xref: /illumos-gate/usr/src/uts/common/disp/thread.c (revision 86ef0a63)
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  * Copyright 2021 Joyent, Inc.
25  */
26 
27 #include <sys/types.h>
28 #include <sys/param.h>
29 #include <sys/sysmacros.h>
30 #include <sys/signal.h>
31 #include <sys/stack.h>
32 #include <sys/pcb.h>
33 #include <sys/user.h>
34 #include <sys/systm.h>
35 #include <sys/sysinfo.h>
36 #include <sys/errno.h>
37 #include <sys/cmn_err.h>
38 #include <sys/cred.h>
39 #include <sys/resource.h>
40 #include <sys/task.h>
41 #include <sys/project.h>
42 #include <sys/proc.h>
43 #include <sys/debug.h>
44 #include <sys/disp.h>
45 #include <sys/class.h>
46 #include <vm/seg_kmem.h>
47 #include <vm/seg_kp.h>
48 #include <sys/machlock.h>
49 #include <sys/kmem.h>
50 #include <sys/varargs.h>
51 #include <sys/turnstile.h>
52 #include <sys/poll.h>
53 #include <sys/vtrace.h>
54 #include <sys/callb.h>
55 #include <c2/audit.h>
56 #include <sys/tnf.h>
57 #include <sys/sobject.h>
58 #include <sys/cpupart.h>
59 #include <sys/pset.h>
60 #include <sys/door.h>
61 #include <sys/spl.h>
62 #include <sys/copyops.h>
63 #include <sys/rctl.h>
64 #include <sys/brand.h>
65 #include <sys/pool.h>
66 #include <sys/zone.h>
67 #include <sys/tsol/label.h>
68 #include <sys/tsol/tndb.h>
69 #include <sys/cpc_impl.h>
70 #include <sys/sdt.h>
71 #include <sys/reboot.h>
72 #include <sys/kdi.h>
73 #include <sys/schedctl.h>
74 #include <sys/waitq.h>
75 #include <sys/cpucaps.h>
76 #include <sys/kiconv.h>
77 #include <sys/ctype.h>
78 #include <sys/smt.h>
79 
80 struct kmem_cache *thread_cache;	/* cache of free threads */
81 struct kmem_cache *lwp_cache;		/* cache of free lwps */
82 struct kmem_cache *turnstile_cache;	/* cache of free turnstiles */
83 
84 /*
85  * allthreads is only for use by kmem_readers.  All kernel loops can use
86  * the current thread as a start/end point.
87  */
88 kthread_t *allthreads = &t0;	/* circular list of all threads */
89 
90 static kcondvar_t reaper_cv;		/* synchronization var */
91 kthread_t	*thread_deathrow;	/* circular list of reapable threads */
92 kthread_t	*lwp_deathrow;		/* circular list of reapable threads */
93 kmutex_t	reaplock;		/* protects lwp and thread deathrows */
94 int	thread_reapcnt = 0;		/* number of threads on deathrow */
95 int	lwp_reapcnt = 0;		/* number of lwps on deathrow */
96 int	reaplimit = 16;			/* delay reaping until reaplimit */
97 
98 thread_free_lock_t	*thread_free_lock;
99 					/* protects tick thread from reaper */
100 
101 extern int nthread;
102 
103 /* System Scheduling classes. */
104 id_t	syscid;				/* system scheduling class ID */
105 id_t	sysdccid = CLASS_UNUSED;	/* reset when SDC loads */
106 
107 void	*segkp_thread;			/* cookie for segkp pool */
108 
109 int lwp_cache_sz = 32;
110 int t_cache_sz = 8;
111 static kt_did_t next_t_id = 1;
112 
113 /* Default mode for thread binding to CPUs and processor sets */
114 int default_binding_mode = TB_ALLHARD;
115 
116 /*
117  * Min/Max stack sizes for stack size parameters
118  */
119 #define	MAX_STKSIZE	(32 * DEFAULTSTKSZ)
120 #define	MIN_STKSIZE	DEFAULTSTKSZ
121 
122 /*
123  * default_stksize overrides lwp_default_stksize if it is set.
124  */
125 int	default_stksize;
126 int	lwp_default_stksize;
127 
128 static zone_key_t zone_thread_key;
129 
130 unsigned int kmem_stackinfo;		/* stackinfo feature on-off */
131 kmem_stkinfo_t *kmem_stkinfo_log;	/* stackinfo circular log */
132 static kmutex_t kmem_stkinfo_lock;	/* protects kmem_stkinfo_log */
133 
134 /*
135  * forward declarations for internal thread specific data (tsd)
136  */
137 static void *tsd_realloc(void *, size_t, size_t);
138 
139 void thread_reaper(void);
140 
141 /* forward declarations for stackinfo feature */
142 static void stkinfo_begin(kthread_t *);
143 static void stkinfo_end(kthread_t *);
144 static size_t stkinfo_percent(caddr_t, caddr_t, caddr_t);
145 
146 /*ARGSUSED*/
147 static int
148 turnstile_constructor(void *buf, void *cdrarg, int kmflags)
149 {
150 	bzero(buf, sizeof (turnstile_t));
151 	return (0);
152 }
153 
154 /*ARGSUSED*/
155 static void
156 turnstile_destructor(void *buf, void *cdrarg)
157 {
158 	turnstile_t *ts = buf;
159 
160 	ASSERT(ts->ts_free == NULL);
161 	ASSERT(ts->ts_waiters == 0);
162 	ASSERT(ts->ts_inheritor == NULL);
163 	ASSERT(ts->ts_sleepq[0].sq_first == NULL);
164 	ASSERT(ts->ts_sleepq[1].sq_first == NULL);
165 }
166 
167 void
168 thread_init(void)
169 {
170 	kthread_t *tp;
171 	extern char sys_name[];
172 	extern void idle();
173 	struct cpu *cpu = CPU;
174 	int i;
175 	kmutex_t *lp;
176 
177 	mutex_init(&reaplock, NULL, MUTEX_SPIN, (void *)ipltospl(DISP_LEVEL));
178 	thread_free_lock =
179 	    kmem_alloc(sizeof (thread_free_lock_t) * THREAD_FREE_NUM, KM_SLEEP);
180 	for (i = 0; i < THREAD_FREE_NUM; i++) {
181 		lp = &thread_free_lock[i].tf_lock;
182 		mutex_init(lp, NULL, MUTEX_DEFAULT, NULL);
183 	}
184 
185 #if defined(__x86)
186 	thread_cache = kmem_cache_create("thread_cache", sizeof (kthread_t),
187 	    PTR24_ALIGN, NULL, NULL, NULL, NULL, NULL, 0);
188 
189 	/*
190 	 * "struct _klwp" includes a "struct pcb", which includes a
191 	 * "struct fpu", which needs to be 64-byte aligned on amd64
192 	 * (and even on i386) for xsave/xrstor.
193 	 */
194 	lwp_cache = kmem_cache_create("lwp_cache", sizeof (klwp_t),
195 	    64, NULL, NULL, NULL, NULL, NULL, 0);
196 #else
197 	/*
198 	 * Allocate thread structures from static_arena.  This prevents
199 	 * issues where a thread tries to relocate its own thread
200 	 * structure and touches it after the mapping has been suspended.
201 	 */
202 	thread_cache = kmem_cache_create("thread_cache", sizeof (kthread_t),
203 	    PTR24_ALIGN, NULL, NULL, NULL, NULL, static_arena, 0);
204 
205 	lwp_stk_cache_init();
206 
207 	lwp_cache = kmem_cache_create("lwp_cache", sizeof (klwp_t),
208 	    0, NULL, NULL, NULL, NULL, NULL, 0);
209 #endif
210 
211 	turnstile_cache = kmem_cache_create("turnstile_cache",
212 	    sizeof (turnstile_t), 0,
213 	    turnstile_constructor, turnstile_destructor, NULL, NULL, NULL, 0);
214 
215 	label_init();
216 	cred_init();
217 
218 	/*
219 	 * Initialize various resource management facilities.
220 	 */
221 	rctl_init();
222 	cpucaps_init();
223 	/*
224 	 * Zone_init() should be called before project_init() so that project ID
225 	 * for the first project is initialized correctly.
226 	 */
227 	zone_init();
228 	project_init();
229 	brand_init();
230 	kiconv_init();
231 	task_init();
232 	tcache_init();
233 	pool_init();
234 
235 	curthread->t_ts = kmem_cache_alloc(turnstile_cache, KM_SLEEP);
236 
237 	/*
238 	 * Originally, we had two parameters to set default stack
239 	 * size: one for lwp's (lwp_default_stksize), and one for
240 	 * kernel-only threads (DEFAULTSTKSZ, a.k.a. _defaultstksz).
241 	 * Now we have a third parameter that overrides both if it is
242 	 * set to a legal stack size, called default_stksize.
243 	 */
244 
245 	if (default_stksize == 0) {
246 		default_stksize = DEFAULTSTKSZ;
247 	} else if (default_stksize % PAGESIZE != 0 ||
248 	    default_stksize > MAX_STKSIZE ||
249 	    default_stksize < MIN_STKSIZE) {
250 		cmn_err(CE_WARN, "Illegal stack size. Using %d",
251 		    (int)DEFAULTSTKSZ);
252 		default_stksize = DEFAULTSTKSZ;
253 	} else {
254 		lwp_default_stksize = default_stksize;
255 	}
256 
257 	if (lwp_default_stksize == 0) {
258 		lwp_default_stksize = default_stksize;
259 	} else if (lwp_default_stksize % PAGESIZE != 0 ||
260 	    lwp_default_stksize > MAX_STKSIZE ||
261 	    lwp_default_stksize < MIN_STKSIZE) {
262 		cmn_err(CE_WARN, "Illegal stack size. Using %d",
263 		    default_stksize);
264 		lwp_default_stksize = default_stksize;
265 	}
266 
267 	segkp_lwp = segkp_cache_init(segkp, lwp_cache_sz,
268 	    lwp_default_stksize,
269 	    (KPD_NOWAIT | KPD_HASREDZONE | KPD_LOCKED));
270 
271 	segkp_thread = segkp_cache_init(segkp, t_cache_sz,
272 	    default_stksize, KPD_HASREDZONE | KPD_LOCKED | KPD_NO_ANON);
273 
274 	(void) getcid(sys_name, &syscid);
275 	curthread->t_cid = syscid;	/* current thread is t0 */
276 
277 	/*
278 	 * Set up the first CPU's idle thread.
