage",         KSTAT_DATA_UINT64 },
 279         { "procovf",            KSTAT_DATA_UINT64 },
 280         { "intrthread",         KSTAT_DATA_UINT64 },
 281         { "intrblk",            KSTAT_DATA_UINT64 },
 282         { "intrunpin",          KSTAT_DATA_UINT64 },
 283         { "idlethread",         KSTAT_DATA_UINT64 },
 284         { "inv_swtch",          KSTAT_DATA_UINT64 },
 285         { "nthreads",           KSTAT_DATA_UINT64 },
 286         { "cpumigrate",         KSTAT_DATA_UINT64 },
 287         { "xcalls",             KSTAT_DATA_UINT64 },
 288         { "mutex_adenters",     KSTAT_DATA_UINT64 },
 289         { "rw_rdfails",         KSTAT_DATA_UINT64 },
 290         { "rw_wrfails",         KSTAT_DATA_UINT64 },
 291         { "modload",            KSTAT_DATA_UINT64 },
 292         { "modunload",          KSTAT_DATA_UINT64 },
 293         { "bawrite",            KSTAT_DATA_UINT64 },
 294         { "iowait",             KSTAT_DATA_UINT64 },
 295 };
 296 
 297 static struct cpu_vm_stats_ks_data {
 298         kstat_named_t pgrec;
 299         kstat_named_t pgfrec;
 300         kstat_named_t pgin;
 301         kstat_named_t pgpgin;
 302         kstat_named_t pgout;
 303         kstat_named_t pgpgout;
 304         kstat_named_t zfod;
 305         kstat_named_t dfree;
 306         kstat_named_t scan;
 307         kstat_named_t rev;
 308         kstat_named_t hat_fault;
 309         kstat_named_t as_fault;
 310         kstat_named_t maj_fault;
 311         kstat_named_t cow_fault;
 312         kstat_named_t prot_fault;
 313         kstat_named_t softlock;
 314         kstat_named_t kernel_asflt;
 315         kstat_named_t pgrrun;
 316         kstat_named_t execpgin;
 317         kstat_named_t execpgout;
 318         kstat_named_t execfree;
 319         kstat_named_t anonpgin;
 320         kstat_named_t anonpgout;
 321         kstat_named_t anonfree;
 322         kstat_named_t fspgin;
 323         kstat_named_t fspgout;
 324         kstat_named_t fsfree;
 325 } cpu_vm_stats_ks_data_template = {
 326         { "pgrec",              KSTAT_DATA_UINT64 },
 327         { "pgfrec",             KSTAT_DATA_UINT64 },
 328         { "pgin",               KSTAT_DATA_UINT64 },
 329         { "pgpgin",             KSTAT_DATA_UINT64 },
 330         { "pgout",              KSTAT_DATA_UINT64 },
 331         { "pgpgout",            KSTAT_DATA_UINT64 },
 332         { "zfod",               KSTAT_DATA_UINT64 },
 333         { "dfree",              KSTAT_DATA_UINT64 },
 334         { "scan",               KSTAT_DATA_UINT64 },
 335         { "rev",                KSTAT_DATA_UINT64 },
 336         { "hat_fault",          KSTAT_DATA_UINT64 },
 337         { "as_fault",           KSTAT_DATA_UINT64 },
 338         { "maj_fault",          KSTAT_DATA_UINT64 },
 339         { "cow_fault",          KSTAT_DATA_UINT64 },
 340         { "prot_fault",         KSTAT_DATA_UINT64 },
 341         { "softlock",           KSTAT_DATA_UINT64 },
 342         { "kernel_asflt",       KSTAT_DATA_UINT64 },
 343         { "pgrrun",             KSTAT_DATA_UINT64 },
 344         { "execpgin",           KSTAT_DATA_UINT64 },
 345         { "execpgout",          KSTAT_DATA_UINT64 },
 346         { "execfree",           KSTAT_DATA_UINT64 },
 347         { "anonpgin",           KSTAT_DATA_UINT64 },
 348         { "anonpgout",          KSTAT_DATA_UINT64 },
 349         { "anonfree",           KSTAT_DATA_UINT64 },
 350         { "fspgin",             KSTAT_DATA_UINT64 },
 351         { "fspgout",            KSTAT_DATA_UINT64 },
 352         { "fsfree",             KSTAT_DATA_UINT64 },
 353 };
 354 
 355 /*
 356  * Force the specified thread to migrate to the appropriate processor.
 357  * Called with thread lock held, returns with it dropped.
 358  */
 359 static void
 360 force_thread_migrate(kthread_id_t tp)
 361 {
 362         ASSERT(THREAD_LOCK_HELD(tp));
 363         if (tp == curthread) {
 364                 THREAD_TRANSITION(tp);
 365                 CL_SETRUN(tp);
 366                 thread_unlock_nopreempt(tp);
 367                 swtch();
 368         } else {
 369                 if (tp->t_state == TS_ONPROC) {
 370                         cpu_surrender(tp);
 371                 } else if (tp->t_state == TS_RUN) {
 372                         (void) dispdeq(tp);
 373                         setbackdq(tp);
 374                 }
 375                 thread_unlock(tp);
 376         }
 377 }
 378 
 379 /*
 380  * Set affinity for a specified CPU.
 381  * A reference count is incremented and the affinity is held until the
 382  * reference count is decremented to zero by thread_affinity_clear().
 383  * This is so regions of code requiring affinity can be nested.
 384  * Caller needs to ensure that cpu_id remains valid, which can be
 385  * done by holding cpu_lock across this call, unless the caller
 386  * specifies CPU_CURRENT in which case the cpu_lock will be acquired
 387  * by thread_affinity_set and CPU->cpu_id will be the target CPU.
 388  */
 389 void
 390 thread_affinity_set(kthread_id_t t, int cpu_id)
 391 {
 392         cpu_t           *cp;
 393         int             c;
 394 
 395         ASSERT(!(t == curthread && t->t_weakbound_cpu != NULL));
 396 
 397         if ((c = cpu_id) == CPU_CURRENT) {
 398                 mutex_enter(&cpu_lock);
 399                 cpu_id = CPU->cpu_id;
 400         }
 401         /*
 402          * We should be asserting that cpu_lock is held here, but
 403          * the NCA code doesn't acquire it.  The following assert
 404          * should be uncommented when the NCA code is fixed.
 405          *
 406          * ASSERT(MUTEX_HELD(&cpu_lock));
 407          */
 408         ASSERT((cpu_id >= 0) && (cpu_id < NCPU));
 409         cp = cpu[cpu_id];
 410         ASSERT(cp != NULL);             /* user must provide a good cpu_id */
 411         /*
 412          * If there is already a hard affinity requested, and this affinity
 413          * conflicts with that, panic.
 414          */
 415         thread_lock(t);
 416         if (t->t_affinitycnt > 0 && t->t_bound_cpu != cp) {
 417                 panic("affinity_set: setting %p but already bound to %p",
 418                     (void *)cp, (void *)t->t_bound_cpu);
 419         }
 420         t->t_affinitycnt++;
 421         t->t_bound_cpu = cp;
 422 
 423         /*
 424          * Make sure we're running on the right CPU.
 425          */
 426         if (cp != t->t_cpu || t != curthread) {
 427                 force_thread_migrate(t);        /* drops thread lock */
 428         } else {
 429                 thread_unlock(t);
 430         }
 431 
 432         if (c == CPU_CURRENT)
 433                 mutex_exit(&cpu_lock);
 434 }
 435 
 436 /*
 437  *      Wrapper for backward compatibility.
 438  */
 439 void
 440 affinity_set(int cpu_id)
 441 {
 442         thread_affinity_set(curthread, cpu_id);
 443 }
 444 
 445 /*
 446  * Decrement the affinity reservation count and if it becomes zero,
 447  * clear the CPU affinity for the current thread, or set it to the user's
 448  * software binding request.
 449  */
 450 void
 451 thread_affinity_clear(kthread_id_t t)
 452 {
 453         register processorid_t binding;
 454 
 455         thread_lock(t);
 456         if (--t->t_affinitycnt == 0) {
 457                 if ((binding = t->t_bind_cpu) == PBIND_NONE) {
 458                         /*
 459                          * Adjust disp_max_unbound_pri if necessary.
 460                          */
 461                         disp_adjust_unbound_pri(t);
 462                         t->t_bound_cpu = NULL;
 463                         if (t->t_cpu->cpu_part != t->t_cpupart) {
 464                                 force_thread_migrate(t);
 465                                 return;
 466                         }
 467                 } else {
 468                         t->t_bound_cpu = cpu[binding];
 469                         /*
 470                          * Make sure the thread is running on the bound CPU.
 471                          */
 472                         if (t->t_cpu != t->t_bound_cpu) {
 473                                 force_thread_migrate(t);
 474                                 return;         /* already dropped lock */
 475                         }
 476                 }
 477         }
 478         thread_unlock(t);
 479 }
 480 
 481 /*
 482  * Wrapper for backward compatibility.
 483  */
 484 void
 485 affinity_clear(void)
 486 {
 487         thread_affinity_clear(curthread);
 488 }
 489 
 490 /*
 491  * Weak cpu affinity.  Bind to the "current" cpu for short periods
 492  * of time during which the thread must not block (but may be preempted).
 493  * Use this instead of kpreempt_disable() when it is only "no migration"
 494  * rather than "no preemption" semantics that are required - disabling
 495  * preemption holds higher priority threads off of cpu and if the
 496  * operation that is protected is more than momentary this is not good
 497  * for realtime etc.
 498  *
 499  * Weakly bound threads will not prevent a cpu from being offlined -
 500  * we'll only run them on the cpu to which they are weakly bound but
 501  * (because they do not block) we'll always be able to move them on to
 502  * another cpu at offline time if we give them just a short moment to
 503  * run during which they will unbind.  To give a cpu a chance of offlining,
 504  * however, we require a barrier to weak bindings that may be raised for a
 505  * given cpu (offline/move code may set this and then wait a short time for
 506  * existing weak bindings to drop); the cpu_inmotion pointer is that barrier.
 507  *
 508  * There are few restrictions on the calling context of thread_nomigrate.
 509  * The caller must not hold the thread lock.  Calls may be nested.
 510  *
 511  * After weakbinding a thread must not perform actions that may block.
 512  * In particular it must not call thread_affinity_set; calling that when
 513  * already weakbound is nonsensical anyway.
 514  *
 515  * If curthread is prevented from migrating for other reasons
 516  * (kernel preemption disabled; high pil; strongly bound; interrupt thread)
 517  * then the weak binding will succeed even if this cpu is the target of an
 518  * offline/move request.
 519  */
 520 void
 521 thread_nomigrate(void)
 522 {
 523         cpu_t *cp;
 524         kthread_id_t t = curthread;
 525 
 526 again:
 527         kpreempt_disable();
 528         cp = CPU;
 529 
 530         /*
 531          * A highlevel interrupt must not modify t_nomigrate or
 532          * t_weakbound_cpu of the thread it has interrupted.  A lowlevel
 533          * interrupt thread cannot migrate and we can avoid the
 534          * thread_lock call below by short-circuiting here.  In either
 535          * case we can just return since no migration is possible and
 536          * the condition will persist (ie, when we test for these again
 537          * in thread_allowmigrate they can't have changed).   Migration
 538          * is also impossible if we're at or above DISP_LEVEL pil.
 539          */
 540         if (CPU_ON_INTR(cp) || t->t_flag & T_INTR_THREAD ||
 541             getpil() >= DISP_LEVEL) {
 542                 kpreempt_enable();
 543                 return;
 544         }
 545 
 546         /*
 547          * We must be consistent with existing weak bindings.  Since we
 548          * may be interrupted between the increment of t_nomigrate and
 549          * the store to t_weakbound_cpu below we cannot assume that
 550          * t_weakbound_cpu will be set if t_nomigrate is.  Note that we
 551          * cannot assert t_weakbound_cpu == t_bind_cpu since that is not
 552          * always the case.
 553          */
 554         if (t->t_nomigrate && t->t_weakbound_cpu && t->t_weakbound_cpu != cp) {
 555                 if (!panicstr)
 556                         panic("thread_nomigrate: binding to %p but already "
 557                             "bound to %p", (void *)cp,
 558                             (void *)t->t_weakbound_cpu);
 559         }
 560 
 561         /*
 562          * At this point we have preemption disabled and we don't yet hold
 563          * the thread lock.  So it's possible that somebody else could
 564          * set t_bind_cpu here and not be able to force us across to the
 565          * new cpu (since we have preemption disabled).
 566          */
 567         thread_lock(curthread);
 568 
 569         /*
 570          * If further weak bindings are being (temporarily) suppressed then
 571          * we'll settle for disabling kernel preemption (which assures
 572          * no migration provided the thread does not block which it is
 573          * not allowed to if using thread_nomigrate).  We must remember
 574          * this disposition so we can take appropriate action in
 575          * thread_allowmigrate.  If this is a nested call and the
 576          * thread is already weakbound then fall through as normal.
 577          * We remember the decision to settle for kpreempt_disable through
 578          * negative nesting counting in t_nomigrate.  Once a thread has had one
 579          * weakbinding request satisfied in this way any further (nested)
 580          * requests will continue to be satisfied in the same way,
 581          * even if weak bindings have recommenced.
 582          */
 583         if (t->t_nomigrate < 0 || weakbindingbarrier && t->t_nomigrate == 0) {
 584                 --t->t_nomigrate;
 585                 thread_unlock(curthread);
 586                 return;         /* with kpreempt_disable still active */
 587         }
 588 
 589         /*
 590          * We hold thread_lock so t_bind_cpu cannot change.  We could,
 591          * however, be running on a different cpu to which we are t_bound_cpu
 592          * to (as explained above).  If we grant the weak binding request
 593          * in that case then the dispatcher must favour our weak binding
 594          * over our strong (in which case, just as when preemption is
 595          * disabled, we can continue to run on a cpu other than the one to
 596          * which we are strongbound; the difference in this case is that
 597          * this thread can be preempted and so can appear on the dispatch
 598          * queues of a cpu other than the one it is strongbound to).
 599          *
 600          * If the cpu we are running on does not appear to be a current
 601          * offline target (we check cpu_inmotion to determine this - since
 602          * we don't hold cpu_lock we may not see a recent store to that,
 603          * so it's possible that we at times can grant a weak binding to a
 604          * cpu that is an offline target, but that one request will not
 605          * prevent the offline from succeeding) then we will always grant
 606          * the weak binding request.  This includes the case above where
 607          * we grant a weakbinding not commensurate with our strong binding.
 608          *
 609          * If our cpu does appear to be an offline target then we're inclined
 610          * not to grant the weakbinding request just yet - we'd prefer to
 611          * migrate to another cpu and grant the request there.  The
 612          * exceptions are those cases where going through preemption code
 613          * will not result in us changing cpu:
 614          *
 615          *      . interrupts have already bypassed this case (see above)
 616          *      . we are already weakbound to this cpu (dispatcher code will
 617          *        always return us to the weakbound cpu)
 618          *      . preemption was disabled even before we disabled it above
 619          *      . we are strongbound to this cpu (if we're strongbound to
 620          *      another and not yet running there the trip through the
 621          *      dispatcher will move us to the strongbound cpu and we
 622          *      will grant the weak binding there)
 623          */
 624         if (cp != cpu_inmotion || t->t_nomigrate > 0 || t->t_preempt > 1 ||
 625             t->t_bound_cpu == cp) {
 626                 /*
 627                  * Don't be tempted to store to t_weakbound_cpu only on
 628                  * the first nested bind request - if we're interrupted
 629                  * after the increment of t_nomigrate and before the
 630                  * store to t_weakbound_cpu and the interrupt calls
 631                  * thread_nomigrate then the assertion in thread_allowmigrate
 632                  * would fail.
 633                  */
 634                 t->t_nomigrate++;
 635                 t->t_weakbound_cpu = cp;
 636                 membar_producer();
 637                 thread_unlock(curthread);
 638                 /*
 639                  * Now that we have dropped the thread_lock another thread
 640                  * can set our t_weakbound_cpu, and will try to migrate us
 641                  * to the strongbound cpu (which will not be prevented by
 642                  * preemption being disabled since we're about to enable
 643                  * preemption).  We have granted the weakbinding to the current
 644                  * cpu, so again we are in the position that is is is possible
 645                  * that our weak and strong bindings differ.  Again this
 646                  * is catered for by dispatcher code which will favour our
 647                  * weak binding.
