Print this page
patch fix-compile3
patch remove-load-flag
patch remove-on-swapq-flag
Split |
Close |
Expand all |
Collapse all |
--- old/usr/src/uts/common/os/cpu.c
+++ new/usr/src/uts/common/os/cpu.c
1 1 /*
2 2 * CDDL HEADER START
3 3 *
4 4 * The contents of this file are subject to the terms of the
5 5 * Common Development and Distribution License (the "License").
6 6 * You may not use this file except in compliance with the License.
7 7 *
8 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 9 * or http://www.opensolaris.org/os/licensing.
10 10 * See the License for the specific language governing permissions
11 11 * and limitations under the License.
12 12 *
13 13 * When distributing Covered Code, include this CDDL HEADER in each
14 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 15 * If applicable, add the following below this CDDL HEADER, with the
16 16 * fields enclosed by brackets "[]" replaced with your own identifying
17 17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 18 *
19 19 * CDDL HEADER END
20 20 */
21 21 /*
22 22 * Copyright (c) 1991, 2010, Oracle and/or its affiliates. All rights reserved.
23 23 * Copyright (c) 2012 by Delphix. All rights reserved.
24 24 */
25 25
26 26 /*
27 27 * Architecture-independent CPU control functions.
28 28 */
29 29
30 30 #include <sys/types.h>
31 31 #include <sys/param.h>
32 32 #include <sys/var.h>
33 33 #include <sys/thread.h>
34 34 #include <sys/cpuvar.h>
35 35 #include <sys/cpu_event.h>
36 36 #include <sys/kstat.h>
37 37 #include <sys/uadmin.h>
38 38 #include <sys/systm.h>
39 39 #include <sys/errno.h>
40 40 #include <sys/cmn_err.h>
41 41 #include <sys/procset.h>
42 42 #include <sys/processor.h>
43 43 #include <sys/debug.h>
44 44 #include <sys/cpupart.h>
45 45 #include <sys/lgrp.h>
46 46 #include <sys/pset.h>
47 47 #include <sys/pghw.h>
48 48 #include <sys/kmem.h>
49 49 #include <sys/kmem_impl.h> /* to set per-cpu kmem_cache offset */
50 50 #include <sys/atomic.h>
51 51 #include <sys/callb.h>
52 52 #include <sys/vtrace.h>
53 53 #include <sys/cyclic.h>
54 54 #include <sys/bitmap.h>
55 55 #include <sys/nvpair.h>
56 56 #include <sys/pool_pset.h>
57 57 #include <sys/msacct.h>
58 58 #include <sys/time.h>
59 59 #include <sys/archsystm.h>
60 60 #include <sys/sdt.h>
61 61 #if defined(__x86) || defined(__amd64)
62 62 #include <sys/x86_archext.h>
63 63 #endif
64 64 #include <sys/callo.h>
65 65
66 66 extern int mp_cpu_start(cpu_t *);
67 67 extern int mp_cpu_stop(cpu_t *);
68 68 extern int mp_cpu_poweron(cpu_t *);
69 69 extern int mp_cpu_poweroff(cpu_t *);
70 70 extern int mp_cpu_configure(int);
71 71 extern int mp_cpu_unconfigure(int);
72 72 extern void mp_cpu_faulted_enter(cpu_t *);
73 73 extern void mp_cpu_faulted_exit(cpu_t *);
74 74
75 75 extern int cmp_cpu_to_chip(processorid_t cpuid);
76 76 #ifdef __sparcv9
77 77 extern char *cpu_fru_fmri(cpu_t *cp);
78 78 #endif
79 79
80 80 static void cpu_add_active_internal(cpu_t *cp);
81 81 static void cpu_remove_active(cpu_t *cp);
82 82 static void cpu_info_kstat_create(cpu_t *cp);
83 83 static void cpu_info_kstat_destroy(cpu_t *cp);
84 84 static void cpu_stats_kstat_create(cpu_t *cp);
85 85 static void cpu_stats_kstat_destroy(cpu_t *cp);
86 86
87 87 static int cpu_sys_stats_ks_update(kstat_t *ksp, int rw);
88 88 static int cpu_vm_stats_ks_update(kstat_t *ksp, int rw);
89 89 static int cpu_stat_ks_update(kstat_t *ksp, int rw);
90 90 static int cpu_state_change_hooks(int, cpu_setup_t, cpu_setup_t);
91 91
92 92 /*
93 93 * cpu_lock protects ncpus, ncpus_online, cpu_flag, cpu_list, cpu_active,
94 94 * max_cpu_seqid_ever, and dispatch queue reallocations. The lock ordering with
95 95 * respect to related locks is:
96 96 *
97 97 * cpu_lock --> thread_free_lock ---> p_lock ---> thread_lock()
98 98 *
99 99 * Warning: Certain sections of code do not use the cpu_lock when
100 100 * traversing the cpu_list (e.g. mutex_vector_enter(), clock()). Since
101 101 * all cpus are paused during modifications to this list, a solution
102 102 * to protect the list is too either disable kernel preemption while
103 103 * walking the list, *or* recheck the cpu_next pointer at each
104 104 * iteration in the loop. Note that in no cases can any cached
105 105 * copies of the cpu pointers be kept as they may become invalid.
106 106 */
107 107 kmutex_t cpu_lock;
108 108 cpu_t *cpu_list; /* list of all CPUs */
109 109 cpu_t *clock_cpu_list; /* used by clock to walk CPUs */
110 110 cpu_t *cpu_active; /* list of active CPUs */
111 111 static cpuset_t cpu_available; /* set of available CPUs */
112 112 cpuset_t cpu_seqid_inuse; /* which cpu_seqids are in use */
113 113
114 114 cpu_t **cpu_seq; /* ptrs to CPUs, indexed by seq_id */
115 115
116 116 /*
117 117 * max_ncpus keeps the max cpus the system can have. Initially
118 118 * it's NCPU, but since most archs scan the devtree for cpus
119 119 * fairly early on during boot, the real max can be known before
120 120 * ncpus is set (useful for early NCPU based allocations).
121 121 */
122 122 int max_ncpus = NCPU;
123 123 /*
124 124 * platforms that set max_ncpus to maxiumum number of cpus that can be
125 125 * dynamically added will set boot_max_ncpus to the number of cpus found
126 126 * at device tree scan time during boot.
127 127 */
128 128 int boot_max_ncpus = -1;
129 129 int boot_ncpus = -1;
130 130 /*
131 131 * Maximum possible CPU id. This can never be >= NCPU since NCPU is
132 132 * used to size arrays that are indexed by CPU id.
133 133 */
134 134 processorid_t max_cpuid = NCPU - 1;
135 135
136 136 /*
137 137 * Maximum cpu_seqid was given. This number can only grow and never shrink. It
138 138 * can be used to optimize NCPU loops to avoid going through CPUs which were
139 139 * never on-line.
140 140 */
141 141 processorid_t max_cpu_seqid_ever = 0;
142 142
143 143 int ncpus = 1;
144 144 int ncpus_online = 1;
145 145
146 146 /*
147 147 * CPU that we're trying to offline. Protected by cpu_lock.
148 148 */
149 149 cpu_t *cpu_inmotion;
150 150
151 151 /*
152 152 * Can be raised to suppress further weakbinding, which are instead
153 153 * satisfied by disabling preemption. Must be raised/lowered under cpu_lock,
154 154 * while individual thread weakbinding synchronization is done under thread
155 155 * lock.
156 156 */
157 157 int weakbindingbarrier;
158 158
159 159 /*
160 160 * Variables used in pause_cpus().
161 161 */
162 162 static volatile char safe_list[NCPU];
163 163
164 164 static struct _cpu_pause_info {
165 165 int cp_spl; /* spl saved in pause_cpus() */
166 166 volatile int cp_go; /* Go signal sent after all ready */
167 167 int cp_count; /* # of CPUs to pause */
168 168 ksema_t cp_sem; /* synch pause_cpus & cpu_pause */
169 169 kthread_id_t cp_paused;
170 170 } cpu_pause_info;
171 171
172 172 static kmutex_t pause_free_mutex;
173 173 static kcondvar_t pause_free_cv;
174 174
175 175 void *(*cpu_pause_func)(void *) = NULL;
176 176
177 177
178 178 static struct cpu_sys_stats_ks_data {
179 179 kstat_named_t cpu_ticks_idle;
180 180 kstat_named_t cpu_ticks_user;
181 181 kstat_named_t cpu_ticks_kernel;
182 182 kstat_named_t cpu_ticks_wait;
183 183 kstat_named_t cpu_nsec_idle;
184 184 kstat_named_t cpu_nsec_user;
185 185 kstat_named_t cpu_nsec_kernel;
186 186 kstat_named_t cpu_nsec_dtrace;
187 187 kstat_named_t cpu_nsec_intr;
188 188 kstat_named_t cpu_load_intr;
189 189 kstat_named_t wait_ticks_io;
190 190 kstat_named_t dtrace_probes;
191 191 kstat_named_t bread;
192 192 kstat_named_t bwrite;
193 193 kstat_named_t lread;
194 194 kstat_named_t lwrite;
195 195 kstat_named_t phread;
196 196 kstat_named_t phwrite;
197 197 kstat_named_t pswitch;
198 198 kstat_named_t trap;
199 199 kstat_named_t intr;
200 200 kstat_named_t syscall;
201 201 kstat_named_t sysread;
202 202 kstat_named_t syswrite;
203 203 kstat_named_t sysfork;
204 204 kstat_named_t sysvfork;
205 205 kstat_named_t sysexec;
206 206 kstat_named_t readch;
207 207 kstat_named_t writech;
208 208 kstat_named_t rcvint;
209 209 kstat_named_t xmtint;
210 210 kstat_named_t mdmint;
211 211 kstat_named_t rawch;
212 212 kstat_named_t canch;
213 213 kstat_named_t outch;
214 214 kstat_named_t msg;
215 215 kstat_named_t sema;
216 216 kstat_named_t namei;
217 217 kstat_named_t ufsiget;
218 218 kstat_named_t ufsdirblk;
219 219 kstat_named_t ufsipage;
220 220 kstat_named_t ufsinopage;
221 221 kstat_named_t procovf;
222 222 kstat_named_t intrthread;
223 223 kstat_named_t intrblk;
224 224 kstat_named_t intrunpin;
225 225 kstat_named_t idlethread;
226 226 kstat_named_t inv_swtch;
227 227 kstat_named_t nthreads;
228 228 kstat_named_t cpumigrate;
229 229 kstat_named_t xcalls;
230 230 kstat_named_t mutex_adenters;
231 231 kstat_named_t rw_rdfails;
232 232 kstat_named_t rw_wrfails;
233 233 kstat_named_t modload;
234 234 kstat_named_t modunload;
235 235 kstat_named_t bawrite;
236 236 kstat_named_t iowait;
237 237 } cpu_sys_stats_ks_data_template = {
238 238 { "cpu_ticks_idle", KSTAT_DATA_UINT64 },
239 239 { "cpu_ticks_user", KSTAT_DATA_UINT64 },
240 240 { "cpu_ticks_kernel", KSTAT_DATA_UINT64 },
241 241 { "cpu_ticks_wait", KSTAT_DATA_UINT64 },
242 242 { "cpu_nsec_idle", KSTAT_DATA_UINT64 },
243 243 { "cpu_nsec_user", KSTAT_DATA_UINT64 },
244 244 { "cpu_nsec_kernel", KSTAT_DATA_UINT64 },
245 245 { "cpu_nsec_dtrace", KSTAT_DATA_UINT64 },
246 246 { "cpu_nsec_intr", KSTAT_DATA_UINT64 },
247 247 { "cpu_load_intr", KSTAT_DATA_UINT64 },
248 248 { "wait_ticks_io", KSTAT_DATA_UINT64 },
249 249 { "dtrace_probes", KSTAT_DATA_UINT64 },
250 250 { "bread", KSTAT_DATA_UINT64 },
251 251 { "bwrite", KSTAT_DATA_UINT64 },
252 252 { "lread", KSTAT_DATA_UINT64 },
253 253 { "lwrite", KSTAT_DATA_UINT64 },
254 254 { "phread", KSTAT_DATA_UINT64 },
255 255 { "phwrite", KSTAT_DATA_UINT64 },
256 256 { "pswitch", KSTAT_DATA_UINT64 },
257 257 { "trap", KSTAT_DATA_UINT64 },
258 258 { "intr", KSTAT_DATA_UINT64 },
259 259 { "syscall", KSTAT_DATA_UINT64 },
260 260 { "sysread", KSTAT_DATA_UINT64 },
261 261 { "syswrite", KSTAT_DATA_UINT64 },
262 262 { "sysfork", KSTAT_DATA_UINT64 },
263 263 { "sysvfork", KSTAT_DATA_UINT64 },
264 264 { "sysexec", KSTAT_DATA_UINT64 },
265 265 { "readch", KSTAT_DATA_UINT64 },
266 266 { "writech", KSTAT_DATA_UINT64 },
267 267 { "rcvint", KSTAT_DATA_UINT64 },
268 268 { "xmtint", KSTAT_DATA_UINT64 },
269 269 { "mdmint", KSTAT_DATA_UINT64 },
270 270 { "rawch", KSTAT_DATA_UINT64 },
271 271 { "canch", KSTAT_DATA_UINT64 },
272 272 { "outch", KSTAT_DATA_UINT64 },
273 273 { "msg", KSTAT_DATA_UINT64 },
274 274 { "sema", KSTAT_DATA_UINT64 },
275 275 { "namei", KSTAT_DATA_UINT64 },
276 276 { "ufsiget", KSTAT_DATA_UINT64 },
277 277 { "ufsdirblk", KSTAT_DATA_UINT64 },
278 278 { "ufsipage", KSTAT_DATA_UINT64 },
279 279 { "ufsinopage", KSTAT_DATA_UINT64 },
280 280 { "procovf", KSTAT_DATA_UINT64 },
281 281 { "intrthread", KSTAT_DATA_UINT64 },
282 282 { "intrblk", KSTAT_DATA_UINT64 },
283 283 { "intrunpin", KSTAT_DATA_UINT64 },
284 284 { "idlethread", KSTAT_DATA_UINT64 },
285 285 { "inv_swtch", KSTAT_DATA_UINT64 },
286 286 { "nthreads", KSTAT_DATA_UINT64 },
287 287 { "cpumigrate", KSTAT_DATA_UINT64 },
288 288 { "xcalls", KSTAT_DATA_UINT64 },
289 289 { "mutex_adenters", KSTAT_DATA_UINT64 },
290 290 { "rw_rdfails", KSTAT_DATA_UINT64 },
291 291 { "rw_wrfails", KSTAT_DATA_UINT64 },
292 292 { "modload", KSTAT_DATA_UINT64 },
293 293 { "modunload", KSTAT_DATA_UINT64 },
294 294 { "bawrite", KSTAT_DATA_UINT64 },
↓ open down ↓ |
294 lines elided |
↑ open up ↑ |
295 295 { "iowait", KSTAT_DATA_UINT64 },
296 296 };
297 297
298 298 static struct cpu_vm_stats_ks_data {
299 299 kstat_named_t pgrec;
300 300 kstat_named_t pgfrec;
301 301 kstat_named_t pgin;
302 302 kstat_named_t pgpgin;
303 303 kstat_named_t pgout;
304 304 kstat_named_t pgpgout;
305 - kstat_named_t swapin;
306 - kstat_named_t pgswapin;
307 - kstat_named_t swapout;
308 - kstat_named_t pgswapout;
309 305 kstat_named_t zfod;
310 306 kstat_named_t dfree;
311 307 kstat_named_t scan;
312 308 kstat_named_t rev;
313 309 kstat_named_t hat_fault;
314 310 kstat_named_t as_fault;
315 311 kstat_named_t maj_fault;
316 312 kstat_named_t cow_fault;
317 313 kstat_named_t prot_fault;
318 314 kstat_named_t softlock;
319 315 kstat_named_t kernel_asflt;
320 316 kstat_named_t pgrrun;
321 317 kstat_named_t execpgin;
322 318 kstat_named_t execpgout;
323 319 kstat_named_t execfree;
324 320 kstat_named_t anonpgin;
325 321 kstat_named_t anonpgout;
326 322 kstat_named_t anonfree;
↓ open down ↓ |
8 lines elided |
↑ open up ↑ |
327 323 kstat_named_t fspgin;
328 324 kstat_named_t fspgout;
329 325 kstat_named_t fsfree;
330 326 } cpu_vm_stats_ks_data_template = {
331 327 { "pgrec", KSTAT_DATA_UINT64 },
332 328 { "pgfrec", KSTAT_DATA_UINT64 },
333 329 { "pgin", KSTAT_DATA_UINT64 },
334 330 { "pgpgin", KSTAT_DATA_UINT64 },
335 331 { "pgout", KSTAT_DATA_UINT64 },
336 332 { "pgpgout", KSTAT_DATA_UINT64 },
337 - { "swapin", KSTAT_DATA_UINT64 },
338 - { "pgswapin", KSTAT_DATA_UINT64 },
339 - { "swapout", KSTAT_DATA_UINT64 },
340 - { "pgswapout", KSTAT_DATA_UINT64 },
341 333 { "zfod", KSTAT_DATA_UINT64 },
342 334 { "dfree", KSTAT_DATA_UINT64 },
343 335 { "scan", KSTAT_DATA_UINT64 },
344 336 { "rev", KSTAT_DATA_UINT64 },
345 337 { "hat_fault", KSTAT_DATA_UINT64 },
346 338 { "as_fault", KSTAT_DATA_UINT64 },
347 339 { "maj_fault", KSTAT_DATA_UINT64 },
348 340 { "cow_fault", KSTAT_DATA_UINT64 },
349 341 { "prot_fault", KSTAT_DATA_UINT64 },
350 342 { "softlock", KSTAT_DATA_UINT64 },
351 343 { "kernel_asflt", KSTAT_DATA_UINT64 },
352 344 { "pgrrun", KSTAT_DATA_UINT64 },
353 345 { "execpgin", KSTAT_DATA_UINT64 },
354 346 { "execpgout", KSTAT_DATA_UINT64 },
355 347 { "execfree", KSTAT_DATA_UINT64 },
356 348 { "anonpgin", KSTAT_DATA_UINT64 },
357 349 { "anonpgout", KSTAT_DATA_UINT64 },
358 350 { "anonfree", KSTAT_DATA_UINT64 },
359 351 { "fspgin", KSTAT_DATA_UINT64 },
360 352 { "fspgout", KSTAT_DATA_UINT64 },
361 353 { "fsfree", KSTAT_DATA_UINT64 },
362 354 };
363 355
364 356 /*
365 357 * Force the specified thread to migrate to the appropriate processor.
366 358 * Called with thread lock held, returns with it dropped.
367 359 */
368 360 static void
369 361 force_thread_migrate(kthread_id_t tp)
370 362 {
371 363 ASSERT(THREAD_LOCK_HELD(tp));
372 364 if (tp == curthread) {
373 365 THREAD_TRANSITION(tp);
374 366 CL_SETRUN(tp);
375 367 thread_unlock_nopreempt(tp);
376 368 swtch();
377 369 } else {
378 370 if (tp->t_state == TS_ONPROC) {
379 371 cpu_surrender(tp);
380 372 } else if (tp->t_state == TS_RUN) {
381 373 (void) dispdeq(tp);
382 374 setbackdq(tp);
383 375 }
384 376 thread_unlock(tp);
385 377 }
386 378 }
387 379
388 380 /*
389 381 * Set affinity for a specified CPU.
390 382 * A reference count is incremented and the affinity is held until the
391 383 * reference count is decremented to zero by thread_affinity_clear().
392 384 * This is so regions of code requiring affinity can be nested.
393 385 * Caller needs to ensure that cpu_id remains valid, which can be
394 386 * done by holding cpu_lock across this call, unless the caller
395 387 * specifies CPU_CURRENT in which case the cpu_lock will be acquired
396 388 * by thread_affinity_set and CPU->cpu_id will be the target CPU.
397 389 */
398 390 void
399 391 thread_affinity_set(kthread_id_t t, int cpu_id)
400 392 {
401 393 cpu_t *cp;
402 394 int c;
403 395
404 396 ASSERT(!(t == curthread && t->t_weakbound_cpu != NULL));
405 397
406 398 if ((c = cpu_id) == CPU_CURRENT) {
407 399 mutex_enter(&cpu_lock);
408 400 cpu_id = CPU->cpu_id;
409 401 }
410 402 /*
411 403 * We should be asserting that cpu_lock is held here, but
412 404 * the NCA code doesn't acquire it. The following assert
413 405 * should be uncommented when the NCA code is fixed.