279 	 * It runs whenever the CPU has nothing worthwhile to do.
280 	 */
281 	tp = thread_create(NULL, 0, idle, NULL, 0, &p0, TS_STOPPED, -1);
282 	cpu->cpu_idle_thread = tp;
283 	tp->t_preempt = 1;
284 	tp->t_disp_queue = cpu->cpu_disp;
285 	ASSERT(tp->t_disp_queue != NULL);
286 	tp->t_bound_cpu = cpu;
287 	tp->t_affinitycnt = 1;
288 
289 	/*
290 	 * Registering a thread in the callback table is usually
291 	 * done in the initialization code of the thread. In this
292 	 * case, we do it right after thread creation to avoid
293 	 * blocking idle thread while registering itself. It also
294 	 * avoids the possibility of reregistration in case a CPU
295 	 * restarts its idle thread.
296 	 */
297 	CALLB_CPR_INIT_SAFE(tp, "idle");
298 
299 	/*
300 	 * Create the thread_reaper daemon. From this point on, exited
301 	 * threads will get reaped.
302 	 */
303 	(void) thread_create(NULL, 0, (void (*)())thread_reaper,
304 	    NULL, 0, &p0, TS_RUN, minclsyspri);
305 
306 	/*
307 	 * Finish initializing the kernel memory allocator now that
308 	 * thread_create() is available.
309 	 */
310 	kmem_thread_init();
311 
312 	if (boothowto & RB_DEBUG)
313 		kdi_dvec_thravail();
314 }
315 
316 /*
317  * Create a thread.
318  *
319  * thread_create() blocks for memory if necessary.  It never fails.
320  *
321  * If stk is NULL, the thread is created at the base of the stack
322  * and cannot be swapped.
323  */
324 kthread_t *
325 thread_create(
326 	caddr_t	stk,
327 	size_t	stksize,
328 	void	(*proc)(),
329 	void	*arg,
330 	size_t	len,
331 	proc_t	 *pp,
332 	int	state,
333 	pri_t	pri)
334 {
335 	kthread_t *t;
336 	extern struct classfuncs sys_classfuncs;
337 	turnstile_t *ts;
338 
339 	/*
340 	 * Every thread keeps a turnstile around in case it needs to block.
341 	 * The only reason the turnstile is not simply part of the thread
342 	 * structure is that we may have to break the association whenever
343 	 * more than one thread blocks on a given synchronization object.
344 	 * From a memory-management standpoint, turnstiles are like the
345 	 * "attached mblks" that hang off dblks in the streams allocator.
346 	 */
347 	ts = kmem_cache_alloc(turnstile_cache, KM_SLEEP);
348 
349 	if (stk == NULL) {
350 		/*
351 		 * alloc both thread and stack in segkp chunk
352 		 */
353 
354 		if (stksize < default_stksize)
355 			stksize = default_stksize;
356 
357 		if (stksize == default_stksize) {
358 			stk = (caddr_t)segkp_cache_get(segkp_thread);
359 		} else {
360 			stksize = roundup(stksize, PAGESIZE);
361 			stk = (caddr_t)segkp_get(segkp, stksize,
362 			    (KPD_HASREDZONE | KPD_NO_ANON | KPD_LOCKED));
363 		}
364 
365 		ASSERT(stk != NULL);
366 
367 		/*
368 		 * The machine-dependent mutex code may require that
369 		 * thread pointers (since they may be used for mutex owner
370 		 * fields) have certain alignment requirements.
371 		 * PTR24_ALIGN is the size of the alignment quanta.
372 		 * XXX - assumes stack grows toward low addresses.
373 		 */
374 		if (stksize <= sizeof (kthread_t) + PTR24_ALIGN)
375 			cmn_err(CE_PANIC, "thread_create: proposed stack size"
376 			    " too small to hold thread.");
377 #ifdef STACK_GROWTH_DOWN
378 		stksize -= SA(sizeof (kthread_t) + PTR24_ALIGN - 1);
379 		stksize &= -PTR24_ALIGN;	/* make thread aligned */
380 		t = (kthread_t *)(stk + stksize);
381 		bzero(t, sizeof (kthread_t));
382 		if (audit_active)
383 			audit_thread_create(t);
384 		t->t_stk = stk + stksize;
385 		t->t_stkbase = stk;
386 #else	/* stack grows to larger addresses */
387 		stksize -= SA(sizeof (kthread_t));
388 		t = (kthread_t *)(stk);
389 		bzero(t, sizeof (kthread_t));
390 		t->t_stk = stk + sizeof (kthread_t);
391 		t->t_stkbase = stk + stksize + sizeof (kthread_t);
392 #endif	/* STACK_GROWTH_DOWN */
393 		t->t_flag |= T_TALLOCSTK;
394 		t->t_swap = stk;
395 	} else {
396 		t = kmem_cache_alloc(thread_cache, KM_SLEEP);
397 		bzero(t, sizeof (kthread_t));
398 		ASSERT(((uintptr_t)t & (PTR24_ALIGN - 1)) == 0);
399 		if (audit_active)
400 			audit_thread_create(t);
401 		/*
402 		 * Initialize t_stk to the kernel stack pointer to use
403 		 * upon entry to the kernel
404 		 */
405 #ifdef STACK_GROWTH_DOWN
406 		t->t_stk = stk + stksize;
407 		t->t_stkbase = stk;
408 #else
409 		t->t_stk = stk;			/* 3b2-like */
410 		t->t_stkbase = stk + stksize;
411 #endif /* STACK_GROWTH_DOWN */
412 	}
413 
414 	if (kmem_stackinfo != 0) {
415 		stkinfo_begin(t);
416 	}
417 
418 	t->t_ts = ts;
419 
420 	/*
421 	 * p_cred could be NULL if it thread_create is called before cred_init
422 	 * is called in main.
423 	 */
424 	mutex_enter(&pp->p_crlock);
425 	if (pp->p_cred)
426 		crhold(t->t_cred = pp->p_cred);
427 	mutex_exit(&pp->p_crlock);
428 	t->t_start = gethrestime_sec();
429 	t->t_startpc = proc;
430 	t->t_procp = pp;
431 	t->t_clfuncs = &sys_classfuncs.thread;
432 	t->t_cid = syscid;
433 	t->t_pri = pri;
434 	t->t_stime = ddi_get_lbolt();
435 	t->t_schedflag = TS_LOAD | TS_DONT_SWAP;
436 	t->t_bind_cpu = PBIND_NONE;
437 	t->t_bindflag = (uchar_t)default_binding_mode;
438 	t->t_bind_pset = PS_NONE;
439 	t->t_plockp = &pp->p_lock;
440 	t->t_copyops = NULL;
441 	t->t_taskq = NULL;
442 	t->t_anttime = 0;
443 	t->t_hatdepth = 0;
444 
445 	t->t_dtrace_vtime = 1;	/* assure vtimestamp is always non-zero */
446 
447 	CPU_STATS_ADDQ(CPU, sys, nthreads, 1);
448 #ifndef NPROBE
449 	/* Kernel probe */
450 	tnf_thread_create(t);
451 #endif /* NPROBE */
452 	LOCK_INIT_CLEAR(&t->t_lock);
453 
454 	/*
455 	 * Callers who give us a NULL proc must do their own
456 	 * stack initialization.  e.g. lwp_create()
457 	 */
458 	if (proc != NULL) {
459 		t->t_stk = thread_stk_init(t->t_stk);
460 		thread_load(t, proc, arg, len);
461 	}
462 
463 	/*
464 	 * Put a hold on project0. If this thread is actually in a
465 	 * different project, then t_proj will be changed later in
466 	 * lwp_create().  All kernel-only threads must be in project 0.
467 	 */
468 	t->t_proj = project_hold(proj0p);
469 
470 	lgrp_affinity_init(&t->t_lgrp_affinity);
471 
472 	mutex_enter(&pidlock);
473 	nthread++;
474 	t->t_did = next_t_id++;
475 	t->t_prev = curthread->t_prev;
476 	t->t_next = curthread;
477 
478 	/*
479 	 * Add the thread to the list of all threads, and initialize
480 	 * its t_cpu pointer.  We need to block preemption since
481 	 * cpu_offline walks the thread list looking for threads
482 	 * with t_cpu pointing to the CPU being offlined.  We want
483 	 * to make sure that the list is consistent and that if t_cpu
484 	 * is set, the thread is on the list.
485 	 */
486 	kpreempt_disable();
487 	curthread->t_prev->t_next = t;
488 	curthread->t_prev = t;
489 
490 	/*
491 	 * We'll always create in the default partition since that's where
492 	 * kernel threads go (we'll change this later if needed, in
493 	 * lwp_create()).
494 	 */
495 	t->t_cpupart = &cp_default;
496 
497 	/*
498 	 * For now, affiliate this thread with the root lgroup.
499 	 * Since the kernel does not (presently) allocate its memory
500 	 * in a locality aware fashion, the root is an appropriate home.
501 	 * If this thread is later associated with an lwp, it will have
502 	 * its lgroup re-assigned at that time.