 648                  */
 649                 kpreempt_enable();
 650         } else {
 651                 /*
 652                  * Move to another cpu before granting the request by
 653                  * forcing this thread through preemption code.  When we
 654                  * get to set{front,back}dq called from CL_PREEMPT()
 655                  * cpu_choose() will be used to select a cpu to queue
 656                  * us on - that will see cpu_inmotion and take
 657                  * steps to avoid returning us to this cpu.
 658                  */
 659                 cp->cpu_kprunrun = 1;
 660                 thread_unlock(curthread);
 661                 kpreempt_enable();      /* will call preempt() */
 662                 goto again;
 663         }
 664 }
 665 
 666 void
 667 thread_allowmigrate(void)
 668 {
 669         kthread_id_t t = curthread;
 670 
 671         ASSERT(t->t_weakbound_cpu == CPU ||
 672             (t->t_nomigrate < 0 && t->t_preempt > 0) ||
 673             CPU_ON_INTR(CPU) || t->t_flag & T_INTR_THREAD ||
 674             getpil() >= DISP_LEVEL);
 675 
 676         if (CPU_ON_INTR(CPU) || (t->t_flag & T_INTR_THREAD) ||
 677             getpil() >= DISP_LEVEL)
 678                 return;
 679 
 680         if (t->t_nomigrate < 0) {
 681                 /*
 682                  * This thread was granted "weak binding" in the
 683                  * stronger form of kernel preemption disabling.
 684                  * Undo a level of nesting for both t_nomigrate
 685                  * and t_preempt.
 686                  */
 687                 ++t->t_nomigrate;
 688                 kpreempt_enable();
 689         } else if (--t->t_nomigrate == 0) {
 690                 /*
 691                  * Time to drop the weak binding.  We need to cater
 692                  * for the case where we're weakbound to a different
 693                  * cpu than that to which we're strongbound (a very
 694                  * temporary arrangement that must only persist until
 695                  * weak binding drops).  We don't acquire thread_lock
 696                  * here so even as this code executes t_bound_cpu
 697                  * may be changing.  So we disable preemption and
 698                  * a) in the case that t_bound_cpu changes while we
 699                  * have preemption disabled kprunrun will be set
 700                  * asynchronously, and b) if before disabling
 701                  * preemption we were already on a different cpu to
 702                  * our t_bound_cpu then we set kprunrun ourselves
 703                  * to force a trip through the dispatcher when
 704                  * preemption is enabled.
 705                  */
 706                 kpreempt_disable();
 707                 if (t->t_bound_cpu &&
 708                     t->t_weakbound_cpu != t->t_bound_cpu)
 709                         CPU->cpu_kprunrun = 1;
 710                 t->t_weakbound_cpu = NULL;
 711                 membar_producer();
 712                 kpreempt_enable();
 713         }
 714 }
 715 
 716 /*
 717  * weakbinding_stop can be used to temporarily cause weakbindings made
 718  * with thread_nomigrate to be satisfied through the stronger action of
 719  * kpreempt_disable.  weakbinding_start recommences normal weakbinding.
 720  */
 721 
 722 void
 723 weakbinding_stop(void)
 724 {
 725         ASSERT(MUTEX_HELD(&cpu_lock));
 726         weakbindingbarrier = 1;
 727         membar_producer();      /* make visible before subsequent thread_lock */
 728 }
 729 
 730 void
 731 weakbinding_start(void)
 732 {
 733         ASSERT(MUTEX_HELD(&cpu_lock));
 734         weakbindingbarrier = 0;
 735 }
 736 
 737 void
 738 null_xcall(void)
 739 {
 740 }
 741 
 742 /*
 743  * This routine is called to place the CPUs in a safe place so that
 744  * one of them can be taken off line or placed on line.  What we are
 745  * trying to do here is prevent a thread from traversing the list
 746  * of active CPUs while we are changing it or from getting placed on
 747  * the run queue of a CPU that has just gone off line.  We do this by
 748  * creating a thread with the highest possible prio for each CPU and
 749  * having it call this routine.  The advantage of this method is that
 750  * we can eliminate all checks for CPU_ACTIVE in the disp routines.
 751  * This makes disp faster at the expense of making p_online() slower
 752  * which is a good trade off.
 753  */
 754 static void
 755 cpu_pause(int index)
 756 {
 757         int s;
 758         struct _cpu_pause_info *cpi = &cpu_pause_info;
 759         volatile char *safe = &safe_list[index];
 760         long    lindex = index;
 761 
 762         ASSERT((curthread->t_bound_cpu != NULL) || (*safe == PAUSE_DIE));
 763 
 764         while (*safe != PAUSE_DIE) {
 765                 *safe = PAUSE_READY;
 766                 membar_enter();         /* make sure stores are flushed */
 767                 sema_v(&cpi->cp_sem);    /* signal requesting thread */
 768 
 769                 /*
 770                  * Wait here until all pause threads are running.  That
 771                  * indicates that it's safe to do the spl.  Until
 772                  * cpu_pause_info.cp_go is set, we don't want to spl
 773                  * because that might block clock interrupts needed
 774                  * to preempt threads on other CPUs.
 775                  */
 776                 while (cpi->cp_go == 0)
 777                         ;
 778                 /*
 779                  * Even though we are at the highest disp prio, we need
 780                  * to block out all interrupts below LOCK_LEVEL so that
 781                  * an intr doesn't come in, wake up a thread, and call
 782                  * setbackdq/setfrontdq.
 783                  */
 784                 s = splhigh();
 785                 /*
 786                  * if cp_func has been set then call it using index as the
 787                  * argument, currently only used by cpr_suspend_cpus().
 788                  * This function is used as the code to execute on the
 789                  * "paused" cpu's when a machine comes out of a sleep state
 790                  * and CPU's were powered off.  (could also be used for
 791                  * hotplugging CPU's).
 792                  */
 793                 if (cpi->cp_func != NULL)
 794                         (*cpi->cp_func)((void *)lindex);
 795 
 796                 mach_cpu_pause(safe);
 797 
 798                 splx(s);
 799                 /*
 800                  * Waiting is at an end. Switch out of cpu_pause
 801                  * loop and resume useful work.
 802                  */
 803                 swtch();
 804         }
 805 
 806         mutex_enter(&pause_free_mutex);
 807         *safe = PAUSE_DEAD;
 808         cv_broadcast(&pause_free_cv);
 809         mutex_exit(&pause_free_mutex);
 810 }
 811 
 812 /*
 813  * Allow the cpus to start running again.
 814  */
 815 void
 816 start_cpus()
 817 {
 818         int i;
 819 
 820         ASSERT(MUTEX_HELD(&cpu_lock));
 821         ASSERT(cpu_pause_info.cp_paused);
 822         cpu_pause_info.cp_paused = NULL;
 823         for (i = 0; i < NCPU; i++)
 824                 safe_list[i] = PAUSE_IDLE;
 825         membar_enter();                 /* make sure stores are flushed */
 826         affinity_clear();
 827         splx(cpu_pause_info.cp_spl);
 828         kpreempt_enable();
 829 }
 830 
 831 /*
 832  * Allocate a pause thread for a CPU.
 833  */
 834 static void
 835 cpu_pause_alloc(cpu_t *cp)
 836 {
 837         kthread_id_t    t;
 838         long            cpun = cp->cpu_id;
 839 
 840         /*
 841          * Note, v.v_nglobpris will not change value as long as I hold
 842          * cpu_lock.
 843          */
 844         t = thread_create(NULL, 0, cpu_pause, (void *)cpun,
 845             0, &p0, TS_STOPPED, v.v_nglobpris - 1);
 846         thread_lock(t);
 847         t->t_bound_cpu = cp;
 848         t->t_disp_queue = cp->cpu_disp;
 849         t->t_affinitycnt = 1;
 850         t->t_preempt = 1;
 851         thread_unlock(t);
 852         cp->cpu_pause_thread = t;
 853         /*
 854          * Registering a thread in the callback table is usually done
 855          * in the initialization code of the thread.  In this
 856          * case, we do it right after thread creation because the
 857          * thread itself may never run, and we need to register the
 858          * fact that it is safe for cpr suspend.
 859          */
 860         CALLB_CPR_INIT_SAFE(t, "cpu_pause");
 861 }
 862 
 863 /*
 864  * Free a pause thread for a CPU.
 865  */
 866 static void
 867 cpu_pause_free(cpu_t *cp)
 868 {
 869         kthread_id_t    t;
 870         int             cpun = cp->cpu_id;
 871 
 872         ASSERT(MUTEX_HELD(&cpu_lock));
 873         /*
 874          * We have to get the thread and tell him to die.
 875          */
 876         if ((t = cp->cpu_pause_thread) == NULL) {
 877                 ASSERT(safe_list[cpun] == PAUSE_IDLE);
 878                 return;
 879         }
 880         thread_lock(t);
 881         t->t_cpu = CPU;              /* disp gets upset if last cpu is quiesced. */
 882         t->t_bound_cpu = NULL;       /* Must un-bind; cpu may not be running. */
 883         t->t_pri = v.v_nglobpris - 1;
 884         ASSERT(safe_list[cpun] == PAUSE_IDLE);
 885         safe_list[cpun] = PAUSE_DIE;
 886         THREAD_TRANSITION(t);
 887         setbackdq(t);
 888         thread_unlock_nopreempt(t);
 889 
 890         /*
 891          * If we don't wait for the thread to actually die, it may try to
 892          * run on the wrong cpu as part of an actual call to pause_cpus().
 893          */
 894         mutex_enter(&pause_free_mutex);
 895         while (safe_list[cpun] != PAUSE_DEAD) {
 896                 cv_wait(&pause_free_cv, &pause_free_mutex);
 897         }
 898         mutex_exit(&pause_free_mutex);
 899         safe_list[cpun] = PAUSE_IDLE;
 900 
 901         cp->cpu_pause_thread = NULL;
 902 }
 903 
 904 /*
 905  * Initialize basic structures for pausing CPUs.
 906  */
 907 void
 908 cpu_pause_init()
 909 {
 910         sema_init(&cpu_pause_info.cp_sem, 0, NULL, SEMA_DEFAULT, NULL);
 911         /*
 912          * Create initial CPU pause thread.
 913          */
 914         cpu_pause_alloc(CPU);
 915 }
 916 
 917 /*
 918  * Start the threads used to pause another CPU.
 919  */
 920 static int
 921 cpu_pause_start(processorid_t cpu_id)
 922 {
 923         int     i;
 924         int     cpu_count = 0;
 925 
 926         for (i = 0; i < NCPU; i++) {
 927                 cpu_t           *cp;
 928                 kthread_id_t    t;
 929 
 930                 cp = cpu[i];
 931                 if (!CPU_IN_SET(cpu_available, i) || (i == cpu_id)) {
 932                         safe_list[i] = PAUSE_WAIT;
 933                         continue;
 934                 }
 935 
 936                 /*
 937                  * Skip CPU if it is quiesced or not yet started.
 938                  */
 939                 if ((cp->cpu_flags & (CPU_QUIESCED | CPU_READY)) != CPU_READY) {
 940                         safe_list[i] = PAUSE_WAIT;
 941                         continue;
 942                 }
 943 
 944                 /*
 945                  * Start this CPU's pause thread.
 946                  */
 947                 t = cp->cpu_pause_thread;
 948                 thread_lock(t);
 949                 /*
 950                  * Reset the priority, since nglobpris may have
 951                  * changed since the thread was created, if someone
 952                  * has loaded the RT (or some other) scheduling
 953                  * class.
 954                  */
 955                 t->t_pri = v.v_nglobpris - 1;
 956                 THREAD_TRANSITION(t);
 957                 setbackdq(t);
 958                 thread_unlock_nopreempt(t);
 959                 ++cpu_count;
 960         }
 961         return (cpu_count);
 962 }
 963 
 964 
 965 /*
 966  * Pause all of the CPUs except the one we are on by creating a high
 967  * priority thread bound to those CPUs.
 968  *
 969  * Note that one must be extremely careful regarding code
 970  * executed while CPUs are paused.  Since a CPU may be paused
 971  * while a thread scheduling on that CPU is holding an adaptive
 972  * lock, code executed with CPUs paused must not acquire adaptive
 973  * (or low-level spin) locks.  Also, such code must not block,
 974  * since the thread that is supposed to initiate the wakeup may
 975  * never run.
 976  *
 977  * With a few exceptions, the restrictions on code executed with CPUs
 978  * paused match those for code executed at high-level interrupt
 979  * context.
 980  */
 981 void
 982 pause_cpus(cpu_t *off_cp, void *(*func)(void *))
 983 {
 984         processorid_t   cpu_id;
 985         int             i;
 986         struct _cpu_pause_info  *cpi = &cpu_pause_info;
 987 
 988         ASSERT(MUTEX_HELD(&cpu_lock));
 989         ASSERT(cpi->cp_paused == NULL);
 990         cpi->cp_count = 0;
 991         cpi->cp_go = 0;
 992         for (i = 0; i < NCPU; i++)
 993                 safe_list[i] = PAUSE_IDLE;
 994         kpreempt_disable();
 995 
 996         cpi->cp_func = func;
 997 
 998         /*
 999          * If running on the cpu that is going offline, get off it.
1000          * This is so that it won't be necessary to rechoose a CPU
1001          * when done.
1002          */
1003         if (CPU == off_cp)
1004                 cpu_id = off_cp->cpu_next_part->cpu_id;
1005         else
1006                 cpu_id = CPU->cpu_id;
1007         affinity_set(cpu_id);
1008 
1009         /*
1010          * Start the pause threads and record how many were started
1011          */
1012         cpi->cp_count = cpu_pause_start(cpu_id);
1013 
1014         /*
1015          * Now wait for all CPUs to be running the pause thread.
1016          */
1017         while (cpi->cp_count > 0) {
1018                 /*
1019                  * Spin reading the count without grabbing the disp
1020                  * lock to make sure we don't prevent the pause
1021                  * threads from getting the lock.
1022                  */
1023                 while (sema_held(&cpi->cp_sem))
1024                         ;
1025                 if (sema_tryp(&cpi->cp_sem))
1026                         --cpi->cp_count;
1027         }
1028         cpi->cp_go = 1;                      /* all have reached cpu_pause */
1029 
1030         /*
1031          * Now wait for all CPUs to spl. (Transition from PAUSE_READY
1032          * to PAUSE_WAIT.)
1033          */
1034         for (i = 0; i < NCPU; i++) {
1035                 while (safe_list[i] != PAUSE_WAIT)
1036                         ;
1037         }
1038         cpi->cp_spl = splhigh();     /* block dispatcher on this CPU */
1039         cpi->cp_paused = curthread;
1040 }
1041 
1042 /*
1043  * Check whether the current thread has CPUs paused
1044  */
1045 int
1046 cpus_paused(void)
1047 {
1048         if (cpu_pause_info.cp_paused != NULL) {
1049                 ASSERT(cpu_pause_info.cp_paused == curthread);
1050                 return (1);
1051         }
1052         return (0);
1053 }
1054 
1055 static cpu_t *
1056 cpu_get_all(processorid_t cpun)
1057 {
1058         ASSERT(MUTEX_HELD(&cpu_lock));
1059 
1060         if (cpun >= NCPU || cpun < 0 || !CPU_IN_SET(cpu_available, cpun))
1061                 return (NULL);
1062         return (cpu[cpun]);
1063 }
1064 
1065 /*
1066  * Check whether cpun is a valid processor id and whether it should be
1067  * visible from the current zone. If it is, return a pointer to the
1068  * associated CPU structure.
1069  */
1070 cpu_t *
1071 cpu_get(processorid_t cpun)
1072 {
1073         cpu_t *c;
1074 
1075         ASSERT(MUTEX_HELD(&cpu_lock));
1076         c = cpu_get_all(cpun);
1077         if (c != NULL && !INGLOBALZONE(curproc) && pool_pset_enabled() &&
1078             zone_pset_get(curproc->p_zone) != cpupart_query_cpu(c))
1079                 return (NULL);
1080         return (c);
1081 }
1082 
1083 /*
1084  * The following functions should be used to check CPU states in the kernel.
1085  * They should be invoked with cpu_lock held.  Kernel subsystems interested
1086  * in CPU states should *not* use cpu_get_state() and various P_ONLINE/etc
1087  * states.  Those are for user-land (and system call) use only.
1088  */
1089 
1090 /*
1091  * Determine whether the CPU is online and handling interrupts.
1092  */
1093 int
1094 cpu_is_online(cpu_t *cpu)
1095 {
1096         ASSERT(MUTEX_HELD(&cpu_lock));
1097         return (cpu_flagged_online(cpu->cpu_flags));
1098 }
1099 
1100 /*
1101  * Determine whether the CPU is offline (this includes spare and faulted).
1102  */
1103 int
1104 cpu_is_offline(cpu_t *cpu)
1105 {
1106         ASSERT(MUTEX_HELD(&cpu_lock));
1107         return (cpu_flagged_offline(cpu->cpu_flags));
1108 }
1109 
1110 /*
1111  * Determine whether the CPU is powered off.
1112  */
1113 int
1114 cpu_is_poweredoff(cpu_t *cpu)
1115 {
1116         ASSERT(MUTEX_HELD(&cpu_lock));
1117         return (cpu_flagged_poweredoff(cpu->cpu_flags));
1118 }
1119 
1120 /*
1121  * Determine whether the CPU is handling interrupts.
1122  */
1123 int
1124 cpu_is_nointr(cpu_t *cpu)
1125 {
1126         ASSERT(MUTEX_HELD(&cpu_lock));
1127         return (cpu_flagged_nointr(cpu->cpu_flags));
1128 }
1129 
1130 /*
1131  * Determine whether the CPU is active (scheduling threads).
1132  */
1133 int
1134 cpu_is_active(cpu_t *cpu)
1135 {
1136         ASSERT(MUTEX_HELD(&cpu_lock));
1137         return (cpu_flagged_active(cpu->cpu_flags));
1138 }
1139 
1140 /*
1141  * Same as above, but these require cpu_flags instead of cpu_t pointers.
1142  */
1143 int
1144 cpu_flagged_online(cpu_flag_t cpu_flags)
1145 {
1146         return (cpu_flagged_active(cpu_flags) &&
1147             (cpu_flags & CPU_ENABLE));
1148 }
1149 
1150 int
1151 cpu_flagged_offline(cpu_flag_t cpu_flags)
1152 {
1153         return (((cpu_flags & CPU_POWEROFF) == 0) &&
1154             ((cpu_flags & (CPU_READY | CPU_OFFLINE)) != CPU_READY));
1155 }
1156 
1157 int
1158 cpu_flagged_poweredoff(cpu_flag_t cpu_flags)
1159 {
1160         return ((cpu_flags & CPU_POWEROFF) == CPU_POWEROFF);
1161 }
1162 
1163 int
1164 cpu_flagged_nointr(cpu_flag_t cpu_flags)
1165 {
1166         return (cpu_flagged_active(cpu_flags) &&
1167             (cpu_flags & CPU_ENABLE) == 0);
1168 }
1169 
1170 int
1171 cpu_flagged_active(cpu_flag_t cpu_flags)
1172 {
1173         return (((cpu_flags & (CPU_POWEROFF | CPU_FAULTED | CPU_SPARE)) == 0) &&
1174             ((cpu_flags & (CPU_READY | CPU_OFFLINE)) == CPU_READY));
1175 }
1176 
1177 /*
1178  * Bring the indicated CPU online.
1179  */
1180 int
1181 cpu_online(cpu_t *cp)
1182 {
1183         int     error = 0;
1184 
1185         /*
1186          * Handle on-line request.
1187          *      This code must put the new CPU on the active list before
1188          *      starting it because it will not be paused, and will start
1189          *      using the active list immediately.  The real start occurs
1190          *      when the CPU_QUIESCED flag is turned off.
1191          */
1192 
1193         ASSERT(MUTEX_HELD(&cpu_lock));
1194 
1195         /*
1196          * Put all the cpus into a known safe place.
1197          * No mutexes can be entered while CPUs are paused.
1198          */
1199         error = mp_cpu_start(cp);       /* arch-dep hook */
1200         if (error == 0) {
1201                 pg_cpupart_in(cp, cp->cpu_part);
1202                 pause_cpus(NULL, NULL);
1203                 cpu_add_active_internal(cp);
1204                 if (cp->cpu_flags & CPU_FAULTED) {
1205                         cp->cpu_flags &= ~CPU_FAULTED;
1206                         mp_cpu_faulted_exit(cp);
1207                 }
1208                 cp->cpu_flags &= ~(CPU_QUIESCED | CPU_OFFLINE | CPU_FROZEN |
1209                     CPU_SPARE);
1210                 CPU_NEW_GENERATION(cp);
1211                 start_cpus();
1212                 cpu_stats_kstat_create(cp);
1213                 cpu_create_intrstat(cp);
1214                 lgrp_kstat_create(cp);
1215                 cpu_state_change_notify(cp->cpu_id, CPU_ON);
1216                 cpu_intr_enable(cp);    /* arch-dep hook */
1217                 cpu_state_change_notify(cp->cpu_id, CPU_INTR_ON);
1218                 cpu_set_state(cp);
1219                 cyclic_online(cp);
1220                 /*
1221                  * This has to be called only after cyclic_online(). This
1222                  * function uses cyclics.
1223                  */
1224                 callout_cpu_online(cp);
1225                 poke_cpu(cp->cpu_id);
1226         }
1227 
1228         return (error);
1229 }
1230 
1231 /*
1232  * Take the indicated CPU offline.
1233  */
1234 int
1235 cpu_offline(cpu_t *cp, int flags)
1236 {
1237         cpupart_t *pp;
1238         int     error = 0;
1239         cpu_t   *ncp;
1240         int     intr_enable;
1241         int     cyclic_off = 0;
1242         int     callout_off = 0;
1243         int     loop_count;
1244         int     no_quiesce = 0;
1245         int     (*bound_func)(struct cpu *, int);
1246         kthread_t *t;
1247         lpl_t   *cpu_lpl;
1248         proc_t  *p;
1249         int     lgrp_diff_lpl;
1250         boolean_t unbind_all_threads = (flags & CPU_FORCED) != 0;
1251 
1252         ASSERT(MUTEX_HELD(&cpu_lock));
1253 
1254         /*
1255          * If we're going from faulted or spare to offline, just
1256          * clear these flags and update CPU state.
1257          */
1258         if (cp->cpu_flags & (CPU_FAULTED | CPU_SPARE)) {
1259                 if (cp->cpu_flags & CPU_FAULTED) {
1260                         cp->cpu_flags &= ~CPU_FAULTED;
1261                         mp_cpu_faulted_exit(cp);
1262                 }
1263                 cp->cpu_flags &= ~CPU_SPARE;
1264                 cpu_set_state(cp);
1265                 return (0);
1266         }
1267 
1268         /*
1269          * Handle off-line request.
1270          */
1271         pp = cp->cpu_part;
1272         /*
1273          * Don't offline last online CPU in partition
1274          */
1275         if (ncpus_online <= 1 || pp->cp_ncpus <= 1 || cpu_intr_count(cp) < 2)
1276                 return (EBUSY);
1277         /*
1278          * Unbind all soft-bound threads bound to our CPU and hard bound threads
1279          * if we were asked to.
1280          */
1281         error = cpu_unbind(cp->cpu_id, unbind_all_threads);
1282         if (error != 0)
1283                 return (error);
1284         /*
1285          * We shouldn't be bound to this CPU ourselves.
1286          */
1287         if (curthread->t_bound_cpu == cp)
1288                 return (EBUSY);
1289 
1290         /*
1291          * Tell interested parties that this CPU is going offline.
1292          */
1293         CPU_NEW_GENERATION(cp);
1294         cpu_state_change_notify(cp->cpu_id, CPU_OFF);
1295 
1296         /*
1297          * Tell the PG subsystem that the CPU is leaving the partition
1298          */
1299         pg_cpupart_out(cp, pp);
1300 
1301         /*
1302          * Take the CPU out of interrupt participation so we won't find
1303          * bound kernel threads.  If the architecture cannot completely
1304          * shut off interrupts on the CPU, don't quiesce it, but don't
1305          * run anything but interrupt thread... this is indicated by
1306          * the CPU_OFFLINE flag being on but the CPU_QUIESCE flag being
1307          * off.
1308          */
1309         intr_enable = cp->cpu_flags & CPU_ENABLE;
1310         if (intr_enable)
1311                 no_quiesce = cpu_intr_disable(cp);
1312 
1313         /*
1314          * Record that we are aiming to offline this cpu.  This acts as
1315          * a barrier to further weak binding requests in thread_nomigrate
1316          * and also causes cpu_choose, disp_lowpri_cpu and setfrontdq to
1317          * lean away from this cpu.  Further strong bindings are already
1318          * avoided since we hold cpu_lock.  Since threads that are set
1319          * runnable around now and others coming off the target cpu are
1320          * directed away from the target, existing strong and weak bindings
1321          * (especially the latter) to the target cpu stand maximum chance of
1322          * being able to unbind during the short delay loop below (if other
1323          * unbound threads compete they may not see cpu in time to unbind
1324          * even if they would do so immediately.
1325          */
1326         cpu_inmotion = cp;
1327         membar_enter();
1328 
1329         /*
1330          * Check for kernel threads (strong or weak) bound to that CPU.
1331          * Strongly bound threads may not unbind, and we'll have to return
1332          * EBUSY.  Weakly bound threads should always disappear - we've
1333          * stopped more weak binding with cpu_inmotion and existing
1334          * bindings will drain imminently (they may not block).  Nonetheless
1335          * we will wait for a fixed period for all bound threads to disappear.
1336          * Inactive interrupt threads are OK (they'll be in TS_FREE
1337          * state).  If test finds some bound threads, wait a few ticks
1338          * to give short-lived threads (such as interrupts) chance to
1339          * complete.  Note that if no_quiesce is set, i.e. this cpu
1340          * is required to service interrupts, then we take the route
1341          * that permits interrupt threads to be active (or bypassed).
1342          */
1343         bound_func = no_quiesce ? disp_bound_threads : disp_bound_anythreads;
1344 
1345 again:  for (loop_count = 0; (*bound_func)(cp, 0); loop_count++) {
1346                 if (loop_count >= 5) {
1347                         error = EBUSY;  /* some threads still bound */
1348                         break;
1349                 }
1350 
1351                 /*
1352                  * If some threads were assigned, give them
1353                  * a chance to complete or move.
1354                  *
1355                  * This assumes that the clock_thread is not bound
1356                  * to any CPU, because the clock_thread is needed to
1357                  * do the delay(hz/100).
1358                  *
1359                  * Note: we still hold the cpu_lock while waiting for
1360                  * the next clock tick.  This is OK since it isn't
1361                  * needed for anything else except processor_bind(2),
1362                  * and system initialization.  If we drop the lock,
1363                  * we would risk another p_online disabling the last
1364                  * processor.
1365                  */
1366                 delay(hz/100);
1367         }
1368 
1369         if (error == 0 && callout_off == 0) {
1370                 callout_cpu_offline(cp);
1371                 callout_off = 1;
1372         }
1373 
1374         if (error == 0 && cyclic_off == 0) {
1375                 if (!cyclic_offline(cp)) {
1376                         /*
1377                          * We must have bound cyclics...
1378                          */
1379                         error = EBUSY;
1380                         goto out;
1381                 }
1382                 cyclic_off = 1;
1383         }
1384 
1385         /*
1386          * Call mp_cpu_stop() to perform any special operations
1387          * needed for this machine architecture to offline a CPU.
1388          */
1389         if (error == 0)
1390                 error = mp_cpu_stop(cp);        /* arch-dep hook */
1391 
1392         /*
1393          * If that all worked, take the CPU offline and decrement
1394          * ncpus_online.
1395          */
1396         if (error == 0) {
1397                 /*
1398                  * Put all the cpus into a known safe place.
1399                  * No mutexes can be entered while CPUs are paused.
1400                  */
1401                 pause_cpus(cp, NULL);
1402                 /*
1403                  * Repeat the operation, if necessary, to make sure that
1404                  * all outstanding low-level interrupts run to completion
1405                  * before we set the CPU_QUIESCED flag.  It's also possible
1406                  * that a thread has weak bound to the cpu despite our raising
1407                  * cpu_inmotion above since it may have loaded that
1408                  * value before the barrier became visible (this would have
1409                  * to be the thread that was on the target cpu at the time
1410                  * we raised the barrier).
1411                  */
1412                 if ((!no_quiesce && cp->cpu_intr_actv != 0) ||
1413                     (*bound_func)(cp, 1)) {
1414                         start_cpus();
1415                         (void) mp_cpu_start(cp);
1416                         goto again;
1417                 }
1418                 ncp = cp->cpu_next_part;
1419                 cpu_lpl = cp->cpu_lpl;
1420                 ASSERT(cpu_lpl != NULL);
1421 
1422                 /*
1423                  * Remove the CPU from the list of active CPUs.
1424                  */
1425                 cpu_remove_active(cp);
1426 
1427                 /*
1428                  * Walk the active process list and look for threads
1429                  * whose home lgroup needs to be updated, or
1430                  * the last CPU they run on is the one being offlined now.
1431                  */
1432 
1433                 ASSERT(curthread->t_cpu != cp);
1434                 for (p = practive; p != NULL; p = p->p_next) {
1435 
1436                         t = p->p_tlist;
1437 
1438                         if (t == NULL)
1439                                 continue;
1440 
1441                         lgrp_diff_lpl = 0;
1442 
1443                         do {
1444                                 ASSERT(t->t_lpl != NULL);
1445                                 /*
1446                                  * Taking last CPU in lpl offline
1447                                  * Rehome thread if it is in this lpl
1448                                  * Otherwise, update the count of how many
1449                                  * threads are in this CPU's lgroup but have
1450                                  * a different lpl.
1451                                  */
1452 
1453                                 if (cpu_lpl->lpl_ncpu == 0) {
1454                                         if (t->t_lpl == cpu_lpl)
1455                                                 lgrp_move_thread(t,
1456                                                     lgrp_choose(t,
1457                                                     t->t_cpupart), 0);
1458                                         else if (t->t_lpl->lpl_lgrpid ==
1459                                             cpu_lpl->lpl_lgrpid)
1460                                                 lgrp_diff_lpl++;
1461                                 }
1462                                 ASSERT(t->t_lpl->lpl_ncpu > 0);
1463 
1464                                 /*
1465                                  * Update CPU last ran on if it was this CPU
1466                                  */
1467                                 if (t->t_cpu == cp && t->t_bound_cpu != cp)
1468                                         t->t_cpu = disp_lowpri_cpu(ncp,
1469                                             t->t_lpl, t->t_pri, NULL);
1470                                 ASSERT(t->t_cpu != cp || t->t_bound_cpu == cp ||
1471                                     t->t_weakbound_cpu == cp);
1472 
1473                                 t = t->t_forw;
1474                         } while (t != p->p_tlist);
1475 
1476                         /*
1477                          * Didn't find any threads in the same lgroup as this
1478                          * CPU with a different lpl, so remove the lgroup from
1479                          * the process lgroup bitmask.
1480                          */
1481 
1482                         if (lgrp_diff_lpl == 0)
1483                                 klgrpset_del(p->p_lgrpset, cpu_lpl->lpl_lgrpid);
1484                 }
1485 
1486                 /*
1487                  * Walk thread list looking for threads that need to be
1488                  * rehomed, since there are some threads that are not in
1489                  * their process's p_tlist.
1490                  */
1491 
1492                 t = curthread;
1493                 do {
1494                         ASSERT(t != NULL && t->t_lpl != NULL);
1495 
1496                         /*
1497                          * Rehome threads with same lpl as this CPU when this
1498                          * is the last CPU in the lpl.
1499                          */
1500 
1501                         if ((cpu_lpl->lpl_ncpu == 0) && (t->t_lpl == cpu_lpl))
1502                                 lgrp_move_thread(t,
1503                                     lgrp_choose(t, t->t_cpupart), 1);
1504 
1505                         ASSERT(t->t_lpl->lpl_ncpu > 0);
1506 
1507                         /*
1508                          * Update CPU last ran on if it was this CPU
1509                          */
1510 
1511                         if (t->t_cpu == cp && t->t_bound_cpu != cp) {
1512                                 t->t_cpu = disp_lowpri_cpu(ncp,
1513                                     t->t_lpl, t->t_pri, NULL);
1514                         }
1515                         ASSERT(t->t_cpu != cp || t->t_bound_cpu == cp ||
1516                             t->t_weakbound_cpu == cp);
1517                         t = t->t_next;
1518 
1519                 } while (t != curthread);
1520                 ASSERT((cp->cpu_flags & (CPU_FAULTED | CPU_SPARE)) == 0);
1521                 cp->cpu_flags |= CPU_OFFLINE;
1522                 disp_cpu_inactive(cp);
1523                 if (!no_quiesce)
1524                         cp->cpu_flags |= CPU_QUIESCED;
1525                 ncpus_online--;
1526                 cpu_set_state(cp);
1527                 cpu_inmotion = NULL;
1528                 start_cpus();
1529                 cpu_stats_kstat_destroy(cp);
1530                 cpu_delete_intrstat(cp);
1531                 lgrp_kstat_destroy(cp);
1532         }
1533 
1534 out:
1535         cpu_inmotion = NULL;
1536 
1537         /*
1538          * If we failed, re-enable interrupts.
1539          * Do this even if cpu_intr_disable returned an error, because
1540          * it may have partially disabled interrupts.
1541          */
1542         if (error && intr_enable)
1543                 cpu_intr_enable(cp);
1544 
1545         /*
1546          * If we failed, but managed to offline the cyclic subsystem on this
1547          * CPU, bring it back online.
1548          */
1549         if (error && cyclic_off)
1550                 cyclic_online(cp);
1551 
1552         /*
1553          * If we failed, but managed to offline callouts on this CPU,
1554          * bring it back online.
1555          */
1556         if (error && callout_off)
1557                 callout_cpu_online(cp);
1558 
1559         /*
1560          * If we failed, tell the PG subsystem that the CPU is back
1561          */
1562         pg_cpupart_in(cp, pp);
1563 
1564         /*
1565          * If we failed, we need to notify everyone that this CPU is back on.
1566          */
1567         if (error != 0) {
1568                 CPU_NEW_GENERATION(cp);
1569                 cpu_state_change_notify(cp->cpu_id, CPU_ON);
1570                 cpu_state_change_notify(cp->cpu_id, CPU_INTR_ON);
1571         }
1572 
1573         return (error);
1574 }
1575 
1576 /*
1577  * Mark the indicated CPU as faulted, taking it offline.
1578  */
1579 int
1580 cpu_faulted(cpu_t *cp, int flags)
1581 {
1582         int     error = 0;
1583 
1584         ASSERT(MUTEX_HELD(&cpu_lock));
1585         ASSERT(!cpu_is_poweredoff(cp));
1586 
1587         if (cpu_is_offline(cp)) {
1588                 cp->cpu_flags &= ~CPU_SPARE;
1589                 cp->cpu_flags |= CPU_FAULTED;
1590                 mp_cpu_faulted_enter(cp);
1591                 cpu_set_state(cp);
1592                 return (0);
1593         }
1594 
1595         if ((error = cpu_offline(cp, flags)) == 0) {
1596                 cp->cpu_flags |= CPU_FAULTED;
1597                 mp_cpu_faulted_enter(cp);
1598                 cpu_set_state(cp);
1599         }
1600 
1601         return (error);
1602 }
1603 
1604 /*
1605  * Mark the indicated CPU as a spare, taking it offline.
1606  */
1607 int
1608 cpu_spare(cpu_t *cp, int flags)
1609 {
1610         int     error = 0;
1611 
1612         ASSERT(MUTEX_HELD(&cpu_lock));
1613         ASSERT(!cpu_is_poweredoff(cp));
1614 
1615         if (cpu_is_offline(cp)) {
1616                 if (cp->cpu_flags & CPU_FAULTED) {
1617                         cp->cpu_flags &= ~CPU_FAULTED;
1618                         mp_cpu_faulted_exit(cp);
1619                 }
1620                 cp->cpu_flags |= CPU_SPARE;
1621                 cpu_set_state(cp);
1622                 return (0);
1623         }
1624 
1625         if ((error = cpu_offline(cp, flags)) == 0) {
1626                 cp->cpu_flags |= CPU_SPARE;
1627                 cpu_set_state(cp);
1628         }
1629 
1630         return (error);
1631 }
1632 
1633 /*
1634  * Take the indicated CPU from poweroff to offline.
1635  */
1636 int
1637 cpu_poweron(cpu_t *cp)
1638 {
1639         int     error = ENOTSUP;
1640 
1641         ASSERT(MUTEX_HELD(&cpu_lock));
1642         ASSERT(cpu_is_poweredoff(cp));
1643 
1644         error = mp_cpu_poweron(cp);     /* arch-dep hook */
1645         if (error == 0)
1646                 cpu_set_state(cp);
1647 
1648         return (error);
1649 }
1650 
1651 /*
1652  * Take the indicated CPU from any inactive state to powered off.
1653  */
1654 int
1655 cpu_poweroff(cpu_t *cp)
1656 {
1657         int     error = ENOTSUP;
1658 
1659         ASSERT(MUTEX_HELD(&cpu_lock));
1660         ASSERT(cpu_is_offline(cp));
1661 
1662         if (!(cp->cpu_flags & CPU_QUIESCED))
1663                 return (EBUSY);         /* not completely idle */
1664 
1665         error = mp_cpu_poweroff(cp);    /* arch-dep hook */
1666         if (error == 0)
1667                 cpu_set_state(cp);
1668 
1669         return (error);
1670 }
1671 
1672 /*
1673  * Initialize the Sequential CPU id lookup table
1674  */
1675 void
1676 cpu_seq_tbl_init()
1677 {
1678         cpu_t   **tbl;
1679 
1680         tbl = kmem_zalloc(sizeof (struct cpu *) * max_ncpus, KM_SLEEP);
1681         tbl[0] = CPU;
1682 
1683         cpu_seq = tbl;
1684 }
1685 
1686 /*
1687  * Initialize the CPU lists for the first CPU.
1688  */
1689 void
1690 cpu_list_init(cpu_t *cp)
1691 {
1692         cp->cpu_next = cp;
1693         cp->cpu_prev = cp;
1694         cpu_list = cp;
1695         clock_cpu_list = cp;
1696 
1697         cp->cpu_next_onln = cp;
1698         cp->cpu_prev_onln = cp;
1699         cpu_active = cp;
1700 
1701         cp->cpu_seqid = 0;
1702         CPUSET_ADD(cpu_seqid_inuse, 0);
1703 
1704         /*
1705          * Bootstrap cpu_seq using cpu_list
1706          * The cpu_seq[] table will be dynamically allocated
1707          * when kmem later becomes available (but before going MP)
1708          */
1709         cpu_seq = &cpu_list;
1710 
1711         cp->cpu_cache_offset = KMEM_CPU_CACHE_OFFSET(cp->cpu_seqid);
1712         cp_default.cp_cpulist = cp;
1713         cp_default.cp_ncpus = 1;
1714         cp->cpu_next_part = cp;
1715         cp->cpu_prev_part = cp;
1716         cp->cpu_part = &cp_default;
1717 
1718         CPUSET_ADD(cpu_available, cp->cpu_id);
1719 }
1720 
1721 /*
1722  * Insert a CPU into the list of available CPUs.
1723  */
1724 void
1725 cpu_add_unit(cpu_t *cp)
1726 {
1727         int seqid;
1728 
1729         ASSERT(MUTEX_HELD(&cpu_lock));
1730         ASSERT(cpu_list != NULL);       /* list started in cpu_list_init */
1731 
1732         lgrp_config(LGRP_CONFIG_CPU_ADD, (uintptr_t)cp, 0);
1733 
1734         /*
1735          * Note: most users of the cpu_list will grab the
1736          * cpu_lock to insure that it isn't modified.  However,
1737          * certain users can't or won't do that.  To allow this
1738          * we pause the other cpus.  Users who walk the list
1739          * without cpu_lock, must disable kernel preemption
1740          * to insure that the list isn't modified underneath
1741          * them.  Also, any cached pointers to cpu structures
1742          * must be revalidated by checking to see if the
1743          * cpu_next pointer points to itself.  This check must
1744          * be done with the cpu_lock held or kernel preemption
1745          * disabled.  This check relies upon the fact that
1746          * old cpu structures are not free'ed or cleared after
1747          * then are removed from the cpu_list.
1748          *
1749          * Note that the clock code walks the cpu list dereferencing
1750          * the cpu_part pointer, so we need to initialize it before
1751          * adding the cpu to the list.
1752          */
1753         cp->cpu_part = &cp_default;
1754         pause_cpus(NULL, NULL);
1755         cp->cpu_next = cpu_list;
1756         cp->cpu_prev = cpu_list->cpu_prev;
1757         cpu_list->cpu_prev->cpu_next = cp;
1758         cpu_list->cpu_prev = cp;
1759         start_cpus();
1760 
1761         for (seqid = 0; CPU_IN_SET(cpu_seqid_inuse, seqid); seqid++)
1762                 continue;
1763         CPUSET_ADD(cpu_seqid_inuse, seqid);
1764         cp->cpu_seqid = seqid;
1765 
1766         if (seqid > max_cpu_seqid_ever)
1767                 max_cpu_seqid_ever = seqid;
1768 
1769         ASSERT(ncpus < max_ncpus);
1770         ncpus++;
1771         cp->cpu_cache_offset = KMEM_CPU_CACHE_OFFSET(cp->cpu_seqid);
1772         cpu[cp->cpu_id] = cp;
1773         CPUSET_ADD(cpu_available, cp->cpu_id);
1774         cpu_seq[cp->cpu_seqid] = cp;
1775 
1776         /*
1777          * allocate a pause thread for this CPU.
1778          */
1779         cpu_pause_alloc(cp);
1780 
1781         /*
1782          * So that new CPUs won't have NULL prev_onln and next_onln pointers,
1783          * link them into a list of just that CPU.
1784          * This is so that disp_lowpri_cpu will work for thread_create in
1785          * pause_cpus() when called from the startup thread in a new CPU.
1786          */
1787         cp->cpu_next_onln = cp;
1788         cp->cpu_prev_onln = cp;
1789         cpu_info_kstat_create(cp);
1790         cp->cpu_next_part = cp;
1791         cp->cpu_prev_part = cp;
1792 
1793         init_cpu_mstate(cp, CMS_SYSTEM);
1794 
1795         pool_pset_mod = gethrtime();
1796 }
1797 
1798 /*
1799  * Do the opposite of cpu_add_unit().
1800  */
1801 void
1802 cpu_del_unit(int cpuid)
1803 {
1804         struct cpu      *cp, *cpnext;
1805 
1806         ASSERT(MUTEX_HELD(&cpu_lock));
1807         cp = cpu[cpuid];
1808         ASSERT(cp != NULL);
1809 
1810         ASSERT(cp->cpu_next_onln == cp);
1811         ASSERT(cp->cpu_prev_onln == cp);
1812         ASSERT(cp->cpu_next_part == cp);
1813         ASSERT(cp->cpu_prev_part == cp);
1814 
1815         /*
1816          * Tear down the CPU's physical ID cache, and update any
1817          * processor groups
1818          */
1819         pg_cpu_fini(cp, NULL);
1820         pghw_physid_destroy(cp);
1821 
1822         /*
1823          * Destroy kstat stuff.
1824          */
1825         cpu_info_kstat_destroy(cp);
1826         term_cpu_mstate(cp);
1827         /*
1828          * Free up pause thread.
1829          */
1830         cpu_pause_free(cp);
1831         CPUSET_DEL(cpu_available, cp->cpu_id);
1832         cpu[cp->cpu_id] = NULL;
1833         cpu_seq[cp->cpu_seqid] = NULL;
1834 
1835         /*
1836          * The clock thread and mutex_vector_enter cannot hold the
1837          * cpu_lock while traversing the cpu list, therefore we pause
1838          * all other threads by pausing the other cpus. These, and any
1839          * other routines holding cpu pointers while possibly sleeping
1840          * must be sure to call kpreempt_disable before processing the
1841          * list and be sure to check that the cpu has not been deleted
1842          * after any sleeps (check cp->cpu_next != NULL). We guarantee
1843          * to keep the deleted cpu structure around.
1844          *
1845          * Note that this MUST be done AFTER cpu_available
1846          * has been updated so that we don't waste time
1847          * trying to pause the cpu we're trying to delete.
1848          */
1849         pause_cpus(NULL, NULL);
1850 
1851         cpnext = cp->cpu_next;
1852         cp->cpu_prev->cpu_next = cp->cpu_next;
1853         cp->cpu_next->cpu_prev = cp->cpu_prev;
1854         if (cp == cpu_list)
1855                 cpu_list = cpnext;
1856 
1857         /*
1858          * Signals that the cpu has been deleted (see above).
1859          */
1860         cp->cpu_next = NULL;
1861         cp->cpu_prev = NULL;
1862 
1863         start_cpus();
1864 
1865         CPUSET_DEL(cpu_seqid_inuse, cp->cpu_seqid);
1866         ncpus--;
1867         lgrp_config(LGRP_CONFIG_CPU_DEL, (uintptr_t)cp, 0);
1868 
1869         pool_pset_mod = gethrtime();
1870 }
1871 
1872 /*
1873  * Add a CPU to the list of active CPUs.
1874  *      This routine must not get any locks, because other CPUs are paused.
1875  */
1876 static void
1877 cpu_add_active_internal(cpu_t *cp)
1878 {
1879         cpupart_t       *pp = cp->cpu_part;
1880 
1881         ASSERT(MUTEX_HELD(&cpu_lock));
1882         ASSERT(cpu_list != NULL);       /* list started in cpu_list_init */
1883 
1884         ncpus_online++;
1885         cpu_set_state(cp);
1886         cp->cpu_next_onln = cpu_active;
1887         cp->cpu_prev_onln = cpu_active->cpu_prev_onln;
1888         cpu_active->cpu_prev_onln->cpu_next_onln = cp;
1889         cpu_active->cpu_prev_onln = cp;
1890 
1891         if (pp->cp_cpulist) {
1892                 cp->cpu_next_part = pp->cp_cpulist;
1893                 cp->cpu_prev_part = pp->cp_cpulist->cpu_prev_part;
1894                 pp->cp_cpulist->cpu_prev_part->cpu_next_part = cp;
1895                 pp->cp_cpulist->cpu_prev_part = cp;
1896         } else {
1897                 ASSERT(pp->cp_ncpus == 0);
1898                 pp->cp_cpulist = cp->cpu_next_part = cp->cpu_prev_part = cp;
1899         }
1900         pp->cp_ncpus++;
1901         if (pp->cp_ncpus == 1) {
1902                 cp_numparts_nonempty++;
1903                 ASSERT(cp_numparts_nonempty != 0);
1904         }
1905 
1906         pg_cpu_active(cp);
1907         lgrp_config(LGRP_CONFIG_CPU_ONLINE, (uintptr_t)cp, 0);
1908 
1909         bzero(&cp->cpu_loadavg, sizeof (cp->cpu_loadavg));
1910 }
1911 
1912 /*
1913  * Add a CPU to the list of active CPUs.
1914  *      This is called from machine-dependent layers when a new CPU is started.
1915  */
1916 void
1917 cpu_add_active(cpu_t *cp)
1918 {
1919         pg_cpupart_in(cp, cp->cpu_part);
1920 
1921         pause_cpus(NULL, NULL);
1922         cpu_add_active_internal(cp);
1923         start_cpus();
1924 
1925         cpu_stats_kstat_create(cp);
1926         cpu_create_intrstat(cp);
1927         lgrp_kstat_create(cp);
1928         cpu_state_change_notify(cp->cpu_id, CPU_INIT);
1929 }
1930 
1931 
1932 /*
1933  * Remove a CPU from the list of active CPUs.
1934  *      This routine must not get any locks, because other CPUs are paused.
1935  */
1936 /* ARGSUSED */
1937 static void
1938 cpu_remove_active(cpu_t *cp)
1939 {
1940         cpupart_t       *pp = cp->cpu_part;
1941 
1942         ASSERT(MUTEX_HELD(&cpu_lock));
1943         ASSERT(cp->cpu_next_onln != cp);     /* not the last one */
1944         ASSERT(cp->cpu_prev_onln != cp);     /* not the last one */
1945 
1946         pg_cpu_inactive(cp);
1947 
1948         lgrp_config(LGRP_CONFIG_CPU_OFFLINE, (uintptr_t)cp, 0);
1949 
1950         if (cp == clock_cpu_list)
1951                 clock_cpu_list = cp->cpu_next_onln;
1952 
1953         cp->cpu_prev_onln->cpu_next_onln = cp->cpu_next_onln;
1954         cp->cpu_next_onln->cpu_prev_onln = cp->cpu_prev_onln;
1955         if (cpu_active == cp) {
1956                 cpu_active = cp->cpu_next_onln;
1957         }
1958         cp->cpu_next_onln = cp;
1959         cp->cpu_prev_onln = cp;
1960 
1961         cp->cpu_prev_part->cpu_next_part = cp->cpu_next_part;
1962         cp->cpu_next_part->cpu_prev_part = cp->cpu_prev_part;
1963         if (pp->cp_cpulist == cp) {
1964                 pp->cp_cpulist = cp->cpu_next_part;
1965                 ASSERT(pp->cp_cpulist != cp);
1966         }
1967         cp->cpu_next_part = cp;
1968         cp->cpu_prev_part = cp;
1969         pp->cp_ncpus--;
1970         if (pp->cp_ncpus == 0) {
1971                 cp_numparts_nonempty--;
1972                 ASSERT(cp_numparts_nonempty != 0);
1973         }
1974 }
1975 
1976 /*
1977  * Routine used to setup a newly inserted CPU in preparation for starting
1978  * it running code.