414 406 *
415 407 * ASSERT(MUTEX_HELD(&cpu_lock));
416 408 */
417 409 ASSERT((cpu_id >= 0) && (cpu_id < NCPU));
418 410 cp = cpu[cpu_id];
419 411 ASSERT(cp != NULL); /* user must provide a good cpu_id */
420 412 /*
421 413 * If there is already a hard affinity requested, and this affinity
422 414 * conflicts with that, panic.
423 415 */
424 416 thread_lock(t);
425 417 if (t->t_affinitycnt > 0 && t->t_bound_cpu != cp) {
426 418 panic("affinity_set: setting %p but already bound to %p",
427 419 (void *)cp, (void *)t->t_bound_cpu);
428 420 }
429 421 t->t_affinitycnt++;
430 422 t->t_bound_cpu = cp;
431 423
432 424 /*
433 425 * Make sure we're running on the right CPU.
434 426 */
435 427 if (cp != t->t_cpu || t != curthread) {
436 428 force_thread_migrate(t); /* drops thread lock */
437 429 } else {
438 430 thread_unlock(t);
439 431 }
440 432
441 433 if (c == CPU_CURRENT)
442 434 mutex_exit(&cpu_lock);
443 435 }
444 436
445 437 /*
446 438 * Wrapper for backward compatibility.
447 439 */
448 440 void
449 441 affinity_set(int cpu_id)
450 442 {
451 443 thread_affinity_set(curthread, cpu_id);
452 444 }
453 445
454 446 /*
455 447 * Decrement the affinity reservation count and if it becomes zero,
456 448 * clear the CPU affinity for the current thread, or set it to the user's
457 449 * software binding request.
458 450 */
459 451 void
460 452 thread_affinity_clear(kthread_id_t t)
461 453 {
462 454 register processorid_t binding;
463 455
464 456 thread_lock(t);
465 457 if (--t->t_affinitycnt == 0) {
466 458 if ((binding = t->t_bind_cpu) == PBIND_NONE) {
467 459 /*
468 460 * Adjust disp_max_unbound_pri if necessary.
469 461 */
470 462 disp_adjust_unbound_pri(t);
471 463 t->t_bound_cpu = NULL;
472 464 if (t->t_cpu->cpu_part != t->t_cpupart) {
473 465 force_thread_migrate(t);
474 466 return;
475 467 }
476 468 } else {
477 469 t->t_bound_cpu = cpu[binding];
478 470 /*
479 471 * Make sure the thread is running on the bound CPU.
480 472 */
481 473 if (t->t_cpu != t->t_bound_cpu) {
482 474 force_thread_migrate(t);
483 475 return; /* already dropped lock */
484 476 }
485 477 }
486 478 }
487 479 thread_unlock(t);
488 480 }
489 481
490 482 /*
491 483 * Wrapper for backward compatibility.
492 484 */
493 485 void
494 486 affinity_clear(void)
495 487 {
496 488 thread_affinity_clear(curthread);
497 489 }
498 490
499 491 /*
500 492 * Weak cpu affinity. Bind to the "current" cpu for short periods
501 493 * of time during which the thread must not block (but may be preempted).
502 494 * Use this instead of kpreempt_disable() when it is only "no migration"
503 495 * rather than "no preemption" semantics that are required - disabling
504 496 * preemption holds higher priority threads off of cpu and if the
505 497 * operation that is protected is more than momentary this is not good
506 498 * for realtime etc.
507 499 *
508 500 * Weakly bound threads will not prevent a cpu from being offlined -
509 501 * we'll only run them on the cpu to which they are weakly bound but
510 502 * (because they do not block) we'll always be able to move them on to
511 503 * another cpu at offline time if we give them just a short moment to
512 504 * run during which they will unbind. To give a cpu a chance of offlining,
513 505 * however, we require a barrier to weak bindings that may be raised for a
514 506 * given cpu (offline/move code may set this and then wait a short time for
515 507 * existing weak bindings to drop); the cpu_inmotion pointer is that barrier.
516 508 *
517 509 * There are few restrictions on the calling context of thread_nomigrate.
518 510 * The caller must not hold the thread lock. Calls may be nested.
519 511 *
520 512 * After weakbinding a thread must not perform actions that may block.
521 513 * In particular it must not call thread_affinity_set; calling that when
522 514 * already weakbound is nonsensical anyway.
523 515 *
524 516 * If curthread is prevented from migrating for other reasons
525 517 * (kernel preemption disabled; high pil; strongly bound; interrupt thread)
526 518 * then the weak binding will succeed even if this cpu is the target of an
527 519 * offline/move request.
528 520 */
529 521 void
530 522 thread_nomigrate(void)
531 523 {
532 524 cpu_t *cp;
533 525 kthread_id_t t = curthread;
534 526
535 527 again:
536 528 kpreempt_disable();
537 529 cp = CPU;
538 530
539 531 /*
540 532 * A highlevel interrupt must not modify t_nomigrate or
541 533 * t_weakbound_cpu of the thread it has interrupted. A lowlevel
542 534 * interrupt thread cannot migrate and we can avoid the
543 535 * thread_lock call below by short-circuiting here. In either
544 536 * case we can just return since no migration is possible and
545 537 * the condition will persist (ie, when we test for these again
546 538 * in thread_allowmigrate they can't have changed). Migration
547 539 * is also impossible if we're at or above DISP_LEVEL pil.
548 540 */
549 541 if (CPU_ON_INTR(cp) || t->t_flag & T_INTR_THREAD ||
550 542 getpil() >= DISP_LEVEL) {
551 543 kpreempt_enable();
552 544 return;
553 545 }
554 546
555 547 /*
556 548 * We must be consistent with existing weak bindings. Since we
557 549 * may be interrupted between the increment of t_nomigrate and
558 550 * the store to t_weakbound_cpu below we cannot assume that
559 551 * t_weakbound_cpu will be set if t_nomigrate is. Note that we
560 552 * cannot assert t_weakbound_cpu == t_bind_cpu since that is not
561 553 * always the case.
562 554 */
563 555 if (t->t_nomigrate && t->t_weakbound_cpu && t->t_weakbound_cpu != cp) {
564 556 if (!panicstr)
565 557 panic("thread_nomigrate: binding to %p but already "
566 558 "bound to %p", (void *)cp,
567 559 (void *)t->t_weakbound_cpu);
568 560 }
569 561
570 562 /*
571 563 * At this point we have preemption disabled and we don't yet hold
572 564 * the thread lock. So it's possible that somebody else could
573 565 * set t_bind_cpu here and not be able to force us across to the
574 566 * new cpu (since we have preemption disabled).
575 567 */
576 568 thread_lock(curthread);
577 569
578 570 /*
579 571 * If further weak bindings are being (temporarily) suppressed then
580 572 * we'll settle for disabling kernel preemption (which assures
581 573 * no migration provided the thread does not block which it is
582 574 * not allowed to if using thread_nomigrate). We must remember
583 575 * this disposition so we can take appropriate action in
584 576 * thread_allowmigrate. If this is a nested call and the
585 577 * thread is already weakbound then fall through as normal.
586 578 * We remember the decision to settle for kpreempt_disable through
587 579 * negative nesting counting in t_nomigrate. Once a thread has had one
588 580 * weakbinding request satisfied in this way any further (nested)
589 581 * requests will continue to be satisfied in the same way,
590 582 * even if weak bindings have recommenced.
591 583 */
592 584 if (t->t_nomigrate < 0 || weakbindingbarrier && t->t_nomigrate == 0) {
593 585 --t->t_nomigrate;
594 586 thread_unlock(curthread);
595 587 return; /* with kpreempt_disable still active */
596 588 }
597 589
598 590 /*
599 591 * We hold thread_lock so t_bind_cpu cannot change. We could,
600 592 * however, be running on a different cpu to which we are t_bound_cpu
601 593 * to (as explained above). If we grant the weak binding request
602 594 * in that case then the dispatcher must favour our weak binding
603 595 * over our strong (in which case, just as when preemption is
604 596 * disabled, we can continue to run on a cpu other than the one to
605 597 * which we are strongbound; the difference in this case is that
606 598 * this thread can be preempted and so can appear on the dispatch
607 599 * queues of a cpu other than the one it is strongbound to).
608 600 *
609 601 * If the cpu we are running on does not appear to be a current
610 602 * offline target (we check cpu_inmotion to determine this - since
611 603 * we don't hold cpu_lock we may not see a recent store to that,
612 604 * so it's possible that we at times can grant a weak binding to a
613 605 * cpu that is an offline target, but that one request will not
614 606 * prevent the offline from succeeding) then we will always grant
615 607 * the weak binding request. This includes the case above where
616 608 * we grant a weakbinding not commensurate with our strong binding.
617 609 *
618 610 * If our cpu does appear to be an offline target then we're inclined
619 611 * not to grant the weakbinding request just yet - we'd prefer to
620 612 * migrate to another cpu and grant the request there. The
621 613 * exceptions are those cases where going through preemption code
622 614 * will not result in us changing cpu:
623 615 *
624 616 * . interrupts have already bypassed this case (see above)
625 617 * . we are already weakbound to this cpu (dispatcher code will
626 618 * always return us to the weakbound cpu)
627 619 * . preemption was disabled even before we disabled it above
628 620 * . we are strongbound to this cpu (if we're strongbound to
629 621 * another and not yet running there the trip through the
630 622 * dispatcher will move us to the strongbound cpu and we
631 623 * will grant the weak binding there)
632 624 */
633 625 if (cp != cpu_inmotion || t->t_nomigrate > 0 || t->t_preempt > 1 ||
634 626 t->t_bound_cpu == cp) {
635 627 /*
636 628 * Don't be tempted to store to t_weakbound_cpu only on
637 629 * the first nested bind request - if we're interrupted
638 630 * after the increment of t_nomigrate and before the
639 631 * store to t_weakbound_cpu and the interrupt calls
640 632 * thread_nomigrate then the assertion in thread_allowmigrate
641 633 * would fail.
642 634 */
643 635 t->t_nomigrate++;
644 636 t->t_weakbound_cpu = cp;
645 637 membar_producer();
646 638 thread_unlock(curthread);
647 639 /*
648 640 * Now that we have dropped the thread_lock another thread
649 641 * can set our t_weakbound_cpu, and will try to migrate us
650 642 * to the strongbound cpu (which will not be prevented by
651 643 * preemption being disabled since we're about to enable
652 644 * preemption). We have granted the weakbinding to the current
653 645 * cpu, so again we are in the position that is is is possible
654 646 * that our weak and strong bindings differ. Again this
655 647 * is catered for by dispatcher code which will favour our
656 648 * weak binding.
657 649 */
658 650 kpreempt_enable();
659 651 } else {
660 652 /*
661 653 * Move to another cpu before granting the request by
662 654 * forcing this thread through preemption code. When we
663 655 * get to set{front,back}dq called from CL_PREEMPT()
664 656 * cpu_choose() will be used to select a cpu to queue
665 657 * us on - that will see cpu_inmotion and take
666 658 * steps to avoid returning us to this cpu.
667 659 */
668 660 cp->cpu_kprunrun = 1;
669 661 thread_unlock(curthread);
670 662 kpreempt_enable(); /* will call preempt() */
671 663 goto again;
672 664 }
673 665 }
674 666
675 667 void
676 668 thread_allowmigrate(void)
677 669 {
678 670 kthread_id_t t = curthread;
679 671
680 672 ASSERT(t->t_weakbound_cpu == CPU ||
681 673 (t->t_nomigrate < 0 && t->t_preempt > 0) ||
682 674 CPU_ON_INTR(CPU) || t->t_flag & T_INTR_THREAD ||
683 675 getpil() >= DISP_LEVEL);
684 676
685 677 if (CPU_ON_INTR(CPU) || (t->t_flag & T_INTR_THREAD) ||
686 678 getpil() >= DISP_LEVEL)
687 679 return;
688 680
689 681 if (t->t_nomigrate < 0) {
690 682 /*
691 683 * This thread was granted "weak binding" in the
692 684 * stronger form of kernel preemption disabling.
693 685 * Undo a level of nesting for both t_nomigrate
694 686 * and t_preempt.
695 687 */
696 688 ++t->t_nomigrate;
697 689 kpreempt_enable();
698 690 } else if (--t->t_nomigrate == 0) {
699 691 /*
700 692 * Time to drop the weak binding. We need to cater
701 693 * for the case where we're weakbound to a different
702 694 * cpu than that to which we're strongbound (a very
703 695 * temporary arrangement that must only persist until
704 696 * weak binding drops). We don't acquire thread_lock
705 697 * here so even as this code executes t_bound_cpu
706 698 * may be changing. So we disable preemption and
707 699 * a) in the case that t_bound_cpu changes while we
708 700 * have preemption disabled kprunrun will be set
709 701 * asynchronously, and b) if before disabling
710 702 * preemption we were already on a different cpu to
711 703 * our t_bound_cpu then we set kprunrun ourselves
712 704 * to force a trip through the dispatcher when
713 705 * preemption is enabled.
714 706 */
715 707 kpreempt_disable();
716 708 if (t->t_bound_cpu &&
717 709 t->t_weakbound_cpu != t->t_bound_cpu)
718 710 CPU->cpu_kprunrun = 1;
719 711 t->t_weakbound_cpu = NULL;
720 712 membar_producer();
721 713 kpreempt_enable();
722 714 }
723 715 }
724 716
725 717 /*
726 718 * weakbinding_stop can be used to temporarily cause weakbindings made
727 719 * with thread_nomigrate to be satisfied through the stronger action of
728 720 * kpreempt_disable. weakbinding_start recommences normal weakbinding.
729 721 */
730 722
731 723 void
732 724 weakbinding_stop(void)
733 725 {
734 726 ASSERT(MUTEX_HELD(&cpu_lock));
735 727 weakbindingbarrier = 1;
736 728 membar_producer(); /* make visible before subsequent thread_lock */
737 729 }
738 730
739 731 void
740 732 weakbinding_start(void)
741 733 {
742 734 ASSERT(MUTEX_HELD(&cpu_lock));
743 735 weakbindingbarrier = 0;
744 736 }
745 737
746 738 void
747 739 null_xcall(void)
748 740 {
749 741 }
750 742
751 743 /*
752 744 * This routine is called to place the CPUs in a safe place so that
753 745 * one of them can be taken off line or placed on line. What we are
754 746 * trying to do here is prevent a thread from traversing the list
755 747 * of active CPUs while we are changing it or from getting placed on
756 748 * the run queue of a CPU that has just gone off line. We do this by
757 749 * creating a thread with the highest possible prio for each CPU and
758 750 * having it call this routine. The advantage of this method is that
759 751 * we can eliminate all checks for CPU_ACTIVE in the disp routines.
760 752 * This makes disp faster at the expense of making p_online() slower
761 753 * which is a good trade off.
762 754 */
763 755 static void
764 756 cpu_pause(int index)
765 757 {
766 758 int s;
767 759 struct _cpu_pause_info *cpi = &cpu_pause_info;
768 760 volatile char *safe = &safe_list[index];
769 761 long lindex = index;
770 762
771 763 ASSERT((curthread->t_bound_cpu != NULL) || (*safe == PAUSE_DIE));
772 764
773 765 while (*safe != PAUSE_DIE) {
774 766 *safe = PAUSE_READY;
775 767 membar_enter(); /* make sure stores are flushed */
776 768 sema_v(&cpi->cp_sem); /* signal requesting thread */
777 769
778 770 /*
779 771 * Wait here until all pause threads are running. That
780 772 * indicates that it's safe to do the spl. Until
781 773 * cpu_pause_info.cp_go is set, we don't want to spl
782 774 * because that might block clock interrupts needed
783 775 * to preempt threads on other CPUs.
784 776 */
785 777 while (cpi->cp_go == 0)
786 778 ;
787 779 /*
788 780 * Even though we are at the highest disp prio, we need
789 781 * to block out all interrupts below LOCK_LEVEL so that
790 782 * an intr doesn't come in, wake up a thread, and call
791 783 * setbackdq/setfrontdq.
792 784 */
793 785 s = splhigh();
794 786 /*
795 787 * if cpu_pause_func() has been set then call it using
796 788 * index as the argument, currently only used by
797 789 * cpr_suspend_cpus(). This function is used as the
798 790 * code to execute on the "paused" cpu's when a machine
799 791 * comes out of a sleep state and CPU's were powered off.
800 792 * (could also be used for hotplugging CPU's).
801 793 */
802 794 if (cpu_pause_func != NULL)
803 795 (*cpu_pause_func)((void *)lindex);
804 796
805 797 mach_cpu_pause(safe);
806 798
807 799 splx(s);
808 800 /*
809 801 * Waiting is at an end. Switch out of cpu_pause
810 802 * loop and resume useful work.
811 803 */
812 804 swtch();
813 805 }
814 806
815 807 mutex_enter(&pause_free_mutex);
816 808 *safe = PAUSE_DEAD;
817 809 cv_broadcast(&pause_free_cv);
818 810 mutex_exit(&pause_free_mutex);
819 811 }
820 812
821 813 /*
822 814 * Allow the cpus to start running again.
823 815 */
824 816 void
825 817 start_cpus()
826 818 {
827 819 int i;
828 820
829 821 ASSERT(MUTEX_HELD(&cpu_lock));
830 822 ASSERT(cpu_pause_info.cp_paused);
831 823 cpu_pause_info.cp_paused = NULL;
832 824 for (i = 0; i < NCPU; i++)
833 825 safe_list[i] = PAUSE_IDLE;
834 826 membar_enter(); /* make sure stores are flushed */
835 827 affinity_clear();
836 828 splx(cpu_pause_info.cp_spl);
837 829 kpreempt_enable();
838 830 }
839 831
840 832 /*
841 833 * Allocate a pause thread for a CPU.
842 834 */
843 835 static void
844 836 cpu_pause_alloc(cpu_t *cp)
845 837 {
846 838 kthread_id_t t;
847 839 long cpun = cp->cpu_id;
848 840
849 841 /*
850 842 * Note, v.v_nglobpris will not change value as long as I hold
851 843 * cpu_lock.
852 844 */
853 845 t = thread_create(NULL, 0, cpu_pause, (void *)cpun,
854 846 0, &p0, TS_STOPPED, v.v_nglobpris - 1);
855 847 thread_lock(t);
856 848 t->t_bound_cpu = cp;
857 849 t->t_disp_queue = cp->cpu_disp;
858 850 t->t_affinitycnt = 1;
859 851 t->t_preempt = 1;
860 852 thread_unlock(t);
861 853 cp->cpu_pause_thread = t;
862 854 /*
863 855 * Registering a thread in the callback table is usually done
864 856 * in the initialization code of the thread. In this
865 857 * case, we do it right after thread creation because the
866 858 * thread itself may never run, and we need to register the
867 859 * fact that it is safe for cpr suspend.
868 860 */
869 861 CALLB_CPR_INIT_SAFE(t, "cpu_pause");
870 862 }
871 863
872 864 /*
873 865 * Free a pause thread for a CPU.
874 866 */
875 867 static void
876 868 cpu_pause_free(cpu_t *cp)
877 869 {
878 870 kthread_id_t t;
879 871 int cpun = cp->cpu_id;
880 872
881 873 ASSERT(MUTEX_HELD(&cpu_lock));
882 874 /*
883 875 * We have to get the thread and tell him to die.
884 876 */
885 877 if ((t = cp->cpu_pause_thread) == NULL) {
886 878 ASSERT(safe_list[cpun] == PAUSE_IDLE);
887 879 return;
888 880 }
889 881 thread_lock(t);
890 882 t->t_cpu = CPU; /* disp gets upset if last cpu is quiesced. */
891 883 t->t_bound_cpu = NULL; /* Must un-bind; cpu may not be running. */
892 884 t->t_pri = v.v_nglobpris - 1;
893 885 ASSERT(safe_list[cpun] == PAUSE_IDLE);
894 886 safe_list[cpun] = PAUSE_DIE;
895 887 THREAD_TRANSITION(t);
896 888 setbackdq(t);
897 889 thread_unlock_nopreempt(t);
898 890
899 891 /*
900 892 * If we don't wait for the thread to actually die, it may try to
901 893 * run on the wrong cpu as part of an actual call to pause_cpus().
902 894 */
903 895 mutex_enter(&pause_free_mutex);
904 896 while (safe_list[cpun] != PAUSE_DEAD) {
905 897 cv_wait(&pause_free_cv, &pause_free_mutex);
906 898 }
907 899 mutex_exit(&pause_free_mutex);
908 900 safe_list[cpun] = PAUSE_IDLE;
909 901
910 902 cp->cpu_pause_thread = NULL;
911 903 }
912 904
913 905 /*
914 906 * Initialize basic structures for pausing CPUs.
915 907 */
916 908 void
917 909 cpu_pause_init()
918 910 {
919 911 sema_init(&cpu_pause_info.cp_sem, 0, NULL, SEMA_DEFAULT, NULL);
920 912 /*
921 913 * Create initial CPU pause thread.
922 914 */
923 915 cpu_pause_alloc(CPU);
924 916 }
925 917
926 918 /*
927 919 * Start the threads used to pause another CPU.
928 920 */
929 921 static int
930 922 cpu_pause_start(processorid_t cpu_id)
931 923 {
932 924 int i;
933 925 int cpu_count = 0;
934 926
935 927 for (i = 0; i < NCPU; i++) {
936 928 cpu_t *cp;
937 929 kthread_id_t t;
938 930
939 931 cp = cpu[i];
940 932 if (!CPU_IN_SET(cpu_available, i) || (i == cpu_id)) {
941 933 safe_list[i] = PAUSE_WAIT;
942 934 continue;
943 935 }
944 936
945 937 /*
946 938 * Skip CPU if it is quiesced or not yet started.
947 939 */
948 940 if ((cp->cpu_flags & (CPU_QUIESCED | CPU_READY)) != CPU_READY) {
949 941 safe_list[i] = PAUSE_WAIT;
950 942 continue;
951 943 }
952 944
953 945 /*
954 946 * Start this CPU's pause thread.
955 947 */
956 948 t = cp->cpu_pause_thread;
957 949 thread_lock(t);
958 950 /*
959 951 * Reset the priority, since nglobpris may have
960 952 * changed since the thread was created, if someone
961 953 * has loaded the RT (or some other) scheduling
962 954 * class.
963 955 */
964 956 t->t_pri = v.v_nglobpris - 1;
965 957 THREAD_TRANSITION(t);
966 958 setbackdq(t);
967 959 thread_unlock_nopreempt(t);
968 960 ++cpu_count;
969 961 }
970 962 return (cpu_count);
971 963 }
972 964
973 965
974 966 /*
975 967 * Pause all of the CPUs except the one we are on by creating a high
976 968 * priority thread bound to those CPUs.
977 969 *
978 970 * Note that one must be extremely careful regarding code
979 971 * executed while CPUs are paused. Since a CPU may be paused
980 972 * while a thread scheduling on that CPU is holding an adaptive
981 973 * lock, code executed with CPUs paused must not acquire adaptive
982 974 * (or low-level spin) locks. Also, such code must not block,
983 975 * since the thread that is supposed to initiate the wakeup may
984 976 * never run.
985 977 *
986 978 * With a few exceptions, the restrictions on code executed with CPUs
987 979 * paused match those for code executed at high-level interrupt
988 980 * context.
989 981 */
990 982 void
991 983 pause_cpus(cpu_t *off_cp)
992 984 {
993 985 processorid_t cpu_id;
994 986 int i;
995 987 struct _cpu_pause_info *cpi = &cpu_pause_info;
996 988
997 989 ASSERT(MUTEX_HELD(&cpu_lock));
998 990 ASSERT(cpi->cp_paused == NULL);
999 991 cpi->cp_count = 0;
1000 992 cpi->cp_go = 0;
1001 993 for (i = 0; i < NCPU; i++)
1002 994 safe_list[i] = PAUSE_IDLE;
1003 995 kpreempt_disable();
1004 996
1005 997 /*
1006 998 * If running on the cpu that is going offline, get off it.
1007 999 * This is so that it won't be necessary to rechoose a CPU
1008 1000 * when done.
1009 1001 */
1010 1002 if (CPU == off_cp)
1011 1003 cpu_id = off_cp->cpu_next_part->cpu_id;
1012 1004 else
1013 1005 cpu_id = CPU->cpu_id;
1014 1006 affinity_set(cpu_id);
1015 1007
1016 1008 /*
1017 1009 * Start the pause threads and record how many were started
1018 1010 */
1019 1011 cpi->cp_count = cpu_pause_start(cpu_id);
1020 1012
1021 1013 /*
1022 1014 * Now wait for all CPUs to be running the pause thread.
1023 1015 */
1024 1016 while (cpi->cp_count > 0) {
1025 1017 /*
1026 1018 * Spin reading the count without grabbing the disp
1027 1019 * lock to make sure we don't prevent the pause
1028 1020 * threads from getting the lock.
1029 1021 */
1030 1022 while (sema_held(&cpi->cp_sem))
1031 1023 ;
1032 1024 if (sema_tryp(&cpi->cp_sem))
1033 1025 --cpi->cp_count;
1034 1026 }
1035 1027 cpi->cp_go = 1; /* all have reached cpu_pause */
1036 1028
1037 1029 /*
1038 1030 * Now wait for all CPUs to spl. (Transition from PAUSE_READY
1039 1031 * to PAUSE_WAIT.)
1040 1032 */
1041 1033 for (i = 0; i < NCPU; i++) {
1042 1034 while (safe_list[i] != PAUSE_WAIT)
1043 1035 ;
1044 1036 }
1045 1037 cpi->cp_spl = splhigh(); /* block dispatcher on this CPU */
1046 1038 cpi->cp_paused = curthread;
1047 1039 }
1048 1040
1049 1041 /*
1050 1042 * Check whether the current thread has CPUs paused
1051 1043 */
1052 1044 int
1053 1045 cpus_paused(void)
1054 1046 {
1055 1047 if (cpu_pause_info.cp_paused != NULL) {
1056 1048 ASSERT(cpu_pause_info.cp_paused == curthread);
1057 1049 return (1);
1058 1050 }
1059 1051 return (0);
1060 1052 }
1061 1053
1062 1054 static cpu_t *
1063 1055 cpu_get_all(processorid_t cpun)
1064 1056 {
1065 1057 ASSERT(MUTEX_HELD(&cpu_lock));
1066 1058
1067 1059 if (cpun >= NCPU || cpun < 0 || !CPU_IN_SET(cpu_available, cpun))
1068 1060 return (NULL);
1069 1061 return (cpu[cpun]);
1070 1062 }
1071 1063
1072 1064 /*
1073 1065 * Check whether cpun is a valid processor id and whether it should be
1074 1066 * visible from the current zone. If it is, return a pointer to the
1075 1067 * associated CPU structure.
1076 1068 */
1077 1069 cpu_t *
1078 1070 cpu_get(processorid_t cpun)
1079 1071 {
1080 1072 cpu_t *c;
1081 1073
1082 1074 ASSERT(MUTEX_HELD(&cpu_lock));
1083 1075 c = cpu_get_all(cpun);
1084 1076 if (c != NULL && !INGLOBALZONE(curproc) && pool_pset_enabled() &&
1085 1077 zone_pset_get(curproc->p_zone) != cpupart_query_cpu(c))
1086 1078 return (NULL);
1087 1079 return (c);
1088 1080 }
1089 1081
1090 1082 /*
1091 1083 * The following functions should be used to check CPU states in the kernel.
1092 1084 * They should be invoked with cpu_lock held. Kernel subsystems interested
1093 1085 * in CPU states should *not* use cpu_get_state() and various P_ONLINE/etc
1094 1086 * states. Those are for user-land (and system call) use only.
1095 1087 */
1096 1088
1097 1089 /*
1098 1090 * Determine whether the CPU is online and handling interrupts.
1099 1091 */
1100 1092 int
1101 1093 cpu_is_online(cpu_t *cpu)
1102 1094 {
1103 1095 ASSERT(MUTEX_HELD(&cpu_lock));
1104 1096 return (cpu_flagged_online(cpu->cpu_flags));
1105 1097 }
1106 1098
1107 1099 /*
1108 1100 * Determine whether the CPU is offline (this includes spare and faulted).
1109 1101 */
1110 1102 int
1111 1103 cpu_is_offline(cpu_t *cpu)
1112 1104 {
1113 1105 ASSERT(MUTEX_HELD(&cpu_lock));
1114 1106 return (cpu_flagged_offline(cpu->cpu_flags));
1115 1107 }
1116 1108
1117 1109 /*
1118 1110 * Determine whether the CPU is powered off.
1119 1111 */
1120 1112 int
1121 1113 cpu_is_poweredoff(cpu_t *cpu)
1122 1114 {
1123 1115 ASSERT(MUTEX_HELD(&cpu_lock));
1124 1116 return (cpu_flagged_poweredoff(cpu->cpu_flags));
1125 1117 }
1126 1118
1127 1119 /*
1128 1120 * Determine whether the CPU is handling interrupts.
1129 1121 */
1130 1122 int
1131 1123 cpu_is_nointr(cpu_t *cpu)
1132 1124 {
1133 1125 ASSERT(MUTEX_HELD(&cpu_lock));
1134 1126 return (cpu_flagged_nointr(cpu->cpu_flags));
1135 1127 }
1136 1128
1137 1129 /*
1138 1130 * Determine whether the CPU is active (scheduling threads).
1139 1131 */
1140 1132 int
1141 1133 cpu_is_active(cpu_t *cpu)
1142 1134 {
1143 1135 ASSERT(MUTEX_HELD(&cpu_lock));
1144 1136 return (cpu_flagged_active(cpu->cpu_flags));
1145 1137 }
1146 1138
1147 1139 /*
1148 1140 * Same as above, but these require cpu_flags instead of cpu_t pointers.
1149 1141 */
1150 1142 int
1151 1143 cpu_flagged_online(cpu_flag_t cpu_flags)
1152 1144 {
1153 1145 return (cpu_flagged_active(cpu_flags) &&
1154 1146 (cpu_flags & CPU_ENABLE));
1155 1147 }
1156 1148
1157 1149 int
1158 1150 cpu_flagged_offline(cpu_flag_t cpu_flags)
1159 1151 {
1160 1152 return (((cpu_flags & CPU_POWEROFF) == 0) &&
1161 1153 ((cpu_flags & (CPU_READY | CPU_OFFLINE)) != CPU_READY));
1162 1154 }
1163 1155
1164 1156 int
1165 1157 cpu_flagged_poweredoff(cpu_flag_t cpu_flags)
1166 1158 {
1167 1159 return ((cpu_flags & CPU_POWEROFF) == CPU_POWEROFF);
1168 1160 }
1169 1161
1170 1162 int
1171 1163 cpu_flagged_nointr(cpu_flag_t cpu_flags)
1172 1164 {
1173 1165 return (cpu_flagged_active(cpu_flags) &&
1174 1166 (cpu_flags & CPU_ENABLE) == 0);
1175 1167 }
1176 1168
1177 1169 int
1178 1170 cpu_flagged_active(cpu_flag_t cpu_flags)
1179 1171 {
1180 1172 return (((cpu_flags & (CPU_POWEROFF | CPU_FAULTED | CPU_SPARE)) == 0) &&
1181 1173 ((cpu_flags & (CPU_READY | CPU_OFFLINE)) == CPU_READY));
1182 1174 }
1183 1175
1184 1176 /*
1185 1177 * Bring the indicated CPU online.
1186 1178 */
1187 1179 int
1188 1180 cpu_online(cpu_t *cp)
1189 1181 {
1190 1182 int error = 0;
1191 1183
1192 1184 /*
1193 1185 * Handle on-line request.
1194 1186 * This code must put the new CPU on the active list before
1195 1187 * starting it because it will not be paused, and will start
1196 1188 * using the active list immediately. The real start occurs
1197 1189 * when the CPU_QUIESCED flag is turned off.
1198 1190 */
1199 1191
1200 1192 ASSERT(MUTEX_HELD(&cpu_lock));
1201 1193
1202 1194 /*
1203 1195 * Put all the cpus into a known safe place.
1204 1196 * No mutexes can be entered while CPUs are paused.
1205 1197 */
1206 1198 error = mp_cpu_start(cp); /* arch-dep hook */
1207 1199 if (error == 0) {
1208 1200 pg_cpupart_in(cp, cp->cpu_part);
1209 1201 pause_cpus(NULL);
1210 1202 cpu_add_active_internal(cp);
1211 1203 if (cp->cpu_flags & CPU_FAULTED) {
1212 1204 cp->cpu_flags &= ~CPU_FAULTED;
1213 1205 mp_cpu_faulted_exit(cp);
1214 1206 }
1215 1207 cp->cpu_flags &= ~(CPU_QUIESCED | CPU_OFFLINE | CPU_FROZEN |
1216 1208 CPU_SPARE);
1217 1209 CPU_NEW_GENERATION(cp);
1218 1210 start_cpus();
1219 1211 cpu_stats_kstat_create(cp);
1220 1212 cpu_create_intrstat(cp);
1221 1213 lgrp_kstat_create(cp);
1222 1214 cpu_state_change_notify(cp->cpu_id, CPU_ON);
1223 1215 cpu_intr_enable(cp); /* arch-dep hook */
1224 1216 cpu_state_change_notify(cp->cpu_id, CPU_INTR_ON);
1225 1217 cpu_set_state(cp);
1226 1218 cyclic_online(cp);
1227 1219 /*
1228 1220 * This has to be called only after cyclic_online(). This
1229 1221 * function uses cyclics.
1230 1222 */
1231 1223 callout_cpu_online(cp);
1232 1224 poke_cpu(cp->cpu_id);
1233 1225 }
1234 1226
1235 1227 return (error);
1236 1228 }
1237 1229
1238 1230 /*
1239 1231 * Take the indicated CPU offline.
1240 1232 */
1241 1233 int
1242 1234 cpu_offline(cpu_t *cp, int flags)
1243 1235 {
1244 1236 cpupart_t *pp;
1245 1237 int error = 0;
1246 1238 cpu_t *ncp;
1247 1239 int intr_enable;
1248 1240 int cyclic_off = 0;
1249 1241 int callout_off = 0;
1250 1242 int loop_count;
1251 1243 int no_quiesce = 0;
1252 1244 int (*bound_func)(struct cpu *, int);
1253 1245 kthread_t *t;
1254 1246 lpl_t *cpu_lpl;
1255 1247 proc_t *p;
1256 1248 int lgrp_diff_lpl;
1257 1249 boolean_t unbind_all_threads = (flags & CPU_FORCED) != 0;
1258 1250
1259 1251 ASSERT(MUTEX_HELD(&cpu_lock));
1260 1252
1261 1253 /*
1262 1254 * If we're going from faulted or spare to offline, just
1263 1255 * clear these flags and update CPU state.
1264 1256 */
1265 1257 if (cp->cpu_flags & (CPU_FAULTED | CPU_SPARE)) {
1266 1258 if (cp->cpu_flags & CPU_FAULTED) {
1267 1259 cp->cpu_flags &= ~CPU_FAULTED;
1268 1260 mp_cpu_faulted_exit(cp);
1269 1261 }
1270 1262 cp->cpu_flags &= ~CPU_SPARE;
1271 1263 cpu_set_state(cp);
1272 1264 return (0);
1273 1265 }
1274 1266
1275 1267 /*
1276 1268 * Handle off-line request.
1277 1269 */
1278 1270 pp = cp->cpu_part;
1279 1271 /*
1280 1272 * Don't offline last online CPU in partition
1281 1273 */
1282 1274 if (ncpus_online <= 1 || pp->cp_ncpus <= 1 || cpu_intr_count(cp) < 2)
1283 1275 return (EBUSY);
1284 1276 /*
1285 1277 * Unbind all soft-bound threads bound to our CPU and hard bound threads
1286 1278 * if we were asked to.
1287 1279 */
1288 1280 error = cpu_unbind(cp->cpu_id, unbind_all_threads);
1289 1281 if (error != 0)
1290 1282 return (error);
1291 1283 /*
1292 1284 * We shouldn't be bound to this CPU ourselves.
1293 1285 */
1294 1286 if (curthread->t_bound_cpu == cp)
1295 1287 return (EBUSY);
1296 1288
1297 1289 /*
1298 1290 * Tell interested parties that this CPU is going offline.
1299 1291 */
1300 1292 CPU_NEW_GENERATION(cp);
1301 1293 cpu_state_change_notify(cp->cpu_id, CPU_OFF);
1302 1294
1303 1295 /*
1304 1296 * Tell the PG subsystem that the CPU is leaving the partition
1305 1297 */
1306 1298 pg_cpupart_out(cp, pp);
1307 1299
1308 1300 /*
1309 1301 * Take the CPU out of interrupt participation so we won't find
1310 1302 * bound kernel threads. If the architecture cannot completely
1311 1303 * shut off interrupts on the CPU, don't quiesce it, but don't
1312 1304 * run anything but interrupt thread... this is indicated by
1313 1305 * the CPU_OFFLINE flag being on but the CPU_QUIESCE flag being
1314 1306 * off.
1315 1307 */
1316 1308 intr_enable = cp->cpu_flags & CPU_ENABLE;
1317 1309 if (intr_enable)
1318 1310 no_quiesce = cpu_intr_disable(cp);
1319 1311
1320 1312 /*
1321 1313 * Record that we are aiming to offline this cpu. This acts as
1322 1314 * a barrier to further weak binding requests in thread_nomigrate
1323 1315 * and also causes cpu_choose, disp_lowpri_cpu and setfrontdq to
1324 1316 * lean away from this cpu. Further strong bindings are already
1325 1317 * avoided since we hold cpu_lock. Since threads that are set
1326 1318 * runnable around now and others coming off the target cpu are
1327 1319 * directed away from the target, existing strong and weak bindings
1328 1320 * (especially the latter) to the target cpu stand maximum chance of
1329 1321 * being able to unbind during the short delay loop below (if other
1330 1322 * unbound threads compete they may not see cpu in time to unbind
1331 1323 * even if they would do so immediately.
1332 1324 */
1333 1325 cpu_inmotion = cp;
1334 1326 membar_enter();
1335 1327
1336 1328 /*
1337 1329 * Check for kernel threads (strong or weak) bound to that CPU.
1338 1330 * Strongly bound threads may not unbind, and we'll have to return
1339 1331 * EBUSY. Weakly bound threads should always disappear - we've
1340 1332 * stopped more weak binding with cpu_inmotion and existing
1341 1333 * bindings will drain imminently (they may not block). Nonetheless
1342 1334 * we will wait for a fixed period for all bound threads to disappear.
1343 1335 * Inactive interrupt threads are OK (they'll be in TS_FREE
1344 1336 * state). If test finds some bound threads, wait a few ticks
1345 1337 * to give short-lived threads (such as interrupts) chance to
1346 1338 * complete. Note that if no_quiesce is set, i.e. this cpu
1347 1339 * is required to service interrupts, then we take the route
1348 1340 * that permits interrupt threads to be active (or bypassed).
1349 1341 */
1350 1342 bound_func = no_quiesce ? disp_bound_threads : disp_bound_anythreads;
1351 1343
1352 1344 again: for (loop_count = 0; (*bound_func)(cp, 0); loop_count++) {
1353 1345 if (loop_count >= 5) {
1354 1346 error = EBUSY; /* some threads still bound */
1355 1347 break;
1356 1348 }
1357 1349
1358 1350 /*
1359 1351 * If some threads were assigned, give them
1360 1352 * a chance to complete or move.
1361 1353 *
1362 1354 * This assumes that the clock_thread is not bound
1363 1355 * to any CPU, because the clock_thread is needed to
1364 1356 * do the delay(hz/100).
1365 1357 *
1366 1358 * Note: we still hold the cpu_lock while waiting for
1367 1359 * the next clock tick. This is OK since it isn't
1368 1360 * needed for anything else except processor_bind(2),
1369 1361 * and system initialization. If we drop the lock,
1370 1362 * we would risk another p_online disabling the last
1371 1363 * processor.
1372 1364 */
1373 1365 delay(hz/100);
1374 1366 }
1375 1367
1376 1368 if (error == 0 && callout_off == 0) {
1377 1369 callout_cpu_offline(cp);
1378 1370 callout_off = 1;
1379 1371 }
1380 1372
1381 1373 if (error == 0 && cyclic_off == 0) {
1382 1374 if (!cyclic_offline(cp)) {
1383 1375 /*
1384 1376 * We must have bound cyclics...
