xref: /illumos-gate/usr/src/uts/common/os/lwp.c (revision 1bc02a70)
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 2010 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 #include <sys/param.h>
28 #include <sys/types.h>
29 #include <sys/sysmacros.h>
30 #include <sys/systm.h>
31 #include <sys/thread.h>
32 #include <sys/proc.h>
33 #include <sys/task.h>
34 #include <sys/project.h>
35 #include <sys/signal.h>
36 #include <sys/errno.h>
37 #include <sys/vmparam.h>
38 #include <sys/stack.h>
39 #include <sys/procfs.h>
40 #include <sys/prsystm.h>
41 #include <sys/cpuvar.h>
42 #include <sys/kmem.h>
43 #include <sys/vtrace.h>
44 #include <sys/door.h>
45 #include <vm/seg_kp.h>
46 #include <sys/debug.h>
47 #include <sys/tnf.h>
48 #include <sys/schedctl.h>
49 #include <sys/poll.h>
50 #include <sys/copyops.h>
51 #include <sys/lwp_upimutex_impl.h>
52 #include <sys/cpupart.h>
53 #include <sys/lgrp.h>
54 #include <sys/rctl.h>
55 #include <sys/contract_impl.h>
56 #include <sys/cpc_impl.h>
57 #include <sys/sdt.h>
58 #include <sys/cmn_err.h>
59 #include <sys/brand.h>
60 #include <sys/cyclic.h>
61 #include <sys/pool.h>
62 
63 /* hash function for the lwpid hash table, p->p_tidhash[] */
64 #define	TIDHASH(tid, hash_sz)	((tid) & ((hash_sz) - 1))
65 
66 void *segkp_lwp;		/* cookie for pool of segkp resources */
67 extern void reapq_move_lq_to_tq(kthread_t *);
68 extern void freectx_ctx(struct ctxop *);
69 
70 /*
71  * Create a kernel thread associated with a particular system process.  Give
72  * it an LWP so that microstate accounting will be available for it.
73  */
74 kthread_t *
75 lwp_kernel_create(proc_t *p, void (*proc)(), void *arg, int state, pri_t pri)
76 {
77 	klwp_t *lwp;
78 
79 	VERIFY((p->p_flag & SSYS) != 0);
80 
81 	lwp = lwp_create(proc, arg, 0, p, state, pri, &t0.t_hold, syscid, 0);
82 
83 	VERIFY(lwp != NULL);
84 
85 	return (lwptot(lwp));
86 }
87 
88 /*
89  * Create a thread that appears to be stopped at sys_rtt.
90  */
91 klwp_t *
92 lwp_create(void (*proc)(), caddr_t arg, size_t len, proc_t *p,
93     int state, int pri, const k_sigset_t *smask, int cid, id_t lwpid)
94 {
95 	klwp_t *lwp = NULL;
96 	kthread_t *t;
97 	kthread_t *tx;
98 	cpupart_t *oldpart = NULL;
99 	size_t	stksize;
100 	caddr_t lwpdata = NULL;
101 	processorid_t	binding;
102 	int err = 0;
103 	kproject_t *oldkpj, *newkpj;
104 	void *bufp = NULL;
105 	klwp_t *curlwp;
106 	lwpent_t *lep;
107 	lwpdir_t *old_dir = NULL;
108 	uint_t old_dirsz = 0;
109 	tidhash_t *old_hash = NULL;
110 	uint_t old_hashsz = 0;
111 	ret_tidhash_t *ret_tidhash = NULL;
112 	int i;
113 	int rctlfail = 0;
114 	boolean_t branded = 0;
115 	struct ctxop *ctx = NULL;
116 
117 	ASSERT(cid != sysdccid);	/* system threads must start in SYS */
118 
119 	ASSERT(p != &p0);		/* No new LWPs in p0. */
120 
121 	mutex_enter(&p->p_lock);
122 	mutex_enter(&p->p_zone->zone_nlwps_lock);
123 	/*
124 	 * don't enforce rctl limits on system processes
125 	 */
126 	if (!CLASS_KERNEL(cid)) {
127 		if (p->p_task->tk_nlwps >= p->p_task->tk_nlwps_ctl)
128 			if (rctl_test(rc_task_lwps, p->p_task->tk_rctls, p,
129 			    1, 0) & RCT_DENY)
130 				rctlfail = 1;
131 		if (p->p_task->tk_proj->kpj_nlwps >=
132 		    p->p_task->tk_proj->kpj_nlwps_ctl)
133 			if (rctl_test(rc_project_nlwps,
134 			    p->p_task->tk_proj->kpj_rctls, p, 1, 0)
135 			    & RCT_DENY)
136 				rctlfail = 1;
137 		if (p->p_zone->zone_nlwps >= p->p_zone->zone_nlwps_ctl)
138 			if (rctl_test(rc_zone_nlwps, p->p_zone->zone_rctls, p,
139 			    1, 0) & RCT_DENY)
140 				rctlfail = 1;
141 	}
142 	if (rctlfail) {
143 		mutex_exit(&p->p_zone->zone_nlwps_lock);
144 		mutex_exit(&p->p_lock);
145 		return (NULL);
146 	}
147 	p->p_task->tk_nlwps++;
148 	p->p_task->tk_proj->kpj_nlwps++;
149 	p->p_zone->zone_nlwps++;
150 	mutex_exit(&p->p_zone->zone_nlwps_lock);
151 	mutex_exit(&p->p_lock);
152 
153 	curlwp = ttolwp(curthread);
154 	if (curlwp == NULL || (stksize = curlwp->lwp_childstksz) == 0)
155 		stksize = lwp_default_stksize;
156 
157 	if (CLASS_KERNEL(cid)) {
158 		/*
159 		 * Since we are creating an LWP in an SSYS process, we do not
160 		 * inherit anything from the current thread's LWP.  We set
161 		 * stksize and lwpdata to 0 in order to let thread_create()
162 		 * allocate a regular kernel thread stack for this thread.
163 		 */
164 		curlwp = NULL;
165 		stksize = 0;
166 		lwpdata = NULL;
167 
168 	} else if (stksize == lwp_default_stksize) {
169 		/*
170 		 * Try to reuse an <lwp,stack> from the LWP deathrow.
171 		 */
172 		if (lwp_reapcnt > 0) {
173 			mutex_enter(&reaplock);
174 			if ((t = lwp_deathrow) != NULL) {
175 				ASSERT(t->t_swap);
176 				lwp_deathrow = t->t_forw;
177 				lwp_reapcnt--;
178 				lwpdata = t->t_swap;
179 				lwp = t->t_lwp;
180 				ctx = t->t_ctx;
181 				t->t_swap = NULL;
182 				t->t_lwp = NULL;
183 				t->t_ctx = NULL;
184 				reapq_move_lq_to_tq(t);
185 			}
186 			mutex_exit(&reaplock);
187 			if (lwp != NULL) {
188 				lwp_stk_fini(lwp);
189 			}
190 			if (ctx != NULL) {
191 				freectx_ctx(ctx);
192 			}
193 		}
194 		if (lwpdata == NULL &&
195 		    (lwpdata = (caddr_t)segkp_cache_get(segkp_lwp)) == NULL) {
196 			mutex_enter(&p->p_lock);
197 			mutex_enter(&p->p_zone->zone_nlwps_lock);
198 			p->p_task->tk_nlwps--;
199 			p->p_task->tk_proj->kpj_nlwps--;
200 			p->p_zone->zone_nlwps--;
201 			mutex_exit(&p->p_zone->zone_nlwps_lock);
202 			mutex_exit(&p->p_lock);
203 			return (NULL);
204 		}
205 	} else {
206 		stksize = roundup(stksize, PAGESIZE);
207 		if ((lwpdata = (caddr_t)segkp_get(segkp, stksize,
208 		    (KPD_NOWAIT | KPD_HASREDZONE | KPD_LOCKED))) == NULL) {
209 			mutex_enter(&p->p_lock);
210 			mutex_enter(&p->p_zone->zone_nlwps_lock);
211 			p->p_task->tk_nlwps--;
212 			p->p_task->tk_proj->kpj_nlwps--;
213 			p->p_zone->zone_nlwps--;
214 			mutex_exit(&p->p_zone->zone_nlwps_lock);
215 			mutex_exit(&p->p_lock);
216 			return (NULL);
217 		}
218 	}
219 
220 	/*
221 	 * Create a thread, initializing the stack pointer
222 	 */
223 	t = thread_create(lwpdata, stksize, NULL, NULL, 0, p, TS_STOPPED, pri);
224 
225 	/*
226 	 * If a non-NULL stack base is passed in, thread_create() assumes
227 	 * that the stack might be statically allocated (as opposed to being
228 	 * allocated from segkp), and so it does not set t_swap.  Since
229 	 * the lwpdata was allocated from segkp, we must set t_swap to point
230 	 * to it ourselves.
231 	 *
232 	 * This would be less confusing if t_swap had a better name; it really
233 	 * indicates that the stack is allocated from segkp, regardless of
234 	 * whether or not it is swappable.
235 	 */
236 	if (lwpdata != NULL) {
237 		ASSERT(!CLASS_KERNEL(cid));
238 		ASSERT(t->t_swap == NULL);
239 		t->t_swap = lwpdata;	/* Start of page-able data */
240 	}
241 
242 	/*
243 	 * If the stack and lwp can be reused, mark the thread as such.
244 	 * When we get to reapq_add() from resume_from_zombie(), these
245 	 * threads will go onto lwp_deathrow instead of thread_deathrow.
246 	 */
247 	if (!CLASS_KERNEL(cid) && stksize == lwp_default_stksize)
248 		t->t_flag |= T_LWPREUSE;
249 
250 	if (lwp == NULL)
251 		lwp = kmem_cache_alloc(lwp_cache, KM_SLEEP);
252 	bzero(lwp, sizeof (*lwp));
253 	t->t_lwp = lwp;
254 
255 	t->t_hold = *smask;
256 	lwp->lwp_thread = t;
257 	lwp->lwp_procp = p;
258 	lwp->lwp_sigaltstack.ss_flags = SS_DISABLE;
259 	if (curlwp != NULL && curlwp->lwp_childstksz != 0)
260 		lwp->lwp_childstksz = curlwp->lwp_childstksz;
261 
262 	t->t_stk = lwp_stk_init(lwp, t->t_stk);
263 	thread_load(t, proc, arg, len);
264 
265 	/*
266 	 * Allocate the SIGPROF buffer if ITIMER_REALPROF is in effect.
267 	 */
268 	if (p->p_rprof_cyclic != CYCLIC_NONE)
269 		t->t_rprof = kmem_zalloc(sizeof (struct rprof), KM_SLEEP);
270 
271 	if (cid != NOCLASS)
272 		(void) CL_ALLOC(&bufp, cid, KM_SLEEP);
273 
274 	/*
275 	 * Allocate an lwp directory entry for the new lwp.
276 	 */
277 	lep = kmem_zalloc(sizeof (*lep), KM_SLEEP);
278 
279 	mutex_enter(&p->p_lock);
280 grow:
281 	/*
282 	 * Grow the lwp (thread) directory and lwpid hash table if necessary.
283 	 * A note on the growth algorithm:
284 	 *	The new lwp directory size is computed as:
285 	 *		new = 2 * old + 2
286 	 *	Starting with an initial size of 2 (see exec_common()),
287 	 *	this yields numbers that are a power of two minus 2:
288 	 *		2, 6, 14, 30, 62, 126, 254, 510, 1022, ...
289 	 *	The size of the lwpid hash table must be a power of two
290 	 *	and must be commensurate in size with the lwp directory
291 	 *	so that hash bucket chains remain short.  Therefore,
292 	 *	the lwpid hash table size is computed as:
293 	 *		hashsz = (dirsz + 2) / 2
294 	 *	which leads to these hash table sizes corresponding to
295 	 *	the above directory sizes:
296 	 *		2, 4, 8, 16, 32, 64, 128, 256, 512, ...
297 	 * A note on growing the hash table:
298 	 *	For performance reasons, code in lwp_unpark() does not
299 	 *	acquire curproc->p_lock when searching the hash table.
300 	 *	Rather, it calls lwp_hash_lookup_and_lock() which
301 	 *	acquires only the individual hash bucket lock, taking
302 	 *	care to deal with reallocation of the hash table
303 	 *	during the time it takes to acquire the lock.
304 	 *
305 	 *	This is sufficient to protect the integrity of the
306 	 *	hash table, but it requires us to acquire all of the
307 	 *	old hash bucket locks before growing the hash table
308 	 *	and to release them afterwards.  It also requires us
309 	 *	not to free the old hash table because some thread
310 	 *	in lwp_hash_lookup_and_lock() might still be trying
311 	 *	to acquire the old bucket lock.
312 	 *
313 	 *	So we adopt the tactic of keeping all of the retired
314 	 *	hash tables on a linked list, so they can be safely
315 	 *	freed when the process exits or execs.
316 	 *
317 	 *	Because the hash table grows in powers of two, the
318 	 *	total size of all of the hash tables will be slightly
319 	 *	less than twice the size of the largest hash table.
320 	 */
321 	while (p->p_lwpfree == NULL) {
322 		uint_t dirsz = p->p_lwpdir_sz;
323 		lwpdir_t *new_dir;
324 		uint_t new_dirsz;
325 		lwpdir_t *ldp;
326 		tidhash_t *new_hash;
327 		uint_t new_hashsz;
328 
329 		mutex_exit(&p->p_lock);
330 
331 		/*
332 		 * Prepare to remember the old p_tidhash for later
333 		 * kmem_free()ing when the process exits or execs.
334 		 */
335 		if (ret_tidhash == NULL)
336 			ret_tidhash = kmem_zalloc(sizeof (ret_tidhash_t),
337 			    KM_SLEEP);
338 		if (old_dir != NULL)
339 			kmem_free(old_dir, old_dirsz * sizeof (*old_dir));
340 		if (old_hash != NULL)
341 			kmem_free(old_hash, old_hashsz * sizeof (*old_hash));
342 
343 		new_dirsz = 2 * dirsz + 2;
344 		new_dir = kmem_zalloc(new_dirsz * sizeof (lwpdir_t), KM_SLEEP);
345 		for (ldp = new_dir, i = 1; i < new_dirsz; i++, ldp++)
346 			ldp->ld_next = ldp + 1;
347 		new_hashsz = (new_dirsz + 2) / 2;
348 		new_hash = kmem_zalloc(new_hashsz * sizeof (tidhash_t),
349 		    KM_SLEEP);
350 
351 		mutex_enter(&p->p_lock);
352 		if (p == curproc)
353 			prbarrier(p);
354 
355 		if (dirsz != p->p_lwpdir_sz || p->p_lwpfree != NULL) {
356 			/*
357 			 * Someone else beat us to it or some lwp exited.
358 			 * Set up to free our memory and take a lap.
359 			 */
360 			old_dir = new_dir;
361 			old_dirsz = new_dirsz;
362 			old_hash = new_hash;
363 			old_hashsz = new_hashsz;
364 		} else {
365 			/*
366 			 * For the benefit of lwp_hash_lookup_and_lock(),
367 			 * called from lwp_unpark(), which searches the
368 			 * tid hash table without acquiring p->p_lock,
369 			 * we must acquire all of the tid hash table
370 			 * locks before replacing p->p_tidhash.
371 			 */
372 			old_hash = p->p_tidhash;
373 			old_hashsz = p->p_tidhash_sz;
374 			for (i = 0; i < old_hashsz; i++) {
375 				mutex_enter(&old_hash[i].th_lock);
376 				mutex_enter(&new_hash[i].th_lock);
377 			}
378 
379 			/*
380 			 * We simply hash in all of the old directory entries.
381 			 * This works because the old directory has no empty
382 			 * slots and the new hash table starts out empty.
383 			 * This reproduces the original directory ordering
384 			 * (required for /proc directory semantics).
385 			 */
386 			old_dir = p->p_lwpdir;
387 			old_dirsz = p->p_lwpdir_sz;
388 			p->p_lwpdir = new_dir;
389 			p->p_lwpfree = new_dir;
390 			p->p_lwpdir_sz = new_dirsz;
391 			for (ldp = old_dir, i = 0; i < old_dirsz; i++, ldp++)
392 				lwp_hash_in(p, ldp->ld_entry,
393 				    new_hash, new_hashsz, 0);
394 
395 			/*
396 			 * Remember the old hash table along with all
397 			 * of the previously-remembered hash tables.
398 			 * We will free them at process exit or exec.
399 			 */
400 			ret_tidhash->rth_tidhash = old_hash;
401 			ret_tidhash->rth_tidhash_sz = old_hashsz;
402 			ret_tidhash->rth_next = p->p_ret_tidhash;
403 			p->p_ret_tidhash = ret_tidhash;
404 
405 			/*
406 			 * Now establish the new tid hash table.
407 			 * As soon as we assign p->p_tidhash,
408 			 * code in lwp_unpark() can start using it.
409 			 */
410 			membar_producer();
411 			p->p_tidhash = new_hash;
412 
413 			/*
414 			 * It is necessary that p_tidhash reach global
415 			 * visibility before p_tidhash_sz.  Otherwise,
416 			 * code in lwp_hash_lookup_and_lock() could
417 			 * index into the old p_tidhash using the new
418 			 * p_tidhash_sz and thereby access invalid data.
419 			 */
420 			membar_producer();
421 			p->p_tidhash_sz = new_hashsz;
422 
423 			/*
424 			 * Release the locks; allow lwp_unpark() to carry on.
425 			 */
426 			for (i = 0; i < old_hashsz; i++) {
427 				mutex_exit(&old_hash[i].th_lock);
428 				mutex_exit(&new_hash[i].th_lock);
429 			}
430 
431 			/*
432 			 * Avoid freeing these objects below.
433 			 */
434 			ret_tidhash = NULL;
435 			old_hash = NULL;
436 			old_hashsz = 0;
437 		}
438 	}
439 
440 	/*
441 	 * Block the process against /proc while we manipulate p->p_tlist,
442 	 * unless lwp_create() was called by /proc for the PCAGENT operation.
443 	 * We want to do this early enough so that we don't drop p->p_lock
444 	 * until the thread is put on the p->p_tlist.
445 	 */
446 	if (p == curproc) {
447 		prbarrier(p);
448 		/*
449 		 * If the current lwp has been requested to stop, do so now.
450 		 * Otherwise we have a race condition between /proc attempting
451 		 * to stop the process and this thread creating a new lwp
452 		 * that was not seen when the /proc PCSTOP request was issued.
453 		 * We rely on stop() to call prbarrier(p) before returning.
454 		 */
455 		while ((curthread->t_proc_flag & TP_PRSTOP) &&
456 		    !ttolwp(curthread)->lwp_nostop) {
457 			/*
458 			 * We called pool_barrier_enter() before calling
459 			 * here to lwp_create(). We have to call
460 			 * pool_barrier_exit() before stopping.
461 			 */
462 			pool_barrier_exit();
463 			prbarrier(p);
464 			stop(PR_REQUESTED, 0);
465 			/*
466 			 * And we have to repeat the call to
467 			 * pool_barrier_enter after stopping.
468 			 */
469 			pool_barrier_enter();
470 			prbarrier(p);
471 		}
472 
473 		/*
474 		 * If process is exiting, there could be a race between
475 		 * the agent lwp creation and the new lwp currently being
476 		 * created. So to prevent this race lwp creation is failed
477 		 * if the process is exiting.
478 		 */
479 		if (p->p_flag & (SEXITLWPS|SKILLED)) {
480 			err = 1;
481 			goto error;
482 		}
483 
484 		/*
485 		 * Since we might have dropped p->p_lock, the
486 		 * lwp directory free list might have changed.
487 		 */
488 		if (p->p_lwpfree == NULL)
489 			goto grow;
490 	}
491 
492 	kpreempt_disable();	/* can't grab cpu_lock here */
493 
494 	/*
495 	 * Inherit processor and processor set bindings from curthread.
496 	 *
497 	 * For kernel LWPs, we do not inherit processor set bindings at
498 	 * process creation time (i.e. when p != curproc).  After the
499 	 * kernel process is created, any subsequent LWPs must be created
500 	 * by threads in the kernel process, at which point we *will*
501 	 * inherit processor set bindings.
502 	 */
503 	if (CLASS_KERNEL(cid) && p != curproc) {
504 		t->t_bind_cpu = binding = PBIND_NONE;
505 		t->t_cpupart = oldpart = &cp_default;
506 		t->t_bind_pset = PS_NONE;
507 		t->t_bindflag = (uchar_t)default_binding_mode;
508 	} else {
509 		binding = curthread->t_bind_cpu;
510 		t->t_bind_cpu = binding;
511 		oldpart = t->t_cpupart;
512 		t->t_cpupart = curthread->t_cpupart;
513 		t->t_bind_pset = curthread->t_bind_pset;
514 		t->t_bindflag = curthread->t_bindflag |
515 		    (uchar_t)default_binding_mode;
516 	}
517 
518 	/*
519 	 * thread_create() initializes this thread's home lgroup to the root.
520 	 * Choose a more suitable lgroup, since this thread is associated
521 	 * with an lwp.
522 	 */
523 	ASSERT(oldpart != NULL);
524 	if (binding != PBIND_NONE && t->t_affinitycnt == 0) {
525 		t->t_bound_cpu = cpu[binding];
526 		if (t->t_lpl != t->t_bound_cpu->cpu_lpl)
527 			lgrp_move_thread(t, t->t_bound_cpu->cpu_lpl, 1);
528 	} else if (CLASS_KERNEL(cid)) {
529 		/*
530 		 * Kernel threads are always in the root lgrp.
531 		 */
532 		lgrp_move_thread(t,
533 		    &t->t_cpupart->cp_lgrploads[LGRP_ROOTID], 1);
534 	} else {
535 		lgrp_move_thread(t, lgrp_choose(t, t->t_cpupart), 1);
536 	}
537 
538 	kpreempt_enable();
539 
540 	/*
541 	 * make sure lpl points to our own partition
542 	 */
543 	ASSERT(t->t_lpl >= t->t_cpupart->cp_lgrploads);
544 	ASSERT(t->t_lpl < t->t_cpupart->cp_lgrploads +
545 	    t->t_cpupart->cp_nlgrploads);
546 
547 	/*
548 	 * It is safe to point the thread to the new project without holding it
549 	 * since we're holding the target process' p_lock here and therefore
550 	 * we're guaranteed that it will not move to another project.
551 	 */
552 	newkpj = p->p_task->tk_proj;
553 	oldkpj = ttoproj(t);
554 	if (newkpj != oldkpj) {
555 		t->t_proj = newkpj;
556 		(void) project_hold(newkpj);
557 		project_rele(oldkpj);
558 	}
559 
560 	if (cid != NOCLASS) {
561 		/*
562 		 * If the lwp is being created in the current process
563 		 * and matches the current thread's scheduling class,
564 		 * we should propagate the current thread's scheduling
565 		 * parameters by calling CL_FORK.  Otherwise just use
566 		 * the defaults by calling CL_ENTERCLASS.
567 		 */
568 		if (p != curproc || curthread->t_cid != cid) {
569 			err = CL_ENTERCLASS(t, cid, NULL, NULL, bufp);
570 			t->t_pri = pri;	/* CL_ENTERCLASS may have changed it */
571 			/*
572 			 * We don't call schedctl_set_cidpri(t) here
573 			 * because the schedctl data is not yet set
574 			 * up for the newly-created lwp.
575 			 */
576 		} else {
577 			t->t_clfuncs = &(sclass[cid].cl_funcs->thread);
578 			err = CL_FORK(curthread, t, bufp);
579 			t->t_cid = cid;
580 		}
581 		if (err)
582 			goto error;
583 		else
584 			bufp = NULL;
585 	}
586 
587 	/*
588 	 * If we were given an lwpid then use it, else allocate one.
589 	 */
590 	if (lwpid != 0)
591 		t->t_tid = lwpid;
592 	else {
593 		/*
594 		 * lwp/thread id 0 is never valid; reserved for special checks.
595 		 * lwp/thread id 1 is reserved for the main thread.
596 		 * Start again at 2 when INT_MAX has been reached
597 		 * (id_t is a signed 32-bit integer).
598 		 */
599 		id_t prev_id = p->p_lwpid;	/* last allocated tid */
600 
601 		do {			/* avoid lwpid duplication */
602 			if (p->p_lwpid == INT_MAX) {
603 				p->p_flag |= SLWPWRAP;
604 				p->p_lwpid = 1;
605 			}
606 			if ((t->t_tid = ++p->p_lwpid) == prev_id) {
607 				/*
608 				 * All lwpids are allocated; fail the request.
609 				 */
610 				err = 1;
611 				goto error;
612 			}
613 			/*
614 			 * We only need to worry about colliding with an id
615 			 * that's already in use if this process has
616 			 * cycled through all available lwp ids.
617 			 */
618 			if ((p->p_flag & SLWPWRAP) == 0)
619 				break;
620 		} while (lwp_hash_lookup(p, t->t_tid) != NULL);
621 	}
622 
623 	/*
624 	 * If this is a branded process, let the brand do any necessary lwp
625 	 * initialization.
626 	 */
627 	if (PROC_IS_BRANDED(p)) {
628 		if (BROP(p)->b_initlwp(lwp)) {
629 			err = 1;
630 			goto error;
631 		}
632 		branded = 1;
633 	}
634 
635 	if (t->t_tid == 1) {
636 		kpreempt_disable();
637 		ASSERT(t->t_lpl != NULL);
638 		p->p_t1_lgrpid = t->t_lpl->lpl_lgrpid;
639 		kpreempt_enable();
640 		if (p->p_tr_lgrpid != LGRP_NONE &&
641 		    p->p_tr_lgrpid != p->p_t1_lgrpid) {
642 			lgrp_update_trthr_migrations(1);
643 		}
644 	}
645 
646 	p->p_lwpcnt++;
647 	t->t_waitfor = -1;
648 
649 	/*
650 	 * Turn microstate accounting on for thread if on for process.
651 	 */
652 	if (p->p_flag & SMSACCT)
653 		t->t_proc_flag |= TP_MSACCT;
654 
655 	/*
656 	 * If the process has watchpoints, mark the new thread as such.
657 	 */
658 	if (pr_watch_active(p))
659 		watch_enable(t);
660 
661 	/*
662 	 * The lwp is being created in the stopped state.
663 	 * We set all the necessary flags to indicate that fact here.
664 	 * We omit the TS_CREATE flag from t_schedflag so that the lwp
665 	 * cannot be set running until the caller is finished with it,
666 	 * even if lwp_continue() is called on it after we drop p->p_lock.
667 	 * When the caller is finished with the newly-created lwp,
668 	 * the caller must call lwp_create_done() to allow the lwp
669 	 * to be set running.  If the TP_HOLDLWP is left set, the
670 	 * lwp will suspend itself after reaching system call exit.
671 	 */
672 	init_mstate(t, LMS_STOPPED);
673 	t->t_proc_flag |= TP_HOLDLWP;
674 	t->t_schedflag |= (TS_ALLSTART & ~(TS_CSTART | TS_CREATE));
675 	t->t_whystop = PR_SUSPENDED;
676 	t->t_whatstop = SUSPEND_NORMAL;
677 	t->t_sig_check = 1;	/* ensure that TP_HOLDLWP is honored */
678 
679 	/*
680 	 * Set system call processing flags in case tracing or profiling
681 	 * is set.  The first system call will evaluate these and turn
682 	 * them off if they aren't needed.
683 	 */
684 	t->t_pre_sys = 1;
685 	t->t_post_sys = 1;
686 
687 	/*
688 	 * Insert the new thread into the list of all threads.
689 	 */
690 	if ((tx = p->p_tlist) == NULL) {
691 		t->t_back = t;
692 		t->t_forw = t;
693 		p->p_tlist = t;
694 	} else {
695 		t->t_forw = tx;
696 		t->t_back = tx->t_back;
697 		tx->t_back->t_forw = t;
698 		tx->t_back = t;
699 	}
700 
701 	/*
702 	 * Insert the new lwp into an lwp directory slot position
703 	 * and into the lwpid hash table.
704 	 */
705 	lep->le_thread = t;
706 	lep->le_lwpid = t->t_tid;
707 	lep->le_start = t->t_start;
708 	lwp_hash_in(p, lep, p->p_tidhash, p->p_tidhash_sz, 1);
709 
710 	if (state == TS_RUN) {
711 		/*
712 		 * We set the new lwp running immediately.
713 		 */
714 		t->t_proc_flag &= ~TP_HOLDLWP;
715 		lwp_create_done(t);
716 	}
717 
718 error:
719 	if (err) {
720 		if (CLASS_KERNEL(cid)) {
721 			/*
722 			 * This should only happen if a system process runs
723 			 * out of lwpids, which shouldn't occur.
724 			 */
725 			panic("Failed to create a system LWP");
726 		}
727 		/*
728 		 * We have failed to create an lwp, so decrement the number
729 		 * of lwps in the task and let the lgroup load averages know
730 		 * that this thread isn't going to show up.
731 		 */
732 		kpreempt_disable();
733 		lgrp_move_thread(t, NULL, 1);
734 		kpreempt_enable();
735 
736 		ASSERT(MUTEX_HELD(&p->p_lock));
737 		mutex_enter(&p->p_zone->zone_nlwps_lock);
738 		p->p_task->tk_nlwps--;
739 		p->p_task->tk_proj->kpj_nlwps--;
740 		p->p_zone->zone_nlwps--;
741 		mutex_exit(&p->p_zone->zone_nlwps_lock);
742 		if (cid != NOCLASS && bufp != NULL)
743 			CL_FREE(cid, bufp);
744 
745 		if (branded)
746 			BROP(p)->b_freelwp(lwp);
747 
748 		mutex_exit(&p->p_lock);
749 		t->t_state = TS_FREE;
750 		thread_rele(t);
751 
752 		/*
753 		 * We need to remove t from the list of all threads
754 		 * because thread_exit()/lwp_exit() isn't called on t.
755 		 */
756 		mutex_enter(&pidlock);
757 		ASSERT(t != t->t_next);		/* t0 never exits */
758 		t->t_next->t_prev = t->t_prev;
759 		t->t_prev->t_next = t->t_next;
760 		mutex_exit(&pidlock);
761 
762 		thread_free(t);
763 		kmem_free(lep, sizeof (*lep));
764 		lwp = NULL;
765 	} else {
766 		mutex_exit(&p->p_lock);
767 	}
768 
769 	if (old_dir != NULL)
770 		kmem_free(old_dir, old_dirsz * sizeof (*old_dir));
771 	if (old_hash != NULL)
772 		kmem_free(old_hash, old_hashsz * sizeof (*old_hash));
773 	if (ret_tidhash != NULL)
774 		kmem_free(ret_tidhash, sizeof (ret_tidhash_t));
775 
776 	DTRACE_PROC1(lwp__create, kthread_t *, t);
777 	return (lwp);
778 }
779 
780 /*
781  * lwp_create_done() is called by the caller of lwp_create() to set the
782  * newly-created lwp running after the caller has finished manipulating it.
783  */
784 void
785 lwp_create_done(kthread_t *t)
786 {
787 	proc_t *p = ttoproc(t);
788 
789 	ASSERT(MUTEX_HELD(&p->p_lock));
790 
791 	/*
792 	 * We set the TS_CREATE and TS_CSTART flags and call setrun_locked().
793 	 * (The absence of the TS_CREATE flag prevents the lwp from running
794 	 * until we are finished with it, even if lwp_continue() is called on
795 	 * it by some other lwp in the process or elsewhere in the kernel.)
796 	 */
797 	thread_lock(t);
798 	ASSERT(t->t_state == TS_STOPPED && !(t->t_schedflag & TS_CREATE));
799 	/*
800 	 * If TS_CSTART is set, lwp_continue(t) has been called and
801 	 * has already incremented p_lwprcnt; avoid doing this twice.
802 	 */
803 	if (!(t->t_schedflag & TS_CSTART))
804 		p->p_lwprcnt++;
805 	t->t_schedflag |= (TS_CSTART | TS_CREATE);
806 	setrun_locked(t);
807 	thread_unlock(t);
808 }
809 
810 /*
811  * Copy an LWP's active templates, and clear the latest contracts.
812  */
813 void
814 lwp_ctmpl_copy(klwp_t *dst, klwp_t *src)
815 {
816 	int i;
817 
818 	for (i = 0; i < ct_ntypes; i++) {
819 		dst->lwp_ct_active[i] = ctmpl_dup(src->lwp_ct_active[i]);
820 		dst->lwp_ct_latest[i] = NULL;
821 	}
822 }
823 
824 /*
825  * Clear an LWP's contract template state.
826  */
827 void
828 lwp_ctmpl_clear(klwp_t *lwp)
829 {
830 	ct_template_t *tmpl;
831 	int i;
832 
833 	for (i = 0; i < ct_ntypes; i++) {
834 		if ((tmpl = lwp->lwp_ct_active[i]) != NULL) {
835 			ctmpl_free(tmpl);
836 			lwp->lwp_ct_active[i] = NULL;
837 		}
838 
839 		if (lwp->lwp_ct_latest[i] != NULL) {
840 			contract_rele(lwp->lwp_ct_latest[i]);
841 			lwp->lwp_ct_latest[i] = NULL;
842 		}
843 	}
844 }
845 
846 /*
847  * Individual lwp exit.
848  * If this is the last lwp, exit the whole process.
849  */
850 void
851 lwp_exit(void)
852 {
853 	kthread_t *t = curthread;
854 	klwp_t *lwp = ttolwp(t);
855 	proc_t *p = ttoproc(t);
856 
857 	ASSERT(MUTEX_HELD(&p->p_lock));
858 
859 	mutex_exit(&p->p_lock);
860 
861 #if defined(__sparc)
862 	/*
863 	 * Ensure that the user stack is fully abandoned..
864 	 */
865 	trash_user_windows();
866 #endif
867 
868 	tsd_exit();			/* free thread specific data */
869 
870 	kcpc_passivate();		/* Clean up performance counter state */
871 
872 	pollcleanup();
873 
874 	if (t->t_door)
875 		door_slam();
876 
877 	if (t->t_schedctl != NULL)
878 		schedctl_lwp_cleanup(t);
879 
880 	if (t->t_upimutex != NULL)
881 		upimutex_cleanup();
882 
883 	/*
884 	 * Perform any brand specific exit processing, then release any
885 	 * brand data associated with the lwp
886 	 */
887 	if (PROC_IS_BRANDED(p))
888 		BROP(p)->b_lwpexit(lwp);
889 
890 	lwp_pcb_exit();
891 
892 	mutex_enter(&p->p_lock);
893 	lwp_cleanup();
894 
895 	/*
896 	 * When this process is dumping core, its lwps are held here
897 	 * until the core dump is finished. Then exitlwps() is called
898 	 * again to release these lwps so that they can finish exiting.
899 	 */
900 	if (p->p_flag & SCOREDUMP)
901 		stop(PR_SUSPENDED, SUSPEND_NORMAL);
902 
903 	/*
904 	 * Block the process against /proc now that we have really acquired
905 	 * p->p_lock (to decrement p_lwpcnt and manipulate p_tlist at least).
906 	 */
907 	prbarrier(p);
908 
909 	/*
910 	 * Call proc_exit() if this is the last non-daemon lwp in the process.
911 	 */
912 	if (!(t->t_proc_flag & TP_DAEMON) &&
913 	    p->p_lwpcnt == p->p_lwpdaemon + 1) {
914 		mutex_exit(&p->p_lock);
915 		if (proc_exit(CLD_EXITED, 0) == 0) {
916 			/* Restarting init. */
917 			return;
918 		}
919 
920 		/*
921 		 * proc_exit() returns a non-zero value when some other
922 		 * lwp got there first.  We just have to continue in
923 		 * lwp_exit().
924 		 */
925 		mutex_enter(&p->p_lock);
926 		ASSERT(curproc->p_flag & SEXITLWPS);
927 		prbarrier(p);
928 	}
929 
930 	DTRACE_PROC(lwp__exit);
931 
932 	/*
933 	 * If the lwp is a detached lwp or if the process is exiting,
934 	 * remove (lwp_hash_out()) the lwp from the lwp directory.
935 	 * Otherwise null out the lwp's le_thread pointer in the lwp
936 	 * directory so that other threads will see it as a zombie lwp.
937 	 */
938 	prlwpexit(t);		/* notify /proc */
939 	if (!(t->t_proc_flag & TP_TWAIT) || (p->p_flag & SEXITLWPS))
940 		lwp_hash_out(p, t->t_tid);
941 	else {
942 		ASSERT(!(t->t_proc_flag & TP_DAEMON));
943 		p->p_lwpdir[t->t_dslot].ld_entry->le_thread = NULL;
944 		p->p_zombcnt++;
945 		cv_broadcast(&p->p_lwpexit);
946 	}
947 	if (t->t_proc_flag & TP_DAEMON) {
948 		p->p_lwpdaemon--;
949 		t->t_proc_flag &= ~TP_DAEMON;
950 	}
951 	t->t_proc_flag &= ~TP_TWAIT;
952 
953 	/*
954 	 * Maintain accurate lwp count for task.max-lwps resource control.
955 	 */
956 	mutex_enter(&p->p_zone->zone_nlwps_lock);
957 	p->p_task->tk_nlwps--;
958 	p->p_task->tk_proj->kpj_nlwps--;
959 	p->p_zone->zone_nlwps--;
960 	mutex_exit(&p->p_zone->zone_nlwps_lock);
961 
962 	CL_EXIT(t);		/* tell the scheduler that t is exiting */
963 	ASSERT(p->p_lwpcnt != 0);
964 	p->p_lwpcnt--;
965 
966 	/*
967 	 * If all remaining non-daemon lwps are waiting in lwp_wait(),
968 	 * wake them up so someone can return EDEADLK.
969 	 * (See the block comment preceeding lwp_wait().)
970 	 */
971 	if (p->p_lwpcnt == p->p_lwpdaemon + (p->p_lwpwait - p->p_lwpdwait))
972 		cv_broadcast(&p->p_lwpexit);
973 
974 	t->t_proc_flag |= TP_LWPEXIT;
975 	term_mstate(t);
976 
977 #ifndef NPROBE
978 	/* Kernel probe */
979 	if (t->t_tnf_tpdp)
980 		tnf_thread_exit();
981 #endif /* NPROBE */
982 
983 	t->t_forw->t_back = t->t_back;
984 	t->t_back->t_forw = t->t_forw;
985 	if (t == p->p_tlist)
986 		p->p_tlist = t->t_forw;
987 
988 	/*
989 	 * Clean up the signal state.
990 	 */
991 	if (t->t_sigqueue != NULL)
992 		sigdelq(p, t, 0);
993 	if (lwp->lwp_curinfo != NULL) {
994 		siginfofree(lwp->lwp_curinfo);
995 		lwp->lwp_curinfo = NULL;
996 	}
997 
998 	thread_rele(t);
999 
1000 	/*
1001 	 * Terminated lwps are associated with process zero and are put onto
1002 	 * death-row by resume().  Avoid preemption after resetting t->t_procp.
1003 	 */
1004 	t->t_preempt++;
1005 
1006 	if (t->t_ctx != NULL)
1007 		exitctx(t);
1008 	if (p->p_pctx != NULL)
1009 		exitpctx(p);
1010 
1011 	t->t_procp = &p0;
1012 
1013 	/*
1014 	 * Notify the HAT about the change of address space
1015 	 */
1016 	hat_thread_exit(t);
1017 	/*
1018 	 * When this is the last running lwp in this process and some lwp is
1019 	 * waiting for this condition to become true, or this thread was being
1020 	 * suspended, then the waiting lwp is awakened.
1021 	 *
1022 	 * Also, if the process is exiting, we may have a thread waiting in
1023 	 * exitlwps() that needs to be notified.
1024 	 */
1025 	if (--p->p_lwprcnt == 0 || (t->t_proc_flag & TP_HOLDLWP) ||
1026 	    (p->p_flag & SEXITLWPS))
1027 		cv_broadcast(&p->p_holdlwps);
1028 
1029 	/*
1030 	 * Need to drop p_lock so we can reacquire pidlock.
1031 	 */
1032 	mutex_exit(&p->p_lock);
1033 	mutex_enter(&pidlock);
1034 
1035 	ASSERT(t != t->t_next);		/* t0 never exits */
1036 	t->t_next->t_prev = t->t_prev;
1037 	t->t_prev->t_next = t->t_next;
1038 	cv_broadcast(&t->t_joincv);	/* wake up anyone in thread_join */
1039 	mutex_exit(&pidlock);
1040 
1041 	t->t_state = TS_ZOMB;
1042 	swtch_from_zombie();
1043 	/* never returns */
1044 }
1045 
1046 
1047 /*
1048  * Cleanup function for an exiting lwp.
1049  * Called both from lwp_exit() and from proc_exit().
1050  * p->p_lock is repeatedly released and grabbed in this function.
1051  */
1052 void
1053 lwp_cleanup(void)
1054 {
1055 	kthread_t *t = curthread;
1056 	proc_t *p = ttoproc(t);
1057 
1058 	ASSERT(MUTEX_HELD(&p->p_lock));
1059 
1060 	/* untimeout any lwp-bound realtime timers */
1061 	if (p->p_itimer != NULL)
1062 		timer_lwpexit();
1063 
1064 	/*
1065 	 * If this is the /proc agent lwp that is exiting, readjust p_lwpid
1066 	 * so it appears that the agent never existed, and clear p_agenttp.
1067 	 */
1068 	if (t == p->p_agenttp) {
1069 		ASSERT(t->t_tid == p->p_lwpid);
1070 		p->p_lwpid--;
1071 		p->p_agenttp = NULL;
1072 	}
1073 
1074 	/*
1075 	 * Do lgroup bookkeeping to account for thread exiting.
1076 	 */
1077 	kpreempt_disable();
1078 	lgrp_move_thread(t, NULL, 1);
1079 	if (t->t_tid == 1) {
1080 		p->p_t1_lgrpid = LGRP_NONE;
1081 	}
1082 	kpreempt_enable();
1083 
1084 	lwp_ctmpl_clear(ttolwp(t));
1085 }
1086 
1087 int
1088 lwp_suspend(kthread_t *t)
1089 {
1090 	int tid;
1091 	proc_t *p = ttoproc(t);
1092 
1093 	ASSERT(MUTEX_HELD(&p->p_lock));
1094 
1095 	/*
1096 	 * Set the thread's TP_HOLDLWP flag so it will stop in holdlwp().
1097 	 * If an lwp is stopping itself, there is no need to wait.
1098 	 */
1099 top:
1100 	t->t_proc_flag |= TP_HOLDLWP;
1101 	if (t == curthread) {
1102 		t->t_sig_check = 1;
1103 	} else {
1104 		/*
1105 		 * Make sure the lwp stops promptly.
1106 		 */
1107 		thread_lock(t);
1108 		t->t_sig_check = 1;
1109 		/*
1110 		 * XXX Should use virtual stop like /proc does instead of
1111 		 * XXX waking the thread to get it to stop.
1112 		 */
1113 		if (ISWAKEABLE(t) || ISWAITING(t)) {
1114 			setrun_locked(t);
1115 		} else if (t->t_state == TS_ONPROC && t->t_cpu != CPU) {
1116 			poke_cpu(t->t_cpu->cpu_id);
1117 		}
1118 
1119 		tid = t->t_tid;	 /* remember thread ID */
1120 		/*
1121 		 * Wait for lwp to stop
1122 		 */
1123 		while (!SUSPENDED(t)) {
1124 			/*
1125 			 * Drop the thread lock before waiting and reacquire it
1126 			 * afterwards, so the thread can change its t_state
1127 			 * field.
1128 			 */
1129 			thread_unlock(t);
1130 
1131 			/*
1132 			 * Check if aborted by exitlwps().
1133 			 */
1134 			if (p->p_flag & SEXITLWPS)
1135 				lwp_exit();
1136 
1137 			/*
1138 			 * Cooperate with jobcontrol signals and /proc stopping
1139 			 * by calling cv_wait_sig() to wait for the target
1140 			 * lwp to stop.  Just using cv_wait() can lead to
1141 			 * deadlock because, if some other lwp has stopped
1142 			 * by either of these mechanisms, then p_lwprcnt will
1143 			 * never become zero if we do a cv_wait().
1144 			 */
1145 			if (!cv_wait_sig(&p->p_holdlwps, &p->p_lock))
1146 				return (EINTR);
1147 
1148 			/*
1149 			 * Check to see if thread died while we were
1150 			 * waiting for it to suspend.
1151 			 */
1152 			if (idtot(p, tid) == NULL)
1153 				return (ESRCH);
1154 
1155 			thread_lock(t);
1156 			/*
1157 			 * If the TP_HOLDLWP flag went away, lwp_continue()
1158 			 * or vfork() must have been called while we were
1159 			 * waiting, so start over again.
1160 			 */
1161 			if ((t->t_proc_flag & TP_HOLDLWP) == 0) {
1162 				thread_unlock(t);
1163 				goto top;
1164 			}
1165 		}
1166 		thread_unlock(t);
1167 	}
1168 	return (0);
1169 }
1170 
1171 /*
1172  * continue a lwp that's been stopped by lwp_suspend().
1173  */
1174 void
1175 lwp_continue(kthread_t *t)
1176 {
1177 	proc_t *p = ttoproc(t);
1178 	int was_suspended = t->t_proc_flag & TP_HOLDLWP;
1179 
1180 	ASSERT(MUTEX_HELD(&p->p_lock));
1181 
1182 	t->t_proc_flag &= ~TP_HOLDLWP;
1183 	thread_lock(t);
1184 	if (SUSPENDED(t) &&
1185 	    !(p->p_flag & (SHOLDFORK | SHOLDFORK1 | SHOLDWATCH))) {
1186 		p->p_lwprcnt++;
1187 		t->t_schedflag |= TS_CSTART;
1188 		setrun_locked(t);
1189 	}
1190 	thread_unlock(t);
1191 	/*
1192 	 * Wakeup anyone waiting for this thread to be suspended
1193 	 */
1194 	if (was_suspended)
1195 		cv_broadcast(&p->p_holdlwps);
1196 }
1197 
1198 /*
1199  * ********************************
1200  *  Miscellaneous lwp routines	  *
1201  * ********************************
1202  */
1203 /*
1204  * When a process is undergoing a forkall(), its p_flag is set to SHOLDFORK.
1205  * This will cause the process's lwps to stop at a hold point.  A hold
1206  * point is where a kernel thread has a flat stack.  This is at the
1207  * return from a system call and at the return from a user level trap.
1208  *
1209  * When a process is undergoing a fork1() or vfork(), its p_flag is set to
1210  * SHOLDFORK1.  This will cause the process's lwps to stop at a modified
1211  * hold point.  The lwps in the process are not being cloned, so they
1212  * are held at the usual hold points and also within issig_forreal().
1213  * This has the side-effect that their system calls do not return
1214  * showing EINTR.
1215  *
1216  * An lwp can also be held.  This is identified by the TP_HOLDLWP flag on
1217  * the thread.  The TP_HOLDLWP flag is set in lwp_suspend(), where the active
1218  * lwp is waiting for the target lwp to be stopped.
1219  */
1220 void
1221 holdlwp(void)
1222 {
1223 	proc_t *p = curproc;
1224 	kthread_t *t = curthread;
1225 
1226 	mutex_enter(&p->p_lock);
1227 	/*
1228 	 * Don't terminate immediately if the process is dumping core.
1229 	 * Once the process has dumped core, all lwps are terminated.
1230 	 */
1231 	if (!(p->p_flag & SCOREDUMP)) {
1232 		if ((p->p_flag & SEXITLWPS) || (t->t_proc_flag & TP_EXITLWP))
1233 			lwp_exit();
1234 	}
1235 	if (!(ISHOLD(p)) && !(p->p_flag & (SHOLDFORK1 | SHOLDWATCH))) {
1236 		mutex_exit(&p->p_lock);
1237 		return;
1238 	}
1239 	/*
1240 	 * stop() decrements p->p_lwprcnt and cv_signal()s &p->p_holdlwps
1241 	 * when p->p_lwprcnt becomes zero.
1242 	 */
1243 	stop(PR_SUSPENDED, SUSPEND_NORMAL);
1244 	if (p->p_flag & SEXITLWPS)
1245 		lwp_exit();
1246 	mutex_exit(&p->p_lock);
1247 }
1248 
1249 /*
1250  * Have all lwps within the process hold at a point where they are
1251  * cloneable (SHOLDFORK) or just safe w.r.t. fork1 (SHOLDFORK1).
1252  */
1253 int
1254 holdlwps(int holdflag)
1255 {
1256 	proc_t *p = curproc;
1257 
1258 	ASSERT(holdflag == SHOLDFORK || holdflag == SHOLDFORK1);
1259 	mutex_enter(&p->p_lock);
1260 	schedctl_finish_sigblock(curthread);
1261 again:
1262 	while (p->p_flag & (SEXITLWPS | SHOLDFORK | SHOLDFORK1 | SHOLDWATCH)) {
1263 		/*
1264 		 * If another lwp is doing a forkall() or proc_exit(), bail out.
1265 		 */
1266 		if (p->p_flag & (SEXITLWPS | SHOLDFORK)) {
1267 			mutex_exit(&p->p_lock);
1268 			return (0);
1269 		}
1270 		/*
1271 		 * Another lwp is doing a fork1() or is undergoing
1272 		 * watchpoint activity.  We hold here for it to complete.
1273 		 */
1274 		stop(PR_SUSPENDED, SUSPEND_NORMAL);
1275 	}
1276 	p->p_flag |= holdflag;
1277 	pokelwps(p);
1278 	--p->p_lwprcnt;
1279 	/*
1280 	 * Wait for the process to become quiescent (p->p_lwprcnt == 0).
1281 	 */
1282 	while (p->p_lwprcnt > 0) {
1283 		/*
1284 		 * Check if aborted by exitlwps().
1285 		 * Also check if SHOLDWATCH is set; it takes precedence.
1286 		 */
1287 		if (p->p_flag & (SEXITLWPS | SHOLDWATCH)) {
1288 			p->p_lwprcnt++;
1289 			p->p_flag &= ~holdflag;
1290 			cv_broadcast(&p->p_holdlwps);
1291 			goto again;
1292 		}
1293 		/*
1294 		 * Cooperate with jobcontrol signals and /proc stopping.
1295 		 * If some other lwp has stopped by either of these
1296 		 * mechanisms, then p_lwprcnt will never become zero
1297 		 * and the process will appear deadlocked unless we
1298 		 * stop here in sympathy with the other lwp before
1299 		 * doing the cv_wait() below.
1300 		 *
1301 		 * If the other lwp stops after we do the cv_wait(), it
1302 		 * will wake us up to loop around and do the sympathy stop.
1303 		 *
1304 		 * Since stop() drops p->p_lock, we must start from
1305 		 * the top again on returning from stop().
1306 		 */
1307 		if (p->p_stopsig | (curthread->t_proc_flag & TP_PRSTOP)) {
1308 			int whystop = p->p_stopsig? PR_JOBCONTROL :
1309 			    PR_REQUESTED;
1310 			p->p_lwprcnt++;
1311 			p->p_flag &= ~holdflag;
1312 			stop(whystop, p->p_stopsig);
1313 			goto again;
1314 		}
1315 		cv_wait(&p->p_holdlwps, &p->p_lock);
1316 	}
1317 	p->p_lwprcnt++;
1318 	p->p_flag &= ~holdflag;
1319 	mutex_exit(&p->p_lock);
1320 	return (1);
1321 }
1322 
1323 /*
1324  * See comments for holdwatch(), below.
1325  */
1326 static int
1327 holdcheck(int clearflags)
1328 {
1329 	proc_t *p = curproc;
1330 
1331 	/*
1332 	 * If we are trying to exit, that takes precedence over anything else.
1333 	 */
1334 	if (p->p_flag & SEXITLWPS) {
1335 		p->p_lwprcnt++;
1336 		p->p_flag &= ~clearflags;
1337 		lwp_exit();
1338 	}
1339 
1340 	/*
1341 	 * If another thread is calling fork1(), stop the current thread so the
1342 	 * other can complete.
1343 	 */
1344 	if (p->p_flag & SHOLDFORK1) {
1345 		p->p_lwprcnt++;
1346 		stop(PR_SUSPENDED, SUSPEND_NORMAL);
1347 		if (p->p_flag & SEXITLWPS) {
1348 			p->p_flag &= ~clearflags;
1349 			lwp_exit();
1350 		}
1351 		return (-1);
1352 	}
1353 
1354 	/*
1355 	 * If another thread is calling fork(), then indicate we are doing
1356 	 * watchpoint activity.  This will cause holdlwps() above to stop the
1357 	 * forking thread, at which point we can continue with watchpoint
1358 	 * activity.
1359 	 */
1360 	if (p->p_flag & SHOLDFORK) {
1361 		p->p_lwprcnt++;
1362 		while (p->p_flag & SHOLDFORK) {
1363 			p->p_flag |= SHOLDWATCH;
1364 			cv_broadcast(&p->p_holdlwps);
1365 			cv_wait(&p->p_holdlwps, &p->p_lock);
1366 			p->p_flag &= ~SHOLDWATCH;
1367 		}
1368 		return (-1);
1369 	}
1370 
1371 	return (0);
1372 }
1373 
1374 /*
1375  * Stop all lwps within the process, holding themselves in the kernel while the
1376  * active lwp undergoes watchpoint activity.  This is more complicated than
1377  * expected because stop() relies on calling holdwatch() in order to copyin data
1378  * from the user's address space.  A double barrier is used to prevent an
1379  * infinite loop.
1380  *
1381  * 	o The first thread into holdwatch() is the 'master' thread and does
1382  *        the following:
1383  *
1384  *              - Sets SHOLDWATCH on the current process
1385  *              - Sets TP_WATCHSTOP on the current thread
1386  *              - Waits for all threads to be either stopped or have
1387  *                TP_WATCHSTOP set.
1388  *              - Sets the SWATCHOK flag on the process
1389  *              - Unsets TP_WATCHSTOP
1390  *              - Waits for the other threads to completely stop
1391  *              - Unsets SWATCHOK
1392  *
1393  * 	o If SHOLDWATCH is already set when we enter this function, then another
1394  *        thread is already trying to stop this thread.  This 'slave' thread
1395  *        does the following:
1396  *
1397  *              - Sets TP_WATCHSTOP on the current thread
1398  *              - Waits for SWATCHOK flag to be set
1399  *              - Calls stop()
1400  *
1401  * 	o If SWATCHOK is set on the process, then this function immediately
1402  *        returns, as we must have been called via stop().
1403  *
1404  * In addition, there are other flags that take precedence over SHOLDWATCH:
1405  *
1406  * 	o If SEXITLWPS is set, exit immediately.
1407  *
1408  * 	o If SHOLDFORK1 is set, wait for fork1() to complete.
1409  *
1410  * 	o If SHOLDFORK is set, then watchpoint activity takes precedence In this
1411  *        case, set SHOLDWATCH, signalling the forking thread to stop first.
1412  *
1413  * 	o If the process is being stopped via /proc (TP_PRSTOP is set), then we
1414  *        stop the current thread.
1415  *
1416  * Returns 0 if all threads have been quiesced.  Returns non-zero if not all
1417  * threads were stopped, or the list of watched pages has changed.
1418  */
1419 int
1420 holdwatch(void)
1421 {
1422 	proc_t *p = curproc;
1423 	kthread_t *t = curthread;
1424 	int ret = 0;
1425 
1426 	mutex_enter(&p->p_lock);
1427 
1428 	p->p_lwprcnt--;
1429 
1430 	/*
1431 	 * Check for bail-out conditions as outlined above.
1432 	 */
1433 	if (holdcheck(0) != 0) {
1434 		mutex_exit(&p->p_lock);
1435 		return (-1);
1436 	}
1437 
1438 	if (!(p->p_flag & SHOLDWATCH)) {
1439 		/*
1440 		 * We are the master watchpoint thread.  Set SHOLDWATCH and poke
1441 		 * the other threads.
1442 		 */
1443 		p->p_flag |= SHOLDWATCH;
1444 		pokelwps(p);
1445 
1446 		/*
1447 		 * Wait for all threads to be stopped or have TP_WATCHSTOP set.
1448 		 */
1449 		while (pr_allstopped(p, 1) > 0) {
1450 			if (holdcheck(SHOLDWATCH) != 0) {
1451 				p->p_flag &= ~SHOLDWATCH;
1452 				mutex_exit(&p->p_lock);
1453 				return (-1);
1454 			}
1455 
1456 			cv_wait(&p->p_holdlwps, &p->p_lock);
1457 		}
1458 
1459 		/*
1460 		 * All threads are now stopped or in the process of stopping.
1461 		 * Set SWATCHOK and let them stop completely.
1462 		 */
1463 		p->p_flag |= SWATCHOK;
1464 		t->t_proc_flag &= ~TP_WATCHSTOP;
1465 		cv_broadcast(&p->p_holdlwps);
1466 
1467 		while (pr_allstopped(p, 0) > 0) {
1468 			/*
1469 			 * At first glance, it may appear that we don't need a
1470 			 * call to holdcheck() here.  But if the process gets a
1471 			 * SIGKILL signal, one of our stopped threads may have
1472 			 * been awakened and is waiting in exitlwps(), which
1473 			 * takes precedence over watchpoints.
1474 			 */
1475 			if (holdcheck(SHOLDWATCH | SWATCHOK) != 0) {
1476 				p->p_flag &= ~(SHOLDWATCH | SWATCHOK);
1477 				mutex_exit(&p->p_lock);
1478 				return (-1);
1479 			}
1480 
1481 			cv_wait(&p->p_holdlwps, &p->p_lock);
1482 		}
1483 
1484 		/*
1485 		 * All threads are now completely stopped.
1486 		 */
1487 		p->p_flag &= ~SWATCHOK;
1488 		p->p_flag &= ~SHOLDWATCH;
1489 		p->p_lwprcnt++;
1490 
1491 	} else if (!(p->p_flag & SWATCHOK)) {
1492 
1493 		/*
1494 		 * SHOLDWATCH is set, so another thread is trying to do
1495 		 * watchpoint activity.  Indicate this thread is stopping, and
1496 		 * wait for the OK from the master thread.
1497 		 */
1498 		t->t_proc_flag |= TP_WATCHSTOP;
1499 		cv_broadcast(&p->p_holdlwps);
1500 
1501 		while (!(p->p_flag & SWATCHOK)) {
1502 			if (holdcheck(0) != 0) {
1503 				t->t_proc_flag &= ~TP_WATCHSTOP;
1504 				mutex_exit(&p->p_lock);
1505 				return (-1);
1506 			}
1507 
1508 			cv_wait(&p->p_holdlwps, &p->p_lock);
1509 		}
1510 
1511 		/*
1512 		 * Once the master thread has given the OK, this thread can
1513 		 * actually call stop().
1514 		 */
1515 		t->t_proc_flag &= ~TP_WATCHSTOP;
1516 		p->p_lwprcnt++;
1517 
1518 		stop(PR_SUSPENDED, SUSPEND_NORMAL);
1519 
1520 		/*
1521 		 * It's not OK to do watchpoint activity, notify caller to
1522 		 * retry.
1523 		 */
1524 		ret = -1;
1525 
1526 	} else {
1527 
1528 		/*
1529 		 * The only way we can hit the case where SHOLDWATCH is set and
1530 		 * SWATCHOK is set is if we are triggering this from within a
1531 		 * stop() call.  Assert that this is the case.
1532 		 */
1533 
1534 		ASSERT(t->t_proc_flag & TP_STOPPING);
1535 		p->p_lwprcnt++;
1536 	}
1537 
1538 	mutex_exit(&p->p_lock);
1539 
1540 	return (ret);
1541 }
1542 
1543 /*
1544  * force all interruptible lwps to trap into the kernel.
1545  */
1546 void
1547 pokelwps(proc_t *p)
1548 {
1549 	kthread_t *t;
1550 
1551 	ASSERT(MUTEX_HELD(&p->p_lock));
1552 
1553 	t = p->p_tlist;
1554 	do {
1555 		if (t == curthread)
1556 			continue;
1557 		thread_lock(t);
1558 		aston(t);	/* make thread trap or do post_syscall */
1559 		if (ISWAKEABLE(t) || ISWAITING(t)) {
1560 			setrun_locked(t);
1561 		} else if (t->t_state == TS_STOPPED) {
1562 			/*
1563 			 * Ensure that proc_exit() is not blocked by lwps
1564 			 * that were stopped via jobcontrol or /proc.
1565 			 */
1566 			if (p->p_flag & SEXITLWPS) {
1567 				p->p_stopsig = 0;
1568 				t->t_schedflag |= (TS_XSTART | TS_PSTART);
1569 				setrun_locked(t);
1570 			}
1571 			/*
1572 			 * If we are holding lwps for a forkall(),
1573 			 * force lwps that have been suspended via
1574 			 * lwp_suspend() and are suspended inside
1575 			 * of a system call to proceed to their
1576 			 * holdlwp() points where they are clonable.
1577 			 */
1578 			if ((p->p_flag & SHOLDFORK) && SUSPENDED(t)) {
1579 				if ((t->t_schedflag & TS_CSTART) == 0) {
1580 					p->p_lwprcnt++;
1581 					t->t_schedflag |= TS_CSTART;
1582 					setrun_locked(t);
1583 				}
1584 			}
1585 		} else if (t->t_state == TS_ONPROC) {
1586 			if (t->t_cpu != CPU)
1587 				poke_cpu(t->t_cpu->cpu_id);
1588 		}
1589 		thread_unlock(t);
1590 	} while ((t = t->t_forw) != p->p_tlist);
1591 }
1592 
1593 /*
1594  * undo the effects of holdlwps() or holdwatch().
1595  */
1596 void
1597 continuelwps(proc_t *p)
1598 {
1599 	kthread_t *t;
1600 
1601 	/*
1602 	 * If this flag is set, then the original holdwatch() didn't actually
1603 	 * stop the process.  See comments for holdwatch().
1604 	 */
1605 	if (p->p_flag & SWATCHOK) {
1606 		ASSERT(curthread->t_proc_flag & TP_STOPPING);
1607 		return;
1608 	}
1609 
1610 	ASSERT(MUTEX_HELD(&p->p_lock));
1611 	ASSERT((p->p_flag & (SHOLDFORK | SHOLDFORK1 | SHOLDWATCH)) == 0);
1612 
1613 	t = p->p_tlist;
1614 	do {
1615 		thread_lock(t);		/* SUSPENDED looks at t_schedflag */
1616 		if (SUSPENDED(t) && !(t->t_proc_flag & TP_HOLDLWP)) {
1617 			p->p_lwprcnt++;
1618 			t->t_schedflag |= TS_CSTART;
1619 			setrun_locked(t);
1620 		}
1621 		thread_unlock(t);
1622 	} while ((t = t->t_forw) != p->p_tlist);
1623 }
1624 
1625 /*
1626  * Force all other LWPs in the current process other than the caller to exit,
1627  * and then cv_wait() on p_holdlwps for them to exit.  The exitlwps() function
1628  * is typically used in these situations:
1629  *
1630  *   (a) prior to an exec() system call
1631  *   (b) prior to dumping a core file
1632  *   (c) prior to a uadmin() shutdown
1633  *
1634  * If the 'coredump' flag is set, other LWPs are quiesced but not destroyed.
1635  * Multiple threads in the process can call this function at one time by
1636  * triggering execs or core dumps simultaneously, so the SEXITLWPS bit is used
1637  * to declare one particular thread the winner who gets to kill the others.
1638  * If a thread wins the exitlwps() dance, zero is returned; otherwise an
1639  * appropriate errno value is returned to caller for its system call to return.
1640  */
1641 int
1642 exitlwps(int coredump)
1643 {
1644 	proc_t *p = curproc;
1645 	int heldcnt;
1646 
1647 	if (curthread->t_door)
1648 		door_slam();
1649 	if (p->p_door_list)
1650 		door_revoke_all();
1651 	if (curthread->t_schedctl != NULL)
1652 		schedctl_lwp_cleanup(curthread);
1653 
1654 	/*
1655 	 * Ensure that before starting to wait for other lwps to exit,
1656 	 * cleanup all upimutexes held by curthread. Otherwise, some other
1657 	 * lwp could be waiting (uninterruptibly) for a upimutex held by
1658 	 * curthread, and the call to pokelwps() below would deadlock.
1659 	 * Even if a blocked upimutex_lock is made interruptible,
1660 	 * curthread's upimutexes need to be unlocked: do it here.
1661 	 */
1662 	if (curthread->t_upimutex != NULL)
1663 		upimutex_cleanup();
1664 
1665 	/*
1666 	 * Grab p_lock in order to check and set SEXITLWPS to declare a winner.
1667 	 * We must also block any further /proc access from this point forward.
1668 	 */
1669 	mutex_enter(&p->p_lock);
1670 	prbarrier(p);
1671 
1672 	if (p->p_flag & SEXITLWPS) {
1673 		mutex_exit(&p->p_lock);
1674 		aston(curthread);	/* force a trip through post_syscall */
1675 		return (set_errno(EINTR));
1676 	}
1677 
1678 	p->p_flag |= SEXITLWPS;
1679 	if (coredump)		/* tell other lwps to stop, not exit */
1680 		p->p_flag |= SCOREDUMP;
1681 
1682 	/*
1683 	 * Give precedence to exitlwps() if a holdlwps() is
1684 	 * in progress. The lwp doing the holdlwps() operation
1685 	 * is aborted when it is awakened.
1686 	 */
1687 	while (p->p_flag & (SHOLDFORK | SHOLDFORK1 | SHOLDWATCH)) {
1688 		cv_broadcast(&p->p_holdlwps);
1689 		cv_wait(&p->p_holdlwps, &p->p_lock);
1690 		prbarrier(p);
1691 	}
1692 	p->p_flag |= SHOLDFORK;
1693 	pokelwps(p);
1694 
1695 	/*
1696 	 * Wait for process to become quiescent.
1697 	 */
1698 	--p->p_lwprcnt;
1699 	while (p->p_lwprcnt > 0) {
1700 		cv_wait(&p->p_holdlwps, &p->p_lock);
1701 		prbarrier(p);
1702 	}
1703 	p->p_lwprcnt++;
1704 	ASSERT(p->p_lwprcnt == 1);
1705 
1706 	/*
1707 	 * The SCOREDUMP flag puts the process into a quiescent
1708 	 * state.  The process's lwps remain attached to this
1709 	 * process until exitlwps() is called again without the
1710 	 * 'coredump' flag set, then the lwps are terminated
1711 	 * and the process can exit.
1712 	 */
1713 	if (coredump) {
1714 		p->p_flag &= ~(SCOREDUMP | SHOLDFORK | SEXITLWPS);
1715 		goto out;
1716 	}
1717 
1718 	/*
1719 	 * Determine if there are any lwps left dangling in
1720 	 * the stopped state.  This happens when exitlwps()
1721 	 * aborts a holdlwps() operation.
1722 	 */
1723 	p->p_flag &= ~SHOLDFORK;
1724 	if ((heldcnt = p->p_lwpcnt) > 1) {
1725 		kthread_t *t;
1726 		for (t = curthread->t_forw; --heldcnt > 0; t = t->t_forw) {
1727 			t->t_proc_flag &= ~TP_TWAIT;
1728 			lwp_continue(t);
1729 		}
1730 	}
1731 
1732 	/*
1733 	 * Wait for all other lwps to exit.
1734 	 */
1735 	--p->p_lwprcnt;
1736 	while (p->p_lwpcnt > 1) {
1737 		cv_wait(&p->p_holdlwps, &p->p_lock);
1738 		prbarrier(p);
1739 	}
1740 	++p->p_lwprcnt;
1741 	ASSERT(p->p_lwpcnt == 1 && p->p_lwprcnt == 1);
1742 
1743 	p->p_flag &= ~SEXITLWPS;
1744 	curthread->t_proc_flag &= ~TP_TWAIT;
1745 
1746 out:
1747 	if (!coredump && p->p_zombcnt) {	/* cleanup the zombie lwps */
1748 		lwpdir_t *ldp;
1749 		lwpent_t *lep;
1750 		int i;
1751 
1752 		for (ldp = p->p_lwpdir, i = 0; i < p->p_lwpdir_sz; i++, ldp++) {
1753 			lep = ldp->ld_entry;
1754 			if (lep != NULL && lep->le_thread != curthread) {
1755 				ASSERT(lep->le_thread == NULL);
1756 				p->p_zombcnt--;
1757 				lwp_hash_out(p, lep->le_lwpid);
1758 			}
1759 		}
1760 		ASSERT(p->p_zombcnt == 0);
1761 	}
1762 
1763 	/*
1764 	 * If some other LWP in the process wanted us to suspend ourself,
1765 	 * then we will not do it.  The other LWP is now terminated and
1766 	 * no one will ever continue us again if we suspend ourself.
1767 	 */
1768 	curthread->t_proc_flag &= ~TP_HOLDLWP;
1769 	p->p_flag &= ~(SHOLDFORK | SHOLDFORK1 | SHOLDWATCH | SLWPWRAP);
1770 	mutex_exit(&p->p_lock);
1771 	return (0);
1772 }
1773 
1774 /*
1775  * duplicate a lwp.
1776  */
1777 klwp_t *
1778 forklwp(klwp_t *lwp, proc_t *cp, id_t lwpid)
1779 {
1780 	klwp_t *clwp;
1781 	void *tregs, *tfpu;
1782 	kthread_t *t = lwptot(lwp);
1783 	kthread_t *ct;
1784 	proc_t *p = lwptoproc(lwp);
1785 	int cid;
1786 	void *bufp;
1787 	void *brand_data;
1788 	int val;
1789 
1790 	ASSERT(p == curproc);
1791 	ASSERT(t == curthread || (SUSPENDED(t) && lwp->lwp_asleep == 0));
1792 
1793 #if defined(__sparc)
1794 	if (t == curthread)
1795 		(void) flush_user_windows_to_stack(NULL);
1796 #endif
1797 
1798 	if (t == curthread)
1799 		/* copy args out of registers first */
1800 		(void) save_syscall_args();
1801 
1802 	clwp = lwp_create(cp->p_lwpcnt == 0 ? lwp_rtt_initial : lwp_rtt,
1803 	    NULL, 0, cp, TS_STOPPED, t->t_pri, &t->t_hold, NOCLASS, lwpid);
1804 	if (clwp == NULL)
1805 		return (NULL);
1806 
1807 	/*
1808 	 * most of the parent's lwp can be copied to its duplicate,
1809 	 * except for the fields that are unique to each lwp, like
1810 	 * lwp_thread, lwp_procp, lwp_regs, and lwp_ap.
1811 	 */
1812 	ct = clwp->lwp_thread;
1813 	tregs = clwp->lwp_regs;
1814 	tfpu = clwp->lwp_fpu;
1815 	brand_data = clwp->lwp_brand;
1816 
1817 	/*
1818 	 * Copy parent lwp to child lwp.  Hold child's p_lock to prevent
1819 	 * mstate_aggr_state() from reading stale mstate entries copied
1820 	 * from lwp to clwp.
1821 	 */
1822 	mutex_enter(&cp->p_lock);
1823 	*clwp = *lwp;
1824 
1825 	/* clear microstate and resource usage data in new lwp */
1826 	init_mstate(ct, LMS_STOPPED);
1827 	bzero(&clwp->lwp_ru, sizeof (clwp->lwp_ru));
1828 	mutex_exit(&cp->p_lock);
1829 
1830 	/* fix up child's lwp */
1831 
1832 	clwp->lwp_pcb.pcb_flags = 0;
1833 #if defined(__sparc)
1834 	clwp->lwp_pcb.pcb_step = STEP_NONE;
1835 #endif
1836 	clwp->lwp_cursig = 0;
1837 	clwp->lwp_extsig = 0;
1838 	clwp->lwp_curinfo = (struct sigqueue *)0;
1839 	clwp->lwp_thread = ct;
1840 	ct->t_sysnum = t->t_sysnum;
1841 	clwp->lwp_regs = tregs;
1842 	clwp->lwp_fpu = tfpu;
1843 	clwp->lwp_brand = brand_data;
1844 	clwp->lwp_ap = clwp->lwp_arg;
1845 	clwp->lwp_procp = cp;
1846 	bzero(clwp->lwp_timer, sizeof (clwp->lwp_timer));
1847 	clwp->lwp_lastfault = 0;
1848 	clwp->lwp_lastfaddr = 0;
1849 
1850 	/* copy parent's struct regs to child. */
1851 	lwp_forkregs(lwp, clwp);
1852 
1853 	/*
1854 	 * Fork thread context ops, if any.
1855 	 */
1856 	if (t->t_ctx)
1857 		forkctx(t, ct);
1858 
1859 	/* fix door state in the child */
1860 	if (t->t_door)
1861 		door_fork(t, ct);
1862 
1863 	/* copy current contract templates, clear latest contracts */
1864 	lwp_ctmpl_copy(clwp, lwp);
1865 
1866 	mutex_enter(&cp->p_lock);
1867 	/* lwp_create() set the TP_HOLDLWP flag */
1868 	if (!(t->t_proc_flag & TP_HOLDLWP))
1869 		ct->t_proc_flag &= ~TP_HOLDLWP;
1870 	if (cp->p_flag & SMSACCT)
1871 		ct->t_proc_flag |= TP_MSACCT;
1872 	mutex_exit(&cp->p_lock);
1873 
1874 	/* Allow brand to propagate brand-specific state */
1875 	if (PROC_IS_BRANDED(p))
1876 		BROP(p)->b_forklwp(lwp, clwp);
1877 
1878 retry:
1879 	cid = t->t_cid;
1880 
1881 	val = CL_ALLOC(&bufp, cid, KM_SLEEP);
1882 	ASSERT(val == 0);
1883 
1884 	mutex_enter(&p->p_lock);
1885 	if (cid != t->t_cid) {
1886 		/*
1887 		 * Someone just changed this thread's scheduling class,
1888 		 * so try pre-allocating the buffer again.  Hopefully we
1889 		 * don't hit this often.
1890 		 */
1891 		mutex_exit(&p->p_lock);
1892 		CL_FREE(cid, bufp);
1893 		goto retry;
1894 	}
1895 
1896 	ct->t_unpark = t->t_unpark;
1897 	ct->t_clfuncs = t->t_clfuncs;
1898 	CL_FORK(t, ct, bufp);
1899 	ct->t_cid = t->t_cid;	/* after data allocated so prgetpsinfo works */
1900 	mutex_exit(&p->p_lock);
1901 
1902 	return (clwp);
1903 }
1904 
1905 /*
1906  * Add a new lwp entry to the lwp directory and to the lwpid hash table.
1907  */
1908 void
1909 lwp_hash_in(proc_t *p, lwpent_t *lep, tidhash_t *tidhash, uint_t tidhash_sz,
1910     int do_lock)
1911 {
1912 	tidhash_t *thp = &tidhash[TIDHASH(lep->le_lwpid, tidhash_sz)];
1913 	lwpdir_t **ldpp;
1914 	lwpdir_t *ldp;
1915 	kthread_t *t;
1916 
1917 	/*
1918 	 * Allocate a directory element from the free list.
1919 	 * Code elsewhere guarantees a free slot.
1920 	 */
1921 	ldp = p->p_lwpfree;
1922 	p->p_lwpfree = ldp->ld_next;
1923 	ASSERT(ldp->ld_entry == NULL);
1924 	ldp->ld_entry = lep;
1925 
1926 	if (do_lock)
1927 		mutex_enter(&thp->th_lock);
1928 
1929 	/*
1930 	 * Insert it into the lwpid hash table.
1931 	 */
1932 	ldpp = &thp->th_list;
1933 	ldp->ld_next = *ldpp;
1934 	*ldpp = ldp;
1935 
1936 	/*
1937 	 * Set the active thread's directory slot entry.
1938 	 */
1939 	if ((t = lep->le_thread) != NULL) {
1940 		ASSERT(lep->le_lwpid == t->t_tid);
1941 		t->t_dslot = (int)(ldp - p->p_lwpdir);
1942 	}
1943 
1944 	if (do_lock)
1945 		mutex_exit(&thp->th_lock);
1946 }
1947 
1948 /*
1949  * Remove an lwp from the lwpid hash table and free its directory entry.
1950  * This is done when a detached lwp exits in lwp_exit() or
1951  * when a non-detached lwp is waited for in lwp_wait() or
1952  * when a zombie lwp is detached in lwp_detach().
1953  */
1954 void
1955 lwp_hash_out(proc_t *p, id_t lwpid)
1956 {
1957 	tidhash_t *thp = &p->p_tidhash[TIDHASH(lwpid, p->p_tidhash_sz)];
1958 	lwpdir_t **ldpp;
1959 	lwpdir_t *ldp;
1960 	lwpent_t *lep;
1961 
1962 	mutex_enter(&thp->th_lock);
1963 	for (ldpp = &thp->th_list;
1964 	    (ldp = *ldpp) != NULL; ldpp = &ldp->ld_next) {
1965 		lep = ldp->ld_entry;
1966 		if (lep->le_lwpid == lwpid) {
1967 			prlwpfree(p, lep);	/* /proc deals with le_trace */
1968 			*ldpp = ldp->ld_next;
1969 			ldp->ld_entry = NULL;
1970 			ldp->ld_next = p->p_lwpfree;
1971 			p->p_lwpfree = ldp;
1972 			kmem_free(lep, sizeof (*lep));
1973 			break;
1974 		}
1975 	}
1976 	mutex_exit(&thp->th_lock);
1977 }
1978 
1979 /*
1980  * Lookup an lwp in the lwpid hash table by lwpid.
1981  */
1982 lwpdir_t *
1983 lwp_hash_lookup(proc_t *p, id_t lwpid)
1984 {
1985 	tidhash_t *thp;
1986 	lwpdir_t *ldp;
1987 
1988 	/*
1989 	 * The process may be exiting, after p_tidhash has been set to NULL in
1990 	 * proc_exit() but before prfee() has been called.  Return failure in
1991 	 * this case.
1992 	 */
1993 	if (p->p_tidhash == NULL)
1994 		return (NULL);
1995 
1996 	thp = &p->p_tidhash[TIDHASH(lwpid, p->p_tidhash_sz)];
1997 	for (ldp = thp->th_list; ldp != NULL; ldp = ldp->ld_next) {
1998 		if (ldp->ld_entry->le_lwpid == lwpid)
1999 			return (ldp);
2000 	}
2001 
2002 	return (NULL);
2003 }
2004 
2005 /*
2006  * Same as lwp_hash_lookup(), but acquire and return
2007  * the tid hash table entry lock on success.
2008  */
2009 lwpdir_t *
2010 lwp_hash_lookup_and_lock(proc_t *p, id_t lwpid, kmutex_t **mpp)
2011 {
2012 	tidhash_t *tidhash;
2013 	uint_t tidhash_sz;
2014 	tidhash_t *thp;
2015 	lwpdir_t *ldp;
2016 
2017 top:
2018 	tidhash_sz = p->p_tidhash_sz;
2019 	membar_consumer();
2020 	if ((tidhash = p->p_tidhash) == NULL)
2021 		return (NULL);
2022 
2023 	thp = &tidhash[TIDHASH(lwpid, tidhash_sz)];
2024 	mutex_enter(&thp->th_lock);
2025 
2026 	/*
2027 	 * Since we are not holding p->p_lock, the tid hash table
2028 	 * may have changed.  If so, start over.  If not, then
2029 	 * it cannot change until after we drop &thp->th_lock;
2030 	 */
2031 	if (tidhash != p->p_tidhash || tidhash_sz != p->p_tidhash_sz) {
2032 		mutex_exit(&thp->th_lock);
2033 		goto top;
2034 	}
2035 
2036 	for (ldp = thp->th_list; ldp != NULL; ldp = ldp->ld_next) {
2037 		if (ldp->ld_entry->le_lwpid == lwpid) {
2038 			*mpp = &thp->th_lock;
2039 			return (ldp);
2040 		}
2041 	}
2042 
2043 	mutex_exit(&thp->th_lock);
2044 	return (NULL);
2045 }
2046 
2047 /*
2048  * Update the indicated LWP usage statistic for the current LWP.
2049  */
2050 void
2051 lwp_stat_update(lwp_stat_id_t lwp_stat_id, long inc)
2052 {
2053 	klwp_t *lwp = ttolwp(curthread);
2054 
2055 	if (lwp == NULL)
2056 		return;
2057 
2058 	switch (lwp_stat_id) {
2059 	case LWP_STAT_INBLK:
2060 		lwp->lwp_ru.inblock += inc;
2061 		break;
2062 	case LWP_STAT_OUBLK:
2063 		lwp->lwp_ru.oublock += inc;
2064 		break;
2065 	case LWP_STAT_MSGRCV:
2066 		lwp->lwp_ru.msgrcv += inc;
2067 		break;
2068 	case LWP_STAT_MSGSND:
2069 		lwp->lwp_ru.msgsnd += inc;
2070 		break;
2071 	default:
2072 		panic("lwp_stat_update: invalid lwp_stat_id 0x%x", lwp_stat_id);
2073 	}
2074 }
2075