1979  */
1980 int
1981 cpu_configure(int cpuid)
1982 {
1983         int retval = 0;
1984 
1985         ASSERT(MUTEX_HELD(&cpu_lock));
1986 
1987         /*
1988          * Some structures are statically allocated based upon
1989          * the maximum number of cpus the system supports.  Do not
1990          * try to add anything beyond this limit.
1991          */
1992         if (cpuid < 0 || cpuid >= NCPU) {
1993                 return (EINVAL);
1994         }
1995 
1996         if ((cpu[cpuid] != NULL) && (cpu[cpuid]->cpu_flags != 0)) {
1997                 return (EALREADY);
1998         }
1999 
2000         if ((retval = mp_cpu_configure(cpuid)) != 0) {
2001                 return (retval);
2002         }
2003 
2004         cpu[cpuid]->cpu_flags = CPU_QUIESCED | CPU_OFFLINE | CPU_POWEROFF;
2005         cpu_set_state(cpu[cpuid]);
2006         retval = cpu_state_change_hooks(cpuid, CPU_CONFIG, CPU_UNCONFIG);
2007         if (retval != 0)
2008                 (void) mp_cpu_unconfigure(cpuid);
2009 
2010         return (retval);
2011 }
2012 
2013 /*
2014  * Routine used to cleanup a CPU that has been powered off.  This will
2015  * destroy all per-cpu information related to this cpu.
2016  */
2017 int
2018 cpu_unconfigure(int cpuid)
2019 {
2020         int error;
2021 
2022         ASSERT(MUTEX_HELD(&cpu_lock));
2023 
2024         if (cpu[cpuid] == NULL) {
2025                 return (ENODEV);
2026         }
2027 
2028         if (cpu[cpuid]->cpu_flags == 0) {
2029                 return (EALREADY);
2030         }
2031 
2032         if ((cpu[cpuid]->cpu_flags & CPU_POWEROFF) == 0) {
2033                 return (EBUSY);
2034         }
2035 
2036         if (cpu[cpuid]->cpu_props != NULL) {
2037                 (void) nvlist_free(cpu[cpuid]->cpu_props);
2038                 cpu[cpuid]->cpu_props = NULL;
2039         }
2040 
2041         error = cpu_state_change_hooks(cpuid, CPU_UNCONFIG, CPU_CONFIG);
2042 
2043         if (error != 0)
2044                 return (error);
2045 
2046         return (mp_cpu_unconfigure(cpuid));
2047 }
2048 
2049 /*
2050  * Routines for registering and de-registering cpu_setup callback functions.
2051  *
2052  * Caller's context
2053  *      These routines must not be called from a driver's attach(9E) or
2054  *      detach(9E) entry point.
2055  *
2056  * NOTE: CPU callbacks should not block. They are called with cpu_lock held.
2057  */
2058 
2059 /*
2060  * Ideally, these would be dynamically allocated and put into a linked
2061  * list; however that is not feasible because the registration routine
2062  * has to be available before the kmem allocator is working (in fact,
2063  * it is called by the kmem allocator init code).  In any case, there
2064  * are quite a few extra entries for future users.
2065  */
2066 #define NCPU_SETUPS     20
2067 
2068 struct cpu_setup {
2069         cpu_setup_func_t *func;
2070         void *arg;
2071 } cpu_setups[NCPU_SETUPS];
2072 
2073 void
2074 register_cpu_setup_func(cpu_setup_func_t *func, void *arg)
2075 {
2076         int i;
2077 
2078         ASSERT(MUTEX_HELD(&cpu_lock));
2079 
2080         for (i = 0; i < NCPU_SETUPS; i++)
2081                 if (cpu_setups[i].func == NULL)
2082                         break;
2083         if (i >= NCPU_SETUPS)
2084                 cmn_err(CE_PANIC, "Ran out of cpu_setup callback entries");
2085 
2086         cpu_setups[i].func = func;
2087         cpu_setups[i].arg = arg;
2088 }
2089 
2090 void
2091 unregister_cpu_setup_func(cpu_setup_func_t *func, void *arg)
2092 {
2093         int i;
2094 
2095         ASSERT(MUTEX_HELD(&cpu_lock));
2096 
2097         for (i = 0; i < NCPU_SETUPS; i++)
2098                 if ((cpu_setups[i].func == func) &&
2099                     (cpu_setups[i].arg == arg))
2100                         break;
2101         if (i >= NCPU_SETUPS)
2102                 cmn_err(CE_PANIC, "Could not find cpu_setup callback to "
2103                     "deregister");
2104 
2105         cpu_setups[i].func = NULL;
2106         cpu_setups[i].arg = 0;
2107 }
2108 
2109 /*
2110  * Call any state change hooks for this CPU, ignore any errors.
2111  */
2112 void
2113 cpu_state_change_notify(int id, cpu_setup_t what)
2114 {
2115         int i;
2116 
2117         ASSERT(MUTEX_HELD(&cpu_lock));
2118 
2119         for (i = 0; i < NCPU_SETUPS; i++) {
2120                 if (cpu_setups[i].func != NULL) {
2121                         cpu_setups[i].func(what, id, cpu_setups[i].arg);
2122                 }
2123         }
2124 }
2125 
2126 /*
2127  * Call any state change hooks for this CPU, undo it if error found.
2128  */
2129 static int
2130 cpu_state_change_hooks(int id, cpu_setup_t what, cpu_setup_t undo)
2131 {
2132         int i;
2133         int retval = 0;
2134 
2135         ASSERT(MUTEX_HELD(&cpu_lock));
2136 
2137         for (i = 0; i < NCPU_SETUPS; i++) {
2138                 if (cpu_setups[i].func != NULL) {
2139                         retval = cpu_setups[i].func(what, id,
2140                             cpu_setups[i].arg);
2141                         if (retval) {
2142                                 for (i--; i >= 0; i--) {
2143                                         if (cpu_setups[i].func != NULL)
2144                                                 cpu_setups[i].func(undo,
2145                                                     id, cpu_setups[i].arg);
2146                                 }
2147                                 break;
2148                         }
2149                 }
2150         }
2151         return (retval);
2152 }
2153 
2154 /*
2155  * Export information about this CPU via the kstat mechanism.
2156  */
2157 static struct {
2158         kstat_named_t ci_state;
2159         kstat_named_t ci_state_begin;
2160         kstat_named_t ci_cpu_type;
2161         kstat_named_t ci_fpu_type;
2162         kstat_named_t ci_clock_MHz;
2163         kstat_named_t ci_chip_id;
2164         kstat_named_t ci_implementation;
2165         kstat_named_t ci_brandstr;
2166         kstat_named_t ci_core_id;
2167         kstat_named_t ci_curr_clock_Hz;
2168         kstat_named_t ci_supp_freq_Hz;
2169         kstat_named_t ci_pg_id;
2170 #if defined(__sparcv9)
2171         kstat_named_t ci_device_ID;
2172         kstat_named_t ci_cpu_fru;
2173 #endif
2174 #if defined(__x86)
2175         kstat_named_t ci_vendorstr;
2176         kstat_named_t ci_family;
2177         kstat_named_t ci_model;
2178         kstat_named_t ci_step;
2179         kstat_named_t ci_clogid;
2180         kstat_named_t ci_pkg_core_id;
2181         kstat_named_t ci_ncpuperchip;
2182         kstat_named_t ci_ncoreperchip;
2183         kstat_named_t ci_max_cstates;
2184         kstat_named_t ci_curr_cstate;
2185         kstat_named_t ci_cacheid;
2186         kstat_named_t ci_sktstr;
2187 #endif
2188 } cpu_info_template = {
2189         { "state",                      KSTAT_DATA_CHAR },
2190         { "state_begin",                KSTAT_DATA_LONG },
2191         { "cpu_type",                   KSTAT_DATA_CHAR },
2192         { "fpu_type",                   KSTAT_DATA_CHAR },
2193         { "clock_MHz",                  KSTAT_DATA_LONG },
2194         { "chip_id",                    KSTAT_DATA_LONG },
2195         { "implementation",             KSTAT_DATA_STRING },
2196         { "brand",                      KSTAT_DATA_STRING },
2197         { "core_id",                    KSTAT_DATA_LONG },
2198         { "current_clock_Hz",           KSTAT_DATA_UINT64 },
2199         { "supported_frequencies_Hz",   KSTAT_DATA_STRING },
2200         { "pg_id",                      KSTAT_DATA_LONG },
2201 #if defined(__sparcv9)
2202         { "device_ID",                  KSTAT_DATA_UINT64 },
2203         { "cpu_fru",                    KSTAT_DATA_STRING },
2204 #endif
2205 #if defined(__x86)
2206         { "vendor_id",                  KSTAT_DATA_STRING },
2207         { "family",                     KSTAT_DATA_INT32 },
2208         { "model",                      KSTAT_DATA_INT32 },
2209         { "stepping",                   KSTAT_DATA_INT32 },
2210         { "clog_id",                    KSTAT_DATA_INT32 },
2211         { "pkg_core_id",                KSTAT_DATA_LONG },
2212         { "ncpu_per_chip",              KSTAT_DATA_INT32 },
2213         { "ncore_per_chip",             KSTAT_DATA_INT32 },
2214         { "supported_max_cstates",      KSTAT_DATA_INT32 },
2215         { "current_cstate",             KSTAT_DATA_INT32 },
2216         { "cache_id",                   KSTAT_DATA_INT32 },
2217         { "socket_type",                KSTAT_DATA_STRING },
2218 #endif
2219 };
2220 
2221 static kmutex_t cpu_info_template_lock;
2222 
2223 static int
2224 cpu_info_kstat_update(kstat_t *ksp, int rw)
2225 {
2226         cpu_t   *cp = ksp->ks_private;
2227         const char *pi_state;
2228 
2229         if (rw == KSTAT_WRITE)
2230                 return (EACCES);
2231 
2232 #if defined(__x86)
2233         /* Is the cpu still initialising itself? */
2234         if (cpuid_checkpass(cp, 1) == 0)
2235                 return (ENXIO);
2236 #endif
2237         switch (cp->cpu_type_info.pi_state) {
2238         case P_ONLINE:
2239                 pi_state = PS_ONLINE;
2240                 break;
2241         case P_POWEROFF:
2242                 pi_state = PS_POWEROFF;
2243                 break;
2244         case P_NOINTR:
2245                 pi_state = PS_NOINTR;
2246                 break;
2247         case P_FAULTED:
2248                 pi_state = PS_FAULTED;
2249                 break;
2250         case P_SPARE:
2251                 pi_state = PS_SPARE;
2252                 break;
2253         case P_OFFLINE:
2254                 pi_state = PS_OFFLINE;
2255                 break;
2256         default:
2257                 pi_state = "unknown";
2258         }
2259         (void) strcpy(cpu_info_template.ci_state.value.c, pi_state);
2260         cpu_info_template.ci_state_begin.value.l = cp->cpu_state_begin;
2261         (void) strncpy(cpu_info_template.ci_cpu_type.value.c,
2262             cp->cpu_type_info.pi_processor_type, 15);
2263         (void) strncpy(cpu_info_template.ci_fpu_type.value.c,
2264             cp->cpu_type_info.pi_fputypes, 15);
2265         cpu_info_template.ci_clock_MHz.value.l = cp->cpu_type_info.pi_clock;
2266         cpu_info_template.ci_chip_id.value.l =
2267             pg_plat_hw_instance_id(cp, PGHW_CHIP);
2268         kstat_named_setstr(&cpu_info_template.ci_implementation,
2269             cp->cpu_idstr);
2270         kstat_named_setstr(&cpu_info_template.ci_brandstr, cp->cpu_brandstr);
2271         cpu_info_template.ci_core_id.value.l = pg_plat_get_core_id(cp);
2272         cpu_info_template.ci_curr_clock_Hz.value.ui64 =
2273             cp->cpu_curr_clock;
2274         cpu_info_template.ci_pg_id.value.l =
2275             cp->cpu_pg && cp->cpu_pg->cmt_lineage ?
2276             cp->cpu_pg->cmt_lineage->pg_id : -1;
2277         kstat_named_setstr(&cpu_info_template.ci_supp_freq_Hz,
2278             cp->cpu_supp_freqs);
2279 #if defined(__sparcv9)
2280         cpu_info_template.ci_device_ID.value.ui64 =
2281             cpunodes[cp->cpu_id].device_id;
2282         kstat_named_setstr(&cpu_info_template.ci_cpu_fru, cpu_fru_fmri(cp));
2283 #endif
2284 #if defined(__x86)
2285         kstat_named_setstr(&cpu_info_template.ci_vendorstr,
2286             cpuid_getvendorstr(cp));
2287         cpu_info_template.ci_family.value.l = cpuid_getfamily(cp);
2288         cpu_info_template.ci_model.value.l = cpuid_getmodel(cp);
2289         cpu_info_template.ci_step.value.l = cpuid_getstep(cp);
2290         cpu_info_template.ci_clogid.value.l = cpuid_get_clogid(cp);
2291         cpu_info_template.ci_ncpuperchip.value.l = cpuid_get_ncpu_per_chip(cp);
2292         cpu_info_template.ci_ncoreperchip.value.l =
2293             cpuid_get_ncore_per_chip(cp);
2294         cpu_info_template.ci_pkg_core_id.value.l = cpuid_get_pkgcoreid(cp);
2295         cpu_info_template.ci_max_cstates.value.l = cp->cpu_m.max_cstates;
2296         cpu_info_template.ci_curr_cstate.value.l = cpu_idle_get_cpu_state(cp);
2297         cpu_info_template.ci_cacheid.value.i32 = cpuid_get_cacheid(cp);
2298         kstat_named_setstr(&cpu_info_template.ci_sktstr,
2299             cpuid_getsocketstr(cp));
2300 #endif
2301 
2302         return (0);
2303 }
2304 
2305 static void
2306 cpu_info_kstat_create(cpu_t *cp)
2307 {
2308         zoneid_t zoneid;
2309 
2310         ASSERT(MUTEX_HELD(&cpu_lock));
2311 
2312         if (pool_pset_enabled())
2313                 zoneid = GLOBAL_ZONEID;
2314         else
2315                 zoneid = ALL_ZONES;
2316         if ((cp->cpu_info_kstat = kstat_create_zone("cpu_info", cp->cpu_id,
2317             NULL, "misc", KSTAT_TYPE_NAMED,
2318             sizeof (cpu_info_template) / sizeof (kstat_named_t),
2319             KSTAT_FLAG_VIRTUAL | KSTAT_FLAG_VAR_SIZE, zoneid)) != NULL) {
2320                 cp->cpu_info_kstat->ks_data_size += 2 * CPU_IDSTRLEN;
2321 #if defined(__sparcv9)
2322                 cp->cpu_info_kstat->ks_data_size +=
2323                     strlen(cpu_fru_fmri(cp)) + 1;
2324 #endif
2325 #if defined(__x86)
2326                 cp->cpu_info_kstat->ks_data_size += X86_VENDOR_STRLEN;
2327 #endif
2328                 if (cp->cpu_supp_freqs != NULL)
2329                         cp->cpu_info_kstat->ks_data_size +=
2330                             strlen(cp->cpu_supp_freqs) + 1;
2331                 cp->cpu_info_kstat->ks_lock = &cpu_info_template_lock;
2332                 cp->cpu_info_kstat->ks_data = &cpu_info_template;
2333                 cp->cpu_info_kstat->ks_private = cp;
2334                 cp->cpu_info_kstat->ks_update = cpu_info_kstat_update;
2335                 kstat_install(cp->cpu_info_kstat);
2336         }
2337 }
2338 
2339 static void
2340 cpu_info_kstat_destroy(cpu_t *cp)
2341 {
2342         ASSERT(MUTEX_HELD(&cpu_lock));
2343 
2344         kstat_delete(cp->cpu_info_kstat);
2345         cp->cpu_info_kstat = NULL;
2346 }
2347 
2348 /*
2349  * Create and install kstats for the boot CPU.
