OS-7125 Need mitigation of L1TF (CVE-2018-3646)
Reviewed by: Robert Mustacchi <rm@joyent.com>
Reviewed by: Jerry Jelinek <jerry.jelinek@joyent.com>
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 * Copyright 2018 Joyent, Inc.
25 */
26
27 /*
28 * Architecture-independent CPU control functions.
29 */
30
31 #include <sys/types.h>
32 #include <sys/param.h>
33 #include <sys/var.h>
34 #include <sys/thread.h>
35 #include <sys/cpuvar.h>
36 #include <sys/cpu_event.h>
37 #include <sys/kstat.h>
38 #include <sys/uadmin.h>
39 #include <sys/systm.h>
40 #include <sys/errno.h>
41 #include <sys/cmn_err.h>
42 #include <sys/procset.h>
43 #include <sys/processor.h>
44 #include <sys/debug.h>
45 #include <sys/cpupart.h>
46 #include <sys/lgrp.h>
47 #include <sys/pset.h>
48 #include <sys/pghw.h>
49 #include <sys/kmem.h>
50 #include <sys/kmem_impl.h> /* to set per-cpu kmem_cache offset */
51 #include <sys/atomic.h>
52 #include <sys/callb.h>
53 #include <sys/vtrace.h>
54 #include <sys/cyclic.h>
55 #include <sys/bitmap.h>
56 #include <sys/nvpair.h>
57 #include <sys/pool_pset.h>
58 #include <sys/msacct.h>
59 #include <sys/time.h>
60 #include <sys/archsystm.h>
61 #include <sys/sdt.h>
62 #if defined(__x86) || defined(__amd64)
63 #include <sys/x86_archext.h>
64 #endif
65 #include <sys/callo.h>
66
67 extern int mp_cpu_start(cpu_t *);
68 extern int mp_cpu_stop(cpu_t *);
69 extern int mp_cpu_poweron(cpu_t *);
70 extern int mp_cpu_poweroff(cpu_t *);
71 extern int mp_cpu_configure(int);
72 extern int mp_cpu_unconfigure(int);
73 extern void mp_cpu_faulted_enter(cpu_t *);
74 extern void mp_cpu_faulted_exit(cpu_t *);
75
76 extern int cmp_cpu_to_chip(processorid_t cpuid);
77 #ifdef __sparcv9
78 extern char *cpu_fru_fmri(cpu_t *cp);
79 #endif
80
81 static void cpu_add_active_internal(cpu_t *cp);
82 static void cpu_remove_active(cpu_t *cp);
83 static void cpu_info_kstat_create(cpu_t *cp);
84 static void cpu_info_kstat_destroy(cpu_t *cp);
85 static void cpu_stats_kstat_create(cpu_t *cp);
86 static void cpu_stats_kstat_destroy(cpu_t *cp);
87
88 static int cpu_sys_stats_ks_update(kstat_t *ksp, int rw);
89 static int cpu_vm_stats_ks_update(kstat_t *ksp, int rw);
90 static int cpu_stat_ks_update(kstat_t *ksp, int rw);
91 static int cpu_state_change_hooks(int, cpu_setup_t, cpu_setup_t);
92
93 /*
94 * cpu_lock protects ncpus, ncpus_online, cpu_flag, cpu_list, cpu_active,
95 * max_cpu_seqid_ever, and dispatch queue reallocations. The lock ordering with
96 * respect to related locks is:
97 *
98 * cpu_lock --> thread_free_lock ---> p_lock ---> thread_lock()
99 *
100 * Warning: Certain sections of code do not use the cpu_lock when
101 * traversing the cpu_list (e.g. mutex_vector_enter(), clock()). Since
102 * all cpus are paused during modifications to this list, a solution
103 * to protect the list is too either disable kernel preemption while
104 * walking the list, *or* recheck the cpu_next pointer at each
105 * iteration in the loop. Note that in no cases can any cached
106 * copies of the cpu pointers be kept as they may become invalid.
107 */
108 kmutex_t cpu_lock;
109 cpu_t *cpu_list; /* list of all CPUs */
110 cpu_t *clock_cpu_list; /* used by clock to walk CPUs */
111 cpu_t *cpu_active; /* list of active CPUs */
112 static cpuset_t cpu_available; /* set of available CPUs */
113 cpuset_t cpu_seqid_inuse; /* which cpu_seqids are in use */
114
115 cpu_t **cpu_seq; /* ptrs to CPUs, indexed by seq_id */
116
117 /*
118 * max_ncpus keeps the max cpus the system can have. Initially
119 * it's NCPU, but since most archs scan the devtree for cpus
120 * fairly early on during boot, the real max can be known before
121 * ncpus is set (useful for early NCPU based allocations).
122 */
123 int max_ncpus = NCPU;
124 /*
125 * platforms that set max_ncpus to maxiumum number of cpus that can be
126 * dynamically added will set boot_max_ncpus to the number of cpus found
127 * at device tree scan time during boot.
128 */
129 int boot_max_ncpus = -1;
130 int boot_ncpus = -1;
131 /*
132 * Maximum possible CPU id. This can never be >= NCPU since NCPU is
133 * used to size arrays that are indexed by CPU id.
134 */
135 processorid_t max_cpuid = NCPU - 1;
136
137 /*
138 * Maximum cpu_seqid was given. This number can only grow and never shrink. It
139 * can be used to optimize NCPU loops to avoid going through CPUs which were
140 * never on-line.
141 */
142 processorid_t max_cpu_seqid_ever = 0;
143
144 int ncpus = 1;
145 int ncpus_online = 1;
146
147 /*
148 * CPU that we're trying to offline. Protected by cpu_lock.
149 */
150 cpu_t *cpu_inmotion;
151
152 /*
153 * Can be raised to suppress further weakbinding, which are instead
154 * satisfied by disabling preemption. Must be raised/lowered under cpu_lock,
155 * while individual thread weakbinding synchronization is done under thread
156 * lock.
157 */
158 int weakbindingbarrier;
159
160 /*
161 * Variables used in pause_cpus().
162 */
163 static volatile char safe_list[NCPU];
164
165 static struct _cpu_pause_info {
166 int cp_spl; /* spl saved in pause_cpus() */
167 volatile int cp_go; /* Go signal sent after all ready */
168 int cp_count; /* # of CPUs to pause */
169 ksema_t cp_sem; /* synch pause_cpus & cpu_pause */
170 kthread_id_t cp_paused;
171 void *(*cp_func)(void *);
172 } cpu_pause_info;
173
174 static kmutex_t pause_free_mutex;
175 static kcondvar_t pause_free_cv;
176
177
178 static struct cpu_sys_stats_ks_data {
179 kstat_named_t cpu_ticks_idle;
180 kstat_named_t cpu_ticks_user;
181 kstat_named_t cpu_ticks_kernel;
182 kstat_named_t cpu_ticks_wait;
183 kstat_named_t cpu_nsec_idle;
184 kstat_named_t cpu_nsec_user;
185 kstat_named_t cpu_nsec_kernel;
186 kstat_named_t cpu_nsec_dtrace;
187 kstat_named_t cpu_nsec_intr;
188 kstat_named_t cpu_load_intr;
189 kstat_named_t wait_ticks_io;
190 kstat_named_t dtrace_probes;
191 kstat_named_t bread;
192 kstat_named_t bwrite;
193 kstat_named_t lread;
194 kstat_named_t lwrite;
195 kstat_named_t phread;
196 kstat_named_t phwrite;
197 kstat_named_t pswitch;
198 kstat_named_t trap;
199 kstat_named_t intr;
200 kstat_named_t syscall;
201 kstat_named_t sysread;
202 kstat_named_t syswrite;
203 kstat_named_t sysfork;
204 kstat_named_t sysvfork;
205 kstat_named_t sysexec;
206 kstat_named_t readch;
207 kstat_named_t writech;
208 kstat_named_t rcvint;
209 kstat_named_t xmtint;
210 kstat_named_t mdmint;
211 kstat_named_t rawch;
212 kstat_named_t canch;
213 kstat_named_t outch;
214 kstat_named_t msg;
215 kstat_named_t sema;
216 kstat_named_t namei;
217 kstat_named_t ufsiget;
218 kstat_named_t ufsdirblk;
219 kstat_named_t ufsipage;
220 kstat_named_t ufsinopage;
221 kstat_named_t procovf;
222 kstat_named_t intrthread;
223 kstat_named_t intrblk;
224 kstat_named_t intrunpin;
225 kstat_named_t idlethread;
226 kstat_named_t inv_swtch;
227 kstat_named_t nthreads;
228 kstat_named_t cpumigrate;
229 kstat_named_t xcalls;
230 kstat_named_t mutex_adenters;
231 kstat_named_t rw_rdfails;
232 kstat_named_t rw_wrfails;
233 kstat_named_t modload;
234 kstat_named_t modunload;
235 kstat_named_t bawrite;
236 kstat_named_t iowait;
237 } cpu_sys_stats_ks_data_template = {
238 { "cpu_ticks_idle", KSTAT_DATA_UINT64 },
239 { "cpu_ticks_user", KSTAT_DATA_UINT64 },
240 { "cpu_ticks_kernel", KSTAT_DATA_UINT64 },
241 { "cpu_ticks_wait", KSTAT_DATA_UINT64 },
242 { "cpu_nsec_idle", KSTAT_DATA_UINT64 },
243 { "cpu_nsec_user", KSTAT_DATA_UINT64 },
244 { "cpu_nsec_kernel", KSTAT_DATA_UINT64 },
245 { "cpu_nsec_dtrace", KSTAT_DATA_UINT64 },
246 { "cpu_nsec_intr", KSTAT_DATA_UINT64 },
247 { "cpu_load_intr", KSTAT_DATA_UINT64 },
248 { "wait_ticks_io", KSTAT_DATA_UINT64 },
249 { "dtrace_probes", KSTAT_DATA_UINT64 },
250 { "bread", KSTAT_DATA_UINT64 },
251 { "bwrite", KSTAT_DATA_UINT64 },
252 { "lread", KSTAT_DATA_UINT64 },
253 { "lwrite", KSTAT_DATA_UINT64 },
254 { "phread", KSTAT_DATA_UINT64 },
255 { "phwrite", KSTAT_DATA_UINT64 },
256 { "pswitch", KSTAT_DATA_UINT64 },
257 { "trap", KSTAT_DATA_UINT64 },
258 { "intr", KSTAT_DATA_UINT64 },
259 { "syscall", KSTAT_DATA_UINT64 },
260 { "sysread", KSTAT_DATA_UINT64 },
261 { "syswrite", KSTAT_DATA_UINT64 },
262 { "sysfork", KSTAT_DATA_UINT64 },
263 { "sysvfork", KSTAT_DATA_UINT64 },
264 { "sysexec", KSTAT_DATA_UINT64 },
265 { "readch", KSTAT_DATA_UINT64 },
266 { "writech", KSTAT_DATA_UINT64 },
267 { "rcvint", KSTAT_DATA_UINT64 },
268 { "xmtint", KSTAT_DATA_UINT64 },
269 { "mdmint", KSTAT_DATA_UINT64 },
270 { "rawch", KSTAT_DATA_UINT64 },
271 { "canch", KSTAT_DATA_UINT64 },
272 { "outch", KSTAT_DATA_UINT64 },
273 { "msg", KSTAT_DATA_UINT64 },
274 { "sema", KSTAT_DATA_UINT64 },
275 { "namei", KSTAT_DATA_UINT64 },
276 { "ufsiget", KSTAT_DATA_UINT64 },
277 { "ufsdirblk", KSTAT_DATA_UINT64 },
278 { "ufsipage", KSTAT_DATA_UINT64 },
279 { "ufsinopage", KSTAT_DATA_UINT64 },
280 { "procovf", KSTAT_DATA_UINT64 },
281 { "intrthread", KSTAT_DATA_UINT64 },
282 { "intrblk", KSTAT_DATA_UINT64 },
283 { "intrunpin", KSTAT_DATA_UINT64 },
284 { "idlethread", KSTAT_DATA_UINT64 },
285 { "inv_swtch", KSTAT_DATA_UINT64 },
286 { "nthreads", KSTAT_DATA_UINT64 },
287 { "cpumigrate", KSTAT_DATA_UINT64 },
288 { "xcalls", KSTAT_DATA_UINT64 },
289 { "mutex_adenters", KSTAT_DATA_UINT64 },
290 { "rw_rdfails", KSTAT_DATA_UINT64 },
291 { "rw_wrfails", KSTAT_DATA_UINT64 },
292 { "modload", KSTAT_DATA_UINT64 },
293 { "modunload", KSTAT_DATA_UINT64 },
294 { "bawrite", KSTAT_DATA_UINT64 },
295 { "iowait", KSTAT_DATA_UINT64 },
296 };
297
298 static struct cpu_vm_stats_ks_data {
299 kstat_named_t pgrec;
300 kstat_named_t pgfrec;
301 kstat_named_t pgin;
302 kstat_named_t pgpgin;
303 kstat_named_t pgout;
304 kstat_named_t pgpgout;
305 kstat_named_t swapin;
306 kstat_named_t pgswapin;
307 kstat_named_t swapout;
308 kstat_named_t pgswapout;
309 kstat_named_t zfod;
310 kstat_named_t dfree;
311 kstat_named_t scan;
312 kstat_named_t rev;
313 kstat_named_t hat_fault;
314 kstat_named_t as_fault;
315 kstat_named_t maj_fault;
316 kstat_named_t cow_fault;
317 kstat_named_t prot_fault;
318 kstat_named_t softlock;
319 kstat_named_t kernel_asflt;
320 kstat_named_t pgrrun;
321 kstat_named_t execpgin;
322 kstat_named_t execpgout;
323 kstat_named_t execfree;
324 kstat_named_t anonpgin;
325 kstat_named_t anonpgout;
326 kstat_named_t anonfree;
327 kstat_named_t fspgin;
328 kstat_named_t fspgout;
329 kstat_named_t fsfree;
330 } cpu_vm_stats_ks_data_template = {
331 { "pgrec", KSTAT_DATA_UINT64 },
332 { "pgfrec", KSTAT_DATA_UINT64 },
333 { "pgin", KSTAT_DATA_UINT64 },
334 { "pgpgin", KSTAT_DATA_UINT64 },
335 { "pgout", KSTAT_DATA_UINT64 },
336 { "pgpgout", KSTAT_DATA_UINT64 },
337 { "swapin", KSTAT_DATA_UINT64 },
338 { "pgswapin", KSTAT_DATA_UINT64 },
339 { "swapout", KSTAT_DATA_UINT64 },
340 { "pgswapout", KSTAT_DATA_UINT64 },
341 { "zfod", KSTAT_DATA_UINT64 },
342 { "dfree", KSTAT_DATA_UINT64 },
343 { "scan", KSTAT_DATA_UINT64 },
344 { "rev", KSTAT_DATA_UINT64 },
345 { "hat_fault", KSTAT_DATA_UINT64 },
346 { "as_fault", KSTAT_DATA_UINT64 },
347 { "maj_fault", KSTAT_DATA_UINT64 },
348 { "cow_fault", KSTAT_DATA_UINT64 },
349 { "prot_fault", KSTAT_DATA_UINT64 },
350 { "softlock", KSTAT_DATA_UINT64 },
351 { "kernel_asflt", KSTAT_DATA_UINT64 },
352 { "pgrrun", KSTAT_DATA_UINT64 },
353 { "execpgin", KSTAT_DATA_UINT64 },
354 { "execpgout", KSTAT_DATA_UINT64 },
355 { "execfree", KSTAT_DATA_UINT64 },
356 { "anonpgin", KSTAT_DATA_UINT64 },
357 { "anonpgout", KSTAT_DATA_UINT64 },
358 { "anonfree", KSTAT_DATA_UINT64 },
359 { "fspgin", KSTAT_DATA_UINT64 },
360 { "fspgout", KSTAT_DATA_UINT64 },
361 { "fsfree", KSTAT_DATA_UINT64 },
362 };
363
364 /*
365 * Force the specified thread to migrate to the appropriate processor.
366 * Called with thread lock held, returns with it dropped.
367 */
368 static void
369 force_thread_migrate(kthread_id_t tp)
370 {
371 ASSERT(THREAD_LOCK_HELD(tp));
372 if (tp == curthread) {
373 THREAD_TRANSITION(tp);
374 CL_SETRUN(tp);
375 thread_unlock_nopreempt(tp);
376 swtch();
377 } else {
378 if (tp->t_state == TS_ONPROC) {
379 cpu_surrender(tp);
380 } else if (tp->t_state == TS_RUN) {
381 (void) dispdeq(tp);
382 setbackdq(tp);
383 }
384 thread_unlock(tp);
385 }
386 }
387
388 /*
389 * Set affinity for a specified CPU.
390 *
391 * Specifying a cpu_id of CPU_CURRENT, allowed _only_ when setting affinity for
392 * curthread, will set affinity to the CPU on which the thread is currently
393 * running. For other cpu_id values, the caller must ensure that the
394 * referenced CPU remains valid, which can be done by holding cpu_lock across
395 * this call.
396 *
397 * CPU affinity is guaranteed after return of thread_affinity_set(). If a
398 * caller setting affinity to CPU_CURRENT requires that its thread not migrate
399 * CPUs prior to a successful return, it should take extra precautions (such as
400 * their own call to kpreempt_disable) to ensure that safety.
401 *
402 * A CPU affinity reference count is maintained by thread_affinity_set and
403 * thread_affinity_clear (incrementing and decrementing it, respectively),
404 * maintaining CPU affinity while the count is non-zero, and allowing regions
405 * of code which require affinity to be nested.