503 	 */
504 	lgrp_move_thread(t, &cp_default.cp_lgrploads[LGRP_ROOTID], 1);
505 
506 	/*
507 	 * If the current CPU is in the default cpupart, use it.  Otherwise,
508 	 * pick one that is; before entering the dispatcher code, we'll
509 	 * make sure to keep the invariant that ->t_cpu is set.  (In fact, we
510 	 * rely on this, in ht_should_run(), in the call tree of
511 	 * disp_lowpri_cpu().)
512 	 */
513 	if (CPU->cpu_part == &cp_default) {
514 		t->t_cpu = CPU;
515 	} else {
516 		t->t_cpu = cp_default.cp_cpulist;
517 		t->t_cpu = disp_lowpri_cpu(t->t_cpu, t, t->t_pri);
518 	}
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 	if (t->t_name != NULL) {
794 		kmem_free(t->t_name, THREAD_NAME_MAX);
795 		t->t_name = NULL;
796 	}
797 
798 	/*
799 	 * Free thread, lwp and stack.  This needs to be done carefully, since
800 	 * if T_TALLOCSTK is set, the thread is part of the stack.
801 	 */
802 	t->t_lwp = NULL;
803 	t->t_swap = NULL;
804 
805 	if (swap) {
806 		segkp_release(segkp, swap);
807 	}
808 	if (lwp) {
809 		kmem_cache_free(lwp_cache, lwp);
810 	}
811 	if (!allocstk) {
812 		kmem_cache_free(thread_cache, t);
813 	}
814 }
815 
816 /*
817  * Removes threads associated with the given zone from a deathrow queue.
818  * tp is a pointer to the head of the deathrow queue, and countp is a
819  * pointer to the current deathrow count.  Returns a linked list of
820  * threads removed from the list.
821  */
822 static kthread_t *
823 thread_zone_cleanup(kthread_t **tp, int *countp, zoneid_t zoneid)
824 {
825 	kthread_t *tmp, *list = NULL;
826 	cred_t *cr;
827 
828 	ASSERT(MUTEX_HELD(&reaplock));
829 	while (*tp != NULL) {
830 		if ((cr = (*tp)->t_cred) != NULL && crgetzoneid(cr) == zoneid) {
831 			tmp = *tp;
832 			*tp = tmp->t_forw;
833 			tmp->t_forw = list;
834 			list = tmp;
835 			(*countp)--;
836 		} else {
837 			tp = &(*tp)->t_forw;
838 		}
839 	}
840 	return (list);
841 }
842 
843 static void
844 thread_reap_list(kthread_t *t)
845 {
846 	kthread_t *next;
847 
848 	while (t != NULL) {
849 		next = t->t_forw;
850 		thread_free(t);
851 		t = next;
852 	}
853 }
854 
855 /* ARGSUSED */
856 static void
857 thread_zone_destroy(zoneid_t zoneid, void *unused)
858 {
859 	kthread_t *t, *l;
860 
861 	mutex_enter(&reaplock);
862 	/*
863 	 * Pull threads and lwps associated with zone off deathrow lists.
864 	 */
865 	t = thread_zone_cleanup(&thread_deathrow, &thread_reapcnt, zoneid);
866 	l = thread_zone_cleanup(&lwp_deathrow, &lwp_reapcnt, zoneid);
867 	mutex_exit(&reaplock);
868 
869 	/*
870 	 * Guard against race condition in mutex_owner_running:
871 	 *	thread=owner(mutex)
872 	 *	<interrupt>
873 	 *				thread exits mutex
874 	 *				thread exits
875 	 *				thread reaped
876 	 *				thread struct freed
877 	 * cpu = thread->t_cpu <- BAD POINTER DEREFERENCE.
878 	 * A cross call to all cpus will cause the interrupt handler
879 	 * to reset the PC if it is in mutex_owner_running, refreshing
880 	 * stale thread pointers.
881 	 */
882 	mutex_sync();   /* sync with mutex code */
883 
884 	/*
885 	 * Reap threads
886 	 */
887 	thread_reap_list(t);
888 
889 	/*
890 	 * Reap lwps
891 	 */
892 	thread_reap_list(l);
893 }
894 
895 /*
896  * cleanup zombie threads that are on deathrow.
897  */
898 void
899 thread_reaper()
900 {
901 	kthread_t *t, *l;
902 	callb_cpr_t cprinfo;
903 
904 	/*
905 	 * Register callback to clean up threads when zone is destroyed.
906 	 */
907 	zone_key_create(&zone_thread_key, NULL, NULL, thread_zone_destroy);
908 
909 	CALLB_CPR_INIT(&cprinfo, &reaplock, callb_generic_cpr, "t_reaper");
910 	for (;;) {
911 		mutex_enter(&reaplock);
912 		while (thread_deathrow == NULL && lwp_deathrow == NULL) {
913 			CALLB_CPR_SAFE_BEGIN(&cprinfo);
914 			cv_wait(&reaper_cv, &reaplock);
915 			CALLB_CPR_SAFE_END(&cprinfo, &reaplock);
916 		}
917 		/*
918 		 * mutex_sync() needs to be called when reaping, but
919 		 * not too often.  We limit reaping rate to once
920 		 * per second.  Reaplimit is max rate at which threads can
921 		 * be freed. Does not impact thread destruction/creation.
922 		 */
923 		t = thread_deathrow;
924 		l = lwp_deathrow;
925 		thread_deathrow = NULL;
926 		lwp_deathrow = NULL;
927 		thread_reapcnt = 0;
928 		lwp_reapcnt = 0;
929 		mutex_exit(&reaplock);
930 
931 		/*
932 		 * Guard against race condition in mutex_owner_running:
933 		 *	thread=owner(mutex)
934 		 *	<interrupt>
935 		 *				thread exits mutex
936 		 *				thread exits
937 		 *				thread reaped
938 		 *				thread struct freed
939 		 * cpu = thread->t_cpu <- BAD POINTER DEREFERENCE.
940 		 * A cross call to all cpus will cause the interrupt handler
941 		 * to reset the PC if it is in mutex_owner_running, refreshing
942 		 * stale thread pointers.
943 		 */
944 		mutex_sync();   /* sync with mutex code */
945 		/*
946 		 * Reap threads
947 		 */
948 		thread_reap_list(t);
949 
950 		/*
951 		 * Reap lwps
952 		 */
953 		thread_reap_list(l);
954 		delay(hz);
955 	}
956 }
957 
958 /*
959  * This is called by lwpcreate, etc.() to put a lwp_deathrow thread onto
960  * thread_deathrow. The thread's state is changed already TS_FREE to indicate
961  * that is reapable. The thread already holds the reaplock, and was already
962  * freed.
963  */
964 void
965 reapq_move_lq_to_tq(kthread_t *t)
966 {
967 	ASSERT(t->t_state == TS_FREE);
968 	ASSERT(MUTEX_HELD(&reaplock));
969 	t->t_forw = thread_deathrow;
970 	thread_deathrow = t;
971 	thread_reapcnt++;
972 	if (lwp_reapcnt + thread_reapcnt > reaplimit)
973 		cv_signal(&reaper_cv);  /* wake the reaper */
974 }
975 
976 /*
977  * This is called by resume() to put a zombie thread onto deathrow.
978  * The thread's state is changed to TS_FREE to indicate that is reapable.
979  * This is called from the idle thread so it must not block - just spin.
980  */
981 void
982 reapq_add(kthread_t *t)
983 {
984 	mutex_enter(&reaplock);
985 
986 	/*
987 	 * lwp_deathrow contains threads with lwp linkage and
988 	 * swappable thread stacks which have the default stacksize.
989 	 * These threads' lwps and stacks may be reused by lwp_create().
990 	 *
991 	 * Anything else goes on thread_deathrow(), where it will eventually
992 	 * be thread_free()d.
993 	 */
994 	if (t->t_flag & T_LWPREUSE) {
995 		ASSERT(ttolwp(t) != NULL);
996 		t->t_forw = lwp_deathrow;
997 		lwp_deathrow = t;
998 		lwp_reapcnt++;
999 	} else {
1000 		t->t_forw = thread_deathrow;
1001 		thread_deathrow = t;
1002 		thread_reapcnt++;
1003 	}
1004 	if (lwp_reapcnt + thread_reapcnt > reaplimit)
1005 		cv_signal(&reaper_cv);	/* wake the reaper */
1006 	t->t_state = TS_FREE;
1007 	lock_clear(&t->t_lock);
1008 
1009 	/*
1010 	 * Before we return, we need to grab and drop the thread lock for
1011 	 * the dead thread.  At this point, the current thread is the idle
1012 	 * thread, and the dead thread's CPU lock points to the current
1013 	 * CPU -- and we must grab and drop the lock to synchronize with
1014 	 * a racing thread walking a blocking chain that the zombie thread
1015 	 * was recently in.  By this point, that blocking chain is (by
1016 	 * definition) stale:  the dead thread is not holding any locks, and
1017 	 * is therefore not in any blocking chains -- but if we do not regrab
1018 	 * our lock before freeing the dead thread's data structures, the
1019 	 * thread walking the (stale) blocking chain will die on memory
1020 	 * corruption when it attempts to drop the dead thread's lock.  We
1021 	 * only need do this once because there is no way for the dead thread
1022 	 * to ever again be on a blocking chain:  once we have grabbed and
1023 	 * dropped the thread lock, we are guaranteed that anyone that could
1024 	 * have seen this thread in a blocking chain can no longer see it.
1025 	 */
1026 	thread_lock(t);
1027 	thread_unlock(t);
1028 
1029 	mutex_exit(&reaplock);
1030 }
1031 
1032 /*
1033  * Provide an allocation function for callers of installctx() that, for
1034  * reasons of incomplete context-op initialization, must call installctx()
1035  * in a kpreempt_disable() block.  The caller, therefore, must call this
1036  * without being in such a block.
1037  */
1038 struct ctxop *
1039 installctx_preallocate(void)
1040 {
1041 	/*
1042 	 * NOTE: We could ASSERT/VERIFY that we are not in a place where
1043 	 * a KM_SLEEP allocation could block indefinitely.
1044 	 *
1045 	 * ASSERT(curthread->t_preempt == 0);
1046 	 */
1047 
1048 	return (kmem_alloc(sizeof (struct ctxop), KM_SLEEP));
1049 }
1050 
1051 /*
1052  * Install thread context ops for the current thread.
1053  * The caller can pass in a preallocated struct ctxop, eliminating the need
1054  * for the requirement of entering with kernel preemption still enabled.
1055  */
1056 void
1057 installctx(
1058 	kthread_t *t,
1059 	void	*arg,
1060 	void	(*save)(void *),
1061 	void	(*restore)(void *),
1062 	void	(*fork)(void *, void *),
1063 	void	(*lwp_create)(void *, void *),
1064 	void	(*exit)(void *),
1065 	void	(*free)(void *, int),
1066 	struct ctxop *ctx)
1067 {
1068 	if (ctx == NULL)
1069 		ctx = kmem_alloc(sizeof (struct ctxop), KM_SLEEP);
1070 
1071 	ctx->save_op = save;
1072 	ctx->restore_op = restore;
1073 	ctx->fork_op = fork;
1074 	ctx->lwp_create_op = lwp_create;
1075 	ctx->exit_op = exit;
1076 	ctx->free_op = free;
1077 	ctx->arg = arg;
1078 	ctx->save_ts = 0;
1079 	ctx->restore_ts = 0;
1080 
1081 	/*
1082 	 * Keep ctxops in a doubly-linked list to allow traversal in both
1083 	 * directions.  Using only the newest-to-oldest ordering was adequate
1084 	 * previously, but reversing the order for restore_op actions is
1085 	 * necessary if later-added ctxops depends on earlier ones.
1086 	 *
1087 	 * One example of such a dependency:  Hypervisor software handling the
1088 	 * guest FPU expects that it save FPU state prior to host FPU handling
1089 	 * and consequently handle the guest logic _after_ the host FPU has
1090 	 * been restored.
1091 	 *
1092 	 * The t_ctx member points to the most recently added ctxop or is NULL
1093 	 * if no ctxops are associated with the thread.  The 'next' pointers
1094 	 * form a loop of the ctxops in newest-to-oldest order.  The 'prev'
1095 	 * pointers form a loop in the reverse direction, where t_ctx->prev is
1096 	 * the oldest entry associated with the thread.
1097 	 *
1098 	 * The protection of kpreempt_disable is required to safely perform the
1099 	 * list insertion, since there are inconsistent states between some of
1100 	 * the pointer assignments.
1101 	 */
1102 	kpreempt_disable();
1103 	if (t->t_ctx == NULL) {
1104 		ctx->next = ctx;
1105 		ctx->prev = ctx;
1106 	} else {
1107 		struct ctxop *head = t->t_ctx, *tail = t->t_ctx->prev;
1108 
1109 		ctx->next = head;
1110 		ctx->prev = tail;
1111 		head->prev = ctx;
1112 		tail->next = ctx;
1113 	}
1114 	t->t_ctx = ctx;
1115 	kpreempt_enable();
1116 }
1117 
1118 /*
1119  * Remove the thread context ops from a thread.
1120  */
1121 int
1122 removectx(
1123 	kthread_t *t,
1124 	void	*arg,
1125 	void	(*save)(void *),
1126 	void	(*restore)(void *),
1127 	void	(*fork)(void *, void *),
1128 	void	(*lwp_create)(void *, void *),
1129 	void	(*exit)(void *),
1130 	void	(*free)(void *, int))
1131 {
1132 	struct ctxop *ctx, *head;
1133 
1134 	/*
1135 	 * The incoming kthread_t (which is the thread for which the
1136 	 * context ops will be removed) should be one of the following:
1137 	 *
1138 	 * a) the current thread,
1139 	 *
1140 	 * b) a thread of a process that's being forked (SIDL),
1141 	 *
1142 	 * c) a thread that belongs to the same process as the current
1143 	 *    thread and for which the current thread is the agent thread,
1144 	 *
1145 	 * d) a thread that is TS_STOPPED which is indicative of it
1146 	 *    being (if curthread is not an agent) a thread being created
1147 	 *    as part of an lwp creation.
1148 	 */
1149 	ASSERT(t == curthread || ttoproc(t)->p_stat == SIDL ||
1150 	    ttoproc(t)->p_agenttp == curthread || t->t_state == TS_STOPPED);
1151 
1152 	/*
1153 	 * Serialize modifications to t->t_ctx to prevent the agent thread
1154 	 * and the target thread from racing with each other during lwp exit.
1155 	 */
1156 	mutex_enter(&t->t_ctx_lock);
1157 	kpreempt_disable();
1158 
1159 	if (t->t_ctx == NULL) {
1160 		mutex_exit(&t->t_ctx_lock);
1161 		kpreempt_enable();
1162 		return (0);
1163 	}
1164 
1165 	ctx = head = t->t_ctx;
1166 	do {
1167 		if (ctx->save_op == save && ctx->restore_op == restore &&
1168 		    ctx->fork_op == fork && ctx->lwp_create_op == lwp_create &&
1169 		    ctx->exit_op == exit && ctx->free_op == free &&
1170 		    ctx->arg == arg) {
1171 			ctx->prev->next = ctx->next;
1172 			ctx->next->prev = ctx->prev;
1173 			if (ctx->next == ctx) {
1174 				/* last remaining item */
1175 				t->t_ctx = NULL;
1176 			} else if (ctx == t->t_ctx) {
1177 				/* fix up head of list */
1178 				t->t_ctx = ctx->next;
1179 			}
1180 			ctx->next = ctx->prev = NULL;
1181 
1182 			mutex_exit(&t->t_ctx_lock);
1183 			if (ctx->free_op != NULL)
1184 				(ctx->free_op)(ctx->arg, 0);
1185 			kmem_free(ctx, sizeof (struct ctxop));
1186 			kpreempt_enable();
1187 			return (1);
1188 		}
1189 
1190 		ctx = ctx->next;
1191 	} while (ctx != head);
1192 
1193 	mutex_exit(&t->t_ctx_lock);
1194 	kpreempt_enable();
1195 	return (0);
1196 }
1197 
1198 void
1199 savectx(kthread_t *t)
1200 {
1201 	ASSERT(t == curthread);
1202 
1203 	if (t->t_ctx != NULL) {
1204 		struct ctxop *ctx, *head;
1205 
1206 		/* Forward traversal */
1207 		ctx = head = t->t_ctx;
1208 		do {
1209 			if (ctx->save_op != NULL) {
1210 				ctx->save_ts = gethrtime_unscaled();
1211 				(ctx->save_op)(ctx->arg);
1212 			}
1213 			ctx = ctx->next;
1214 		} while (ctx != head);
1215 	}
1216 }
1217 
1218 void
1219 restorectx(kthread_t *t)
1220 {
1221 	ASSERT(t == curthread);
1222 
1223 	if (t->t_ctx != NULL) {
1224 		struct ctxop *ctx, *tail;
1225 
1226 		/* Backward traversal (starting at the tail) */
1227 		ctx = tail = t->t_ctx->prev;
1228 		do {
1229 			if (ctx->restore_op != NULL) {
1230 				ctx->restore_ts = gethrtime_unscaled();
1231 				(ctx->restore_op)(ctx->arg);
1232 			}
1233 			ctx = ctx->prev;
1234 		} while (ctx != tail);
1235 	}
1236 }
1237 
1238 void
1239 forkctx(kthread_t *t, kthread_t *ct)
1240 {
1241 	if (t->t_ctx != NULL) {
1242 		struct ctxop *ctx, *head;
1243 
1244 		/* Forward traversal */
1245 		ctx = head = t->t_ctx;
1246 		do {
1247 			if (ctx->fork_op != NULL) {
1248 				(ctx->fork_op)(t, ct);
1249 			}
1250 			ctx = ctx->next;
1251 		} while (ctx != head);
1252 	}
1253 }
1254 
1255 /*
1256  * Note that this operator is only invoked via the _lwp_create
1257  * system call.  The system may have other reasons to create lwps
1258  * e.g. the agent lwp or the doors unreferenced lwp.
1259  */
1260 void
1261 lwp_createctx(kthread_t *t, kthread_t *ct)
1262 {
1263 	if (t->t_ctx != NULL) {
1264 		struct ctxop *ctx, *head;
1265 
1266 		/* Forward traversal */
1267 		ctx = head = t->t_ctx;
1268 		do {
1269 			if (ctx->lwp_create_op != NULL) {
1270 				(ctx->lwp_create_op)(t, ct);
1271 			}
1272 			ctx = ctx->next;
1273 		} while (ctx != head);
1274 	}
1275 }
1276 
1277 /*
1278  * exitctx is called from thread_exit() and lwp_exit() to perform any actions
1279  * needed when the thread/LWP leaves the processor for the last time. This
1280  * routine is not intended to deal with freeing memory; freectx() is used for
1281  * that purpose during thread_free(). This routine is provided to allow for
1282  * clean-up that can't wait until thread_free().
1283  */
1284 void
1285 exitctx(kthread_t *t)
1286 {
1287 	if (t->t_ctx != NULL) {
1288 		struct ctxop *ctx, *head;
1289 
1290 		/* Forward traversal */
1291 		ctx = head = t->t_ctx;
1292 		do {
1293 			if (ctx->exit_op != NULL) {
1294 				(ctx->exit_op)(t);
1295 			}
1296 			ctx = ctx->next;
1297 		} while (ctx != head);
1298 	}
1299 }
1300 
1301 /*
1302  * freectx is called from thread_free() and exec() to get
1303  * rid of old thread context ops.
1304  */
1305 void
1306 freectx(kthread_t *t, int isexec)
1307 {
1308 	kpreempt_disable();
1309 	if (t->t_ctx != NULL) {
1310 		struct ctxop *ctx, *head;
1311 
1312 		ctx = head = t->t_ctx;
1313 		t->t_ctx = NULL;
1314 		do {
1315 			struct ctxop *next = ctx->next;
1316 
1317 			if (ctx->free_op != NULL) {
1318 				(ctx->free_op)(ctx->arg, isexec);
1319 			}
1320 			kmem_free(ctx, sizeof (struct ctxop));
1321 			ctx = next;
1322 		} while (ctx != head);
1323 	}
1324 	kpreempt_enable();
1325 }
1326 
1327 /*
1328  * freectx_ctx is called from lwp_create() when lwp is reused from
1329  * lwp_deathrow and its thread structure is added to thread_deathrow.
1330  * The thread structure to which this ctx was attached may be already
1331  * freed by the thread reaper so free_op implementations shouldn't rely
1332  * on thread structure to which this ctx was attached still being around.
1333  */
1334 void
1335 freectx_ctx(struct ctxop *ctx)
1336 {
1337 	struct ctxop *head = ctx;
1338 
1339 	ASSERT(ctx != NULL);
1340 
1341 	kpreempt_disable();
1342 
1343 	head = ctx;
1344 	do {
1345 		struct ctxop *next = ctx->next;
1346 
1347 		if (ctx->free_op != NULL) {
1348 			(ctx->free_op)(ctx->arg, 0);
1349 		}
1350 		kmem_free(ctx, sizeof (struct ctxop));
1351 		ctx = next;
1352 	} while (ctx != head);
1353 	kpreempt_enable();
1354 }
1355 
1356 /*
1357  * Set the thread running; arrange for it to be swapped in if necessary.
1358  */
1359 void
1360 setrun_locked(kthread_t *t)
1361 {
1362 	ASSERT(THREAD_LOCK_HELD(t));
1363 	if (t->t_state == TS_SLEEP) {
1364 		/*
1365 		 * Take off sleep queue.
1366 		 */
1367 		SOBJ_UNSLEEP(t->t_sobj_ops, t);
1368 	} else if (t->t_state & (TS_RUN | TS_ONPROC)) {
1369 		/*
1370 		 * Already on dispatcher queue.
1371 		 */
1372 		return;
1373 	} else if (t->t_state == TS_WAIT) {
1374 		waitq_setrun(t);
1375 	} else if (t->t_state == TS_STOPPED) {
1376 		/*
1377 		 * All of the sending of SIGCONT (TC_XSTART) and /proc
1378 		 * (TC_PSTART) and lwp_continue() (TC_CSTART) must have
1379 		 * requested that the thread be run.
1380 		 * Just calling setrun() is not sufficient to set a stopped
1381 		 * thread running.  TP_TXSTART is always set if the thread
1382 		 * is not stopped by a jobcontrol stop signal.
1383 		 * TP_TPSTART is always set if /proc is not controlling it.
1384 		 * TP_TCSTART is always set if lwp_suspend() didn't stop it.
1385 		 * The thread won't be stopped unless one of these
1386 		 * three mechanisms did it.
1387 		 *
1388 		 * These flags must be set before calling setrun_locked(t).
1389 		 * They can't be passed as arguments because the streams
1390 		 * code calls setrun() indirectly and the mechanism for
1391 		 * doing so admits only one argument.  Note that the
1392 		 * thread must be locked in order to change t_schedflags.
1393 		 */
1394 		if ((t->t_schedflag & TS_ALLSTART) != TS_ALLSTART)
1395 			return;
1396 		/*
1397 		 * Process is no longer stopped (a thread is running).
1398 		 */
1399 		t->t_whystop = 0;
1400 		t->t_whatstop = 0;
1401 		/*
1402 		 * Strictly speaking, we do not have to clear these
1403 		 * flags here; they are cleared on entry to stop().
1404 		 * However, they are confusing when doing kernel
1405 		 * debugging or when they are revealed by ps(1).
1406 		 */
1407 		t->t_schedflag &= ~TS_ALLSTART;
1408 		THREAD_TRANSITION(t);	/* drop stopped-thread lock */
1409 		ASSERT(t->t_lockp == &transition_lock);
1410 		ASSERT(t->t_wchan0 == NULL && t->t_wchan == NULL);
1411 		/*
1412 		 * Let the class put the process on the dispatcher queue.
1413 		 */
1414 		CL_SETRUN(t);
1415 	}
1416 }
1417 
1418 void
1419 setrun(kthread_t *t)
1420 {
1421 	thread_lock(t);
1422 	setrun_locked(t);
1423 	thread_unlock(t);
1424 }
1425 
1426 /*
1427  * Unpin an interrupted thread.
1428  *	When an interrupt occurs, the interrupt is handled on the stack
1429  *	of an interrupt thread, taken from a pool linked to the CPU structure.
1430  *
1431  *	When swtch() is switching away from an interrupt thread because it
1432  *	blocked or was preempted, this routine is called to complete the
1433  *	saving of the interrupted thread state, and returns the interrupted
1434  *	thread pointer so it may be resumed.
1435  *
1436  *	Called by swtch() only at high spl.
1437  */
1438 kthread_t *
1439 thread_unpin()
1440 {
1441 	kthread_t	*t = curthread;	/* current thread */
1442 	kthread_t	*itp;		/* interrupted thread */
1443 	int		i;		/* interrupt level */
1444 	extern int	intr_passivate();
1445 
1446 	ASSERT(t->t_intr != NULL);
1447 
1448 	itp = t->t_intr;		/* interrupted thread */
1449 	t->t_intr = NULL;		/* clear interrupt ptr */
1450 
1451 	smt_end_intr();
1452 
1453 	/*
1454 	 * Get state from interrupt thread for the one
1455 	 * it interrupted.
1456 	 */
1457 
1458 	i = intr_passivate(t, itp);
1459 
1460 	TRACE_5(TR_FAC_INTR, TR_INTR_PASSIVATE,
1461 	    "intr_passivate:level %d curthread %p (%T) ithread %p (%T)",
1462 	    i, t, t, itp, itp);
1463 
1464 	/*
1465 	 * Dissociate the current thread from the interrupted thread's LWP.
1466 	 */
1467 	t->t_lwp = NULL;
1468 
1469 	/*
1470 	 * Interrupt handlers above the level that spinlocks block must
1471 	 * not block.
1472 	 */
1473 #if DEBUG
1474 	if (i < 0 || i > LOCK_LEVEL)
1475 		cmn_err(CE_PANIC, "thread_unpin: ipl out of range %x", i);
1476 #endif
1477 
1478 	/*
1479 	 * Compute the CPU's base interrupt level based on the active
1480 	 * interrupts.
1481 	 */
1482 	ASSERT(CPU->cpu_intr_actv & (1 << i));
1483 	set_base_spl();
1484 
1485 	return (itp);
1486 }
1487 
1488 /*
1489  * Create and initialize an interrupt thread.
1490  *	Returns non-zero on error.
1491  *	Called at spl7() or better.
1492  */
1493 void
1494 thread_create_intr(struct cpu *cp)
1495 {
1496 	kthread_t *tp;
1497 
1498 	tp = thread_create(NULL, 0,
1499 	    (void (*)())thread_create_intr, NULL, 0, &p0, TS_ONPROC, 0);
1500 
1501 	/*
1502 	 * Set the thread in the TS_FREE state.  The state will change
1503 	 * to TS_ONPROC only while the interrupt is active.  Think of these
1504 	 * as being on a private free list for the CPU.  Being TS_FREE keeps
1505 	 * inactive interrupt threads out of debugger thread lists.
1506 	 *
1507 	 * We cannot call thread_create with TS_FREE because of the current
1508 	 * checks there for ONPROC.  Fix this when thread_create takes flags.
1509 	 */
1510 	THREAD_FREEINTR(tp, cp);
1511 
1512 	/*
1513 	 * Nobody should ever reference the credentials of an interrupt
1514 	 * thread so make it NULL to catch any such references.
1515 	 */
1516 	tp->t_cred = NULL;
1517 	tp->t_flag |= T_INTR_THREAD;
1518 	tp->t_cpu = cp;
1519 	tp->t_bound_cpu = cp;
1520 	tp->t_disp_queue = cp->cpu_disp;
1521 	tp->t_affinitycnt = 1;
1522 	tp->t_preempt = 1;
1523 
1524 	/*
1525 	 * Don't make a user-requested binding on this thread so that
1526 	 * the processor can be offlined.
1527 	 */
1528 	tp->t_bind_cpu = PBIND_NONE;	/* no USER-requested binding */
1529 	tp->t_bind_pset = PS_NONE;
1530 
1531 #if defined(__x86)
1532 	tp->t_stk -= STACK_ALIGN;
1533 	*(tp->t_stk) = 0;		/* terminate intr thread stack */
1534 #endif
1535 
1536 	/*
1537 	 * Link onto CPU's interrupt pool.
1538 	 */
1539 	tp->t_link = cp->cpu_intr_thread;
1540 	cp->cpu_intr_thread = tp;
1541 }
1542 
1543 /*
1544  * TSD -- THREAD SPECIFIC DATA
1545  */
1546 static kmutex_t		tsd_mutex;	 /* linked list spin lock */
1547 static uint_t		tsd_nkeys;	 /* size of destructor array */
1548 /* per-key destructor funcs */
1549 static void		(**tsd_destructor)(void *);
1550 /* list of tsd_thread's */
1551 static struct tsd_thread	*tsd_list;
1552 
1553 /*
1554  * Default destructor
1555  *	Needed because NULL destructor means that the key is unused
1556  */
1557 /* ARGSUSED */
1558 void
1559 tsd_defaultdestructor(void *value)
1560 {}
1561 
1562 /*
1563  * Create a key (index into per thread array)
1564  *	Locks out tsd_create, tsd_destroy, and tsd_exit
1565  *	May allocate memory with lock held
1566  */
1567 void
1568 tsd_create(uint_t *keyp, void (*destructor)(void *))
1569 {
1570 	int	i;
1571 	uint_t	nkeys;
1572 
1573 	/*
1574 	 * if key is allocated, do nothing
1575 	 */
1576 	mutex_enter(&tsd_mutex);
1577 	if (*keyp) {
1578 		mutex_exit(&tsd_mutex);
1579 		return;
1580 	}
1581 	/*
1582 	 * find an unused key
1583 	 */
1584 	if (destructor == NULL)
1585 		destructor = tsd_defaultdestructor;
1586 
1587 	for (i = 0; i < tsd_nkeys; ++i)
1588 		if (tsd_destructor[i] == NULL)
1589 			break;
1590 
1591 	/*
1592 	 * if no unused keys, increase the size of the destructor array
1593 	 */
1594 	if (i == tsd_nkeys) {
1595 		if ((nkeys = (tsd_nkeys << 1)) == 0)
1596 			nkeys = 1;
1597 		tsd_destructor =
1598 		    (void (**)(void *))tsd_realloc((void *)tsd_destructor,
1599 		    (size_t)(tsd_nkeys * sizeof (void (*)(void *))),
1600 		    (size_t)(nkeys * sizeof (void (*)(void *))));
1601 		tsd_nkeys = nkeys;
1602 	}
1603 
1604 	/*
1605 	 * allocate the next available unused key
1606 	 */
1607 	tsd_destructor[i] = destructor;
1608 	*keyp = i + 1;
1609 	mutex_exit(&tsd_mutex);
1610 }
1611 
1612 /*
1613  * Destroy a key -- this is for unloadable modules
1614  *
1615  * Assumes that the caller is preventing tsd_set and tsd_get
1616  * Locks out tsd_create, tsd_destroy, and tsd_exit
1617  * May free memory with lock held
1618  */
1619 void
1620 tsd_destroy(uint_t *keyp)
1621 {
1622 	uint_t key;
1623 	struct tsd_thread *tsd;
1624 
1625 	/*
1626 	 * protect the key namespace and our destructor lists
1627 	 */
1628 	mutex_enter(&tsd_mutex);
1629 	key = *keyp;
1630 	*keyp = 0;
1631 
1632 	ASSERT(key <= tsd_nkeys);
1633 
1634 	/*
1635 	 * if the key is valid
1636 	 */
1637 	if (key != 0) {
1638 		uint_t k = key - 1;
1639 		/*
1640 		 * for every thread with TSD, call key's destructor
1641 		 */
1642 		for (tsd = tsd_list; tsd; tsd = tsd->ts_next) {
1643 			/*
1644 			 * no TSD for key in this thread
1645 			 */
1646 			if (key > tsd->ts_nkeys)
1647 				continue;
1648 			/*
1649 			 * call destructor for key
1650 			 */
1651 			if (tsd->ts_value[k] && tsd_destructor[k])
1652 				(*tsd_destructor[k])(tsd->ts_value[k]);
1653 			/*
1654 			 * reset value for key
1655 			 */
1656 			tsd->ts_value[k] = NULL;
1657 		}
1658 		/*
1659 		 * actually free the key (NULL destructor == unused)
1660 		 */
1661 		tsd_destructor[k] = NULL;
1662 	}
1663 
1664 	mutex_exit(&tsd_mutex);
1665 }
1666 
1667 /*
1668  * Quickly return the per thread value that was stored with the specified key
1669  * Assumes the caller is protecting key from tsd_create and tsd_destroy
1670  */
1671 void *
1672 tsd_get(uint_t key)
1673 {
1674 	return (tsd_agent_get(curthread, key));
1675 }
1676 
1677 /*
1678  * Set a per thread value indexed with the specified key
1679  */
1680 int
1681 tsd_set(uint_t key, void *value)
1682 {
1683 	return (tsd_agent_set(curthread, key, value));
1684 }
1685 
1686 /*
1687  * Like tsd_get(), except that the agent lwp can get the tsd of
1688  * another thread in the same process (the agent thread only runs when the
1689  * process is completely stopped by /proc), or syslwp is creating a new lwp.
1690  */
1691 void *
1692 tsd_agent_get(kthread_t *t, uint_t key)
1693 {
1694 	struct tsd_thread *tsd = t->t_tsd;
1695 
1696 	ASSERT(t == curthread ||
1697 	    ttoproc(t)->p_agenttp == curthread || t->t_state == TS_STOPPED);
1698 
1699 	if (key && tsd != NULL && key <= tsd->ts_nkeys)
1700 		return (tsd->ts_value[key - 1]);
1701 	return (NULL);
1702 }
1703 
1704 /*
1705  * Like tsd_set(), except that the agent lwp can set the tsd of
1706  * another thread in the same process, or syslwp can set the tsd
1707  * of a thread it's in the middle of creating.
1708  *
1709  * Assumes the caller is protecting key from tsd_create and tsd_destroy
1710  * May lock out tsd_destroy (and tsd_create), may allocate memory with
1711  * lock held
1712  */
1713 int
1714 tsd_agent_set(kthread_t *t, uint_t key, void *value)
1715 {
1716 	struct tsd_thread *tsd = t->t_tsd;
1717 
1718 	ASSERT(t == curthread ||
1719 	    ttoproc(t)->p_agenttp == curthread || t->t_state == TS_STOPPED);
1720 
1721 	if (key == 0)
1722 		return (EINVAL);
1723 	if (tsd == NULL)
1724 		tsd = t->t_tsd = kmem_zalloc(sizeof (*tsd), KM_SLEEP);
1725 	if (key <= tsd->ts_nkeys) {
1726 		tsd->ts_value[key - 1] = value;
1727 		return (0);
1728 	}
1729 
1730 	ASSERT(key <= tsd_nkeys);
1731 
1732 	/*
1733 	 * lock out tsd_destroy()
1734 	 */
1735 	mutex_enter(&tsd_mutex);
1736 	if (tsd->ts_nkeys == 0) {
1737 		/*
1738 		 * Link onto list of threads with TSD
1739 		 */
1740 		if ((tsd->ts_next = tsd_list) != NULL)
1741 			tsd_list->ts_prev = tsd;
1742 		tsd_list = tsd;
1743 	}
1744 
1745 	/*
1746 	 * Allocate thread local storage and set the value for key
1747 	 */
1748 	tsd->ts_value = tsd_realloc(tsd->ts_value,
1749 	    tsd->ts_nkeys * sizeof (void *),
1750 	    key * sizeof (void *));
1751 	tsd->ts_nkeys = key;
1752 	tsd->ts_value[key - 1] = value;
1753 	mutex_exit(&tsd_mutex);
1754 
1755 	return (0);
1756 }
1757 
1758 
1759 /*
1760  * Return the per thread value that was stored with the specified key
1761  *	If necessary, create the key and the value
1762  *	Assumes the caller is protecting *keyp from tsd_destroy
1763  */
1764 void *
1765 tsd_getcreate(uint_t *keyp, void (*destroy)(void *), void *(*allocate)(void))
1766 {
1767 	void *value;
1768 	uint_t key = *keyp;
1769 	struct tsd_thread *tsd = curthread->t_tsd;
1770 
1771 	if (tsd == NULL)
1772 		tsd = curthread->t_tsd = kmem_zalloc(sizeof (*tsd), KM_SLEEP);
1773 	if (key && key <= tsd->ts_nkeys && (value = tsd->ts_value[key - 1]))
1774 		return (value);
1775 	if (key == 0)
1776 		tsd_create(keyp, destroy);
1777 	(void) tsd_set(*keyp, value = (*allocate)());
1778 
1779 	return (value);
1780 }
1781 
1782 /*
1783  * Called from thread_exit() to run the destructor function for each tsd
1784  *	Locks out tsd_create and tsd_destroy
1785  *	Assumes that the destructor *DOES NOT* use tsd
1786  */
1787 void
1788 tsd_exit(void)
1789 {
1790 	int i;
1791 	struct tsd_thread *tsd = curthread->t_tsd;
1792 
1793 	if (tsd == NULL)
1794 		return;
1795 
1796 	if (tsd->ts_nkeys == 0) {
1797 		kmem_free(tsd, sizeof (*tsd));
1798 		curthread->t_tsd = NULL;
1799 		return;
1800 	}
1801 
1802 	/*
1803 	 * lock out tsd_create and tsd_destroy, call
1804 	 * the destructor, and mark the value as destroyed.
1805 	 */
1806 	mutex_enter(&tsd_mutex);
1807 
1808 	for (i = 0; i < tsd->ts_nkeys; i++) {
1809 		if (tsd->ts_value[i] && tsd_destructor[i])
1810 			(*tsd_destructor[i])(tsd->ts_value[i]);
1811 		tsd->ts_value[i] = NULL;
1812 	}
1813 
1814 	/*
1815 	 * remove from linked list of threads with TSD
1816 	 */
1817 	if (tsd->ts_next)
1818 		tsd->ts_next->ts_prev = tsd->ts_prev;
1819 	if (tsd->ts_prev)
1820 		tsd->ts_prev->ts_next = tsd->ts_next;
1821 	if (tsd_list == tsd)
1822 		tsd_list = tsd->ts_next;
1823 
1824 	mutex_exit(&tsd_mutex);
1825 
1826 	/*
1827 	 * free up the TSD
1828 	 */
1829 	kmem_free(tsd->ts_value, tsd->ts_nkeys * sizeof (void *));
1830 	kmem_free(tsd, sizeof (struct tsd_thread));
1831 	curthread->t_tsd = NULL;
1832 }
1833 
1834 /*
1835  * realloc
1836  */
1837 static void *
1838 tsd_realloc(void *old, size_t osize, size_t nsize)
1839 {
1840 	void *new;
1841 
1842 	new = kmem_zalloc(nsize, KM_SLEEP);
1843 	if (old) {
1844 		bcopy(old, new, osize);
1845 		kmem_free(old, osize);
1846 	}
1847 	return (new);
1848 }
1849 
1850 /*
1851  * Return non-zero if an interrupt is being serviced.
1852  */
1853 int
1854 servicing_interrupt()
1855 {
1856 	int onintr = 0;
1857 
1858 	/* Are we an interrupt thread */
1859 	if (curthread->t_flag & T_INTR_THREAD)
1860 		return (1);
1861 	/* Are we servicing a high level interrupt? */
1862 	if (CPU_ON_INTR(CPU)) {
1863 		kpreempt_disable();
1864 		onintr = CPU_ON_INTR(CPU);
1865 		kpreempt_enable();
1866 	}
1867 	return (onintr);
1868 }
1869 
1870 
1871 /*
1872  * Change the dispatch priority of a thread in the system.
1873  * Used when raising or lowering a thread's priority.
1874  * (E.g., priority inheritance)
1875  *
1876  * Since threads are queued according to their priority, we
1877  * we must check the thread's state to determine whether it
1878  * is on a queue somewhere. If it is, we've got to:
1879  *
1880  *	o Dequeue the thread.
1881  *	o Change its effective priority.
1882  *	o Enqueue the thread.
1883  *
1884  * Assumptions: The thread whose priority we wish to change
1885  * must be locked before we call thread_change_(e)pri().
1886  * The thread_change(e)pri() function doesn't drop the thread
1887  * lock--that must be done by its caller.
1888  */
1889 void
1890 thread_change_epri(kthread_t *t, pri_t disp_pri)
1891 {
1892 	uint_t	state;
1893 
1894 	ASSERT(THREAD_LOCK_HELD(t));
1895 
1896 	/*
1897 	 * If the inherited priority hasn't actually changed,
1898 	 * just return.
1899 	 */
1900 	if (t->t_epri == disp_pri)
1901 		return;
1902 
1903 	state = t->t_state;
1904 
1905 	/*
1906 	 * If it's not on a queue, change the priority with impunity.
1907 	 */
1908 	if ((state & (TS_SLEEP | TS_RUN | TS_WAIT)) == 0) {
1909 		t->t_epri = disp_pri;
1910 		if (state == TS_ONPROC) {
1911 			cpu_t *cp = t->t_disp_queue->disp_cpu;
1912 
1913 			if (t == cp->cpu_dispthread)
1914 				cp->cpu_dispatch_pri = DISP_PRIO(t);
1915 		}
1916 	} else if (state == TS_SLEEP) {
1917 		/*
1918 		 * Take the thread out of its sleep queue.
1919 		 * Change the inherited priority.
1920 		 * Re-enqueue the thread.
1921 		 * Each synchronization object exports a function
1922 		 * to do this in an appropriate manner.
1923 		 */
1924 		SOBJ_CHANGE_EPRI(t->t_sobj_ops, t, disp_pri);
1925 	} else if (state == TS_WAIT) {
1926 		/*
1927 		 * Re-enqueue a thread on the wait queue if its
1928 		 * effective priority needs to change.
1929 		 */
1930 		if (disp_pri != t->t_epri)
1931 			waitq_change_pri(t, disp_pri);
1932 	} else {
1933 		/*
1934 		 * The thread is on a run queue.
1935 		 * Note: setbackdq() may not put the thread
1936 		 * back on the same run queue where it originally
1937 		 * resided.
1938 		 */
1939 		(void) dispdeq(t);
1940 		t->t_epri = disp_pri;
1941 		setbackdq(t);
1942 	}
1943 	schedctl_set_cidpri(t);
1944 }
1945 
1946 /*
1947  * Function: Change the t_pri field of a thread.
1948  * Side Effects: Adjust the thread ordering on a run queue
1949  *		 or sleep queue, if necessary.
1950  * Returns: 1 if the thread was on a run queue, else 0.
1951  */
1952 int
1953 thread_change_pri(kthread_t *t, pri_t disp_pri, int front)
1954 {
1955 	uint_t	state;
1956 	int	on_rq = 0;
1957 
1958 	ASSERT(THREAD_LOCK_HELD(t));
1959 
1960 	state = t->t_state;
1961 	THREAD_WILLCHANGE_PRI(t, disp_pri);
1962 
1963 	/*
1964 	 * If it's not on a queue, change the priority with impunity.
1965 	 */
1966 	if ((state & (TS_SLEEP | TS_RUN | TS_WAIT)) == 0) {
1967 		t->t_pri = disp_pri;
1968 
1969 		if (state == TS_ONPROC) {
1970 			cpu_t *cp = t->t_disp_queue->disp_cpu;
1971 
1972 			if (t == cp->cpu_dispthread)
1973 				cp->cpu_dispatch_pri = DISP_PRIO(t);
1974 		}
1975 	} else if (state == TS_SLEEP) {
1976 		/*
1977 		 * If the priority has changed, take the thread out of
1978 		 * its sleep queue and change the priority.
1979 		 * Re-enqueue the thread.
1980 		 * Each synchronization object exports a function
1981 		 * to do this in an appropriate manner.
1982 		 */
1983 		if (disp_pri != t->t_pri)
1984 			SOBJ_CHANGE_PRI(t->t_sobj_ops, t, disp_pri);
1985 	} else if (state == TS_WAIT) {
1986 		/*
1987 		 * Re-enqueue a thread on the wait queue if its
1988 		 * priority needs to change.
1989 		 */
1990 		if (disp_pri != t->t_pri)
1991 			waitq_change_pri(t, disp_pri);
1992 	} else {
1993 		/*
1994 		 * The thread is on a run queue.
1995 		 * Note: setbackdq() may not put the thread
1996 		 * back on the same run queue where it originally
1997 		 * resided.
1998 		 *
1999 		 * We still requeue the thread even if the priority
2000 		 * is unchanged to preserve round-robin (and other)
2001 		 * effects between threads of the same priority.
2002 		 */
2003 		on_rq = dispdeq(t);
2004 		ASSERT(on_rq);
2005 		t->t_pri = disp_pri;
2006 		if (front) {
2007 			setfrontdq(t);
2008 		} else {
2009 			setbackdq(t);
2010 		}
2011 	}
2012 	schedctl_set_cidpri(t);
2013 	return (on_rq);
2014 }
2015 
2016 /*
2017  * Tunable kmem_stackinfo is set, fill the kernel thread stack with a
2018  * specific pattern.
2019  */
2020 static void
2021 stkinfo_begin(kthread_t *t)
2022 {
2023 	caddr_t	start;	/* stack start */
2024 	caddr_t	end;	/* stack end  */
2025 	uint64_t *ptr;	/* pattern pointer */
2026 
2027 	/*
2028 	 * Stack grows up or down, see thread_create(),
2029 	 * compute stack memory area start and end (start < end).
2030 	 */
2031 	if (t->t_stk > t->t_stkbase) {
2032 		/* stack grows down */
2033 		start = t->t_stkbase;
2034 		end = t->t_stk;
2035 	} else {
2036 		/* stack grows up */
2037 		start = t->t_stk;
2038 		end = t->t_stkbase;
2039 	}
2040 
2041 	/*
2042 	 * Stackinfo pattern size is 8 bytes. Ensure proper 8 bytes
2043 	 * alignement for start and end in stack area boundaries
2044 	 * (protection against corrupt t_stkbase/t_stk data).
2045 	 */
2046 	if ((((uintptr_t)start) & 0x7) != 0) {
2047 		start = (caddr_t)((((uintptr_t)start) & (~0x7)) + 8);
2048 	}
2049 	end = (caddr_t)(((uintptr_t)end) & (~0x7));
2050 
2051 	if ((end <= start) || (end - start) > (1024 * 1024)) {
2052 		/* negative or stack size > 1 meg, assume bogus */
2053 		return;
2054 	}
2055 
2056 	/* fill stack area with a pattern (instead of zeros) */
2057 	ptr = (uint64_t *)((void *)start);
2058 	while (ptr < (uint64_t *)((void *)end)) {
2059 		*ptr++ = KMEM_STKINFO_PATTERN;
2060 	}
2061 }
2062 
2063 
2064 /*
2065  * Tunable kmem_stackinfo is set, create stackinfo log if doesn't already exist,
2066  * compute the percentage of kernel stack really used, and set in the log
2067  * if it's the latest highest percentage.
2068  */
2069 static void
2070 stkinfo_end(kthread_t *t)
2071 {
2072 	caddr_t	start;	/* stack start */
2073 	caddr_t	end;	/* stack end  */
2074 	uint64_t *ptr;	/* pattern pointer */
2075 	size_t stksz;	/* stack size */
2076 	size_t smallest = 0;
2077 	size_t percent = 0;
2078 	uint_t index = 0;
2079 	uint_t i;
2080 	static size_t smallest_percent = (size_t)-1;
2081 	static uint_t full = 0;
2082 
2083 	/* create the stackinfo log, if doesn't already exist */
2084 	mutex_enter(&kmem_stkinfo_lock);
2085 	if (kmem_stkinfo_log == NULL) {
2086 		kmem_stkinfo_log = (kmem_stkinfo_t *)
2087 		    kmem_zalloc(KMEM_STKINFO_LOG_SIZE *
2088 		    (sizeof (kmem_stkinfo_t)), KM_NOSLEEP);
2089 		if (kmem_stkinfo_log == NULL) {
2090 			mutex_exit(&kmem_stkinfo_lock);
2091 			return;
2092 		}
2093 	}
2094 	mutex_exit(&kmem_stkinfo_lock);
2095 
2096 	/*
2097 	 * Stack grows up or down, see thread_create(),
2098 	 * compute stack memory area start and end (start < end).
2099 	 */
2100 	if (t->t_stk > t->t_stkbase) {
2101 		/* stack grows down */
2102 		start = t->t_stkbase;
2103 		end = t->t_stk;
2104 	} else {
2105 		/* stack grows up */
2106 		start = t->t_stk;
2107 		end = t->t_stkbase;
2108 	}
2109 
2110 	/* stack size as found in kthread_t */
2111 	stksz = end - start;
2112 
2113 	/*
2114 	 * Stackinfo pattern size is 8 bytes. Ensure proper 8 bytes
2115 	 * alignement for start and end in stack area boundaries
2116 	 * (protection against corrupt t_stkbase/t_stk data).
2117 	 */
2118 	if ((((uintptr_t)start) & 0x7) != 0) {
2119 		start = (caddr_t)((((uintptr_t)start) & (~0x7)) + 8);
2120 	}
2121 	end = (caddr_t)(((uintptr_t)end) & (~0x7));
2122 
2123 	if ((end <= start) || (end - start) > (1024 * 1024)) {
2124 		/* negative or stack size > 1 meg, assume bogus */
2125 		return;
2126 	}
2127 
2128 	/* search until no pattern in the stack */
2129 	if (t->t_stk > t->t_stkbase) {
2130 		/* stack grows down */
2131 #if defined(__x86)
2132 		/*
2133 		 * 6 longs are pushed on stack, see thread_load(). Skip
2134 		 * them, so if kthread has never run, percent is zero.
2135 		 * 8 bytes alignement is preserved for a 32 bit kernel,
2136 		 * 6 x 4 = 24, 24 is a multiple of 8.
2137 		 *
2138 		 */
2139 		end -= (6 * sizeof (long));
2140 #endif
2141 		ptr = (uint64_t *)((void *)start);
2142 		while (ptr < (uint64_t *)((void *)end)) {
2143 			if (*ptr != KMEM_STKINFO_PATTERN) {
2144 				percent = stkinfo_percent(end,
2145 				    start, (caddr_t)ptr);
2146 				break;
2147 			}
2148 			ptr++;
2149 		}
2150 	} else {
2151 		/* stack grows up */
2152 		ptr = (uint64_t *)((void *)end);
2153 		ptr--;
2154 		while (ptr >= (uint64_t *)((void *)start)) {
2155 			if (*ptr != KMEM_STKINFO_PATTERN) {
2156 				percent = stkinfo_percent(start,
2157 				    end, (caddr_t)ptr);
2158 				break;
2159 			}
2160 			ptr--;
2161 		}
2162 	}
2163 
2164 	DTRACE_PROBE3(stack__usage, kthread_t *, t,
2165 	    size_t, stksz, size_t, percent);
2166 
2167 	if (percent == 0) {
2168 		return;
2169 	}
2170 
2171 	mutex_enter(&kmem_stkinfo_lock);
2172 	if (full == KMEM_STKINFO_LOG_SIZE && percent < smallest_percent) {
2173 		/*
2174 		 * The log is full and already contains the highest values
2175 		 */
2176 		mutex_exit(&kmem_stkinfo_lock);
2177 		return;
2178 	}
2179 
2180 	/* keep a log of the highest used stack */
2181 	for (i = 0; i < KMEM_STKINFO_LOG_SIZE; i++) {
2182 		if (kmem_stkinfo_log[i].percent == 0) {
2183 			index = i;
2184 			full++;
2185 			break;
2186 		}
2187 		if (smallest == 0) {
2188 			smallest = kmem_stkinfo_log[i].percent;
2189 			index = i;
2190 			continue;
2191 		}
2192 		if (kmem_stkinfo_log[i].percent < smallest) {
2193 			smallest = kmem_stkinfo_log[i].percent;
2194 			index = i;
2195 		}
2196 	}
2197 
2198 	if (percent >= kmem_stkinfo_log[index].percent) {
2199 		kmem_stkinfo_log[index].kthread = (caddr_t)t;
2200 		kmem_stkinfo_log[index].t_startpc = (caddr_t)t->t_startpc;
2201 		kmem_stkinfo_log[index].start = start;
2202 		kmem_stkinfo_log[index].stksz = stksz;
2203 		kmem_stkinfo_log[index].percent = percent;
2204 		kmem_stkinfo_log[index].t_tid = t->t_tid;
2205 		kmem_stkinfo_log[index].cmd[0] = '\0';
2206 		if (t->t_tid != 0) {
2207 			stksz = strlen((t->t_procp)->p_user.u_comm);
2208 			if (stksz >= KMEM_STKINFO_STR_SIZE) {
2209 				stksz = KMEM_STKINFO_STR_SIZE - 1;
2210 				kmem_stkinfo_log[index].cmd[stksz] = '\0';
2211 			} else {
2212 				stksz += 1;
2213 			}
2214 			(void) memcpy(kmem_stkinfo_log[index].cmd,
2215 			    (t->t_procp)->p_user.u_comm, stksz);
2216 		}
2217 		if (percent < smallest_percent) {
2218 			smallest_percent = percent;
2219 		}
2220 	}
2221 	mutex_exit(&kmem_stkinfo_lock);
2222 }
2223 
2224 /*
2225  * Tunable kmem_stackinfo is set, compute stack utilization percentage.
2226  */
2227 static size_t
2228 stkinfo_percent(caddr_t t_stk, caddr_t t_stkbase, caddr_t sp)
2229 {
2230 	size_t percent;
2231 	size_t s;
2232 
2233 	if (t_stk > t_stkbase) {
2234 		/* stack grows down */
2235 		if (sp > t_stk) {
2236 			return (0);
2237 		}
2238 		if (sp < t_stkbase) {
2239 			return (100);
2240 		}
2241 		percent = t_stk - sp + 1;
2242 		s = t_stk - t_stkbase + 1;
2243 	} else {
2244 		/* stack grows up */
2245 		if (sp < t_stk) {
2246 			return (0);
2247 		}
2248 		if (sp > t_stkbase) {
2249 			return (100);
2250 		}
2251 		percent = sp - t_stk + 1;
2252 		s = t_stkbase - t_stk + 1;
2253 	}
2254 	percent = ((100 * percent) / s) + 1;
2255 	if (percent > 100) {
2256 		percent = 100;
2257 	}
2258 	return (percent);
2259 }
2260 
2261 /*
2262  * NOTE: This will silently truncate a name > THREAD_NAME_MAX - 1 characters
2263  * long.  It is expected that callers (acting on behalf of userland clients)
2264  * will perform any required checks to return the correct error semantics.
2265  * It is also expected callers on behalf of userland clients have done
2266  * any necessary permission checks.
2267  */
2268 int
2269 thread_setname(kthread_t *t, const char *name)
2270 {
2271 	char *buf = NULL;
2272 
2273 	/*
2274 	 * We optimistically assume that a thread's name will only be set
2275 	 * once and so allocate memory in preparation of setting t_name.
2276 	 * If it turns out a name has already been set, we just discard (free)
2277 	 * the buffer we just allocated and reuse the current buffer
2278 	 * (as all should be THREAD_NAME_MAX large).
2279 	 *
2280 	 * Such an arrangement means over the lifetime of a kthread_t, t_name
2281 	 * is either NULL or has one value (the address of the buffer holding
2282 	 * the current thread name).   The assumption is that most kthread_t
2283 	 * instances will not have a name assigned, so dynamically allocating
2284 	 * the memory should minimize the footprint of this feature, but by
2285 	 * having the buffer persist for the life of the thread, it simplifies
2286 	 * usage in highly constrained situations (e.g. dtrace).
2287 	 */
2288 	if (name != NULL && name[0] != '\0') {
2289 		for (size_t i = 0; name[i] != '\0'; i++) {
2290 			if (!isprint(name[i]))
2291 				return (EINVAL);
2292 		}
2293 
2294 		buf = kmem_zalloc(THREAD_NAME_MAX, KM_SLEEP);
2295 		(void) strlcpy(buf, name, THREAD_NAME_MAX);
2296 	}
2297 
2298 	mutex_enter(&ttoproc(t)->p_lock);
2299 	if (t->t_name == NULL) {
2300 		t->t_name = buf;
2301 	} else {
2302 		if (buf != NULL) {
2303 			(void) strlcpy(t->t_name, name, THREAD_NAME_MAX);
2304 			kmem_free(buf, THREAD_NAME_MAX);
2305 		} else {
2306 			bzero(t->t_name, THREAD_NAME_MAX);
2307 		}
2308 	}
2309 	mutex_exit(&ttoproc(t)->p_lock);
2310 	return (0);
2311 }
2312 
2313 int
2314 thread_vsetname(kthread_t *t, const char *fmt, ...)
2315 {
2316 	char name[THREAD_NAME_MAX];
2317 	va_list va;
2318 	int rc;
2319 
2320 	va_start(va, fmt);
2321 	rc = vsnprintf(name, sizeof (name), fmt, va);
2322 	va_end(va);
2323 
2324 	if (rc < 0)
2325 		return (EINVAL);
2326 
2327 	if (rc >= sizeof (name))
2328 		return (ENAMETOOLONG);
2329 
2330 	return (thread_setname(t, name));
2331 }
2332