2350  */
2351 void
2352 cpu_kstat_init(cpu_t *cp)
2353 {
2354         mutex_enter(&cpu_lock);
2355         cpu_info_kstat_create(cp);
2356         cpu_stats_kstat_create(cp);
2357         cpu_create_intrstat(cp);
2358         cpu_set_state(cp);
2359         mutex_exit(&cpu_lock);
2360 }
2361 
2362 /*
2363  * Make visible to the zone that subset of the cpu information that would be
2364  * initialized when a cpu is configured (but still offline).
2365  */
2366 void
2367 cpu_visibility_configure(cpu_t *cp, zone_t *zone)
2368 {
2369         zoneid_t zoneid = zone ? zone->zone_id : ALL_ZONES;
2370 
2371         ASSERT(MUTEX_HELD(&cpu_lock));
2372         ASSERT(pool_pset_enabled());
2373         ASSERT(cp != NULL);
2374 
2375         if (zoneid != ALL_ZONES && zoneid != GLOBAL_ZONEID) {
2376                 zone->zone_ncpus++;
2377                 ASSERT(zone->zone_ncpus <= ncpus);
2378         }
2379         if (cp->cpu_info_kstat != NULL)
2380                 kstat_zone_add(cp->cpu_info_kstat, zoneid);
2381 }
2382 
2383 /*
2384  * Make visible to the zone that subset of the cpu information that would be
2385  * initialized when a previously configured cpu is onlined.
2386  */
2387 void
2388 cpu_visibility_online(cpu_t *cp, zone_t *zone)
2389 {
2390         kstat_t *ksp;
2391         char name[sizeof ("cpu_stat") + 10];    /* enough for 32-bit cpuids */
2392         zoneid_t zoneid = zone ? zone->zone_id : ALL_ZONES;
2393         processorid_t cpun;
2394 
2395         ASSERT(MUTEX_HELD(&cpu_lock));
2396         ASSERT(pool_pset_enabled());
2397         ASSERT(cp != NULL);
2398         ASSERT(cpu_is_active(cp));
2399 
2400         cpun = cp->cpu_id;
2401         if (zoneid != ALL_ZONES && zoneid != GLOBAL_ZONEID) {
2402                 zone->zone_ncpus_online++;
2403                 ASSERT(zone->zone_ncpus_online <= ncpus_online);
2404         }
2405         (void) snprintf(name, sizeof (name), "cpu_stat%d", cpun);
2406         if ((ksp = kstat_hold_byname("cpu_stat", cpun, name, ALL_ZONES))
2407             != NULL) {
2408                 kstat_zone_add(ksp, zoneid);
2409                 kstat_rele(ksp);
2410         }
2411         if ((ksp = kstat_hold_byname("cpu", cpun, "sys", ALL_ZONES)) != NULL) {
2412                 kstat_zone_add(ksp, zoneid);
2413                 kstat_rele(ksp);
2414         }
2415         if ((ksp = kstat_hold_byname("cpu", cpun, "vm", ALL_ZONES)) != NULL) {
2416                 kstat_zone_add(ksp, zoneid);
2417                 kstat_rele(ksp);
2418         }
2419         if ((ksp = kstat_hold_byname("cpu", cpun, "intrstat", ALL_ZONES)) !=
2420             NULL) {
2421                 kstat_zone_add(ksp, zoneid);
2422                 kstat_rele(ksp);
2423         }
2424 }
2425 
2426 /*
2427  * Update relevant kstats such that cpu is now visible to processes
2428  * executing in specified zone.
2429  */
2430 void
2431 cpu_visibility_add(cpu_t *cp, zone_t *zone)
2432 {
2433         cpu_visibility_configure(cp, zone);
2434         if (cpu_is_active(cp))
2435                 cpu_visibility_online(cp, zone);
2436 }
2437 
2438 /*
2439  * Make invisible to the zone that subset of the cpu information that would be
2440  * torn down when a previously offlined cpu is unconfigured.
2441  */
2442 void
2443 cpu_visibility_unconfigure(cpu_t *cp, zone_t *zone)
2444 {
2445         zoneid_t zoneid = zone ? zone->zone_id : ALL_ZONES;
2446 
2447         ASSERT(MUTEX_HELD(&cpu_lock));
2448         ASSERT(pool_pset_enabled());
2449         ASSERT(cp != NULL);
2450 
2451         if (zoneid != ALL_ZONES && zoneid != GLOBAL_ZONEID) {
2452                 ASSERT(zone->zone_ncpus != 0);
2453                 zone->zone_ncpus--;
2454         }
2455         if (cp->cpu_info_kstat)
2456                 kstat_zone_remove(cp->cpu_info_kstat, zoneid);
2457 }
2458 
2459 /*
2460  * Make invisible to the zone that subset of the cpu information that would be
2461  * torn down when a cpu is offlined (but still configured).
2462  */
2463 void
2464 cpu_visibility_offline(cpu_t *cp, zone_t *zone)
2465 {
2466         kstat_t *ksp;
2467         char name[sizeof ("cpu_stat") + 10];    /* enough for 32-bit cpuids */
2468         zoneid_t zoneid = zone ? zone->zone_id : ALL_ZONES;
2469         processorid_t cpun;
2470 
2471         ASSERT(MUTEX_HELD(&cpu_lock));
2472         ASSERT(pool_pset_enabled());
2473         ASSERT(cp != NULL);
2474         ASSERT(cpu_is_active(cp));
2475 
2476         cpun = cp->cpu_id;
2477         if (zoneid != ALL_ZONES && zoneid != GLOBAL_ZONEID) {
2478                 ASSERT(zone->zone_ncpus_online != 0);
2479                 zone->zone_ncpus_online--;
2480         }
2481 
2482         if ((ksp = kstat_hold_byname("cpu", cpun, "intrstat", ALL_ZONES)) !=
2483             NULL) {
2484                 kstat_zone_remove(ksp, zoneid);
2485                 kstat_rele(ksp);
2486         }
2487         if ((ksp = kstat_hold_byname("cpu", cpun, "vm", ALL_ZONES)) != NULL) {
2488                 kstat_zone_remove(ksp, zoneid);
2489                 kstat_rele(ksp);
2490         }
2491         if ((ksp = kstat_hold_byname("cpu", cpun, "sys", ALL_ZONES)) != NULL) {
2492                 kstat_zone_remove(ksp, zoneid);
2493                 kstat_rele(ksp);
2494         }
2495         (void) snprintf(name, sizeof (name), "cpu_stat%d", cpun);
2496         if ((ksp = kstat_hold_byname("cpu_stat", cpun, name, ALL_ZONES))
2497             != NULL) {
2498                 kstat_zone_remove(ksp, zoneid);
2499                 kstat_rele(ksp);
2500         }
2501 }
2502 
2503 /*
2504  * Update relevant kstats such that cpu is no longer visible to processes
2505  * executing in specified zone.
2506  */
2507 void
2508 cpu_visibility_remove(cpu_t *cp, zone_t *zone)
2509 {
2510         if (cpu_is_active(cp))
2511                 cpu_visibility_offline(cp, zone);
2512         cpu_visibility_unconfigure(cp, zone);
2513 }
2514 
2515 /*
2516  * Bind a thread to a CPU as requested.
2517  */
2518 int
2519 cpu_bind_thread(kthread_id_t tp, processorid_t bind, processorid_t *obind,
2520     int *error)
2521 {
2522         processorid_t   binding;
2523         cpu_t           *cp = NULL;
2524 
2525         ASSERT(MUTEX_HELD(&cpu_lock));
2526         ASSERT(MUTEX_HELD(&ttoproc(tp)->p_lock));
2527 
2528         thread_lock(tp);
2529 
2530         /*
2531          * Record old binding, but change the obind, which was initialized
2532          * to PBIND_NONE, only if this thread has a binding.  This avoids
2533          * reporting PBIND_NONE for a process when some LWPs are bound.
2534          */
2535         binding = tp->t_bind_cpu;
2536         if (binding != PBIND_NONE)
2537                 *obind = binding;       /* record old binding */
2538 
2539         switch (bind) {
2540         case PBIND_QUERY:
2541                 /* Just return the old binding */
2542                 thread_unlock(tp);
2543                 return (0);
2544 
2545         case PBIND_QUERY_TYPE:
2546                 /* Return the binding type */
2547                 *obind = TB_CPU_IS_SOFT(tp) ? PBIND_SOFT : PBIND_HARD;
2548                 thread_unlock(tp);
2549                 return (0);
2550 
2551         case PBIND_SOFT:
2552                 /*
2553                  *  Set soft binding for this thread and return the actual
2554                  *  binding
2555                  */
2556                 TB_CPU_SOFT_SET(tp);
2557                 thread_unlock(tp);
2558                 return (0);
2559 
2560         case PBIND_HARD:
2561                 /*
2562                  *  Set hard binding for this thread and return the actual
2563                  *  binding
2564                  */
2565                 TB_CPU_HARD_SET(tp);
2566                 thread_unlock(tp);
2567                 return (0);
2568 
2569         default:
2570                 break;
2571         }
2572 
2573         /*
2574          * If this thread/LWP cannot be bound because of permission
2575          * problems, just note that and return success so that the
2576          * other threads/LWPs will be bound.  This is the way
2577          * processor_bind() is defined to work.
2578          *
2579          * Binding will get EPERM if the thread is of system class
2580          * or hasprocperm() fails.
2581          */
2582         if (tp->t_cid == 0 || !hasprocperm(tp->t_cred, CRED())) {
2583                 *error = EPERM;
2584                 thread_unlock(tp);
2585                 return (0);
2586         }
2587 
2588         binding = bind;
2589         if (binding != PBIND_NONE) {
2590                 cp = cpu_get((processorid_t)binding);
2591                 /*
2592                  * Make sure binding is valid and is in right partition.
2593                  */
2594                 if (cp == NULL || tp->t_cpupart != cp->cpu_part) {
2595                         *error = EINVAL;
2596                         thread_unlock(tp);
2597                         return (0);
2598                 }
2599         }
2600         tp->t_bind_cpu = binding;    /* set new binding */
2601 
2602         /*
2603          * If there is no system-set reason for affinity, set
2604          * the t_bound_cpu field to reflect the binding.
2605          */
2606         if (tp->t_affinitycnt == 0) {
2607                 if (binding == PBIND_NONE) {
2608                         /*
2609                          * We may need to adjust disp_max_unbound_pri
2610                          * since we're becoming unbound.
2611                          */
2612                         disp_adjust_unbound_pri(tp);
2613 
2614                         tp->t_bound_cpu = NULL;      /* set new binding */
2615 
2616                         /*
2617                          * Move thread to lgroup with strongest affinity
2618                          * after unbinding
2619                          */
2620                         if (tp->t_lgrp_affinity)
2621                                 lgrp_move_thread(tp,
2622                                     lgrp_choose(tp, tp->t_cpupart), 1);
2623 
2624                         if (tp->t_state == TS_ONPROC &&
2625                             tp->t_cpu->cpu_part != tp->t_cpupart)
2626                                 cpu_surrender(tp);
2627                 } else {
2628                         lpl_t   *lpl;
2629 
2630                         tp->t_bound_cpu = cp;
2631                         ASSERT(cp->cpu_lpl != NULL);
2632 
2633                         /*
2634                          * Set home to lgroup with most affinity containing CPU
2635                          * that thread is being bound or minimum bounding
2636                          * lgroup if no affinities set
2637                          */
2638                         if (tp->t_lgrp_affinity)
2639                                 lpl = lgrp_affinity_best(tp, tp->t_cpupart,
2640                                     LGRP_NONE, B_FALSE);
2641                         else
2642                                 lpl = cp->cpu_lpl;
2643 
2644                         if (tp->t_lpl != lpl) {
2645                                 /* can't grab cpu_lock */
2646                                 lgrp_move_thread(tp, lpl, 1);
2647                         }
2648 
2649                         /*
2650                          * Make the thread switch to the bound CPU.
2651                          * If the thread is runnable, we need to
2652                          * requeue it even if t_cpu is already set
2653                          * to the right CPU, since it may be on a
2654                          * kpreempt queue and need to move to a local
2655                          * queue.  We could check t_disp_queue to
2656                          * avoid unnecessary overhead if it's already
2657                          * on the right queue, but since this isn't
2658                          * a performance-critical operation it doesn't
2659                          * seem worth the extra code and complexity.
2660                          *
2661                          * If the thread is weakbound to the cpu then it will
2662                          * resist the new binding request until the weak
2663                          * binding drops.  The cpu_surrender or requeueing
2664                          * below could be skipped in such cases (since it
2665                          * will have no effect), but that would require
2666                          * thread_allowmigrate to acquire thread_lock so
2667                          * we'll take the very occasional hit here instead.
2668                          */
2669                         if (tp->t_state == TS_ONPROC) {
2670                                 cpu_surrender(tp);
2671                         } else if (tp->t_state == TS_RUN) {
2672                                 cpu_t *ocp = tp->t_cpu;
2673 
2674                                 (void) dispdeq(tp);
2675                                 setbackdq(tp);
2676                                 /*
2677                                  * On the bound CPU's disp queue now.
2678                                  */
2679                                 ASSERT(tp->t_disp_queue == cp->cpu_disp ||
2680                                     tp->t_weakbound_cpu == ocp);
2681                         }
2682                 }
2683         }
2684 
2685         /*
2686          * Our binding has changed; set TP_CHANGEBIND.
2687          */
2688         tp->t_proc_flag |= TP_CHANGEBIND;
2689         aston(tp);
2690 
2691         thread_unlock(tp);
2692 
2693         return (0);
2694 }
2695 
2696 #if CPUSET_WORDS > 1
2697 
2698 /*
2699  * Functions for implementing cpuset operations when a cpuset is more
2700  * than one word.  On platforms where a cpuset is a single word these
2701  * are implemented as macros in cpuvar.h.
2702  */
2703 
2704 void
2705 cpuset_all(cpuset_t *s)
2706 {
2707         int i;
2708 
2709         for (i = 0; i < CPUSET_WORDS; i++)
2710                 s->cpub[i] = ~0UL;
2711 }
2712 
2713 void
2714 cpuset_all_but(cpuset_t *s, uint_t cpu)
2715 {
2716         cpuset_all(s);
2717         CPUSET_DEL(*s, cpu);
2718 }
2719 
2720 void
2721 cpuset_only(cpuset_t *s, uint_t cpu)
2722 {
2723         CPUSET_ZERO(*s);
2724         CPUSET_ADD(*s, cpu);
2725 }
2726 
2727 int
2728 cpuset_isnull(cpuset_t *s)
2729 {
2730         int i;
2731 
2732         for (i = 0; i < CPUSET_WORDS; i++)
2733                 if (s->cpub[i] != 0)
2734                         return (0);
2735         return (1);
2736 }
2737 
2738 int
2739 cpuset_cmp(cpuset_t *s1, cpuset_t *s2)
2740 {
2741         int i;
2742 
2743         for (i = 0; i < CPUSET_WORDS; i++)
2744                 if (s1->cpub[i] != s2->cpub[i])
2745                         return (0);
2746         return (1);
2747 }
2748 
2749 uint_t
2750 cpuset_find(cpuset_t *s)
2751 {
2752 
2753         uint_t  i;
2754         uint_t  cpu = (uint_t)-1;
2755 
2756         /*
2757          * Find a cpu in the cpuset
2758          */
2759         for (i = 0; i < CPUSET_WORDS; i++) {
2760                 cpu = (uint_t)(lowbit(s->cpub[i]) - 1);
2761                 if (cpu != (uint_t)-1) {
2762                         cpu += i * BT_NBIPUL;
2763                         break;
2764                 }
2765         }
2766         return (cpu);
2767 }
2768 
2769 void
2770 cpuset_bounds(cpuset_t *s, uint_t *smallestid, uint_t *largestid)
2771 {
2772         int     i, j;
2773         uint_t  bit;
2774 
2775         /*
2776          * First, find the smallest cpu id in the set.
2777          */
2778         for (i = 0; i < CPUSET_WORDS; i++) {
2779                 if (s->cpub[i] != 0) {
2780                         bit = (uint_t)(lowbit(s->cpub[i]) - 1);
2781                         ASSERT(bit != (uint_t)-1);
2782                         *smallestid = bit + (i * BT_NBIPUL);
2783 
2784                         /*
2785                          * Now find the largest cpu id in
2786                          * the set and return immediately.
2787                          * Done in an inner loop to avoid
2788                          * having to break out of the first
2789                          * loop.