406 */
407 void
408 thread_affinity_set(kthread_id_t t, int cpu_id)
409 {
410 cpu_t *cp;
411
412 ASSERT(!(t == curthread && t->t_weakbound_cpu != NULL));
413
414 if (cpu_id == CPU_CURRENT) {
415 VERIFY3P(t, ==, curthread);
416 kpreempt_disable();
417 cp = CPU;
418 } else {
419 /*
420 * We should be asserting that cpu_lock is held here, but
421 * the NCA code doesn't acquire it. The following assert
422 * should be uncommented when the NCA code is fixed.
423 *
424 * ASSERT(MUTEX_HELD(&cpu_lock));
425 */
426 VERIFY((cpu_id >= 0) && (cpu_id < NCPU));
427 cp = cpu[cpu_id];
428
429 /* user must provide a good cpu_id */
430 VERIFY(cp != NULL);
431 }
432
433 /*
434 * If there is already a hard affinity requested, and this affinity
435 * conflicts with that, panic.
436 */
437 thread_lock(t);
438 if (t->t_affinitycnt > 0 && t->t_bound_cpu != cp) {
439 panic("affinity_set: setting %p but already bound to %p",
440 (void *)cp, (void *)t->t_bound_cpu);
441 }
442 t->t_affinitycnt++;
443 t->t_bound_cpu = cp;
444
445 /*
446 * Make sure we're running on the right CPU.
447 */
448 if (cp != t->t_cpu || t != curthread) {
449 ASSERT(cpu_id != CPU_CURRENT);
450 force_thread_migrate(t); /* drops thread lock */
451 } else {
452 thread_unlock(t);
453 }
454
455 if (cpu_id == CPU_CURRENT) {
456 kpreempt_enable();
457 }
458 }
459
460 /*
461 * Wrapper for backward compatibility.
462 */
463 void
464 affinity_set(int cpu_id)
465 {
466 thread_affinity_set(curthread, cpu_id);
467 }
468
469 /*
470 * Decrement the affinity reservation count and if it becomes zero,
471 * clear the CPU affinity for the current thread, or set it to the user's
472 * software binding request.
473 */
474 void
475 thread_affinity_clear(kthread_id_t t)
476 {
477 register processorid_t binding;
478
479 thread_lock(t);
480 if (--t->t_affinitycnt == 0) {
481 if ((binding = t->t_bind_cpu) == PBIND_NONE) {
482 /*
483 * Adjust disp_max_unbound_pri if necessary.
484 */
485 disp_adjust_unbound_pri(t);
486 t->t_bound_cpu = NULL;
487 if (t->t_cpu->cpu_part != t->t_cpupart) {
488 force_thread_migrate(t);
489 return;
490 }
491 } else {
492 t->t_bound_cpu = cpu[binding];
493 /*
494 * Make sure the thread is running on the bound CPU.
495 */
496 if (t->t_cpu != t->t_bound_cpu) {
497 force_thread_migrate(t);
498 return; /* already dropped lock */
499 }
500 }
501 }
502 thread_unlock(t);
503 }
504
505 /*
506 * Wrapper for backward compatibility.
507 */
508 void
509 affinity_clear(void)
510 {
511 thread_affinity_clear(curthread);
512 }
513
514 /*
515 * Weak cpu affinity. Bind to the "current" cpu for short periods
516 * of time during which the thread must not block (but may be preempted).
517 * Use this instead of kpreempt_disable() when it is only "no migration"
518 * rather than "no preemption" semantics that are required - disabling
519 * preemption holds higher priority threads off of cpu and if the
520 * operation that is protected is more than momentary this is not good
521 * for realtime etc.
522 *
523 * Weakly bound threads will not prevent a cpu from being offlined -
524 * we'll only run them on the cpu to which they are weakly bound but
525 * (because they do not block) we'll always be able to move them on to
526 * another cpu at offline time if we give them just a short moment to
527 * run during which they will unbind. To give a cpu a chance of offlining,
528 * however, we require a barrier to weak bindings that may be raised for a
529 * given cpu (offline/move code may set this and then wait a short time for
530 * existing weak bindings to drop); the cpu_inmotion pointer is that barrier.
531 *
532 * There are few restrictions on the calling context of thread_nomigrate.
533 * The caller must not hold the thread lock. Calls may be nested.
534 *
535 * After weakbinding a thread must not perform actions that may block.
536 * In particular it must not call thread_affinity_set; calling that when
537 * already weakbound is nonsensical anyway.
538 *
539 * If curthread is prevented from migrating for other reasons
540 * (kernel preemption disabled; high pil; strongly bound; interrupt thread)
541 * then the weak binding will succeed even if this cpu is the target of an
542 * offline/move request.
543 */
544 void
545 thread_nomigrate(void)
546 {
547 cpu_t *cp;
548 kthread_id_t t = curthread;
549
550 again:
551 kpreempt_disable();
552 cp = CPU;
553
554 /*
555 * A highlevel interrupt must not modify t_nomigrate or
556 * t_weakbound_cpu of the thread it has interrupted. A lowlevel
557 * interrupt thread cannot migrate and we can avoid the
558 * thread_lock call below by short-circuiting here. In either
559 * case we can just return since no migration is possible and
560 * the condition will persist (ie, when we test for these again
561 * in thread_allowmigrate they can't have changed). Migration
562 * is also impossible if we're at or above DISP_LEVEL pil.
563 */
564 if (CPU_ON_INTR(cp) || t->t_flag & T_INTR_THREAD ||
565 getpil() >= DISP_LEVEL) {
566 kpreempt_enable();
567 return;
568 }
569
570 /*
571 * We must be consistent with existing weak bindings. Since we
572 * may be interrupted between the increment of t_nomigrate and
573 * the store to t_weakbound_cpu below we cannot assume that
574 * t_weakbound_cpu will be set if t_nomigrate is. Note that we
575 * cannot assert t_weakbound_cpu == t_bind_cpu since that is not
576 * always the case.
577 */
578 if (t->t_nomigrate && t->t_weakbound_cpu && t->t_weakbound_cpu != cp) {
579 if (!panicstr)
580 panic("thread_nomigrate: binding to %p but already "
581 "bound to %p", (void *)cp,
582 (void *)t->t_weakbound_cpu);
583 }
584
585 /*
586 * At this point we have preemption disabled and we don't yet hold
587 * the thread lock. So it's possible that somebody else could
588 * set t_bind_cpu here and not be able to force us across to the
589 * new cpu (since we have preemption disabled).
590 */
591 thread_lock(curthread);
592
593 /*
594 * If further weak bindings are being (temporarily) suppressed then
595 * we'll settle for disabling kernel preemption (which assures
596 * no migration provided the thread does not block which it is
597 * not allowed to if using thread_nomigrate). We must remember
598 * this disposition so we can take appropriate action in
599 * thread_allowmigrate. If this is a nested call and the
600 * thread is already weakbound then fall through as normal.
601 * We remember the decision to settle for kpreempt_disable through
602 * negative nesting counting in t_nomigrate. Once a thread has had one
603 * weakbinding request satisfied in this way any further (nested)
604 * requests will continue to be satisfied in the same way,
605 * even if weak bindings have recommenced.
606 */
607 if (t->t_nomigrate < 0 || weakbindingbarrier && t->t_nomigrate == 0) {
608 --t->t_nomigrate;
609 thread_unlock(curthread);
610 return; /* with kpreempt_disable still active */
611 }
612
613 /*
614 * We hold thread_lock so t_bind_cpu cannot change. We could,
615 * however, be running on a different cpu to which we are t_bound_cpu
616 * to (as explained above). If we grant the weak binding request
617 * in that case then the dispatcher must favour our weak binding
618 * over our strong (in which case, just as when preemption is
619 * disabled, we can continue to run on a cpu other than the one to
620 * which we are strongbound; the difference in this case is that
621 * this thread can be preempted and so can appear on the dispatch
622 * queues of a cpu other than the one it is strongbound to).
623 *
624 * If the cpu we are running on does not appear to be a current
625 * offline target (we check cpu_inmotion to determine this - since
626 * we don't hold cpu_lock we may not see a recent store to that,
627 * so it's possible that we at times can grant a weak binding to a
628 * cpu that is an offline target, but that one request will not
629 * prevent the offline from succeeding) then we will always grant
630 * the weak binding request. This includes the case above where
631 * we grant a weakbinding not commensurate with our strong binding.
632 *
633 * If our cpu does appear to be an offline target then we're inclined
634 * not to grant the weakbinding request just yet - we'd prefer to
635 * migrate to another cpu and grant the request there. The
636 * exceptions are those cases where going through preemption code
637 * will not result in us changing cpu:
638 *
639 * . interrupts have already bypassed this case (see above)
640 * . we are already weakbound to this cpu (dispatcher code will
641 * always return us to the weakbound cpu)
642 * . preemption was disabled even before we disabled it above
643 * . we are strongbound to this cpu (if we're strongbound to
644 * another and not yet running there the trip through the
645 * dispatcher will move us to the strongbound cpu and we
646 * will grant the weak binding there)
647 */
648 if (cp != cpu_inmotion || t->t_nomigrate > 0 || t->t_preempt > 1 ||
649 t->t_bound_cpu == cp) {
650 /*
651 * Don't be tempted to store to t_weakbound_cpu only on
652 * the first nested bind request - if we're interrupted
653 * after the increment of t_nomigrate and before the
654 * store to t_weakbound_cpu and the interrupt calls
655 * thread_nomigrate then the assertion in thread_allowmigrate
656 * would fail.
657 */
658 t->t_nomigrate++;
659 t->t_weakbound_cpu = cp;
660 membar_producer();
661 thread_unlock(curthread);
662 /*
663 * Now that we have dropped the thread_lock another thread
664 * can set our t_weakbound_cpu, and will try to migrate us
665 * to the strongbound cpu (which will not be prevented by
666 * preemption being disabled since we're about to enable
667 * preemption). We have granted the weakbinding to the current
668 * cpu, so again we are in the position that is is is possible
669 * that our weak and strong bindings differ. Again this
670 * is catered for by dispatcher code which will favour our
671 * weak binding.
672 */
673 kpreempt_enable();
674 } else {
675 /*
676 * Move to another cpu before granting the request by
677 * forcing this thread through preemption code. When we
678 * get to set{front,back}dq called from CL_PREEMPT()
679 * cpu_choose() will be used to select a cpu to queue
680 * us on - that will see cpu_inmotion and take
681 * steps to avoid returning us to this cpu.
682 */
683 cp->cpu_kprunrun = 1;
684 thread_unlock(curthread);
685 kpreempt_enable(); /* will call preempt() */
686 goto again;
687 }
688 }
689
690 void
691 thread_allowmigrate(void)
692 {
693 kthread_id_t t = curthread;
694
695 ASSERT(t->t_weakbound_cpu == CPU ||
696 (t->t_nomigrate < 0 && t->t_preempt > 0) ||
697 CPU_ON_INTR(CPU) || t->t_flag & T_INTR_THREAD ||
698 getpil() >= DISP_LEVEL);
699
700 if (CPU_ON_INTR(CPU) || (t->t_flag & T_INTR_THREAD) ||
701 getpil() >= DISP_LEVEL)
702 return;
703
704 if (t->t_nomigrate < 0) {
705 /*
706 * This thread was granted "weak binding" in the
707 * stronger form of kernel preemption disabling.
708 * Undo a level of nesting for both t_nomigrate
709 * and t_preempt.
710 */
711 ++t->t_nomigrate;
712 kpreempt_enable();
713 } else if (--t->t_nomigrate == 0) {
714 /*
715 * Time to drop the weak binding. We need to cater
716 * for the case where we're weakbound to a different
717 * cpu than that to which we're strongbound (a very
718 * temporary arrangement that must only persist until
719 * weak binding drops). We don't acquire thread_lock
720 * here so even as this code executes t_bound_cpu
721 * may be changing. So we disable preemption and
722 * a) in the case that t_bound_cpu changes while we
723 * have preemption disabled kprunrun will be set
724 * asynchronously, and b) if before disabling
725 * preemption we were already on a different cpu to
726 * our t_bound_cpu then we set kprunrun ourselves
727 * to force a trip through the dispatcher when
728 * preemption is enabled.
729 */
730 kpreempt_disable();
731 if (t->t_bound_cpu &&
732 t->t_weakbound_cpu != t->t_bound_cpu)
733 CPU->cpu_kprunrun = 1;
734 t->t_weakbound_cpu = NULL;
735 membar_producer();
736 kpreempt_enable();
737 }
738 }
739
740 /*
741 * weakbinding_stop can be used to temporarily cause weakbindings made
742 * with thread_nomigrate to be satisfied through the stronger action of
743 * kpreempt_disable. weakbinding_start recommences normal weakbinding.
744 */
745
746 void
747 weakbinding_stop(void)
748 {
749 ASSERT(MUTEX_HELD(&cpu_lock));
750 weakbindingbarrier = 1;
751 membar_producer(); /* make visible before subsequent thread_lock */
752 }
753
754 void
755 weakbinding_start(void)
756 {
757 ASSERT(MUTEX_HELD(&cpu_lock));
758 weakbindingbarrier = 0;
759 }
760
761 void
762 null_xcall(void)
763 {
764 }
765
766 /*
767 * This routine is called to place the CPUs in a safe place so that
768 * one of them can be taken off line or placed on line. What we are
769 * trying to do here is prevent a thread from traversing the list
770 * of active CPUs while we are changing it or from getting placed on
771 * the run queue of a CPU that has just gone off line. We do this by
772 * creating a thread with the highest possible prio for each CPU and
773 * having it call this routine. The advantage of this method is that
774 * we can eliminate all checks for CPU_ACTIVE in the disp routines.
775 * This makes disp faster at the expense of making p_online() slower
776 * which is a good trade off.
777 */
778 static void
779 cpu_pause(int index)
780 {
781 int s;
782 struct _cpu_pause_info *cpi = &cpu_pause_info;
783 volatile char *safe = &safe_list[index];
784 long lindex = index;
785
786 ASSERT((curthread->t_bound_cpu != NULL) || (*safe == PAUSE_DIE));
787
788 while (*safe != PAUSE_DIE) {
789 *safe = PAUSE_READY;
790 membar_enter(); /* make sure stores are flushed */
791 sema_v(&cpi->cp_sem); /* signal requesting thread */
792
793 /*
794 * Wait here until all pause threads are running. That
795 * indicates that it's safe to do the spl. Until
796 * cpu_pause_info.cp_go is set, we don't want to spl
797 * because that might block clock interrupts needed
798 * to preempt threads on other CPUs.
799 */
800 while (cpi->cp_go == 0)
801 ;
802 /*
803 * Even though we are at the highest disp prio, we need
804 * to block out all interrupts below LOCK_LEVEL so that
805 * an intr doesn't come in, wake up a thread, and call
806 * setbackdq/setfrontdq.
807 */
808 s = splhigh();
809 /*
810 * if cp_func has been set then call it using index as the
811 * argument, currently only used by cpr_suspend_cpus().
812 * This function is used as the code to execute on the
813 * "paused" cpu's when a machine comes out of a sleep state
814 * and CPU's were powered off. (could also be used for
815 * hotplugging CPU's).
816 */
817 if (cpi->cp_func != NULL)
818 (*cpi->cp_func)((void *)lindex);
819
820 mach_cpu_pause(safe);
821
822 splx(s);
823 /*
824 * Waiting is at an end. Switch out of cpu_pause
825 * loop and resume useful work.
826 */
827 swtch();
828 }
829
830 mutex_enter(&pause_free_mutex);
831 *safe = PAUSE_DEAD;
832 cv_broadcast(&pause_free_cv);
833 mutex_exit(&pause_free_mutex);
834 }
835
836 /*
837 * Allow the cpus to start running again.
838 */
839 void
840 start_cpus()
841 {
842 int i;
843
844 ASSERT(MUTEX_HELD(&cpu_lock));
845 ASSERT(cpu_pause_info.cp_paused);
846 cpu_pause_info.cp_paused = NULL;
847 for (i = 0; i < NCPU; i++)
848 safe_list[i] = PAUSE_IDLE;
849 membar_enter(); /* make sure stores are flushed */
850 affinity_clear();
851 splx(cpu_pause_info.cp_spl);
852 kpreempt_enable();
853 }
854
855 /*
856 * Allocate a pause thread for a CPU.
857 */
858 static void
859 cpu_pause_alloc(cpu_t *cp)
860 {
861 kthread_id_t t;
862 long cpun = cp->cpu_id;
863
864 /*
865 * Note, v.v_nglobpris will not change value as long as I hold
866 * cpu_lock.
867 */
868 t = thread_create(NULL, 0, cpu_pause, (void *)cpun,
869 0, &p0, TS_STOPPED, v.v_nglobpris - 1);
870 thread_lock(t);
871 t->t_bound_cpu = cp;
872 t->t_disp_queue = cp->cpu_disp;
873 t->t_affinitycnt = 1;
874 t->t_preempt = 1;
875 thread_unlock(t);
876 cp->cpu_pause_thread = t;
877 /*
878 * Registering a thread in the callback table is usually done
879 * in the initialization code of the thread. In this
880 * case, we do it right after thread creation because the
881 * thread itself may never run, and we need to register the
882 * fact that it is safe for cpr suspend.
883 */
884 CALLB_CPR_INIT_SAFE(t, "cpu_pause");
885 }
886
887 /*
888 * Free a pause thread for a CPU.