1385 1377 */
1386 1378 error = EBUSY;
1387 1379 goto out;
1388 1380 }
1389 1381 cyclic_off = 1;
1390 1382 }
1391 1383
1392 1384 /*
1393 1385 * Call mp_cpu_stop() to perform any special operations
1394 1386 * needed for this machine architecture to offline a CPU.
1395 1387 */
1396 1388 if (error == 0)
1397 1389 error = mp_cpu_stop(cp); /* arch-dep hook */
1398 1390
1399 1391 /*
1400 1392 * If that all worked, take the CPU offline and decrement
1401 1393 * ncpus_online.
1402 1394 */
1403 1395 if (error == 0) {
1404 1396 /*
1405 1397 * Put all the cpus into a known safe place.
1406 1398 * No mutexes can be entered while CPUs are paused.
1407 1399 */
1408 1400 pause_cpus(cp);
1409 1401 /*
1410 1402 * Repeat the operation, if necessary, to make sure that
1411 1403 * all outstanding low-level interrupts run to completion
1412 1404 * before we set the CPU_QUIESCED flag. It's also possible
1413 1405 * that a thread has weak bound to the cpu despite our raising
1414 1406 * cpu_inmotion above since it may have loaded that
1415 1407 * value before the barrier became visible (this would have
1416 1408 * to be the thread that was on the target cpu at the time
1417 1409 * we raised the barrier).
1418 1410 */
1419 1411 if ((!no_quiesce && cp->cpu_intr_actv != 0) ||
1420 1412 (*bound_func)(cp, 1)) {
1421 1413 start_cpus();
1422 1414 (void) mp_cpu_start(cp);
1423 1415 goto again;
1424 1416 }
1425 1417 ncp = cp->cpu_next_part;
1426 1418 cpu_lpl = cp->cpu_lpl;
1427 1419 ASSERT(cpu_lpl != NULL);
1428 1420
1429 1421 /*
1430 1422 * Remove the CPU from the list of active CPUs.
1431 1423 */
1432 1424 cpu_remove_active(cp);
1433 1425
1434 1426 /*
1435 1427 * Walk the active process list and look for threads
1436 1428 * whose home lgroup needs to be updated, or
1437 1429 * the last CPU they run on is the one being offlined now.
1438 1430 */
1439 1431
1440 1432 ASSERT(curthread->t_cpu != cp);
1441 1433 for (p = practive; p != NULL; p = p->p_next) {
1442 1434
1443 1435 t = p->p_tlist;
1444 1436
1445 1437 if (t == NULL)
1446 1438 continue;
1447 1439
1448 1440 lgrp_diff_lpl = 0;
1449 1441
1450 1442 do {
1451 1443 ASSERT(t->t_lpl != NULL);
1452 1444 /*
1453 1445 * Taking last CPU in lpl offline
1454 1446 * Rehome thread if it is in this lpl
1455 1447 * Otherwise, update the count of how many
1456 1448 * threads are in this CPU's lgroup but have
1457 1449 * a different lpl.
1458 1450 */
1459 1451
1460 1452 if (cpu_lpl->lpl_ncpu == 0) {
1461 1453 if (t->t_lpl == cpu_lpl)
1462 1454 lgrp_move_thread(t,
1463 1455 lgrp_choose(t,
1464 1456 t->t_cpupart), 0);
1465 1457 else if (t->t_lpl->lpl_lgrpid ==
1466 1458 cpu_lpl->lpl_lgrpid)
1467 1459 lgrp_diff_lpl++;
1468 1460 }
1469 1461 ASSERT(t->t_lpl->lpl_ncpu > 0);
1470 1462
1471 1463 /*
1472 1464 * Update CPU last ran on if it was this CPU
1473 1465 */
1474 1466 if (t->t_cpu == cp && t->t_bound_cpu != cp)
1475 1467 t->t_cpu = disp_lowpri_cpu(ncp,
1476 1468 t->t_lpl, t->t_pri, NULL);
1477 1469 ASSERT(t->t_cpu != cp || t->t_bound_cpu == cp ||
1478 1470 t->t_weakbound_cpu == cp);
1479 1471
1480 1472 t = t->t_forw;
1481 1473 } while (t != p->p_tlist);
1482 1474
1483 1475 /*
1484 1476 * Didn't find any threads in the same lgroup as this
1485 1477 * CPU with a different lpl, so remove the lgroup from
1486 1478 * the process lgroup bitmask.
1487 1479 */
1488 1480
1489 1481 if (lgrp_diff_lpl == 0)
1490 1482 klgrpset_del(p->p_lgrpset, cpu_lpl->lpl_lgrpid);
1491 1483 }
1492 1484
1493 1485 /*
1494 1486 * Walk thread list looking for threads that need to be
1495 1487 * rehomed, since there are some threads that are not in
1496 1488 * their process's p_tlist.
1497 1489 */
1498 1490
1499 1491 t = curthread;
1500 1492 do {
1501 1493 ASSERT(t != NULL && t->t_lpl != NULL);
1502 1494
1503 1495 /*
1504 1496 * Rehome threads with same lpl as this CPU when this
1505 1497 * is the last CPU in the lpl.
1506 1498 */
1507 1499
1508 1500 if ((cpu_lpl->lpl_ncpu == 0) && (t->t_lpl == cpu_lpl))
1509 1501 lgrp_move_thread(t,
1510 1502 lgrp_choose(t, t->t_cpupart), 1);
1511 1503
1512 1504 ASSERT(t->t_lpl->lpl_ncpu > 0);
1513 1505
1514 1506 /*
1515 1507 * Update CPU last ran on if it was this CPU
1516 1508 */
1517 1509
1518 1510 if (t->t_cpu == cp && t->t_bound_cpu != cp) {
1519 1511 t->t_cpu = disp_lowpri_cpu(ncp,
1520 1512 t->t_lpl, t->t_pri, NULL);
1521 1513 }
1522 1514 ASSERT(t->t_cpu != cp || t->t_bound_cpu == cp ||
1523 1515 t->t_weakbound_cpu == cp);
1524 1516 t = t->t_next;
1525 1517
1526 1518 } while (t != curthread);
1527 1519 ASSERT((cp->cpu_flags & (CPU_FAULTED | CPU_SPARE)) == 0);
1528 1520 cp->cpu_flags |= CPU_OFFLINE;
1529 1521 disp_cpu_inactive(cp);
1530 1522 if (!no_quiesce)
1531 1523 cp->cpu_flags |= CPU_QUIESCED;
1532 1524 ncpus_online--;
1533 1525 cpu_set_state(cp);
1534 1526 cpu_inmotion = NULL;
1535 1527 start_cpus();
1536 1528 cpu_stats_kstat_destroy(cp);
1537 1529 cpu_delete_intrstat(cp);
1538 1530 lgrp_kstat_destroy(cp);
1539 1531 }
1540 1532
1541 1533 out:
1542 1534 cpu_inmotion = NULL;
1543 1535
1544 1536 /*
1545 1537 * If we failed, re-enable interrupts.
1546 1538 * Do this even if cpu_intr_disable returned an error, because
1547 1539 * it may have partially disabled interrupts.
1548 1540 */
1549 1541 if (error && intr_enable)
1550 1542 cpu_intr_enable(cp);
1551 1543
1552 1544 /*
1553 1545 * If we failed, but managed to offline the cyclic subsystem on this
1554 1546 * CPU, bring it back online.
1555 1547 */
1556 1548 if (error && cyclic_off)
1557 1549 cyclic_online(cp);
1558 1550
1559 1551 /*
1560 1552 * If we failed, but managed to offline callouts on this CPU,
1561 1553 * bring it back online.
1562 1554 */
1563 1555 if (error && callout_off)
1564 1556 callout_cpu_online(cp);
1565 1557
1566 1558 /*
1567 1559 * If we failed, tell the PG subsystem that the CPU is back
1568 1560 */
1569 1561 pg_cpupart_in(cp, pp);
1570 1562
1571 1563 /*
1572 1564 * If we failed, we need to notify everyone that this CPU is back on.
1573 1565 */
1574 1566 if (error != 0) {
1575 1567 CPU_NEW_GENERATION(cp);
1576 1568 cpu_state_change_notify(cp->cpu_id, CPU_ON);
1577 1569 cpu_state_change_notify(cp->cpu_id, CPU_INTR_ON);
1578 1570 }
1579 1571
1580 1572 return (error);
1581 1573 }
1582 1574
1583 1575 /*
1584 1576 * Mark the indicated CPU as faulted, taking it offline.
1585 1577 */
1586 1578 int
1587 1579 cpu_faulted(cpu_t *cp, int flags)
1588 1580 {
1589 1581 int error = 0;
1590 1582
1591 1583 ASSERT(MUTEX_HELD(&cpu_lock));
1592 1584 ASSERT(!cpu_is_poweredoff(cp));
1593 1585
1594 1586 if (cpu_is_offline(cp)) {
1595 1587 cp->cpu_flags &= ~CPU_SPARE;
1596 1588 cp->cpu_flags |= CPU_FAULTED;
1597 1589 mp_cpu_faulted_enter(cp);
1598 1590 cpu_set_state(cp);
1599 1591 return (0);
1600 1592 }
1601 1593
1602 1594 if ((error = cpu_offline(cp, flags)) == 0) {
1603 1595 cp->cpu_flags |= CPU_FAULTED;
1604 1596 mp_cpu_faulted_enter(cp);
1605 1597 cpu_set_state(cp);
1606 1598 }
1607 1599
1608 1600 return (error);
1609 1601 }
1610 1602
1611 1603 /*
1612 1604 * Mark the indicated CPU as a spare, taking it offline.
1613 1605 */
1614 1606 int
1615 1607 cpu_spare(cpu_t *cp, int flags)
1616 1608 {
1617 1609 int error = 0;
1618 1610
1619 1611 ASSERT(MUTEX_HELD(&cpu_lock));
1620 1612 ASSERT(!cpu_is_poweredoff(cp));
1621 1613
1622 1614 if (cpu_is_offline(cp)) {
1623 1615 if (cp->cpu_flags & CPU_FAULTED) {
1624 1616 cp->cpu_flags &= ~CPU_FAULTED;
1625 1617 mp_cpu_faulted_exit(cp);
1626 1618 }
1627 1619 cp->cpu_flags |= CPU_SPARE;
1628 1620 cpu_set_state(cp);
1629 1621 return (0);
1630 1622 }
1631 1623
1632 1624 if ((error = cpu_offline(cp, flags)) == 0) {
1633 1625 cp->cpu_flags |= CPU_SPARE;
1634 1626 cpu_set_state(cp);
1635 1627 }
1636 1628
1637 1629 return (error);
1638 1630 }
1639 1631
1640 1632 /*
1641 1633 * Take the indicated CPU from poweroff to offline.
1642 1634 */
1643 1635 int
1644 1636 cpu_poweron(cpu_t *cp)
1645 1637 {
1646 1638 int error = ENOTSUP;
1647 1639
1648 1640 ASSERT(MUTEX_HELD(&cpu_lock));
1649 1641 ASSERT(cpu_is_poweredoff(cp));
1650 1642
1651 1643 error = mp_cpu_poweron(cp); /* arch-dep hook */
1652 1644 if (error == 0)
1653 1645 cpu_set_state(cp);
1654 1646
1655 1647 return (error);
1656 1648 }
1657 1649
1658 1650 /*
1659 1651 * Take the indicated CPU from any inactive state to powered off.
1660 1652 */
1661 1653 int
1662 1654 cpu_poweroff(cpu_t *cp)
1663 1655 {
1664 1656 int error = ENOTSUP;
1665 1657
1666 1658 ASSERT(MUTEX_HELD(&cpu_lock));
1667 1659 ASSERT(cpu_is_offline(cp));
1668 1660
1669 1661 if (!(cp->cpu_flags & CPU_QUIESCED))
1670 1662 return (EBUSY); /* not completely idle */
1671 1663
1672 1664 error = mp_cpu_poweroff(cp); /* arch-dep hook */
1673 1665 if (error == 0)
1674 1666 cpu_set_state(cp);
1675 1667
1676 1668 return (error);
1677 1669 }
1678 1670
1679 1671 /*
1680 1672 * Initialize the Sequential CPU id lookup table
1681 1673 */
1682 1674 void
1683 1675 cpu_seq_tbl_init()
1684 1676 {
1685 1677 cpu_t **tbl;
1686 1678
1687 1679 tbl = kmem_zalloc(sizeof (struct cpu *) * max_ncpus, KM_SLEEP);
1688 1680 tbl[0] = CPU;
1689 1681
1690 1682 cpu_seq = tbl;
1691 1683 }
1692 1684
1693 1685 /*
1694 1686 * Initialize the CPU lists for the first CPU.
1695 1687 */
1696 1688 void
1697 1689 cpu_list_init(cpu_t *cp)
1698 1690 {
1699 1691 cp->cpu_next = cp;
1700 1692 cp->cpu_prev = cp;
1701 1693 cpu_list = cp;
1702 1694 clock_cpu_list = cp;
1703 1695
1704 1696 cp->cpu_next_onln = cp;
1705 1697 cp->cpu_prev_onln = cp;
1706 1698 cpu_active = cp;
1707 1699
1708 1700 cp->cpu_seqid = 0;
1709 1701 CPUSET_ADD(cpu_seqid_inuse, 0);
1710 1702
1711 1703 /*
1712 1704 * Bootstrap cpu_seq using cpu_list
1713 1705 * The cpu_seq[] table will be dynamically allocated
1714 1706 * when kmem later becomes available (but before going MP)
1715 1707 */
1716 1708 cpu_seq = &cpu_list;
1717 1709
1718 1710 cp->cpu_cache_offset = KMEM_CPU_CACHE_OFFSET(cp->cpu_seqid);
1719 1711 cp_default.cp_cpulist = cp;
1720 1712 cp_default.cp_ncpus = 1;
1721 1713 cp->cpu_next_part = cp;
1722 1714 cp->cpu_prev_part = cp;
1723 1715 cp->cpu_part = &cp_default;
1724 1716
1725 1717 CPUSET_ADD(cpu_available, cp->cpu_id);
1726 1718 }
1727 1719
1728 1720 /*
1729 1721 * Insert a CPU into the list of available CPUs.
1730 1722 */
1731 1723 void
1732 1724 cpu_add_unit(cpu_t *cp)
1733 1725 {
1734 1726 int seqid;
1735 1727
1736 1728 ASSERT(MUTEX_HELD(&cpu_lock));
1737 1729 ASSERT(cpu_list != NULL); /* list started in cpu_list_init */
1738 1730
1739 1731 lgrp_config(LGRP_CONFIG_CPU_ADD, (uintptr_t)cp, 0);
1740 1732
1741 1733 /*
1742 1734 * Note: most users of the cpu_list will grab the
1743 1735 * cpu_lock to insure that it isn't modified. However,
1744 1736 * certain users can't or won't do that. To allow this
1745 1737 * we pause the other cpus. Users who walk the list
1746 1738 * without cpu_lock, must disable kernel preemption
1747 1739 * to insure that the list isn't modified underneath
1748 1740 * them. Also, any cached pointers to cpu structures
1749 1741 * must be revalidated by checking to see if the
1750 1742 * cpu_next pointer points to itself. This check must
1751 1743 * be done with the cpu_lock held or kernel preemption
1752 1744 * disabled. This check relies upon the fact that
1753 1745 * old cpu structures are not free'ed or cleared after
1754 1746 * then are removed from the cpu_list.
1755 1747 *
1756 1748 * Note that the clock code walks the cpu list dereferencing
1757 1749 * the cpu_part pointer, so we need to initialize it before
1758 1750 * adding the cpu to the list.
1759 1751 */
1760 1752 cp->cpu_part = &cp_default;
1761 1753 (void) pause_cpus(NULL);
1762 1754 cp->cpu_next = cpu_list;
1763 1755 cp->cpu_prev = cpu_list->cpu_prev;
1764 1756 cpu_list->cpu_prev->cpu_next = cp;
1765 1757 cpu_list->cpu_prev = cp;
1766 1758 start_cpus();
1767 1759
1768 1760 for (seqid = 0; CPU_IN_SET(cpu_seqid_inuse, seqid); seqid++)
1769 1761 continue;
1770 1762 CPUSET_ADD(cpu_seqid_inuse, seqid);
1771 1763 cp->cpu_seqid = seqid;
1772 1764
1773 1765 if (seqid > max_cpu_seqid_ever)
1774 1766 max_cpu_seqid_ever = seqid;
1775 1767
1776 1768 ASSERT(ncpus < max_ncpus);
1777 1769 ncpus++;
1778 1770 cp->cpu_cache_offset = KMEM_CPU_CACHE_OFFSET(cp->cpu_seqid);
1779 1771 cpu[cp->cpu_id] = cp;
1780 1772 CPUSET_ADD(cpu_available, cp->cpu_id);
1781 1773 cpu_seq[cp->cpu_seqid] = cp;
1782 1774
1783 1775 /*
1784 1776 * allocate a pause thread for this CPU.
1785 1777 */
1786 1778 cpu_pause_alloc(cp);
1787 1779
1788 1780 /*
1789 1781 * So that new CPUs won't have NULL prev_onln and next_onln pointers,
1790 1782 * link them into a list of just that CPU.
1791 1783 * This is so that disp_lowpri_cpu will work for thread_create in
1792 1784 * pause_cpus() when called from the startup thread in a new CPU.
1793 1785 */
1794 1786 cp->cpu_next_onln = cp;
1795 1787 cp->cpu_prev_onln = cp;
1796 1788 cpu_info_kstat_create(cp);
1797 1789 cp->cpu_next_part = cp;
1798 1790 cp->cpu_prev_part = cp;
1799 1791
1800 1792 init_cpu_mstate(cp, CMS_SYSTEM);
1801 1793
1802 1794 pool_pset_mod = gethrtime();
1803 1795 }
1804 1796
1805 1797 /*
1806 1798 * Do the opposite of cpu_add_unit().
1807 1799 */
1808 1800 void
1809 1801 cpu_del_unit(int cpuid)
1810 1802 {
1811 1803 struct cpu *cp, *cpnext;
1812 1804
1813 1805 ASSERT(MUTEX_HELD(&cpu_lock));
1814 1806 cp = cpu[cpuid];
1815 1807 ASSERT(cp != NULL);
1816 1808
1817 1809 ASSERT(cp->cpu_next_onln == cp);
1818 1810 ASSERT(cp->cpu_prev_onln == cp);
1819 1811 ASSERT(cp->cpu_next_part == cp);
1820 1812 ASSERT(cp->cpu_prev_part == cp);
1821 1813
1822 1814 /*
1823 1815 * Tear down the CPU's physical ID cache, and update any
1824 1816 * processor groups
1825 1817 */
1826 1818 pg_cpu_fini(cp, NULL);
1827 1819 pghw_physid_destroy(cp);
1828 1820
1829 1821 /*
1830 1822 * Destroy kstat stuff.
1831 1823 */
1832 1824 cpu_info_kstat_destroy(cp);
1833 1825 term_cpu_mstate(cp);
1834 1826 /*
1835 1827 * Free up pause thread.
1836 1828 */
1837 1829 cpu_pause_free(cp);
1838 1830 CPUSET_DEL(cpu_available, cp->cpu_id);
1839 1831 cpu[cp->cpu_id] = NULL;
1840 1832 cpu_seq[cp->cpu_seqid] = NULL;
1841 1833
1842 1834 /*
1843 1835 * The clock thread and mutex_vector_enter cannot hold the
1844 1836 * cpu_lock while traversing the cpu list, therefore we pause
1845 1837 * all other threads by pausing the other cpus. These, and any
1846 1838 * other routines holding cpu pointers while possibly sleeping
1847 1839 * must be sure to call kpreempt_disable before processing the
1848 1840 * list and be sure to check that the cpu has not been deleted
1849 1841 * after any sleeps (check cp->cpu_next != NULL). We guarantee
1850 1842 * to keep the deleted cpu structure around.
1851 1843 *
1852 1844 * Note that this MUST be done AFTER cpu_available
1853 1845 * has been updated so that we don't waste time
1854 1846 * trying to pause the cpu we're trying to delete.
1855 1847 */
1856 1848 (void) pause_cpus(NULL);
1857 1849
1858 1850 cpnext = cp->cpu_next;
1859 1851 cp->cpu_prev->cpu_next = cp->cpu_next;
1860 1852 cp->cpu_next->cpu_prev = cp->cpu_prev;
1861 1853 if (cp == cpu_list)
1862 1854 cpu_list = cpnext;
1863 1855
1864 1856 /*
1865 1857 * Signals that the cpu has been deleted (see above).
1866 1858 */
1867 1859 cp->cpu_next = NULL;
1868 1860 cp->cpu_prev = NULL;
1869 1861
1870 1862 start_cpus();
1871 1863
1872 1864 CPUSET_DEL(cpu_seqid_inuse, cp->cpu_seqid);
1873 1865 ncpus--;
1874 1866 lgrp_config(LGRP_CONFIG_CPU_DEL, (uintptr_t)cp, 0);
1875 1867
1876 1868 pool_pset_mod = gethrtime();
1877 1869 }
1878 1870
1879 1871 /*
1880 1872 * Add a CPU to the list of active CPUs.
1881 1873 * This routine must not get any locks, because other CPUs are paused.
1882 1874 */
1883 1875 static void
1884 1876 cpu_add_active_internal(cpu_t *cp)
1885 1877 {
1886 1878 cpupart_t *pp = cp->cpu_part;
1887 1879
1888 1880 ASSERT(MUTEX_HELD(&cpu_lock));
1889 1881 ASSERT(cpu_list != NULL); /* list started in cpu_list_init */
1890 1882
1891 1883 ncpus_online++;
1892 1884 cpu_set_state(cp);
1893 1885 cp->cpu_next_onln = cpu_active;
1894 1886 cp->cpu_prev_onln = cpu_active->cpu_prev_onln;
1895 1887 cpu_active->cpu_prev_onln->cpu_next_onln = cp;
1896 1888 cpu_active->cpu_prev_onln = cp;
1897 1889
1898 1890 if (pp->cp_cpulist) {
1899 1891 cp->cpu_next_part = pp->cp_cpulist;
1900 1892 cp->cpu_prev_part = pp->cp_cpulist->cpu_prev_part;
1901 1893 pp->cp_cpulist->cpu_prev_part->cpu_next_part = cp;
1902 1894 pp->cp_cpulist->cpu_prev_part = cp;
1903 1895 } else {
1904 1896 ASSERT(pp->cp_ncpus == 0);
1905 1897 pp->cp_cpulist = cp->cpu_next_part = cp->cpu_prev_part = cp;
1906 1898 }
1907 1899 pp->cp_ncpus++;
1908 1900 if (pp->cp_ncpus == 1) {
1909 1901 cp_numparts_nonempty++;
1910 1902 ASSERT(cp_numparts_nonempty != 0);
1911 1903 }
1912 1904
1913 1905 pg_cpu_active(cp);
1914 1906 lgrp_config(LGRP_CONFIG_CPU_ONLINE, (uintptr_t)cp, 0);
1915 1907
1916 1908 bzero(&cp->cpu_loadavg, sizeof (cp->cpu_loadavg));
1917 1909 }
1918 1910
1919 1911 /*
1920 1912 * Add a CPU to the list of active CPUs.
1921 1913 * This is called from machine-dependent layers when a new CPU is started.
1922 1914 */
1923 1915 void
1924 1916 cpu_add_active(cpu_t *cp)
1925 1917 {
1926 1918 pg_cpupart_in(cp, cp->cpu_part);
1927 1919
1928 1920 pause_cpus(NULL);
1929 1921 cpu_add_active_internal(cp);
1930 1922 start_cpus();
1931 1923
1932 1924 cpu_stats_kstat_create(cp);
1933 1925 cpu_create_intrstat(cp);
1934 1926 lgrp_kstat_create(cp);
1935 1927 cpu_state_change_notify(cp->cpu_id, CPU_INIT);
1936 1928 }
1937 1929
1938 1930
1939 1931 /*
1940 1932 * Remove a CPU from the list of active CPUs.
1941 1933 * This routine must not get any locks, because other CPUs are paused.
1942 1934 */
1943 1935 /* ARGSUSED */
1944 1936 static void
1945 1937 cpu_remove_active(cpu_t *cp)
1946 1938 {
1947 1939 cpupart_t *pp = cp->cpu_part;
1948 1940
1949 1941 ASSERT(MUTEX_HELD(&cpu_lock));
1950 1942 ASSERT(cp->cpu_next_onln != cp); /* not the last one */
1951 1943 ASSERT(cp->cpu_prev_onln != cp); /* not the last one */
1952 1944
1953 1945 pg_cpu_inactive(cp);
1954 1946
1955 1947 lgrp_config(LGRP_CONFIG_CPU_OFFLINE, (uintptr_t)cp, 0);
1956 1948
1957 1949 if (cp == clock_cpu_list)
1958 1950 clock_cpu_list = cp->cpu_next_onln;
1959 1951
1960 1952 cp->cpu_prev_onln->cpu_next_onln = cp->cpu_next_onln;
1961 1953 cp->cpu_next_onln->cpu_prev_onln = cp->cpu_prev_onln;
1962 1954 if (cpu_active == cp) {
1963 1955 cpu_active = cp->cpu_next_onln;
1964 1956 }
1965 1957 cp->cpu_next_onln = cp;
1966 1958 cp->cpu_prev_onln = cp;
1967 1959
1968 1960 cp->cpu_prev_part->cpu_next_part = cp->cpu_next_part;
1969 1961 cp->cpu_next_part->cpu_prev_part = cp->cpu_prev_part;
1970 1962 if (pp->cp_cpulist == cp) {
1971 1963 pp->cp_cpulist = cp->cpu_next_part;
1972 1964 ASSERT(pp->cp_cpulist != cp);
1973 1965 }
1974 1966 cp->cpu_next_part = cp;
1975 1967 cp->cpu_prev_part = cp;
1976 1968 pp->cp_ncpus--;
1977 1969 if (pp->cp_ncpus == 0) {
1978 1970 cp_numparts_nonempty--;
1979 1971 ASSERT(cp_numparts_nonempty != 0);
1980 1972 }
1981 1973 }
1982 1974
1983 1975 /*
1984 1976 * Routine used to setup a newly inserted CPU in preparation for starting
1985 1977 * it running code.
1986 1978 */
1987 1979 int
1988 1980 cpu_configure(int cpuid)
1989 1981 {
1990 1982 int retval = 0;
1991 1983
1992 1984 ASSERT(MUTEX_HELD(&cpu_lock));
1993 1985
1994 1986 /*
1995 1987 * Some structures are statically allocated based upon
1996 1988 * the maximum number of cpus the system supports. Do not
1997 1989 * try to add anything beyond this limit.
1998 1990 */
1999 1991 if (cpuid < 0 || cpuid >= NCPU) {
2000 1992 return (EINVAL);
2001 1993 }
2002 1994
2003 1995 if ((cpu[cpuid] != NULL) && (cpu[cpuid]->cpu_flags != 0)) {
2004 1996 return (EALREADY);
2005 1997 }
2006 1998
2007 1999 if ((retval = mp_cpu_configure(cpuid)) != 0) {
2008 2000 return (retval);
2009 2001 }
2010 2002
2011 2003 cpu[cpuid]->cpu_flags = CPU_QUIESCED | CPU_OFFLINE | CPU_POWEROFF;
2012 2004 cpu_set_state(cpu[cpuid]);
2013 2005 retval = cpu_state_change_hooks(cpuid, CPU_CONFIG, CPU_UNCONFIG);
2014 2006 if (retval != 0)
2015 2007 (void) mp_cpu_unconfigure(cpuid);
2016 2008
2017 2009 return (retval);
2018 2010 }
2019 2011
2020 2012 /*
2021 2013 * Routine used to cleanup a CPU that has been powered off. This will
2022 2014 * destroy all per-cpu information related to this cpu.
2023 2015 */
2024 2016 int
2025 2017 cpu_unconfigure(int cpuid)
2026 2018 {
2027 2019 int error;
2028 2020
2029 2021 ASSERT(MUTEX_HELD(&cpu_lock));
2030 2022
2031 2023 if (cpu[cpuid] == NULL) {
2032 2024 return (ENODEV);
2033 2025 }
2034 2026
2035 2027 if (cpu[cpuid]->cpu_flags == 0) {
2036 2028 return (EALREADY);
2037 2029 }
2038 2030
2039 2031 if ((cpu[cpuid]->cpu_flags & CPU_POWEROFF) == 0) {
2040 2032 return (EBUSY);
2041 2033 }
2042 2034
2043 2035 if (cpu[cpuid]->cpu_props != NULL) {
2044 2036 (void) nvlist_free(cpu[cpuid]->cpu_props);
2045 2037 cpu[cpuid]->cpu_props = NULL;
2046 2038 }
2047 2039
2048 2040 error = cpu_state_change_hooks(cpuid, CPU_UNCONFIG, CPU_CONFIG);
2049 2041
2050 2042 if (error != 0)
2051 2043 return (error);
2052 2044
2053 2045 return (mp_cpu_unconfigure(cpuid));
2054 2046 }
2055 2047
2056 2048 /*
2057 2049 * Routines for registering and de-registering cpu_setup callback functions.
2058 2050 *
2059 2051 * Caller's context
2060 2052 * These routines must not be called from a driver's attach(9E) or
2061 2053 * detach(9E) entry point.
2062 2054 *
2063 2055 * NOTE: CPU callbacks should not block. They are called with cpu_lock held.
2064 2056 */
2065 2057
2066 2058 /*
2067 2059 * Ideally, these would be dynamically allocated and put into a linked
2068 2060 * list; however that is not feasible because the registration routine
2069 2061 * has to be available before the kmem allocator is working (in fact,
2070 2062 * it is called by the kmem allocator init code). In any case, there
2071 2063 * are quite a few extra entries for future users.
2072 2064 */
2073 2065 #define NCPU_SETUPS 20
2074 2066
2075 2067 struct cpu_setup {
2076 2068 cpu_setup_func_t *func;
2077 2069 void *arg;
2078 2070 } cpu_setups[NCPU_SETUPS];
2079 2071
2080 2072 void
2081 2073 register_cpu_setup_func(cpu_setup_func_t *func, void *arg)
2082 2074 {
2083 2075 int i;
2084 2076
2085 2077 ASSERT(MUTEX_HELD(&cpu_lock));
2086 2078
2087 2079 for (i = 0; i < NCPU_SETUPS; i++)
2088 2080 if (cpu_setups[i].func == NULL)
2089 2081 break;
2090 2082 if (i >= NCPU_SETUPS)
2091 2083 cmn_err(CE_PANIC, "Ran out of cpu_setup callback entries");
2092 2084
2093 2085 cpu_setups[i].func = func;
2094 2086 cpu_setups[i].arg = arg;
2095 2087 }
2096 2088
2097 2089 void
2098 2090 unregister_cpu_setup_func(cpu_setup_func_t *func, void *arg)
2099 2091 {
2100 2092 int i;
2101 2093
2102 2094 ASSERT(MUTEX_HELD(&cpu_lock));
2103 2095
2104 2096 for (i = 0; i < NCPU_SETUPS; i++)
2105 2097 if ((cpu_setups[i].func == func) &&
2106 2098 (cpu_setups[i].arg == arg))
2107 2099 break;
2108 2100 if (i >= NCPU_SETUPS)
2109 2101 cmn_err(CE_PANIC, "Could not find cpu_setup callback to "
2110 2102 "deregister");
2111 2103
2112 2104 cpu_setups[i].func = NULL;
2113 2105 cpu_setups[i].arg = 0;
2114 2106 }
2115 2107
2116 2108 /*
2117 2109 * Call any state change hooks for this CPU, ignore any errors.
2118 2110 */
2119 2111 void
2120 2112 cpu_state_change_notify(int id, cpu_setup_t what)
2121 2113 {
2122 2114 int i;
2123 2115
2124 2116 ASSERT(MUTEX_HELD(&cpu_lock));
2125 2117
2126 2118 for (i = 0; i < NCPU_SETUPS; i++) {
2127 2119 if (cpu_setups[i].func != NULL) {
2128 2120 cpu_setups[i].func(what, id, cpu_setups[i].arg);
2129 2121 }
2130 2122 }
2131 2123 }
2132 2124
2133 2125 /*
2134 2126 * Call any state change hooks for this CPU, undo it if error found.
2135 2127 */
2136 2128 static int
2137 2129 cpu_state_change_hooks(int id, cpu_setup_t what, cpu_setup_t undo)
2138 2130 {
2139 2131 int i;
2140 2132 int retval = 0;
2141 2133
2142 2134 ASSERT(MUTEX_HELD(&cpu_lock));
2143 2135
2144 2136 for (i = 0; i < NCPU_SETUPS; i++) {
2145 2137 if (cpu_setups[i].func != NULL) {
2146 2138 retval = cpu_setups[i].func(what, id,
2147 2139 cpu_setups[i].arg);
2148 2140 if (retval) {
2149 2141 for (i--; i >= 0; i--) {
2150 2142 if (cpu_setups[i].func != NULL)
2151 2143 cpu_setups[i].func(undo,
2152 2144 id, cpu_setups[i].arg);
2153 2145 }
2154 2146 break;
2155 2147 }
2156 2148 }
2157 2149 }
2158 2150 return (retval);
2159 2151 }
2160 2152
2161 2153 /*
2162 2154 * Export information about this CPU via the kstat mechanism.
2163 2155 */
2164 2156 static struct {
2165 2157 kstat_named_t ci_state;
2166 2158 kstat_named_t ci_state_begin;
2167 2159 kstat_named_t ci_cpu_type;
2168 2160 kstat_named_t ci_fpu_type;
2169 2161 kstat_named_t ci_clock_MHz;
2170 2162 kstat_named_t ci_chip_id;
2171 2163 kstat_named_t ci_implementation;
2172 2164 kstat_named_t ci_brandstr;
2173 2165 kstat_named_t ci_core_id;
2174 2166 kstat_named_t ci_curr_clock_Hz;
2175 2167 kstat_named_t ci_supp_freq_Hz;
2176 2168 kstat_named_t ci_pg_id;
2177 2169 #if defined(__sparcv9)
2178 2170 kstat_named_t ci_device_ID;
2179 2171 kstat_named_t ci_cpu_fru;
2180 2172 #endif
2181 2173 #if defined(__x86)
2182 2174 kstat_named_t ci_vendorstr;
2183 2175 kstat_named_t ci_family;
2184 2176 kstat_named_t ci_model;
2185 2177 kstat_named_t ci_step;
2186 2178 kstat_named_t ci_clogid;
2187 2179 kstat_named_t ci_pkg_core_id;
2188 2180 kstat_named_t ci_ncpuperchip;
2189 2181 kstat_named_t ci_ncoreperchip;
2190 2182 kstat_named_t ci_max_cstates;
2191 2183 kstat_named_t ci_curr_cstate;
2192 2184 kstat_named_t ci_cacheid;
2193 2185 kstat_named_t ci_sktstr;
2194 2186 #endif
2195 2187 } cpu_info_template = {
2196 2188 { "state", KSTAT_DATA_CHAR },
2197 2189 { "state_begin", KSTAT_DATA_LONG },
2198 2190 { "cpu_type", KSTAT_DATA_CHAR },
2199 2191 { "fpu_type", KSTAT_DATA_CHAR },
2200 2192 { "clock_MHz", KSTAT_DATA_LONG },
2201 2193 { "chip_id", KSTAT_DATA_LONG },
2202 2194 { "implementation", KSTAT_DATA_STRING },
2203 2195 { "brand", KSTAT_DATA_STRING },
2204 2196 { "core_id", KSTAT_DATA_LONG },
2205 2197 { "current_clock_Hz", KSTAT_DATA_UINT64 },
2206 2198 { "supported_frequencies_Hz", KSTAT_DATA_STRING },
2207 2199 { "pg_id", KSTAT_DATA_LONG },
2208 2200 #if defined(__sparcv9)
2209 2201 { "device_ID", KSTAT_DATA_UINT64 },
2210 2202 { "cpu_fru", KSTAT_DATA_STRING },
2211 2203 #endif
2212 2204 #if defined(__x86)
2213 2205 { "vendor_id", KSTAT_DATA_STRING },
2214 2206 { "family", KSTAT_DATA_INT32 },
2215 2207 { "model", KSTAT_DATA_INT32 },
2216 2208 { "stepping", KSTAT_DATA_INT32 },
2217 2209 { "clog_id", KSTAT_DATA_INT32 },
2218 2210 { "pkg_core_id", KSTAT_DATA_LONG },
2219 2211 { "ncpu_per_chip", KSTAT_DATA_INT32 },
2220 2212 { "ncore_per_chip", KSTAT_DATA_INT32 },
2221 2213 { "supported_max_cstates", KSTAT_DATA_INT32 },
2222 2214 { "current_cstate", KSTAT_DATA_INT32 },
2223 2215 { "cache_id", KSTAT_DATA_INT32 },
2224 2216 { "socket_type", KSTAT_DATA_STRING },
2225 2217 #endif
2226 2218 };
2227 2219
2228 2220 static kmutex_t cpu_info_template_lock;
2229 2221
2230 2222 static int
2231 2223 cpu_info_kstat_update(kstat_t *ksp, int rw)
2232 2224 {
2233 2225 cpu_t *cp = ksp->ks_private;
2234 2226 const char *pi_state;
2235 2227
2236 2228 if (rw == KSTAT_WRITE)
2237 2229 return (EACCES);
2238 2230
2239 2231 #if defined(__x86)
2240 2232 /* Is the cpu still initialising itself? */
2241 2233 if (cpuid_checkpass(cp, 1) == 0)
2242 2234 return (ENXIO);
2243 2235 #endif
2244 2236 switch (cp->cpu_type_info.pi_state) {
2245 2237 case P_ONLINE:
2246 2238 pi_state = PS_ONLINE;
2247 2239 break;
2248 2240 case P_POWEROFF:
2249 2241 pi_state = PS_POWEROFF;
2250 2242 break;
2251 2243 case P_NOINTR:
2252 2244 pi_state = PS_NOINTR;
2253 2245 break;
2254 2246 case P_FAULTED:
2255 2247 pi_state = PS_FAULTED;
2256 2248 break;
2257 2249 case P_SPARE:
2258 2250 pi_state = PS_SPARE;
2259 2251 break;
2260 2252 case P_OFFLINE:
2261 2253 pi_state = PS_OFFLINE;
2262 2254 break;
2263 2255 default:
2264 2256 pi_state = "unknown";
2265 2257 }
2266 2258 (void) strcpy(cpu_info_template.ci_state.value.c, pi_state);
2267 2259 cpu_info_template.ci_state_begin.value.l = cp->cpu_state_begin;
2268 2260 (void) strncpy(cpu_info_template.ci_cpu_type.value.c,
2269 2261 cp->cpu_type_info.pi_processor_type, 15);
2270 2262 (void) strncpy(cpu_info_template.ci_fpu_type.value.c,
2271 2263 cp->cpu_type_info.pi_fputypes, 15);
2272 2264 cpu_info_template.ci_clock_MHz.value.l = cp->cpu_type_info.pi_clock;
2273 2265 cpu_info_template.ci_chip_id.value.l =
2274 2266 pg_plat_hw_instance_id(cp, PGHW_CHIP);
2275 2267 kstat_named_setstr(&cpu_info_template.ci_implementation,
2276 2268 cp->cpu_idstr);
2277 2269 kstat_named_setstr(&cpu_info_template.ci_brandstr, cp->cpu_brandstr);
2278 2270 cpu_info_template.ci_core_id.value.l = pg_plat_get_core_id(cp);
2279 2271 cpu_info_template.ci_curr_clock_Hz.value.ui64 =
2280 2272 cp->cpu_curr_clock;
2281 2273 cpu_info_template.ci_pg_id.value.l =
2282 2274 cp->cpu_pg && cp->cpu_pg->cmt_lineage ?
2283 2275 cp->cpu_pg->cmt_lineage->pg_id : -1;
2284 2276 kstat_named_setstr(&cpu_info_template.ci_supp_freq_Hz,
2285 2277 cp->cpu_supp_freqs);
2286 2278 #if defined(__sparcv9)
2287 2279 cpu_info_template.ci_device_ID.value.ui64 =
2288 2280 cpunodes[cp->cpu_id].device_id;
2289 2281 kstat_named_setstr(&cpu_info_template.ci_cpu_fru, cpu_fru_fmri(cp));
2290 2282 #endif
2291 2283 #if defined(__x86)
2292 2284 kstat_named_setstr(&cpu_info_template.ci_vendorstr,
2293 2285 cpuid_getvendorstr(cp));
2294 2286 cpu_info_template.ci_family.value.l = cpuid_getfamily(cp);
2295 2287 cpu_info_template.ci_model.value.l = cpuid_getmodel(cp);
2296 2288 cpu_info_template.ci_step.value.l = cpuid_getstep(cp);
2297 2289 cpu_info_template.ci_clogid.value.l = cpuid_get_clogid(cp);
2298 2290 cpu_info_template.ci_ncpuperchip.value.l = cpuid_get_ncpu_per_chip(cp);
2299 2291 cpu_info_template.ci_ncoreperchip.value.l =
2300 2292 cpuid_get_ncore_per_chip(cp);
2301 2293 cpu_info_template.ci_pkg_core_id.value.l = cpuid_get_pkgcoreid(cp);
2302 2294 cpu_info_template.ci_max_cstates.value.l = cp->cpu_m.max_cstates;
2303 2295 cpu_info_template.ci_curr_cstate.value.l = cpu_idle_get_cpu_state(cp);
2304 2296 cpu_info_template.ci_cacheid.value.i32 = cpuid_get_cacheid(cp);
2305 2297 kstat_named_setstr(&cpu_info_template.ci_sktstr,
2306 2298 cpuid_getsocketstr(cp));
2307 2299 #endif
2308 2300
2309 2301 return (0);
2310 2302 }
2311 2303
2312 2304 static void
2313 2305 cpu_info_kstat_create(cpu_t *cp)
2314 2306 {
2315 2307 zoneid_t zoneid;
2316 2308
2317 2309 ASSERT(MUTEX_HELD(&cpu_lock));
2318 2310
2319 2311 if (pool_pset_enabled())
2320 2312 zoneid = GLOBAL_ZONEID;
2321 2313 else
2322 2314 zoneid = ALL_ZONES;
2323 2315 if ((cp->cpu_info_kstat = kstat_create_zone("cpu_info", cp->cpu_id,
2324 2316 NULL, "misc", KSTAT_TYPE_NAMED,
2325 2317 sizeof (cpu_info_template) / sizeof (kstat_named_t),
2326 2318 KSTAT_FLAG_VIRTUAL | KSTAT_FLAG_VAR_SIZE, zoneid)) != NULL) {
2327 2319 cp->cpu_info_kstat->ks_data_size += 2 * CPU_IDSTRLEN;
2328 2320 #if defined(__sparcv9)
2329 2321 cp->cpu_info_kstat->ks_data_size +=
2330 2322 strlen(cpu_fru_fmri(cp)) + 1;
2331 2323 #endif
2332 2324 #if defined(__x86)
2333 2325 cp->cpu_info_kstat->ks_data_size += X86_VENDOR_STRLEN;
2334 2326 #endif
2335 2327 if (cp->cpu_supp_freqs != NULL)
2336 2328 cp->cpu_info_kstat->ks_data_size +=
2337 2329 strlen(cp->cpu_supp_freqs) + 1;
2338 2330 cp->cpu_info_kstat->ks_lock = &cpu_info_template_lock;
2339 2331 cp->cpu_info_kstat->ks_data = &cpu_info_template;
2340 2332 cp->cpu_info_kstat->ks_private = cp;
2341 2333 cp->cpu_info_kstat->ks_update = cpu_info_kstat_update;
2342 2334 kstat_install(cp->cpu_info_kstat);
2343 2335 }
2344 2336 }
2345 2337
2346 2338 static void
2347 2339 cpu_info_kstat_destroy(cpu_t *cp)
2348 2340 {
2349 2341 ASSERT(MUTEX_HELD(&cpu_lock));
2350 2342
2351 2343 kstat_delete(cp->cpu_info_kstat);
2352 2344 cp->cpu_info_kstat = NULL;
2353 2345 }
2354 2346
2355 2347 /*
2356 2348 * Create and install kstats for the boot CPU.
2357 2349 */
2358 2350 void
2359 2351 cpu_kstat_init(cpu_t *cp)
2360 2352 {
2361 2353 mutex_enter(&cpu_lock);
2362 2354 cpu_info_kstat_create(cp);
2363 2355 cpu_stats_kstat_create(cp);
2364 2356 cpu_create_intrstat(cp);
2365 2357 cpu_set_state(cp);
2366 2358 mutex_exit(&cpu_lock);
2367 2359 }
2368 2360
2369 2361 /*
2370 2362 * Make visible to the zone that subset of the cpu information that would be
2371 2363 * initialized when a cpu is configured (but still offline).
2372 2364 */
2373 2365 void
2374 2366 cpu_visibility_configure(cpu_t *cp, zone_t *zone)
2375 2367 {
2376 2368 zoneid_t zoneid = zone ? zone->zone_id : ALL_ZONES;
2377 2369
2378 2370 ASSERT(MUTEX_HELD(&cpu_lock));
2379 2371 ASSERT(pool_pset_enabled());
2380 2372 ASSERT(cp != NULL);
2381 2373
2382 2374 if (zoneid != ALL_ZONES && zoneid != GLOBAL_ZONEID) {
2383 2375 zone->zone_ncpus++;
2384 2376 ASSERT(zone->zone_ncpus <= ncpus);
2385 2377 }
2386 2378 if (cp->cpu_info_kstat != NULL)
2387 2379 kstat_zone_add(cp->cpu_info_kstat, zoneid);
2388 2380 }
2389 2381
2390 2382 /*
2391 2383 * Make visible to the zone that subset of the cpu information that would be
2392 2384 * initialized when a previously configured cpu is onlined.
2393 2385 */
2394 2386 void
2395 2387 cpu_visibility_online(cpu_t *cp, zone_t *zone)
2396 2388 {
2397 2389 kstat_t *ksp;
2398 2390 char name[sizeof ("cpu_stat") + 10]; /* enough for 32-bit cpuids */
2399 2391 zoneid_t zoneid = zone ? zone->zone_id : ALL_ZONES;
2400 2392 processorid_t cpun;
2401 2393
2402 2394 ASSERT(MUTEX_HELD(&cpu_lock));
2403 2395 ASSERT(pool_pset_enabled());
2404 2396 ASSERT(cp != NULL);
2405 2397 ASSERT(cpu_is_active(cp));
2406 2398
2407 2399 cpun = cp->cpu_id;
2408 2400 if (zoneid != ALL_ZONES && zoneid != GLOBAL_ZONEID) {
2409 2401 zone->zone_ncpus_online++;
2410 2402 ASSERT(zone->zone_ncpus_online <= ncpus_online);
2411 2403 }
2412 2404 (void) snprintf(name, sizeof (name), "cpu_stat%d", cpun);
2413 2405 if ((ksp = kstat_hold_byname("cpu_stat", cpun, name, ALL_ZONES))
2414 2406 != NULL) {
2415 2407 kstat_zone_add(ksp, zoneid);
2416 2408 kstat_rele(ksp);
2417 2409 }
2418 2410 if ((ksp = kstat_hold_byname("cpu", cpun, "sys", ALL_ZONES)) != NULL) {
2419 2411 kstat_zone_add(ksp, zoneid);
2420 2412 kstat_rele(ksp);
2421 2413 }
2422 2414 if ((ksp = kstat_hold_byname("cpu", cpun, "vm", ALL_ZONES)) != NULL) {
2423 2415 kstat_zone_add(ksp, zoneid);
2424 2416 kstat_rele(ksp);
2425 2417 }
2426 2418 if ((ksp = kstat_hold_byname("cpu", cpun, "intrstat", ALL_ZONES)) !=
2427 2419 NULL) {
2428 2420 kstat_zone_add(ksp, zoneid);
2429 2421 kstat_rele(ksp);
2430 2422 }
2431 2423 }
2432 2424
2433 2425 /*
2434 2426 * Update relevant kstats such that cpu is now visible to processes
2435 2427 * executing in specified zone.
2436 2428 */
2437 2429 void
2438 2430 cpu_visibility_add(cpu_t *cp, zone_t *zone)
2439 2431 {
2440 2432 cpu_visibility_configure(cp, zone);
2441 2433 if (cpu_is_active(cp))
2442 2434 cpu_visibility_online(cp, zone);
2443 2435 }
2444 2436
2445 2437 /*
2446 2438 * Make invisible to the zone that subset of the cpu information that would be
2447 2439 * torn down when a previously offlined cpu is unconfigured.
2448 2440 */
2449 2441 void
2450 2442 cpu_visibility_unconfigure(cpu_t *cp, zone_t *zone)
2451 2443 {
2452 2444 zoneid_t zoneid = zone ? zone->zone_id : ALL_ZONES;
2453 2445
2454 2446 ASSERT(MUTEX_HELD(&cpu_lock));
2455 2447 ASSERT(pool_pset_enabled());
2456 2448 ASSERT(cp != NULL);
2457 2449
2458 2450 if (zoneid != ALL_ZONES && zoneid != GLOBAL_ZONEID) {
2459 2451 ASSERT(zone->zone_ncpus != 0);
2460 2452 zone->zone_ncpus--;
2461 2453 }
2462 2454 if (cp->cpu_info_kstat)
2463 2455 kstat_zone_remove(cp->cpu_info_kstat, zoneid);
2464 2456 }
2465 2457
2466 2458 /*
2467 2459 * Make invisible to the zone that subset of the cpu information that would be
2468 2460 * torn down when a cpu is offlined (but still configured).
2469 2461 */
2470 2462 void
2471 2463 cpu_visibility_offline(cpu_t *cp, zone_t *zone)
2472 2464 {
2473 2465 kstat_t *ksp;
2474 2466 char name[sizeof ("cpu_stat") + 10]; /* enough for 32-bit cpuids */
2475 2467 zoneid_t zoneid = zone ? zone->zone_id : ALL_ZONES;
2476 2468 processorid_t cpun;
2477 2469
2478 2470 ASSERT(MUTEX_HELD(&cpu_lock));
2479 2471 ASSERT(pool_pset_enabled());
2480 2472 ASSERT(cp != NULL);
2481 2473 ASSERT(cpu_is_active(cp));
2482 2474
2483 2475 cpun = cp->cpu_id;
2484 2476 if (zoneid != ALL_ZONES && zoneid != GLOBAL_ZONEID) {
2485 2477 ASSERT(zone->zone_ncpus_online != 0);
2486 2478 zone->zone_ncpus_online--;
2487 2479 }
2488 2480
2489 2481 if ((ksp = kstat_hold_byname("cpu", cpun, "intrstat", ALL_ZONES)) !=
2490 2482 NULL) {
2491 2483 kstat_zone_remove(ksp, zoneid);
2492 2484 kstat_rele(ksp);
2493 2485 }
2494 2486 if ((ksp = kstat_hold_byname("cpu", cpun, "vm", ALL_ZONES)) != NULL) {
2495 2487 kstat_zone_remove(ksp, zoneid);
2496 2488 kstat_rele(ksp);
2497 2489 }
2498 2490 if ((ksp = kstat_hold_byname("cpu", cpun, "sys", ALL_ZONES)) != NULL) {
2499 2491 kstat_zone_remove(ksp, zoneid);
2500 2492 kstat_rele(ksp);
2501 2493 }
2502 2494 (void) snprintf(name, sizeof (name), "cpu_stat%d", cpun);
2503 2495 if ((ksp = kstat_hold_byname("cpu_stat", cpun, name, ALL_ZONES))
2504 2496 != NULL) {
2505 2497 kstat_zone_remove(ksp, zoneid);
2506 2498 kstat_rele(ksp);
2507 2499 }
2508 2500 }
2509 2501
2510 2502 /*
2511 2503 * Update relevant kstats such that cpu is no longer visible to processes
2512 2504 * executing in specified zone.
2513 2505 */
2514 2506 void
2515 2507 cpu_visibility_remove(cpu_t *cp, zone_t *zone)
2516 2508 {
2517 2509 if (cpu_is_active(cp))
2518 2510 cpu_visibility_offline(cp, zone);
2519 2511 cpu_visibility_unconfigure(cp, zone);
2520 2512 }
2521 2513
2522 2514 /*
2523 2515 * Bind a thread to a CPU as requested.
2524 2516 */
2525 2517 int
2526 2518 cpu_bind_thread(kthread_id_t tp, processorid_t bind, processorid_t *obind,
2527 2519 int *error)
2528 2520 {
2529 2521 processorid_t binding;
2530 2522 cpu_t *cp = NULL;
2531 2523
2532 2524 ASSERT(MUTEX_HELD(&cpu_lock));
2533 2525 ASSERT(MUTEX_HELD(&ttoproc(tp)->p_lock));
2534 2526
2535 2527 thread_lock(tp);
2536 2528
2537 2529 /*
2538 2530 * Record old binding, but change the obind, which was initialized
2539 2531 * to PBIND_NONE, only if this thread has a binding. This avoids
2540 2532 * reporting PBIND_NONE for a process when some LWPs are bound.
2541 2533 */
2542 2534 binding = tp->t_bind_cpu;
2543 2535 if (binding != PBIND_NONE)
2544 2536 *obind = binding; /* record old binding */
2545 2537
2546 2538 switch (bind) {
2547 2539 case PBIND_QUERY:
2548 2540 /* Just return the old binding */
2549 2541 thread_unlock(tp);
2550 2542 return (0);
2551 2543
2552 2544 case PBIND_QUERY_TYPE:
2553 2545 /* Return the binding type */
2554 2546 *obind = TB_CPU_IS_SOFT(tp) ? PBIND_SOFT : PBIND_HARD;
2555 2547 thread_unlock(tp);
2556 2548 return (0);
2557 2549
2558 2550 case PBIND_SOFT:
2559 2551 /*
2560 2552 * Set soft binding for this thread and return the actual
2561 2553 * binding
2562 2554 */
2563 2555 TB_CPU_SOFT_SET(tp);
2564 2556 thread_unlock(tp);
2565 2557 return (0);
2566 2558
2567 2559 case PBIND_HARD:
2568 2560 /*
2569 2561 * Set hard binding for this thread and return the actual
2570 2562 * binding
2571 2563 */
2572 2564 TB_CPU_HARD_SET(tp);
2573 2565 thread_unlock(tp);
2574 2566 return (0);
2575 2567
2576 2568 default:
2577 2569 break;
2578 2570 }
2579 2571
2580 2572 /*
2581 2573 * If this thread/LWP cannot be bound because of permission
2582 2574 * problems, just note that and return success so that the
2583 2575 * other threads/LWPs will be bound. This is the way
2584 2576 * processor_bind() is defined to work.
2585 2577 *
2586 2578 * Binding will get EPERM if the thread is of system class
2587 2579 * or hasprocperm() fails.
2588 2580 */
2589 2581 if (tp->t_cid == 0 || !hasprocperm(tp->t_cred, CRED())) {
2590 2582 *error = EPERM;
2591 2583 thread_unlock(tp);
2592 2584 return (0);
2593 2585 }
2594 2586
2595 2587 binding = bind;
2596 2588 if (binding != PBIND_NONE) {
2597 2589 cp = cpu_get((processorid_t)binding);
2598 2590 /*
2599 2591 * Make sure binding is valid and is in right partition.
2600 2592 */
2601 2593 if (cp == NULL || tp->t_cpupart != cp->cpu_part) {
2602 2594 *error = EINVAL;
2603 2595 thread_unlock(tp);
2604 2596 return (0);
2605 2597 }
2606 2598 }
2607 2599 tp->t_bind_cpu = binding; /* set new binding */
2608 2600
2609 2601 /*
2610 2602 * If there is no system-set reason for affinity, set
2611 2603 * the t_bound_cpu field to reflect the binding.
2612 2604 */
2613 2605 if (tp->t_affinitycnt == 0) {
2614 2606 if (binding == PBIND_NONE) {
2615 2607 /*
2616 2608 * We may need to adjust disp_max_unbound_pri
2617 2609 * since we're becoming unbound.
2618 2610 */
2619 2611 disp_adjust_unbound_pri(tp);
2620 2612
2621 2613 tp->t_bound_cpu = NULL; /* set new binding */
2622 2614
2623 2615 /*
2624 2616 * Move thread to lgroup with strongest affinity
2625 2617 * after unbinding
2626 2618 */
2627 2619 if (tp->t_lgrp_affinity)
2628 2620 lgrp_move_thread(tp,
2629 2621 lgrp_choose(tp, tp->t_cpupart), 1);
2630 2622
2631 2623 if (tp->t_state == TS_ONPROC &&
2632 2624 tp->t_cpu->cpu_part != tp->t_cpupart)
2633 2625 cpu_surrender(tp);
2634 2626 } else {
2635 2627 lpl_t *lpl;
2636 2628
2637 2629 tp->t_bound_cpu = cp;
2638 2630 ASSERT(cp->cpu_lpl != NULL);
2639 2631
2640 2632 /*
2641 2633 * Set home to lgroup with most affinity containing CPU
2642 2634 * that thread is being bound or minimum bounding
2643 2635 * lgroup if no affinities set
2644 2636 */
2645 2637 if (tp->t_lgrp_affinity)
2646 2638 lpl = lgrp_affinity_best(tp, tp->t_cpupart,
2647 2639 LGRP_NONE, B_FALSE);
2648 2640 else
2649 2641 lpl = cp->cpu_lpl;
2650 2642
2651 2643 if (tp->t_lpl != lpl) {
2652 2644 /* can't grab cpu_lock */
2653 2645 lgrp_move_thread(tp, lpl, 1);
2654 2646 }
2655 2647
2656 2648 /*
2657 2649 * Make the thread switch to the bound CPU.
2658 2650 * If the thread is runnable, we need to
2659 2651 * requeue it even if t_cpu is already set
2660 2652 * to the right CPU, since it may be on a
2661 2653 * kpreempt queue and need to move to a local
2662 2654 * queue. We could check t_disp_queue to
2663 2655 * avoid unnecessary overhead if it's already
2664 2656 * on the right queue, but since this isn't
2665 2657 * a performance-critical operation it doesn't
2666 2658 * seem worth the extra code and complexity.
2667 2659 *
2668 2660 * If the thread is weakbound to the cpu then it will
2669 2661 * resist the new binding request until the weak
2670 2662 * binding drops. The cpu_surrender or requeueing
2671 2663 * below could be skipped in such cases (since it
2672 2664 * will have no effect), but that would require
2673 2665 * thread_allowmigrate to acquire thread_lock so
↓ open down ↓ |
2323 lines elided |
↑ open up ↑ |
2674 2666 * we'll take the very occasional hit here instead.
2675 2667 */
2676 2668 if (tp->t_state == TS_ONPROC) {
2677 2669 cpu_surrender(tp);
2678 2670 } else if (tp->t_state == TS_RUN) {
2679 2671 cpu_t *ocp = tp->t_cpu;
2680 2672
2681 2673 (void) dispdeq(tp);
2682 2674 setbackdq(tp);
2683 2675 /*
2684 - * Either on the bound CPU's disp queue now,
2685 - * or swapped out or on the swap queue.
2676 + * On the bound CPU's disp queue now.
2686 2677 */
2687 2678 ASSERT(tp->t_disp_queue == cp->cpu_disp ||
2688 - tp->t_weakbound_cpu == ocp ||
2689 - (tp->t_schedflag & (TS_LOAD | TS_ON_SWAPQ))
2690 - != TS_LOAD);
2679 + tp->t_weakbound_cpu == ocp);
2691 2680 }
2692 2681 }
2693 2682 }
2694 2683
2695 2684 /*
2696 2685 * Our binding has changed; set TP_CHANGEBIND.
2697 2686 */
2698 2687 tp->t_proc_flag |= TP_CHANGEBIND;
2699 2688 aston(tp);
2700 2689
2701 2690 thread_unlock(tp);
2702 2691
2703 2692 return (0);
2704 2693 }
2705 2694
2706 2695 #if CPUSET_WORDS > 1
2707 2696
2708 2697 /*
2709 2698 * Functions for implementing cpuset operations when a cpuset is more
2710 2699 * than one word. On platforms where a cpuset is a single word these
2711 2700 * are implemented as macros in cpuvar.h.