2790                          */
2791                         for (j = CPUSET_WORDS - 1; j >= i; j--) {
2792                                 if (s->cpub[j] != 0) {
2793                                         bit = (uint_t)(highbit(s->cpub[j]) - 1);
2794                                         ASSERT(bit != (uint_t)-1);
2795                                         *largestid = bit + (j * BT_NBIPUL);
2796                                         ASSERT(*largestid >= *smallestid);
2797                                         return;
2798                                 }
2799                         }
2800 
2801                         /*
2802                          * If this code is reached, a
2803                          * smallestid was found, but not a
2804                          * largestid. The cpuset must have
2805                          * been changed during the course
2806                          * of this function call.
2807                          */
2808                         ASSERT(0);
2809                 }
2810         }
2811         *smallestid = *largestid = CPUSET_NOTINSET;
2812 }
2813 
2814 #endif  /* CPUSET_WORDS */
2815 
2816 /*
2817  * Unbind threads bound to specified CPU.
2818  *
2819  * If `unbind_all_threads' is true, unbind all user threads bound to a given
2820  * CPU. Otherwise unbind all soft-bound user threads.
2821  */
2822 int
2823 cpu_unbind(processorid_t cpu, boolean_t unbind_all_threads)
2824 {
2825         processorid_t obind;
2826         kthread_t *tp;
2827         int ret = 0;
2828         proc_t *pp;
2829         int err, berr = 0;
2830 
2831         ASSERT(MUTEX_HELD(&cpu_lock));
2832 
2833         mutex_enter(&pidlock);
2834         for (pp = practive; pp != NULL; pp = pp->p_next) {
2835                 mutex_enter(&pp->p_lock);
2836                 tp = pp->p_tlist;
2837                 /*
2838                  * Skip zombies, kernel processes, and processes in
2839                  * other zones, if called from a non-global zone.
2840                  */
2841                 if (tp == NULL || (pp->p_flag & SSYS) ||
2842                     !HASZONEACCESS(curproc, pp->p_zone->zone_id)) {
2843                         mutex_exit(&pp->p_lock);
2844                         continue;
2845                 }
2846                 do {
2847                         if (tp->t_bind_cpu != cpu)
2848                                 continue;
2849                         /*
2850                          * Skip threads with hard binding when
2851                          * `unbind_all_threads' is not specified.
2852                          */
2853                         if (!unbind_all_threads && TB_CPU_IS_HARD(tp))
2854                                 continue;
2855                         err = cpu_bind_thread(tp, PBIND_NONE, &obind, &berr);
2856                         if (ret == 0)
2857                                 ret = err;
2858                 } while ((tp = tp->t_forw) != pp->p_tlist);
2859                 mutex_exit(&pp->p_lock);
2860         }
2861         mutex_exit(&pidlock);
2862         if (ret == 0)
2863                 ret = berr;
2864         return (ret);
2865 }
2866 
2867 
2868 /*
2869  * Destroy all remaining bound threads on a cpu.
2870  */
2871 void
2872 cpu_destroy_bound_threads(cpu_t *cp)
2873 {
2874         extern id_t syscid;
2875         register kthread_id_t   t, tlist, tnext;
2876 
2877         /*
2878          * Destroy all remaining bound threads on the cpu.  This
2879          * should include both the interrupt threads and the idle thread.
2880          * This requires some care, since we need to traverse the
2881          * thread list with the pidlock mutex locked, but thread_free
2882          * also locks the pidlock mutex.  So, we collect the threads
2883          * we're going to reap in a list headed by "tlist", then we
2884          * unlock the pidlock mutex and traverse the tlist list,
2885          * doing thread_free's on the thread's.  Simple, n'est pas?
2886          * Also, this depends on thread_free not mucking with the
2887          * t_next and t_prev links of the thread.
2888          */
2889 
2890         if ((t = curthread) != NULL) {
2891 
2892                 tlist = NULL;
2893                 mutex_enter(&pidlock);
2894                 do {
2895                         tnext = t->t_next;
2896                         if (t->t_bound_cpu == cp) {
2897 
2898                                 /*
2899                                  * We've found a bound thread, carefully unlink
2900                                  * it out of the thread list, and add it to
2901                                  * our "tlist".  We "know" we don't have to
2902                                  * worry about unlinking curthread (the thread
2903                                  * that is executing this code).
2904                                  */
2905                                 t->t_next->t_prev = t->t_prev;
2906                                 t->t_prev->t_next = t->t_next;
2907                                 t->t_next = tlist;
2908                                 tlist = t;
2909                                 ASSERT(t->t_cid == syscid);
2910                                 /* wake up anyone blocked in thread_join */
2911                                 cv_broadcast(&t->t_joincv);
2912                                 /*
2913                                  * t_lwp set by interrupt threads and not
2914                                  * cleared.
2915                                  */
2916                                 t->t_lwp = NULL;
2917                                 /*
2918                                  * Pause and idle threads always have
2919                                  * t_state set to TS_ONPROC.
2920                                  */
2921                                 t->t_state = TS_FREE;
2922                                 t->t_prev = NULL;    /* Just in case */
2923                         }
2924 
2925                 } while ((t = tnext) != curthread);
2926 
2927                 mutex_exit(&pidlock);
2928 
2929                 mutex_sync();
2930                 for (t = tlist; t != NULL; t = tnext) {
2931                         tnext = t->t_next;
2932                         thread_free(t);
2933                 }
2934         }
2935 }
2936 
2937 /*
2938  * Update the cpu_supp_freqs of this cpu. This information is returned
2939  * as part of cpu_info kstats. If the cpu_info_kstat exists already, then
2940  * maintain the kstat data size.
2941  */
2942 void
2943 cpu_set_supp_freqs(cpu_t *cp, const char *freqs)
2944 {
2945         char clkstr[sizeof ("18446744073709551615") + 1]; /* ui64 MAX */
2946         const char *lfreqs = clkstr;
2947         boolean_t kstat_exists = B_FALSE;
2948         kstat_t *ksp;
2949         size_t len;
2950 
2951         /*
2952          * A NULL pointer means we only support one speed.
2953          */
2954         if (freqs == NULL)
2955                 (void) snprintf(clkstr, sizeof (clkstr), "%"PRIu64,
2956                     cp->cpu_curr_clock);
2957         else
2958                 lfreqs = freqs;
2959 
2960         /*
2961          * Make sure the frequency doesn't change while a snapshot is
2962          * going on. Of course, we only need to worry about this if
2963          * the kstat exists.
2964          */
2965         if ((ksp = cp->cpu_info_kstat) != NULL) {
2966                 mutex_enter(ksp->ks_lock);
2967                 kstat_exists = B_TRUE;
2968         }
2969 
2970         /*
2971          * Free any previously allocated string and if the kstat
2972          * already exists, then update its data size.
2973          */
2974         if (cp->cpu_supp_freqs != NULL) {
2975                 len = strlen(cp->cpu_supp_freqs) + 1;
2976                 kmem_free(cp->cpu_supp_freqs, len);
2977                 if (kstat_exists)
2978                         ksp->ks_data_size -= len;
2979         }
2980 
2981         /*
2982          * Allocate the new string and set the pointer.
2983          */
2984         len = strlen(lfreqs) + 1;
2985         cp->cpu_supp_freqs = kmem_alloc(len, KM_SLEEP);
2986         (void) strcpy(cp->cpu_supp_freqs, lfreqs);
2987 
2988         /*
2989          * If the kstat already exists then update the data size and
2990          * free the lock.
2991          */
2992         if (kstat_exists) {
2993                 ksp->ks_data_size += len;
2994                 mutex_exit(ksp->ks_lock);
2995         }
2996 }
2997 
2998 /*
2999  * Indicate the current CPU's clock freqency (in Hz).
3000  * The calling context must be such that CPU references are safe.
3001  */
3002 void
3003 cpu_set_curr_clock(uint64_t new_clk)
3004 {
3005         uint64_t old_clk;
3006 
3007         old_clk = CPU->cpu_curr_clock;
3008         CPU->cpu_curr_clock = new_clk;
3009 
3010         /*
3011          * The cpu-change-speed DTrace probe exports the frequency in Hz
3012          */
3013         DTRACE_PROBE3(cpu__change__speed, processorid_t, CPU->cpu_id,
3014             uint64_t, old_clk, uint64_t, new_clk);
3015 }
3016 
3017 /*
3018  * processor_info(2) and p_online(2) status support functions
3019  *   The constants returned by the cpu_get_state() and cpu_get_state_str() are
3020  *   for use in communicating processor state information to userland.  Kernel
3021  *   subsystems should only be using the cpu_flags value directly.  Subsystems
3022  *   modifying cpu_flags should record the state change via a call to the
3023  *   cpu_set_state().
3024  */
3025 
3026 /*
3027  * Update the pi_state of this CPU.  This function provides the CPU status for
3028  * the information returned by processor_info(2).
3029  */
3030 void
3031 cpu_set_state(cpu_t *cpu)
3032 {
3033         ASSERT(MUTEX_HELD(&cpu_lock));
3034         cpu->cpu_type_info.pi_state = cpu_get_state(cpu);
3035         cpu->cpu_state_begin = gethrestime_sec();
3036         pool_cpu_mod = gethrtime();
3037 }
3038 
3039 /*
3040  * Return offline/online/other status for the indicated CPU.  Use only for
3041  * communication with user applications; cpu_flags provides the in-kernel
3042  * interface.
3043  */
3044 int
3045 cpu_get_state(cpu_t *cpu)
3046 {
3047         ASSERT(MUTEX_HELD(&cpu_lock));
3048         if (cpu->cpu_flags & CPU_POWEROFF)
3049                 return (P_POWEROFF);
3050         else if (cpu->cpu_flags & CPU_FAULTED)
3051                 return (P_FAULTED);
3052         else if (cpu->cpu_flags & CPU_SPARE)
3053                 return (P_SPARE);
3054         else if ((cpu->cpu_flags & (CPU_READY | CPU_OFFLINE)) != CPU_READY)
3055                 return (P_OFFLINE);
3056         else if (cpu->cpu_flags & CPU_ENABLE)
3057                 return (P_ONLINE);
3058         else
3059                 return (P_NOINTR);
3060 }
3061 
3062 /*
3063  * Return processor_info(2) state as a string.
3064  */
3065 const char *
3066 cpu_get_state_str(cpu_t *cpu)
3067 {
3068         const char *string;
3069 
3070         switch (cpu_get_state(cpu)) {
3071         case P_ONLINE:
3072                 string = PS_ONLINE;
3073                 break;
3074         case P_POWEROFF:
3075                 string = PS_POWEROFF;
3076                 break;
3077         case P_NOINTR:
3078                 string = PS_NOINTR;
3079                 break;
3080         case P_SPARE:
3081                 string = PS_SPARE;
3082                 break;
3083         case P_FAULTED:
3084                 string = PS_FAULTED;
3085                 break;
3086         case P_OFFLINE:
3087                 string = PS_OFFLINE;
3088                 break;
3089         default:
3090                 string = "unknown";
3091                 break;
3092         }
3093         return (string);
3094 }
3095 
3096 /*
3097  * Export this CPU's statistics (cpu_stat_t and cpu_stats_t) as raw and named
3098  * kstats, respectively.  This is done when a CPU is initialized or placed
3099  * online via p_online(2).
3100  */
3101 static void
3102 cpu_stats_kstat_create(cpu_t *cp)
3103 {
3104         int     instance = cp->cpu_id;
3105         char    *module = "cpu";
3106         char    *class = "misc";
3107         kstat_t *ksp;
3108         zoneid_t zoneid;
3109 
3110         ASSERT(MUTEX_HELD(&cpu_lock));
3111 
3112         if (pool_pset_enabled())
3113                 zoneid = GLOBAL_ZONEID;
3114         else
3115                 zoneid = ALL_ZONES;
3116         /*
3117          * Create named kstats
3118          */
3119 #define CPU_STATS_KS_CREATE(name, tsize, update_func)                    \
3120         ksp = kstat_create_zone(module, instance, (name), class,         \
3121             KSTAT_TYPE_NAMED, (tsize) / sizeof (kstat_named_t), 0,       \
3122             zoneid);                                                     \
3123         if (ksp != NULL) {                                               \
3124                 ksp->ks_private = cp;                                    \
3125                 ksp->ks_update = (update_func);                          \
3126                 kstat_install(ksp);                                      \
3127         } else                                                           \
3128                 cmn_err(CE_WARN, "cpu: unable to create %s:%d:%s kstat", \
3129                     module, instance, (name));
3130 
3131         CPU_STATS_KS_CREATE("sys", sizeof (cpu_sys_stats_ks_data_template),
3132             cpu_sys_stats_ks_update);
3133         CPU_STATS_KS_CREATE("vm", sizeof (cpu_vm_stats_ks_data_template),
3134             cpu_vm_stats_ks_update);
3135 
3136         /*
3137          * Export the familiar cpu_stat_t KSTAT_TYPE_RAW kstat.
3138          */
3139         ksp = kstat_create_zone("cpu_stat", cp->cpu_id, NULL,
3140             "misc", KSTAT_TYPE_RAW, sizeof (cpu_stat_t), 0, zoneid);
3141         if (ksp != NULL) {
3142                 ksp->ks_update = cpu_stat_ks_update;
3143                 ksp->ks_private = cp;
3144                 kstat_install(ksp);
3145         }
3146 }
3147 
3148 static void
3149 cpu_stats_kstat_destroy(cpu_t *cp)
3150 {
3151         char ks_name[KSTAT_STRLEN];
3152 
3153         (void) sprintf(ks_name, "cpu_stat%d", cp->cpu_id);
3154         kstat_delete_byname("cpu_stat", cp->cpu_id, ks_name);
3155 
3156         kstat_delete_byname("cpu", cp->cpu_id, "sys");
3157         kstat_delete_byname("cpu", cp->cpu_id, "vm");
3158 }
3159 
3160 static int
3161 cpu_sys_stats_ks_update(kstat_t *ksp, int rw)
3162 {
3163         cpu_t *cp = (cpu_t *)ksp->ks_private;
3164         struct cpu_sys_stats_ks_data *csskd;
3165         cpu_sys_stats_t *css;
3166         hrtime_t msnsecs[NCMSTATES];
3167         int     i;
3168 
3169         if (rw == KSTAT_WRITE)
3170                 return (EACCES);
3171 
3172         csskd = ksp->ks_data;
3173         css = &cp->cpu_stats.sys;
3174 
3175         /*
3176          * Read CPU mstate, but compare with the last values we
3177          * received to make sure that the returned kstats never
3178          * decrease.