889 */
890 static void
891 cpu_pause_free(cpu_t *cp)
892 {
893 kthread_id_t t;
894 int cpun = cp->cpu_id;
895
896 ASSERT(MUTEX_HELD(&cpu_lock));
897 /*
898 * We have to get the thread and tell it to die.
899 */
900 if ((t = cp->cpu_pause_thread) == NULL) {
901 ASSERT(safe_list[cpun] == PAUSE_IDLE);
902 return;
903 }
904 thread_lock(t);
905 t->t_cpu = CPU; /* disp gets upset if last cpu is quiesced. */
906 t->t_bound_cpu = NULL; /* Must un-bind; cpu may not be running. */
907 t->t_pri = v.v_nglobpris - 1;
908 ASSERT(safe_list[cpun] == PAUSE_IDLE);
909 safe_list[cpun] = PAUSE_DIE;
910 THREAD_TRANSITION(t);
911 setbackdq(t);
912 thread_unlock_nopreempt(t);
913
914 /*
915 * If we don't wait for the thread to actually die, it may try to
916 * run on the wrong cpu as part of an actual call to pause_cpus().
917 */
918 mutex_enter(&pause_free_mutex);
919 while (safe_list[cpun] != PAUSE_DEAD) {
920 cv_wait(&pause_free_cv, &pause_free_mutex);
921 }
922 mutex_exit(&pause_free_mutex);
923 safe_list[cpun] = PAUSE_IDLE;
924
925 cp->cpu_pause_thread = NULL;
926 }
927
928 /*
929 * Initialize basic structures for pausing CPUs.
930 */
931 void
932 cpu_pause_init()
933 {
934 sema_init(&cpu_pause_info.cp_sem, 0, NULL, SEMA_DEFAULT, NULL);
935 /*
936 * Create initial CPU pause thread.
937 */
938 cpu_pause_alloc(CPU);
939 }
940
941 /*
942 * Start the threads used to pause another CPU.
943 */
944 static int
945 cpu_pause_start(processorid_t cpu_id)
946 {
947 int i;
948 int cpu_count = 0;
949
950 for (i = 0; i < NCPU; i++) {
951 cpu_t *cp;
952 kthread_id_t t;
953
954 cp = cpu[i];
955 if (!CPU_IN_SET(cpu_available, i) || (i == cpu_id)) {
956 safe_list[i] = PAUSE_WAIT;
957 continue;
958 }
959
960 /*
961 * Skip CPU if it is quiesced or not yet started.
962 */
963 if ((cp->cpu_flags & (CPU_QUIESCED | CPU_READY)) != CPU_READY) {
964 safe_list[i] = PAUSE_WAIT;
965 continue;
966 }
967
968 /*
969 * Start this CPU's pause thread.
970 */
971 t = cp->cpu_pause_thread;
972 thread_lock(t);
973 /*
974 * Reset the priority, since nglobpris may have
975 * changed since the thread was created, if someone
976 * has loaded the RT (or some other) scheduling
977 * class.
978 */
979 t->t_pri = v.v_nglobpris - 1;
980 THREAD_TRANSITION(t);
981 setbackdq(t);
982 thread_unlock_nopreempt(t);
983 ++cpu_count;
984 }
985 return (cpu_count);
986 }
987
988
989 /*
990 * Pause all of the CPUs except the one we are on by creating a high
991 * priority thread bound to those CPUs.
992 *
993 * Note that one must be extremely careful regarding code
994 * executed while CPUs are paused. Since a CPU may be paused
995 * while a thread scheduling on that CPU is holding an adaptive
996 * lock, code executed with CPUs paused must not acquire adaptive
997 * (or low-level spin) locks. Also, such code must not block,
998 * since the thread that is supposed to initiate the wakeup may
999 * never run.
1000 *
1001 * With a few exceptions, the restrictions on code executed with CPUs
1002 * paused match those for code executed at high-level interrupt
1003 * context.
1004 */
1005 void
1006 pause_cpus(cpu_t *off_cp, void *(*func)(void *))
1007 {
1008 processorid_t cpu_id;
1009 int i;
1010 struct _cpu_pause_info *cpi = &cpu_pause_info;
1011
1012 ASSERT(MUTEX_HELD(&cpu_lock));
1013 ASSERT(cpi->cp_paused == NULL);
1014 cpi->cp_count = 0;
1015 cpi->cp_go = 0;
1016 for (i = 0; i < NCPU; i++)
1017 safe_list[i] = PAUSE_IDLE;
1018 kpreempt_disable();
1019
1020 cpi->cp_func = func;
1021
1022 /*
1023 * If running on the cpu that is going offline, get off it.
1024 * This is so that it won't be necessary to rechoose a CPU
1025 * when done.
1026 */
1027 if (CPU == off_cp)
1028 cpu_id = off_cp->cpu_next_part->cpu_id;
1029 else
1030 cpu_id = CPU->cpu_id;
1031 affinity_set(cpu_id);
1032
1033 /*
1034 * Start the pause threads and record how many were started
1035 */
1036 cpi->cp_count = cpu_pause_start(cpu_id);
1037
1038 /*
1039 * Now wait for all CPUs to be running the pause thread.
1040 */
1041 while (cpi->cp_count > 0) {
1042 /*
1043 * Spin reading the count without grabbing the disp
1044 * lock to make sure we don't prevent the pause
1045 * threads from getting the lock.
1046 */
1047 while (sema_held(&cpi->cp_sem))
1048 ;
1049 if (sema_tryp(&cpi->cp_sem))
1050 --cpi->cp_count;
1051 }
1052 cpi->cp_go = 1; /* all have reached cpu_pause */
1053
1054 /*
1055 * Now wait for all CPUs to spl. (Transition from PAUSE_READY
1056 * to PAUSE_WAIT.)
1057 */
1058 for (i = 0; i < NCPU; i++) {
1059 while (safe_list[i] != PAUSE_WAIT)
1060 ;
1061 }
1062 cpi->cp_spl = splhigh(); /* block dispatcher on this CPU */
1063 cpi->cp_paused = curthread;
1064 }
1065
1066 /*
1067 * Check whether the current thread has CPUs paused
1068 */
1069 int
1070 cpus_paused(void)
1071 {
1072 if (cpu_pause_info.cp_paused != NULL) {
1073 ASSERT(cpu_pause_info.cp_paused == curthread);
1074 return (1);
1075 }
1076 return (0);
1077 }
1078
1079 static cpu_t *
1080 cpu_get_all(processorid_t cpun)
1081 {
1082 ASSERT(MUTEX_HELD(&cpu_lock));
1083
1084 if (cpun >= NCPU || cpun < 0 || !CPU_IN_SET(cpu_available, cpun))
1085 return (NULL);
1086 return (cpu[cpun]);
1087 }
1088
1089 /*
1090 * Check whether cpun is a valid processor id and whether it should be
1091 * visible from the current zone. If it is, return a pointer to the
1092 * associated CPU structure.
1093 */
1094 cpu_t *
1095 cpu_get(processorid_t cpun)
1096 {
1097 cpu_t *c;
1098
1099 ASSERT(MUTEX_HELD(&cpu_lock));
1100 c = cpu_get_all(cpun);
1101 if (c != NULL && !INGLOBALZONE(curproc) && pool_pset_enabled() &&
1102 zone_pset_get(curproc->p_zone) != cpupart_query_cpu(c))
1103 return (NULL);
1104 return (c);
1105 }
1106
1107 /*
1108 * The following functions should be used to check CPU states in the kernel.
1109 * They should be invoked with cpu_lock held. Kernel subsystems interested
1110 * in CPU states should *not* use cpu_get_state() and various P_ONLINE/etc
1111 * states. Those are for user-land (and system call) use only.
1112 */
1113
1114 /*
1115 * Determine whether the CPU is online and handling interrupts.
1116 */
1117 int
1118 cpu_is_online(cpu_t *cpu)
1119 {
1120 ASSERT(MUTEX_HELD(&cpu_lock));
1121 return (cpu_flagged_online(cpu->cpu_flags));
1122 }
1123
1124 /*
1125 * Determine whether the CPU is offline (this includes spare and faulted).
1126 */
1127 int
1128 cpu_is_offline(cpu_t *cpu)
1129 {
1130 ASSERT(MUTEX_HELD(&cpu_lock));
1131 return (cpu_flagged_offline(cpu->cpu_flags));
1132 }
1133
1134 /*
1135 * Determine whether the CPU is powered off.
1136 */
1137 int
1138 cpu_is_poweredoff(cpu_t *cpu)
1139 {
1140 ASSERT(MUTEX_HELD(&cpu_lock));
1141 return (cpu_flagged_poweredoff(cpu->cpu_flags));
1142 }
1143
1144 /*
1145 * Determine whether the CPU is handling interrupts.
1146 */
1147 int
1148 cpu_is_nointr(cpu_t *cpu)
1149 {
1150 ASSERT(MUTEX_HELD(&cpu_lock));
1151 return (cpu_flagged_nointr(cpu->cpu_flags));
1152 }
1153
1154 /*
1155 * Determine whether the CPU is active (scheduling threads).
1156 */
1157 int
1158 cpu_is_active(cpu_t *cpu)
1159 {
1160 ASSERT(MUTEX_HELD(&cpu_lock));
1161 return (cpu_flagged_active(cpu->cpu_flags));
1162 }
1163
1164 /*
1165 * Same as above, but these require cpu_flags instead of cpu_t pointers.
1166 */
1167 int
1168 cpu_flagged_online(cpu_flag_t cpu_flags)
1169 {
1170 return (cpu_flagged_active(cpu_flags) &&
1171 (cpu_flags & CPU_ENABLE));
1172 }
1173
1174 int
1175 cpu_flagged_offline(cpu_flag_t cpu_flags)
1176 {
1177 return (((cpu_flags & CPU_POWEROFF) == 0) &&
1178 ((cpu_flags & (CPU_READY | CPU_OFFLINE)) != CPU_READY));
1179 }
1180
1181 int
1182 cpu_flagged_poweredoff(cpu_flag_t cpu_flags)
1183 {
1184 return ((cpu_flags & CPU_POWEROFF) == CPU_POWEROFF);
1185 }
1186
1187 int
1188 cpu_flagged_nointr(cpu_flag_t cpu_flags)
1189 {
1190 return (cpu_flagged_active(cpu_flags) &&
1191 (cpu_flags & CPU_ENABLE) == 0);
1192 }
1193
1194 int
1195 cpu_flagged_active(cpu_flag_t cpu_flags)
1196 {
1197 return (((cpu_flags & (CPU_POWEROFF | CPU_FAULTED | CPU_SPARE)) == 0) &&
1198 ((cpu_flags & (CPU_READY | CPU_OFFLINE)) == CPU_READY));
1199 }
1200
1201 /*
1202 * Bring the indicated CPU online.
1203 */
1204 int
1205 cpu_online(cpu_t *cp)
1206 {
1207 int error = 0;
1208
1209 /*
1210 * Handle on-line request.
1211 * This code must put the new CPU on the active list before
1212 * starting it because it will not be paused, and will start
1213 * using the active list immediately. The real start occurs
1214 * when the CPU_QUIESCED flag is turned off.
1215 */
1216
1217 ASSERT(MUTEX_HELD(&cpu_lock));
1218
1219 /*
1220 * Put all the cpus into a known safe place.
1221 * No mutexes can be entered while CPUs are paused.
1222 */
1223 error = mp_cpu_start(cp); /* arch-dep hook */
1224 if (error == 0) {
1225 pg_cpupart_in(cp, cp->cpu_part);
1226 pause_cpus(NULL, NULL);
1227 cpu_add_active_internal(cp);
1228 if (cp->cpu_flags & CPU_FAULTED) {
1229 cp->cpu_flags &= ~CPU_FAULTED;
1230 mp_cpu_faulted_exit(cp);
1231 }
1232 cp->cpu_flags &= ~(CPU_QUIESCED | CPU_OFFLINE | CPU_FROZEN |
1233 CPU_SPARE);
1234 CPU_NEW_GENERATION(cp);
1235 start_cpus();
1236 cpu_stats_kstat_create(cp);
1237 cpu_create_intrstat(cp);
1238 lgrp_kstat_create(cp);
1239 cpu_state_change_notify(cp->cpu_id, CPU_ON);
1240 cpu_intr_enable(cp); /* arch-dep hook */
1241 cpu_state_change_notify(cp->cpu_id, CPU_INTR_ON);
1242 cpu_set_state(cp);
1243 cyclic_online(cp);
1244 /*
1245 * This has to be called only after cyclic_online(). This
1246 * function uses cyclics.
1247 */
1248 callout_cpu_online(cp);
1249 poke_cpu(cp->cpu_id);
1250 }
1251
1252 return (error);
1253 }
1254
1255 /*
1256 * Take the indicated CPU offline.
1257 */
1258 int
1259 cpu_offline(cpu_t *cp, int flags)
1260 {
1261 cpupart_t *pp;
1262 int error = 0;
1263 cpu_t *ncp;
1264 int intr_enable;
1265 int cyclic_off = 0;
1266 int callout_off = 0;
1267 int loop_count;
1268 int no_quiesce = 0;
1269 int (*bound_func)(struct cpu *, int);
1270 kthread_t *t;
1271 lpl_t *cpu_lpl;
1272 proc_t *p;
1273 int lgrp_diff_lpl;
1274 boolean_t unbind_all_threads = (flags & CPU_FORCED) != 0;
1275
1276 ASSERT(MUTEX_HELD(&cpu_lock));
1277
1278 /*
1279 * If we're going from faulted or spare to offline, just
1280 * clear these flags and update CPU state.
1281 */
1282 if (cp->cpu_flags & (CPU_FAULTED | CPU_SPARE)) {
1283 if (cp->cpu_flags & CPU_FAULTED) {
1284 cp->cpu_flags &= ~CPU_FAULTED;
1285 mp_cpu_faulted_exit(cp);
1286 }
1287 cp->cpu_flags &= ~CPU_SPARE;
1288 cpu_set_state(cp);
1289 return (0);
1290 }
1291
1292 /*
1293 * Handle off-line request.
1294 */
1295 pp = cp->cpu_part;
1296 /*
1297 * Don't offline last online CPU in partition
1298 */
1299 if (ncpus_online <= 1 || pp->cp_ncpus <= 1 || cpu_intr_count(cp) < 2)
1300 return (EBUSY);
1301 /*
1302 * Unbind all soft-bound threads bound to our CPU and hard bound threads
1303 * if we were asked to.
1304 */
1305 error = cpu_unbind(cp->cpu_id, unbind_all_threads);
1306 if (error != 0)
1307 return (error);
1308 /*
1309 * We shouldn't be bound to this CPU ourselves.
1310 */
1311 if (curthread->t_bound_cpu == cp)
1312 return (EBUSY);
1313
1314 /*
1315 * Tell interested parties that this CPU is going offline.
1316 */
1317 CPU_NEW_GENERATION(cp);
1318 cpu_state_change_notify(cp->cpu_id, CPU_OFF);
1319
1320 /*
1321 * Tell the PG subsystem that the CPU is leaving the partition
1322 */
1323 pg_cpupart_out(cp, pp);
1324
1325 /*
1326 * Take the CPU out of interrupt participation so we won't find
1327 * bound kernel threads. If the architecture cannot completely
1328 * shut off interrupts on the CPU, don't quiesce it, but don't
1329 * run anything but interrupt thread... this is indicated by
1330 * the CPU_OFFLINE flag being on but the CPU_QUIESCE flag being
1331 * off.
1332 */
1333 intr_enable = cp->cpu_flags & CPU_ENABLE;
1334 if (intr_enable)
1335 no_quiesce = cpu_intr_disable(cp);
1336
1337 /*
1338 * Record that we are aiming to offline this cpu. This acts as
1339 * a barrier to further weak binding requests in thread_nomigrate
1340 * and also causes cpu_choose, disp_lowpri_cpu and setfrontdq to
1341 * lean away from this cpu. Further strong bindings are already
1342 * avoided since we hold cpu_lock. Since threads that are set
1343 * runnable around now and others coming off the target cpu are
1344 * directed away from the target, existing strong and weak bindings
1345 * (especially the latter) to the target cpu stand maximum chance of
1346 * being able to unbind during the short delay loop below (if other
1347 * unbound threads compete they may not see cpu in time to unbind
1348 * even if they would do so immediately.
1349 */
1350 cpu_inmotion = cp;
1351 membar_enter();
1352
1353 /*
1354 * Check for kernel threads (strong or weak) bound to that CPU.
1355 * Strongly bound threads may not unbind, and we'll have to return
1356 * EBUSY. Weakly bound threads should always disappear - we've
1357 * stopped more weak binding with cpu_inmotion and existing
1358 * bindings will drain imminently (they may not block). Nonetheless
1359 * we will wait for a fixed period for all bound threads to disappear.
1360 * Inactive interrupt threads are OK (they'll be in TS_FREE
1361 * state). If test finds some bound threads, wait a few ticks
1362 * to give short-lived threads (such as interrupts) chance to
1363 * complete. Note that if no_quiesce is set, i.e. this cpu
1364 * is required to service interrupts, then we take the route
1365 * that permits interrupt threads to be active (or bypassed).
1366 */
1367 bound_func = no_quiesce ? disp_bound_threads : disp_bound_anythreads;
1368
1369 again: for (loop_count = 0; (*bound_func)(cp, 0); loop_count++) {
1370 if (loop_count >= 5) {
1371 error = EBUSY; /* some threads still bound */
1372 break;
1373 }
1374
1375 /*
1376 * If some threads were assigned, give them
1377 * a chance to complete or move.
