1 /*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21
22 /*
23 * Copyright (c) 1991, 2010, Oracle and/or its affiliates. All rights reserved.
24 */
25
26 #include <sys/types.h>
27 #include <sys/param.h>
28 #include <sys/sysmacros.h>
29 #include <sys/signal.h>
30 #include <sys/stack.h>
31 #include <sys/pcb.h>
32 #include <sys/user.h>
33 #include <sys/systm.h>
34 #include <sys/sysinfo.h>
35 #include <sys/errno.h>
36 #include <sys/cmn_err.h>
37 #include <sys/cred.h>
38 #include <sys/resource.h>
39 #include <sys/task.h>
40 #include <sys/project.h>
41 #include <sys/proc.h>
42 #include <sys/debug.h>
43 #include <sys/disp.h>
44 #include <sys/class.h>
45 #include <vm/seg_kmem.h>
46 #include <vm/seg_kp.h>
47 #include <sys/machlock.h>
48 #include <sys/kmem.h>
49 #include <sys/varargs.h>
50 #include <sys/turnstile.h>
51 #include <sys/poll.h>
52 #include <sys/vtrace.h>
53 #include <sys/callb.h>
54 #include <c2/audit.h>
55 #include <sys/tnf.h>
56 #include <sys/sobject.h>
57 #include <sys/cpupart.h>
58 #include <sys/pset.h>
59 #include <sys/door.h>
60 #include <sys/spl.h>
61 #include <sys/copyops.h>
62 #include <sys/rctl.h>
63 #include <sys/brand.h>
64 #include <sys/pool.h>
65 #include <sys/zone.h>
66 #include <sys/tsol/label.h>
67 #include <sys/tsol/tndb.h>
68 #include <sys/cpc_impl.h>
69 #include <sys/sdt.h>
70 #include <sys/reboot.h>
71 #include <sys/kdi.h>
72 #include <sys/schedctl.h>
73 #include <sys/waitq.h>
74 #include <sys/cpucaps.h>
75 #include <sys/kiconv.h>
76
77 struct kmem_cache *thread_cache; /* cache of free threads */
78 struct kmem_cache *lwp_cache; /* cache of free lwps */
79 struct kmem_cache *turnstile_cache; /* cache of free turnstiles */
80
81 /*
82 * allthreads is only for use by kmem_readers. All kernel loops can use
83 * the current thread as a start/end point.
84 */
85 static kthread_t *allthreads = &t0; /* circular list of all threads */
86
87 static kcondvar_t reaper_cv; /* synchronization var */
88 kthread_t *thread_deathrow; /* circular list of reapable threads */
89 kthread_t *lwp_deathrow; /* circular list of reapable threads */
90 kmutex_t reaplock; /* protects lwp and thread deathrows */
91 int thread_reapcnt = 0; /* number of threads on deathrow */
92 int lwp_reapcnt = 0; /* number of lwps on deathrow */
93 int reaplimit = 16; /* delay reaping until reaplimit */
94
95 thread_free_lock_t *thread_free_lock;
96 /* protects tick thread from reaper */
97
98 extern int nthread;
99
100 /* System Scheduling classes. */
101 id_t syscid; /* system scheduling class ID */
102 id_t sysdccid = CLASS_UNUSED; /* reset when SDC loads */
103
104 void *segkp_thread; /* cookie for segkp pool */
105
106 int lwp_cache_sz = 32;
107 int t_cache_sz = 8;
108 static kt_did_t next_t_id = 1;
109
110 /* Default mode for thread binding to CPUs and processor sets */
111 int default_binding_mode = TB_ALLHARD;
112
113 /*
114 * Min/Max stack sizes for stack size parameters
115 */
116 #define MAX_STKSIZE (32 * DEFAULTSTKSZ)
117 #define MIN_STKSIZE DEFAULTSTKSZ
118
119 /*
120 * default_stksize overrides lwp_default_stksize if it is set.
121 */
122 int default_stksize;
123 int lwp_default_stksize;
124
125 static zone_key_t zone_thread_key;
126
127 unsigned int kmem_stackinfo; /* stackinfo feature on-off */
128 kmem_stkinfo_t *kmem_stkinfo_log; /* stackinfo circular log */
129 static kmutex_t kmem_stkinfo_lock; /* protects kmem_stkinfo_log */
130
131 /*
132 * forward declarations for internal thread specific data (tsd)
133 */
134 static void *tsd_realloc(void *, size_t, size_t);
135
136 void thread_reaper(void);
137
138 /* forward declarations for stackinfo feature */
139 static void stkinfo_begin(kthread_t *);
140 static void stkinfo_end(kthread_t *);
141 static size_t stkinfo_percent(caddr_t, caddr_t, caddr_t);
142
143 /*ARGSUSED*/
144 static int
145 turnstile_constructor(void *buf, void *cdrarg, int kmflags)
146 {
147 bzero(buf, sizeof (turnstile_t));
148 return (0);
149 }
150
151 /*ARGSUSED*/
152 static void
153 turnstile_destructor(void *buf, void *cdrarg)
154 {
155 turnstile_t *ts = buf;
156
157 ASSERT(ts->ts_free == NULL);
158 ASSERT(ts->ts_waiters == 0);
159 ASSERT(ts->ts_inheritor == NULL);
160 ASSERT(ts->ts_sleepq[0].sq_first == NULL);
161 ASSERT(ts->ts_sleepq[1].sq_first == NULL);
162 }
163
164 void
165 thread_init(void)
166 {
167 kthread_t *tp;
168 extern char sys_name[];
169 extern void idle();
170 struct cpu *cpu = CPU;
171 int i;
172 kmutex_t *lp;
173
174 mutex_init(&reaplock, NULL, MUTEX_SPIN, (void *)ipltospl(DISP_LEVEL));
175 thread_free_lock =
176 kmem_alloc(sizeof (thread_free_lock_t) * THREAD_FREE_NUM, KM_SLEEP);
177 for (i = 0; i < THREAD_FREE_NUM; i++) {
178 lp = &thread_free_lock[i].tf_lock;
179 mutex_init(lp, NULL, MUTEX_DEFAULT, NULL);
180 }
181
182 #if defined(__i386) || defined(__amd64)
183 thread_cache = kmem_cache_create("thread_cache", sizeof (kthread_t),
184 PTR24_ALIGN, NULL, NULL, NULL, NULL, NULL, 0);
185
186 /*
187 * "struct _klwp" includes a "struct pcb", which includes a
188 * "struct fpu", which needs to be 64-byte aligned on amd64
189 * (and even on i386) for xsave/xrstor.
190 */
191 lwp_cache = kmem_cache_create("lwp_cache", sizeof (klwp_t),
192 64, NULL, NULL, NULL, NULL, NULL, 0);
193 #else
194 /*
195 * Allocate thread structures from static_arena. This prevents
196 * issues where a thread tries to relocate its own thread
197 * structure and touches it after the mapping has been suspended.
198 */
199 thread_cache = kmem_cache_create("thread_cache", sizeof (kthread_t),
200 PTR24_ALIGN, NULL, NULL, NULL, NULL, static_arena, 0);
201
202 lwp_stk_cache_init();
203
204 lwp_cache = kmem_cache_create("lwp_cache", sizeof (klwp_t),
205 0, NULL, NULL, NULL, NULL, NULL, 0);
206 #endif
207
208 turnstile_cache = kmem_cache_create("turnstile_cache",
209 sizeof (turnstile_t), 0,
210 turnstile_constructor, turnstile_destructor, NULL, NULL, NULL, 0);
211
212 label_init();
213 cred_init();
214
215 /*
216 * Initialize various resource management facilities.
217 */
218 rctl_init();
219 cpucaps_init();
220 /*
221 * Zone_init() should be called before project_init() so that project ID
222 * for the first project is initialized correctly.
223 */
224 zone_init();
225 project_init();
226 brand_init();
227 kiconv_init();
228 task_init();
229 tcache_init();
230 pool_init();
231
232 curthread->t_ts = kmem_cache_alloc(turnstile_cache, KM_SLEEP);
233
234 /*
235 * Originally, we had two parameters to set default stack
236 * size: one for lwp's (lwp_default_stksize), and one for
237 * kernel-only threads (DEFAULTSTKSZ, a.k.a. _defaultstksz).
238 * Now we have a third parameter that overrides both if it is
239 * set to a legal stack size, called default_stksize.
240 */
241
242 if (default_stksize == 0) {
243 default_stksize = DEFAULTSTKSZ;
244 } else if (default_stksize % PAGESIZE != 0 ||
245 default_stksize > MAX_STKSIZE ||
246 default_stksize < MIN_STKSIZE) {
247 cmn_err(CE_WARN, "Illegal stack size. Using %d",
248 (int)DEFAULTSTKSZ);
249 default_stksize = DEFAULTSTKSZ;
250 } else {
251 lwp_default_stksize = default_stksize;
252 }
253
254 if (lwp_default_stksize == 0) {
255 lwp_default_stksize = default_stksize;
256 } else if (lwp_default_stksize % PAGESIZE != 0 ||
257 lwp_default_stksize > MAX_STKSIZE ||
258 lwp_default_stksize < MIN_STKSIZE) {
259 cmn_err(CE_WARN, "Illegal stack size. Using %d",
260 default_stksize);
261 lwp_default_stksize = default_stksize;
262 }
263
264 segkp_lwp = segkp_cache_init(segkp, lwp_cache_sz,
265 lwp_default_stksize,
266 (KPD_NOWAIT | KPD_HASREDZONE | KPD_LOCKED));
267
268 segkp_thread = segkp_cache_init(segkp, t_cache_sz,
269 default_stksize, KPD_HASREDZONE | KPD_LOCKED | KPD_NO_ANON);
270
271 (void) getcid(sys_name, &syscid);
272 curthread->t_cid = syscid; /* current thread is t0 */
273
274 /*
275 * Set up the first CPU's idle thread.