3179          */
3180 
3181         get_cpu_mstate(cp, msnsecs);
3182         if (csskd->cpu_nsec_idle.value.ui64 > msnsecs[CMS_IDLE])
3183                 msnsecs[CMS_IDLE] = csskd->cpu_nsec_idle.value.ui64;
3184         if (csskd->cpu_nsec_user.value.ui64 > msnsecs[CMS_USER])
3185                 msnsecs[CMS_USER] = csskd->cpu_nsec_user.value.ui64;
3186         if (csskd->cpu_nsec_kernel.value.ui64 > msnsecs[CMS_SYSTEM])
3187                 msnsecs[CMS_SYSTEM] = csskd->cpu_nsec_kernel.value.ui64;
3188 
3189         bcopy(&cpu_sys_stats_ks_data_template, ksp->ks_data,
3190             sizeof (cpu_sys_stats_ks_data_template));
3191 
3192         csskd->cpu_ticks_wait.value.ui64 = 0;
3193         csskd->wait_ticks_io.value.ui64 = 0;
3194 
3195         csskd->cpu_nsec_idle.value.ui64 = msnsecs[CMS_IDLE];
3196         csskd->cpu_nsec_user.value.ui64 = msnsecs[CMS_USER];
3197         csskd->cpu_nsec_kernel.value.ui64 = msnsecs[CMS_SYSTEM];
3198         csskd->cpu_ticks_idle.value.ui64 =
3199             NSEC_TO_TICK(csskd->cpu_nsec_idle.value.ui64);
3200         csskd->cpu_ticks_user.value.ui64 =
3201             NSEC_TO_TICK(csskd->cpu_nsec_user.value.ui64);
3202         csskd->cpu_ticks_kernel.value.ui64 =
3203             NSEC_TO_TICK(csskd->cpu_nsec_kernel.value.ui64);
3204         csskd->cpu_nsec_dtrace.value.ui64 = cp->cpu_dtrace_nsec;
3205         csskd->dtrace_probes.value.ui64 = cp->cpu_dtrace_probes;
3206         csskd->cpu_nsec_intr.value.ui64 = cp->cpu_intrlast;
3207         csskd->cpu_load_intr.value.ui64 = cp->cpu_intrload;
3208         csskd->bread.value.ui64 = css->bread;
3209         csskd->bwrite.value.ui64 = css->bwrite;
3210         csskd->lread.value.ui64 = css->lread;
3211         csskd->lwrite.value.ui64 = css->lwrite;
3212         csskd->phread.value.ui64 = css->phread;
3213         csskd->phwrite.value.ui64 = css->phwrite;
3214         csskd->pswitch.value.ui64 = css->pswitch;
3215         csskd->trap.value.ui64 = css->trap;
3216         csskd->intr.value.ui64 = 0;
3217         for (i = 0; i < PIL_MAX; i++)
3218                 csskd->intr.value.ui64 += css->intr[i];
3219         csskd->syscall.value.ui64 = css->syscall;
3220         csskd->sysread.value.ui64 = css->sysread;
3221         csskd->syswrite.value.ui64 = css->syswrite;
3222         csskd->sysfork.value.ui64 = css->sysfork;
3223         csskd->sysvfork.value.ui64 = css->sysvfork;
3224         csskd->sysexec.value.ui64 = css->sysexec;
3225         csskd->readch.value.ui64 = css->readch;
3226         csskd->writech.value.ui64 = css->writech;
3227         csskd->rcvint.value.ui64 = css->rcvint;
3228         csskd->xmtint.value.ui64 = css->xmtint;
3229         csskd->mdmint.value.ui64 = css->mdmint;
3230         csskd->rawch.value.ui64 = css->rawch;
3231         csskd->canch.value.ui64 = css->canch;
3232         csskd->outch.value.ui64 = css->outch;
3233         csskd->msg.value.ui64 = css->msg;
3234         csskd->sema.value.ui64 = css->sema;
3235         csskd->namei.value.ui64 = css->namei;
3236         csskd->ufsiget.value.ui64 = css->ufsiget;
3237         csskd->ufsdirblk.value.ui64 = css->ufsdirblk;
3238         csskd->ufsipage.value.ui64 = css->ufsipage;
3239         csskd->ufsinopage.value.ui64 = css->ufsinopage;
3240         csskd->procovf.value.ui64 = css->procovf;
3241         csskd->intrthread.value.ui64 = 0;
3242         for (i = 0; i < LOCK_LEVEL - 1; i++)
3243                 csskd->intrthread.value.ui64 += css->intr[i];
3244         csskd->intrblk.value.ui64 = css->intrblk;
3245         csskd->intrunpin.value.ui64 = css->intrunpin;
3246         csskd->idlethread.value.ui64 = css->idlethread;
3247         csskd->inv_swtch.value.ui64 = css->inv_swtch;
3248         csskd->nthreads.value.ui64 = css->nthreads;
3249         csskd->cpumigrate.value.ui64 = css->cpumigrate;
3250         csskd->xcalls.value.ui64 = css->xcalls;
3251         csskd->mutex_adenters.value.ui64 = css->mutex_adenters;
3252         csskd->rw_rdfails.value.ui64 = css->rw_rdfails;
3253         csskd->rw_wrfails.value.ui64 = css->rw_wrfails;
3254         csskd->modload.value.ui64 = css->modload;
3255         csskd->modunload.value.ui64 = css->modunload;
3256         csskd->bawrite.value.ui64 = css->bawrite;
3257         csskd->iowait.value.ui64 = css->iowait;
3258 
3259         return (0);
3260 }
3261 
3262 static int
3263 cpu_vm_stats_ks_update(kstat_t *ksp, int rw)
3264 {
3265         cpu_t *cp = (cpu_t *)ksp->ks_private;
3266         struct cpu_vm_stats_ks_data *cvskd;
3267         cpu_vm_stats_t *cvs;
3268 
3269         if (rw == KSTAT_WRITE)
3270                 return (EACCES);
3271 
3272         cvs = &cp->cpu_stats.vm;
3273         cvskd = ksp->ks_data;
3274 
3275         bcopy(&cpu_vm_stats_ks_data_template, ksp->ks_data,
3276             sizeof (cpu_vm_stats_ks_data_template));
3277         cvskd->pgrec.value.ui64 = cvs->pgrec;
3278         cvskd->pgfrec.value.ui64 = cvs->pgfrec;
3279         cvskd->pgin.value.ui64 = cvs->pgin;
3280         cvskd->pgpgin.value.ui64 = cvs->pgpgin;
3281         cvskd->pgout.value.ui64 = cvs->pgout;
3282         cvskd->pgpgout.value.ui64 = cvs->pgpgout;
3283         cvskd->zfod.value.ui64 = cvs->zfod;
3284         cvskd->dfree.value.ui64 = cvs->dfree;
3285         cvskd->scan.value.ui64 = cvs->scan;
3286         cvskd->rev.value.ui64 = cvs->rev;
3287         cvskd->hat_fault.value.ui64 = cvs->hat_fault;
3288         cvskd->as_fault.value.ui64 = cvs->as_fault;
3289         cvskd->maj_fault.value.ui64 = cvs->maj_fault;
3290         cvskd->cow_fault.value.ui64 = cvs->cow_fault;
3291         cvskd->prot_fault.value.ui64 = cvs->prot_fault;
3292         cvskd->softlock.value.ui64 = cvs->softlock;
3293         cvskd->kernel_asflt.value.ui64 = cvs->kernel_asflt;
3294         cvskd->pgrrun.value.ui64 = cvs->pgrrun;
3295         cvskd->execpgin.value.ui64 = cvs->execpgin;
3296         cvskd->execpgout.value.ui64 = cvs->execpgout;
3297         cvskd->execfree.value.ui64 = cvs->execfree;
3298         cvskd->anonpgin.value.ui64 = cvs->anonpgin;
3299         cvskd->anonpgout.value.ui64 = cvs->anonpgout;
3300         cvskd->anonfree.value.ui64 = cvs->anonfree;
3301         cvskd->fspgin.value.ui64 = cvs->fspgin;
3302         cvskd->fspgout.value.ui64 = cvs->fspgout;
3303         cvskd->fsfree.value.ui64 = cvs->fsfree;
3304 
3305         return (0);
3306 }
3307 
3308 static int
3309 cpu_stat_ks_update(kstat_t *ksp, int rw)
3310 {
3311         cpu_stat_t *cso;
3312         cpu_t *cp;
3313         int i;
3314         hrtime_t msnsecs[NCMSTATES];
3315 
3316         cso = (cpu_stat_t *)ksp->ks_data;
3317         cp = (cpu_t *)ksp->ks_private;
3318 
3319         if (rw == KSTAT_WRITE)
3320                 return (EACCES);
3321 
3322         /*
3323          * Read CPU mstate, but compare with the last values we
3324          * received to make sure that the returned kstats never
3325          * decrease.
3326          */
3327 
3328         get_cpu_mstate(cp, msnsecs);
3329         msnsecs[CMS_IDLE] = NSEC_TO_TICK(msnsecs[CMS_IDLE]);
3330         msnsecs[CMS_USER] = NSEC_TO_TICK(msnsecs[CMS_USER]);
3331         msnsecs[CMS_SYSTEM] = NSEC_TO_TICK(msnsecs[CMS_SYSTEM]);
3332         if (cso->cpu_sysinfo.cpu[CPU_IDLE] < msnsecs[CMS_IDLE])
3333                 cso->cpu_sysinfo.cpu[CPU_IDLE] = msnsecs[CMS_IDLE];
3334         if (cso->cpu_sysinfo.cpu[CPU_USER] < msnsecs[CMS_USER])
3335                 cso->cpu_sysinfo.cpu[CPU_USER] = msnsecs[CMS_USER];
3336         if (cso->cpu_sysinfo.cpu[CPU_KERNEL] < msnsecs[CMS_SYSTEM])
3337                 cso->cpu_sysinfo.cpu[CPU_KERNEL] = msnsecs[CMS_SYSTEM];
3338         cso->cpu_sysinfo.cpu[CPU_WAIT]       = 0;
3339         cso->cpu_sysinfo.wait[W_IO]  = 0;
3340         cso->cpu_sysinfo.wait[W_SWAP]        = 0;
3341         cso->cpu_sysinfo.wait[W_PIO] = 0;
3342         cso->cpu_sysinfo.bread               = CPU_STATS(cp, sys.bread);
3343         cso->cpu_sysinfo.bwrite      = CPU_STATS(cp, sys.bwrite);
3344         cso->cpu_sysinfo.lread               = CPU_STATS(cp, sys.lread);
3345         cso->cpu_sysinfo.lwrite      = CPU_STATS(cp, sys.lwrite);
3346         cso->cpu_sysinfo.phread      = CPU_STATS(cp, sys.phread);
3347         cso->cpu_sysinfo.phwrite     = CPU_STATS(cp, sys.phwrite);
3348         cso->cpu_sysinfo.pswitch     = CPU_STATS(cp, sys.pswitch);
3349         cso->cpu_sysinfo.trap                = CPU_STATS(cp, sys.trap);
3350         cso->cpu_sysinfo.intr                = 0;
3351         for (i = 0; i < PIL_MAX; i++)
3352                 cso->cpu_sysinfo.intr += CPU_STATS(cp, sys.intr[i]);
3353         cso->cpu_sysinfo.syscall     = CPU_STATS(cp, sys.syscall);
3354         cso->cpu_sysinfo.sysread     = CPU_STATS(cp, sys.sysread);
3355         cso->cpu_sysinfo.syswrite    = CPU_STATS(cp, sys.syswrite);
3356         cso->cpu_sysinfo.sysfork     = CPU_STATS(cp, sys.sysfork);
3357         cso->cpu_sysinfo.sysvfork    = CPU_STATS(cp, sys.sysvfork);
3358         cso->cpu_sysinfo.sysexec     = CPU_STATS(cp, sys.sysexec);
3359         cso->cpu_sysinfo.readch              = CPU_STATS(cp, sys.readch);
3360         cso->cpu_sysinfo.writech     = CPU_STATS(cp, sys.writech);
3361         cso->cpu_sysinfo.rcvint              = CPU_STATS(cp, sys.rcvint);
3362         cso->cpu_sysinfo.xmtint              = CPU_STATS(cp, sys.xmtint);
3363         cso->cpu_sysinfo.mdmint              = CPU_STATS(cp, sys.mdmint);
3364         cso->cpu_sysinfo.rawch               = CPU_STATS(cp, sys.rawch);
3365         cso->cpu_sysinfo.canch               = CPU_STATS(cp, sys.canch);
3366         cso->cpu_sysinfo.outch               = CPU_STATS(cp, sys.outch);
3367         cso->cpu_sysinfo.msg         = CPU_STATS(cp, sys.msg);
3368         cso->cpu_sysinfo.sema                = CPU_STATS(cp, sys.sema);
3369         cso->cpu_sysinfo.namei               = CPU_STATS(cp, sys.namei);
3370         cso->cpu_sysinfo.ufsiget     = CPU_STATS(cp, sys.ufsiget);
3371         cso->cpu_sysinfo.ufsdirblk   = CPU_STATS(cp, sys.ufsdirblk);
3372         cso->cpu_sysinfo.ufsipage    = CPU_STATS(cp, sys.ufsipage);
3373         cso->cpu_sysinfo.ufsinopage  = CPU_STATS(cp, sys.ufsinopage);
3374         cso->cpu_sysinfo.inodeovf    = 0;
3375         cso->cpu_sysinfo.fileovf     = 0;
3376         cso->cpu_sysinfo.procovf     = CPU_STATS(cp, sys.procovf);
3377         cso->cpu_sysinfo.intrthread  = 0;
3378         for (i = 0; i < LOCK_LEVEL - 1; i++)
3379                 cso->cpu_sysinfo.intrthread += CPU_STATS(cp, sys.intr[i]);
3380         cso->cpu_sysinfo.intrblk     = CPU_STATS(cp, sys.intrblk);
3381         cso->cpu_sysinfo.idlethread  = CPU_STATS(cp, sys.idlethread);
3382         cso->cpu_sysinfo.inv_swtch   = CPU_STATS(cp, sys.inv_swtch);
3383         cso->cpu_sysinfo.nthreads    = CPU_STATS(cp, sys.nthreads);
3384         cso->cpu_sysinfo.cpumigrate  = CPU_STATS(cp, sys.cpumigrate);
3385         cso->cpu_sysinfo.xcalls              = CPU_STATS(cp, sys.xcalls);
3386         cso->cpu_sysinfo.mutex_adenters      = CPU_STATS(cp, sys.mutex_adenters);
3387         cso->cpu_sysinfo.rw_rdfails  = CPU_STATS(cp, sys.rw_rdfails);
3388         cso->cpu_sysinfo.rw_wrfails  = CPU_STATS(cp, sys.rw_wrfails);
3389         cso->cpu_sysinfo.modload     = CPU_STATS(cp, sys.modload);
3390         cso->cpu_sysinfo.modunload   = CPU_STATS(cp, sys.modunload);
3391         cso->cpu_sysinfo.bawrite     = CPU_STATS(cp, sys.bawrite);
3392         cso->cpu_sysinfo.rw_enters   = 0;
3393         cso->cpu_sysinfo.win_uo_cnt  = 0;
3394         cso->cpu_sysinfo.win_uu_cnt  = 0;
3395         cso->cpu_sysinfo.win_so_cnt  = 0;
3396         cso->cpu_sysinfo.win_su_cnt  = 0;
3397         cso->cpu_sysinfo.win_suo_cnt = 0;
3398 
3399         cso->cpu_syswait.iowait              = CPU_STATS(cp, sys.iowait);
3400         cso->cpu_syswait.swap                = 0;
3401         cso->cpu_syswait.physio              = 0;
3402 
3403         cso->cpu_vminfo.pgrec                = CPU_STATS(cp, vm.pgrec);
3404         cso->cpu_vminfo.pgfrec               = CPU_STATS(cp, vm.pgfrec);
3405         cso->cpu_vminfo.pgin         = CPU_STATS(cp, vm.pgin);
3406         cso->cpu_vminfo.pgpgin               = CPU_STATS(cp, vm.pgpgin);
3407         cso->cpu_vminfo.pgout                = CPU_STATS(cp, vm.pgout);
3408         cso->cpu_vminfo.pgpgout              = CPU_STATS(cp, vm.pgpgout);
3409         cso->cpu_vminfo.zfod         = CPU_STATS(cp, vm.zfod);
3410         cso->cpu_vminfo.dfree                = CPU_STATS(cp, vm.dfree);
3411         cso->cpu_vminfo.scan         = CPU_STATS(cp, vm.scan);
3412         cso->cpu_vminfo.rev          = CPU_STATS(cp, vm.rev);
3413         cso->cpu_vminfo.hat_fault    = CPU_STATS(cp, vm.hat_fault);
3414         cso->cpu_vminfo.as_fault     = CPU_STATS(cp, vm.as_fault);
3415         cso->cpu_vminfo.maj_fault    = CPU_STATS(cp, vm.maj_fault);
3416         cso->cpu_vminfo.cow_fault    = CPU_STATS(cp, vm.cow_fault);
3417         cso->cpu_vminfo.prot_fault   = CPU_STATS(cp, vm.prot_fault);
3418         cso->cpu_vminfo.softlock     = CPU_STATS(cp, vm.softlock);
3419         cso->cpu_vminfo.kernel_asflt = CPU_STATS(cp, vm.kernel_asflt);
3420         cso->cpu_vminfo.pgrrun               = CPU_STATS(cp, vm.pgrrun);
3421         cso->cpu_vminfo.execpgin     = CPU_STATS(cp, vm.execpgin);
3422         cso->cpu_vminfo.execpgout    = CPU_STATS(cp, vm.execpgout);
3423         cso->cpu_vminfo.execfree     = CPU_STATS(cp, vm.execfree);
3424         cso->cpu_vminfo.anonpgin     = CPU_STATS(cp, vm.anonpgin);
3425         cso->cpu_vminfo.anonpgout    = CPU_STATS(cp, vm.anonpgout);
3426         cso->cpu_vminfo.anonfree     = CPU_STATS(cp, vm.anonfree);
3427         cso->cpu_vminfo.fspgin               = CPU_STATS(cp, vm.fspgin);
3428         cso->cpu_vminfo.fspgout              = CPU_STATS(cp, vm.fspgout);
3429         cso->cpu_vminfo.fsfree               = CPU_STATS(cp, vm.fsfree);
3430 
3431         return (0);
3432 }