1378 *
1379 * This assumes that the clock_thread is not bound
1380 * to any CPU, because the clock_thread is needed to
1381 * do the delay(hz/100).
1382 *
1383 * Note: we still hold the cpu_lock while waiting for
1384 * the next clock tick. This is OK since it isn't
1385 * needed for anything else except processor_bind(2),
1386 * and system initialization. If we drop the lock,
1387 * we would risk another p_online disabling the last
1388 * processor.
1389 */
1390 delay(hz/100);
1391 }
1392
1393 if (error == 0 && callout_off == 0) {
1394 callout_cpu_offline(cp);
1395 callout_off = 1;
1396 }
1397
1398 if (error == 0 && cyclic_off == 0) {
1399 if (!cyclic_offline(cp)) {
1400 /*
1401 * We must have bound cyclics...
1402 */
1403 error = EBUSY;
1404 goto out;
1405 }
1406 cyclic_off = 1;
1407 }
1408
1409 /*
1410 * Call mp_cpu_stop() to perform any special operations
1411 * needed for this machine architecture to offline a CPU.
1412 */
1413 if (error == 0)
1414 error = mp_cpu_stop(cp); /* arch-dep hook */
1415
1416 /*
1417 * If that all worked, take the CPU offline and decrement
1418 * ncpus_online.
1419 */
1420 if (error == 0) {
1421 /*
1422 * Put all the cpus into a known safe place.
1423 * No mutexes can be entered while CPUs are paused.
1424 */
1425 pause_cpus(cp, NULL);
1426 /*
1427 * Repeat the operation, if necessary, to make sure that
1428 * all outstanding low-level interrupts run to completion
1429 * before we set the CPU_QUIESCED flag. It's also possible
1430 * that a thread has weak bound to the cpu despite our raising
1431 * cpu_inmotion above since it may have loaded that
1432 * value before the barrier became visible (this would have
1433 * to be the thread that was on the target cpu at the time
1434 * we raised the barrier).
1435 */
1436 if ((!no_quiesce && cp->cpu_intr_actv != 0) ||
1437 (*bound_func)(cp, 1)) {
1438 start_cpus();
1439 (void) mp_cpu_start(cp);
1440 goto again;
1441 }
1442 ncp = cp->cpu_next_part;
1443 cpu_lpl = cp->cpu_lpl;
1444 ASSERT(cpu_lpl != NULL);
1445
1446 /*
1447 * Remove the CPU from the list of active CPUs.
1448 */
1449 cpu_remove_active(cp);
1450
1451 /*
1452 * Walk the active process list and look for threads
1453 * whose home lgroup needs to be updated, or
1454 * the last CPU they run on is the one being offlined now.
1455 */
1456
1457 ASSERT(curthread->t_cpu != cp);
1458 for (p = practive; p != NULL; p = p->p_next) {
1459
1460 t = p->p_tlist;
1461
1462 if (t == NULL)
1463 continue;
1464
1465 lgrp_diff_lpl = 0;
1466
1467 do {
1468 ASSERT(t->t_lpl != NULL);
1469 /*
1470 * Taking last CPU in lpl offline
1471 * Rehome thread if it is in this lpl
1472 * Otherwise, update the count of how many
1473 * threads are in this CPU's lgroup but have
1474 * a different lpl.
1475 */
1476
1477 if (cpu_lpl->lpl_ncpu == 0) {
1478 if (t->t_lpl == cpu_lpl)
1479 lgrp_move_thread(t,
1480 lgrp_choose(t,
1481 t->t_cpupart), 0);
1482 else if (t->t_lpl->lpl_lgrpid ==
1483 cpu_lpl->lpl_lgrpid)
1484 lgrp_diff_lpl++;
1485 }
1486 ASSERT(t->t_lpl->lpl_ncpu > 0);
1487
1488 /*
1489 * Update CPU last ran on if it was this CPU
1490 */
1491 if (t->t_cpu == cp && t->t_bound_cpu != cp)
1492 t->t_cpu = disp_lowpri_cpu(ncp,
1493 t->t_lpl, t->t_pri, NULL);
1494 ASSERT(t->t_cpu != cp || t->t_bound_cpu == cp ||
1495 t->t_weakbound_cpu == cp);
1496
1497 t = t->t_forw;
1498 } while (t != p->p_tlist);
1499
1500 /*
1501 * Didn't find any threads in the same lgroup as this
1502 * CPU with a different lpl, so remove the lgroup from
1503 * the process lgroup bitmask.
1504 */
1505
1506 if (lgrp_diff_lpl == 0)
1507 klgrpset_del(p->p_lgrpset, cpu_lpl->lpl_lgrpid);
1508 }
1509
1510 /*
1511 * Walk thread list looking for threads that need to be
1512 * rehomed, since there are some threads that are not in
1513 * their process's p_tlist.
1514 */
1515
1516 t = curthread;
1517 do {
1518 ASSERT(t != NULL && t->t_lpl != NULL);
1519
1520 /*
1521 * Rehome threads with same lpl as this CPU when this
1522 * is the last CPU in the lpl.
1523 */
1524
1525 if ((cpu_lpl->lpl_ncpu == 0) && (t->t_lpl == cpu_lpl))
1526 lgrp_move_thread(t,
1527 lgrp_choose(t, t->t_cpupart), 1);
1528
1529 ASSERT(t->t_lpl->lpl_ncpu > 0);
1530
1531 /*
1532 * Update CPU last ran on if it was this CPU
1533 */
1534
1535 if (t->t_cpu == cp && t->t_bound_cpu != cp) {
1536 t->t_cpu = disp_lowpri_cpu(ncp,
1537 t->t_lpl, t->t_pri, NULL);
1538 }
1539 ASSERT(t->t_cpu != cp || t->t_bound_cpu == cp ||
1540 t->t_weakbound_cpu == cp);
1541 t = t->t_next;
1542
1543 } while (t != curthread);
1544 ASSERT((cp->cpu_flags & (CPU_FAULTED | CPU_SPARE)) == 0);
1545 cp->cpu_flags |= CPU_OFFLINE;
1546 disp_cpu_inactive(cp);
1547 if (!no_quiesce)
1548 cp->cpu_flags |= CPU_QUIESCED;
1549 ncpus_online--;
1550 cpu_set_state(cp);
1551 cpu_inmotion = NULL;
1552 start_cpus();
1553 cpu_stats_kstat_destroy(cp);
1554 cpu_delete_intrstat(cp);
1555 lgrp_kstat_destroy(cp);
1556 }
1557
1558 out:
1559 cpu_inmotion = NULL;
1560
1561 /*
1562 * If we failed, re-enable interrupts.
1563 * Do this even if cpu_intr_disable returned an error, because
1564 * it may have partially disabled interrupts.
1565 */
1566 if (error && intr_enable)
1567 cpu_intr_enable(cp);
1568
1569 /*
1570 * If we failed, but managed to offline the cyclic subsystem on this
1571 * CPU, bring it back online.
1572 */
1573 if (error && cyclic_off)
1574 cyclic_online(cp);
1575
1576 /*
1577 * If we failed, but managed to offline callouts on this CPU,
1578 * bring it back online.
1579 */
1580 if (error && callout_off)
1581 callout_cpu_online(cp);
1582
1583 /*
1584 * If we failed, tell the PG subsystem that the CPU is back
1585 */
1586 pg_cpupart_in(cp, pp);
1587
1588 /*
1589 * If we failed, we need to notify everyone that this CPU is back on.
1590 */
1591 if (error != 0) {
1592 CPU_NEW_GENERATION(cp);
1593 cpu_state_change_notify(cp->cpu_id, CPU_ON);
1594 cpu_state_change_notify(cp->cpu_id, CPU_INTR_ON);
1595 }
1596
1597 return (error);
1598 }
1599
1600 /*
1601 * Mark the indicated CPU as faulted, taking it offline.
1602 */
1603 int
1604 cpu_faulted(cpu_t *cp, int flags)
1605 {
1606 int error = 0;
1607
1608 ASSERT(MUTEX_HELD(&cpu_lock));
1609 ASSERT(!cpu_is_poweredoff(cp));
1610
1611 if (cpu_is_offline(cp)) {
1612 cp->cpu_flags &= ~CPU_SPARE;
1613 cp->cpu_flags |= CPU_FAULTED;
1614 mp_cpu_faulted_enter(cp);
1615 cpu_set_state(cp);
1616 return (0);
1617 }
1618
1619 if ((error = cpu_offline(cp, flags)) == 0) {
1620 cp->cpu_flags |= CPU_FAULTED;
1621 mp_cpu_faulted_enter(cp);
1622 cpu_set_state(cp);
1623 }
1624
1625 return (error);
1626 }
1627
1628 /*
1629 * Mark the indicated CPU as a spare, taking it offline.
1630 */
1631 int
1632 cpu_spare(cpu_t *cp, int flags)
1633 {
1634 int error = 0;
1635
1636 ASSERT(MUTEX_HELD(&cpu_lock));
1637 ASSERT(!cpu_is_poweredoff(cp));
1638
1639 if (cpu_is_offline(cp)) {
1640 if (cp->cpu_flags & CPU_FAULTED) {
1641 cp->cpu_flags &= ~CPU_FAULTED;
1642 mp_cpu_faulted_exit(cp);
1643 }
1644 cp->cpu_flags |= CPU_SPARE;
1645 cpu_set_state(cp);
1646 return (0);
1647 }
1648
1649 if ((error = cpu_offline(cp, flags)) == 0) {
1650 cp->cpu_flags |= CPU_SPARE;
1651 cpu_set_state(cp);
1652 }
1653
1654 return (error);
1655 }
1656
1657 /*
1658 * Take the indicated CPU from poweroff to offline.
1659 */
1660 int
1661 cpu_poweron(cpu_t *cp)
1662 {
1663 int error = ENOTSUP;
1664
1665 ASSERT(MUTEX_HELD(&cpu_lock));
1666 ASSERT(cpu_is_poweredoff(cp));
1667
1668 error = mp_cpu_poweron(cp); /* arch-dep hook */
1669 if (error == 0)
1670 cpu_set_state(cp);
1671
1672 return (error);
1673 }
1674
1675 /*
1676 * Take the indicated CPU from any inactive state to powered off.
1677 */
1678 int
1679 cpu_poweroff(cpu_t *cp)
1680 {
1681 int error = ENOTSUP;
1682
1683 ASSERT(MUTEX_HELD(&cpu_lock));
1684 ASSERT(cpu_is_offline(cp));
1685
1686 if (!(cp->cpu_flags & CPU_QUIESCED))
1687 return (EBUSY); /* not completely idle */
1688
1689 error = mp_cpu_poweroff(cp); /* arch-dep hook */
1690 if (error == 0)
1691 cpu_set_state(cp);
1692
1693 return (error);
1694 }
1695
1696 /*
1697 * Initialize the Sequential CPU id lookup table
1698 */
1699 void
1700 cpu_seq_tbl_init()
1701 {
1702 cpu_t **tbl;
1703
1704 tbl = kmem_zalloc(sizeof (struct cpu *) * max_ncpus, KM_SLEEP);
1705 tbl[0] = CPU;
1706
1707 cpu_seq = tbl;
1708 }
1709
1710 /*
1711 * Initialize the CPU lists for the first CPU.
1712 */
1713 void
1714 cpu_list_init(cpu_t *cp)
1715 {
1716 cp->cpu_next = cp;
1717 cp->cpu_prev = cp;
1718 cpu_list = cp;
1719 clock_cpu_list = cp;
1720
1721 cp->cpu_next_onln = cp;
1722 cp->cpu_prev_onln = cp;
1723 cpu_active = cp;
1724
1725 cp->cpu_seqid = 0;
1726 CPUSET_ADD(cpu_seqid_inuse, 0);
1727
1728 /*
1729 * Bootstrap cpu_seq using cpu_list
1730 * The cpu_seq[] table will be dynamically allocated
1731 * when kmem later becomes available (but before going MP)
1732 */
1733 cpu_seq = &cpu_list;
1734
1735 cp->cpu_cache_offset = KMEM_CPU_CACHE_OFFSET(cp->cpu_seqid);
1736 cp_default.cp_cpulist = cp;
1737 cp_default.cp_ncpus = 1;
1738 cp->cpu_next_part = cp;
1739 cp->cpu_prev_part = cp;
1740 cp->cpu_part = &cp_default;
1741
1742 CPUSET_ADD(cpu_available, cp->cpu_id);
1743 }
1744
1745 /*
1746 * Insert a CPU into the list of available CPUs.
1747 */
1748 void
1749 cpu_add_unit(cpu_t *cp)
1750 {
1751 int seqid;
1752
1753 ASSERT(MUTEX_HELD(&cpu_lock));
1754 ASSERT(cpu_list != NULL); /* list started in cpu_list_init */
1755
1756 lgrp_config(LGRP_CONFIG_CPU_ADD, (uintptr_t)cp, 0);
1757
1758 /*
1759 * Note: most users of the cpu_list will grab the
1760 * cpu_lock to insure that it isn't modified. However,
1761 * certain users can't or won't do that. To allow this
1762 * we pause the other cpus. Users who walk the list
1763 * without cpu_lock, must disable kernel preemption
1764 * to insure that the list isn't modified underneath
1765 * them. Also, any cached pointers to cpu structures
1766 * must be revalidated by checking to see if the
1767 * cpu_next pointer points to itself. This check must
1768 * be done with the cpu_lock held or kernel preemption
1769 * disabled. This check relies upon the fact that
1770 * old cpu structures are not free'ed or cleared after
1771 * then are removed from the cpu_list.
1772 *
1773 * Note that the clock code walks the cpu list dereferencing
1774 * the cpu_part pointer, so we need to initialize it before
1775 * adding the cpu to the list.
1776 */
1777 cp->cpu_part = &cp_default;
1778 pause_cpus(NULL, NULL);
1779 cp->cpu_next = cpu_list;
1780 cp->cpu_prev = cpu_list->cpu_prev;
1781 cpu_list->cpu_prev->cpu_next = cp;
1782 cpu_list->cpu_prev = cp;
1783 start_cpus();
1784
1785 for (seqid = 0; CPU_IN_SET(cpu_seqid_inuse, seqid); seqid++)
1786 continue;
1787 CPUSET_ADD(cpu_seqid_inuse, seqid);
1788 cp->cpu_seqid = seqid;
1789
1790 if (seqid > max_cpu_seqid_ever)
1791 max_cpu_seqid_ever = seqid;
1792
1793 ASSERT(ncpus < max_ncpus);
1794 ncpus++;
1795 cp->cpu_cache_offset = KMEM_CPU_CACHE_OFFSET(cp->cpu_seqid);
1796 cpu[cp->cpu_id] = cp;
1797 CPUSET_ADD(cpu_available, cp->cpu_id);
1798 cpu_seq[cp->cpu_seqid] = cp;
1799
1800 /*
1801 * allocate a pause thread for this CPU.
1802 */
1803 cpu_pause_alloc(cp);
1804
1805 /*
1806 * So that new CPUs won't have NULL prev_onln and next_onln pointers,
1807 * link them into a list of just that CPU.
1808 * This is so that disp_lowpri_cpu will work for thread_create in
1809 * pause_cpus() when called from the startup thread in a new CPU.
1810 */
1811 cp->cpu_next_onln = cp;
1812 cp->cpu_prev_onln = cp;
1813 cpu_info_kstat_create(cp);
1814 cp->cpu_next_part = cp;
1815 cp->cpu_prev_part = cp;
1816
1817 init_cpu_mstate(cp, CMS_SYSTEM);
1818
1819 pool_pset_mod = gethrtime();
1820 }
1821
1822 /*
1823 * Do the opposite of cpu_add_unit().
1824 */
1825 void
1826 cpu_del_unit(int cpuid)
1827 {
1828 struct cpu *cp, *cpnext;
1829
1830 ASSERT(MUTEX_HELD(&cpu_lock));
1831 cp = cpu[cpuid];
1832 ASSERT(cp != NULL);
1833
1834 ASSERT(cp->cpu_next_onln == cp);
1835 ASSERT(cp->cpu_prev_onln == cp);
1836 ASSERT(cp->cpu_next_part == cp);
1837 ASSERT(cp->cpu_prev_part == cp);
1838
1839 /*
1840 * Tear down the CPU's physical ID cache, and update any
1841 * processor groups
1842 */
1843 pg_cpu_fini(cp, NULL);
1844 pghw_physid_destroy(cp);
1845
1846 /*
1847 * Destroy kstat stuff.
1848 */
1849 cpu_info_kstat_destroy(cp);
1850 term_cpu_mstate(cp);
1851 /*
1852 * Free up pause thread.
1853 */
1854 cpu_pause_free(cp);
1855 CPUSET_DEL(cpu_available, cp->cpu_id);
1856 cpu[cp->cpu_id] = NULL;
1857 cpu_seq[cp->cpu_seqid] = NULL;
1858
1859 /*
1860 * The clock thread and mutex_vector_enter cannot hold the
1861 * cpu_lock while traversing the cpu list, therefore we pause
1862 * all other threads by pausing the other cpus. These, and any
1863 * other routines holding cpu pointers while possibly sleeping
1864 * must be sure to call kpreempt_disable before processing the
1865 * list and be sure to check that the cpu has not been deleted
1866 * after any sleeps (check cp->cpu_next != NULL). We guarantee
1867 * to keep the deleted cpu structure around.
1868 *
1869 * Note that this MUST be done AFTER cpu_available
1870 * has been updated so that we don't waste time
1871 * trying to pause the cpu we're trying to delete.
1872 */
1873 pause_cpus(NULL, NULL);
1874
1875 cpnext = cp->cpu_next;
1876 cp->cpu_prev->cpu_next = cp->cpu_next;
1877 cp->cpu_next->cpu_prev = cp->cpu_prev;
1878 if (cp == cpu_list)
1879 cpu_list = cpnext;
1880
1881 /*
1882 * Signals that the cpu has been deleted (see above).
1883 */
1884 cp->cpu_next = NULL;
1885 cp->cpu_prev = NULL;
1886
1887 start_cpus();
1888
1889 CPUSET_DEL(cpu_seqid_inuse, cp->cpu_seqid);
1890 ncpus--;
1891 lgrp_config(LGRP_CONFIG_CPU_DEL, (uintptr_t)cp, 0);
1892
1893 pool_pset_mod = gethrtime();
1894 }
1895
1896 /*
1897 * Add a CPU to the list of active CPUs.
1898 * This routine must not get any locks, because other CPUs are paused.
1899 */
1900 static void
1901 cpu_add_active_internal(cpu_t *cp)
1902 {
1903 cpupart_t *pp = cp->cpu_part;
1904
1905 ASSERT(MUTEX_HELD(&cpu_lock));
1906 ASSERT(cpu_list != NULL); /* list started in cpu_list_init */
1907
1908 ncpus_online++;
1909 cpu_set_state(cp);
1910 cp->cpu_next_onln = cpu_active;
1911 cp->cpu_prev_onln = cpu_active->cpu_prev_onln;
1912 cpu_active->cpu_prev_onln->cpu_next_onln = cp;
1913 cpu_active->cpu_prev_onln = cp;
1914
1915 if (pp->cp_cpulist) {
1916 cp->cpu_next_part = pp->cp_cpulist;
1917 cp->cpu_prev_part = pp->cp_cpulist->cpu_prev_part;
1918 pp->cp_cpulist->cpu_prev_part->cpu_next_part = cp;
1919 pp->cp_cpulist->cpu_prev_part = cp;
1920 } else {
1921 ASSERT(pp->cp_ncpus == 0);
1922 pp->cp_cpulist = cp->cpu_next_part = cp->cpu_prev_part = cp;
1923 }
1924 pp->cp_ncpus++;
1925 if (pp->cp_ncpus == 1) {
1926 cp_numparts_nonempty++;
1927 ASSERT(cp_numparts_nonempty != 0);
1928 }
1929
1930 pg_cpu_active(cp);
1931 lgrp_config(LGRP_CONFIG_CPU_ONLINE, (uintptr_t)cp, 0);
1932
1933 bzero(&cp->cpu_loadavg, sizeof (cp->cpu_loadavg));
1934 }
1935
1936 /*
1937 * Add a CPU to the list of active CPUs.
1938 * This is called from machine-dependent layers when a new CPU is started.
1939 */
1940 void
1941 cpu_add_active(cpu_t *cp)
1942 {
1943 pg_cpupart_in(cp, cp->cpu_part);
1944
1945 pause_cpus(NULL, NULL);
1946 cpu_add_active_internal(cp);
1947 start_cpus();
1948
1949 cpu_stats_kstat_create(cp);
1950 cpu_create_intrstat(cp);
1951 lgrp_kstat_create(cp);
1952 cpu_state_change_notify(cp->cpu_id, CPU_INIT);
1953 }
1954
1955
1956 /*
1957 * Remove a CPU from the list of active CPUs.
1958 * This routine must not get any locks, because other CPUs are paused.
1959 */
1960 /* ARGSUSED */
1961 static void
1962 cpu_remove_active(cpu_t *cp)
1963 {
1964 cpupart_t *pp = cp->cpu_part;
1965
1966 ASSERT(MUTEX_HELD(&cpu_lock));
1967 ASSERT(cp->cpu_next_onln != cp); /* not the last one */
1968 ASSERT(cp->cpu_prev_onln != cp); /* not the last one */
1969
1970 pg_cpu_inactive(cp);
1971
1972 lgrp_config(LGRP_CONFIG_CPU_OFFLINE, (uintptr_t)cp, 0);
1973
1974 if (cp == clock_cpu_list)
1975 clock_cpu_list = cp->cpu_next_onln;
1976
1977 cp->cpu_prev_onln->cpu_next_onln = cp->cpu_next_onln;
1978 cp->cpu_next_onln->cpu_prev_onln = cp->cpu_prev_onln;
1979 if (cpu_active == cp) {
1980 cpu_active = cp->cpu_next_onln;
1981 }
1982 cp->cpu_next_onln = cp;
1983 cp->cpu_prev_onln = cp;
1984
1985 cp->cpu_prev_part->cpu_next_part = cp->cpu_next_part;
1986 cp->cpu_next_part->cpu_prev_part = cp->cpu_prev_part;
1987 if (pp->cp_cpulist == cp) {
1988 pp->cp_cpulist = cp->cpu_next_part;
1989 ASSERT(pp->cp_cpulist != cp);
1990 }
1991 cp->cpu_next_part = cp;
1992 cp->cpu_prev_part = cp;
1993 pp->cp_ncpus--;
1994 if (pp->cp_ncpus == 0) {
1995 cp_numparts_nonempty--;
1996 ASSERT(cp_numparts_nonempty != 0);
1997 }
1998 }
1999
2000 /*
2001 * Routine used to setup a newly inserted CPU in preparation for starting
2002 * it running code.
2003 */
2004 int
2005 cpu_configure(int cpuid)
2006 {
2007 int retval = 0;
2008
2009 ASSERT(MUTEX_HELD(&cpu_lock));
2010
2011 /*
2012 * Some structures are statically allocated based upon
2013 * the maximum number of cpus the system supports. Do not
2014 * try to add anything beyond this limit.
2015 */
2016 if (cpuid < 0 || cpuid >= NCPU) {
2017 return (EINVAL);
2018 }
2019
2020 if ((cpu[cpuid] != NULL) && (cpu[cpuid]->cpu_flags != 0)) {
2021 return (EALREADY);
2022 }
2023
2024 if ((retval = mp_cpu_configure(cpuid)) != 0) {
2025 return (retval);
2026 }
2027
2028 cpu[cpuid]->cpu_flags = CPU_QUIESCED | CPU_OFFLINE | CPU_POWEROFF;
2029 cpu_set_state(cpu[cpuid]);
2030 retval = cpu_state_change_hooks(cpuid, CPU_CONFIG, CPU_UNCONFIG);
2031 if (retval != 0)
2032 (void) mp_cpu_unconfigure(cpuid);
2033
2034 return (retval);
2035 }
2036
2037 /*
2038 * Routine used to cleanup a CPU that has been powered off. This will
2039 * destroy all per-cpu information related to this cpu.
2040 */
2041 int
2042 cpu_unconfigure(int cpuid)
2043 {
2044 int error;
2045
2046 ASSERT(MUTEX_HELD(&cpu_lock));
2047
2048 if (cpu[cpuid] == NULL) {
2049 return (ENODEV);
2050 }
2051
2052 if (cpu[cpuid]->cpu_flags == 0) {
2053 return (EALREADY);
2054 }
2055
2056 if ((cpu[cpuid]->cpu_flags & CPU_POWEROFF) == 0) {
2057 return (EBUSY);
2058 }
2059
2060 if (cpu[cpuid]->cpu_props != NULL) {
2061 (void) nvlist_free(cpu[cpuid]->cpu_props);
2062 cpu[cpuid]->cpu_props = NULL;
2063 }
2064
2065 error = cpu_state_change_hooks(cpuid, CPU_UNCONFIG, CPU_CONFIG);
2066
2067 if (error != 0)
2068 return (error);
2069
2070 return (mp_cpu_unconfigure(cpuid));
2071 }
2072
2073 /*
2074 * Routines for registering and de-registering cpu_setup callback functions.
2075 *
2076 * Caller's context
2077 * These routines must not be called from a driver's attach(9E) or
2078 * detach(9E) entry point.
2079 *
2080 * NOTE: CPU callbacks should not block. They are called with cpu_lock held.
2081 */
2082
2083 /*
2084 * Ideally, these would be dynamically allocated and put into a linked
2085 * list; however that is not feasible because the registration routine
2086 * has to be available before the kmem allocator is working (in fact,
2087 * it is called by the kmem allocator init code). In any case, there
2088 * are quite a few extra entries for future users.
2089 */
2090 #define NCPU_SETUPS 20
2091
2092 struct cpu_setup {
2093 cpu_setup_func_t *func;
2094 void *arg;
2095 } cpu_setups[NCPU_SETUPS];
2096
2097 void
2098 register_cpu_setup_func(cpu_setup_func_t *func, void *arg)
2099 {
2100 int i;
2101
2102 ASSERT(MUTEX_HELD(&cpu_lock));
2103
2104 for (i = 0; i < NCPU_SETUPS; i++)
2105 if (cpu_setups[i].func == NULL)
2106 break;
2107 if (i >= NCPU_SETUPS)
2108 cmn_err(CE_PANIC, "Ran out of cpu_setup callback entries");
2109
2110 cpu_setups[i].func = func;
2111 cpu_setups[i].arg = arg;
2112 }
2113
2114 void
2115 unregister_cpu_setup_func(cpu_setup_func_t *func, void *arg)
2116 {
2117 int i;
2118
2119 ASSERT(MUTEX_HELD(&cpu_lock));
2120
2121 for (i = 0; i < NCPU_SETUPS; i++)
2122 if ((cpu_setups[i].func == func) &&
2123 (cpu_setups[i].arg == arg))
2124 break;
2125 if (i >= NCPU_SETUPS)
2126 cmn_err(CE_PANIC, "Could not find cpu_setup callback to "
2127 "deregister");
2128
2129 cpu_setups[i].func = NULL;
2130 cpu_setups[i].arg = 0;
2131 }
2132
2133 /*
2134 * Call any state change hooks for this CPU, ignore any errors.
2135 */
2136 void
2137 cpu_state_change_notify(int id, cpu_setup_t what)
2138 {
2139 int i;
2140
2141 ASSERT(MUTEX_HELD(&cpu_lock));
2142
2143 for (i = 0; i < NCPU_SETUPS; i++) {
2144 if (cpu_setups[i].func != NULL) {
2145 cpu_setups[i].func(what, id, cpu_setups[i].arg);
2146 }
2147 }
2148 }
2149
2150 /*
2151 * Call any state change hooks for this CPU, undo it if error found.
2152 */
2153 static int
2154 cpu_state_change_hooks(int id, cpu_setup_t what, cpu_setup_t undo)
2155 {
2156 int i;
2157 int retval = 0;
2158
2159 ASSERT(MUTEX_HELD(&cpu_lock));
2160
2161 for (i = 0; i < NCPU_SETUPS; i++) {
2162 if (cpu_setups[i].func != NULL) {
2163 retval = cpu_setups[i].func(what, id,
2164 cpu_setups[i].arg);
2165 if (retval) {
2166 for (i--; i >= 0; i--) {
2167 if (cpu_setups[i].func != NULL)
2168 cpu_setups[i].func(undo,
2169 id, cpu_setups[i].arg);
2170 }
2171 break;
2172 }
2173 }
2174 }
2175 return (retval);
2176 }
2177
2178 /*
2179 * Export information about this CPU via the kstat mechanism.
2180 */
2181 static struct {
2182 kstat_named_t ci_state;
2183 kstat_named_t ci_state_begin;
2184 kstat_named_t ci_cpu_type;
2185 kstat_named_t ci_fpu_type;
2186 kstat_named_t ci_clock_MHz;
2187 kstat_named_t ci_chip_id;
2188 kstat_named_t ci_implementation;
2189 kstat_named_t ci_brandstr;
2190 kstat_named_t ci_core_id;
2191 kstat_named_t ci_curr_clock_Hz;
2192 kstat_named_t ci_supp_freq_Hz;
2193 kstat_named_t ci_pg_id;
2194 #if defined(__sparcv9)
2195 kstat_named_t ci_device_ID;
2196 kstat_named_t ci_cpu_fru;
2197 #endif
2198 #if defined(__x86)
2199 kstat_named_t ci_vendorstr;
2200 kstat_named_t ci_family;
2201 kstat_named_t ci_model;
2202 kstat_named_t ci_step;
2203 kstat_named_t ci_clogid;
2204 kstat_named_t ci_pkg_core_id;
2205 kstat_named_t ci_ncpuperchip;
2206 kstat_named_t ci_ncoreperchip;
2207 kstat_named_t ci_max_cstates;
2208 kstat_named_t ci_curr_cstate;
2209 kstat_named_t ci_cacheid;
2210 kstat_named_t ci_sktstr;
2211 #endif
2212 } cpu_info_template = {
2213 { "state", KSTAT_DATA_CHAR },
2214 { "state_begin", KSTAT_DATA_LONG },
2215 { "cpu_type", KSTAT_DATA_CHAR },
2216 { "fpu_type", KSTAT_DATA_CHAR },
2217 { "clock_MHz", KSTAT_DATA_LONG },
2218 { "chip_id", KSTAT_DATA_LONG },
2219 { "implementation", KSTAT_DATA_STRING },
2220 { "brand", KSTAT_DATA_STRING },
2221 { "core_id", KSTAT_DATA_LONG },
2222 { "current_clock_Hz", KSTAT_DATA_UINT64 },
2223 { "supported_frequencies_Hz", KSTAT_DATA_STRING },
2224 { "pg_id", KSTAT_DATA_LONG },
2225 #if defined(__sparcv9)
2226 { "device_ID", KSTAT_DATA_UINT64 },
2227 { "cpu_fru", KSTAT_DATA_STRING },
2228 #endif
2229 #if defined(__x86)
2230 { "vendor_id", KSTAT_DATA_STRING },
2231 { "family", KSTAT_DATA_INT32 },
2232 { "model", KSTAT_DATA_INT32 },
2233 { "stepping", KSTAT_DATA_INT32 },
2234 { "clog_id", KSTAT_DATA_INT32 },
2235 { "pkg_core_id", KSTAT_DATA_LONG },
2236 { "ncpu_per_chip", KSTAT_DATA_INT32 },
2237 { "ncore_per_chip", KSTAT_DATA_INT32 },
2238 { "supported_max_cstates", KSTAT_DATA_INT32 },
2239 { "current_cstate", KSTAT_DATA_INT32 },
2240 { "cache_id", KSTAT_DATA_INT32 },
2241 { "socket_type", KSTAT_DATA_STRING },
2242 #endif
2243 };
2244
2245 static kmutex_t cpu_info_template_lock;
2246
2247 static int
2248 cpu_info_kstat_update(kstat_t *ksp, int rw)
2249 {
2250 cpu_t *cp = ksp->ks_private;
2251 const char *pi_state;
2252
2253 if (rw == KSTAT_WRITE)
2254 return (EACCES);
2255
2256 #if defined(__x86)
2257 /* Is the cpu still initialising itself? */
2258 if (cpuid_checkpass(cp, 1) == 0)
2259 return (ENXIO);
2260 #endif
2261 switch (cp->cpu_type_info.pi_state) {
2262 case P_ONLINE:
2263 pi_state = PS_ONLINE;
2264 break;
2265 case P_POWEROFF:
2266 pi_state = PS_POWEROFF;
2267 break;
2268 case P_NOINTR:
2269 pi_state = PS_NOINTR;
2270 break;
2271 case P_FAULTED:
2272 pi_state = PS_FAULTED;
2273 break;
2274 case P_SPARE:
2275 pi_state = PS_SPARE;
2276 break;
2277 case P_OFFLINE:
2278 pi_state = PS_OFFLINE;
2279 break;
2280 default:
2281 pi_state = "unknown";
2282 }
2283 (void) strcpy(cpu_info_template.ci_state.value.c, pi_state);
2284 cpu_info_template.ci_state_begin.value.l = cp->cpu_state_begin;
2285 (void) strncpy(cpu_info_template.ci_cpu_type.value.c,
2286 cp->cpu_type_info.pi_processor_type, 15);
2287 (void) strncpy(cpu_info_template.ci_fpu_type.value.c,
2288 cp->cpu_type_info.pi_fputypes, 15);
2289 cpu_info_template.ci_clock_MHz.value.l = cp->cpu_type_info.pi_clock;
2290 cpu_info_template.ci_chip_id.value.l =
2291 pg_plat_hw_instance_id(cp, PGHW_CHIP);
2292 kstat_named_setstr(&cpu_info_template.ci_implementation,
2293 cp->cpu_idstr);
2294 kstat_named_setstr(&cpu_info_template.ci_brandstr, cp->cpu_brandstr);
2295 cpu_info_template.ci_core_id.value.l = pg_plat_get_core_id(cp);
2296 cpu_info_template.ci_curr_clock_Hz.value.ui64 =
2297 cp->cpu_curr_clock;
2298 cpu_info_template.ci_pg_id.value.l =
2299 cp->cpu_pg && cp->cpu_pg->cmt_lineage ?
2300 cp->cpu_pg->cmt_lineage->pg_id : -1;
2301 kstat_named_setstr(&cpu_info_template.ci_supp_freq_Hz,
2302 cp->cpu_supp_freqs);
2303 #if defined(__sparcv9)
2304 cpu_info_template.ci_device_ID.value.ui64 =
2305 cpunodes[cp->cpu_id].device_id;
2306 kstat_named_setstr(&cpu_info_template.ci_cpu_fru, cpu_fru_fmri(cp));
2307 #endif
2308 #if defined(__x86)
2309 kstat_named_setstr(&cpu_info_template.ci_vendorstr,
2310 cpuid_getvendorstr(cp));
2311 cpu_info_template.ci_family.value.l = cpuid_getfamily(cp);
2312 cpu_info_template.ci_model.value.l = cpuid_getmodel(cp);
2313 cpu_info_template.ci_step.value.l = cpuid_getstep(cp);
2314 cpu_info_template.ci_clogid.value.l = cpuid_get_clogid(cp);
2315 cpu_info_template.ci_ncpuperchip.value.l = cpuid_get_ncpu_per_chip(cp);
2316 cpu_info_template.ci_ncoreperchip.value.l =
2317 cpuid_get_ncore_per_chip(cp);
2318 cpu_info_template.ci_pkg_core_id.value.l = cpuid_get_pkgcoreid(cp);
2319 cpu_info_template.ci_max_cstates.value.l = cp->cpu_m.max_cstates;
2320 cpu_info_template.ci_curr_cstate.value.l = cpu_idle_get_cpu_state(cp);
2321 cpu_info_template.ci_cacheid.value.i32 = cpuid_get_cacheid(cp);
2322 kstat_named_setstr(&cpu_info_template.ci_sktstr,
2323 cpuid_getsocketstr(cp));
2324 #endif
2325
2326 return (0);
2327 }
2328
2329 static void
2330 cpu_info_kstat_create(cpu_t *cp)
2331 {
2332 zoneid_t zoneid;
2333
2334 ASSERT(MUTEX_HELD(&cpu_lock));
2335
2336 if (pool_pset_enabled())
2337 zoneid = GLOBAL_ZONEID;
2338 else
2339 zoneid = ALL_ZONES;
2340 if ((cp->cpu_info_kstat = kstat_create_zone("cpu_info", cp->cpu_id,
2341 NULL, "misc", KSTAT_TYPE_NAMED,
2342 sizeof (cpu_info_template) / sizeof (kstat_named_t),
2343 KSTAT_FLAG_VIRTUAL | KSTAT_FLAG_VAR_SIZE, zoneid)) != NULL) {
2344 cp->cpu_info_kstat->ks_data_size += 2 * CPU_IDSTRLEN;
2345 #if defined(__sparcv9)
2346 cp->cpu_info_kstat->ks_data_size +=
2347 strlen(cpu_fru_fmri(cp)) + 1;
2348 #endif
2349 #if defined(__x86)
2350 cp->cpu_info_kstat->ks_data_size += X86_VENDOR_STRLEN;
2351 #endif
2352 if (cp->cpu_supp_freqs != NULL)
2353 cp->cpu_info_kstat->ks_data_size +=
2354 strlen(cp->cpu_supp_freqs) + 1;
2355 cp->cpu_info_kstat->ks_lock = &cpu_info_template_lock;
2356 cp->cpu_info_kstat->ks_data = &cpu_info_template;
2357 cp->cpu_info_kstat->ks_private = cp;
2358 cp->cpu_info_kstat->ks_update = cpu_info_kstat_update;
2359 kstat_install(cp->cpu_info_kstat);
2360 }
2361 }
2362
2363 static void
2364 cpu_info_kstat_destroy(cpu_t *cp)
2365 {
2366 ASSERT(MUTEX_HELD(&cpu_lock));
2367
2368 kstat_delete(cp->cpu_info_kstat);
2369 cp->cpu_info_kstat = NULL;
2370 }
2371
2372 /*
2373 * Create and install kstats for the boot CPU.
2374 */
2375 void
2376 cpu_kstat_init(cpu_t *cp)
2377 {
2378 mutex_enter(&cpu_lock);
2379 cpu_info_kstat_create(cp);
2380 cpu_stats_kstat_create(cp);
2381 cpu_create_intrstat(cp);
2382 cpu_set_state(cp);
2383 mutex_exit(&cpu_lock);
2384 }
2385
2386 /*
2387 * Make visible to the zone that subset of the cpu information that would be
2388 * initialized when a cpu is configured (but still offline).
2389 */
2390 void
2391 cpu_visibility_configure(cpu_t *cp, zone_t *zone)
2392 {
2393 zoneid_t zoneid = zone ? zone->zone_id : ALL_ZONES;
2394
2395 ASSERT(MUTEX_HELD(&cpu_lock));
2396 ASSERT(pool_pset_enabled());
2397 ASSERT(cp != NULL);
2398
2399 if (zoneid != ALL_ZONES && zoneid != GLOBAL_ZONEID) {
2400 zone->zone_ncpus++;
2401 ASSERT(zone->zone_ncpus <= ncpus);
2402 }
2403 if (cp->cpu_info_kstat != NULL)
2404 kstat_zone_add(cp->cpu_info_kstat, zoneid);
2405 }
2406
2407 /*
2408 * Make visible to the zone that subset of the cpu information that would be
2409 * initialized when a previously configured cpu is onlined.
2410 */
2411 void
2412 cpu_visibility_online(cpu_t *cp, zone_t *zone)
2413 {
2414 kstat_t *ksp;
2415 char name[sizeof ("cpu_stat") + 10]; /* enough for 32-bit cpuids */
2416 zoneid_t zoneid = zone ? zone->zone_id : ALL_ZONES;
2417 processorid_t cpun;
2418
2419 ASSERT(MUTEX_HELD(&cpu_lock));
2420 ASSERT(pool_pset_enabled());
2421 ASSERT(cp != NULL);
2422 ASSERT(cpu_is_active(cp));
2423
2424 cpun = cp->cpu_id;
2425 if (zoneid != ALL_ZONES && zoneid != GLOBAL_ZONEID) {
2426 zone->zone_ncpus_online++;
2427 ASSERT(zone->zone_ncpus_online <= ncpus_online);
2428 }
2429 (void) snprintf(name, sizeof (name), "cpu_stat%d", cpun);
2430 if ((ksp = kstat_hold_byname("cpu_stat", cpun, name, ALL_ZONES))
2431 != NULL) {
2432 kstat_zone_add(ksp, zoneid);
2433 kstat_rele(ksp);
2434 }
2435 if ((ksp = kstat_hold_byname("cpu", cpun, "sys", ALL_ZONES)) != NULL) {
2436 kstat_zone_add(ksp, zoneid);
2437 kstat_rele(ksp);
2438 }
2439 if ((ksp = kstat_hold_byname("cpu", cpun, "vm", ALL_ZONES)) != NULL) {
2440 kstat_zone_add(ksp, zoneid);
2441 kstat_rele(ksp);
2442 }
2443 if ((ksp = kstat_hold_byname("cpu", cpun, "intrstat", ALL_ZONES)) !=
2444 NULL) {
2445 kstat_zone_add(ksp, zoneid);
2446 kstat_rele(ksp);
2447 }
2448 }
2449
2450 /*
2451 * Update relevant kstats such that cpu is now visible to processes
2452 * executing in specified zone.
2453 */
2454 void
2455 cpu_visibility_add(cpu_t *cp, zone_t *zone)
2456 {
2457 cpu_visibility_configure(cp, zone);
2458 if (cpu_is_active(cp))
2459 cpu_visibility_online(cp, zone);
2460 }
2461
2462 /*
2463 * Make invisible to the zone that subset of the cpu information that would be
2464 * torn down when a previously offlined cpu is unconfigured.
2465 */
2466 void
2467 cpu_visibility_unconfigure(cpu_t *cp, zone_t *zone)
2468 {
2469 zoneid_t zoneid = zone ? zone->zone_id : ALL_ZONES;
2470
2471 ASSERT(MUTEX_HELD(&cpu_lock));
2472 ASSERT(pool_pset_enabled());
2473 ASSERT(cp != NULL);
2474
2475 if (zoneid != ALL_ZONES && zoneid != GLOBAL_ZONEID) {
2476 ASSERT(zone->zone_ncpus != 0);
2477 zone->zone_ncpus--;
2478 }
2479 if (cp->cpu_info_kstat)
2480 kstat_zone_remove(cp->cpu_info_kstat, zoneid);
2481 }
2482
2483 /*
2484 * Make invisible to the zone that subset of the cpu information that would be
2485 * torn down when a cpu is offlined (but still configured).
2486 */
2487 void
2488 cpu_visibility_offline(cpu_t *cp, zone_t *zone)
2489 {
2490 kstat_t *ksp;
2491 char name[sizeof ("cpu_stat") + 10]; /* enough for 32-bit cpuids */
2492 zoneid_t zoneid = zone ? zone->zone_id : ALL_ZONES;
2493 processorid_t cpun;
2494
2495 ASSERT(MUTEX_HELD(&cpu_lock));
2496 ASSERT(pool_pset_enabled());
2497 ASSERT(cp != NULL);
2498 ASSERT(cpu_is_active(cp));
2499
2500 cpun = cp->cpu_id;
2501 if (zoneid != ALL_ZONES && zoneid != GLOBAL_ZONEID) {
2502 ASSERT(zone->zone_ncpus_online != 0);
2503 zone->zone_ncpus_online--;
2504 }
2505
2506 if ((ksp = kstat_hold_byname("cpu", cpun, "intrstat", ALL_ZONES)) !=
2507 NULL) {
2508 kstat_zone_remove(ksp, zoneid);
2509 kstat_rele(ksp);
2510 }
2511 if ((ksp = kstat_hold_byname("cpu", cpun, "vm", ALL_ZONES)) != NULL) {
2512 kstat_zone_remove(ksp, zoneid);
2513 kstat_rele(ksp);
2514 }
2515 if ((ksp = kstat_hold_byname("cpu", cpun, "sys", ALL_ZONES)) != NULL) {
2516 kstat_zone_remove(ksp, zoneid);
2517 kstat_rele(ksp);
2518 }
2519 (void) snprintf(name, sizeof (name), "cpu_stat%d", cpun);
2520 if ((ksp = kstat_hold_byname("cpu_stat", cpun, name, ALL_ZONES))
2521 != NULL) {
2522 kstat_zone_remove(ksp, zoneid);
2523 kstat_rele(ksp);
2524 }
2525 }
2526
2527 /*
2528 * Update relevant kstats such that cpu is no longer visible to processes
2529 * executing in specified zone.
2530 */
2531 void
2532 cpu_visibility_remove(cpu_t *cp, zone_t *zone)
2533 {
2534 if (cpu_is_active(cp))
2535 cpu_visibility_offline(cp, zone);
2536 cpu_visibility_unconfigure(cp, zone);
2537 }
2538
2539 /*
2540 * Bind a thread to a CPU as requested.
2541 */
2542 int
2543 cpu_bind_thread(kthread_id_t tp, processorid_t bind, processorid_t *obind,
2544 int *error)
2545 {
2546 processorid_t binding;
2547 cpu_t *cp = NULL;
2548
2549 ASSERT(MUTEX_HELD(&cpu_lock));
2550 ASSERT(MUTEX_HELD(&ttoproc(tp)->p_lock));
2551
2552 thread_lock(tp);
2553
2554 /*
2555 * Record old binding, but change the obind, which was initialized
2556 * to PBIND_NONE, only if this thread has a binding. This avoids
2557 * reporting PBIND_NONE for a process when some LWPs are bound.
2558 */
2559 binding = tp->t_bind_cpu;
2560 if (binding != PBIND_NONE)
2561 *obind = binding; /* record old binding */
2562
2563 switch (bind) {
2564 case PBIND_QUERY:
2565 /* Just return the old binding */
2566 thread_unlock(tp);
2567 return (0);
2568
2569 case PBIND_QUERY_TYPE:
2570 /* Return the binding type */
2571 *obind = TB_CPU_IS_SOFT(tp) ? PBIND_SOFT : PBIND_HARD;
2572 thread_unlock(tp);
2573 return (0);
2574
2575 case PBIND_SOFT:
2576 /*
2577 * Set soft binding for this thread and return the actual
2578 * binding
2579 */
2580 TB_CPU_SOFT_SET(tp);
2581 thread_unlock(tp);
2582 return (0);
2583
2584 case PBIND_HARD:
2585 /*
2586 * Set hard binding for this thread and return the actual
2587 * binding
2588 */
2589 TB_CPU_HARD_SET(tp);
2590 thread_unlock(tp);
2591 return (0);
2592
2593 default:
2594 break;
2595 }
2596
2597 /*
2598 * If this thread/LWP cannot be bound because of permission
2599 * problems, just note that and return success so that the
2600 * other threads/LWPs will be bound. This is the way
2601 * processor_bind() is defined to work.
2602 *
2603 * Binding will get EPERM if the thread is of system class
2604 * or hasprocperm() fails.
2605 */
2606 if (tp->t_cid == 0 || !hasprocperm(tp->t_cred, CRED())) {
2607 *error = EPERM;
2608 thread_unlock(tp);
2609 return (0);
2610 }
2611
2612 binding = bind;
2613 if (binding != PBIND_NONE) {
2614 cp = cpu_get((processorid_t)binding);
2615 /*
2616 * Make sure binding is valid and is in right partition.
2617 */
2618 if (cp == NULL || tp->t_cpupart != cp->cpu_part) {
2619 *error = EINVAL;
2620 thread_unlock(tp);
2621 return (0);
2622 }
2623 }
2624 tp->t_bind_cpu = binding; /* set new binding */
2625
2626 /*
2627 * If there is no system-set reason for affinity, set
2628 * the t_bound_cpu field to reflect the binding.
2629 */
2630 if (tp->t_affinitycnt == 0) {
2631 if (binding == PBIND_NONE) {
2632 /*
2633 * We may need to adjust disp_max_unbound_pri
2634 * since we're becoming unbound.
2635 */
2636 disp_adjust_unbound_pri(tp);
2637
2638 tp->t_bound_cpu = NULL; /* set new binding */
2639
2640 /*
2641 * Move thread to lgroup with strongest affinity
2642 * after unbinding
2643 */
2644 if (tp->t_lgrp_affinity)
2645 lgrp_move_thread(tp,
2646 lgrp_choose(tp, tp->t_cpupart), 1);
2647
2648 if (tp->t_state == TS_ONPROC &&
2649 tp->t_cpu->cpu_part != tp->t_cpupart)
2650 cpu_surrender(tp);
2651 } else {
2652 lpl_t *lpl;
2653
2654 tp->t_bound_cpu = cp;
2655 ASSERT(cp->cpu_lpl != NULL);
2656
2657 /*
2658 * Set home to lgroup with most affinity containing CPU
2659 * that thread is being bound or minimum bounding
2660 * lgroup if no affinities set
2661 */
2662 if (tp->t_lgrp_affinity)
2663 lpl = lgrp_affinity_best(tp, tp->t_cpupart,
2664 LGRP_NONE, B_FALSE);
2665 else
2666 lpl = cp->cpu_lpl;
2667
2668 if (tp->t_lpl != lpl) {
2669 /* can't grab cpu_lock */
2670 lgrp_move_thread(tp, lpl, 1);
2671 }
2672
2673 /*
2674 * Make the thread switch to the bound CPU.
2675 * If the thread is runnable, we need to
2676 * requeue it even if t_cpu is already set
2677 * to the right CPU, since it may be on a
2678 * kpreempt queue and need to move to a local
2679 * queue. We could check t_disp_queue to
2680 * avoid unnecessary overhead if it's already
2681 * on the right queue, but since this isn't
2682 * a performance-critical operation it doesn't
2683 * seem worth the extra code and complexity.
2684 *
2685 * If the thread is weakbound to the cpu then it will
2686 * resist the new binding request until the weak
2687 * binding drops. The cpu_surrender or requeueing
2688 * below could be skipped in such cases (since it
2689 * will have no effect), but that would require
2690 * thread_allowmigrate to acquire thread_lock so
2691 * we'll take the very occasional hit here instead.
2692 */
2693 if (tp->t_state == TS_ONPROC) {
2694 cpu_surrender(tp);
2695 } else if (tp->t_state == TS_RUN) {
2696 cpu_t *ocp = tp->t_cpu;
2697
2698 (void) dispdeq(tp);
2699 setbackdq(tp);
2700 /*
2701 * Either on the bound CPU's disp queue now,
2702 * or swapped out or on the swap queue.
2703 */
2704 ASSERT(tp->t_disp_queue == cp->cpu_disp ||
2705 tp->t_weakbound_cpu == ocp ||
2706 (tp->t_schedflag & (TS_LOAD | TS_ON_SWAPQ))
2707 != TS_LOAD);
2708 }
2709 }
2710 }
2711
2712 /*
2713 * Our binding has changed; set TP_CHANGEBIND.
2714 */
2715 tp->t_proc_flag |= TP_CHANGEBIND;
2716 aston(tp);
2717
2718 thread_unlock(tp);
2719
2720 return (0);
2721 }
2722
2723 #if CPUSET_WORDS > 1
2724
2725 /*
2726 * Functions for implementing cpuset operations when a cpuset is more
2727 * than one word. On platforms where a cpuset is a single word these
2728 * are implemented as macros in cpuvar.h.
2729 */
2730
2731 void
2732 cpuset_all(cpuset_t *s)
2733 {
2734 int i;
2735
2736 for (i = 0; i < CPUSET_WORDS; i++)
2737 s->cpub[i] = ~0UL;
2738 }
2739
2740 void
2741 cpuset_all_but(cpuset_t *s, uint_t cpu)
2742 {
2743 cpuset_all(s);
2744 CPUSET_DEL(*s, cpu);
2745 }
2746
2747 void
2748 cpuset_only(cpuset_t *s, uint_t cpu)
2749 {
2750 CPUSET_ZERO(*s);
2751 CPUSET_ADD(*s, cpu);
2752 }
2753
2754 int
2755 cpuset_isnull(cpuset_t *s)
2756 {
2757 int i;
2758
2759 for (i = 0; i < CPUSET_WORDS; i++)
2760 if (s->cpub[i] != 0)
2761 return (0);
2762 return (1);
2763 }
2764
2765 int
2766 cpuset_cmp(cpuset_t *s1, cpuset_t *s2)
2767 {
2768 int i;
2769
2770 for (i = 0; i < CPUSET_WORDS; i++)
2771 if (s1->cpub[i] != s2->cpub[i])
2772 return (0);
2773 return (1);
2774 }
2775
2776 uint_t
2777 cpuset_find(cpuset_t *s)
2778 {
2779
2780 uint_t i;
2781 uint_t cpu = (uint_t)-1;
2782
2783 /*
2784 * Find a cpu in the cpuset
2785 */
2786 for (i = 0; i < CPUSET_WORDS; i++) {
2787 cpu = (uint_t)(lowbit(s->cpub[i]) - 1);
2788 if (cpu != (uint_t)-1) {
2789 cpu += i * BT_NBIPUL;
2790 break;
2791 }
2792 }
2793 return (cpu);
2794 }
2795
2796 void
2797 cpuset_bounds(cpuset_t *s, uint_t *smallestid, uint_t *largestid)
2798 {
2799 int i, j;
2800 uint_t bit;
2801
2802 /*
2803 * First, find the smallest cpu id in the set.
2804 */
2805 for (i = 0; i < CPUSET_WORDS; i++) {
2806 if (s->cpub[i] != 0) {
2807 bit = (uint_t)(lowbit(s->cpub[i]) - 1);
2808 ASSERT(bit != (uint_t)-1);
2809 *smallestid = bit + (i * BT_NBIPUL);
2810
2811 /*
2812 * Now find the largest cpu id in
2813 * the set and return immediately.
2814 * Done in an inner loop to avoid
2815 * having to break out of the first
2816 * loop.
2817 */
2818 for (j = CPUSET_WORDS - 1; j >= i; j--) {
2819 if (s->cpub[j] != 0) {
2820 bit = (uint_t)(highbit(s->cpub[j]) - 1);
2821 ASSERT(bit != (uint_t)-1);
2822 *largestid = bit + (j * BT_NBIPUL);
2823 ASSERT(*largestid >= *smallestid);
2824 return;
2825 }
2826 }
2827
2828 /*
2829 * If this code is reached, a
2830 * smallestid was found, but not a
2831 * largestid. The cpuset must have
2832 * been changed during the course
2833 * of this function call.
2834 */
2835 ASSERT(0);
2836 }
2837 }
2838 *smallestid = *largestid = CPUSET_NOTINSET;
2839 }
2840
2841 #endif /* CPUSET_WORDS */
2842
2843 /*
2844 * Unbind threads bound to specified CPU.
2845 *
2846 * If `unbind_all_threads' is true, unbind all user threads bound to a given
2847 * CPU. Otherwise unbind all soft-bound user threads.
2848 */
2849 int
2850 cpu_unbind(processorid_t cpu, boolean_t unbind_all_threads)
2851 {
2852 processorid_t obind;
2853 kthread_t *tp;
2854 int ret = 0;
2855 proc_t *pp;
2856 int err, berr = 0;
2857
2858 ASSERT(MUTEX_HELD(&cpu_lock));
2859
2860 mutex_enter(&pidlock);
2861 for (pp = practive; pp != NULL; pp = pp->p_next) {
2862 mutex_enter(&pp->p_lock);
2863 tp = pp->p_tlist;
2864 /*
2865 * Skip zombies, kernel processes, and processes in
2866 * other zones, if called from a non-global zone.
2867 */
2868 if (tp == NULL || (pp->p_flag & SSYS) ||
2869 !HASZONEACCESS(curproc, pp->p_zone->zone_id)) {
2870 mutex_exit(&pp->p_lock);
2871 continue;
2872 }
2873 do {
2874 if (tp->t_bind_cpu != cpu)
2875 continue;
2876 /*
2877 * Skip threads with hard binding when
2878 * `unbind_all_threads' is not specified.
2879 */
2880 if (!unbind_all_threads && TB_CPU_IS_HARD(tp))
2881 continue;
2882 err = cpu_bind_thread(tp, PBIND_NONE, &obind, &berr);
2883 if (ret == 0)
2884 ret = err;
2885 } while ((tp = tp->t_forw) != pp->p_tlist);
2886 mutex_exit(&pp->p_lock);
2887 }
2888 mutex_exit(&pidlock);
2889 if (ret == 0)
2890 ret = berr;
2891 return (ret);
2892 }
2893
2894
2895 /*
2896 * Destroy all remaining bound threads on a cpu.
2897 */
2898 void
2899 cpu_destroy_bound_threads(cpu_t *cp)
2900 {
2901 extern id_t syscid;
2902 register kthread_id_t t, tlist, tnext;
2903
2904 /*
2905 * Destroy all remaining bound threads on the cpu. This
2906 * should include both the interrupt threads and the idle thread.
2907 * This requires some care, since we need to traverse the
2908 * thread list with the pidlock mutex locked, but thread_free
2909 * also locks the pidlock mutex. So, we collect the threads
2910 * we're going to reap in a list headed by "tlist", then we
2911 * unlock the pidlock mutex and traverse the tlist list,
2912 * doing thread_free's on the thread's. Simple, n'est pas?
2913 * Also, this depends on thread_free not mucking with the
2914 * t_next and t_prev links of the thread.
2915 */
2916
2917 if ((t = curthread) != NULL) {
2918
2919 tlist = NULL;
2920 mutex_enter(&pidlock);
2921 do {
2922 tnext = t->t_next;
2923 if (t->t_bound_cpu == cp) {
2924
2925 /*
2926 * We've found a bound thread, carefully unlink
2927 * it out of the thread list, and add it to
2928 * our "tlist". We "know" we don't have to
2929 * worry about unlinking curthread (the thread
2930 * that is executing this code).
2931 */
2932 t->t_next->t_prev = t->t_prev;
2933 t->t_prev->t_next = t->t_next;
2934 t->t_next = tlist;
2935 tlist = t;
2936 ASSERT(t->t_cid == syscid);
2937 /* wake up anyone blocked in thread_join */
2938 cv_broadcast(&t->t_joincv);
2939 /*
2940 * t_lwp set by interrupt threads and not
2941 * cleared.
2942 */
2943 t->t_lwp = NULL;
2944 /*
2945 * Pause and idle threads always have
2946 * t_state set to TS_ONPROC.
2947 */
2948 t->t_state = TS_FREE;
2949 t->t_prev = NULL; /* Just in case */
2950 }
2951
2952 } while ((t = tnext) != curthread);
2953
2954 mutex_exit(&pidlock);
2955
2956 mutex_sync();
2957 for (t = tlist; t != NULL; t = tnext) {
2958 tnext = t->t_next;
2959 thread_free(t);
2960 }
2961 }
2962 }
2963
2964 /*
2965 * Update the cpu_supp_freqs of this cpu. This information is returned
2966 * as part of cpu_info kstats. If the cpu_info_kstat exists already, then
2967 * maintain the kstat data size.
2968 */
2969 void
2970 cpu_set_supp_freqs(cpu_t *cp, const char *freqs)
2971 {
2972 char clkstr[sizeof ("18446744073709551615") + 1]; /* ui64 MAX */
2973 const char *lfreqs = clkstr;
2974 boolean_t kstat_exists = B_FALSE;
2975 kstat_t *ksp;
2976 size_t len;
2977
2978 /*
2979 * A NULL pointer means we only support one speed.
2980 */
2981 if (freqs == NULL)
2982 (void) snprintf(clkstr, sizeof (clkstr), "%"PRIu64,
2983 cp->cpu_curr_clock);
2984 else
2985 lfreqs = freqs;
2986
2987 /*
2988 * Make sure the frequency doesn't change while a snapshot is
2989 * going on. Of course, we only need to worry about this if
2990 * the kstat exists.
2991 */
2992 if ((ksp = cp->cpu_info_kstat) != NULL) {
2993 mutex_enter(ksp->ks_lock);
2994 kstat_exists = B_TRUE;
2995 }
2996
2997 /*
2998 * Free any previously allocated string and if the kstat
2999 * already exists, then update its data size.
3000 */
3001 if (cp->cpu_supp_freqs != NULL) {
3002 len = strlen(cp->cpu_supp_freqs) + 1;
3003 kmem_free(cp->cpu_supp_freqs, len);
3004 if (kstat_exists)
3005 ksp->ks_data_size -= len;
3006 }
3007
3008 /*
3009 * Allocate the new string and set the pointer.
3010 */
3011 len = strlen(lfreqs) + 1;
3012 cp->cpu_supp_freqs = kmem_alloc(len, KM_SLEEP);
3013 (void) strcpy(cp->cpu_supp_freqs, lfreqs);
3014
3015 /*
3016 * If the kstat already exists then update the data size and
3017 * free the lock.
3018 */
3019 if (kstat_exists) {
3020 ksp->ks_data_size += len;
3021 mutex_exit(ksp->ks_lock);
3022 }
3023 }
3024
3025 /*
3026 * Indicate the current CPU's clock freqency (in Hz).
3027 * The calling context must be such that CPU references are safe.
3028 */
3029 void
3030 cpu_set_curr_clock(uint64_t new_clk)
3031 {
3032 uint64_t old_clk;
3033
3034 old_clk = CPU->cpu_curr_clock;
3035 CPU->cpu_curr_clock = new_clk;
3036
3037 /*
3038 * The cpu-change-speed DTrace probe exports the frequency in Hz
3039 */
3040 DTRACE_PROBE3(cpu__change__speed, processorid_t, CPU->cpu_id,
3041 uint64_t, old_clk, uint64_t, new_clk);
3042 }
3043
3044 /*
3045 * processor_info(2) and p_online(2) status support functions
3046 * The constants returned by the cpu_get_state() and cpu_get_state_str() are
3047 * for use in communicating processor state information to userland. Kernel
3048 * subsystems should only be using the cpu_flags value directly. Subsystems
3049 * modifying cpu_flags should record the state change via a call to the
3050 * cpu_set_state().
3051 */
3052
3053 /*
3054 * Update the pi_state of this CPU. This function provides the CPU status for
3055 * the information returned by processor_info(2).
3056 */
3057 void
3058 cpu_set_state(cpu_t *cpu)
3059 {
3060 ASSERT(MUTEX_HELD(&cpu_lock));
3061 cpu->cpu_type_info.pi_state = cpu_get_state(cpu);
3062 cpu->cpu_state_begin = gethrestime_sec();
3063 pool_cpu_mod = gethrtime();
3064 }
3065
3066 /*
3067 * Return offline/online/other status for the indicated CPU. Use only for
3068 * communication with user applications; cpu_flags provides the in-kernel
3069 * interface.
3070 */
3071 int
3072 cpu_get_state(cpu_t *cpu)
3073 {
3074 ASSERT(MUTEX_HELD(&cpu_lock));
3075 if (cpu->cpu_flags & CPU_POWEROFF)
3076 return (P_POWEROFF);
3077 else if (cpu->cpu_flags & CPU_FAULTED)
3078 return (P_FAULTED);
3079 else if (cpu->cpu_flags & CPU_SPARE)
3080 return (P_SPARE);
3081 else if ((cpu->cpu_flags & (CPU_READY | CPU_OFFLINE)) != CPU_READY)
3082 return (P_OFFLINE);
3083 else if (cpu->cpu_flags & CPU_ENABLE)
3084 return (P_ONLINE);
3085 else
3086 return (P_NOINTR);
3087 }
3088
3089 /*
3090 * Return processor_info(2) state as a string.
3091 */
3092 const char *
3093 cpu_get_state_str(cpu_t *cpu)
3094 {
3095 const char *string;
3096
3097 switch (cpu_get_state(cpu)) {
3098 case P_ONLINE:
3099 string = PS_ONLINE;
3100 break;
3101 case P_POWEROFF:
3102 string = PS_POWEROFF;
3103 break;
3104 case P_NOINTR:
3105 string = PS_NOINTR;
3106 break;
3107 case P_SPARE:
3108 string = PS_SPARE;
3109 break;
3110 case P_FAULTED:
3111 string = PS_FAULTED;
3112 break;
3113 case P_OFFLINE:
3114 string = PS_OFFLINE;
3115 break;
3116 default:
3117 string = "unknown";
3118 break;
3119 }
3120 return (string);
3121 }
3122
3123 /*
3124 * Export this CPU's statistics (cpu_stat_t and cpu_stats_t) as raw and named
3125 * kstats, respectively. This is done when a CPU is initialized or placed
3126 * online via p_online(2).
3127 */
3128 static void
3129 cpu_stats_kstat_create(cpu_t *cp)
3130 {
3131 int instance = cp->cpu_id;
3132 char *module = "cpu";
3133 char *class = "misc";
3134 kstat_t *ksp;
3135 zoneid_t zoneid;
3136
3137 ASSERT(MUTEX_HELD(&cpu_lock));
3138
3139 if (pool_pset_enabled())
3140 zoneid = GLOBAL_ZONEID;
3141 else
3142 zoneid = ALL_ZONES;
3143 /*
3144 * Create named kstats
3145 */
3146 #define CPU_STATS_KS_CREATE(name, tsize, update_func) \
3147 ksp = kstat_create_zone(module, instance, (name), class, \
3148 KSTAT_TYPE_NAMED, (tsize) / sizeof (kstat_named_t), 0, \
3149 zoneid); \
3150 if (ksp != NULL) { \
3151 ksp->ks_private = cp; \
3152 ksp->ks_update = (update_func); \
3153 kstat_install(ksp); \
3154 } else \
3155 cmn_err(CE_WARN, "cpu: unable to create %s:%d:%s kstat", \
3156 module, instance, (name));
3157
3158 CPU_STATS_KS_CREATE("sys", sizeof (cpu_sys_stats_ks_data_template),
3159 cpu_sys_stats_ks_update);
3160 CPU_STATS_KS_CREATE("vm", sizeof (cpu_vm_stats_ks_data_template),
3161 cpu_vm_stats_ks_update);
3162
3163 /*
3164 * Export the familiar cpu_stat_t KSTAT_TYPE_RAW kstat.
3165 */
3166 ksp = kstat_create_zone("cpu_stat", cp->cpu_id, NULL,
3167 "misc", KSTAT_TYPE_RAW, sizeof (cpu_stat_t), 0, zoneid);
3168 if (ksp != NULL) {
3169 ksp->ks_update = cpu_stat_ks_update;
3170 ksp->ks_private = cp;
3171 kstat_install(ksp);
3172 }
3173 }
3174
3175 static void
3176 cpu_stats_kstat_destroy(cpu_t *cp)
3177 {
3178 char ks_name[KSTAT_STRLEN];
3179
3180 (void) sprintf(ks_name, "cpu_stat%d", cp->cpu_id);
3181 kstat_delete_byname("cpu_stat", cp->cpu_id, ks_name);
3182
3183 kstat_delete_byname("cpu", cp->cpu_id, "sys");
3184 kstat_delete_byname("cpu", cp->cpu_id, "vm");
3185 }
3186
3187 static int
3188 cpu_sys_stats_ks_update(kstat_t *ksp, int rw)
3189 {
3190 cpu_t *cp = (cpu_t *)ksp->ks_private;
3191 struct cpu_sys_stats_ks_data *csskd;
3192 cpu_sys_stats_t *css;
3193 hrtime_t msnsecs[NCMSTATES];
3194 int i;
3195
3196 if (rw == KSTAT_WRITE)
3197 return (EACCES);
3198
3199 csskd = ksp->ks_data;
3200 css = &cp->cpu_stats.sys;
3201
3202 /*
3203 * Read CPU mstate, but compare with the last values we
3204 * received to make sure that the returned kstats never
3205 * decrease.
3206 */
3207
3208 get_cpu_mstate(cp, msnsecs);
3209 if (csskd->cpu_nsec_idle.value.ui64 > msnsecs[CMS_IDLE])
3210 msnsecs[CMS_IDLE] = csskd->cpu_nsec_idle.value.ui64;
3211 if (csskd->cpu_nsec_user.value.ui64 > msnsecs[CMS_USER])
3212 msnsecs[CMS_USER] = csskd->cpu_nsec_user.value.ui64;
3213 if (csskd->cpu_nsec_kernel.value.ui64 > msnsecs[CMS_SYSTEM])
3214 msnsecs[CMS_SYSTEM] = csskd->cpu_nsec_kernel.value.ui64;
3215
3216 bcopy(&cpu_sys_stats_ks_data_template, ksp->ks_data,
3217 sizeof (cpu_sys_stats_ks_data_template));
3218
3219 csskd->cpu_ticks_wait.value.ui64 = 0;
3220 csskd->wait_ticks_io.value.ui64 = 0;
3221
3222 csskd->cpu_nsec_idle.value.ui64 = msnsecs[CMS_IDLE];
3223 csskd->cpu_nsec_user.value.ui64 = msnsecs[CMS_USER];
3224 csskd->cpu_nsec_kernel.value.ui64 = msnsecs[CMS_SYSTEM];
3225 csskd->cpu_ticks_idle.value.ui64 =
3226 NSEC_TO_TICK(csskd->cpu_nsec_idle.value.ui64);
3227 csskd->cpu_ticks_user.value.ui64 =
3228 NSEC_TO_TICK(csskd->cpu_nsec_user.value.ui64);
3229 csskd->cpu_ticks_kernel.value.ui64 =
3230 NSEC_TO_TICK(csskd->cpu_nsec_kernel.value.ui64);
3231 csskd->cpu_nsec_dtrace.value.ui64 = cp->cpu_dtrace_nsec;
3232 csskd->dtrace_probes.value.ui64 = cp->cpu_dtrace_probes;
3233 csskd->cpu_nsec_intr.value.ui64 = cp->cpu_intrlast;
3234 csskd->cpu_load_intr.value.ui64 = cp->cpu_intrload;
3235 csskd->bread.value.ui64 = css->bread;
3236 csskd->bwrite.value.ui64 = css->bwrite;
3237 csskd->lread.value.ui64 = css->lread;
3238 csskd->lwrite.value.ui64 = css->lwrite;
3239 csskd->phread.value.ui64 = css->phread;
3240 csskd->phwrite.value.ui64 = css->phwrite;
3241 csskd->pswitch.value.ui64 = css->pswitch;
3242 csskd->trap.value.ui64 = css->trap;
3243 csskd->intr.value.ui64 = 0;
3244 for (i = 0; i < PIL_MAX; i++)
3245 csskd->intr.value.ui64 += css->intr[i];
3246 csskd->syscall.value.ui64 = css->syscall;
3247 csskd->sysread.value.ui64 = css->sysread;
3248 csskd->syswrite.value.ui64 = css->syswrite;
3249 csskd->sysfork.value.ui64 = css->sysfork;
3250 csskd->sysvfork.value.ui64 = css->sysvfork;
3251 csskd->sysexec.value.ui64 = css->sysexec;
3252 csskd->readch.value.ui64 = css->readch;
3253 csskd->writech.value.ui64 = css->writech;
3254 csskd->rcvint.value.ui64 = css->rcvint;
3255 csskd->xmtint.value.ui64 = css->xmtint;
3256 csskd->mdmint.value.ui64 = css->mdmint;
3257 csskd->rawch.value.ui64 = css->rawch;
3258 csskd->canch.value.ui64 = css->canch;
3259 csskd->outch.value.ui64 = css->outch;
3260 csskd->msg.value.ui64 = css->msg;
3261 csskd->sema.value.ui64 = css->sema;
3262 csskd->namei.value.ui64 = css->namei;
3263 csskd->ufsiget.value.ui64 = css->ufsiget;
3264 csskd->ufsdirblk.value.ui64 = css->ufsdirblk;
3265 csskd->ufsipage.value.ui64 = css->ufsipage;
3266 csskd->ufsinopage.value.ui64 = css->ufsinopage;
3267 csskd->procovf.value.ui64 = css->procovf;
3268 csskd->intrthread.value.ui64 = 0;
3269 for (i = 0; i < LOCK_LEVEL - 1; i++)
3270 csskd->intrthread.value.ui64 += css->intr[i];
3271 csskd->intrblk.value.ui64 = css->intrblk;
3272 csskd->intrunpin.value.ui64 = css->intrunpin;
3273 csskd->idlethread.value.ui64 = css->idlethread;
3274 csskd->inv_swtch.value.ui64 = css->inv_swtch;
3275 csskd->nthreads.value.ui64 = css->nthreads;
3276 csskd->cpumigrate.value.ui64 = css->cpumigrate;
3277 csskd->xcalls.value.ui64 = css->xcalls;
3278 csskd->mutex_adenters.value.ui64 = css->mutex_adenters;
3279 csskd->rw_rdfails.value.ui64 = css->rw_rdfails;
3280 csskd->rw_wrfails.value.ui64 = css->rw_wrfails;
3281 csskd->modload.value.ui64 = css->modload;
3282 csskd->modunload.value.ui64 = css->modunload;
3283 csskd->bawrite.value.ui64 = css->bawrite;
3284 csskd->iowait.value.ui64 = css->iowait;
3285
3286 return (0);
3287 }
3288
3289 static int
3290 cpu_vm_stats_ks_update(kstat_t *ksp, int rw)
3291 {
3292 cpu_t *cp = (cpu_t *)ksp->ks_private;
3293 struct cpu_vm_stats_ks_data *cvskd;
3294 cpu_vm_stats_t *cvs;
3295
3296 if (rw == KSTAT_WRITE)
3297 return (EACCES);
3298
3299 cvs = &cp->cpu_stats.vm;
3300 cvskd = ksp->ks_data;
3301
3302 bcopy(&cpu_vm_stats_ks_data_template, ksp->ks_data,
3303 sizeof (cpu_vm_stats_ks_data_template));
3304 cvskd->pgrec.value.ui64 = cvs->pgrec;
3305 cvskd->pgfrec.value.ui64 = cvs->pgfrec;
3306 cvskd->pgin.value.ui64 = cvs->pgin;
3307 cvskd->pgpgin.value.ui64 = cvs->pgpgin;
3308 cvskd->pgout.value.ui64 = cvs->pgout;
3309 cvskd->pgpgout.value.ui64 = cvs->pgpgout;
3310 cvskd->swapin.value.ui64 = cvs->swapin;
3311 cvskd->pgswapin.value.ui64 = cvs->pgswapin;
3312 cvskd->swapout.value.ui64 = cvs->swapout;
3313 cvskd->pgswapout.value.ui64 = cvs->pgswapout;
3314 cvskd->zfod.value.ui64 = cvs->zfod;
3315 cvskd->dfree.value.ui64 = cvs->dfree;
3316 cvskd->scan.value.ui64 = cvs->scan;
3317 cvskd->rev.value.ui64 = cvs->rev;
3318 cvskd->hat_fault.value.ui64 = cvs->hat_fault;
3319 cvskd->as_fault.value.ui64 = cvs->as_fault;
3320 cvskd->maj_fault.value.ui64 = cvs->maj_fault;
3321 cvskd->cow_fault.value.ui64 = cvs->cow_fault;
3322 cvskd->prot_fault.value.ui64 = cvs->prot_fault;
3323 cvskd->softlock.value.ui64 = cvs->softlock;
3324 cvskd->kernel_asflt.value.ui64 = cvs->kernel_asflt;
3325 cvskd->pgrrun.value.ui64 = cvs->pgrrun;
3326 cvskd->execpgin.value.ui64 = cvs->execpgin;
3327 cvskd->execpgout.value.ui64 = cvs->execpgout;
3328 cvskd->execfree.value.ui64 = cvs->execfree;
3329 cvskd->anonpgin.value.ui64 = cvs->anonpgin;
3330 cvskd->anonpgout.value.ui64 = cvs->anonpgout;
3331 cvskd->anonfree.value.ui64 = cvs->anonfree;
3332 cvskd->fspgin.value.ui64 = cvs->fspgin;
3333 cvskd->fspgout.value.ui64 = cvs->fspgout;
3334 cvskd->fsfree.value.ui64 = cvs->fsfree;
3335
3336 return (0);
3337 }
3338
3339 static int
3340 cpu_stat_ks_update(kstat_t *ksp, int rw)
3341 {
3342 cpu_stat_t *cso;
3343 cpu_t *cp;
3344 int i;
3345 hrtime_t msnsecs[NCMSTATES];
3346
3347 cso = (cpu_stat_t *)ksp->ks_data;
3348 cp = (cpu_t *)ksp->ks_private;
3349
3350 if (rw == KSTAT_WRITE)
3351 return (EACCES);
3352
3353 /*
3354 * Read CPU mstate, but compare with the last values we
3355 * received to make sure that the returned kstats never
3356 * decrease.
3357 */
3358
3359 get_cpu_mstate(cp, msnsecs);
3360 msnsecs[CMS_IDLE] = NSEC_TO_TICK(msnsecs[CMS_IDLE]);
3361 msnsecs[CMS_USER] = NSEC_TO_TICK(msnsecs[CMS_USER]);
3362 msnsecs[CMS_SYSTEM] = NSEC_TO_TICK(msnsecs[CMS_SYSTEM]);
3363 if (cso->cpu_sysinfo.cpu[CPU_IDLE] < msnsecs[CMS_IDLE])
3364 cso->cpu_sysinfo.cpu[CPU_IDLE] = msnsecs[CMS_IDLE];
3365 if (cso->cpu_sysinfo.cpu[CPU_USER] < msnsecs[CMS_USER])
3366 cso->cpu_sysinfo.cpu[CPU_USER] = msnsecs[CMS_USER];
3367 if (cso->cpu_sysinfo.cpu[CPU_KERNEL] < msnsecs[CMS_SYSTEM])
3368 cso->cpu_sysinfo.cpu[CPU_KERNEL] = msnsecs[CMS_SYSTEM];
3369 cso->cpu_sysinfo.cpu[CPU_WAIT] = 0;
3370 cso->cpu_sysinfo.wait[W_IO] = 0;
3371 cso->cpu_sysinfo.wait[W_SWAP] = 0;
3372 cso->cpu_sysinfo.wait[W_PIO] = 0;
3373 cso->cpu_sysinfo.bread = CPU_STATS(cp, sys.bread);
3374 cso->cpu_sysinfo.bwrite = CPU_STATS(cp, sys.bwrite);
3375 cso->cpu_sysinfo.lread = CPU_STATS(cp, sys.lread);
3376 cso->cpu_sysinfo.lwrite = CPU_STATS(cp, sys.lwrite);
3377 cso->cpu_sysinfo.phread = CPU_STATS(cp, sys.phread);
3378 cso->cpu_sysinfo.phwrite = CPU_STATS(cp, sys.phwrite);
3379 cso->cpu_sysinfo.pswitch = CPU_STATS(cp, sys.pswitch);
3380 cso->cpu_sysinfo.trap = CPU_STATS(cp, sys.trap);
3381 cso->cpu_sysinfo.intr = 0;
3382 for (i = 0; i < PIL_MAX; i++)
3383 cso->cpu_sysinfo.intr += CPU_STATS(cp, sys.intr[i]);
3384 cso->cpu_sysinfo.syscall = CPU_STATS(cp, sys.syscall);
3385 cso->cpu_sysinfo.sysread = CPU_STATS(cp, sys.sysread);
3386 cso->cpu_sysinfo.syswrite = CPU_STATS(cp, sys.syswrite);
3387 cso->cpu_sysinfo.sysfork = CPU_STATS(cp, sys.sysfork);
3388 cso->cpu_sysinfo.sysvfork = CPU_STATS(cp, sys.sysvfork);
3389 cso->cpu_sysinfo.sysexec = CPU_STATS(cp, sys.sysexec);
3390 cso->cpu_sysinfo.readch = CPU_STATS(cp, sys.readch);
3391 cso->cpu_sysinfo.writech = CPU_STATS(cp, sys.writech);
3392 cso->cpu_sysinfo.rcvint = CPU_STATS(cp, sys.rcvint);
3393 cso->cpu_sysinfo.xmtint = CPU_STATS(cp, sys.xmtint);
3394 cso->cpu_sysinfo.mdmint = CPU_STATS(cp, sys.mdmint);
3395 cso->cpu_sysinfo.rawch = CPU_STATS(cp, sys.rawch);
3396 cso->cpu_sysinfo.canch = CPU_STATS(cp, sys.canch);
3397 cso->cpu_sysinfo.outch = CPU_STATS(cp, sys.outch);
3398 cso->cpu_sysinfo.msg = CPU_STATS(cp, sys.msg);
3399 cso->cpu_sysinfo.sema = CPU_STATS(cp, sys.sema);
3400 cso->cpu_sysinfo.namei = CPU_STATS(cp, sys.namei);
3401 cso->cpu_sysinfo.ufsiget = CPU_STATS(cp, sys.ufsiget);
3402 cso->cpu_sysinfo.ufsdirblk = CPU_STATS(cp, sys.ufsdirblk);
3403 cso->cpu_sysinfo.ufsipage = CPU_STATS(cp, sys.ufsipage);
3404 cso->cpu_sysinfo.ufsinopage = CPU_STATS(cp, sys.ufsinopage);
3405 cso->cpu_sysinfo.inodeovf = 0;
3406 cso->cpu_sysinfo.fileovf = 0;
3407 cso->cpu_sysinfo.procovf = CPU_STATS(cp, sys.procovf);
3408 cso->cpu_sysinfo.intrthread = 0;
3409 for (i = 0; i < LOCK_LEVEL - 1; i++)
3410 cso->cpu_sysinfo.intrthread += CPU_STATS(cp, sys.intr[i]);
3411 cso->cpu_sysinfo.intrblk = CPU_STATS(cp, sys.intrblk);
3412 cso->cpu_sysinfo.idlethread = CPU_STATS(cp, sys.idlethread);
3413 cso->cpu_sysinfo.inv_swtch = CPU_STATS(cp, sys.inv_swtch);
3414 cso->cpu_sysinfo.nthreads = CPU_STATS(cp, sys.nthreads);
3415 cso->cpu_sysinfo.cpumigrate = CPU_STATS(cp, sys.cpumigrate);
3416 cso->cpu_sysinfo.xcalls = CPU_STATS(cp, sys.xcalls);
3417 cso->cpu_sysinfo.mutex_adenters = CPU_STATS(cp, sys.mutex_adenters);
3418 cso->cpu_sysinfo.rw_rdfails = CPU_STATS(cp, sys.rw_rdfails);
3419 cso->cpu_sysinfo.rw_wrfails = CPU_STATS(cp, sys.rw_wrfails);
3420 cso->cpu_sysinfo.modload = CPU_STATS(cp, sys.modload);
3421 cso->cpu_sysinfo.modunload = CPU_STATS(cp, sys.modunload);
3422 cso->cpu_sysinfo.bawrite = CPU_STATS(cp, sys.bawrite);
3423 cso->cpu_sysinfo.rw_enters = 0;
3424 cso->cpu_sysinfo.win_uo_cnt = 0;
3425 cso->cpu_sysinfo.win_uu_cnt = 0;
3426 cso->cpu_sysinfo.win_so_cnt = 0;
3427 cso->cpu_sysinfo.win_su_cnt = 0;
3428 cso->cpu_sysinfo.win_suo_cnt = 0;
3429
3430 cso->cpu_syswait.iowait = CPU_STATS(cp, sys.iowait);
3431 cso->cpu_syswait.swap = 0;
3432 cso->cpu_syswait.physio = 0;
3433
3434 cso->cpu_vminfo.pgrec = CPU_STATS(cp, vm.pgrec);
3435 cso->cpu_vminfo.pgfrec = CPU_STATS(cp, vm.pgfrec);
3436 cso->cpu_vminfo.pgin = CPU_STATS(cp, vm.pgin);
3437 cso->cpu_vminfo.pgpgin = CPU_STATS(cp, vm.pgpgin);
3438 cso->cpu_vminfo.pgout = CPU_STATS(cp, vm.pgout);
3439 cso->cpu_vminfo.pgpgout = CPU_STATS(cp, vm.pgpgout);
3440 cso->cpu_vminfo.swapin = CPU_STATS(cp, vm.swapin);
3441 cso->cpu_vminfo.pgswapin = CPU_STATS(cp, vm.pgswapin);
3442 cso->cpu_vminfo.swapout = CPU_STATS(cp, vm.swapout);
3443 cso->cpu_vminfo.pgswapout = CPU_STATS(cp, vm.pgswapout);
3444 cso->cpu_vminfo.zfod = CPU_STATS(cp, vm.zfod);
3445 cso->cpu_vminfo.dfree = CPU_STATS(cp, vm.dfree);
3446 cso->cpu_vminfo.scan = CPU_STATS(cp, vm.scan);
3447 cso->cpu_vminfo.rev = CPU_STATS(cp, vm.rev);
3448 cso->cpu_vminfo.hat_fault = CPU_STATS(cp, vm.hat_fault);
3449 cso->cpu_vminfo.as_fault = CPU_STATS(cp, vm.as_fault);
3450 cso->cpu_vminfo.maj_fault = CPU_STATS(cp, vm.maj_fault);
3451 cso->cpu_vminfo.cow_fault = CPU_STATS(cp, vm.cow_fault);
3452 cso->cpu_vminfo.prot_fault = CPU_STATS(cp, vm.prot_fault);
3453 cso->cpu_vminfo.softlock = CPU_STATS(cp, vm.softlock);
3454 cso->cpu_vminfo.kernel_asflt = CPU_STATS(cp, vm.kernel_asflt);
3455 cso->cpu_vminfo.pgrrun = CPU_STATS(cp, vm.pgrrun);
3456 cso->cpu_vminfo.execpgin = CPU_STATS(cp, vm.execpgin);
3457 cso->cpu_vminfo.execpgout = CPU_STATS(cp, vm.execpgout);
3458 cso->cpu_vminfo.execfree = CPU_STATS(cp, vm.execfree);
3459 cso->cpu_vminfo.anonpgin = CPU_STATS(cp, vm.anonpgin);
3460 cso->cpu_vminfo.anonpgout = CPU_STATS(cp, vm.anonpgout);
3461 cso->cpu_vminfo.anonfree = CPU_STATS(cp, vm.anonfree);
3462 cso->cpu_vminfo.fspgin = CPU_STATS(cp, vm.fspgin);
3463 cso->cpu_vminfo.fspgout = CPU_STATS(cp, vm.fspgout);
3464 cso->cpu_vminfo.fsfree = CPU_STATS(cp, vm.fsfree);
3465
3466 return (0);
3467 }
--- EOF ---