276 * It runs whenever the CPU has nothing worthwhile to do.
277 */
278 tp = thread_create(NULL, 0, idle, NULL, 0, &p0, TS_STOPPED, -1);
279 cpu->cpu_idle_thread = tp;
280 tp->t_preempt = 1;
281 tp->t_disp_queue = cpu->cpu_disp;
282 ASSERT(tp->t_disp_queue != NULL);
283 tp->t_bound_cpu = cpu;
284 tp->t_affinitycnt = 1;
285
286 /*
287 * Registering a thread in the callback table is usually
288 * done in the initialization code of the thread. In this
289 * case, we do it right after thread creation to avoid
290 * blocking idle thread while registering itself. It also
291 * avoids the possibility of reregistration in case a CPU
292 * restarts its idle thread.
293 */
294 CALLB_CPR_INIT_SAFE(tp, "idle");
295
296 /*
297 * Create the thread_reaper daemon. From this point on, exited
298 * threads will get reaped.
299 */
300 (void) thread_create(NULL, 0, (void (*)())thread_reaper,
301 NULL, 0, &p0, TS_RUN, minclsyspri);
302
303 /*
304 * Finish initializing the kernel memory allocator now that
305 * thread_create() is available.
306 */
307 kmem_thread_init();
308
309 if (boothowto & RB_DEBUG)
310 kdi_dvec_thravail();
311 }
312
313 /*
314 * Create a thread.
315 *
316 * thread_create() blocks for memory if necessary. It never fails.
317 *
318 * If stk is NULL, the thread is created at the base of the stack
319 * and cannot be swapped.
320 */
321 kthread_t *
322 thread_create(
323 caddr_t stk,
324 size_t stksize,
325 void (*proc)(),
326 void *arg,
327 size_t len,
328 proc_t *pp,
329 int state,
330 pri_t pri)
331 {
332 kthread_t *t;
333 extern struct classfuncs sys_classfuncs;
334 turnstile_t *ts;
335
336 /*
337 * Every thread keeps a turnstile around in case it needs to block.
338 * The only reason the turnstile is not simply part of the thread
339 * structure is that we may have to break the association whenever
340 * more than one thread blocks on a given synchronization object.
341 * From a memory-management standpoint, turnstiles are like the
342 * "attached mblks" that hang off dblks in the streams allocator.
343 */
344 ts = kmem_cache_alloc(turnstile_cache, KM_SLEEP);
345
346 if (stk == NULL) {
347 /*
348 * alloc both thread and stack in segkp chunk
349 */
350
351 if (stksize < default_stksize)
352 stksize = default_stksize;
353
354 if (stksize == default_stksize) {
355 stk = (caddr_t)segkp_cache_get(segkp_thread);
356 } else {
357 stksize = roundup(stksize, PAGESIZE);
358 stk = (caddr_t)segkp_get(segkp, stksize,
359 (KPD_HASREDZONE | KPD_NO_ANON | KPD_LOCKED));
360 }
361
362 ASSERT(stk != NULL);
363
364 /*
365 * The machine-dependent mutex code may require that
366 * thread pointers (since they may be used for mutex owner
367 * fields) have certain alignment requirements.
368 * PTR24_ALIGN is the size of the alignment quanta.
369 * XXX - assumes stack grows toward low addresses.
370 */
371 if (stksize <= sizeof (kthread_t) + PTR24_ALIGN)
372 cmn_err(CE_PANIC, "thread_create: proposed stack size"
373 " too small to hold thread.");
374 #ifdef STACK_GROWTH_DOWN
375 stksize -= SA(sizeof (kthread_t) + PTR24_ALIGN - 1);
376 stksize &= -PTR24_ALIGN; /* make thread aligned */
377 t = (kthread_t *)(stk + stksize);
378 bzero(t, sizeof (kthread_t));
379 if (audit_active)
380 audit_thread_create(t);
381 t->t_stk = stk + stksize;
382 t->t_stkbase = stk;
383 #else /* stack grows to larger addresses */
384 stksize -= SA(sizeof (kthread_t));
385 t = (kthread_t *)(stk);
386 bzero(t, sizeof (kthread_t));
387 t->t_stk = stk + sizeof (kthread_t);
388 t->t_stkbase = stk + stksize + sizeof (kthread_t);
389 #endif /* STACK_GROWTH_DOWN */
390 t->t_flag |= T_TALLOCSTK;
391 t->t_swap = stk;
392 } else {
393 t = kmem_cache_alloc(thread_cache, KM_SLEEP);
394 bzero(t, sizeof (kthread_t));
395 ASSERT(((uintptr_t)t & (PTR24_ALIGN - 1)) == 0);
396 if (audit_active)
397 audit_thread_create(t);
398 /*
399 * Initialize t_stk to the kernel stack pointer to use
400 * upon entry to the kernel
401 */
402 #ifdef STACK_GROWTH_DOWN
403 t->t_stk = stk + stksize;
404 t->t_stkbase = stk;
405 #else
406 t->t_stk = stk; /* 3b2-like */
407 t->t_stkbase = stk + stksize;
408 #endif /* STACK_GROWTH_DOWN */
409 }
410
411 if (kmem_stackinfo != 0) {
412 stkinfo_begin(t);
413 }
414
415 t->t_ts = ts;
416
417 /*
418 * p_cred could be NULL if it thread_create is called before cred_init
419 * is called in main.
420 */
421 mutex_enter(&pp->p_crlock);
422 if (pp->p_cred)
423 crhold(t->t_cred = pp->p_cred);
424 mutex_exit(&pp->p_crlock);
425 t->t_start = gethrestime_sec();
426 t->t_startpc = proc;
427 t->t_procp = pp;
428 t->t_clfuncs = &sys_classfuncs.thread;
429 t->t_cid = syscid;
430 t->t_pri = pri;
431 t->t_stime = ddi_get_lbolt();
432 t->t_schedflag = TS_LOAD | TS_DONT_SWAP;
433 t->t_bind_cpu = PBIND_NONE;
434 t->t_bindflag = (uchar_t)default_binding_mode;
435 t->t_bind_pset = PS_NONE;
436 t->t_plockp = &pp->p_lock;
437 t->t_copyops = NULL;
438 t->t_taskq = NULL;
439 t->t_anttime = 0;
440 t->t_hatdepth = 0;
441
442 t->t_dtrace_vtime = 1; /* assure vtimestamp is always non-zero */
443
444 CPU_STATS_ADDQ(CPU, sys, nthreads, 1);
445 #ifndef NPROBE
446 /* Kernel probe */
447 tnf_thread_create(t);
448 #endif /* NPROBE */
449 LOCK_INIT_CLEAR(&t->t_lock);
450
451 /*
452 * Callers who give us a NULL proc must do their own
453 * stack initialization. e.g. lwp_create()
454 */
455 if (proc != NULL) {
456 t->t_stk = thread_stk_init(t->t_stk);
457 thread_load(t, proc, arg, len);
458 }
459
460 /*
461 * Put a hold on project0. If this thread is actually in a
462 * different project, then t_proj will be changed later in
463 * lwp_create(). All kernel-only threads must be in project 0.
464 */
465 t->t_proj = project_hold(proj0p);
466
467 lgrp_affinity_init(&t->t_lgrp_affinity);
468
469 mutex_enter(&pidlock);
470 nthread++;
471 t->t_did = next_t_id++;
472 t->t_prev = curthread->t_prev;
473 t->t_next = curthread;
474
475 /*
476 * Add the thread to the list of all threads, and initialize
477 * its t_cpu pointer. We need to block preemption since
478 * cpu_offline walks the thread list looking for threads
479 * with t_cpu pointing to the CPU being offlined. We want
480 * to make sure that the list is consistent and that if t_cpu
481 * is set, the thread is on the list.
482 */
483 kpreempt_disable();
484 curthread->t_prev->t_next = t;
485 curthread->t_prev = t;
486
487 /*
488 * Threads should never have a NULL t_cpu pointer so assign it
489 * here. If the thread is being created with state TS_RUN a
490 * better CPU may be chosen when it is placed on the run queue.
491 *
492 * We need to keep kernel preemption disabled when setting all
493 * three fields to keep them in sync. Also, always create in
494 * the default partition since that's where kernel threads go
495 * (if this isn't a kernel thread, t_cpupart will be changed
496 * in lwp_create before setting the thread runnable).
497 */
498 t->t_cpupart = &cp_default;
499
500 /*
501 * For now, affiliate this thread with the root lgroup.
502 * Since the kernel does not (presently) allocate its memory
503 * in a locality aware fashion, the root is an appropriate home.
504 * If this thread is later associated with an lwp, it will have
505 * it's lgroup re-assigned at that time.
506 */
507 lgrp_move_thread(t, &cp_default.cp_lgrploads[LGRP_ROOTID], 1);
508
509 /*
510 * Inherit the current cpu. If this cpu isn't part of the chosen
511 * lgroup, a new cpu will be chosen by cpu_choose when the thread
512 * is ready to run.
513 */
514 if (CPU->cpu_part == &cp_default)
515 t->t_cpu = CPU;
516 else
517 t->t_cpu = disp_lowpri_cpu(cp_default.cp_cpulist, t->t_lpl,
518 t->t_pri, NULL);
519
520 t->t_disp_queue = t->t_cpu->cpu_disp;
521 kpreempt_enable();
522
523 /*
524 * Initialize thread state and the dispatcher lock pointer.
525 * Need to hold onto pidlock to block allthreads walkers until
526 * the state is set.
527 */
528 switch (state) {
529 case TS_RUN:
530 curthread->t_oldspl = splhigh(); /* get dispatcher spl */
531 THREAD_SET_STATE(t, TS_STOPPED, &transition_lock);
532 CL_SETRUN(t);
533 thread_unlock(t);
534 break;
535
536 case TS_ONPROC:
537 THREAD_ONPROC(t, t->t_cpu);
538 break;
539
540 case TS_FREE:
541 /*
542 * Free state will be used for intr threads.
543 * The interrupt routine must set the thread dispatcher
544 * lock pointer (t_lockp) if starting on a CPU
545 * other than the current one.
546 */
547 THREAD_FREEINTR(t, CPU);
548 break;
549
550 case TS_STOPPED:
551 THREAD_SET_STATE(t, TS_STOPPED, &stop_lock);
552 break;
553
554 default: /* TS_SLEEP, TS_ZOMB or TS_TRANS */
555 cmn_err(CE_PANIC, "thread_create: invalid state %d", state);
556 }
557 mutex_exit(&pidlock);
558 return (t);
559 }
560
561 /*
562 * Move thread to project0 and take care of project reference counters.
563 */
564 void
565 thread_rele(kthread_t *t)
566 {
567 kproject_t *kpj;
568
569 thread_lock(t);
570
571 ASSERT(t == curthread || t->t_state == TS_FREE || t->t_procp == &p0);
572 kpj = ttoproj(t);
573 t->t_proj = proj0p;
574
575 thread_unlock(t);
576
577 if (kpj != proj0p) {
578 project_rele(kpj);
579 (void) project_hold(proj0p);
580 }
581 }
582
583 void
584 thread_exit(void)
585 {
586 kthread_t *t = curthread;
587
588 if ((t->t_proc_flag & TP_ZTHREAD) != 0)
589 cmn_err(CE_PANIC, "thread_exit: zthread_exit() not called");
590
591 tsd_exit(); /* Clean up this thread's TSD */
592
593 kcpc_passivate(); /* clean up performance counter state */
594
595 /*
596 * No kernel thread should have called poll() without arranging
597 * calling pollcleanup() here.
598 */
599 ASSERT(t->t_pollstate == NULL);
600 ASSERT(t->t_schedctl == NULL);
601 if (t->t_door)
602 door_slam(); /* in case thread did an upcall */
603
604 #ifndef NPROBE
605 /* Kernel probe */
606 if (t->t_tnf_tpdp)
607 tnf_thread_exit();
608 #endif /* NPROBE */
609
610 thread_rele(t);
611 t->t_preempt++;
612
613 /*
614 * remove thread from the all threads list so that
615 * death-row can use the same pointers.
616 */
617 mutex_enter(&pidlock);
618 t->t_next->t_prev = t->t_prev;
619 t->t_prev->t_next = t->t_next;
620 ASSERT(allthreads != t); /* t0 never exits */
621 cv_broadcast(&t->t_joincv); /* wake up anyone in thread_join */
622 mutex_exit(&pidlock);
623
624 if (t->t_ctx != NULL)
625 exitctx(t);
626 if (t->t_procp->p_pctx != NULL)
627 exitpctx(t->t_procp);
628
629 if (kmem_stackinfo != 0) {
630 stkinfo_end(t);
631 }
632
633 t->t_state = TS_ZOMB; /* set zombie thread */
634
635 swtch_from_zombie(); /* give up the CPU */
636 /* NOTREACHED */
637 }
638
639 /*
640 * Check to see if the specified thread is active (defined as being on
641 * the thread list). This is certainly a slow way to do this; if there's
642 * ever a reason to speed it up, we could maintain a hash table of active
643 * threads indexed by their t_did.
644 */
645 static kthread_t *
646 did_to_thread(kt_did_t tid)
647 {
648 kthread_t *t;
649
650 ASSERT(MUTEX_HELD(&pidlock));
651 for (t = curthread->t_next; t != curthread; t = t->t_next) {
652 if (t->t_did == tid)
653 break;
654 }
655 if (t->t_did == tid)
656 return (t);
657 else
658 return (NULL);
659 }
660
661 /*
662 * Wait for specified thread to exit. Returns immediately if the thread
663 * could not be found, meaning that it has either already exited or never
664 * existed.
665 */
666 void
667 thread_join(kt_did_t tid)
668 {
669 kthread_t *t;
670
671 ASSERT(tid != curthread->t_did);
672 ASSERT(tid != t0.t_did);
673
674 mutex_enter(&pidlock);
675 /*
676 * Make sure we check that the thread is on the thread list
677 * before blocking on it; otherwise we could end up blocking on
678 * a cv that's already been freed. In other words, don't cache
679 * the thread pointer across calls to cv_wait.
680 *
681 * The choice of loop invariant means that whenever a thread
682 * is taken off the allthreads list, a cv_broadcast must be
683 * performed on that thread's t_joincv to wake up any waiters.
684 * The broadcast doesn't have to happen right away, but it
685 * shouldn't be postponed indefinitely (e.g., by doing it in
686 * thread_free which may only be executed when the deathrow
687 * queue is processed.
688 */
689 while (t = did_to_thread(tid))
690 cv_wait(&t->t_joincv, &pidlock);
691 mutex_exit(&pidlock);
692 }
693
694 void
695 thread_free_prevent(kthread_t *t)
696 {
697 kmutex_t *lp;
698
699 lp = &thread_free_lock[THREAD_FREE_HASH(t)].tf_lock;
700 mutex_enter(lp);
701 }
702
703 void
704 thread_free_allow(kthread_t *t)
705 {
706 kmutex_t *lp;
707
708 lp = &thread_free_lock[THREAD_FREE_HASH(t)].tf_lock;
709 mutex_exit(lp);
710 }
711
712 static void
713 thread_free_barrier(kthread_t *t)
714 {
715 kmutex_t *lp;
716
717 lp = &thread_free_lock[THREAD_FREE_HASH(t)].tf_lock;
718 mutex_enter(lp);
719 mutex_exit(lp);
720 }
721
722 void
723 thread_free(kthread_t *t)
724 {
725 boolean_t allocstk = (t->t_flag & T_TALLOCSTK);
726 klwp_t *lwp = t->t_lwp;
727 caddr_t swap = t->t_swap;
728
729 ASSERT(t != &t0 && t->t_state == TS_FREE);
730 ASSERT(t->t_door == NULL);
731 ASSERT(t->t_schedctl == NULL);
732 ASSERT(t->t_pollstate == NULL);
733
734 t->t_pri = 0;
735 t->t_pc = 0;
736 t->t_sp = 0;
737 t->t_wchan0 = NULL;
738 t->t_wchan = NULL;
739 if (t->t_cred != NULL) {
740 crfree(t->t_cred);
741 t->t_cred = 0;
742 }
743 if (t->t_pdmsg) {
744 kmem_free(t->t_pdmsg, strlen(t->t_pdmsg) + 1);
745 t->t_pdmsg = NULL;
746 }
747 if (audit_active)
748 audit_thread_free(t);
749 #ifndef NPROBE
750 if (t->t_tnf_tpdp)
751 tnf_thread_free(t);
752 #endif /* NPROBE */
753 if (t->t_cldata) {
754 CL_EXITCLASS(t->t_cid, (caddr_t *)t->t_cldata);
755 }
756 if (t->t_rprof != NULL) {
757 kmem_free(t->t_rprof, sizeof (*t->t_rprof));
758 t->t_rprof = NULL;
759 }
760 t->t_lockp = NULL; /* nothing should try to lock this thread now */
761 if (lwp)
762 lwp_freeregs(lwp, 0);
763 if (t->t_ctx)
764 freectx(t, 0);
765 t->t_stk = NULL;
766 if (lwp)
767 lwp_stk_fini(lwp);
768 lock_clear(&t->t_lock);
769
770 if (t->t_ts->ts_waiters > 0)
771 panic("thread_free: turnstile still active");
772
773 kmem_cache_free(turnstile_cache, t->t_ts);
774
775 free_afd(&t->t_activefd);
776
777 /*
778 * Barrier for the tick accounting code. The tick accounting code
779 * holds this lock to keep the thread from going away while it's
780 * looking at it.
781 */
782 thread_free_barrier(t);
783
784 ASSERT(ttoproj(t) == proj0p);
785 project_rele(ttoproj(t));
786
787 lgrp_affinity_free(&t->t_lgrp_affinity);
788
789 mutex_enter(&pidlock);
790 nthread--;
791 mutex_exit(&pidlock);
792
793 /*
794 * Free thread, lwp and stack. This needs to be done carefully, since
795 * if T_TALLOCSTK is set, the thread is part of the stack.
796 */
797 t->t_lwp = NULL;
798 t->t_swap = NULL;
799
800 if (swap) {
801 segkp_release(segkp, swap);
802 }
803 if (lwp) {
804 kmem_cache_free(lwp_cache, lwp);
805 }
806 if (!allocstk) {
807 kmem_cache_free(thread_cache, t);
808 }
809 }
810
811 /*
812 * Removes threads associated with the given zone from a deathrow queue.
813 * tp is a pointer to the head of the deathrow queue, and countp is a
814 * pointer to the current deathrow count. Returns a linked list of
815 * threads removed from the list.
816 */
817 static kthread_t *
818 thread_zone_cleanup(kthread_t **tp, int *countp, zoneid_t zoneid)
819 {
820 kthread_t *tmp, *list = NULL;
821 cred_t *cr;
822
823 ASSERT(MUTEX_HELD(&reaplock));
824 while (*tp != NULL) {
825 if ((cr = (*tp)->t_cred) != NULL && crgetzoneid(cr) == zoneid) {
826 tmp = *tp;
827 *tp = tmp->t_forw;
828 tmp->t_forw = list;
829 list = tmp;
830 (*countp)--;
831 } else {
832 tp = &(*tp)->t_forw;
833 }
834 }
835 return (list);
836 }
837
838 static void
839 thread_reap_list(kthread_t *t)
840 {
841 kthread_t *next;
842
843 while (t != NULL) {
844 next = t->t_forw;
845 thread_free(t);
846 t = next;
847 }
848 }
849
850 /* ARGSUSED */
851 static void
852 thread_zone_destroy(zoneid_t zoneid, void *unused)
853 {
854 kthread_t *t, *l;
855
856 mutex_enter(&reaplock);
857 /*
858 * Pull threads and lwps associated with zone off deathrow lists.
859 */
860 t = thread_zone_cleanup(&thread_deathrow, &thread_reapcnt, zoneid);
861 l = thread_zone_cleanup(&lwp_deathrow, &lwp_reapcnt, zoneid);
862 mutex_exit(&reaplock);
863
864 /*
865 * Guard against race condition in mutex_owner_running:
866 * thread=owner(mutex)
867 * <interrupt>
868 * thread exits mutex
869 * thread exits
870 * thread reaped
871 * thread struct freed
872 * cpu = thread->t_cpu <- BAD POINTER DEREFERENCE.
873 * A cross call to all cpus will cause the interrupt handler
874 * to reset the PC if it is in mutex_owner_running, refreshing
875 * stale thread pointers.
876 */
877 mutex_sync(); /* sync with mutex code */
878
879 /*
880 * Reap threads
881 */
882 thread_reap_list(t);
883
884 /*
885 * Reap lwps
886 */
887 thread_reap_list(l);
888 }
889
890 /*
891 * cleanup zombie threads that are on deathrow.
892 */
893 void
894 thread_reaper()
895 {
896 kthread_t *t, *l;
897 callb_cpr_t cprinfo;
898
899 /*
900 * Register callback to clean up threads when zone is destroyed.
901 */
902 zone_key_create(&zone_thread_key, NULL, NULL, thread_zone_destroy);
903
904 CALLB_CPR_INIT(&cprinfo, &reaplock, callb_generic_cpr, "t_reaper");
905 for (;;) {
906 mutex_enter(&reaplock);
907 while (thread_deathrow == NULL && lwp_deathrow == NULL) {
908 CALLB_CPR_SAFE_BEGIN(&cprinfo);
909 cv_wait(&reaper_cv, &reaplock);
910 CALLB_CPR_SAFE_END(&cprinfo, &reaplock);
911 }
912 /*
913 * mutex_sync() needs to be called when reaping, but
914 * not too often. We limit reaping rate to once
915 * per second. Reaplimit is max rate at which threads can
916 * be freed. Does not impact thread destruction/creation.
917 */
918 t = thread_deathrow;
919 l = lwp_deathrow;
920 thread_deathrow = NULL;
921 lwp_deathrow = NULL;
922 thread_reapcnt = 0;
923 lwp_reapcnt = 0;
924 mutex_exit(&reaplock);
925
926 /*
927 * Guard against race condition in mutex_owner_running:
928 * thread=owner(mutex)
929 * <interrupt>
930 * thread exits mutex
931 * thread exits
932 * thread reaped
933 * thread struct freed
934 * cpu = thread->t_cpu <- BAD POINTER DEREFERENCE.
935 * A cross call to all cpus will cause the interrupt handler
936 * to reset the PC if it is in mutex_owner_running, refreshing
937 * stale thread pointers.
938 */
939 mutex_sync(); /* sync with mutex code */
940 /*
941 * Reap threads
942 */
943 thread_reap_list(t);
944
945 /*
946 * Reap lwps
947 */
948 thread_reap_list(l);
949 delay(hz);
950 }
951 }
952
953 /*
954 * This is called by lwpcreate, etc.() to put a lwp_deathrow thread onto
955 * thread_deathrow. The thread's state is changed already TS_FREE to indicate
956 * that is reapable. The thread already holds the reaplock, and was already
957 * freed.
958 */
959 void
960 reapq_move_lq_to_tq(kthread_t *t)
961 {
962 ASSERT(t->t_state == TS_FREE);
963 ASSERT(MUTEX_HELD(&reaplock));
964 t->t_forw = thread_deathrow;
965 thread_deathrow = t;
966 thread_reapcnt++;
967 if (lwp_reapcnt + thread_reapcnt > reaplimit)
968 cv_signal(&reaper_cv); /* wake the reaper */
969 }
970
971 /*
972 * This is called by resume() to put a zombie thread onto deathrow.
973 * The thread's state is changed to TS_FREE to indicate that is reapable.
974 * This is called from the idle thread so it must not block - just spin.
975 */
976 void
977 reapq_add(kthread_t *t)
978 {
979 mutex_enter(&reaplock);
980
981 /*
982 * lwp_deathrow contains threads with lwp linkage and
983 * swappable thread stacks which have the default stacksize.
984 * These threads' lwps and stacks may be reused by lwp_create().
985 *
986 * Anything else goes on thread_deathrow(), where it will eventually
987 * be thread_free()d.
988 */
989 if (t->t_flag & T_LWPREUSE) {
990 ASSERT(ttolwp(t) != NULL);
991 t->t_forw = lwp_deathrow;
992 lwp_deathrow = t;
993 lwp_reapcnt++;
994 } else {
995 t->t_forw = thread_deathrow;
996 thread_deathrow = t;
997 thread_reapcnt++;
998 }
999 if (lwp_reapcnt + thread_reapcnt > reaplimit)
1000 cv_signal(&reaper_cv); /* wake the reaper */
1001 t->t_state = TS_FREE;
1002 lock_clear(&t->t_lock);
1003
1004 /*
1005 * Before we return, we need to grab and drop the thread lock for
1006 * the dead thread. At this point, the current thread is the idle
1007 * thread, and the dead thread's CPU lock points to the current
1008 * CPU -- and we must grab and drop the lock to synchronize with
1009 * a racing thread walking a blocking chain that the zombie thread
1010 * was recently in. By this point, that blocking chain is (by
1011 * definition) stale: the dead thread is not holding any locks, and
1012 * is therefore not in any blocking chains -- but if we do not regrab
1013 * our lock before freeing the dead thread's data structures, the
1014 * thread walking the (stale) blocking chain will die on memory
1015 * corruption when it attempts to drop the dead thread's lock. We
1016 * only need do this once because there is no way for the dead thread
1017 * to ever again be on a blocking chain: once we have grabbed and
1018 * dropped the thread lock, we are guaranteed that anyone that could
1019 * have seen this thread in a blocking chain can no longer see it.
1020 */
1021 thread_lock(t);
1022 thread_unlock(t);
1023
1024 mutex_exit(&reaplock);
1025 }
1026
1027 /*
1028 * Install thread context ops for the current thread.
1029 */
1030 void
1031 installctx(
1032 kthread_t *t,
1033 void *arg,
1034 void (*save)(void *),
1035 void (*restore)(void *),
1036 void (*fork)(void *, void *),
1037 void (*lwp_create)(void *, void *),
1038 void (*exit)(void *),
1039 void (*free)(void *, int))
1040 {
1041 struct ctxop *ctx;
1042
1043 ctx = kmem_alloc(sizeof (struct ctxop), KM_SLEEP);
1044 ctx->save_op = save;
1045 ctx->restore_op = restore;
1046 ctx->fork_op = fork;
1047 ctx->lwp_create_op = lwp_create;
1048 ctx->exit_op = exit;
1049 ctx->free_op = free;
1050 ctx->arg = arg;
1051 ctx->next = t->t_ctx;
1052 t->t_ctx = ctx;
1053 }
1054
1055 /*
1056 * Remove the thread context ops from a thread.
1057 */
1058 int
1059 removectx(
1060 kthread_t *t,
1061 void *arg,
1062 void (*save)(void *),
1063 void (*restore)(void *),
1064 void (*fork)(void *, void *),
1065 void (*lwp_create)(void *, void *),
1066 void (*exit)(void *),
1067 void (*free)(void *, int))
1068 {
1069 struct ctxop *ctx, *prev_ctx;
1070
1071 /*
1072 * The incoming kthread_t (which is the thread for which the
1073 * context ops will be removed) should be one of the following:
1074 *
1075 * a) the current thread,
1076 *
1077 * b) a thread of a process that's being forked (SIDL),
1078 *
1079 * c) a thread that belongs to the same process as the current
1080 * thread and for which the current thread is the agent thread,
1081 *
1082 * d) a thread that is TS_STOPPED which is indicative of it
1083 * being (if curthread is not an agent) a thread being created
1084 * as part of an lwp creation.
1085 */
1086 ASSERT(t == curthread || ttoproc(t)->p_stat == SIDL ||
1087 ttoproc(t)->p_agenttp == curthread || t->t_state == TS_STOPPED);
1088
1089 /*
1090 * Serialize modifications to t->t_ctx to prevent the agent thread
1091 * and the target thread from racing with each other during lwp exit.
1092 */
1093 mutex_enter(&t->t_ctx_lock);
1094 prev_ctx = NULL;
1095 for (ctx = t->t_ctx; ctx != NULL; ctx = ctx->next) {
1096 if (ctx->save_op == save && ctx->restore_op == restore &&
1097 ctx->fork_op == fork && ctx->lwp_create_op == lwp_create &&
1098 ctx->exit_op == exit && ctx->free_op == free &&
1099 ctx->arg == arg) {
1100 if (prev_ctx)
1101 prev_ctx->next = ctx->next;
1102 else
1103 t->t_ctx = ctx->next;
1104 mutex_exit(&t->t_ctx_lock);
1105 if (ctx->free_op != NULL)
1106 (ctx->free_op)(ctx->arg, 0);
1107 kmem_free(ctx, sizeof (struct ctxop));
1108 return (1);
1109 }
1110 prev_ctx = ctx;
1111 }
1112 mutex_exit(&t->t_ctx_lock);
1113
1114 return (0);
1115 }
1116
1117 void
1118 savectx(kthread_t *t)
1119 {
1120 struct ctxop *ctx;
1121
1122 ASSERT(t == curthread);
1123 for (ctx = t->t_ctx; ctx != 0; ctx = ctx->next)
1124 if (ctx->save_op != NULL)
1125 (ctx->save_op)(ctx->arg);
1126 }
1127
1128 void
1129 restorectx(kthread_t *t)
1130 {
1131 struct ctxop *ctx;
1132
1133 ASSERT(t == curthread);
1134 for (ctx = t->t_ctx; ctx != 0; ctx = ctx->next)
1135 if (ctx->restore_op != NULL)
1136 (ctx->restore_op)(ctx->arg);
1137 }
1138
1139 void
1140 forkctx(kthread_t *t, kthread_t *ct)
1141 {
1142 struct ctxop *ctx;
1143
1144 for (ctx = t->t_ctx; ctx != NULL; ctx = ctx->next)
1145 if (ctx->fork_op != NULL)
1146 (ctx->fork_op)(t, ct);
1147 }
1148
1149 /*
1150 * Note that this operator is only invoked via the _lwp_create
1151 * system call. The system may have other reasons to create lwps
1152 * e.g. the agent lwp or the doors unreferenced lwp.
1153 */
1154 void
1155 lwp_createctx(kthread_t *t, kthread_t *ct)
1156 {
1157 struct ctxop *ctx;
1158
1159 for (ctx = t->t_ctx; ctx != NULL; ctx = ctx->next)
1160 if (ctx->lwp_create_op != NULL)
1161 (ctx->lwp_create_op)(t, ct);
1162 }
1163
1164 /*
1165 * exitctx is called from thread_exit() and lwp_exit() to perform any actions
1166 * needed when the thread/LWP leaves the processor for the last time. This
1167 * routine is not intended to deal with freeing memory; freectx() is used for
1168 * that purpose during thread_free(). This routine is provided to allow for
1169 * clean-up that can't wait until thread_free().
1170 */
1171 void
1172 exitctx(kthread_t *t)
1173 {
1174 struct ctxop *ctx;
1175
1176 for (ctx = t->t_ctx; ctx != NULL; ctx = ctx->next)
1177 if (ctx->exit_op != NULL)
1178 (ctx->exit_op)(t);
1179 }
1180
1181 /*
1182 * freectx is called from thread_free() and exec() to get
1183 * rid of old thread context ops.
1184 */
1185 void
1186 freectx(kthread_t *t, int isexec)
1187 {
1188 struct ctxop *ctx;
1189
1190 while ((ctx = t->t_ctx) != NULL) {
1191 t->t_ctx = ctx->next;
1192 if (ctx->free_op != NULL)
1193 (ctx->free_op)(ctx->arg, isexec);
1194 kmem_free(ctx, sizeof (struct ctxop));
1195 }
1196 }
1197
1198 /*
1199 * freectx_ctx is called from lwp_create() when lwp is reused from
1200 * lwp_deathrow and its thread structure is added to thread_deathrow.
1201 * The thread structure to which this ctx was attached may be already
1202 * freed by the thread reaper so free_op implementations shouldn't rely
1203 * on thread structure to which this ctx was attached still being around.
1204 */
1205 void
1206 freectx_ctx(struct ctxop *ctx)
1207 {
1208 struct ctxop *nctx;
1209
1210 ASSERT(ctx != NULL);
1211
1212 do {
1213 nctx = ctx->next;
1214 if (ctx->free_op != NULL)
1215 (ctx->free_op)(ctx->arg, 0);
1216 kmem_free(ctx, sizeof (struct ctxop));
1217 } while ((ctx = nctx) != NULL);
1218 }
1219
1220 /*
1221 * Set the thread running; arrange for it to be swapped in if necessary.
1222 */
1223 void
1224 setrun_locked(kthread_t *t)
1225 {
1226 ASSERT(THREAD_LOCK_HELD(t));
1227 if (t->t_state == TS_SLEEP) {
1228 /*
1229 * Take off sleep queue.
1230 */
1231 SOBJ_UNSLEEP(t->t_sobj_ops, t);
1232 } else if (t->t_state & (TS_RUN | TS_ONPROC)) {
1233 /*
1234 * Already on dispatcher queue.
1235 */
1236 return;
1237 } else if (t->t_state == TS_WAIT) {
1238 waitq_setrun(t);
1239 } else if (t->t_state == TS_STOPPED) {
1240 /*
1241 * All of the sending of SIGCONT (TC_XSTART) and /proc
1242 * (TC_PSTART) and lwp_continue() (TC_CSTART) must have
1243 * requested that the thread be run.
1244 * Just calling setrun() is not sufficient to set a stopped
1245 * thread running. TP_TXSTART is always set if the thread
1246 * is not stopped by a jobcontrol stop signal.
1247 * TP_TPSTART is always set if /proc is not controlling it.
1248 * TP_TCSTART is always set if lwp_suspend() didn't stop it.
1249 * The thread won't be stopped unless one of these
1250 * three mechanisms did it.
1251 *
1252 * These flags must be set before calling setrun_locked(t).
1253 * They can't be passed as arguments because the streams
1254 * code calls setrun() indirectly and the mechanism for
1255 * doing so admits only one argument. Note that the
1256 * thread must be locked in order to change t_schedflags.
1257 */
1258 if ((t->t_schedflag & TS_ALLSTART) != TS_ALLSTART)
1259 return;
1260 /*
1261 * Process is no longer stopped (a thread is running).
1262 */
1263 t->t_whystop = 0;
1264 t->t_whatstop = 0;
1265 /*
1266 * Strictly speaking, we do not have to clear these
1267 * flags here; they are cleared on entry to stop().
1268 * However, they are confusing when doing kernel
1269 * debugging or when they are revealed by ps(1).
1270 */
1271 t->t_schedflag &= ~TS_ALLSTART;
1272 THREAD_TRANSITION(t); /* drop stopped-thread lock */
1273 ASSERT(t->t_lockp == &transition_lock);
1274 ASSERT(t->t_wchan0 == NULL && t->t_wchan == NULL);
1275 /*
1276 * Let the class put the process on the dispatcher queue.
1277 */
1278 CL_SETRUN(t);
1279 }
1280 }
1281
1282 void
1283 setrun(kthread_t *t)
1284 {
1285 thread_lock(t);
1286 setrun_locked(t);
1287 thread_unlock(t);
1288 }
1289
1290 /*
1291 * Unpin an interrupted thread.
1292 * When an interrupt occurs, the interrupt is handled on the stack
1293 * of an interrupt thread, taken from a pool linked to the CPU structure.
1294 *
1295 * When swtch() is switching away from an interrupt thread because it
1296 * blocked or was preempted, this routine is called to complete the
1297 * saving of the interrupted thread state, and returns the interrupted
1298 * thread pointer so it may be resumed.
1299 *
1300 * Called by swtch() only at high spl.
1301 */
1302 kthread_t *
1303 thread_unpin()
1304 {
1305 kthread_t *t = curthread; /* current thread */
1306 kthread_t *itp; /* interrupted thread */
1307 int i; /* interrupt level */
1308 extern int intr_passivate();
1309
1310 ASSERT(t->t_intr != NULL);
1311
1312 itp = t->t_intr; /* interrupted thread */
1313 t->t_intr = NULL; /* clear interrupt ptr */
1314
1315 /*
1316 * Get state from interrupt thread for the one
1317 * it interrupted.
1318 */
1319
1320 i = intr_passivate(t, itp);
1321
1322 TRACE_5(TR_FAC_INTR, TR_INTR_PASSIVATE,
1323 "intr_passivate:level %d curthread %p (%T) ithread %p (%T)",
1324 i, t, t, itp, itp);
1325
1326 /*
1327 * Dissociate the current thread from the interrupted thread's LWP.
1328 */
1329 t->t_lwp = NULL;
1330
1331 /*
1332 * Interrupt handlers above the level that spinlocks block must
1333 * not block.
1334 */
1335 #if DEBUG
1336 if (i < 0 || i > LOCK_LEVEL)
1337 cmn_err(CE_PANIC, "thread_unpin: ipl out of range %x", i);
1338 #endif
1339
1340 /*
1341 * Compute the CPU's base interrupt level based on the active
1342 * interrupts.
1343 */
1344 ASSERT(CPU->cpu_intr_actv & (1 << i));
1345 set_base_spl();
1346
1347 return (itp);
1348 }
1349
1350 /*
1351 * Create and initialize an interrupt thread.
1352 * Returns non-zero on error.
1353 * Called at spl7() or better.
1354 */
1355 void
1356 thread_create_intr(struct cpu *cp)
1357 {
1358 kthread_t *tp;
1359
1360 tp = thread_create(NULL, 0,
1361 (void (*)())thread_create_intr, NULL, 0, &p0, TS_ONPROC, 0);
1362
1363 /*
1364 * Set the thread in the TS_FREE state. The state will change
1365 * to TS_ONPROC only while the interrupt is active. Think of these
1366 * as being on a private free list for the CPU. Being TS_FREE keeps
1367 * inactive interrupt threads out of debugger thread lists.
1368 *
1369 * We cannot call thread_create with TS_FREE because of the current
1370 * checks there for ONPROC. Fix this when thread_create takes flags.
1371 */
1372 THREAD_FREEINTR(tp, cp);
1373
1374 /*
1375 * Nobody should ever reference the credentials of an interrupt
1376 * thread so make it NULL to catch any such references.
1377 */
1378 tp->t_cred = NULL;
1379 tp->t_flag |= T_INTR_THREAD;
1380 tp->t_cpu = cp;
1381 tp->t_bound_cpu = cp;
1382 tp->t_disp_queue = cp->cpu_disp;
1383 tp->t_affinitycnt = 1;
1384 tp->t_preempt = 1;
1385
1386 /*
1387 * Don't make a user-requested binding on this thread so that
1388 * the processor can be offlined.
1389 */
1390 tp->t_bind_cpu = PBIND_NONE; /* no USER-requested binding */
1391 tp->t_bind_pset = PS_NONE;
1392
1393 #if defined(__i386) || defined(__amd64)
1394 tp->t_stk -= STACK_ALIGN;
1395 *(tp->t_stk) = 0; /* terminate intr thread stack */
1396 #endif
1397
1398 /*
1399 * Link onto CPU's interrupt pool.
1400 */
1401 tp->t_link = cp->cpu_intr_thread;
1402 cp->cpu_intr_thread = tp;
1403 }
1404
1405 /*
1406 * TSD -- THREAD SPECIFIC DATA
1407 */
1408 static kmutex_t tsd_mutex; /* linked list spin lock */
1409 static uint_t tsd_nkeys; /* size of destructor array */
1410 /* per-key destructor funcs */
1411 static void (**tsd_destructor)(void *);
1412 /* list of tsd_thread's */
1413 static struct tsd_thread *tsd_list;
1414
1415 /*
1416 * Default destructor
1417 * Needed because NULL destructor means that the key is unused
1418 */
1419 /* ARGSUSED */
1420 void
1421 tsd_defaultdestructor(void *value)
1422 {}
1423
1424 /*
1425 * Create a key (index into per thread array)
1426 * Locks out tsd_create, tsd_destroy, and tsd_exit
1427 * May allocate memory with lock held
1428 */
1429 void
1430 tsd_create(uint_t *keyp, void (*destructor)(void *))
1431 {
1432 int i;
1433 uint_t nkeys;
1434
1435 /*
1436 * if key is allocated, do nothing
1437 */
1438 mutex_enter(&tsd_mutex);
1439 if (*keyp) {
1440 mutex_exit(&tsd_mutex);
1441 return;
1442 }
1443 /*
1444 * find an unused key
1445 */
1446 if (destructor == NULL)
1447 destructor = tsd_defaultdestructor;
1448
1449 for (i = 0; i < tsd_nkeys; ++i)
1450 if (tsd_destructor[i] == NULL)
1451 break;
1452
1453 /*
1454 * if no unused keys, increase the size of the destructor array
1455 */
1456 if (i == tsd_nkeys) {
1457 if ((nkeys = (tsd_nkeys << 1)) == 0)
1458 nkeys = 1;
1459 tsd_destructor =
1460 (void (**)(void *))tsd_realloc((void *)tsd_destructor,
1461 (size_t)(tsd_nkeys * sizeof (void (*)(void *))),
1462 (size_t)(nkeys * sizeof (void (*)(void *))));
1463 tsd_nkeys = nkeys;
1464 }
1465
1466 /*
1467 * allocate the next available unused key
1468 */
1469 tsd_destructor[i] = destructor;
1470 *keyp = i + 1;
1471 mutex_exit(&tsd_mutex);
1472 }
1473
1474 /*
1475 * Destroy a key -- this is for unloadable modules
1476 *
1477 * Assumes that the caller is preventing tsd_set and tsd_get
1478 * Locks out tsd_create, tsd_destroy, and tsd_exit
1479 * May free memory with lock held
1480 */
1481 void
1482 tsd_destroy(uint_t *keyp)
1483 {
1484 uint_t key;
1485 struct tsd_thread *tsd;
1486
1487 /*
1488 * protect the key namespace and our destructor lists
1489 */
1490 mutex_enter(&tsd_mutex);
1491 key = *keyp;
1492 *keyp = 0;
1493
1494 ASSERT(key <= tsd_nkeys);
1495
1496 /*
1497 * if the key is valid
1498 */
1499 if (key != 0) {
1500 uint_t k = key - 1;
1501 /*
1502 * for every thread with TSD, call key's destructor
1503 */
1504 for (tsd = tsd_list; tsd; tsd = tsd->ts_next) {
1505 /*
1506 * no TSD for key in this thread
1507 */
1508 if (key > tsd->ts_nkeys)
1509 continue;
1510 /*
1511 * call destructor for key
1512 */
1513 if (tsd->ts_value[k] && tsd_destructor[k])
1514 (*tsd_destructor[k])(tsd->ts_value[k]);
1515 /*
1516 * reset value for key
1517 */
1518 tsd->ts_value[k] = NULL;
1519 }
1520 /*
1521 * actually free the key (NULL destructor == unused)
1522 */
1523 tsd_destructor[k] = NULL;
1524 }
1525
1526 mutex_exit(&tsd_mutex);
1527 }
1528
1529 /*
1530 * Quickly return the per thread value that was stored with the specified key
1531 * Assumes the caller is protecting key from tsd_create and tsd_destroy
1532 */
1533 void *
1534 tsd_get(uint_t key)
1535 {
1536 return (tsd_agent_get(curthread, key));
1537 }
1538
1539 /*
1540 * Set a per thread value indexed with the specified key
1541 */
1542 int
1543 tsd_set(uint_t key, void *value)
1544 {
1545 return (tsd_agent_set(curthread, key, value));
1546 }
1547
1548 /*
1549 * Like tsd_get(), except that the agent lwp can get the tsd of
1550 * another thread in the same process (the agent thread only runs when the
1551 * process is completely stopped by /proc), or syslwp is creating a new lwp.
1552 */
1553 void *
1554 tsd_agent_get(kthread_t *t, uint_t key)
1555 {
1556 struct tsd_thread *tsd = t->t_tsd;
1557
1558 ASSERT(t == curthread ||
1559 ttoproc(t)->p_agenttp == curthread || t->t_state == TS_STOPPED);
1560
1561 if (key && tsd != NULL && key <= tsd->ts_nkeys)
1562 return (tsd->ts_value[key - 1]);
1563 return (NULL);
1564 }
1565
1566 /*
1567 * Like tsd_set(), except that the agent lwp can set the tsd of
1568 * another thread in the same process, or syslwp can set the tsd
1569 * of a thread it's in the middle of creating.
1570 *
1571 * Assumes the caller is protecting key from tsd_create and tsd_destroy
1572 * May lock out tsd_destroy (and tsd_create), may allocate memory with
1573 * lock held
1574 */
1575 int
1576 tsd_agent_set(kthread_t *t, uint_t key, void *value)
1577 {
1578 struct tsd_thread *tsd = t->t_tsd;
1579
1580 ASSERT(t == curthread ||
1581 ttoproc(t)->p_agenttp == curthread || t->t_state == TS_STOPPED);
1582
1583 if (key == 0)
1584 return (EINVAL);
1585 if (tsd == NULL)
1586 tsd = t->t_tsd = kmem_zalloc(sizeof (*tsd), KM_SLEEP);
1587 if (key <= tsd->ts_nkeys) {
1588 tsd->ts_value[key - 1] = value;
1589 return (0);
1590 }
1591
1592 ASSERT(key <= tsd_nkeys);
1593
1594 /*
1595 * lock out tsd_destroy()
1596 */
1597 mutex_enter(&tsd_mutex);
1598 if (tsd->ts_nkeys == 0) {
1599 /*
1600 * Link onto list of threads with TSD
1601 */
1602 if ((tsd->ts_next = tsd_list) != NULL)
1603 tsd_list->ts_prev = tsd;
1604 tsd_list = tsd;
1605 }
1606
1607 /*
1608 * Allocate thread local storage and set the value for key
1609 */
1610 tsd->ts_value = tsd_realloc(tsd->ts_value,
1611 tsd->ts_nkeys * sizeof (void *),
1612 key * sizeof (void *));
1613 tsd->ts_nkeys = key;
1614 tsd->ts_value[key - 1] = value;
1615 mutex_exit(&tsd_mutex);
1616
1617 return (0);
1618 }
1619
1620
1621 /*
1622 * Return the per thread value that was stored with the specified key
1623 * If necessary, create the key and the value
1624 * Assumes the caller is protecting *keyp from tsd_destroy
1625 */
1626 void *
1627 tsd_getcreate(uint_t *keyp, void (*destroy)(void *), void *(*allocate)(void))
1628 {
1629 void *value;
1630 uint_t key = *keyp;
1631 struct tsd_thread *tsd = curthread->t_tsd;
1632
1633 if (tsd == NULL)
1634 tsd = curthread->t_tsd = kmem_zalloc(sizeof (*tsd), KM_SLEEP);
1635 if (key && key <= tsd->ts_nkeys && (value = tsd->ts_value[key - 1]))
1636 return (value);
1637 if (key == 0)
1638 tsd_create(keyp, destroy);
1639 (void) tsd_set(*keyp, value = (*allocate)());
1640
1641 return (value);
1642 }
1643
1644 /*
1645 * Called from thread_exit() to run the destructor function for each tsd
1646 * Locks out tsd_create and tsd_destroy
1647 * Assumes that the destructor *DOES NOT* use tsd
1648 */
1649 void
1650 tsd_exit(void)
1651 {
1652 int i;
1653 struct tsd_thread *tsd = curthread->t_tsd;
1654
1655 if (tsd == NULL)
1656 return;
1657
1658 if (tsd->ts_nkeys == 0) {
1659 kmem_free(tsd, sizeof (*tsd));
1660 curthread->t_tsd = NULL;
1661 return;
1662 }
1663
1664 /*
1665 * lock out tsd_create and tsd_destroy, call
1666 * the destructor, and mark the value as destroyed.
1667 */
1668 mutex_enter(&tsd_mutex);
1669
1670 for (i = 0; i < tsd->ts_nkeys; i++) {
1671 if (tsd->ts_value[i] && tsd_destructor[i])
1672 (*tsd_destructor[i])(tsd->ts_value[i]);
1673 tsd->ts_value[i] = NULL;
1674 }
1675
1676 /*
1677 * remove from linked list of threads with TSD
1678 */
1679 if (tsd->ts_next)
1680 tsd->ts_next->ts_prev = tsd->ts_prev;
1681 if (tsd->ts_prev)
1682 tsd->ts_prev->ts_next = tsd->ts_next;
1683 if (tsd_list == tsd)
1684 tsd_list = tsd->ts_next;
1685
1686 mutex_exit(&tsd_mutex);
1687
1688 /*
1689 * free up the TSD
1690 */
1691 kmem_free(tsd->ts_value, tsd->ts_nkeys * sizeof (void *));
1692 kmem_free(tsd, sizeof (struct tsd_thread));
1693 curthread->t_tsd = NULL;
1694 }
1695
1696 /*
1697 * realloc
1698 */
1699 static void *
1700 tsd_realloc(void *old, size_t osize, size_t nsize)
1701 {
1702 void *new;
1703
1704 new = kmem_zalloc(nsize, KM_SLEEP);
1705 if (old) {
1706 bcopy(old, new, osize);
1707 kmem_free(old, osize);
1708 }
1709 return (new);
1710 }
1711
1712 /*
1713 * Return non-zero if an interrupt is being serviced.
1714 */
1715 int
1716 servicing_interrupt()
1717 {
1718 int onintr = 0;
1719
1720 /* Are we an interrupt thread */
1721 if (curthread->t_flag & T_INTR_THREAD)
1722 return (1);
1723 /* Are we servicing a high level interrupt? */
1724 if (CPU_ON_INTR(CPU)) {
1725 kpreempt_disable();
1726 onintr = CPU_ON_INTR(CPU);
1727 kpreempt_enable();
1728 }
1729 return (onintr);
1730 }
1731
1732
1733 /*
1734 * Change the dispatch priority of a thread in the system.
1735 * Used when raising or lowering a thread's priority.
1736 * (E.g., priority inheritance)
1737 *
1738 * Since threads are queued according to their priority, we
1739 * we must check the thread's state to determine whether it
1740 * is on a queue somewhere. If it is, we've got to:
1741 *
1742 * o Dequeue the thread.
1743 * o Change its effective priority.
1744 * o Enqueue the thread.
1745 *
1746 * Assumptions: The thread whose priority we wish to change
1747 * must be locked before we call thread_change_(e)pri().
1748 * The thread_change(e)pri() function doesn't drop the thread
1749 * lock--that must be done by its caller.
1750 */
1751 void
1752 thread_change_epri(kthread_t *t, pri_t disp_pri)
1753 {
1754 uint_t state;
1755
1756 ASSERT(THREAD_LOCK_HELD(t));
1757
1758 /*
1759 * If the inherited priority hasn't actually changed,
1760 * just return.
1761 */
1762 if (t->t_epri == disp_pri)
1763 return;
1764
1765 state = t->t_state;
1766
1767 /*
1768 * If it's not on a queue, change the priority with impunity.
1769 */
1770 if ((state & (TS_SLEEP | TS_RUN | TS_WAIT)) == 0) {
1771 t->t_epri = disp_pri;
1772 if (state == TS_ONPROC) {
1773 cpu_t *cp = t->t_disp_queue->disp_cpu;
1774
1775 if (t == cp->cpu_dispthread)
1776 cp->cpu_dispatch_pri = DISP_PRIO(t);
1777 }
1778 } else if (state == TS_SLEEP) {
1779 /*
1780 * Take the thread out of its sleep queue.
1781 * Change the inherited priority.
1782 * Re-enqueue the thread.
1783 * Each synchronization object exports a function
1784 * to do this in an appropriate manner.
1785 */
1786 SOBJ_CHANGE_EPRI(t->t_sobj_ops, t, disp_pri);
1787 } else if (state == TS_WAIT) {
1788 /*
1789 * Re-enqueue a thread on the wait queue if its
1790 * effective priority needs to change.
1791 */
1792 if (disp_pri != t->t_epri)
1793 waitq_change_pri(t, disp_pri);
1794 } else {
1795 /*
1796 * The thread is on a run queue.
1797 * Note: setbackdq() may not put the thread
1798 * back on the same run queue where it originally
1799 * resided.
1800 */
1801 (void) dispdeq(t);
1802 t->t_epri = disp_pri;
1803 setbackdq(t);
1804 }
1805 schedctl_set_cidpri(t);
1806 }
1807
1808 /*
1809 * Function: Change the t_pri field of a thread.
1810 * Side Effects: Adjust the thread ordering on a run queue
1811 * or sleep queue, if necessary.
1812 * Returns: 1 if the thread was on a run queue, else 0.
1813 */
1814 int
1815 thread_change_pri(kthread_t *t, pri_t disp_pri, int front)
1816 {
1817 uint_t state;
1818 int on_rq = 0;
1819
1820 ASSERT(THREAD_LOCK_HELD(t));
1821
1822 state = t->t_state;
1823 THREAD_WILLCHANGE_PRI(t, disp_pri);
1824
1825 /*
1826 * If it's not on a queue, change the priority with impunity.
1827 */
1828 if ((state & (TS_SLEEP | TS_RUN | TS_WAIT)) == 0) {
1829 t->t_pri = disp_pri;
1830
1831 if (state == TS_ONPROC) {
1832 cpu_t *cp = t->t_disp_queue->disp_cpu;
1833
1834 if (t == cp->cpu_dispthread)
1835 cp->cpu_dispatch_pri = DISP_PRIO(t);
1836 }
1837 } else if (state == TS_SLEEP) {
1838 /*
1839 * If the priority has changed, take the thread out of
1840 * its sleep queue and change the priority.
1841 * Re-enqueue the thread.
1842 * Each synchronization object exports a function
1843 * to do this in an appropriate manner.
1844 */
1845 if (disp_pri != t->t_pri)
1846 SOBJ_CHANGE_PRI(t->t_sobj_ops, t, disp_pri);
1847 } else if (state == TS_WAIT) {
1848 /*
1849 * Re-enqueue a thread on the wait queue if its
1850 * priority needs to change.
1851 */
1852 if (disp_pri != t->t_pri)
1853 waitq_change_pri(t, disp_pri);
1854 } else {
1855 /*
1856 * The thread is on a run queue.
1857 * Note: setbackdq() may not put the thread
1858 * back on the same run queue where it originally
1859 * resided.
1860 *
1861 * We still requeue the thread even if the priority
1862 * is unchanged to preserve round-robin (and other)
1863 * effects between threads of the same priority.
1864 */
1865 on_rq = dispdeq(t);
1866 ASSERT(on_rq);
1867 t->t_pri = disp_pri;
1868 if (front) {
1869 setfrontdq(t);
1870 } else {
1871 setbackdq(t);
1872 }
1873 }
1874 schedctl_set_cidpri(t);
1875 return (on_rq);
1876 }
1877
1878 /*
1879 * Tunable kmem_stackinfo is set, fill the kernel thread stack with a
1880 * specific pattern.
1881 */
1882 static void
1883 stkinfo_begin(kthread_t *t)
1884 {
1885 caddr_t start; /* stack start */
1886 caddr_t end; /* stack end */
1887 uint64_t *ptr; /* pattern pointer */
1888
1889 /*
1890 * Stack grows up or down, see thread_create(),
1891 * compute stack memory area start and end (start < end).
1892 */
1893 if (t->t_stk > t->t_stkbase) {
1894 /* stack grows down */
1895 start = t->t_stkbase;
1896 end = t->t_stk;
1897 } else {
1898 /* stack grows up */
1899 start = t->t_stk;
1900 end = t->t_stkbase;
1901 }
1902
1903 /*
1904 * Stackinfo pattern size is 8 bytes. Ensure proper 8 bytes
1905 * alignement for start and end in stack area boundaries
1906 * (protection against corrupt t_stkbase/t_stk data).
1907 */
1908 if ((((uintptr_t)start) & 0x7) != 0) {
1909 start = (caddr_t)((((uintptr_t)start) & (~0x7)) + 8);
1910 }
1911 end = (caddr_t)(((uintptr_t)end) & (~0x7));
1912
1913 if ((end <= start) || (end - start) > (1024 * 1024)) {
1914 /* negative or stack size > 1 meg, assume bogus */
1915 return;
1916 }
1917
1918 /* fill stack area with a pattern (instead of zeros) */
1919 ptr = (uint64_t *)((void *)start);
1920 while (ptr < (uint64_t *)((void *)end)) {
1921 *ptr++ = KMEM_STKINFO_PATTERN;
1922 }
1923 }
1924
1925
1926 /*
1927 * Tunable kmem_stackinfo is set, create stackinfo log if doesn't already exist,
1928 * compute the percentage of kernel stack really used, and set in the log
1929 * if it's the latest highest percentage.
1930 */
1931 static void
1932 stkinfo_end(kthread_t *t)
1933 {
1934 caddr_t start; /* stack start */
1935 caddr_t end; /* stack end */
1936 uint64_t *ptr; /* pattern pointer */
1937 size_t stksz; /* stack size */
1938 size_t smallest = 0;
1939 size_t percent = 0;
1940 uint_t index = 0;
1941 uint_t i;
1942 static size_t smallest_percent = (size_t)-1;
1943 static uint_t full = 0;
1944
1945 /* create the stackinfo log, if doesn't already exist */
1946 mutex_enter(&kmem_stkinfo_lock);
1947 if (kmem_stkinfo_log == NULL) {
1948 kmem_stkinfo_log = (kmem_stkinfo_t *)
1949 kmem_zalloc(KMEM_STKINFO_LOG_SIZE *
1950 (sizeof (kmem_stkinfo_t)), KM_NOSLEEP);
1951 if (kmem_stkinfo_log == NULL) {
1952 mutex_exit(&kmem_stkinfo_lock);
1953 return;
1954 }
1955 }
1956 mutex_exit(&kmem_stkinfo_lock);
1957
1958 /*
1959 * Stack grows up or down, see thread_create(),
1960 * compute stack memory area start and end (start < end).
1961 */
1962 if (t->t_stk > t->t_stkbase) {
1963 /* stack grows down */
1964 start = t->t_stkbase;
1965 end = t->t_stk;
1966 } else {
1967 /* stack grows up */
1968 start = t->t_stk;
1969 end = t->t_stkbase;
1970 }
1971
1972 /* stack size as found in kthread_t */
1973 stksz = end - start;
1974
1975 /*
1976 * Stackinfo pattern size is 8 bytes. Ensure proper 8 bytes
1977 * alignement for start and end in stack area boundaries
1978 * (protection against corrupt t_stkbase/t_stk data).
1979 */
1980 if ((((uintptr_t)start) & 0x7) != 0) {
1981 start = (caddr_t)((((uintptr_t)start) & (~0x7)) + 8);
1982 }
1983 end = (caddr_t)(((uintptr_t)end) & (~0x7));
1984
1985 if ((end <= start) || (end - start) > (1024 * 1024)) {
1986 /* negative or stack size > 1 meg, assume bogus */
1987 return;
1988 }
1989
1990 /* search until no pattern in the stack */
1991 if (t->t_stk > t->t_stkbase) {
1992 /* stack grows down */
1993 #if defined(__i386) || defined(__amd64)
1994 /*
1995 * 6 longs are pushed on stack, see thread_load(). Skip
1996 * them, so if kthread has never run, percent is zero.
1997 * 8 bytes alignement is preserved for a 32 bit kernel,
1998 * 6 x 4 = 24, 24 is a multiple of 8.
1999 *
2000 */
2001 end -= (6 * sizeof (long));
2002 #endif
2003 ptr = (uint64_t *)((void *)start);
2004 while (ptr < (uint64_t *)((void *)end)) {
2005 if (*ptr != KMEM_STKINFO_PATTERN) {
2006 percent = stkinfo_percent(end,
2007 start, (caddr_t)ptr);
2008 break;
2009 }
2010 ptr++;
2011 }
2012 } else {
2013 /* stack grows up */
2014 ptr = (uint64_t *)((void *)end);
2015 ptr--;
2016 while (ptr >= (uint64_t *)((void *)start)) {
2017 if (*ptr != KMEM_STKINFO_PATTERN) {
2018 percent = stkinfo_percent(start,
2019 end, (caddr_t)ptr);
2020 break;
2021 }
2022 ptr--;
2023 }
2024 }
2025
2026 DTRACE_PROBE3(stack__usage, kthread_t *, t,
2027 size_t, stksz, size_t, percent);
2028
2029 if (percent == 0) {
2030 return;
2031 }
2032
2033 mutex_enter(&kmem_stkinfo_lock);
2034 if (full == KMEM_STKINFO_LOG_SIZE && percent < smallest_percent) {
2035 /*
2036 * The log is full and already contains the highest values
2037 */
2038 mutex_exit(&kmem_stkinfo_lock);
2039 return;
2040 }
2041
2042 /* keep a log of the highest used stack */
2043 for (i = 0; i < KMEM_STKINFO_LOG_SIZE; i++) {
2044 if (kmem_stkinfo_log[i].percent == 0) {
2045 index = i;
2046 full++;
2047 break;
2048 }
2049 if (smallest == 0) {
2050 smallest = kmem_stkinfo_log[i].percent;
2051 index = i;
2052 continue;
2053 }
2054 if (kmem_stkinfo_log[i].percent < smallest) {
2055 smallest = kmem_stkinfo_log[i].percent;
2056 index = i;
2057 }
2058 }
2059
2060 if (percent >= kmem_stkinfo_log[index].percent) {
2061 kmem_stkinfo_log[index].kthread = (caddr_t)t;
2062 kmem_stkinfo_log[index].t_startpc = (caddr_t)t->t_startpc;
2063 kmem_stkinfo_log[index].start = start;
2064 kmem_stkinfo_log[index].stksz = stksz;
2065 kmem_stkinfo_log[index].percent = percent;
2066 kmem_stkinfo_log[index].t_tid = t->t_tid;
2067 kmem_stkinfo_log[index].cmd[0] = '\0';
2068 if (t->t_tid != 0) {
2069 stksz = strlen((t->t_procp)->p_user.u_comm);
2070 if (stksz >= KMEM_STKINFO_STR_SIZE) {
2071 stksz = KMEM_STKINFO_STR_SIZE - 1;
2072 kmem_stkinfo_log[index].cmd[stksz] = '\0';
2073 } else {
2074 stksz += 1;
2075 }
2076 (void) memcpy(kmem_stkinfo_log[index].cmd,
2077 (t->t_procp)->p_user.u_comm, stksz);
2078 }
2079 if (percent < smallest_percent) {
2080 smallest_percent = percent;
2081 }
2082 }
2083 mutex_exit(&kmem_stkinfo_lock);
2084 }
2085
2086 /*
2087 * Tunable kmem_stackinfo is set, compute stack utilization percentage.
2088 */
2089 static size_t
2090 stkinfo_percent(caddr_t t_stk, caddr_t t_stkbase, caddr_t sp)
2091 {
2092 size_t percent;
2093 size_t s;
2094
2095 if (t_stk > t_stkbase) {
2096 /* stack grows down */
2097 if (sp > t_stk) {
2098 return (0);
2099 }
2100 if (sp < t_stkbase) {
2101 return (100);
2102 }
2103 percent = t_stk - sp + 1;
2104 s = t_stk - t_stkbase + 1;
2105 } else {
2106 /* stack grows up */
2107 if (sp < t_stk) {
2108 return (0);
2109 }
2110 if (sp > t_stkbase) {
2111 return (100);
2112 }
2113 percent = sp - t_stk + 1;
2114 s = t_stkbase - t_stk + 1;
2115 }
2116 percent = ((100 * percent) / s) + 1;
2117 if (percent > 100) {
2118 percent = 100;
2119 }
2120 return (percent);
2121 }