2712 2701 */
2713 2702
2714 2703 void
2715 2704 cpuset_all(cpuset_t *s)
2716 2705 {
2717 2706 int i;
2718 2707
2719 2708 for (i = 0; i < CPUSET_WORDS; i++)
2720 2709 s->cpub[i] = ~0UL;
2721 2710 }
2722 2711
2723 2712 void
2724 2713 cpuset_all_but(cpuset_t *s, uint_t cpu)
2725 2714 {
2726 2715 cpuset_all(s);
2727 2716 CPUSET_DEL(*s, cpu);
2728 2717 }
2729 2718
2730 2719 void
2731 2720 cpuset_only(cpuset_t *s, uint_t cpu)
2732 2721 {
2733 2722 CPUSET_ZERO(*s);
2734 2723 CPUSET_ADD(*s, cpu);
2735 2724 }
2736 2725
2737 2726 int
2738 2727 cpuset_isnull(cpuset_t *s)
2739 2728 {
2740 2729 int i;
2741 2730
2742 2731 for (i = 0; i < CPUSET_WORDS; i++)
2743 2732 if (s->cpub[i] != 0)
2744 2733 return (0);
2745 2734 return (1);
2746 2735 }
2747 2736
2748 2737 int
2749 2738 cpuset_cmp(cpuset_t *s1, cpuset_t *s2)
2750 2739 {
2751 2740 int i;
2752 2741
2753 2742 for (i = 0; i < CPUSET_WORDS; i++)
2754 2743 if (s1->cpub[i] != s2->cpub[i])
2755 2744 return (0);
2756 2745 return (1);
2757 2746 }
2758 2747
2759 2748 uint_t
2760 2749 cpuset_find(cpuset_t *s)
2761 2750 {
2762 2751
2763 2752 uint_t i;
2764 2753 uint_t cpu = (uint_t)-1;
2765 2754
2766 2755 /*
2767 2756 * Find a cpu in the cpuset
2768 2757 */
2769 2758 for (i = 0; i < CPUSET_WORDS; i++) {
2770 2759 cpu = (uint_t)(lowbit(s->cpub[i]) - 1);
2771 2760 if (cpu != (uint_t)-1) {
2772 2761 cpu += i * BT_NBIPUL;
2773 2762 break;
2774 2763 }
2775 2764 }
2776 2765 return (cpu);
2777 2766 }
2778 2767
2779 2768 void
2780 2769 cpuset_bounds(cpuset_t *s, uint_t *smallestid, uint_t *largestid)
2781 2770 {
2782 2771 int i, j;
2783 2772 uint_t bit;
2784 2773
2785 2774 /*
2786 2775 * First, find the smallest cpu id in the set.
2787 2776 */
2788 2777 for (i = 0; i < CPUSET_WORDS; i++) {
2789 2778 if (s->cpub[i] != 0) {
2790 2779 bit = (uint_t)(lowbit(s->cpub[i]) - 1);
2791 2780 ASSERT(bit != (uint_t)-1);
2792 2781 *smallestid = bit + (i * BT_NBIPUL);
2793 2782
2794 2783 /*
2795 2784 * Now find the largest cpu id in
2796 2785 * the set and return immediately.
2797 2786 * Done in an inner loop to avoid
2798 2787 * having to break out of the first
2799 2788 * loop.
2800 2789 */
2801 2790 for (j = CPUSET_WORDS - 1; j >= i; j--) {
2802 2791 if (s->cpub[j] != 0) {
2803 2792 bit = (uint_t)(highbit(s->cpub[j]) - 1);
2804 2793 ASSERT(bit != (uint_t)-1);
2805 2794 *largestid = bit + (j * BT_NBIPUL);
2806 2795 ASSERT(*largestid >= *smallestid);
2807 2796 return;
2808 2797 }
2809 2798 }
2810 2799
2811 2800 /*
2812 2801 * If this code is reached, a
2813 2802 * smallestid was found, but not a
2814 2803 * largestid. The cpuset must have
2815 2804 * been changed during the course
2816 2805 * of this function call.
2817 2806 */
2818 2807 ASSERT(0);
2819 2808 }
2820 2809 }
2821 2810 *smallestid = *largestid = CPUSET_NOTINSET;
2822 2811 }
2823 2812
2824 2813 #endif /* CPUSET_WORDS */
2825 2814
2826 2815 /*
2827 2816 * Unbind threads bound to specified CPU.
2828 2817 *
2829 2818 * If `unbind_all_threads' is true, unbind all user threads bound to a given
2830 2819 * CPU. Otherwise unbind all soft-bound user threads.
2831 2820 */
2832 2821 int
2833 2822 cpu_unbind(processorid_t cpu, boolean_t unbind_all_threads)
2834 2823 {
2835 2824 processorid_t obind;
2836 2825 kthread_t *tp;
2837 2826 int ret = 0;
2838 2827 proc_t *pp;
2839 2828 int err, berr = 0;
2840 2829
2841 2830 ASSERT(MUTEX_HELD(&cpu_lock));
2842 2831
2843 2832 mutex_enter(&pidlock);
2844 2833 for (pp = practive; pp != NULL; pp = pp->p_next) {
2845 2834 mutex_enter(&pp->p_lock);
2846 2835 tp = pp->p_tlist;
2847 2836 /*
2848 2837 * Skip zombies, kernel processes, and processes in
2849 2838 * other zones, if called from a non-global zone.
2850 2839 */
2851 2840 if (tp == NULL || (pp->p_flag & SSYS) ||
2852 2841 !HASZONEACCESS(curproc, pp->p_zone->zone_id)) {
2853 2842 mutex_exit(&pp->p_lock);
2854 2843 continue;
2855 2844 }
2856 2845 do {
2857 2846 if (tp->t_bind_cpu != cpu)
2858 2847 continue;
2859 2848 /*
2860 2849 * Skip threads with hard binding when
2861 2850 * `unbind_all_threads' is not specified.
2862 2851 */
2863 2852 if (!unbind_all_threads && TB_CPU_IS_HARD(tp))
2864 2853 continue;
2865 2854 err = cpu_bind_thread(tp, PBIND_NONE, &obind, &berr);
2866 2855 if (ret == 0)
2867 2856 ret = err;
2868 2857 } while ((tp = tp->t_forw) != pp->p_tlist);
2869 2858 mutex_exit(&pp->p_lock);
2870 2859 }
2871 2860 mutex_exit(&pidlock);
2872 2861 if (ret == 0)
2873 2862 ret = berr;
2874 2863 return (ret);
2875 2864 }
2876 2865
2877 2866
2878 2867 /*
2879 2868 * Destroy all remaining bound threads on a cpu.
2880 2869 */
2881 2870 void
2882 2871 cpu_destroy_bound_threads(cpu_t *cp)
2883 2872 {
2884 2873 extern id_t syscid;
2885 2874 register kthread_id_t t, tlist, tnext;
2886 2875
2887 2876 /*
2888 2877 * Destroy all remaining bound threads on the cpu. This
2889 2878 * should include both the interrupt threads and the idle thread.
2890 2879 * This requires some care, since we need to traverse the
2891 2880 * thread list with the pidlock mutex locked, but thread_free
2892 2881 * also locks the pidlock mutex. So, we collect the threads
2893 2882 * we're going to reap in a list headed by "tlist", then we
2894 2883 * unlock the pidlock mutex and traverse the tlist list,
2895 2884 * doing thread_free's on the thread's. Simple, n'est pas?
2896 2885 * Also, this depends on thread_free not mucking with the
2897 2886 * t_next and t_prev links of the thread.
2898 2887 */
2899 2888
2900 2889 if ((t = curthread) != NULL) {
2901 2890
2902 2891 tlist = NULL;
2903 2892 mutex_enter(&pidlock);
2904 2893 do {
2905 2894 tnext = t->t_next;
2906 2895 if (t->t_bound_cpu == cp) {
2907 2896
2908 2897 /*
2909 2898 * We've found a bound thread, carefully unlink
2910 2899 * it out of the thread list, and add it to
2911 2900 * our "tlist". We "know" we don't have to
2912 2901 * worry about unlinking curthread (the thread
2913 2902 * that is executing this code).
2914 2903 */
2915 2904 t->t_next->t_prev = t->t_prev;
2916 2905 t->t_prev->t_next = t->t_next;
2917 2906 t->t_next = tlist;
2918 2907 tlist = t;
2919 2908 ASSERT(t->t_cid == syscid);
2920 2909 /* wake up anyone blocked in thread_join */
2921 2910 cv_broadcast(&t->t_joincv);
2922 2911 /*
2923 2912 * t_lwp set by interrupt threads and not
2924 2913 * cleared.
2925 2914 */
2926 2915 t->t_lwp = NULL;
2927 2916 /*
2928 2917 * Pause and idle threads always have
2929 2918 * t_state set to TS_ONPROC.
2930 2919 */
2931 2920 t->t_state = TS_FREE;
2932 2921 t->t_prev = NULL; /* Just in case */
2933 2922 }
2934 2923
2935 2924 } while ((t = tnext) != curthread);
2936 2925
2937 2926 mutex_exit(&pidlock);
2938 2927
2939 2928 mutex_sync();
2940 2929 for (t = tlist; t != NULL; t = tnext) {
2941 2930 tnext = t->t_next;
2942 2931 thread_free(t);
2943 2932 }
2944 2933 }
2945 2934 }
2946 2935
2947 2936 /*
2948 2937 * Update the cpu_supp_freqs of this cpu. This information is returned
2949 2938 * as part of cpu_info kstats. If the cpu_info_kstat exists already, then
2950 2939 * maintain the kstat data size.
2951 2940 */
2952 2941 void
2953 2942 cpu_set_supp_freqs(cpu_t *cp, const char *freqs)
2954 2943 {
2955 2944 char clkstr[sizeof ("18446744073709551615") + 1]; /* ui64 MAX */
2956 2945 const char *lfreqs = clkstr;
2957 2946 boolean_t kstat_exists = B_FALSE;
2958 2947 kstat_t *ksp;
2959 2948 size_t len;
2960 2949
2961 2950 /*
2962 2951 * A NULL pointer means we only support one speed.
2963 2952 */
2964 2953 if (freqs == NULL)
2965 2954 (void) snprintf(clkstr, sizeof (clkstr), "%"PRIu64,
2966 2955 cp->cpu_curr_clock);
2967 2956 else
2968 2957 lfreqs = freqs;
2969 2958
2970 2959 /*
2971 2960 * Make sure the frequency doesn't change while a snapshot is
2972 2961 * going on. Of course, we only need to worry about this if
2973 2962 * the kstat exists.
2974 2963 */
2975 2964 if ((ksp = cp->cpu_info_kstat) != NULL) {
2976 2965 mutex_enter(ksp->ks_lock);
2977 2966 kstat_exists = B_TRUE;
2978 2967 }
2979 2968
2980 2969 /*
2981 2970 * Free any previously allocated string and if the kstat
2982 2971 * already exists, then update its data size.
2983 2972 */
2984 2973 if (cp->cpu_supp_freqs != NULL) {
2985 2974 len = strlen(cp->cpu_supp_freqs) + 1;
2986 2975 kmem_free(cp->cpu_supp_freqs, len);
2987 2976 if (kstat_exists)
2988 2977 ksp->ks_data_size -= len;
2989 2978 }
2990 2979
2991 2980 /*
2992 2981 * Allocate the new string and set the pointer.
2993 2982 */
2994 2983 len = strlen(lfreqs) + 1;
2995 2984 cp->cpu_supp_freqs = kmem_alloc(len, KM_SLEEP);
2996 2985 (void) strcpy(cp->cpu_supp_freqs, lfreqs);
2997 2986
2998 2987 /*
2999 2988 * If the kstat already exists then update the data size and
3000 2989 * free the lock.
3001 2990 */
3002 2991 if (kstat_exists) {
3003 2992 ksp->ks_data_size += len;
3004 2993 mutex_exit(ksp->ks_lock);
3005 2994 }
3006 2995 }
3007 2996
3008 2997 /*
3009 2998 * Indicate the current CPU's clock freqency (in Hz).
3010 2999 * The calling context must be such that CPU references are safe.
3011 3000 */
3012 3001 void
3013 3002 cpu_set_curr_clock(uint64_t new_clk)
3014 3003 {
3015 3004 uint64_t old_clk;
3016 3005
3017 3006 old_clk = CPU->cpu_curr_clock;
3018 3007 CPU->cpu_curr_clock = new_clk;
3019 3008
3020 3009 /*
3021 3010 * The cpu-change-speed DTrace probe exports the frequency in Hz
3022 3011 */
3023 3012 DTRACE_PROBE3(cpu__change__speed, processorid_t, CPU->cpu_id,
3024 3013 uint64_t, old_clk, uint64_t, new_clk);
3025 3014 }
3026 3015
3027 3016 /*
3028 3017 * processor_info(2) and p_online(2) status support functions
3029 3018 * The constants returned by the cpu_get_state() and cpu_get_state_str() are
3030 3019 * for use in communicating processor state information to userland. Kernel
3031 3020 * subsystems should only be using the cpu_flags value directly. Subsystems
3032 3021 * modifying cpu_flags should record the state change via a call to the
3033 3022 * cpu_set_state().
3034 3023 */
3035 3024
3036 3025 /*
3037 3026 * Update the pi_state of this CPU. This function provides the CPU status for
3038 3027 * the information returned by processor_info(2).
3039 3028 */
3040 3029 void
3041 3030 cpu_set_state(cpu_t *cpu)
3042 3031 {
3043 3032 ASSERT(MUTEX_HELD(&cpu_lock));
3044 3033 cpu->cpu_type_info.pi_state = cpu_get_state(cpu);
3045 3034 cpu->cpu_state_begin = gethrestime_sec();
3046 3035 pool_cpu_mod = gethrtime();
3047 3036 }
3048 3037
3049 3038 /*
3050 3039 * Return offline/online/other status for the indicated CPU. Use only for
3051 3040 * communication with user applications; cpu_flags provides the in-kernel
3052 3041 * interface.
3053 3042 */
3054 3043 int
3055 3044 cpu_get_state(cpu_t *cpu)
3056 3045 {
3057 3046 ASSERT(MUTEX_HELD(&cpu_lock));
3058 3047 if (cpu->cpu_flags & CPU_POWEROFF)
3059 3048 return (P_POWEROFF);
3060 3049 else if (cpu->cpu_flags & CPU_FAULTED)
3061 3050 return (P_FAULTED);
3062 3051 else if (cpu->cpu_flags & CPU_SPARE)
3063 3052 return (P_SPARE);
3064 3053 else if ((cpu->cpu_flags & (CPU_READY | CPU_OFFLINE)) != CPU_READY)
3065 3054 return (P_OFFLINE);
3066 3055 else if (cpu->cpu_flags & CPU_ENABLE)
3067 3056 return (P_ONLINE);
3068 3057 else
3069 3058 return (P_NOINTR);
3070 3059 }
3071 3060
3072 3061 /*
3073 3062 * Return processor_info(2) state as a string.
3074 3063 */
3075 3064 const char *
3076 3065 cpu_get_state_str(cpu_t *cpu)
3077 3066 {
3078 3067 const char *string;
3079 3068
3080 3069 switch (cpu_get_state(cpu)) {
3081 3070 case P_ONLINE:
3082 3071 string = PS_ONLINE;
3083 3072 break;
3084 3073 case P_POWEROFF:
3085 3074 string = PS_POWEROFF;
3086 3075 break;
3087 3076 case P_NOINTR:
3088 3077 string = PS_NOINTR;
3089 3078 break;
3090 3079 case P_SPARE:
3091 3080 string = PS_SPARE;
3092 3081 break;
3093 3082 case P_FAULTED:
3094 3083 string = PS_FAULTED;
3095 3084 break;
3096 3085 case P_OFFLINE:
3097 3086 string = PS_OFFLINE;
3098 3087 break;
3099 3088 default:
3100 3089 string = "unknown";
3101 3090 break;
3102 3091 }
3103 3092 return (string);
3104 3093 }
3105 3094
3106 3095 /*
3107 3096 * Export this CPU's statistics (cpu_stat_t and cpu_stats_t) as raw and named
3108 3097 * kstats, respectively. This is done when a CPU is initialized or placed
3109 3098 * online via p_online(2).
3110 3099 */
3111 3100 static void
3112 3101 cpu_stats_kstat_create(cpu_t *cp)
3113 3102 {
3114 3103 int instance = cp->cpu_id;
3115 3104 char *module = "cpu";
3116 3105 char *class = "misc";
3117 3106 kstat_t *ksp;
3118 3107 zoneid_t zoneid;
3119 3108
3120 3109 ASSERT(MUTEX_HELD(&cpu_lock));
3121 3110
3122 3111 if (pool_pset_enabled())
3123 3112 zoneid = GLOBAL_ZONEID;
3124 3113 else
3125 3114 zoneid = ALL_ZONES;
3126 3115 /*
3127 3116 * Create named kstats
3128 3117 */
3129 3118 #define CPU_STATS_KS_CREATE(name, tsize, update_func) \
3130 3119 ksp = kstat_create_zone(module, instance, (name), class, \
3131 3120 KSTAT_TYPE_NAMED, (tsize) / sizeof (kstat_named_t), 0, \
3132 3121 zoneid); \
3133 3122 if (ksp != NULL) { \
3134 3123 ksp->ks_private = cp; \
3135 3124 ksp->ks_update = (update_func); \
3136 3125 kstat_install(ksp); \
3137 3126 } else \
3138 3127 cmn_err(CE_WARN, "cpu: unable to create %s:%d:%s kstat", \
3139 3128 module, instance, (name));
3140 3129
3141 3130 CPU_STATS_KS_CREATE("sys", sizeof (cpu_sys_stats_ks_data_template),
3142 3131 cpu_sys_stats_ks_update);
3143 3132 CPU_STATS_KS_CREATE("vm", sizeof (cpu_vm_stats_ks_data_template),
3144 3133 cpu_vm_stats_ks_update);
3145 3134
3146 3135 /*
3147 3136 * Export the familiar cpu_stat_t KSTAT_TYPE_RAW kstat.
3148 3137 */
3149 3138 ksp = kstat_create_zone("cpu_stat", cp->cpu_id, NULL,
3150 3139 "misc", KSTAT_TYPE_RAW, sizeof (cpu_stat_t), 0, zoneid);
3151 3140 if (ksp != NULL) {
3152 3141 ksp->ks_update = cpu_stat_ks_update;
3153 3142 ksp->ks_private = cp;
3154 3143 kstat_install(ksp);
3155 3144 }
3156 3145 }
3157 3146
3158 3147 static void
3159 3148 cpu_stats_kstat_destroy(cpu_t *cp)
3160 3149 {
3161 3150 char ks_name[KSTAT_STRLEN];
3162 3151
3163 3152 (void) sprintf(ks_name, "cpu_stat%d", cp->cpu_id);
3164 3153 kstat_delete_byname("cpu_stat", cp->cpu_id, ks_name);
3165 3154
3166 3155 kstat_delete_byname("cpu", cp->cpu_id, "sys");
3167 3156 kstat_delete_byname("cpu", cp->cpu_id, "vm");
3168 3157 }
3169 3158
3170 3159 static int
3171 3160 cpu_sys_stats_ks_update(kstat_t *ksp, int rw)
3172 3161 {
3173 3162 cpu_t *cp = (cpu_t *)ksp->ks_private;
3174 3163 struct cpu_sys_stats_ks_data *csskd;
3175 3164 cpu_sys_stats_t *css;
3176 3165 hrtime_t msnsecs[NCMSTATES];
3177 3166 int i;
3178 3167
3179 3168 if (rw == KSTAT_WRITE)
3180 3169 return (EACCES);
3181 3170
3182 3171 csskd = ksp->ks_data;
3183 3172 css = &cp->cpu_stats.sys;
3184 3173
3185 3174 /*
3186 3175 * Read CPU mstate, but compare with the last values we
3187 3176 * received to make sure that the returned kstats never
3188 3177 * decrease.
3189 3178 */
3190 3179
3191 3180 get_cpu_mstate(cp, msnsecs);
3192 3181 if (csskd->cpu_nsec_idle.value.ui64 > msnsecs[CMS_IDLE])
3193 3182 msnsecs[CMS_IDLE] = csskd->cpu_nsec_idle.value.ui64;
3194 3183 if (csskd->cpu_nsec_user.value.ui64 > msnsecs[CMS_USER])
3195 3184 msnsecs[CMS_USER] = csskd->cpu_nsec_user.value.ui64;
3196 3185 if (csskd->cpu_nsec_kernel.value.ui64 > msnsecs[CMS_SYSTEM])
3197 3186 msnsecs[CMS_SYSTEM] = csskd->cpu_nsec_kernel.value.ui64;
3198 3187
3199 3188 bcopy(&cpu_sys_stats_ks_data_template, ksp->ks_data,
3200 3189 sizeof (cpu_sys_stats_ks_data_template));
3201 3190
3202 3191 csskd->cpu_ticks_wait.value.ui64 = 0;
3203 3192 csskd->wait_ticks_io.value.ui64 = 0;
3204 3193
3205 3194 csskd->cpu_nsec_idle.value.ui64 = msnsecs[CMS_IDLE];
3206 3195 csskd->cpu_nsec_user.value.ui64 = msnsecs[CMS_USER];
3207 3196 csskd->cpu_nsec_kernel.value.ui64 = msnsecs[CMS_SYSTEM];
3208 3197 csskd->cpu_ticks_idle.value.ui64 =
3209 3198 NSEC_TO_TICK(csskd->cpu_nsec_idle.value.ui64);
3210 3199 csskd->cpu_ticks_user.value.ui64 =
3211 3200 NSEC_TO_TICK(csskd->cpu_nsec_user.value.ui64);
3212 3201 csskd->cpu_ticks_kernel.value.ui64 =
3213 3202 NSEC_TO_TICK(csskd->cpu_nsec_kernel.value.ui64);
3214 3203 csskd->cpu_nsec_dtrace.value.ui64 = cp->cpu_dtrace_nsec;
3215 3204 csskd->dtrace_probes.value.ui64 = cp->cpu_dtrace_probes;
3216 3205 csskd->cpu_nsec_intr.value.ui64 = cp->cpu_intrlast;
3217 3206 csskd->cpu_load_intr.value.ui64 = cp->cpu_intrload;
3218 3207 csskd->bread.value.ui64 = css->bread;
3219 3208 csskd->bwrite.value.ui64 = css->bwrite;
3220 3209 csskd->lread.value.ui64 = css->lread;
3221 3210 csskd->lwrite.value.ui64 = css->lwrite;
3222 3211 csskd->phread.value.ui64 = css->phread;
3223 3212 csskd->phwrite.value.ui64 = css->phwrite;
3224 3213 csskd->pswitch.value.ui64 = css->pswitch;
3225 3214 csskd->trap.value.ui64 = css->trap;
3226 3215 csskd->intr.value.ui64 = 0;
3227 3216 for (i = 0; i < PIL_MAX; i++)
3228 3217 csskd->intr.value.ui64 += css->intr[i];
3229 3218 csskd->syscall.value.ui64 = css->syscall;
3230 3219 csskd->sysread.value.ui64 = css->sysread;
3231 3220 csskd->syswrite.value.ui64 = css->syswrite;
3232 3221 csskd->sysfork.value.ui64 = css->sysfork;
3233 3222 csskd->sysvfork.value.ui64 = css->sysvfork;
3234 3223 csskd->sysexec.value.ui64 = css->sysexec;
3235 3224 csskd->readch.value.ui64 = css->readch;
3236 3225 csskd->writech.value.ui64 = css->writech;
3237 3226 csskd->rcvint.value.ui64 = css->rcvint;
3238 3227 csskd->xmtint.value.ui64 = css->xmtint;
3239 3228 csskd->mdmint.value.ui64 = css->mdmint;
3240 3229 csskd->rawch.value.ui64 = css->rawch;
3241 3230 csskd->canch.value.ui64 = css->canch;
3242 3231 csskd->outch.value.ui64 = css->outch;
3243 3232 csskd->msg.value.ui64 = css->msg;
3244 3233 csskd->sema.value.ui64 = css->sema;
3245 3234 csskd->namei.value.ui64 = css->namei;
3246 3235 csskd->ufsiget.value.ui64 = css->ufsiget;
3247 3236 csskd->ufsdirblk.value.ui64 = css->ufsdirblk;
3248 3237 csskd->ufsipage.value.ui64 = css->ufsipage;
3249 3238 csskd->ufsinopage.value.ui64 = css->ufsinopage;
3250 3239 csskd->procovf.value.ui64 = css->procovf;
3251 3240 csskd->intrthread.value.ui64 = 0;
3252 3241 for (i = 0; i < LOCK_LEVEL - 1; i++)
3253 3242 csskd->intrthread.value.ui64 += css->intr[i];
3254 3243 csskd->intrblk.value.ui64 = css->intrblk;
3255 3244 csskd->intrunpin.value.ui64 = css->intrunpin;
3256 3245 csskd->idlethread.value.ui64 = css->idlethread;
3257 3246 csskd->inv_swtch.value.ui64 = css->inv_swtch;
3258 3247 csskd->nthreads.value.ui64 = css->nthreads;
3259 3248 csskd->cpumigrate.value.ui64 = css->cpumigrate;
3260 3249 csskd->xcalls.value.ui64 = css->xcalls;
3261 3250 csskd->mutex_adenters.value.ui64 = css->mutex_adenters;
3262 3251 csskd->rw_rdfails.value.ui64 = css->rw_rdfails;
3263 3252 csskd->rw_wrfails.value.ui64 = css->rw_wrfails;
3264 3253 csskd->modload.value.ui64 = css->modload;
3265 3254 csskd->modunload.value.ui64 = css->modunload;
3266 3255 csskd->bawrite.value.ui64 = css->bawrite;
3267 3256 csskd->iowait.value.ui64 = css->iowait;
3268 3257
3269 3258 return (0);
3270 3259 }
3271 3260
3272 3261 static int
3273 3262 cpu_vm_stats_ks_update(kstat_t *ksp, int rw)
3274 3263 {
3275 3264 cpu_t *cp = (cpu_t *)ksp->ks_private;
3276 3265 struct cpu_vm_stats_ks_data *cvskd;
3277 3266 cpu_vm_stats_t *cvs;
3278 3267
3279 3268 if (rw == KSTAT_WRITE)
3280 3269 return (EACCES);
3281 3270
3282 3271 cvs = &cp->cpu_stats.vm;
↓ open down ↓ |
582 lines elided |
↑ open up ↑ |
3283 3272 cvskd = ksp->ks_data;
3284 3273
3285 3274 bcopy(&cpu_vm_stats_ks_data_template, ksp->ks_data,
3286 3275 sizeof (cpu_vm_stats_ks_data_template));
3287 3276 cvskd->pgrec.value.ui64 = cvs->pgrec;
3288 3277 cvskd->pgfrec.value.ui64 = cvs->pgfrec;
3289 3278 cvskd->pgin.value.ui64 = cvs->pgin;
3290 3279 cvskd->pgpgin.value.ui64 = cvs->pgpgin;
3291 3280 cvskd->pgout.value.ui64 = cvs->pgout;
3292 3281 cvskd->pgpgout.value.ui64 = cvs->pgpgout;
3293 - cvskd->swapin.value.ui64 = cvs->swapin;
3294 - cvskd->pgswapin.value.ui64 = cvs->pgswapin;
3295 - cvskd->swapout.value.ui64 = cvs->swapout;
3296 - cvskd->pgswapout.value.ui64 = cvs->pgswapout;
3297 3282 cvskd->zfod.value.ui64 = cvs->zfod;
3298 3283 cvskd->dfree.value.ui64 = cvs->dfree;
3299 3284 cvskd->scan.value.ui64 = cvs->scan;
3300 3285 cvskd->rev.value.ui64 = cvs->rev;
3301 3286 cvskd->hat_fault.value.ui64 = cvs->hat_fault;
3302 3287 cvskd->as_fault.value.ui64 = cvs->as_fault;
3303 3288 cvskd->maj_fault.value.ui64 = cvs->maj_fault;
3304 3289 cvskd->cow_fault.value.ui64 = cvs->cow_fault;
3305 3290 cvskd->prot_fault.value.ui64 = cvs->prot_fault;
3306 3291 cvskd->softlock.value.ui64 = cvs->softlock;
3307 3292 cvskd->kernel_asflt.value.ui64 = cvs->kernel_asflt;
3308 3293 cvskd->pgrrun.value.ui64 = cvs->pgrrun;
3309 3294 cvskd->execpgin.value.ui64 = cvs->execpgin;
3310 3295 cvskd->execpgout.value.ui64 = cvs->execpgout;
3311 3296 cvskd->execfree.value.ui64 = cvs->execfree;
3312 3297 cvskd->anonpgin.value.ui64 = cvs->anonpgin;
3313 3298 cvskd->anonpgout.value.ui64 = cvs->anonpgout;
3314 3299 cvskd->anonfree.value.ui64 = cvs->anonfree;
3315 3300 cvskd->fspgin.value.ui64 = cvs->fspgin;
3316 3301 cvskd->fspgout.value.ui64 = cvs->fspgout;
3317 3302 cvskd->fsfree.value.ui64 = cvs->fsfree;
3318 3303
3319 3304 return (0);
3320 3305 }
3321 3306
3322 3307 static int
3323 3308 cpu_stat_ks_update(kstat_t *ksp, int rw)
3324 3309 {
3325 3310 cpu_stat_t *cso;
3326 3311 cpu_t *cp;
3327 3312 int i;
3328 3313 hrtime_t msnsecs[NCMSTATES];
3329 3314
3330 3315 cso = (cpu_stat_t *)ksp->ks_data;
3331 3316 cp = (cpu_t *)ksp->ks_private;
3332 3317
3333 3318 if (rw == KSTAT_WRITE)
3334 3319 return (EACCES);
3335 3320
3336 3321 /*
3337 3322 * Read CPU mstate, but compare with the last values we
3338 3323 * received to make sure that the returned kstats never
3339 3324 * decrease.
3340 3325 */
3341 3326
3342 3327 get_cpu_mstate(cp, msnsecs);
3343 3328 msnsecs[CMS_IDLE] = NSEC_TO_TICK(msnsecs[CMS_IDLE]);
3344 3329 msnsecs[CMS_USER] = NSEC_TO_TICK(msnsecs[CMS_USER]);
3345 3330 msnsecs[CMS_SYSTEM] = NSEC_TO_TICK(msnsecs[CMS_SYSTEM]);
3346 3331 if (cso->cpu_sysinfo.cpu[CPU_IDLE] < msnsecs[CMS_IDLE])
3347 3332 cso->cpu_sysinfo.cpu[CPU_IDLE] = msnsecs[CMS_IDLE];
3348 3333 if (cso->cpu_sysinfo.cpu[CPU_USER] < msnsecs[CMS_USER])
3349 3334 cso->cpu_sysinfo.cpu[CPU_USER] = msnsecs[CMS_USER];
3350 3335 if (cso->cpu_sysinfo.cpu[CPU_KERNEL] < msnsecs[CMS_SYSTEM])
3351 3336 cso->cpu_sysinfo.cpu[CPU_KERNEL] = msnsecs[CMS_SYSTEM];
3352 3337 cso->cpu_sysinfo.cpu[CPU_WAIT] = 0;
3353 3338 cso->cpu_sysinfo.wait[W_IO] = 0;
3354 3339 cso->cpu_sysinfo.wait[W_SWAP] = 0;
3355 3340 cso->cpu_sysinfo.wait[W_PIO] = 0;
3356 3341 cso->cpu_sysinfo.bread = CPU_STATS(cp, sys.bread);
3357 3342 cso->cpu_sysinfo.bwrite = CPU_STATS(cp, sys.bwrite);
3358 3343 cso->cpu_sysinfo.lread = CPU_STATS(cp, sys.lread);
3359 3344 cso->cpu_sysinfo.lwrite = CPU_STATS(cp, sys.lwrite);
3360 3345 cso->cpu_sysinfo.phread = CPU_STATS(cp, sys.phread);
3361 3346 cso->cpu_sysinfo.phwrite = CPU_STATS(cp, sys.phwrite);
3362 3347 cso->cpu_sysinfo.pswitch = CPU_STATS(cp, sys.pswitch);
3363 3348 cso->cpu_sysinfo.trap = CPU_STATS(cp, sys.trap);
3364 3349 cso->cpu_sysinfo.intr = 0;
3365 3350 for (i = 0; i < PIL_MAX; i++)
3366 3351 cso->cpu_sysinfo.intr += CPU_STATS(cp, sys.intr[i]);
3367 3352 cso->cpu_sysinfo.syscall = CPU_STATS(cp, sys.syscall);
3368 3353 cso->cpu_sysinfo.sysread = CPU_STATS(cp, sys.sysread);
3369 3354 cso->cpu_sysinfo.syswrite = CPU_STATS(cp, sys.syswrite);
3370 3355 cso->cpu_sysinfo.sysfork = CPU_STATS(cp, sys.sysfork);
3371 3356 cso->cpu_sysinfo.sysvfork = CPU_STATS(cp, sys.sysvfork);
3372 3357 cso->cpu_sysinfo.sysexec = CPU_STATS(cp, sys.sysexec);
3373 3358 cso->cpu_sysinfo.readch = CPU_STATS(cp, sys.readch);
3374 3359 cso->cpu_sysinfo.writech = CPU_STATS(cp, sys.writech);
3375 3360 cso->cpu_sysinfo.rcvint = CPU_STATS(cp, sys.rcvint);
3376 3361 cso->cpu_sysinfo.xmtint = CPU_STATS(cp, sys.xmtint);
3377 3362 cso->cpu_sysinfo.mdmint = CPU_STATS(cp, sys.mdmint);
3378 3363 cso->cpu_sysinfo.rawch = CPU_STATS(cp, sys.rawch);
3379 3364 cso->cpu_sysinfo.canch = CPU_STATS(cp, sys.canch);
3380 3365 cso->cpu_sysinfo.outch = CPU_STATS(cp, sys.outch);
3381 3366 cso->cpu_sysinfo.msg = CPU_STATS(cp, sys.msg);
3382 3367 cso->cpu_sysinfo.sema = CPU_STATS(cp, sys.sema);
3383 3368 cso->cpu_sysinfo.namei = CPU_STATS(cp, sys.namei);
3384 3369 cso->cpu_sysinfo.ufsiget = CPU_STATS(cp, sys.ufsiget);
3385 3370 cso->cpu_sysinfo.ufsdirblk = CPU_STATS(cp, sys.ufsdirblk);
3386 3371 cso->cpu_sysinfo.ufsipage = CPU_STATS(cp, sys.ufsipage);
3387 3372 cso->cpu_sysinfo.ufsinopage = CPU_STATS(cp, sys.ufsinopage);
3388 3373 cso->cpu_sysinfo.inodeovf = 0;
3389 3374 cso->cpu_sysinfo.fileovf = 0;
3390 3375 cso->cpu_sysinfo.procovf = CPU_STATS(cp, sys.procovf);
3391 3376 cso->cpu_sysinfo.intrthread = 0;
3392 3377 for (i = 0; i < LOCK_LEVEL - 1; i++)
3393 3378 cso->cpu_sysinfo.intrthread += CPU_STATS(cp, sys.intr[i]);
3394 3379 cso->cpu_sysinfo.intrblk = CPU_STATS(cp, sys.intrblk);
3395 3380 cso->cpu_sysinfo.idlethread = CPU_STATS(cp, sys.idlethread);
3396 3381 cso->cpu_sysinfo.inv_swtch = CPU_STATS(cp, sys.inv_swtch);
3397 3382 cso->cpu_sysinfo.nthreads = CPU_STATS(cp, sys.nthreads);
3398 3383 cso->cpu_sysinfo.cpumigrate = CPU_STATS(cp, sys.cpumigrate);
3399 3384 cso->cpu_sysinfo.xcalls = CPU_STATS(cp, sys.xcalls);
3400 3385 cso->cpu_sysinfo.mutex_adenters = CPU_STATS(cp, sys.mutex_adenters);
3401 3386 cso->cpu_sysinfo.rw_rdfails = CPU_STATS(cp, sys.rw_rdfails);
3402 3387 cso->cpu_sysinfo.rw_wrfails = CPU_STATS(cp, sys.rw_wrfails);
3403 3388 cso->cpu_sysinfo.modload = CPU_STATS(cp, sys.modload);
3404 3389 cso->cpu_sysinfo.modunload = CPU_STATS(cp, sys.modunload);
3405 3390 cso->cpu_sysinfo.bawrite = CPU_STATS(cp, sys.bawrite);
3406 3391 cso->cpu_sysinfo.rw_enters = 0;
3407 3392 cso->cpu_sysinfo.win_uo_cnt = 0;
3408 3393 cso->cpu_sysinfo.win_uu_cnt = 0;
3409 3394 cso->cpu_sysinfo.win_so_cnt = 0;
3410 3395 cso->cpu_sysinfo.win_su_cnt = 0;
3411 3396 cso->cpu_sysinfo.win_suo_cnt = 0;
3412 3397
↓ open down ↓ |
106 lines elided |
↑ open up ↑ |
3413 3398 cso->cpu_syswait.iowait = CPU_STATS(cp, sys.iowait);
3414 3399 cso->cpu_syswait.swap = 0;
3415 3400 cso->cpu_syswait.physio = 0;
3416 3401
3417 3402 cso->cpu_vminfo.pgrec = CPU_STATS(cp, vm.pgrec);
3418 3403 cso->cpu_vminfo.pgfrec = CPU_STATS(cp, vm.pgfrec);
3419 3404 cso->cpu_vminfo.pgin = CPU_STATS(cp, vm.pgin);
3420 3405 cso->cpu_vminfo.pgpgin = CPU_STATS(cp, vm.pgpgin);
3421 3406 cso->cpu_vminfo.pgout = CPU_STATS(cp, vm.pgout);
3422 3407 cso->cpu_vminfo.pgpgout = CPU_STATS(cp, vm.pgpgout);
3423 - cso->cpu_vminfo.swapin = CPU_STATS(cp, vm.swapin);
3424 - cso->cpu_vminfo.pgswapin = CPU_STATS(cp, vm.pgswapin);
3425 - cso->cpu_vminfo.swapout = CPU_STATS(cp, vm.swapout);
3426 - cso->cpu_vminfo.pgswapout = CPU_STATS(cp, vm.pgswapout);
3427 3408 cso->cpu_vminfo.zfod = CPU_STATS(cp, vm.zfod);
3428 3409 cso->cpu_vminfo.dfree = CPU_STATS(cp, vm.dfree);
3429 3410 cso->cpu_vminfo.scan = CPU_STATS(cp, vm.scan);
3430 3411 cso->cpu_vminfo.rev = CPU_STATS(cp, vm.rev);
3431 3412 cso->cpu_vminfo.hat_fault = CPU_STATS(cp, vm.hat_fault);
3432 3413 cso->cpu_vminfo.as_fault = CPU_STATS(cp, vm.as_fault);
3433 3414 cso->cpu_vminfo.maj_fault = CPU_STATS(cp, vm.maj_fault);
3434 3415 cso->cpu_vminfo.cow_fault = CPU_STATS(cp, vm.cow_fault);
3435 3416 cso->cpu_vminfo.prot_fault = CPU_STATS(cp, vm.prot_fault);
3436 3417 cso->cpu_vminfo.softlock = CPU_STATS(cp, vm.softlock);
3437 3418 cso->cpu_vminfo.kernel_asflt = CPU_STATS(cp, vm.kernel_asflt);
3438 3419 cso->cpu_vminfo.pgrrun = CPU_STATS(cp, vm.pgrrun);
3439 3420 cso->cpu_vminfo.execpgin = CPU_STATS(cp, vm.execpgin);
3440 3421 cso->cpu_vminfo.execpgout = CPU_STATS(cp, vm.execpgout);
3441 3422 cso->cpu_vminfo.execfree = CPU_STATS(cp, vm.execfree);
3442 3423 cso->cpu_vminfo.anonpgin = CPU_STATS(cp, vm.anonpgin);
3443 3424 cso->cpu_vminfo.anonpgout = CPU_STATS(cp, vm.anonpgout);
3444 3425 cso->cpu_vminfo.anonfree = CPU_STATS(cp, vm.anonfree);
3445 3426 cso->cpu_vminfo.fspgin = CPU_STATS(cp, vm.fspgin);
3446 3427 cso->cpu_vminfo.fspgout = CPU_STATS(cp, vm.fspgout);
3447 3428 cso->cpu_vminfo.fsfree = CPU_STATS(cp, vm.fsfree);
3448 3429
3449 3430 return (0);
3450 3431 }
↓ open down ↓ |
14 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX