Print this page
re #13613 rb4516 Tunables needs volatile keyword
| Split |
Close |
| Expand all |
| Collapse all |
--- old/usr/src/uts/common/disp/disp.c
+++ new/usr/src/uts/common/disp/disp.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.
|
↓ open down ↓ |
14 lines elided |
↑ open up ↑ |
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 2009 Sun Microsystems, Inc. All rights reserved.
23 23 * Use is subject to license terms.
24 24 */
25 +/*
26 + * Copyright 2013 Nexenta Systems, Inc. All rights reserved.
27 + */
25 28
26 29 /* Copyright (c) 1984, 1986, 1987, 1988, 1989 AT&T */
27 30 /* All Rights Reserved */
28 31
29 32
30 33 #include <sys/types.h>
31 34 #include <sys/param.h>
32 35 #include <sys/sysmacros.h>
33 36 #include <sys/signal.h>
34 37 #include <sys/user.h>
35 38 #include <sys/systm.h>
36 39 #include <sys/sysinfo.h>
37 40 #include <sys/var.h>
38 41 #include <sys/errno.h>
39 42 #include <sys/cmn_err.h>
40 43 #include <sys/debug.h>
41 44 #include <sys/inline.h>
42 45 #include <sys/disp.h>
43 46 #include <sys/class.h>
44 47 #include <sys/bitmap.h>
45 48 #include <sys/kmem.h>
46 49 #include <sys/cpuvar.h>
47 50 #include <sys/vtrace.h>
48 51 #include <sys/tnf.h>
49 52 #include <sys/cpupart.h>
50 53 #include <sys/lgrp.h>
51 54 #include <sys/pg.h>
52 55 #include <sys/cmt.h>
53 56 #include <sys/bitset.h>
54 57 #include <sys/schedctl.h>
55 58 #include <sys/atomic.h>
56 59 #include <sys/dtrace.h>
57 60 #include <sys/sdt.h>
58 61 #include <sys/archsystm.h>
59 62
60 63 #include <vm/as.h>
61 64
62 65 #define BOUND_CPU 0x1
63 66 #define BOUND_PARTITION 0x2
64 67 #define BOUND_INTR 0x4
65 68
66 69 /* Dispatch queue allocation structure and functions */
67 70 struct disp_queue_info {
68 71 disp_t *dp;
69 72 dispq_t *olddispq;
70 73 dispq_t *newdispq;
71 74 ulong_t *olddqactmap;
72 75 ulong_t *newdqactmap;
73 76 int oldnglobpris;
74 77 };
75 78 static void disp_dq_alloc(struct disp_queue_info *dptr, int numpris,
76 79 disp_t *dp);
77 80 static void disp_dq_assign(struct disp_queue_info *dptr, int numpris);
78 81 static void disp_dq_free(struct disp_queue_info *dptr);
79 82
80 83 /* platform-specific routine to call when processor is idle */
81 84 static void generic_idle_cpu();
82 85 void (*idle_cpu)() = generic_idle_cpu;
83 86
84 87 /* routines invoked when a CPU enters/exits the idle loop */
85 88 static void idle_enter();
86 89 static void idle_exit();
87 90
88 91 /* platform-specific routine to call when thread is enqueued */
89 92 static void generic_enq_thread(cpu_t *, int);
90 93 void (*disp_enq_thread)(cpu_t *, int) = generic_enq_thread;
91 94
92 95 pri_t kpreemptpri; /* priority where kernel preemption applies */
93 96 pri_t upreemptpri = 0; /* priority where normal preemption applies */
94 97 pri_t intr_pri; /* interrupt thread priority base level */
95 98
96 99 #define KPQPRI -1 /* pri where cpu affinity is dropped for kpq */
97 100 pri_t kpqpri = KPQPRI; /* can be set in /etc/system */
98 101 disp_t cpu0_disp; /* boot CPU's dispatch queue */
99 102 disp_lock_t swapped_lock; /* lock swapped threads and swap queue */
100 103 int nswapped; /* total number of swapped threads */
101 104 void disp_swapped_enq(kthread_t *tp);
102 105 static void disp_swapped_setrun(kthread_t *tp);
103 106 static void cpu_resched(cpu_t *cp, pri_t tpri);
104 107
105 108 /*
106 109 * If this is set, only interrupt threads will cause kernel preemptions.
107 110 * This is done by changing the value of kpreemptpri. kpreemptpri
108 111 * will either be the max sysclass pri + 1 or the min interrupt pri.
109 112 */
110 113 int only_intr_kpreempt;
111 114
112 115 extern void set_idle_cpu(int cpun);
|
↓ open down ↓ |
78 lines elided |
↑ open up ↑ |
113 116 extern void unset_idle_cpu(int cpun);
114 117 static void setkpdq(kthread_t *tp, int borf);
115 118 #define SETKP_BACK 0
116 119 #define SETKP_FRONT 1
117 120 /*
118 121 * Parameter that determines how recently a thread must have run
119 122 * on the CPU to be considered loosely-bound to that CPU to reduce
120 123 * cold cache effects. The interval is in hertz.
121 124 */
122 125 #define RECHOOSE_INTERVAL 3
123 -int rechoose_interval = RECHOOSE_INTERVAL;
126 +volatile int rechoose_interval = RECHOOSE_INTERVAL;
124 127
125 128 /*
126 129 * Parameter that determines how long (in nanoseconds) a thread must
127 130 * be sitting on a run queue before it can be stolen by another CPU
128 131 * to reduce migrations. The interval is in nanoseconds.
129 132 *
130 133 * The nosteal_nsec should be set by platform code cmp_set_nosteal_interval()
131 134 * to an appropriate value. nosteal_nsec is set to NOSTEAL_UNINITIALIZED
132 135 * here indicating it is uninitiallized.
133 136 * Setting nosteal_nsec to 0 effectively disables the nosteal 'protection'.
134 137 *
135 138 */
136 139 #define NOSTEAL_UNINITIALIZED (-1)
137 140 hrtime_t nosteal_nsec = NOSTEAL_UNINITIALIZED;
138 141 extern void cmp_set_nosteal_interval(void);
139 142
140 143 id_t defaultcid; /* system "default" class; see dispadmin(1M) */
141 144
142 145 disp_lock_t transition_lock; /* lock on transitioning threads */
143 146 disp_lock_t stop_lock; /* lock on stopped threads */
144 147
145 148 static void cpu_dispqalloc(int numpris);
146 149
147 150 /*
148 151 * This gets returned by disp_getwork/disp_getbest if we couldn't steal
149 152 * a thread because it was sitting on its run queue for a very short
150 153 * period of time.
151 154 */
152 155 #define T_DONTSTEAL (kthread_t *)(-1) /* returned by disp_getwork/getbest */
153 156
154 157 static kthread_t *disp_getwork(cpu_t *to);
155 158 static kthread_t *disp_getbest(disp_t *from);
156 159 static kthread_t *disp_ratify(kthread_t *tp, disp_t *kpq);
157 160
158 161 void swtch_to(kthread_t *);
159 162
160 163 /*
161 164 * dispatcher and scheduler initialization
162 165 */
163 166
164 167 /*
165 168 * disp_setup - Common code to calculate and allocate dispatcher
166 169 * variables and structures based on the maximum priority.
167 170 */
168 171 static void
169 172 disp_setup(pri_t maxglobpri, pri_t oldnglobpris)
170 173 {
171 174 pri_t newnglobpris;
172 175
173 176 ASSERT(MUTEX_HELD(&cpu_lock));
174 177
175 178 newnglobpris = maxglobpri + 1 + LOCK_LEVEL;
176 179
177 180 if (newnglobpris > oldnglobpris) {
178 181 /*
179 182 * Allocate new kp queues for each CPU partition.
180 183 */
181 184 cpupart_kpqalloc(newnglobpris);
182 185
183 186 /*
184 187 * Allocate new dispatch queues for each CPU.
185 188 */
186 189 cpu_dispqalloc(newnglobpris);
187 190
188 191 /*
189 192 * compute new interrupt thread base priority
190 193 */
191 194 intr_pri = maxglobpri;
192 195 if (only_intr_kpreempt) {
193 196 kpreemptpri = intr_pri + 1;
194 197 if (kpqpri == KPQPRI)
195 198 kpqpri = kpreemptpri;
196 199 }
197 200 v.v_nglobpris = newnglobpris;
198 201 }
199 202 }
200 203
201 204 /*
202 205 * dispinit - Called to initialize all loaded classes and the
203 206 * dispatcher framework.
204 207 */
205 208 void
206 209 dispinit(void)
207 210 {
208 211 id_t cid;
209 212 pri_t maxglobpri;
210 213 pri_t cl_maxglobpri;
211 214
212 215 maxglobpri = -1;
213 216
214 217 /*
215 218 * Initialize transition lock, which will always be set.
216 219 */
217 220 DISP_LOCK_INIT(&transition_lock);
218 221 disp_lock_enter_high(&transition_lock);
219 222 DISP_LOCK_INIT(&stop_lock);
220 223
221 224 mutex_enter(&cpu_lock);
222 225 CPU->cpu_disp->disp_maxrunpri = -1;
223 226 CPU->cpu_disp->disp_max_unbound_pri = -1;
224 227
225 228 /*
226 229 * Initialize the default CPU partition.
227 230 */
228 231 cpupart_initialize_default();
229 232 /*
230 233 * Call the class specific initialization functions for
231 234 * all pre-installed schedulers.
232 235 *
233 236 * We pass the size of a class specific parameter
234 237 * buffer to each of the initialization functions
235 238 * to try to catch problems with backward compatibility
236 239 * of class modules.
237 240 *
238 241 * For example a new class module running on an old system
239 242 * which didn't provide sufficiently large parameter buffers
240 243 * would be bad news. Class initialization modules can check for
241 244 * this and take action if they detect a problem.
242 245 */
243 246
244 247 for (cid = 0; cid < nclass; cid++) {
245 248 sclass_t *sc;
246 249
247 250 sc = &sclass[cid];
248 251 if (SCHED_INSTALLED(sc)) {
249 252 cl_maxglobpri = sc->cl_init(cid, PC_CLPARMSZ,
250 253 &sc->cl_funcs);
251 254 if (cl_maxglobpri > maxglobpri)
252 255 maxglobpri = cl_maxglobpri;
253 256 }
254 257 }
255 258 kpreemptpri = (pri_t)v.v_maxsyspri + 1;
256 259 if (kpqpri == KPQPRI)
257 260 kpqpri = kpreemptpri;
258 261
259 262 ASSERT(maxglobpri >= 0);
260 263 disp_setup(maxglobpri, 0);
261 264
262 265 mutex_exit(&cpu_lock);
263 266
264 267 /*
265 268 * Platform specific sticky scheduler setup.
266 269 */
267 270 if (nosteal_nsec == NOSTEAL_UNINITIALIZED)
268 271 cmp_set_nosteal_interval();
269 272
270 273 /*
271 274 * Get the default class ID; this may be later modified via
272 275 * dispadmin(1M). This will load the class (normally TS) and that will
273 276 * call disp_add(), which is why we had to drop cpu_lock first.
274 277 */
275 278 if (getcid(defaultclass, &defaultcid) != 0) {
276 279 cmn_err(CE_PANIC, "Couldn't load default scheduling class '%s'",
277 280 defaultclass);
278 281 }
279 282 }
280 283
281 284 /*
282 285 * disp_add - Called with class pointer to initialize the dispatcher
283 286 * for a newly loaded class.
284 287 */
285 288 void
286 289 disp_add(sclass_t *clp)
287 290 {
288 291 pri_t maxglobpri;
289 292 pri_t cl_maxglobpri;
290 293
291 294 mutex_enter(&cpu_lock);
292 295 /*
293 296 * Initialize the scheduler class.
294 297 */
295 298 maxglobpri = (pri_t)(v.v_nglobpris - LOCK_LEVEL - 1);
296 299 cl_maxglobpri = clp->cl_init(clp - sclass, PC_CLPARMSZ, &clp->cl_funcs);
297 300 if (cl_maxglobpri > maxglobpri)
298 301 maxglobpri = cl_maxglobpri;
299 302
300 303 /*
301 304 * Save old queue information. Since we're initializing a
302 305 * new scheduling class which has just been loaded, then
303 306 * the size of the dispq may have changed. We need to handle
304 307 * that here.
305 308 */
306 309 disp_setup(maxglobpri, v.v_nglobpris);
307 310
308 311 mutex_exit(&cpu_lock);
309 312 }
310 313
311 314
312 315 /*
313 316 * For each CPU, allocate new dispatch queues
314 317 * with the stated number of priorities.
315 318 */
316 319 static void
317 320 cpu_dispqalloc(int numpris)
318 321 {
319 322 cpu_t *cpup;
320 323 struct disp_queue_info *disp_mem;
321 324 int i, num;
322 325
323 326 ASSERT(MUTEX_HELD(&cpu_lock));
324 327
325 328 disp_mem = kmem_zalloc(NCPU *
326 329 sizeof (struct disp_queue_info), KM_SLEEP);
327 330
328 331 /*
329 332 * This routine must allocate all of the memory before stopping
330 333 * the cpus because it must not sleep in kmem_alloc while the
331 334 * CPUs are stopped. Locks they hold will not be freed until they
332 335 * are restarted.
333 336 */
334 337 i = 0;
335 338 cpup = cpu_list;
336 339 do {
337 340 disp_dq_alloc(&disp_mem[i], numpris, cpup->cpu_disp);
338 341 i++;
339 342 cpup = cpup->cpu_next;
340 343 } while (cpup != cpu_list);
341 344 num = i;
342 345
343 346 pause_cpus(NULL, NULL);
344 347 for (i = 0; i < num; i++)
345 348 disp_dq_assign(&disp_mem[i], numpris);
346 349 start_cpus();
347 350
348 351 /*
349 352 * I must free all of the memory after starting the cpus because
350 353 * I can not risk sleeping in kmem_free while the cpus are stopped.
351 354 */
352 355 for (i = 0; i < num; i++)
353 356 disp_dq_free(&disp_mem[i]);
354 357
355 358 kmem_free(disp_mem, NCPU * sizeof (struct disp_queue_info));
356 359 }
357 360
358 361 static void
359 362 disp_dq_alloc(struct disp_queue_info *dptr, int numpris, disp_t *dp)
360 363 {
361 364 dptr->newdispq = kmem_zalloc(numpris * sizeof (dispq_t), KM_SLEEP);
362 365 dptr->newdqactmap = kmem_zalloc(((numpris / BT_NBIPUL) + 1) *
363 366 sizeof (long), KM_SLEEP);
364 367 dptr->dp = dp;
365 368 }
366 369
367 370 static void
368 371 disp_dq_assign(struct disp_queue_info *dptr, int numpris)
369 372 {
370 373 disp_t *dp;
371 374
372 375 dp = dptr->dp;
373 376 dptr->olddispq = dp->disp_q;
374 377 dptr->olddqactmap = dp->disp_qactmap;
375 378 dptr->oldnglobpris = dp->disp_npri;
376 379
377 380 ASSERT(dptr->oldnglobpris < numpris);
378 381
379 382 if (dptr->olddispq != NULL) {
380 383 /*
381 384 * Use kcopy because bcopy is platform-specific
382 385 * and could block while we might have paused the cpus.
383 386 */
384 387 (void) kcopy(dptr->olddispq, dptr->newdispq,
385 388 dptr->oldnglobpris * sizeof (dispq_t));
386 389 (void) kcopy(dptr->olddqactmap, dptr->newdqactmap,
387 390 ((dptr->oldnglobpris / BT_NBIPUL) + 1) *
388 391 sizeof (long));
389 392 }
390 393 dp->disp_q = dptr->newdispq;
391 394 dp->disp_qactmap = dptr->newdqactmap;
392 395 dp->disp_q_limit = &dptr->newdispq[numpris];
393 396 dp->disp_npri = numpris;
394 397 }
395 398
396 399 static void
397 400 disp_dq_free(struct disp_queue_info *dptr)
398 401 {
399 402 if (dptr->olddispq != NULL)
400 403 kmem_free(dptr->olddispq,
401 404 dptr->oldnglobpris * sizeof (dispq_t));
402 405 if (dptr->olddqactmap != NULL)
403 406 kmem_free(dptr->olddqactmap,
404 407 ((dptr->oldnglobpris / BT_NBIPUL) + 1) * sizeof (long));
405 408 }
406 409
407 410 /*
408 411 * For a newly created CPU, initialize the dispatch queue.
409 412 * This is called before the CPU is known through cpu[] or on any lists.
410 413 */
411 414 void
412 415 disp_cpu_init(cpu_t *cp)
413 416 {
414 417 disp_t *dp;
415 418 dispq_t *newdispq;
416 419 ulong_t *newdqactmap;
417 420
418 421 ASSERT(MUTEX_HELD(&cpu_lock)); /* protect dispatcher queue sizes */
419 422
420 423 if (cp == cpu0_disp.disp_cpu)
421 424 dp = &cpu0_disp;
422 425 else
423 426 dp = kmem_alloc(sizeof (disp_t), KM_SLEEP);
424 427 bzero(dp, sizeof (disp_t));
425 428 cp->cpu_disp = dp;
426 429 dp->disp_cpu = cp;
427 430 dp->disp_maxrunpri = -1;
428 431 dp->disp_max_unbound_pri = -1;
429 432 DISP_LOCK_INIT(&cp->cpu_thread_lock);
430 433 /*
431 434 * Allocate memory for the dispatcher queue headers
432 435 * and the active queue bitmap.
433 436 */
434 437 newdispq = kmem_zalloc(v.v_nglobpris * sizeof (dispq_t), KM_SLEEP);
435 438 newdqactmap = kmem_zalloc(((v.v_nglobpris / BT_NBIPUL) + 1) *
436 439 sizeof (long), KM_SLEEP);
437 440 dp->disp_q = newdispq;
438 441 dp->disp_qactmap = newdqactmap;
439 442 dp->disp_q_limit = &newdispq[v.v_nglobpris];
440 443 dp->disp_npri = v.v_nglobpris;
441 444 }
442 445
443 446 void
444 447 disp_cpu_fini(cpu_t *cp)
445 448 {
446 449 ASSERT(MUTEX_HELD(&cpu_lock));
447 450
448 451 disp_kp_free(cp->cpu_disp);
449 452 if (cp->cpu_disp != &cpu0_disp)
450 453 kmem_free(cp->cpu_disp, sizeof (disp_t));
451 454 }
452 455
453 456 /*
454 457 * Allocate new, larger kpreempt dispatch queue to replace the old one.
455 458 */
456 459 void
457 460 disp_kp_alloc(disp_t *dq, pri_t npri)
458 461 {
459 462 struct disp_queue_info mem_info;
460 463
461 464 if (npri > dq->disp_npri) {
462 465 /*
463 466 * Allocate memory for the new array.
464 467 */
465 468 disp_dq_alloc(&mem_info, npri, dq);
466 469
467 470 /*
468 471 * We need to copy the old structures to the new
469 472 * and free the old.
470 473 */
471 474 disp_dq_assign(&mem_info, npri);
472 475 disp_dq_free(&mem_info);
473 476 }
474 477 }
475 478
476 479 /*
477 480 * Free dispatch queue.
478 481 * Used for the kpreempt queues for a removed CPU partition and
479 482 * for the per-CPU queues of deleted CPUs.
480 483 */
481 484 void
482 485 disp_kp_free(disp_t *dq)
483 486 {
484 487 struct disp_queue_info mem_info;
485 488
486 489 mem_info.olddispq = dq->disp_q;
487 490 mem_info.olddqactmap = dq->disp_qactmap;
488 491 mem_info.oldnglobpris = dq->disp_npri;
489 492 disp_dq_free(&mem_info);
490 493 }
491 494
492 495 /*
493 496 * End dispatcher and scheduler initialization.
494 497 */
495 498
496 499 /*
497 500 * See if there's anything to do other than remain idle.
498 501 * Return non-zero if there is.
499 502 *
500 503 * This function must be called with high spl, or with
501 504 * kernel preemption disabled to prevent the partition's
502 505 * active cpu list from changing while being traversed.
503 506 *
504 507 * This is essentially a simpler version of disp_getwork()
505 508 * to be called by CPUs preparing to "halt".
506 509 */
507 510 int
508 511 disp_anywork(void)
509 512 {
510 513 cpu_t *cp = CPU;
511 514 cpu_t *ocp;
512 515 volatile int *local_nrunnable = &cp->cpu_disp->disp_nrunnable;
513 516
514 517 if (!(cp->cpu_flags & CPU_OFFLINE)) {
515 518 if (CP_MAXRUNPRI(cp->cpu_part) >= 0)
516 519 return (1);
517 520
518 521 for (ocp = cp->cpu_next_part; ocp != cp;
519 522 ocp = ocp->cpu_next_part) {
520 523 ASSERT(CPU_ACTIVE(ocp));
521 524
522 525 /*
523 526 * Something has appeared on the local run queue.
524 527 */
525 528 if (*local_nrunnable > 0)
526 529 return (1);
527 530 /*
528 531 * If we encounter another idle CPU that will
529 532 * soon be trolling around through disp_anywork()
530 533 * terminate our walk here and let this other CPU
531 534 * patrol the next part of the list.
532 535 */
533 536 if (ocp->cpu_dispatch_pri == -1 &&
534 537 (ocp->cpu_disp_flags & CPU_DISP_HALTED) == 0)
535 538 return (0);
536 539 /*
537 540 * Work can be taken from another CPU if:
538 541 * - There is unbound work on the run queue
539 542 * - That work isn't a thread undergoing a
540 543 * - context switch on an otherwise empty queue.
541 544 * - The CPU isn't running the idle loop.
542 545 */
543 546 if (ocp->cpu_disp->disp_max_unbound_pri != -1 &&
544 547 !((ocp->cpu_disp_flags & CPU_DISP_DONTSTEAL) &&
545 548 ocp->cpu_disp->disp_nrunnable == 1) &&
546 549 ocp->cpu_dispatch_pri != -1)
547 550 return (1);
548 551 }
549 552 }
550 553 return (0);
551 554 }
552 555
553 556 /*
554 557 * Called when CPU enters the idle loop
555 558 */
556 559 static void
557 560 idle_enter()
558 561 {
559 562 cpu_t *cp = CPU;
560 563
561 564 new_cpu_mstate(CMS_IDLE, gethrtime_unscaled());
562 565 CPU_STATS_ADDQ(cp, sys, idlethread, 1);
563 566 set_idle_cpu(cp->cpu_id); /* arch-dependent hook */
564 567 }
565 568
566 569 /*
567 570 * Called when CPU exits the idle loop
568 571 */
569 572 static void
570 573 idle_exit()
571 574 {
572 575 cpu_t *cp = CPU;
573 576
574 577 new_cpu_mstate(CMS_SYSTEM, gethrtime_unscaled());
575 578 unset_idle_cpu(cp->cpu_id); /* arch-dependent hook */
576 579 }
577 580
578 581 /*
579 582 * Idle loop.
580 583 */
581 584 void
582 585 idle()
583 586 {
584 587 struct cpu *cp = CPU; /* pointer to this CPU */
585 588 kthread_t *t; /* taken thread */
586 589
587 590 idle_enter();
588 591
589 592 /*
590 593 * Uniprocessor version of idle loop.
591 594 * Do this until notified that we're on an actual multiprocessor.
592 595 */
593 596 while (ncpus == 1) {
594 597 if (cp->cpu_disp->disp_nrunnable == 0) {
595 598 (*idle_cpu)();
596 599 continue;
597 600 }
598 601 idle_exit();
599 602 swtch();
600 603
601 604 idle_enter(); /* returned from swtch */
602 605 }
603 606
604 607 /*
605 608 * Multiprocessor idle loop.
606 609 */
607 610 for (;;) {
608 611 /*
609 612 * If CPU is completely quiesced by p_online(2), just wait
610 613 * here with minimal bus traffic until put online.
611 614 */
612 615 while (cp->cpu_flags & CPU_QUIESCED)
613 616 (*idle_cpu)();
614 617
615 618 if (cp->cpu_disp->disp_nrunnable != 0) {
616 619 idle_exit();
617 620 swtch();
618 621 } else {
619 622 if (cp->cpu_flags & CPU_OFFLINE)
620 623 continue;
621 624 if ((t = disp_getwork(cp)) == NULL) {
622 625 if (cp->cpu_chosen_level != -1) {
623 626 disp_t *dp = cp->cpu_disp;
624 627 disp_t *kpq;
625 628
626 629 disp_lock_enter(&dp->disp_lock);
627 630 /*
628 631 * Set kpq under lock to prevent
629 632 * migration between partitions.
630 633 */
631 634 kpq = &cp->cpu_part->cp_kp_queue;
632 635 if (kpq->disp_maxrunpri == -1)
633 636 cp->cpu_chosen_level = -1;
634 637 disp_lock_exit(&dp->disp_lock);
635 638 }
636 639 (*idle_cpu)();
637 640 continue;
638 641 }
639 642 /*
640 643 * If there was a thread but we couldn't steal
641 644 * it, then keep trying.
642 645 */
643 646 if (t == T_DONTSTEAL)
644 647 continue;
645 648 idle_exit();
646 649 swtch_to(t);
647 650 }
648 651 idle_enter(); /* returned from swtch/swtch_to */
649 652 }
650 653 }
651 654
652 655
653 656 /*
654 657 * Preempt the currently running thread in favor of the highest
655 658 * priority thread. The class of the current thread controls
656 659 * where it goes on the dispatcher queues. If panicking, turn
657 660 * preemption off.
658 661 */
659 662 void
660 663 preempt()
661 664 {
662 665 kthread_t *t = curthread;
663 666 klwp_t *lwp = ttolwp(curthread);
664 667
665 668 if (panicstr)
666 669 return;
667 670
668 671 TRACE_0(TR_FAC_DISP, TR_PREEMPT_START, "preempt_start");
669 672
670 673 thread_lock(t);
671 674
672 675 if (t->t_state != TS_ONPROC || t->t_disp_queue != CPU->cpu_disp) {
673 676 /*
674 677 * this thread has already been chosen to be run on
675 678 * another CPU. Clear kprunrun on this CPU since we're
676 679 * already headed for swtch().
677 680 */
678 681 CPU->cpu_kprunrun = 0;
679 682 thread_unlock_nopreempt(t);
680 683 TRACE_0(TR_FAC_DISP, TR_PREEMPT_END, "preempt_end");
681 684 } else {
682 685 if (lwp != NULL)
683 686 lwp->lwp_ru.nivcsw++;
684 687 CPU_STATS_ADDQ(CPU, sys, inv_swtch, 1);
685 688 THREAD_TRANSITION(t);
686 689 CL_PREEMPT(t);
687 690 DTRACE_SCHED(preempt);
688 691 thread_unlock_nopreempt(t);
689 692
690 693 TRACE_0(TR_FAC_DISP, TR_PREEMPT_END, "preempt_end");
691 694
692 695 swtch(); /* clears CPU->cpu_runrun via disp() */
693 696 }
694 697 }
695 698
696 699 extern kthread_t *thread_unpin();
697 700
698 701 /*
699 702 * disp() - find the highest priority thread for this processor to run, and
700 703 * set it in TS_ONPROC state so that resume() can be called to run it.
701 704 */
702 705 static kthread_t *
703 706 disp()
704 707 {
705 708 cpu_t *cpup;
706 709 disp_t *dp;
707 710 kthread_t *tp;
708 711 dispq_t *dq;
709 712 int maxrunword;
710 713 pri_t pri;
711 714 disp_t *kpq;
712 715
713 716 TRACE_0(TR_FAC_DISP, TR_DISP_START, "disp_start");
714 717
715 718 cpup = CPU;
716 719 /*
717 720 * Find the highest priority loaded, runnable thread.
718 721 */
719 722 dp = cpup->cpu_disp;
720 723
721 724 reschedule:
722 725 /*
723 726 * If there is more important work on the global queue with a better
724 727 * priority than the maximum on this CPU, take it now.
725 728 */
726 729 kpq = &cpup->cpu_part->cp_kp_queue;
727 730 while ((pri = kpq->disp_maxrunpri) >= 0 &&
728 731 pri >= dp->disp_maxrunpri &&
729 732 (cpup->cpu_flags & CPU_OFFLINE) == 0 &&
730 733 (tp = disp_getbest(kpq)) != NULL) {
731 734 if (disp_ratify(tp, kpq) != NULL) {
732 735 TRACE_1(TR_FAC_DISP, TR_DISP_END,
733 736 "disp_end:tid %p", tp);
734 737 return (tp);
735 738 }
736 739 }
737 740
738 741 disp_lock_enter(&dp->disp_lock);
739 742 pri = dp->disp_maxrunpri;
740 743
741 744 /*
742 745 * If there is nothing to run, look at what's runnable on other queues.
743 746 * Choose the idle thread if the CPU is quiesced.
744 747 * Note that CPUs that have the CPU_OFFLINE flag set can still run
745 748 * interrupt threads, which will be the only threads on the CPU's own
746 749 * queue, but cannot run threads from other queues.
747 750 */
748 751 if (pri == -1) {
749 752 if (!(cpup->cpu_flags & CPU_OFFLINE)) {
750 753 disp_lock_exit(&dp->disp_lock);
751 754 if ((tp = disp_getwork(cpup)) == NULL ||
752 755 tp == T_DONTSTEAL) {
753 756 tp = cpup->cpu_idle_thread;
754 757 (void) splhigh();
755 758 THREAD_ONPROC(tp, cpup);
756 759 cpup->cpu_dispthread = tp;
757 760 cpup->cpu_dispatch_pri = -1;
758 761 cpup->cpu_runrun = cpup->cpu_kprunrun = 0;
759 762 cpup->cpu_chosen_level = -1;
760 763 }
761 764 } else {
762 765 disp_lock_exit_high(&dp->disp_lock);
763 766 tp = cpup->cpu_idle_thread;
764 767 THREAD_ONPROC(tp, cpup);
765 768 cpup->cpu_dispthread = tp;
766 769 cpup->cpu_dispatch_pri = -1;
767 770 cpup->cpu_runrun = cpup->cpu_kprunrun = 0;
768 771 cpup->cpu_chosen_level = -1;
769 772 }
770 773 TRACE_1(TR_FAC_DISP, TR_DISP_END,
771 774 "disp_end:tid %p", tp);
772 775 return (tp);
773 776 }
774 777
775 778 dq = &dp->disp_q[pri];
776 779 tp = dq->dq_first;
777 780
778 781 ASSERT(tp != NULL);
779 782 ASSERT(tp->t_schedflag & TS_LOAD); /* thread must be swapped in */
780 783
781 784 DTRACE_SCHED2(dequeue, kthread_t *, tp, disp_t *, dp);
782 785
783 786 /*
784 787 * Found it so remove it from queue.
785 788 */
786 789 dp->disp_nrunnable--;
787 790 dq->dq_sruncnt--;
788 791 if ((dq->dq_first = tp->t_link) == NULL) {
789 792 ulong_t *dqactmap = dp->disp_qactmap;
790 793
791 794 ASSERT(dq->dq_sruncnt == 0);
792 795 dq->dq_last = NULL;
793 796
794 797 /*
795 798 * The queue is empty, so the corresponding bit needs to be
796 799 * turned off in dqactmap. If nrunnable != 0 just took the
797 800 * last runnable thread off the
798 801 * highest queue, so recompute disp_maxrunpri.
799 802 */
800 803 maxrunword = pri >> BT_ULSHIFT;
801 804 dqactmap[maxrunword] &= ~BT_BIW(pri);
802 805
803 806 if (dp->disp_nrunnable == 0) {
804 807 dp->disp_max_unbound_pri = -1;
805 808 dp->disp_maxrunpri = -1;
806 809 } else {
807 810 int ipri;
808 811
809 812 ipri = bt_gethighbit(dqactmap, maxrunword);
810 813 dp->disp_maxrunpri = ipri;
811 814 if (ipri < dp->disp_max_unbound_pri)
812 815 dp->disp_max_unbound_pri = ipri;
813 816 }
814 817 } else {
815 818 tp->t_link = NULL;
816 819 }
817 820
818 821 /*
819 822 * Set TS_DONT_SWAP flag to prevent another processor from swapping
820 823 * out this thread before we have a chance to run it.
821 824 * While running, it is protected against swapping by t_lock.
822 825 */
823 826 tp->t_schedflag |= TS_DONT_SWAP;
824 827 cpup->cpu_dispthread = tp; /* protected by spl only */
825 828 cpup->cpu_dispatch_pri = pri;
826 829 ASSERT(pri == DISP_PRIO(tp));
827 830 thread_onproc(tp, cpup); /* set t_state to TS_ONPROC */
828 831 disp_lock_exit_high(&dp->disp_lock); /* drop run queue lock */
829 832
830 833 ASSERT(tp != NULL);
831 834 TRACE_1(TR_FAC_DISP, TR_DISP_END,
832 835 "disp_end:tid %p", tp);
833 836
834 837 if (disp_ratify(tp, kpq) == NULL)
835 838 goto reschedule;
836 839
837 840 return (tp);
838 841 }
839 842
840 843 /*
841 844 * swtch()
842 845 * Find best runnable thread and run it.
843 846 * Called with the current thread already switched to a new state,
844 847 * on a sleep queue, run queue, stopped, and not zombied.
845 848 * May be called at any spl level less than or equal to LOCK_LEVEL.
846 849 * Always drops spl to the base level (spl0()).
847 850 */
848 851 void
849 852 swtch()
850 853 {
851 854 kthread_t *t = curthread;
852 855 kthread_t *next;
853 856 cpu_t *cp;
854 857
855 858 TRACE_0(TR_FAC_DISP, TR_SWTCH_START, "swtch_start");
856 859
857 860 if (t->t_flag & T_INTR_THREAD)
858 861 cpu_intr_swtch_enter(t);
859 862
860 863 if (t->t_intr != NULL) {
861 864 /*
862 865 * We are an interrupt thread. Setup and return
863 866 * the interrupted thread to be resumed.
864 867 */
865 868 (void) splhigh(); /* block other scheduler action */
866 869 cp = CPU; /* now protected against migration */
867 870 ASSERT(CPU_ON_INTR(cp) == 0); /* not called with PIL > 10 */
868 871 CPU_STATS_ADDQ(cp, sys, pswitch, 1);
869 872 CPU_STATS_ADDQ(cp, sys, intrblk, 1);
870 873 next = thread_unpin();
871 874 TRACE_0(TR_FAC_DISP, TR_RESUME_START, "resume_start");
872 875 resume_from_intr(next);
873 876 } else {
874 877 #ifdef DEBUG
875 878 if (t->t_state == TS_ONPROC &&
876 879 t->t_disp_queue->disp_cpu == CPU &&
877 880 t->t_preempt == 0) {
878 881 thread_lock(t);
879 882 ASSERT(t->t_state != TS_ONPROC ||
880 883 t->t_disp_queue->disp_cpu != CPU ||
881 884 t->t_preempt != 0); /* cannot migrate */
882 885 thread_unlock_nopreempt(t);
883 886 }
884 887 #endif /* DEBUG */
885 888 cp = CPU;
886 889 next = disp(); /* returns with spl high */
887 890 ASSERT(CPU_ON_INTR(cp) == 0); /* not called with PIL > 10 */
888 891
889 892 /* OK to steal anything left on run queue */
890 893 cp->cpu_disp_flags &= ~CPU_DISP_DONTSTEAL;
891 894
892 895 if (next != t) {
893 896 hrtime_t now;
894 897
895 898 now = gethrtime_unscaled();
896 899 pg_ev_thread_swtch(cp, now, t, next);
897 900
898 901 /*
899 902 * If t was previously in the TS_ONPROC state,
900 903 * setfrontdq and setbackdq won't have set its t_waitrq.
901 904 * Since we now finally know that we're switching away
902 905 * from this thread, set its t_waitrq if it is on a run
903 906 * queue.
904 907 */
905 908 if ((t->t_state == TS_RUN) && (t->t_waitrq == 0)) {
906 909 t->t_waitrq = now;
907 910 }
908 911
909 912 /*
910 913 * restore mstate of thread that we are switching to
911 914 */
912 915 restore_mstate(next);
913 916
914 917 CPU_STATS_ADDQ(cp, sys, pswitch, 1);
915 918 cp->cpu_last_swtch = t->t_disp_time = ddi_get_lbolt();
916 919 TRACE_0(TR_FAC_DISP, TR_RESUME_START, "resume_start");
917 920
918 921 if (dtrace_vtime_active)
919 922 dtrace_vtime_switch(next);
920 923
921 924 resume(next);
922 925 /*
923 926 * The TR_RESUME_END and TR_SWTCH_END trace points
924 927 * appear at the end of resume(), because we may not
925 928 * return here
926 929 */
927 930 } else {
928 931 if (t->t_flag & T_INTR_THREAD)
929 932 cpu_intr_swtch_exit(t);
930 933 /*
931 934 * Threads that enqueue themselves on a run queue defer
932 935 * setting t_waitrq. It is then either set in swtch()
933 936 * when the CPU is actually yielded, or not at all if it
934 937 * is remaining on the CPU.
935 938 * There is however a window between where the thread
936 939 * placed itself on a run queue, and where it selects
937 940 * itself in disp(), where a third party (eg. clock()
938 941 * doing tick processing) may have re-enqueued this
939 942 * thread, setting t_waitrq in the process. We detect
940 943 * this race by noticing that despite switching to
941 944 * ourself, our t_waitrq has been set, and should be
942 945 * cleared.
943 946 */
944 947 if (t->t_waitrq != 0)
945 948 t->t_waitrq = 0;
946 949
947 950 pg_ev_thread_remain(cp, t);
948 951
949 952 DTRACE_SCHED(remain__cpu);
950 953 TRACE_0(TR_FAC_DISP, TR_SWTCH_END, "swtch_end");
951 954 (void) spl0();
952 955 }
953 956 }
954 957 }
955 958
956 959 /*
957 960 * swtch_from_zombie()
958 961 * Special case of swtch(), which allows checks for TS_ZOMB to be
959 962 * eliminated from normal resume.
960 963 * Find best runnable thread and run it.
961 964 * Called with the current thread zombied.
962 965 * Zombies cannot migrate, so CPU references are safe.
963 966 */
964 967 void
965 968 swtch_from_zombie()
966 969 {
967 970 kthread_t *next;
968 971 cpu_t *cpu = CPU;
969 972
970 973 TRACE_0(TR_FAC_DISP, TR_SWTCH_START, "swtch_start");
971 974
972 975 ASSERT(curthread->t_state == TS_ZOMB);
973 976
974 977 next = disp(); /* returns with spl high */
975 978 ASSERT(CPU_ON_INTR(CPU) == 0); /* not called with PIL > 10 */
976 979 CPU_STATS_ADDQ(CPU, sys, pswitch, 1);
977 980 ASSERT(next != curthread);
978 981 TRACE_0(TR_FAC_DISP, TR_RESUME_START, "resume_start");
979 982
980 983 pg_ev_thread_swtch(cpu, gethrtime_unscaled(), curthread, next);
981 984
982 985 restore_mstate(next);
983 986
984 987 if (dtrace_vtime_active)
985 988 dtrace_vtime_switch(next);
986 989
987 990 resume_from_zombie(next);
988 991 /*
989 992 * The TR_RESUME_END and TR_SWTCH_END trace points
990 993 * appear at the end of resume(), because we certainly will not
991 994 * return here
992 995 */
993 996 }
994 997
995 998 #if defined(DEBUG) && (defined(DISP_DEBUG) || defined(lint))
996 999
997 1000 /*
998 1001 * search_disp_queues()
999 1002 * Search the given dispatch queues for thread tp.
1000 1003 * Return 1 if tp is found, otherwise return 0.
1001 1004 */
1002 1005 static int
1003 1006 search_disp_queues(disp_t *dp, kthread_t *tp)
1004 1007 {
1005 1008 dispq_t *dq;
1006 1009 dispq_t *eq;
1007 1010
1008 1011 disp_lock_enter_high(&dp->disp_lock);
1009 1012
1010 1013 for (dq = dp->disp_q, eq = dp->disp_q_limit; dq < eq; ++dq) {
1011 1014 kthread_t *rp;
1012 1015
1013 1016 ASSERT(dq->dq_last == NULL || dq->dq_last->t_link == NULL);
1014 1017
1015 1018 for (rp = dq->dq_first; rp; rp = rp->t_link)
1016 1019 if (tp == rp) {
1017 1020 disp_lock_exit_high(&dp->disp_lock);
1018 1021 return (1);
1019 1022 }
1020 1023 }
1021 1024 disp_lock_exit_high(&dp->disp_lock);
1022 1025
1023 1026 return (0);
1024 1027 }
1025 1028
1026 1029 /*
1027 1030 * thread_on_queue()
1028 1031 * Search all per-CPU dispatch queues and all partition-wide kpreempt
1029 1032 * queues for thread tp. Return 1 if tp is found, otherwise return 0.
1030 1033 */
1031 1034 static int
1032 1035 thread_on_queue(kthread_t *tp)
1033 1036 {
1034 1037 cpu_t *cp;
1035 1038 struct cpupart *part;
1036 1039
1037 1040 ASSERT(getpil() >= DISP_LEVEL);
1038 1041
1039 1042 /*
1040 1043 * Search the per-CPU dispatch queues for tp.
1041 1044 */
1042 1045 cp = CPU;
1043 1046 do {
1044 1047 if (search_disp_queues(cp->cpu_disp, tp))
1045 1048 return (1);
1046 1049 } while ((cp = cp->cpu_next_onln) != CPU);
1047 1050
1048 1051 /*
1049 1052 * Search the partition-wide kpreempt queues for tp.
1050 1053 */
1051 1054 part = CPU->cpu_part;
1052 1055 do {
1053 1056 if (search_disp_queues(&part->cp_kp_queue, tp))
1054 1057 return (1);
1055 1058 } while ((part = part->cp_next) != CPU->cpu_part);
1056 1059
1057 1060 return (0);
1058 1061 }
1059 1062
1060 1063 #else
1061 1064
1062 1065 #define thread_on_queue(tp) 0 /* ASSERT must be !thread_on_queue */
1063 1066
1064 1067 #endif /* DEBUG */
1065 1068
1066 1069 /*
1067 1070 * like swtch(), but switch to a specified thread taken from another CPU.
1068 1071 * called with spl high..
1069 1072 */
1070 1073 void
1071 1074 swtch_to(kthread_t *next)
1072 1075 {
1073 1076 cpu_t *cp = CPU;
1074 1077 hrtime_t now;
1075 1078
1076 1079 TRACE_0(TR_FAC_DISP, TR_SWTCH_START, "swtch_start");
1077 1080
1078 1081 /*
1079 1082 * Update context switch statistics.
1080 1083 */
1081 1084 CPU_STATS_ADDQ(cp, sys, pswitch, 1);
1082 1085
1083 1086 TRACE_0(TR_FAC_DISP, TR_RESUME_START, "resume_start");
1084 1087
1085 1088 now = gethrtime_unscaled();
1086 1089 pg_ev_thread_swtch(cp, now, curthread, next);
1087 1090
1088 1091 /* OK to steal anything left on run queue */
1089 1092 cp->cpu_disp_flags &= ~CPU_DISP_DONTSTEAL;
1090 1093
1091 1094 /* record last execution time */
1092 1095 cp->cpu_last_swtch = curthread->t_disp_time = ddi_get_lbolt();
1093 1096
1094 1097 /*
1095 1098 * If t was previously in the TS_ONPROC state, setfrontdq and setbackdq
1096 1099 * won't have set its t_waitrq. Since we now finally know that we're
1097 1100 * switching away from this thread, set its t_waitrq if it is on a run
1098 1101 * queue.
1099 1102 */
1100 1103 if ((curthread->t_state == TS_RUN) && (curthread->t_waitrq == 0)) {
1101 1104 curthread->t_waitrq = now;
1102 1105 }
1103 1106
1104 1107 /* restore next thread to previously running microstate */
1105 1108 restore_mstate(next);
1106 1109
1107 1110 if (dtrace_vtime_active)
1108 1111 dtrace_vtime_switch(next);
1109 1112
1110 1113 resume(next);
1111 1114 /*
1112 1115 * The TR_RESUME_END and TR_SWTCH_END trace points
1113 1116 * appear at the end of resume(), because we may not
1114 1117 * return here
1115 1118 */
1116 1119 }
1117 1120
1118 1121 #define CPU_IDLING(pri) ((pri) == -1)
1119 1122
1120 1123 static void
1121 1124 cpu_resched(cpu_t *cp, pri_t tpri)
1122 1125 {
1123 1126 int call_poke_cpu = 0;
1124 1127 pri_t cpupri = cp->cpu_dispatch_pri;
1125 1128
1126 1129 if (!CPU_IDLING(cpupri) && (cpupri < tpri)) {
1127 1130 TRACE_2(TR_FAC_DISP, TR_CPU_RESCHED,
1128 1131 "CPU_RESCHED:Tpri %d Cpupri %d", tpri, cpupri);
1129 1132 if (tpri >= upreemptpri && cp->cpu_runrun == 0) {
1130 1133 cp->cpu_runrun = 1;
1131 1134 aston(cp->cpu_dispthread);
1132 1135 if (tpri < kpreemptpri && cp != CPU)
1133 1136 call_poke_cpu = 1;
1134 1137 }
1135 1138 if (tpri >= kpreemptpri && cp->cpu_kprunrun == 0) {
1136 1139 cp->cpu_kprunrun = 1;
1137 1140 if (cp != CPU)
1138 1141 call_poke_cpu = 1;
1139 1142 }
1140 1143 }
1141 1144
1142 1145 /*
1143 1146 * Propagate cpu_runrun, and cpu_kprunrun to global visibility.
1144 1147 */
1145 1148 membar_enter();
1146 1149
1147 1150 if (call_poke_cpu)
1148 1151 poke_cpu(cp->cpu_id);
1149 1152 }
1150 1153
1151 1154 /*
1152 1155 * setbackdq() keeps runqs balanced such that the difference in length
1153 1156 * between the chosen runq and the next one is no more than RUNQ_MAX_DIFF.
1154 1157 * For threads with priorities below RUNQ_MATCH_PRI levels, the runq's lengths
1155 1158 * must match. When per-thread TS_RUNQMATCH flag is set, setbackdq() will
1156 1159 * try to keep runqs perfectly balanced regardless of the thread priority.
1157 1160 */
1158 1161 #define RUNQ_MATCH_PRI 16 /* pri below which queue lengths must match */
1159 1162 #define RUNQ_MAX_DIFF 2 /* maximum runq length difference */
1160 1163 #define RUNQ_LEN(cp, pri) ((cp)->cpu_disp->disp_q[pri].dq_sruncnt)
1161 1164
1162 1165 /*
1163 1166 * Macro that evaluates to true if it is likely that the thread has cache
1164 1167 * warmth. This is based on the amount of time that has elapsed since the
1165 1168 * thread last ran. If that amount of time is less than "rechoose_interval"
1166 1169 * ticks, then we decide that the thread has enough cache warmth to warrant
1167 1170 * some affinity for t->t_cpu.
1168 1171 */
1169 1172 #define THREAD_HAS_CACHE_WARMTH(thread) \
1170 1173 ((thread == curthread) || \
1171 1174 ((ddi_get_lbolt() - thread->t_disp_time) <= rechoose_interval))
1172 1175 /*
1173 1176 * Put the specified thread on the back of the dispatcher
1174 1177 * queue corresponding to its current priority.
1175 1178 *
1176 1179 * Called with the thread in transition, onproc or stopped state
1177 1180 * and locked (transition implies locked) and at high spl.
1178 1181 * Returns with the thread in TS_RUN state and still locked.
1179 1182 */
1180 1183 void
1181 1184 setbackdq(kthread_t *tp)
1182 1185 {
1183 1186 dispq_t *dq;
1184 1187 disp_t *dp;
1185 1188 cpu_t *cp;
1186 1189 pri_t tpri;
1187 1190 int bound;
1188 1191 boolean_t self;
1189 1192
1190 1193 ASSERT(THREAD_LOCK_HELD(tp));
1191 1194 ASSERT((tp->t_schedflag & TS_ALLSTART) == 0);
1192 1195 ASSERT(!thread_on_queue(tp)); /* make sure tp isn't on a runq */
1193 1196
1194 1197 /*
1195 1198 * If thread is "swapped" or on the swap queue don't
1196 1199 * queue it, but wake sched.
1197 1200 */
1198 1201 if ((tp->t_schedflag & (TS_LOAD | TS_ON_SWAPQ)) != TS_LOAD) {
1199 1202 disp_swapped_setrun(tp);
1200 1203 return;
1201 1204 }
1202 1205
1203 1206 self = (tp == curthread);
1204 1207
1205 1208 if (tp->t_bound_cpu || tp->t_weakbound_cpu)
1206 1209 bound = 1;
1207 1210 else
1208 1211 bound = 0;
1209 1212
1210 1213 tpri = DISP_PRIO(tp);
1211 1214 if (ncpus == 1)
1212 1215 cp = tp->t_cpu;
1213 1216 else if (!bound) {
1214 1217 if (tpri >= kpqpri) {
1215 1218 setkpdq(tp, SETKP_BACK);
1216 1219 return;
1217 1220 }
1218 1221
1219 1222 /*
1220 1223 * We'll generally let this thread continue to run where
1221 1224 * it last ran...but will consider migration if:
1222 1225 * - We thread probably doesn't have much cache warmth.
1223 1226 * - The CPU where it last ran is the target of an offline
1224 1227 * request.
1225 1228 * - The thread last ran outside it's home lgroup.
1226 1229 */
1227 1230 if ((!THREAD_HAS_CACHE_WARMTH(tp)) ||
1228 1231 (tp->t_cpu == cpu_inmotion)) {
1229 1232 cp = disp_lowpri_cpu(tp->t_cpu, tp->t_lpl, tpri, NULL);
1230 1233 } else if (!LGRP_CONTAINS_CPU(tp->t_lpl->lpl_lgrp, tp->t_cpu)) {
1231 1234 cp = disp_lowpri_cpu(tp->t_cpu, tp->t_lpl, tpri,
1232 1235 self ? tp->t_cpu : NULL);
1233 1236 } else {
1234 1237 cp = tp->t_cpu;
1235 1238 }
1236 1239
1237 1240 if (tp->t_cpupart == cp->cpu_part) {
1238 1241 int qlen;
1239 1242
1240 1243 /*
1241 1244 * Perform any CMT load balancing
1242 1245 */
1243 1246 cp = cmt_balance(tp, cp);
1244 1247
1245 1248 /*
1246 1249 * Balance across the run queues
1247 1250 */
1248 1251 qlen = RUNQ_LEN(cp, tpri);
1249 1252 if (tpri >= RUNQ_MATCH_PRI &&
1250 1253 !(tp->t_schedflag & TS_RUNQMATCH))
1251 1254 qlen -= RUNQ_MAX_DIFF;
1252 1255 if (qlen > 0) {
1253 1256 cpu_t *newcp;
1254 1257
1255 1258 if (tp->t_lpl->lpl_lgrpid == LGRP_ROOTID) {
1256 1259 newcp = cp->cpu_next_part;
1257 1260 } else if ((newcp = cp->cpu_next_lpl) == cp) {
1258 1261 newcp = cp->cpu_next_part;
1259 1262 }
1260 1263
1261 1264 if (RUNQ_LEN(newcp, tpri) < qlen) {
1262 1265 DTRACE_PROBE3(runq__balance,
1263 1266 kthread_t *, tp,
1264 1267 cpu_t *, cp, cpu_t *, newcp);
1265 1268 cp = newcp;
1266 1269 }
1267 1270 }
1268 1271 } else {
1269 1272 /*
1270 1273 * Migrate to a cpu in the new partition.
1271 1274 */
1272 1275 cp = disp_lowpri_cpu(tp->t_cpupart->cp_cpulist,
1273 1276 tp->t_lpl, tp->t_pri, NULL);
1274 1277 }
1275 1278 ASSERT((cp->cpu_flags & CPU_QUIESCED) == 0);
1276 1279 } else {
1277 1280 /*
1278 1281 * It is possible that t_weakbound_cpu != t_bound_cpu (for
1279 1282 * a short time until weak binding that existed when the
1280 1283 * strong binding was established has dropped) so we must
1281 1284 * favour weak binding over strong.
1282 1285 */
1283 1286 cp = tp->t_weakbound_cpu ?
1284 1287 tp->t_weakbound_cpu : tp->t_bound_cpu;
1285 1288 }
1286 1289 /*
1287 1290 * A thread that is ONPROC may be temporarily placed on the run queue
1288 1291 * but then chosen to run again by disp. If the thread we're placing on
1289 1292 * the queue is in TS_ONPROC state, don't set its t_waitrq until a
1290 1293 * replacement process is actually scheduled in swtch(). In this
1291 1294 * situation, curthread is the only thread that could be in the ONPROC
1292 1295 * state.
1293 1296 */
1294 1297 if ((!self) && (tp->t_waitrq == 0)) {
1295 1298 hrtime_t curtime;
1296 1299
1297 1300 curtime = gethrtime_unscaled();
1298 1301 (void) cpu_update_pct(tp, curtime);
1299 1302 tp->t_waitrq = curtime;
1300 1303 } else {
1301 1304 (void) cpu_update_pct(tp, gethrtime_unscaled());
1302 1305 }
1303 1306
1304 1307 dp = cp->cpu_disp;
1305 1308 disp_lock_enter_high(&dp->disp_lock);
1306 1309
1307 1310 DTRACE_SCHED3(enqueue, kthread_t *, tp, disp_t *, dp, int, 0);
1308 1311 TRACE_3(TR_FAC_DISP, TR_BACKQ, "setbackdq:pri %d cpu %p tid %p",
1309 1312 tpri, cp, tp);
1310 1313
1311 1314 #ifndef NPROBE
1312 1315 /* Kernel probe */
1313 1316 if (tnf_tracing_active)
1314 1317 tnf_thread_queue(tp, cp, tpri);
1315 1318 #endif /* NPROBE */
1316 1319
1317 1320 ASSERT(tpri >= 0 && tpri < dp->disp_npri);
1318 1321
1319 1322 THREAD_RUN(tp, &dp->disp_lock); /* set t_state to TS_RUN */
1320 1323 tp->t_disp_queue = dp;
1321 1324 tp->t_link = NULL;
1322 1325
1323 1326 dq = &dp->disp_q[tpri];
1324 1327 dp->disp_nrunnable++;
1325 1328 if (!bound)
1326 1329 dp->disp_steal = 0;
1327 1330 membar_enter();
1328 1331
1329 1332 if (dq->dq_sruncnt++ != 0) {
1330 1333 ASSERT(dq->dq_first != NULL);
1331 1334 dq->dq_last->t_link = tp;
1332 1335 dq->dq_last = tp;
1333 1336 } else {
1334 1337 ASSERT(dq->dq_first == NULL);
1335 1338 ASSERT(dq->dq_last == NULL);
1336 1339 dq->dq_first = dq->dq_last = tp;
1337 1340 BT_SET(dp->disp_qactmap, tpri);
1338 1341 if (tpri > dp->disp_maxrunpri) {
1339 1342 dp->disp_maxrunpri = tpri;
1340 1343 membar_enter();
1341 1344 cpu_resched(cp, tpri);
1342 1345 }
1343 1346 }
1344 1347
1345 1348 if (!bound && tpri > dp->disp_max_unbound_pri) {
1346 1349 if (self && dp->disp_max_unbound_pri == -1 && cp == CPU) {
1347 1350 /*
1348 1351 * If there are no other unbound threads on the
1349 1352 * run queue, don't allow other CPUs to steal
1350 1353 * this thread while we are in the middle of a
1351 1354 * context switch. We may just switch to it
1352 1355 * again right away. CPU_DISP_DONTSTEAL is cleared
1353 1356 * in swtch and swtch_to.
1354 1357 */
1355 1358 cp->cpu_disp_flags |= CPU_DISP_DONTSTEAL;
1356 1359 }
1357 1360 dp->disp_max_unbound_pri = tpri;
1358 1361 }
1359 1362 (*disp_enq_thread)(cp, bound);
1360 1363 }
1361 1364
1362 1365 /*
1363 1366 * Put the specified thread on the front of the dispatcher
1364 1367 * queue corresponding to its current priority.
1365 1368 *
1366 1369 * Called with the thread in transition, onproc or stopped state
1367 1370 * and locked (transition implies locked) and at high spl.
1368 1371 * Returns with the thread in TS_RUN state and still locked.
1369 1372 */
1370 1373 void
1371 1374 setfrontdq(kthread_t *tp)
1372 1375 {
1373 1376 disp_t *dp;
1374 1377 dispq_t *dq;
1375 1378 cpu_t *cp;
1376 1379 pri_t tpri;
1377 1380 int bound;
1378 1381
1379 1382 ASSERT(THREAD_LOCK_HELD(tp));
1380 1383 ASSERT((tp->t_schedflag & TS_ALLSTART) == 0);
1381 1384 ASSERT(!thread_on_queue(tp)); /* make sure tp isn't on a runq */
1382 1385
1383 1386 /*
1384 1387 * If thread is "swapped" or on the swap queue don't
1385 1388 * queue it, but wake sched.
1386 1389 */
1387 1390 if ((tp->t_schedflag & (TS_LOAD | TS_ON_SWAPQ)) != TS_LOAD) {
1388 1391 disp_swapped_setrun(tp);
1389 1392 return;
1390 1393 }
1391 1394
1392 1395 if (tp->t_bound_cpu || tp->t_weakbound_cpu)
1393 1396 bound = 1;
1394 1397 else
1395 1398 bound = 0;
1396 1399
1397 1400 tpri = DISP_PRIO(tp);
1398 1401 if (ncpus == 1)
1399 1402 cp = tp->t_cpu;
1400 1403 else if (!bound) {
1401 1404 if (tpri >= kpqpri) {
1402 1405 setkpdq(tp, SETKP_FRONT);
1403 1406 return;
1404 1407 }
1405 1408 cp = tp->t_cpu;
1406 1409 if (tp->t_cpupart == cp->cpu_part) {
1407 1410 /*
1408 1411 * We'll generally let this thread continue to run
1409 1412 * where it last ran, but will consider migration if:
1410 1413 * - The thread last ran outside it's home lgroup.
1411 1414 * - The CPU where it last ran is the target of an
1412 1415 * offline request (a thread_nomigrate() on the in
1413 1416 * motion CPU relies on this when forcing a preempt).
1414 1417 * - The thread isn't the highest priority thread where
1415 1418 * it last ran, and it is considered not likely to
1416 1419 * have significant cache warmth.
1417 1420 */
1418 1421 if ((!LGRP_CONTAINS_CPU(tp->t_lpl->lpl_lgrp, cp)) ||
1419 1422 (cp == cpu_inmotion)) {
1420 1423 cp = disp_lowpri_cpu(tp->t_cpu, tp->t_lpl, tpri,
1421 1424 (tp == curthread) ? cp : NULL);
1422 1425 } else if ((tpri < cp->cpu_disp->disp_maxrunpri) &&
1423 1426 (!THREAD_HAS_CACHE_WARMTH(tp))) {
1424 1427 cp = disp_lowpri_cpu(tp->t_cpu, tp->t_lpl, tpri,
1425 1428 NULL);
1426 1429 }
1427 1430 } else {
1428 1431 /*
1429 1432 * Migrate to a cpu in the new partition.
1430 1433 */
1431 1434 cp = disp_lowpri_cpu(tp->t_cpupart->cp_cpulist,
1432 1435 tp->t_lpl, tp->t_pri, NULL);
1433 1436 }
1434 1437 ASSERT((cp->cpu_flags & CPU_QUIESCED) == 0);
1435 1438 } else {
1436 1439 /*
1437 1440 * It is possible that t_weakbound_cpu != t_bound_cpu (for
1438 1441 * a short time until weak binding that existed when the
1439 1442 * strong binding was established has dropped) so we must
1440 1443 * favour weak binding over strong.
1441 1444 */
1442 1445 cp = tp->t_weakbound_cpu ?
1443 1446 tp->t_weakbound_cpu : tp->t_bound_cpu;
1444 1447 }
1445 1448
1446 1449 /*
1447 1450 * A thread that is ONPROC may be temporarily placed on the run queue
1448 1451 * but then chosen to run again by disp. If the thread we're placing on
1449 1452 * the queue is in TS_ONPROC state, don't set its t_waitrq until a
1450 1453 * replacement process is actually scheduled in swtch(). In this
1451 1454 * situation, curthread is the only thread that could be in the ONPROC
1452 1455 * state.
1453 1456 */
1454 1457 if ((tp != curthread) && (tp->t_waitrq == 0)) {
1455 1458 hrtime_t curtime;
1456 1459
1457 1460 curtime = gethrtime_unscaled();
1458 1461 (void) cpu_update_pct(tp, curtime);
1459 1462 tp->t_waitrq = curtime;
1460 1463 } else {
1461 1464 (void) cpu_update_pct(tp, gethrtime_unscaled());
1462 1465 }
1463 1466
1464 1467 dp = cp->cpu_disp;
1465 1468 disp_lock_enter_high(&dp->disp_lock);
1466 1469
1467 1470 TRACE_2(TR_FAC_DISP, TR_FRONTQ, "frontq:pri %d tid %p", tpri, tp);
1468 1471 DTRACE_SCHED3(enqueue, kthread_t *, tp, disp_t *, dp, int, 1);
1469 1472
1470 1473 #ifndef NPROBE
1471 1474 /* Kernel probe */
1472 1475 if (tnf_tracing_active)
1473 1476 tnf_thread_queue(tp, cp, tpri);
1474 1477 #endif /* NPROBE */
1475 1478
1476 1479 ASSERT(tpri >= 0 && tpri < dp->disp_npri);
1477 1480
1478 1481 THREAD_RUN(tp, &dp->disp_lock); /* set TS_RUN state and lock */
1479 1482 tp->t_disp_queue = dp;
1480 1483
1481 1484 dq = &dp->disp_q[tpri];
1482 1485 dp->disp_nrunnable++;
1483 1486 if (!bound)
1484 1487 dp->disp_steal = 0;
1485 1488 membar_enter();
1486 1489
1487 1490 if (dq->dq_sruncnt++ != 0) {
1488 1491 ASSERT(dq->dq_last != NULL);
1489 1492 tp->t_link = dq->dq_first;
1490 1493 dq->dq_first = tp;
1491 1494 } else {
1492 1495 ASSERT(dq->dq_last == NULL);
1493 1496 ASSERT(dq->dq_first == NULL);
1494 1497 tp->t_link = NULL;
1495 1498 dq->dq_first = dq->dq_last = tp;
1496 1499 BT_SET(dp->disp_qactmap, tpri);
1497 1500 if (tpri > dp->disp_maxrunpri) {
1498 1501 dp->disp_maxrunpri = tpri;
1499 1502 membar_enter();
1500 1503 cpu_resched(cp, tpri);
1501 1504 }
1502 1505 }
1503 1506
1504 1507 if (!bound && tpri > dp->disp_max_unbound_pri) {
1505 1508 if (tp == curthread && dp->disp_max_unbound_pri == -1 &&
1506 1509 cp == CPU) {
1507 1510 /*
1508 1511 * If there are no other unbound threads on the
1509 1512 * run queue, don't allow other CPUs to steal
1510 1513 * this thread while we are in the middle of a
1511 1514 * context switch. We may just switch to it
1512 1515 * again right away. CPU_DISP_DONTSTEAL is cleared
1513 1516 * in swtch and swtch_to.
1514 1517 */
1515 1518 cp->cpu_disp_flags |= CPU_DISP_DONTSTEAL;
1516 1519 }
1517 1520 dp->disp_max_unbound_pri = tpri;
1518 1521 }
1519 1522 (*disp_enq_thread)(cp, bound);
1520 1523 }
1521 1524
1522 1525 /*
1523 1526 * Put a high-priority unbound thread on the kp queue
1524 1527 */
1525 1528 static void
1526 1529 setkpdq(kthread_t *tp, int borf)
1527 1530 {
1528 1531 dispq_t *dq;
1529 1532 disp_t *dp;
1530 1533 cpu_t *cp;
1531 1534 pri_t tpri;
1532 1535
1533 1536 tpri = DISP_PRIO(tp);
1534 1537
1535 1538 dp = &tp->t_cpupart->cp_kp_queue;
1536 1539 disp_lock_enter_high(&dp->disp_lock);
1537 1540
1538 1541 TRACE_2(TR_FAC_DISP, TR_FRONTQ, "frontq:pri %d tid %p", tpri, tp);
1539 1542
1540 1543 ASSERT(tpri >= 0 && tpri < dp->disp_npri);
1541 1544 DTRACE_SCHED3(enqueue, kthread_t *, tp, disp_t *, dp, int, borf);
1542 1545 THREAD_RUN(tp, &dp->disp_lock); /* set t_state to TS_RUN */
1543 1546 tp->t_disp_queue = dp;
1544 1547 dp->disp_nrunnable++;
1545 1548 dq = &dp->disp_q[tpri];
1546 1549
1547 1550 if (dq->dq_sruncnt++ != 0) {
1548 1551 if (borf == SETKP_BACK) {
1549 1552 ASSERT(dq->dq_first != NULL);
1550 1553 tp->t_link = NULL;
1551 1554 dq->dq_last->t_link = tp;
1552 1555 dq->dq_last = tp;
1553 1556 } else {
1554 1557 ASSERT(dq->dq_last != NULL);
1555 1558 tp->t_link = dq->dq_first;
1556 1559 dq->dq_first = tp;
1557 1560 }
1558 1561 } else {
1559 1562 if (borf == SETKP_BACK) {
1560 1563 ASSERT(dq->dq_first == NULL);
1561 1564 ASSERT(dq->dq_last == NULL);
1562 1565 dq->dq_first = dq->dq_last = tp;
1563 1566 } else {
1564 1567 ASSERT(dq->dq_last == NULL);
1565 1568 ASSERT(dq->dq_first == NULL);
1566 1569 tp->t_link = NULL;
1567 1570 dq->dq_first = dq->dq_last = tp;
1568 1571 }
1569 1572 BT_SET(dp->disp_qactmap, tpri);
1570 1573 if (tpri > dp->disp_max_unbound_pri)
1571 1574 dp->disp_max_unbound_pri = tpri;
1572 1575 if (tpri > dp->disp_maxrunpri) {
1573 1576 dp->disp_maxrunpri = tpri;
1574 1577 membar_enter();
1575 1578 }
1576 1579 }
1577 1580
1578 1581 cp = tp->t_cpu;
1579 1582 if (tp->t_cpupart != cp->cpu_part) {
1580 1583 /* migrate to a cpu in the new partition */
1581 1584 cp = tp->t_cpupart->cp_cpulist;
1582 1585 }
1583 1586 cp = disp_lowpri_cpu(cp, tp->t_lpl, tp->t_pri, NULL);
1584 1587 disp_lock_enter_high(&cp->cpu_disp->disp_lock);
1585 1588 ASSERT((cp->cpu_flags & CPU_QUIESCED) == 0);
1586 1589
1587 1590 #ifndef NPROBE
1588 1591 /* Kernel probe */
1589 1592 if (tnf_tracing_active)
1590 1593 tnf_thread_queue(tp, cp, tpri);
1591 1594 #endif /* NPROBE */
1592 1595
1593 1596 if (cp->cpu_chosen_level < tpri)
1594 1597 cp->cpu_chosen_level = tpri;
1595 1598 cpu_resched(cp, tpri);
1596 1599 disp_lock_exit_high(&cp->cpu_disp->disp_lock);
1597 1600 (*disp_enq_thread)(cp, 0);
1598 1601 }
1599 1602
1600 1603 /*
1601 1604 * Remove a thread from the dispatcher queue if it is on it.
1602 1605 * It is not an error if it is not found but we return whether
1603 1606 * or not it was found in case the caller wants to check.
1604 1607 */
1605 1608 int
1606 1609 dispdeq(kthread_t *tp)
1607 1610 {
1608 1611 disp_t *dp;
1609 1612 dispq_t *dq;
1610 1613 kthread_t *rp;
1611 1614 kthread_t *trp;
1612 1615 kthread_t **ptp;
1613 1616 int tpri;
1614 1617
1615 1618 ASSERT(THREAD_LOCK_HELD(tp));
1616 1619
1617 1620 if (tp->t_state != TS_RUN)
1618 1621 return (0);
1619 1622
1620 1623 /*
1621 1624 * The thread is "swapped" or is on the swap queue and
1622 1625 * hence no longer on the run queue, so return true.
1623 1626 */
1624 1627 if ((tp->t_schedflag & (TS_LOAD | TS_ON_SWAPQ)) != TS_LOAD)
1625 1628 return (1);
1626 1629
1627 1630 tpri = DISP_PRIO(tp);
1628 1631 dp = tp->t_disp_queue;
1629 1632 ASSERT(tpri < dp->disp_npri);
1630 1633 dq = &dp->disp_q[tpri];
1631 1634 ptp = &dq->dq_first;
1632 1635 rp = *ptp;
1633 1636 trp = NULL;
1634 1637
1635 1638 ASSERT(dq->dq_last == NULL || dq->dq_last->t_link == NULL);
1636 1639
1637 1640 /*
1638 1641 * Search for thread in queue.
1639 1642 * Double links would simplify this at the expense of disp/setrun.
1640 1643 */
1641 1644 while (rp != tp && rp != NULL) {
1642 1645 trp = rp;
1643 1646 ptp = &trp->t_link;
1644 1647 rp = trp->t_link;
1645 1648 }
1646 1649
1647 1650 if (rp == NULL) {
1648 1651 panic("dispdeq: thread not on queue");
1649 1652 }
1650 1653
1651 1654 DTRACE_SCHED2(dequeue, kthread_t *, tp, disp_t *, dp);
1652 1655
1653 1656 /*
1654 1657 * Found it so remove it from queue.
1655 1658 */
1656 1659 if ((*ptp = rp->t_link) == NULL)
1657 1660 dq->dq_last = trp;
1658 1661
1659 1662 dp->disp_nrunnable--;
1660 1663 if (--dq->dq_sruncnt == 0) {
1661 1664 dp->disp_qactmap[tpri >> BT_ULSHIFT] &= ~BT_BIW(tpri);
1662 1665 if (dp->disp_nrunnable == 0) {
1663 1666 dp->disp_max_unbound_pri = -1;
1664 1667 dp->disp_maxrunpri = -1;
1665 1668 } else if (tpri == dp->disp_maxrunpri) {
1666 1669 int ipri;
1667 1670
1668 1671 ipri = bt_gethighbit(dp->disp_qactmap,
1669 1672 dp->disp_maxrunpri >> BT_ULSHIFT);
1670 1673 if (ipri < dp->disp_max_unbound_pri)
1671 1674 dp->disp_max_unbound_pri = ipri;
1672 1675 dp->disp_maxrunpri = ipri;
1673 1676 }
1674 1677 }
1675 1678 tp->t_link = NULL;
1676 1679 THREAD_TRANSITION(tp); /* put in intermediate state */
1677 1680 return (1);
1678 1681 }
1679 1682
1680 1683
1681 1684 /*
1682 1685 * dq_sruninc and dq_srundec are public functions for
1683 1686 * incrementing/decrementing the sruncnts when a thread on
1684 1687 * a dispatcher queue is made schedulable/unschedulable by
1685 1688 * resetting the TS_LOAD flag.
1686 1689 *
1687 1690 * The caller MUST have the thread lock and therefore the dispatcher
1688 1691 * queue lock so that the operation which changes
1689 1692 * the flag, the operation that checks the status of the thread to
1690 1693 * determine if it's on a disp queue AND the call to this function
1691 1694 * are one atomic operation with respect to interrupts.
1692 1695 */
1693 1696
1694 1697 /*
1695 1698 * Called by sched AFTER TS_LOAD flag is set on a swapped, runnable thread.
1696 1699 */
1697 1700 void
1698 1701 dq_sruninc(kthread_t *t)
1699 1702 {
1700 1703 ASSERT(t->t_state == TS_RUN);
1701 1704 ASSERT(t->t_schedflag & TS_LOAD);
1702 1705
1703 1706 THREAD_TRANSITION(t);
1704 1707 setfrontdq(t);
1705 1708 }
1706 1709
1707 1710 /*
1708 1711 * See comment on calling conventions above.
1709 1712 * Called by sched BEFORE TS_LOAD flag is cleared on a runnable thread.
1710 1713 */
1711 1714 void
1712 1715 dq_srundec(kthread_t *t)
1713 1716 {
1714 1717 ASSERT(t->t_schedflag & TS_LOAD);
1715 1718
1716 1719 (void) dispdeq(t);
1717 1720 disp_swapped_enq(t);
1718 1721 }
1719 1722
1720 1723 /*
1721 1724 * Change the dispatcher lock of thread to the "swapped_lock"
1722 1725 * and return with thread lock still held.
1723 1726 *
1724 1727 * Called with thread_lock held, in transition state, and at high spl.
1725 1728 */
1726 1729 void
1727 1730 disp_swapped_enq(kthread_t *tp)
1728 1731 {
1729 1732 ASSERT(THREAD_LOCK_HELD(tp));
1730 1733 ASSERT(tp->t_schedflag & TS_LOAD);
1731 1734
1732 1735 switch (tp->t_state) {
1733 1736 case TS_RUN:
1734 1737 disp_lock_enter_high(&swapped_lock);
1735 1738 THREAD_SWAP(tp, &swapped_lock); /* set TS_RUN state and lock */
1736 1739 break;
1737 1740 case TS_ONPROC:
1738 1741 disp_lock_enter_high(&swapped_lock);
1739 1742 THREAD_TRANSITION(tp);
1740 1743 wake_sched_sec = 1; /* tell clock to wake sched */
1741 1744 THREAD_SWAP(tp, &swapped_lock); /* set TS_RUN state and lock */
1742 1745 break;
1743 1746 default:
1744 1747 panic("disp_swapped: tp: %p bad t_state", (void *)tp);
1745 1748 }
1746 1749 }
1747 1750
1748 1751 /*
1749 1752 * This routine is called by setbackdq/setfrontdq if the thread is
1750 1753 * not loaded or loaded and on the swap queue.
1751 1754 *
1752 1755 * Thread state TS_SLEEP implies that a swapped thread
1753 1756 * has been woken up and needs to be swapped in by the swapper.
1754 1757 *
1755 1758 * Thread state TS_RUN, it implies that the priority of a swapped
1756 1759 * thread is being increased by scheduling class (e.g. ts_update).
1757 1760 */
1758 1761 static void
1759 1762 disp_swapped_setrun(kthread_t *tp)
1760 1763 {
1761 1764 ASSERT(THREAD_LOCK_HELD(tp));
1762 1765 ASSERT((tp->t_schedflag & (TS_LOAD | TS_ON_SWAPQ)) != TS_LOAD);
1763 1766
1764 1767 switch (tp->t_state) {
1765 1768 case TS_SLEEP:
1766 1769 disp_lock_enter_high(&swapped_lock);
1767 1770 /*
1768 1771 * Wakeup sched immediately (i.e., next tick) if the
1769 1772 * thread priority is above maxclsyspri.
1770 1773 */
1771 1774 if (DISP_PRIO(tp) > maxclsyspri)
1772 1775 wake_sched = 1;
1773 1776 else
1774 1777 wake_sched_sec = 1;
1775 1778 THREAD_RUN(tp, &swapped_lock); /* set TS_RUN state and lock */
1776 1779 break;
1777 1780 case TS_RUN: /* called from ts_update */
1778 1781 break;
1779 1782 default:
1780 1783 panic("disp_swapped_setrun: tp: %p bad t_state", (void *)tp);
1781 1784 }
1782 1785 }
1783 1786
1784 1787 /*
1785 1788 * Make a thread give up its processor. Find the processor on
1786 1789 * which this thread is executing, and have that processor
1787 1790 * preempt.
1788 1791 *
1789 1792 * We allow System Duty Cycle (SDC) threads to be preempted even if
1790 1793 * they are running at kernel priorities. To implement this, we always
1791 1794 * set cpu_kprunrun; this ensures preempt() will be called. Since SDC
1792 1795 * calls cpu_surrender() very often, we only preempt if there is anyone
1793 1796 * competing with us.
1794 1797 */
1795 1798 void
1796 1799 cpu_surrender(kthread_t *tp)
1797 1800 {
1798 1801 cpu_t *cpup;
1799 1802 int max_pri;
1800 1803 int max_run_pri;
1801 1804 klwp_t *lwp;
1802 1805
1803 1806 ASSERT(THREAD_LOCK_HELD(tp));
1804 1807
1805 1808 if (tp->t_state != TS_ONPROC)
1806 1809 return;
1807 1810 cpup = tp->t_disp_queue->disp_cpu; /* CPU thread dispatched to */
1808 1811 max_pri = cpup->cpu_disp->disp_maxrunpri; /* best pri of that CPU */
1809 1812 max_run_pri = CP_MAXRUNPRI(cpup->cpu_part);
1810 1813 if (max_pri < max_run_pri)
1811 1814 max_pri = max_run_pri;
1812 1815
1813 1816 if (tp->t_cid == sysdccid) {
1814 1817 uint_t t_pri = DISP_PRIO(tp);
1815 1818 if (t_pri > max_pri)
1816 1819 return; /* we are not competing w/ anyone */
1817 1820 cpup->cpu_runrun = cpup->cpu_kprunrun = 1;
1818 1821 } else {
1819 1822 cpup->cpu_runrun = 1;
1820 1823 if (max_pri >= kpreemptpri && cpup->cpu_kprunrun == 0) {
1821 1824 cpup->cpu_kprunrun = 1;
1822 1825 }
1823 1826 }
1824 1827
1825 1828 /*
1826 1829 * Propagate cpu_runrun, and cpu_kprunrun to global visibility.
1827 1830 */
1828 1831 membar_enter();
1829 1832
1830 1833 DTRACE_SCHED1(surrender, kthread_t *, tp);
1831 1834
1832 1835 /*
1833 1836 * Make the target thread take an excursion through trap()
1834 1837 * to do preempt() (unless we're already in trap or post_syscall,
1835 1838 * calling cpu_surrender via CL_TRAPRET).
1836 1839 */
1837 1840 if (tp != curthread || (lwp = tp->t_lwp) == NULL ||
1838 1841 lwp->lwp_state != LWP_USER) {
1839 1842 aston(tp);
1840 1843 if (cpup != CPU)
1841 1844 poke_cpu(cpup->cpu_id);
1842 1845 }
1843 1846 TRACE_2(TR_FAC_DISP, TR_CPU_SURRENDER,
1844 1847 "cpu_surrender:tid %p cpu %p", tp, cpup);
1845 1848 }
1846 1849
1847 1850 /*
1848 1851 * Commit to and ratify a scheduling decision
1849 1852 */
1850 1853 /*ARGSUSED*/
1851 1854 static kthread_t *
1852 1855 disp_ratify(kthread_t *tp, disp_t *kpq)
1853 1856 {
1854 1857 pri_t tpri, maxpri;
1855 1858 pri_t maxkpri;
1856 1859 cpu_t *cpup;
1857 1860
1858 1861 ASSERT(tp != NULL);
1859 1862 /*
1860 1863 * Commit to, then ratify scheduling decision
1861 1864 */
1862 1865 cpup = CPU;
1863 1866 if (cpup->cpu_runrun != 0)
1864 1867 cpup->cpu_runrun = 0;
1865 1868 if (cpup->cpu_kprunrun != 0)
1866 1869 cpup->cpu_kprunrun = 0;
1867 1870 if (cpup->cpu_chosen_level != -1)
1868 1871 cpup->cpu_chosen_level = -1;
1869 1872 membar_enter();
1870 1873 tpri = DISP_PRIO(tp);
1871 1874 maxpri = cpup->cpu_disp->disp_maxrunpri;
1872 1875 maxkpri = kpq->disp_maxrunpri;
1873 1876 if (maxpri < maxkpri)
1874 1877 maxpri = maxkpri;
1875 1878 if (tpri < maxpri) {
1876 1879 /*
1877 1880 * should have done better
1878 1881 * put this one back and indicate to try again
1879 1882 */
1880 1883 cpup->cpu_dispthread = curthread; /* fixup dispthread */
1881 1884 cpup->cpu_dispatch_pri = DISP_PRIO(curthread);
1882 1885 thread_lock_high(tp);
1883 1886 THREAD_TRANSITION(tp);
1884 1887 setfrontdq(tp);
1885 1888 thread_unlock_nopreempt(tp);
1886 1889
1887 1890 tp = NULL;
1888 1891 }
1889 1892 return (tp);
1890 1893 }
1891 1894
1892 1895 /*
1893 1896 * See if there is any work on the dispatcher queue for other CPUs.
1894 1897 * If there is, dequeue the best thread and return.
1895 1898 */
1896 1899 static kthread_t *
1897 1900 disp_getwork(cpu_t *cp)
1898 1901 {
1899 1902 cpu_t *ocp; /* other CPU */
1900 1903 cpu_t *ocp_start;
1901 1904 cpu_t *tcp; /* target local CPU */
1902 1905 kthread_t *tp;
1903 1906 kthread_t *retval = NULL;
1904 1907 pri_t maxpri;
1905 1908 disp_t *kpq; /* kp queue for this partition */
1906 1909 lpl_t *lpl, *lpl_leaf;
1907 1910 int leafidx, startidx;
1908 1911 hrtime_t stealtime;
1909 1912 lgrp_id_t local_id;
1910 1913
1911 1914 maxpri = -1;
1912 1915 tcp = NULL;
1913 1916
1914 1917 kpq = &cp->cpu_part->cp_kp_queue;
1915 1918 while (kpq->disp_maxrunpri >= 0) {
1916 1919 /*
1917 1920 * Try to take a thread from the kp_queue.
1918 1921 */
1919 1922 tp = (disp_getbest(kpq));
1920 1923 if (tp)
1921 1924 return (disp_ratify(tp, kpq));
1922 1925 }
1923 1926
1924 1927 kpreempt_disable(); /* protect the cpu_active list */
1925 1928
1926 1929 /*
1927 1930 * Try to find something to do on another CPU's run queue.
1928 1931 * Loop through all other CPUs looking for the one with the highest
1929 1932 * priority unbound thread.
1930 1933 *
1931 1934 * On NUMA machines, the partition's CPUs are consulted in order of
1932 1935 * distance from the current CPU. This way, the first available
1933 1936 * work found is also the closest, and will suffer the least
1934 1937 * from being migrated.
1935 1938 */
1936 1939 lpl = lpl_leaf = cp->cpu_lpl;
1937 1940 local_id = lpl_leaf->lpl_lgrpid;
1938 1941 leafidx = startidx = 0;
1939 1942
1940 1943 /*
1941 1944 * This loop traverses the lpl hierarchy. Higher level lpls represent
1942 1945 * broader levels of locality
1943 1946 */
1944 1947 do {
1945 1948 /* This loop iterates over the lpl's leaves */
1946 1949 do {
1947 1950 if (lpl_leaf != cp->cpu_lpl)
1948 1951 ocp = lpl_leaf->lpl_cpus;
1949 1952 else
1950 1953 ocp = cp->cpu_next_lpl;
1951 1954
1952 1955 /* This loop iterates over the CPUs in the leaf */
1953 1956 ocp_start = ocp;
1954 1957 do {
1955 1958 pri_t pri;
1956 1959
1957 1960 ASSERT(CPU_ACTIVE(ocp));
1958 1961
1959 1962 /*
1960 1963 * End our stroll around this lpl if:
1961 1964 *
1962 1965 * - Something became runnable on the local
1963 1966 * queue...which also ends our stroll around
1964 1967 * the partition.
1965 1968 *
1966 1969 * - We happen across another idle CPU.
1967 1970 * Since it is patrolling the next portion
1968 1971 * of the lpl's list (assuming it's not
1969 1972 * halted, or busy servicing an interrupt),
1970 1973 * move to the next higher level of locality.
1971 1974 */
1972 1975 if (cp->cpu_disp->disp_nrunnable != 0) {
1973 1976 kpreempt_enable();
1974 1977 return (NULL);
1975 1978 }
1976 1979 if (ocp->cpu_dispatch_pri == -1) {
1977 1980 if (ocp->cpu_disp_flags &
1978 1981 CPU_DISP_HALTED ||
1979 1982 ocp->cpu_intr_actv != 0)
1980 1983 continue;
1981 1984 else
1982 1985 goto next_level;
1983 1986 }
1984 1987
1985 1988 /*
1986 1989 * If there's only one thread and the CPU
1987 1990 * is in the middle of a context switch,
1988 1991 * or it's currently running the idle thread,
1989 1992 * don't steal it.
1990 1993 */
1991 1994 if ((ocp->cpu_disp_flags &
1992 1995 CPU_DISP_DONTSTEAL) &&
1993 1996 ocp->cpu_disp->disp_nrunnable == 1)
1994 1997 continue;
1995 1998
1996 1999 pri = ocp->cpu_disp->disp_max_unbound_pri;
1997 2000 if (pri > maxpri) {
1998 2001 /*
1999 2002 * Don't steal threads that we attempted
2000 2003 * to steal recently until they're ready
2001 2004 * to be stolen again.
2002 2005 */
2003 2006 stealtime = ocp->cpu_disp->disp_steal;
2004 2007 if (stealtime == 0 ||
2005 2008 stealtime - gethrtime() <= 0) {
2006 2009 maxpri = pri;
2007 2010 tcp = ocp;
2008 2011 } else {
2009 2012 /*
2010 2013 * Don't update tcp, just set
2011 2014 * the retval to T_DONTSTEAL, so
2012 2015 * that if no acceptable CPUs
2013 2016 * are found the return value
2014 2017 * will be T_DONTSTEAL rather
2015 2018 * then NULL.
2016 2019 */
2017 2020 retval = T_DONTSTEAL;
2018 2021 }
2019 2022 }
2020 2023 } while ((ocp = ocp->cpu_next_lpl) != ocp_start);
2021 2024
2022 2025 /*
2023 2026 * Iterate to the next leaf lpl in the resource set
2024 2027 * at this level of locality. If we hit the end of
2025 2028 * the set, wrap back around to the beginning.
2026 2029 *
2027 2030 * Note: This iteration is NULL terminated for a reason
2028 2031 * see lpl_topo_bootstrap() in lgrp.c for details.
2029 2032 */
2030 2033 if ((lpl_leaf = lpl->lpl_rset[++leafidx]) == NULL) {
2031 2034 leafidx = 0;
2032 2035 lpl_leaf = lpl->lpl_rset[leafidx];
2033 2036 }
2034 2037 } while (leafidx != startidx);
2035 2038
2036 2039 next_level:
2037 2040 /*
2038 2041 * Expand the search to include farther away CPUs (next
2039 2042 * locality level). The closer CPUs that have already been
2040 2043 * checked will be checked again. In doing so, idle CPUs
2041 2044 * will tend to be more aggresive about stealing from CPUs
2042 2045 * that are closer (since the closer CPUs will be considered
2043 2046 * more often).
2044 2047 * Begin at this level with the CPUs local leaf lpl.
2045 2048 */
2046 2049 if ((lpl = lpl->lpl_parent) != NULL) {
2047 2050 leafidx = startidx = lpl->lpl_id2rset[local_id];
2048 2051 lpl_leaf = lpl->lpl_rset[leafidx];
2049 2052 }
2050 2053 } while (!tcp && lpl);
2051 2054
2052 2055 kpreempt_enable();
2053 2056
2054 2057 /*
2055 2058 * If another queue looks good, and there is still nothing on
2056 2059 * the local queue, try to transfer one or more threads
2057 2060 * from it to our queue.
2058 2061 */
2059 2062 if (tcp && cp->cpu_disp->disp_nrunnable == 0) {
2060 2063 tp = disp_getbest(tcp->cpu_disp);
2061 2064 if (tp == NULL || tp == T_DONTSTEAL)
2062 2065 return (tp);
2063 2066 return (disp_ratify(tp, kpq));
2064 2067 }
2065 2068 return (retval);
2066 2069 }
2067 2070
2068 2071
2069 2072 /*
2070 2073 * disp_fix_unbound_pri()
2071 2074 * Determines the maximum priority of unbound threads on the queue.
2072 2075 * The priority is kept for the queue, but is only increased, never
2073 2076 * reduced unless some CPU is looking for something on that queue.
2074 2077 *
2075 2078 * The priority argument is the known upper limit.
2076 2079 *
2077 2080 * Perhaps this should be kept accurately, but that probably means
2078 2081 * separate bitmaps for bound and unbound threads. Since only idled
2079 2082 * CPUs will have to do this recalculation, it seems better this way.
2080 2083 */
2081 2084 static void
2082 2085 disp_fix_unbound_pri(disp_t *dp, pri_t pri)
2083 2086 {
2084 2087 kthread_t *tp;
2085 2088 dispq_t *dq;
2086 2089 ulong_t *dqactmap = dp->disp_qactmap;
2087 2090 ulong_t mapword;
2088 2091 int wx;
2089 2092
2090 2093 ASSERT(DISP_LOCK_HELD(&dp->disp_lock));
2091 2094
2092 2095 ASSERT(pri >= 0); /* checked by caller */
2093 2096
2094 2097 /*
2095 2098 * Start the search at the next lowest priority below the supplied
2096 2099 * priority. This depends on the bitmap implementation.
2097 2100 */
2098 2101 do {
2099 2102 wx = pri >> BT_ULSHIFT; /* index of word in map */
2100 2103
2101 2104 /*
2102 2105 * Form mask for all lower priorities in the word.
2103 2106 */
2104 2107 mapword = dqactmap[wx] & (BT_BIW(pri) - 1);
2105 2108
2106 2109 /*
2107 2110 * Get next lower active priority.
2108 2111 */
2109 2112 if (mapword != 0) {
2110 2113 pri = (wx << BT_ULSHIFT) + highbit(mapword) - 1;
2111 2114 } else if (wx > 0) {
2112 2115 pri = bt_gethighbit(dqactmap, wx - 1); /* sign extend */
2113 2116 if (pri < 0)
2114 2117 break;
2115 2118 } else {
2116 2119 pri = -1;
2117 2120 break;
2118 2121 }
2119 2122
2120 2123 /*
2121 2124 * Search the queue for unbound, runnable threads.
2122 2125 */
2123 2126 dq = &dp->disp_q[pri];
2124 2127 tp = dq->dq_first;
2125 2128
2126 2129 while (tp && (tp->t_bound_cpu || tp->t_weakbound_cpu)) {
2127 2130 tp = tp->t_link;
2128 2131 }
2129 2132
2130 2133 /*
2131 2134 * If a thread was found, set the priority and return.
2132 2135 */
2133 2136 } while (tp == NULL);
2134 2137
2135 2138 /*
2136 2139 * pri holds the maximum unbound thread priority or -1.
2137 2140 */
2138 2141 if (dp->disp_max_unbound_pri != pri)
2139 2142 dp->disp_max_unbound_pri = pri;
2140 2143 }
2141 2144
2142 2145 /*
2143 2146 * disp_adjust_unbound_pri() - thread is becoming unbound, so we should
2144 2147 * check if the CPU to which is was previously bound should have
2145 2148 * its disp_max_unbound_pri increased.
2146 2149 */
2147 2150 void
2148 2151 disp_adjust_unbound_pri(kthread_t *tp)
2149 2152 {
2150 2153 disp_t *dp;
2151 2154 pri_t tpri;
2152 2155
2153 2156 ASSERT(THREAD_LOCK_HELD(tp));
2154 2157
2155 2158 /*
2156 2159 * Don't do anything if the thread is not bound, or
2157 2160 * currently not runnable or swapped out.
2158 2161 */
2159 2162 if (tp->t_bound_cpu == NULL ||
2160 2163 tp->t_state != TS_RUN ||
2161 2164 tp->t_schedflag & TS_ON_SWAPQ)
2162 2165 return;
2163 2166
2164 2167 tpri = DISP_PRIO(tp);
2165 2168 dp = tp->t_bound_cpu->cpu_disp;
2166 2169 ASSERT(tpri >= 0 && tpri < dp->disp_npri);
2167 2170 if (tpri > dp->disp_max_unbound_pri)
2168 2171 dp->disp_max_unbound_pri = tpri;
2169 2172 }
2170 2173
2171 2174 /*
2172 2175 * disp_getbest()
2173 2176 * De-queue the highest priority unbound runnable thread.
2174 2177 * Returns with the thread unlocked and onproc but at splhigh (like disp()).
2175 2178 * Returns NULL if nothing found.
2176 2179 * Returns T_DONTSTEAL if the thread was not stealable.
2177 2180 * so that the caller will try again later.
2178 2181 *
2179 2182 * Passed a pointer to a dispatch queue not associated with this CPU, and
2180 2183 * its type.
2181 2184 */
2182 2185 static kthread_t *
2183 2186 disp_getbest(disp_t *dp)
2184 2187 {
2185 2188 kthread_t *tp;
2186 2189 dispq_t *dq;
2187 2190 pri_t pri;
2188 2191 cpu_t *cp, *tcp;
2189 2192 boolean_t allbound;
2190 2193
2191 2194 disp_lock_enter(&dp->disp_lock);
2192 2195
2193 2196 /*
2194 2197 * If there is nothing to run, or the CPU is in the middle of a
2195 2198 * context switch of the only thread, return NULL.
2196 2199 */
2197 2200 tcp = dp->disp_cpu;
2198 2201 cp = CPU;
2199 2202 pri = dp->disp_max_unbound_pri;
2200 2203 if (pri == -1 ||
2201 2204 (tcp != NULL && (tcp->cpu_disp_flags & CPU_DISP_DONTSTEAL) &&
2202 2205 tcp->cpu_disp->disp_nrunnable == 1)) {
2203 2206 disp_lock_exit_nopreempt(&dp->disp_lock);
2204 2207 return (NULL);
2205 2208 }
2206 2209
2207 2210 dq = &dp->disp_q[pri];
2208 2211
2209 2212
2210 2213 /*
2211 2214 * Assume that all threads are bound on this queue, and change it
2212 2215 * later when we find out that it is not the case.
2213 2216 */
2214 2217 allbound = B_TRUE;
2215 2218 for (tp = dq->dq_first; tp != NULL; tp = tp->t_link) {
2216 2219 hrtime_t now, nosteal, rqtime;
2217 2220
2218 2221 /*
2219 2222 * Skip over bound threads which could be here even
2220 2223 * though disp_max_unbound_pri indicated this level.
2221 2224 */
2222 2225 if (tp->t_bound_cpu || tp->t_weakbound_cpu)
2223 2226 continue;
2224 2227
2225 2228 /*
2226 2229 * We've got some unbound threads on this queue, so turn
2227 2230 * the allbound flag off now.
2228 2231 */
2229 2232 allbound = B_FALSE;
2230 2233
2231 2234 /*
2232 2235 * The thread is a candidate for stealing from its run queue. We
2233 2236 * don't want to steal threads that became runnable just a
2234 2237 * moment ago. This improves CPU affinity for threads that get
2235 2238 * preempted for short periods of time and go back on the run
2236 2239 * queue.
2237 2240 *
2238 2241 * We want to let it stay on its run queue if it was only placed
2239 2242 * there recently and it was running on the same CPU before that
2240 2243 * to preserve its cache investment. For the thread to remain on
2241 2244 * its run queue, ALL of the following conditions must be
2242 2245 * satisfied:
2243 2246 *
2244 2247 * - the disp queue should not be the kernel preemption queue
2245 2248 * - delayed idle stealing should not be disabled
2246 2249 * - nosteal_nsec should be non-zero
2247 2250 * - it should run with user priority
2248 2251 * - it should be on the run queue of the CPU where it was
2249 2252 * running before being placed on the run queue
2250 2253 * - it should be the only thread on the run queue (to prevent
2251 2254 * extra scheduling latency for other threads)
2252 2255 * - it should sit on the run queue for less than per-chip
2253 2256 * nosteal interval or global nosteal interval
2254 2257 * - in case of CPUs with shared cache it should sit in a run
2255 2258 * queue of a CPU from a different chip
2256 2259 *
2257 2260 * The checks are arranged so that the ones that are faster are
2258 2261 * placed earlier.
2259 2262 */
2260 2263 if (tcp == NULL ||
2261 2264 pri >= minclsyspri ||
2262 2265 tp->t_cpu != tcp)
2263 2266 break;
2264 2267
2265 2268 /*
2266 2269 * Steal immediately if, due to CMT processor architecture
2267 2270 * migraiton between cp and tcp would incur no performance
2268 2271 * penalty.
2269 2272 */
2270 2273 if (pg_cmt_can_migrate(cp, tcp))
2271 2274 break;
2272 2275
2273 2276 nosteal = nosteal_nsec;
2274 2277 if (nosteal == 0)
2275 2278 break;
2276 2279
2277 2280 /*
2278 2281 * Calculate time spent sitting on run queue
2279 2282 */
2280 2283 now = gethrtime_unscaled();
2281 2284 rqtime = now - tp->t_waitrq;
2282 2285 scalehrtime(&rqtime);
2283 2286
2284 2287 /*
2285 2288 * Steal immediately if the time spent on this run queue is more
2286 2289 * than allowed nosteal delay.
2287 2290 *
2288 2291 * Negative rqtime check is needed here to avoid infinite
2289 2292 * stealing delays caused by unlikely but not impossible
2290 2293 * drifts between CPU times on different CPUs.
2291 2294 */
2292 2295 if (rqtime > nosteal || rqtime < 0)
2293 2296 break;
2294 2297
2295 2298 DTRACE_PROBE4(nosteal, kthread_t *, tp,
2296 2299 cpu_t *, tcp, cpu_t *, cp, hrtime_t, rqtime);
2297 2300 scalehrtime(&now);
2298 2301 /*
2299 2302 * Calculate when this thread becomes stealable
2300 2303 */
2301 2304 now += (nosteal - rqtime);
2302 2305
2303 2306 /*
2304 2307 * Calculate time when some thread becomes stealable
2305 2308 */
2306 2309 if (now < dp->disp_steal)
2307 2310 dp->disp_steal = now;
2308 2311 }
2309 2312
2310 2313 /*
2311 2314 * If there were no unbound threads on this queue, find the queue
2312 2315 * where they are and then return later. The value of
2313 2316 * disp_max_unbound_pri is not always accurate because it isn't
2314 2317 * reduced until another idle CPU looks for work.
2315 2318 */
2316 2319 if (allbound)
2317 2320 disp_fix_unbound_pri(dp, pri);
2318 2321
2319 2322 /*
2320 2323 * If we reached the end of the queue and found no unbound threads
2321 2324 * then return NULL so that other CPUs will be considered. If there
2322 2325 * are unbound threads but they cannot yet be stolen, then
2323 2326 * return T_DONTSTEAL and try again later.
2324 2327 */
2325 2328 if (tp == NULL) {
2326 2329 disp_lock_exit_nopreempt(&dp->disp_lock);
2327 2330 return (allbound ? NULL : T_DONTSTEAL);
2328 2331 }
2329 2332
2330 2333 /*
2331 2334 * Found a runnable, unbound thread, so remove it from queue.
2332 2335 * dispdeq() requires that we have the thread locked, and we do,
2333 2336 * by virtue of holding the dispatch queue lock. dispdeq() will
2334 2337 * put the thread in transition state, thereby dropping the dispq
2335 2338 * lock.
2336 2339 */
2337 2340
2338 2341 #ifdef DEBUG
2339 2342 {
2340 2343 int thread_was_on_queue;
2341 2344
2342 2345 thread_was_on_queue = dispdeq(tp); /* drops disp_lock */
2343 2346 ASSERT(thread_was_on_queue);
2344 2347 }
2345 2348
2346 2349 #else /* DEBUG */
2347 2350 (void) dispdeq(tp); /* drops disp_lock */
2348 2351 #endif /* DEBUG */
2349 2352
2350 2353 /*
2351 2354 * Reset the disp_queue steal time - we do not know what is the smallest
2352 2355 * value across the queue is.
2353 2356 */
2354 2357 dp->disp_steal = 0;
2355 2358
2356 2359 tp->t_schedflag |= TS_DONT_SWAP;
2357 2360
2358 2361 /*
2359 2362 * Setup thread to run on the current CPU.
2360 2363 */
2361 2364 tp->t_disp_queue = cp->cpu_disp;
2362 2365
2363 2366 cp->cpu_dispthread = tp; /* protected by spl only */
2364 2367 cp->cpu_dispatch_pri = pri;
2365 2368
2366 2369 /*
2367 2370 * There can be a memory synchronization race between disp_getbest()
2368 2371 * and disp_ratify() vs cpu_resched() where cpu_resched() is trying
2369 2372 * to preempt the current thread to run the enqueued thread while
2370 2373 * disp_getbest() and disp_ratify() are changing the current thread
2371 2374 * to the stolen thread. This may lead to a situation where
2372 2375 * cpu_resched() tries to preempt the wrong thread and the
2373 2376 * stolen thread continues to run on the CPU which has been tagged
2374 2377 * for preemption.
2375 2378 * Later the clock thread gets enqueued but doesn't get to run on the
2376 2379 * CPU causing the system to hang.
2377 2380 *
2378 2381 * To avoid this, grabbing and dropping the disp_lock (which does
2379 2382 * a memory barrier) is needed to synchronize the execution of
2380 2383 * cpu_resched() with disp_getbest() and disp_ratify() and
2381 2384 * synchronize the memory read and written by cpu_resched(),
2382 2385 * disp_getbest(), and disp_ratify() with each other.
2383 2386 * (see CR#6482861 for more details).
2384 2387 */
2385 2388 disp_lock_enter_high(&cp->cpu_disp->disp_lock);
2386 2389 disp_lock_exit_high(&cp->cpu_disp->disp_lock);
2387 2390
2388 2391 ASSERT(pri == DISP_PRIO(tp));
2389 2392
2390 2393 DTRACE_PROBE3(steal, kthread_t *, tp, cpu_t *, tcp, cpu_t *, cp);
2391 2394
2392 2395 thread_onproc(tp, cp); /* set t_state to TS_ONPROC */
2393 2396
2394 2397 /*
2395 2398 * Return with spl high so that swtch() won't need to raise it.
2396 2399 * The disp_lock was dropped by dispdeq().
2397 2400 */
2398 2401
2399 2402 return (tp);
2400 2403 }
2401 2404
2402 2405 /*
2403 2406 * disp_bound_common() - common routine for higher level functions
2404 2407 * that check for bound threads under certain conditions.
2405 2408 * If 'threadlistsafe' is set then there is no need to acquire
2406 2409 * pidlock to stop the thread list from changing (eg, if
2407 2410 * disp_bound_* is called with cpus paused).
2408 2411 */
2409 2412 static int
2410 2413 disp_bound_common(cpu_t *cp, int threadlistsafe, int flag)
2411 2414 {
2412 2415 int found = 0;
2413 2416 kthread_t *tp;
2414 2417
2415 2418 ASSERT(flag);
2416 2419
2417 2420 if (!threadlistsafe)
2418 2421 mutex_enter(&pidlock);
2419 2422 tp = curthread; /* faster than allthreads */
2420 2423 do {
2421 2424 if (tp->t_state != TS_FREE) {
2422 2425 /*
2423 2426 * If an interrupt thread is busy, but the
2424 2427 * caller doesn't care (i.e. BOUND_INTR is off),
2425 2428 * then just ignore it and continue through.
2426 2429 */
2427 2430 if ((tp->t_flag & T_INTR_THREAD) &&
2428 2431 !(flag & BOUND_INTR))
2429 2432 continue;
2430 2433
2431 2434 /*
2432 2435 * Skip the idle thread for the CPU
2433 2436 * we're about to set offline.
2434 2437 */
2435 2438 if (tp == cp->cpu_idle_thread)
2436 2439 continue;
2437 2440
2438 2441 /*
2439 2442 * Skip the pause thread for the CPU
2440 2443 * we're about to set offline.
2441 2444 */
2442 2445 if (tp == cp->cpu_pause_thread)
2443 2446 continue;
2444 2447
2445 2448 if ((flag & BOUND_CPU) &&
2446 2449 (tp->t_bound_cpu == cp ||
2447 2450 tp->t_bind_cpu == cp->cpu_id ||
2448 2451 tp->t_weakbound_cpu == cp)) {
2449 2452 found = 1;
2450 2453 break;
2451 2454 }
2452 2455
2453 2456 if ((flag & BOUND_PARTITION) &&
2454 2457 (tp->t_cpupart == cp->cpu_part)) {
2455 2458 found = 1;
2456 2459 break;
2457 2460 }
2458 2461 }
2459 2462 } while ((tp = tp->t_next) != curthread && found == 0);
2460 2463 if (!threadlistsafe)
2461 2464 mutex_exit(&pidlock);
2462 2465 return (found);
2463 2466 }
2464 2467
2465 2468 /*
2466 2469 * disp_bound_threads - return nonzero if threads are bound to the processor.
2467 2470 * Called infrequently. Keep this simple.
2468 2471 * Includes threads that are asleep or stopped but not onproc.
2469 2472 */
2470 2473 int
2471 2474 disp_bound_threads(cpu_t *cp, int threadlistsafe)
2472 2475 {
2473 2476 return (disp_bound_common(cp, threadlistsafe, BOUND_CPU));
2474 2477 }
2475 2478
2476 2479 /*
2477 2480 * disp_bound_anythreads - return nonzero if _any_ threads are bound
2478 2481 * to the given processor, including interrupt threads.
2479 2482 */
2480 2483 int
2481 2484 disp_bound_anythreads(cpu_t *cp, int threadlistsafe)
2482 2485 {
2483 2486 return (disp_bound_common(cp, threadlistsafe, BOUND_CPU | BOUND_INTR));
2484 2487 }
2485 2488
2486 2489 /*
2487 2490 * disp_bound_partition - return nonzero if threads are bound to the same
2488 2491 * partition as the processor.
2489 2492 * Called infrequently. Keep this simple.
2490 2493 * Includes threads that are asleep or stopped but not onproc.
2491 2494 */
2492 2495 int
2493 2496 disp_bound_partition(cpu_t *cp, int threadlistsafe)
2494 2497 {
2495 2498 return (disp_bound_common(cp, threadlistsafe, BOUND_PARTITION));
2496 2499 }
2497 2500
2498 2501 /*
2499 2502 * disp_cpu_inactive - make a CPU inactive by moving all of its unbound
2500 2503 * threads to other CPUs.
2501 2504 */
2502 2505 void
2503 2506 disp_cpu_inactive(cpu_t *cp)
2504 2507 {
2505 2508 kthread_t *tp;
2506 2509 disp_t *dp = cp->cpu_disp;
2507 2510 dispq_t *dq;
2508 2511 pri_t pri;
2509 2512 int wasonq;
2510 2513
2511 2514 disp_lock_enter(&dp->disp_lock);
2512 2515 while ((pri = dp->disp_max_unbound_pri) != -1) {
2513 2516 dq = &dp->disp_q[pri];
2514 2517 tp = dq->dq_first;
2515 2518
2516 2519 /*
2517 2520 * Skip over bound threads.
2518 2521 */
2519 2522 while (tp != NULL && tp->t_bound_cpu != NULL) {
2520 2523 tp = tp->t_link;
2521 2524 }
2522 2525
2523 2526 if (tp == NULL) {
2524 2527 /* disp_max_unbound_pri must be inaccurate, so fix it */
2525 2528 disp_fix_unbound_pri(dp, pri);
2526 2529 continue;
2527 2530 }
2528 2531
2529 2532 wasonq = dispdeq(tp); /* drops disp_lock */
2530 2533 ASSERT(wasonq);
2531 2534 ASSERT(tp->t_weakbound_cpu == NULL);
2532 2535
2533 2536 setbackdq(tp);
2534 2537 /*
2535 2538 * Called from cpu_offline:
2536 2539 *
2537 2540 * cp has already been removed from the list of active cpus
2538 2541 * and tp->t_cpu has been changed so there is no risk of
2539 2542 * tp ending up back on cp.
2540 2543 *
2541 2544 * Called from cpupart_move_cpu:
2542 2545 *
2543 2546 * The cpu has moved to a new cpupart. Any threads that
2544 2547 * were on it's dispatch queues before the move remain
2545 2548 * in the old partition and can't run in the new partition.
2546 2549 */
2547 2550 ASSERT(tp->t_cpu != cp);
2548 2551 thread_unlock(tp);
2549 2552
2550 2553 disp_lock_enter(&dp->disp_lock);
2551 2554 }
2552 2555 disp_lock_exit(&dp->disp_lock);
2553 2556 }
2554 2557
2555 2558 /*
2556 2559 * disp_lowpri_cpu - find CPU running the lowest priority thread.
2557 2560 * The hint passed in is used as a starting point so we don't favor
2558 2561 * CPU 0 or any other CPU. The caller should pass in the most recently
2559 2562 * used CPU for the thread.
2560 2563 *
2561 2564 * The lgroup and priority are used to determine the best CPU to run on
2562 2565 * in a NUMA machine. The lgroup specifies which CPUs are closest while
2563 2566 * the thread priority will indicate whether the thread will actually run
2564 2567 * there. To pick the best CPU, the CPUs inside and outside of the given
2565 2568 * lgroup which are running the lowest priority threads are found. The
2566 2569 * remote CPU is chosen only if the thread will not run locally on a CPU
2567 2570 * within the lgroup, but will run on the remote CPU. If the thread
2568 2571 * cannot immediately run on any CPU, the best local CPU will be chosen.
2569 2572 *
2570 2573 * The lpl specified also identifies the cpu partition from which
2571 2574 * disp_lowpri_cpu should select a CPU.
2572 2575 *
2573 2576 * curcpu is used to indicate that disp_lowpri_cpu is being called on
2574 2577 * behalf of the current thread. (curthread is looking for a new cpu)
2575 2578 * In this case, cpu_dispatch_pri for this thread's cpu should be
2576 2579 * ignored.
2577 2580 *
2578 2581 * If a cpu is the target of an offline request then try to avoid it.
2579 2582 *
2580 2583 * This function must be called at either high SPL, or with preemption
2581 2584 * disabled, so that the "hint" CPU cannot be removed from the online
2582 2585 * CPU list while we are traversing it.
2583 2586 */
2584 2587 cpu_t *
2585 2588 disp_lowpri_cpu(cpu_t *hint, lpl_t *lpl, pri_t tpri, cpu_t *curcpu)
2586 2589 {
2587 2590 cpu_t *bestcpu;
2588 2591 cpu_t *besthomecpu;
2589 2592 cpu_t *cp, *cpstart;
2590 2593
2591 2594 pri_t bestpri;
2592 2595 pri_t cpupri;
2593 2596
2594 2597 klgrpset_t done;
2595 2598 klgrpset_t cur_set;
2596 2599
2597 2600 lpl_t *lpl_iter, *lpl_leaf;
2598 2601 int i;
2599 2602
2600 2603 /*
2601 2604 * Scan for a CPU currently running the lowest priority thread.
2602 2605 * Cannot get cpu_lock here because it is adaptive.
2603 2606 * We do not require lock on CPU list.
2604 2607 */
2605 2608 ASSERT(hint != NULL);
2606 2609 ASSERT(lpl != NULL);
2607 2610 ASSERT(lpl->lpl_ncpu > 0);
2608 2611
2609 2612 /*
2610 2613 * First examine local CPUs. Note that it's possible the hint CPU
2611 2614 * passed in in remote to the specified home lgroup. If our priority
2612 2615 * isn't sufficient enough such that we can run immediately at home,
2613 2616 * then examine CPUs remote to our home lgroup.
2614 2617 * We would like to give preference to CPUs closest to "home".
2615 2618 * If we can't find a CPU where we'll run at a given level
2616 2619 * of locality, we expand our search to include the next level.
2617 2620 */
2618 2621 bestcpu = besthomecpu = NULL;
2619 2622 klgrpset_clear(done);
2620 2623 /* start with lpl we were passed */
2621 2624
2622 2625 lpl_iter = lpl;
2623 2626
2624 2627 do {
2625 2628
2626 2629 bestpri = SHRT_MAX;
2627 2630 klgrpset_clear(cur_set);
2628 2631
2629 2632 for (i = 0; i < lpl_iter->lpl_nrset; i++) {
2630 2633 lpl_leaf = lpl_iter->lpl_rset[i];
2631 2634 if (klgrpset_ismember(done, lpl_leaf->lpl_lgrpid))
2632 2635 continue;
2633 2636
2634 2637 klgrpset_add(cur_set, lpl_leaf->lpl_lgrpid);
2635 2638
2636 2639 if (hint->cpu_lpl == lpl_leaf)
2637 2640 cp = cpstart = hint;
2638 2641 else
2639 2642 cp = cpstart = lpl_leaf->lpl_cpus;
2640 2643
2641 2644 do {
2642 2645 if (cp == curcpu)
2643 2646 cpupri = -1;
2644 2647 else if (cp == cpu_inmotion)
2645 2648 cpupri = SHRT_MAX;
2646 2649 else
2647 2650 cpupri = cp->cpu_dispatch_pri;
2648 2651 if (cp->cpu_disp->disp_maxrunpri > cpupri)
2649 2652 cpupri = cp->cpu_disp->disp_maxrunpri;
2650 2653 if (cp->cpu_chosen_level > cpupri)
2651 2654 cpupri = cp->cpu_chosen_level;
2652 2655 if (cpupri < bestpri) {
2653 2656 if (CPU_IDLING(cpupri)) {
2654 2657 ASSERT((cp->cpu_flags &
2655 2658 CPU_QUIESCED) == 0);
2656 2659 return (cp);
2657 2660 }
2658 2661 bestcpu = cp;
2659 2662 bestpri = cpupri;
2660 2663 }
2661 2664 } while ((cp = cp->cpu_next_lpl) != cpstart);
2662 2665 }
2663 2666
2664 2667 if (bestcpu && (tpri > bestpri)) {
2665 2668 ASSERT((bestcpu->cpu_flags & CPU_QUIESCED) == 0);
2666 2669 return (bestcpu);
2667 2670 }
2668 2671 if (besthomecpu == NULL)
2669 2672 besthomecpu = bestcpu;
2670 2673 /*
2671 2674 * Add the lgrps we just considered to the "done" set
2672 2675 */
2673 2676 klgrpset_or(done, cur_set);
2674 2677
2675 2678 } while ((lpl_iter = lpl_iter->lpl_parent) != NULL);
2676 2679
2677 2680 /*
2678 2681 * The specified priority isn't high enough to run immediately
2679 2682 * anywhere, so just return the best CPU from the home lgroup.
2680 2683 */
2681 2684 ASSERT((besthomecpu->cpu_flags & CPU_QUIESCED) == 0);
2682 2685 return (besthomecpu);
2683 2686 }
2684 2687
2685 2688 /*
2686 2689 * This routine provides the generic idle cpu function for all processors.
2687 2690 * If a processor has some specific code to execute when idle (say, to stop
2688 2691 * the pipeline and save power) then that routine should be defined in the
2689 2692 * processors specific code (module_xx.c) and the global variable idle_cpu
2690 2693 * set to that function.
2691 2694 */
2692 2695 static void
2693 2696 generic_idle_cpu(void)
2694 2697 {
2695 2698 }
2696 2699
2697 2700 /*ARGSUSED*/
2698 2701 static void
2699 2702 generic_enq_thread(cpu_t *cpu, int bound)
2700 2703 {
2701 2704 }
|
↓ open down ↓ |
2568 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX