xref: /illumos-gate/usr/src/uts/common/os/lwp.c (revision 2570281c)
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 /*
28  * Copyright (c) 2013, Joyent, Inc. All rights reserved.
29  */
30 
31 #include <sys/param.h>
32 #include <sys/types.h>
33 #include <sys/sysmacros.h>
34 #include <sys/systm.h>
35 #include <sys/thread.h>
36 #include <sys/proc.h>
37 #include <sys/task.h>
38 #include <sys/project.h>
39 #include <sys/signal.h>
40 #include <sys/errno.h>
41 #include <sys/vmparam.h>
42 #include <sys/stack.h>
43 #include <sys/procfs.h>
44 #include <sys/prsystm.h>
45 #include <sys/cpuvar.h>
46 #include <sys/kmem.h>
47 #include <sys/vtrace.h>
48 #include <sys/door.h>
49 #include <vm/seg_kp.h>
50 #include <sys/debug.h>
51 #include <sys/schedctl.h>
52 #include <sys/poll.h>
53 #include <sys/copyops.h>
54 #include <sys/lwp_upimutex_impl.h>
55 #include <sys/cpupart.h>
56 #include <sys/lgrp.h>
57 #include <sys/rctl.h>
58 #include <sys/contract_impl.h>
59 #include <sys/cpc_impl.h>
60 #include <sys/sdt.h>
61 #include <sys/cmn_err.h>
62 #include <sys/brand.h>
63 #include <sys/cyclic.h>
64 #include <sys/pool.h>
65 
66 /* hash function for the lwpid hash table, p->p_tidhash[] */
67 #define	TIDHASH(tid, hash_sz)	((tid) & ((hash_sz) - 1))
68 
69 void *segkp_lwp;		/* cookie for pool of segkp resources */
70 extern void reapq_move_lq_to_tq(kthread_t *);
71 extern void freectx_ctx(struct ctxop *);
72 
73 /*
74  * Create a kernel thread associated with a particular system process.  Give
75  * it an LWP so that microstate accounting will be available for it.
76  */
77 kthread_t *
lwp_kernel_create(proc_t * p,void (* proc)(),void * arg,int state,pri_t pri)78 lwp_kernel_create(proc_t *p, void (*proc)(), void *arg, int state, pri_t pri)
79 {
80 	klwp_t *lwp;
81 
82 	VERIFY((p->p_flag & SSYS) != 0);
83 
84 	lwp = lwp_create(proc, arg, 0, p, state, pri, &t0.t_hold, syscid, 0);
85 
86 	VERIFY(lwp != NULL);
87 
88 	return (lwptot(lwp));
89 }
90 
91 /*
92  * Create a thread that appears to be stopped at sys_rtt.
93  */
94 klwp_t *
lwp_create(void (* proc)(),caddr_t arg,size_t len,proc_t * p,int state,int pri,const k_sigset_t * smask,int cid,id_t lwpid)95 lwp_create(void (*proc)(), caddr_t arg, size_t len, proc_t *p,
96     int state, int pri, const k_sigset_t *smask, int cid, id_t lwpid)
97 {
98 	klwp_t *lwp = NULL;
99 	kthread_t *t;
100 	kthread_t *tx;
101 	cpupart_t *oldpart = NULL;
102 	size_t	stksize;
103 	caddr_t lwpdata = NULL;
104 	processorid_t	binding;
105 	int err = 0;
106 	kproject_t *oldkpj, *newkpj;
107 	void *bufp = NULL;
108 	klwp_t *curlwp;
109 	lwpent_t *lep;
110 	lwpdir_t *old_dir = NULL;
111 	uint_t old_dirsz = 0;
112 	tidhash_t *old_hash = NULL;
113 	uint_t old_hashsz = 0;
114 	ret_tidhash_t *ret_tidhash = NULL;
115 	int i;
116 	int rctlfail = 0;
117 	boolean_t branded = 0;
118 	struct ctxop *ctx = NULL;
119 
120 	ASSERT(cid != sysdccid);	/* system threads must start in SYS */
121 
122 	ASSERT(p != &p0);		/* No new LWPs in p0. */
123 
124 	mutex_enter(&p->p_lock);
125 	mutex_enter(&p->p_zone->zone_nlwps_lock);
126 	/*
127 	 * don't enforce rctl limits on system processes
128 	 */
129 	if (!CLASS_KERNEL(cid)) {
130 		if (p->p_task->tk_nlwps >= p->p_task->tk_nlwps_ctl)
131 			if (rctl_test(rc_task_lwps, p->p_task->tk_rctls, p,
132 			    1, 0) & RCT_DENY)
133 				rctlfail = 1;
134 		if (p->p_task->tk_proj->kpj_nlwps >=
135 		    p->p_task->tk_proj->kpj_nlwps_ctl)
136 			if (rctl_test(rc_project_nlwps,
137 			    p->p_task->tk_proj->kpj_rctls, p, 1, 0)
138 			    & RCT_DENY)
139 				rctlfail = 1;
140 		if (p->p_zone->zone_nlwps >= p->p_zone->zone_nlwps_ctl)
141 			if (rctl_test(rc_zone_nlwps, p->p_zone->zone_rctls, p,
142 			    1, 0) & RCT_DENY)
143 				rctlfail = 1;
144 	}
145 	if (rctlfail) {
146 		mutex_exit(&p->p_zone->zone_nlwps_lock);
147 		mutex_exit(&p->p_lock);
148 		atomic_inc_32(&p->p_zone->zone_ffcap);
149 		return (NULL);
150 	}
151 	p->p_task->tk_nlwps++;
152 	p->p_task->tk_proj->kpj_nlwps++;
153 	p->p_zone->zone_nlwps++;
154 	mutex_exit(&p->p_zone->zone_nlwps_lock);
155 	mutex_exit(&p->p_lock);
156 
157 	curlwp = ttolwp(curthread);
158 	if (curlwp == NULL || (stksize = curlwp->lwp_childstksz) == 0)
159 		stksize = lwp_default_stksize;
160 
161 	if (CLASS_KERNEL(cid)) {
162 		/*
163 		 * Since we are creating an LWP in an SSYS process, we do not
164 		 * inherit anything from the current thread's LWP.  We set
165 		 * stksize and lwpdata to 0 in order to let thread_create()
166 		 * allocate a regular kernel thread stack for this thread.
167 		 */
168 		curlwp = NULL;
169 		stksize = 0;
170 		lwpdata = NULL;
171 
172 	} else if (stksize == lwp_default_stksize) {
173 		/*
174 		 * Try to reuse an <lwp,stack> from the LWP deathrow.
175 		 */
176 		if (lwp_reapcnt > 0) {
177 			mutex_enter(&reaplock);
178 			if ((t = lwp_deathrow) != NULL) {
179 				ASSERT(t->t_swap);
180 				lwp_deathrow = t->t_forw;
181 				lwp_reapcnt--;
182 				lwpdata = t->t_swap;
183 				lwp = t->t_lwp;
184 				ctx = t->t_ctx;
185 				t->t_swap = NULL;
186 				t->t_lwp = NULL;
187 				t->t_ctx = NULL;
188 				reapq_move_lq_to_tq(t);
189 			}
190 			mutex_exit(&reaplock);
191 			if (lwp != NULL) {
192 				lwp_stk_fini(lwp);
193 			}
194 			if (ctx != NULL) {
195 				freectx_ctx(ctx);
196 			}
197 		}
198 		if (lwpdata == NULL &&
199 		    (lwpdata = (caddr_t)segkp_cache_get(segkp_lwp)) == NULL) {
200 			mutex_enter(&p->p_lock);
201 			mutex_enter(&p->p_zone->zone_nlwps_lock);
202 			p->p_task->tk_nlwps--;
203 			p->p_task->tk_proj->kpj_nlwps--;
204 			p->p_zone->zone_nlwps--;
205 			mutex_exit(&p->p_zone->zone_nlwps_lock);
206 			mutex_exit(&p->p_lock);
207 			atomic_inc_32(&p->p_zone->zone_ffnomem);
208 			return (NULL);
209 		}
210 	} else {
211 		stksize = roundup(stksize, PAGESIZE);
212 		if ((lwpdata = (caddr_t)segkp_get(segkp, stksize,
213 		    (KPD_NOWAIT | KPD_HASREDZONE | KPD_LOCKED))) == NULL) {
214 			mutex_enter(&p->p_lock);
215 			mutex_enter(&p->p_zone->zone_nlwps_lock);
216 			p->p_task->tk_nlwps--;
217 			p->p_task->tk_proj->kpj_nlwps--;
218 			p->p_zone->zone_nlwps--;
219 			mutex_exit(&p->p_zone->zone_nlwps_lock);
220 			mutex_exit(&p->p_lock);
221 			atomic_inc_32(&p->p_zone->zone_ffnomem);
222 			return (NULL);
223 		}
224 	}
225 
226 	/*
227 	 * Create a thread, initializing the stack pointer
228 	 */
229 	t = thread_create(lwpdata, stksize, NULL, NULL, 0, p, TS_STOPPED, pri);
230 
231 	/*
232 	 * If a non-NULL stack base is passed in, thread_create() assumes
233 	 * that the stack might be statically allocated (as opposed to being
234 	 * allocated from segkp), and so it does not set t_swap.  Since
235 	 * the lwpdata was allocated from segkp, we must set t_swap to point
236 	 * to it ourselves.
237 	 *
238 	 * This would be less confusing if t_swap had a better name; it really
239 	 * indicates that the stack is allocated from segkp, regardless of
240 	 * whether or not it is swappable.
241 	 */
242 	if (lwpdata != NULL) {
243 		ASSERT(!CLASS_KERNEL(cid));
244 		ASSERT(t->t_swap == NULL);
245 		t->t_swap = lwpdata;	/* Start of page-able data */
246 	}
247 
248 	/*
249 	 * If the stack and lwp can be reused, mark the thread as such.
250 	 * When we get to reapq_add() from resume_from_zombie(), these
251 	 * threads will go onto lwp_deathrow instead of thread_deathrow.
252 	 */
253 	if (!CLASS_KERNEL(cid) && stksize == lwp_default_stksize)
254 		t->t_flag |= T_LWPREUSE;
255 
256 	if (lwp == NULL)
257 		lwp = kmem_cache_alloc(lwp_cache, KM_SLEEP);
258 	bzero(lwp, sizeof (*lwp));
259 	t->t_lwp = lwp;
260 
261 	t->t_hold = *smask;
262 	lwp->lwp_thread = t;
263 	lwp->lwp_procp = p;
264 	lwp->lwp_sigaltstack.ss_flags = SS_DISABLE;
265 	if (curlwp != NULL && curlwp->lwp_childstksz != 0)
266 		lwp->lwp_childstksz = curlwp->lwp_childstksz;
267 
268 	t->t_stk = lwp_stk_init(lwp, t->t_stk);
269 	thread_load(t, proc, arg, len);
270 
271 	/*
272 	 * Allocate the SIGPROF buffer if ITIMER_REALPROF is in effect.
273 	 */
274 	if (p->p_rprof_cyclic != CYCLIC_NONE)
275 		t->t_rprof = kmem_zalloc(sizeof (struct rprof), KM_SLEEP);
276 
277 	if (cid != NOCLASS)
278 		(void) CL_ALLOC(&bufp, cid, KM_SLEEP);
279 
280 	/*
281 	 * Allocate an lwp directory entry for the new lwp.
282 	 */
283 	lep = kmem_zalloc(sizeof (*lep), KM_SLEEP);
284 
285 	mutex_enter(&p->p_lock);
286 grow:
287 	/*
288 	 * Grow the lwp (thread) directory and lwpid hash table if necessary.
289 	 * A note on the growth algorithm:
290 	 *	The new lwp directory size is computed as:
291 	 *		new = 2 * old + 2
292 	 *	Starting with an initial size of 2 (see exec_common()),
293 	 *	this yields numbers that are a power of two minus 2:
294 	 *		2, 6, 14, 30, 62, 126, 254, 510, 1022, ...
295 	 *	The size of the lwpid hash table must be a power of two
296 	 *	and must be commensurate in size with the lwp directory
297 	 *	so that hash bucket chains remain short.  Therefore,
298 	 *	the lwpid hash table size is computed as:
299 	 *		hashsz = (dirsz + 2) / 2
300 	 *	which leads to these hash table sizes corresponding to
301 	 *	the above directory sizes:
302 	 *		2, 4, 8, 16, 32, 64, 128, 256, 512, ...
303 	 * A note on growing the hash table:
304 	 *	For performance reasons, code in lwp_unpark() does not
305 	 *	acquire curproc->p_lock when searching the hash table.
306 	 *	Rather, it calls lwp_hash_lookup_and_lock() which
307 	 *	acquires only the individual hash bucket lock, taking
308 	 *	care to deal with reallocation of the hash table
309 	 *	during the time it takes to acquire the lock.
310 	 *
311 	 *	This is sufficient to protect the integrity of the
312 	 *	hash table, but it requires us to acquire all of the
313 	 *	old hash bucket locks before growing the hash table
314 	 *	and to release them afterwards.  It also requires us
315 	 *	not to free the old hash table because some thread
316 	 *	in lwp_hash_lookup_and_lock() might still be trying
317 	 *	to acquire the old bucket lock.
318 	 *
319 	 *	So we adopt the tactic of keeping all of the retired
320 	 *	hash tables on a linked list, so they can be safely
321 	 *	freed when the process exits or execs.
322 	 *
323 	 *	Because the hash table grows in powers of two, the
324 	 *	total size of all of the hash tables will be slightly
325 	 *	less than twice the size of the largest hash table.
326 	 */
327 	while (p->p_lwpfree == NULL) {
328 		uint_t dirsz = p->p_lwpdir_sz;
329 		lwpdir_t *new_dir;
330 		uint_t new_dirsz;
331 		lwpdir_t *ldp;
332 		tidhash_t *new_hash;
333 		uint_t new_hashsz;
334 
335 		mutex_exit(&p->p_lock);
336 
337 		/*
338 		 * Prepare to remember the old p_tidhash for later
339 		 * kmem_free()ing when the process exits or execs.
340 		 */
341 		if (ret_tidhash == NULL)
342 			ret_tidhash = kmem_zalloc(sizeof (ret_tidhash_t),
343 			    KM_SLEEP);
344 		if (old_dir != NULL)
345 			kmem_free(old_dir, old_dirsz * sizeof (*old_dir));
346 		if (old_hash != NULL)
347 			kmem_free(old_hash, old_hashsz * sizeof (*old_hash));
348 
349 		new_dirsz = 2 * dirsz + 2;
350 		new_dir = kmem_zalloc(new_dirsz * sizeof (lwpdir_t), KM_SLEEP);
351 		for (ldp = new_dir, i = 1; i < new_dirsz; i++, ldp++)
352 			ldp->ld_next = ldp + 1;
353 		new_hashsz = (new_dirsz + 2) / 2;
354 		new_hash = kmem_zalloc(new_hashsz * sizeof (tidhash_t),
355 		    KM_SLEEP);
356 
357 		mutex_enter(&p->p_lock);
358 		if (p == curproc)
359 			prbarrier(p);
360 
361 		if (dirsz != p->p_lwpdir_sz || p->p_lwpfree != NULL) {
362 			/*
363 			 * Someone else beat us to it or some lwp exited.
364 			 * Set up to free our memory and take a lap.
365 			 */
366 			old_dir = new_dir;
367 			old_dirsz = new_dirsz;
368 			old_hash = new_hash;
369 			old_hashsz = new_hashsz;
370 		} else {
371 			/*
372 			 * For the benefit of lwp_hash_lookup_and_lock(),
373 			 * called from lwp_unpark(), which searches the
374 			 * tid hash table without acquiring p->p_lock,
375 			 * we must acquire all of the tid hash table
376 			 * locks before replacing p->p_tidhash.
377 			 */
378 			old_hash = p->p_tidhash;
379 			old_hashsz = p->p_tidhash_sz;
380 			for (i = 0; i < old_hashsz; i++) {
381 				mutex_enter(&old_hash[i].th_lock);
382 				mutex_enter(&new_hash[i].th_lock);
383 			}
384 
385 			/*
386 			 * We simply hash in all of the old directory entries.
387 			 * This works because the old directory has no empty
388 			 * slots and the new hash table starts out empty.
389 			 * This reproduces the original directory ordering
390 			 * (required for /proc directory semantics).
391 			 */
392 			old_dir = p->p_lwpdir;
393 			old_dirsz = p->p_lwpdir_sz;
394 			p->p_lwpdir = new_dir;
395 			p->p_lwpfree = new_dir;
396 			p->p_lwpdir_sz = new_dirsz;
397 			for (ldp = old_dir, i = 0; i < old_dirsz; i++, ldp++)
398 				lwp_hash_in(p, ldp->ld_entry,
399 				    new_hash, new_hashsz, 0);
400 
401 			/*
402 			 * Remember the old hash table along with all
403 			 * of the previously-remembered hash tables.
404 			 * We will free them at process exit or exec.
405 			 */
406 			ret_tidhash->rth_tidhash = old_hash;
407 			ret_tidhash->rth_tidhash_sz = old_hashsz;
408 			ret_tidhash->rth_next = p->p_ret_tidhash;
409 			p->p_ret_tidhash = ret_tidhash;
410 
411 			/*
412 			 * Now establish the new tid hash table.
413 			 * As soon as we assign p->p_tidhash,
414 			 * code in lwp_unpark() can start using it.
415 			 */
416 			membar_producer();
417 			p->p_tidhash = new_hash;
418 
419 			/*
420 			 * It is necessary that p_tidhash reach global
421 			 * visibility before p_tidhash_sz.  Otherwise,
422 			 * code in lwp_hash_lookup_and_lock() could
423 			 * index into the old p_tidhash using the new
424 			 * p_tidhash_sz and thereby access invalid data.
425 			 */
426 			membar_producer();
427 			p->p_tidhash_sz = new_hashsz;
428 
429 			/*
430 			 * Release the locks; allow lwp_unpark() to carry on.
431 			 */
432 			for (i = 0; i < old_hashsz; i++) {
433 				mutex_exit(&old_hash[i].th_lock);
434 				mutex_exit(&new_hash[i].th_lock);
435 			}
436 
437 			/*
438 			 * Avoid freeing these objects below.
439 			 */
440 			ret_tidhash = NULL;
441 			old_hash = NULL;
442 			old_hashsz = 0;
443 		}
444 	}
445 
446 	/*
447 	 * Block the process against /proc while we manipulate p->p_tlist,
448 	 * unless lwp_create() was called by /proc for the PCAGENT operation.
449 	 * We want to do this early enough so that we don't drop p->p_lock
450 	 * until the thread is put on the p->p_tlist.
451 	 */
452 	if (p == curproc) {
453 		prbarrier(p);
454 		/*
455 		 * If the current lwp has been requested to stop, do so now.
456 		 * Otherwise we have a race condition between /proc attempting
457 		 * to stop the process and this thread creating a new lwp
458 		 * that was not seen when the /proc PCSTOP request was issued.
459 		 * We rely on stop() to call prbarrier(p) before returning.
460 		 */
461 		while ((curthread->t_proc_flag & TP_PRSTOP) &&
462 		    !ttolwp(curthread)->lwp_nostop) {
463 			/*
464 			 * We called pool_barrier_enter() before calling
465 			 * here to lwp_create(). We have to call
466 			 * pool_barrier_exit() before stopping.
467 			 */
468 			pool_barrier_exit();
469 			prbarrier(p);
470 			stop(PR_REQUESTED, 0);
471 			/*
472 			 * And we have to repeat the call to
473 			 * pool_barrier_enter after stopping.
474 			 */
475 			pool_barrier_enter();
476 			prbarrier(p);
477 		}
478 
479 		/*
480 		 * If process is exiting, there could be a race between
481 		 * the agent lwp creation and the new lwp currently being
482 		 * created. So to prevent this race lwp creation is failed
483 		 * if the process is exiting.
484 		 */
485 		if (p->p_flag & (SEXITLWPS|SKILLED)) {
486 			err = 1;
487 			goto error;
488 		}
489 
490 		/*
491 		 * Since we might have dropped p->p_lock, the
492 		 * lwp directory free list might have changed.
493 		 */
494 		if (p->p_lwpfree == NULL)
495 			goto grow;
496 	}
497 
498 	kpreempt_disable();	/* can't grab cpu_lock here */
499 
500 	/*
501 	 * Inherit processor and processor set bindings from curthread.
502 	 *
503 	 * For kernel LWPs, we do not inherit processor set bindings at
504 	 * process creation time (i.e. when p != curproc).  After the
505 	 * kernel process is created, any subsequent LWPs must be created
506 	 * by threads in the kernel process, at which point we *will*
507 	 * inherit processor set bindings.
508 	 */
509 	if (CLASS_KERNEL(cid) && p != curproc) {
510 		t->t_bind_cpu = binding = PBIND_NONE;
511 		t->t_cpupart = oldpart = &cp_default;
512 		t->t_bind_pset = PS_NONE;
513 		t->t_bindflag = (uchar_t)default_binding_mode;
514 	} else {
515 		binding = curthread->t_bind_cpu;
516 		t->t_bind_cpu = binding;
517 		oldpart = t->t_cpupart;
518 		t->t_cpupart = curthread->t_cpupart;
519 		t->t_bind_pset = curthread->t_bind_pset;
520 		t->t_bindflag = curthread->t_bindflag |
521 		    (uchar_t)default_binding_mode;
522 	}
523 
524 	/*
525 	 * thread_create() initializes this thread's home lgroup to the root.
526 	 * Choose a more suitable lgroup, since this thread is associated
527 	 * with an lwp.
528 	 */
529 	ASSERT(oldpart != NULL);
530 	if (binding != PBIND_NONE && t->t_affinitycnt == 0) {
531 		t->t_bound_cpu = cpu[binding];
532 		if (t->t_lpl != t->t_bound_cpu->cpu_lpl)
533 			lgrp_move_thread(t, t->t_bound_cpu->cpu_lpl, 1);
534 	} else if (CLASS_KERNEL(cid)) {
535 		/*
536 		 * Kernel threads are always in the root lgrp.
537 		 */
538 		lgrp_move_thread(t,
539 		    &t->t_cpupart->cp_lgrploads[LGRP_ROOTID], 1);
540 	} else {
541 		lgrp_move_thread(t, lgrp_choose(t, t->t_cpupart), 1);
542 	}
543 
544 	kpreempt_enable();
545 
546 	/*
547 	 * make sure lpl points to our own partition
548 	 */
549 	ASSERT(t->t_lpl >= t->t_cpupart->cp_lgrploads);
550 	ASSERT(t->t_lpl < t->t_cpupart->cp_lgrploads +
551 	    t->t_cpupart->cp_nlgrploads);
552 
553 	/*
554 	 * It is safe to point the thread to the new project without holding it
555 	 * since we're holding the target process' p_lock here and therefore
556 	 * we're guaranteed that it will not move to another project.
557 	 */
558 	newkpj = p->p_task->tk_proj;
559 	oldkpj = ttoproj(t);
560 	if (newkpj != oldkpj) {
561 		t->t_proj = newkpj;
562 		(void) project_hold(newkpj);
563 		project_rele(oldkpj);
564 	}
565 
566 	if (cid != NOCLASS) {
567 		/*
568 		 * If the lwp is being created in the current process
569 		 * and matches the current thread's scheduling class,
570 		 * we should propagate the current thread's scheduling
571 		 * parameters by calling CL_FORK.  Otherwise just use
572 		 * the defaults by calling CL_ENTERCLASS.
573 		 */
574 		if (p != curproc || curthread->t_cid != cid) {
575 			err = CL_ENTERCLASS(t, cid, NULL, NULL, bufp);
576 			t->t_pri = pri;	/* CL_ENTERCLASS may have changed it */
577 			/*
578 			 * We don't call schedctl_set_cidpri(t) here
579 			 * because the schedctl data is not yet set
580 			 * up for the newly-created lwp.
581 			 */
582 		} else {
583 			t->t_clfuncs = &(sclass[cid].cl_funcs->thread);
584 			err = CL_FORK(curthread, t, bufp);
585 			t->t_cid = cid;
586 		}
587 		if (err) {
588 			atomic_inc_32(&p->p_zone->zone_ffmisc);
589 			goto error;
590 		} else {
591 			bufp = NULL;
592 		}
593 	}
594 
595 	/*
596 	 * If we were given an lwpid then use it, else allocate one.
597 	 */
598 	if (lwpid != 0)
599 		t->t_tid = lwpid;
600 	else {
601 		/*
602 		 * lwp/thread id 0 is never valid; reserved for special checks.
603 		 * lwp/thread id 1 is reserved for the main thread.
604 		 * Start again at 2 when INT_MAX has been reached
605 		 * (id_t is a signed 32-bit integer).
606 		 */
607 		id_t prev_id = p->p_lwpid;	/* last allocated tid */
608 
609 		do {			/* avoid lwpid duplication */
610 			if (p->p_lwpid == INT_MAX) {
611 				p->p_flag |= SLWPWRAP;
612 				p->p_lwpid = 1;
613 			}
614 			if ((t->t_tid = ++p->p_lwpid) == prev_id) {
615 				/*
616 				 * All lwpids are allocated; fail the request.
617 				 */
618 				err = 1;
619 				atomic_inc_32(&p->p_zone->zone_ffnoproc);
620 				goto error;
621 			}
622 			/*
623 			 * We only need to worry about colliding with an id
624 			 * that's already in use if this process has
625 			 * cycled through all available lwp ids.
626 			 */
627 			if ((p->p_flag & SLWPWRAP) == 0)
628 				break;
629 		} while (lwp_hash_lookup(p, t->t_tid) != NULL);
630 	}
631 
632 	/*
633 	 * If this is a branded process, let the brand do any necessary lwp
634 	 * initialization.
635 	 */
636 	if (PROC_IS_BRANDED(p)) {
637 		if (BROP(p)->b_initlwp(lwp)) {
638 			err = 1;
639 			atomic_inc_32(&p->p_zone->zone_ffmisc);
640 			goto error;
641 		}
642 		branded = 1;
643 	}
644 
645 	if (t->t_tid == 1) {
646 		kpreempt_disable();
647 		ASSERT(t->t_lpl != NULL);
648 		p->p_t1_lgrpid = t->t_lpl->lpl_lgrpid;
649 		kpreempt_enable();
650 		if (p->p_tr_lgrpid != LGRP_NONE &&
651 		    p->p_tr_lgrpid != p->p_t1_lgrpid) {
652 			lgrp_update_trthr_migrations(1);
653 		}
654 	}
655 
656 	p->p_lwpcnt++;
657 	t->t_waitfor = -1;
658 
659 	/*
660 	 * Turn microstate accounting on for thread if on for process.
661 	 */
662 	if (p->p_flag & SMSACCT)
663 		t->t_proc_flag |= TP_MSACCT;
664 
665 	/*
666 	 * If the process has watchpoints, mark the new thread as such.
667 	 */
668 	if (pr_watch_active(p))
669 		watch_enable(t);
670 
671 	/*
672 	 * The lwp is being created in the stopped state.
673 	 * We set all the necessary flags to indicate that fact here.
674 	 * We omit the TS_CREATE flag from t_schedflag so that the lwp
675 	 * cannot be set running until the caller is finished with it,
676 	 * even if lwp_continue() is called on it after we drop p->p_lock.
677 	 * When the caller is finished with the newly-created lwp,
678 	 * the caller must call lwp_create_done() to allow the lwp
679 	 * to be set running.  If the TP_HOLDLWP is left set, the
680 	 * lwp will suspend itself after reaching system call exit.
681 	 */
682 	init_mstate(t, LMS_STOPPED);
683 	t->t_proc_flag |= TP_HOLDLWP;
684 	t->t_schedflag |= (TS_ALLSTART & ~(TS_CSTART | TS_CREATE));
685 	t->t_whystop = PR_SUSPENDED;
686 	t->t_whatstop = SUSPEND_NORMAL;
687 	t->t_sig_check = 1;	/* ensure that TP_HOLDLWP is honored */
688 
689 	/*
690 	 * Set system call processing flags in case tracing or profiling
691 	 * is set.  The first system call will evaluate these and turn
692 	 * them off if they aren't needed.
693 	 */
694 	t->t_pre_sys = 1;
695 	t->t_post_sys = 1;
696 
697 	/*
698 	 * Insert the new thread into the list of all threads.
699 	 */
700 	if ((tx = p->p_tlist) == NULL) {
701 		t->t_back = t;
702 		t->t_forw = t;
703 		p->p_tlist = t;
704 	} else {
705 		t->t_forw = tx;
706 		t->t_back = tx->t_back;
707 		tx->t_back->t_forw = t;
708 		tx->t_back = t;
709 	}
710 
711 	/*
712 	 * Insert the new lwp into an lwp directory slot position
713 	 * and into the lwpid hash table.
714 	 */
715 	lep->le_thread = t;
716 	lep->le_lwpid = t->t_tid;
717 	lep->le_start = t->t_start;
718 	lwp_hash_in(p, lep, p->p_tidhash, p->p_tidhash_sz, 1);
719 
720 	lwp_fp_init(lwp);
721 
722 	if (state == TS_RUN) {
723 		/*
724 		 * We set the new lwp running immediately.
725 		 */
726 		t->t_proc_flag &= ~TP_HOLDLWP;
727 		lwp_create_done(t);
728 	}
729 
730 error:
731 	if (err) {
732 		if (CLASS_KERNEL(cid)) {
733 			/*
734 			 * This should only happen if a system process runs
735 			 * out of lwpids, which shouldn't occur.
736 			 */
737 			panic("Failed to create a system LWP");
738 		}
739 		/*
740 		 * We have failed to create an lwp, so decrement the number
741 		 * of lwps in the task and let the lgroup load averages know
742 		 * that this thread isn't going to show up.
743 		 */
744 		kpreempt_disable();
745 		lgrp_move_thread(t, NULL, 1);
746 		kpreempt_enable();
747 
748 		ASSERT(MUTEX_HELD(&p->p_lock));
749 		mutex_enter(&p->p_zone->zone_nlwps_lock);
750 		p->p_task->tk_nlwps--;
751 		p->p_task->tk_proj->kpj_nlwps--;
752 		p->p_zone->zone_nlwps--;
753 		mutex_exit(&p->p_zone->zone_nlwps_lock);
754 		if (cid != NOCLASS && bufp != NULL)
755 			CL_FREE(cid, bufp);
756 
757 		if (branded)
758 			BROP(p)->b_freelwp(lwp);
759 
760 		mutex_exit(&p->p_lock);
761 		t->t_state = TS_FREE;
762 		thread_rele(t);
763 
764 		/*
765 		 * We need to remove t from the list of all threads
766 		 * because thread_exit()/lwp_exit() isn't called on t.
767 		 */
768 		mutex_enter(&pidlock);
769 		ASSERT(t != t->t_next);		/* t0 never exits */
770 		t->t_next->t_prev = t->t_prev;
771 		t->t_prev->t_next = t->t_next;
772 		mutex_exit(&pidlock);
773 
774 		thread_free(t);
775 		kmem_free(lep, sizeof (*lep));
776 		lwp = NULL;
777 	} else {
778 		mutex_exit(&p->p_lock);
779 	}
780 
781 	if (old_dir != NULL)
782 		kmem_free(old_dir, old_dirsz * sizeof (*old_dir));
783 	if (old_hash != NULL)
784 		kmem_free(old_hash, old_hashsz * sizeof (*old_hash));
785 	if (ret_tidhash != NULL)
786 		kmem_free(ret_tidhash, sizeof (ret_tidhash_t));
787 
788 	DTRACE_PROC1(lwp__create, kthread_t *, t);
789 	return (lwp);
790 }
791 
792 /*
793  * lwp_create_done() is called by the caller of lwp_create() to set the
794  * newly-created lwp running after the caller has finished manipulating it.
795  */
796 void
lwp_create_done(kthread_t * t)797 lwp_create_done(kthread_t *t)
798 {
799 	proc_t *p = ttoproc(t);
800 
801 	ASSERT(MUTEX_HELD(&p->p_lock));
802 
803 	/*
804 	 * We set the TS_CREATE and TS_CSTART flags and call setrun_locked().
805 	 * (The absence of the TS_CREATE flag prevents the lwp from running
806 	 * until we are finished with it, even if lwp_continue() is called on
807 	 * it by some other lwp in the process or elsewhere in the kernel.)
808 	 */
809 	thread_lock(t);
810 	ASSERT(t->t_state == TS_STOPPED && !(t->t_schedflag & TS_CREATE));
811 	/*
812 	 * If TS_CSTART is set, lwp_continue(t) has been called and
813 	 * has already incremented p_lwprcnt; avoid doing this twice.
814 	 */
815 	if (!(t->t_schedflag & TS_CSTART))
816 		p->p_lwprcnt++;
817 	t->t_schedflag |= (TS_CSTART | TS_CREATE);
818 	setrun_locked(t);
819 	thread_unlock(t);
820 }
821 
822 /*
823  * Copy an LWP's active templates, and clear the latest contracts.
824  */
825 void
lwp_ctmpl_copy(klwp_t * dst,klwp_t * src)826 lwp_ctmpl_copy(klwp_t *dst, klwp_t *src)
827 {
828 	int i;
829 
830 	for (i = 0; i < ct_ntypes; i++) {
831 		dst->lwp_ct_active[i] = ctmpl_dup(src->lwp_ct_active[i]);
832 		dst->lwp_ct_latest[i] = NULL;
833 	}
834 }
835 
836 /*
837  * Clear an LWP's contract template state.
838  */
839 void
lwp_ctmpl_clear(klwp_t * lwp)840 lwp_ctmpl_clear(klwp_t *lwp)
841 {
842 	ct_template_t *tmpl;
843 	int i;
844 
845 	for (i = 0; i < ct_ntypes; i++) {
846 		if ((tmpl = lwp->lwp_ct_active[i]) != NULL) {
847 			ctmpl_free(tmpl);
848 			lwp->lwp_ct_active[i] = NULL;
849 		}
850 
851 		if (lwp->lwp_ct_latest[i] != NULL) {
852 			contract_rele(lwp->lwp_ct_latest[i]);
853 			lwp->lwp_ct_latest[i] = NULL;
854 		}
855 	}
856 }
857 
858 /*
859  * Individual lwp exit.
860  * If this is the last lwp, exit the whole process.
861  */
862 void
lwp_exit(void)863 lwp_exit(void)
864 {
865 	kthread_t *t = curthread;
866 	klwp_t *lwp = ttolwp(t);
867 	proc_t *p = ttoproc(t);
868 
869 	ASSERT(MUTEX_HELD(&p->p_lock));
870 
871 	mutex_exit(&p->p_lock);
872 
873 #if defined(__sparc)
874 	/*
875 	 * Ensure that the user stack is fully abandoned..
876 	 */
877 	trash_user_windows();
878 #endif
879 
880 	tsd_exit();			/* free thread specific data */
881 
882 	kcpc_passivate();		/* Clean up performance counter state */
883 
884 	pollcleanup();
885 
886 	if (t->t_door)
887 		door_slam();
888 
889 	if (t->t_schedctl != NULL)
890 		schedctl_lwp_cleanup(t);
891 
892 	if (t->t_upimutex != NULL)
893 		upimutex_cleanup();
894 
895 	/*
896 	 * Perform any brand specific exit processing, then release any
897 	 * brand data associated with the lwp
898 	 */
899 	if (PROC_IS_BRANDED(p))
900 		BROP(p)->b_lwpexit(lwp);
901 
902 	lwp_pcb_exit();
903 
904 	mutex_enter(&p->p_lock);
905 	lwp_cleanup();
906 
907 	/*
908 	 * When this process is dumping core, its lwps are held here
909 	 * until the core dump is finished. Then exitlwps() is called
910 	 * again to release these lwps so that they can finish exiting.
911 	 */
912 	if (p->p_flag & SCOREDUMP)
913 		stop(PR_SUSPENDED, SUSPEND_NORMAL);
914 
915 	/*
916 	 * Block the process against /proc now that we have really acquired
917 	 * p->p_lock (to decrement p_lwpcnt and manipulate p_tlist at least).
918 	 */
919 	prbarrier(p);
920 
921 	/*
922 	 * Call proc_exit() if this is the last non-daemon lwp in the process.
923 	 */
924 	if (!(t->t_proc_flag & TP_DAEMON) &&
925 	    p->p_lwpcnt == p->p_lwpdaemon + 1) {
926 		mutex_exit(&p->p_lock);
927 		if (proc_exit(CLD_EXITED, 0) == 0) {
928 			/* Restarting init. */
929 			return;
930 		}
931 
932 		/*
933 		 * proc_exit() returns a non-zero value when some other
934 		 * lwp got there first.  We just have to continue in
935 		 * lwp_exit().
936 		 */
937 		mutex_enter(&p->p_lock);
938 		ASSERT(curproc->p_flag & SEXITLWPS);
939 		prbarrier(p);
940 	}
941 
942 	DTRACE_PROC(lwp__exit);
943 
944 	/*
945 	 * If the lwp is a detached lwp or if the process is exiting,
946 	 * remove (lwp_hash_out()) the lwp from the lwp directory.
947 	 * Otherwise null out the lwp's le_thread pointer in the lwp
948 	 * directory so that other threads will see it as a zombie lwp.
949 	 */
950 	prlwpexit(t);		/* notify /proc */
951 	if (!(t->t_proc_flag & TP_TWAIT) || (p->p_flag & SEXITLWPS))
952 		lwp_hash_out(p, t->t_tid);
953 	else {
954 		ASSERT(!(t->t_proc_flag & TP_DAEMON));
955 		p->p_lwpdir[t->t_dslot].ld_entry->le_thread = NULL;
956 		p->p_zombcnt++;
957 		cv_broadcast(&p->p_lwpexit);
958 	}
959 	if (t->t_proc_flag & TP_DAEMON) {
960 		p->p_lwpdaemon--;
961 		t->t_proc_flag &= ~TP_DAEMON;
962 	}
963 	t->t_proc_flag &= ~TP_TWAIT;
964 
965 	/*
966 	 * Maintain accurate lwp count for task.max-lwps resource control.
967 	 */
968 	mutex_enter(&p->p_zone->zone_nlwps_lock);
969 	p->p_task->tk_nlwps--;
970 	p->p_task->tk_proj->kpj_nlwps--;
971 	p->p_zone->zone_nlwps--;
972 	mutex_exit(&p->p_zone->zone_nlwps_lock);
973 
974 	CL_EXIT(t);		/* tell the scheduler that t is exiting */
975 	ASSERT(p->p_lwpcnt != 0);
976 	p->p_lwpcnt--;
977 
978 	/*
979 	 * If all remaining non-daemon lwps are waiting in lwp_wait(),
980 	 * wake them up so someone can return EDEADLK.
981 	 * (See the block comment preceeding lwp_wait().)
982 	 */
983 	if (p->p_lwpcnt == p->p_lwpdaemon + (p->p_lwpwait - p->p_lwpdwait))
984 		cv_broadcast(&p->p_lwpexit);
985 
986 	t->t_proc_flag |= TP_LWPEXIT;
987 	term_mstate(t);
988 
989 	t->t_forw->t_back = t->t_back;
990 	t->t_back->t_forw = t->t_forw;
991 	if (t == p->p_tlist)
992 		p->p_tlist = t->t_forw;
993 
994 	/*
995 	 * Clean up the signal state.
996 	 */
997 	if (t->t_sigqueue != NULL)
998 		sigdelq(p, t, 0);
999 	if (lwp->lwp_curinfo != NULL) {
1000 		siginfofree(lwp->lwp_curinfo);
1001 		lwp->lwp_curinfo = NULL;
1002 	}
1003 
1004 	/*
1005 	 * If we have spymaster information (that is, if we're an agent LWP),
1006 	 * free that now.
1007 	 */
1008 	if (lwp->lwp_spymaster != NULL) {
1009 		kmem_free(lwp->lwp_spymaster, sizeof (psinfo_t));
1010 		lwp->lwp_spymaster = NULL;
1011 	}
1012 
1013 	thread_rele(t);
1014 
1015 	/*
1016 	 * Terminated lwps are associated with process zero and are put onto
1017 	 * death-row by resume().  Avoid preemption after resetting t->t_procp.
1018 	 */
1019 	t->t_preempt++;
1020 
1021 	if (t->t_ctx != NULL)
1022 		exitctx(t);
1023 	if (p->p_pctx != NULL)
1024 		exitpctx(p);
1025 
1026 	t->t_procp = &p0;
1027 
1028 	/*
1029 	 * Notify the HAT about the change of address space
1030 	 */
1031 	hat_thread_exit(t);
1032 	/*
1033 	 * When this is the last running lwp in this process and some lwp is
1034 	 * waiting for this condition to become true, or this thread was being
1035 	 * suspended, then the waiting lwp is awakened.
1036 	 *
1037 	 * Also, if the process is exiting, we may have a thread waiting in
1038 	 * exitlwps() that needs to be notified.
1039 	 */
1040 	if (--p->p_lwprcnt == 0 || (t->t_proc_flag & TP_HOLDLWP) ||
1041 	    (p->p_flag & SEXITLWPS))
1042 		cv_broadcast(&p->p_holdlwps);
1043 
1044 	/*
1045 	 * Need to drop p_lock so we can reacquire pidlock.
1046 	 */
1047 	mutex_exit(&p->p_lock);
1048 	mutex_enter(&pidlock);
1049 
1050 	ASSERT(t != t->t_next);		/* t0 never exits */
1051 	t->t_next->t_prev = t->t_prev;
1052 	t->t_prev->t_next = t->t_next;
1053 	cv_broadcast(&t->t_joincv);	/* wake up anyone in thread_join */
1054 	mutex_exit(&pidlock);
1055 
1056 	t->t_state = TS_ZOMB;
1057 	swtch_from_zombie();
1058 	/* never returns */
1059 }
1060 
1061 
1062 /*
1063  * Cleanup function for an exiting lwp.
1064  * Called both from lwp_exit() and from proc_exit().
1065  * p->p_lock is repeatedly released and grabbed in this function.
1066  */
1067 void
lwp_cleanup(void)1068 lwp_cleanup(void)
1069 {
1070 	kthread_t *t = curthread;
1071 	proc_t *p = ttoproc(t);
1072 
1073 	ASSERT(MUTEX_HELD(&p->p_lock));
1074 
1075 	/* untimeout any lwp-bound realtime timers */
1076 	if (p->p_itimer != NULL)
1077 		timer_lwpexit();
1078 
1079 	/*
1080 	 * If this is the /proc agent lwp that is exiting, readjust p_lwpid
1081 	 * so it appears that the agent never existed, and clear p_agenttp.
1082 	 */
1083 	if (t == p->p_agenttp) {
1084 		ASSERT(t->t_tid == p->p_lwpid);
1085 		p->p_lwpid--;
1086 		p->p_agenttp = NULL;
1087 	}
1088 
1089 	/*
1090 	 * Do lgroup bookkeeping to account for thread exiting.
1091 	 */
1092 	kpreempt_disable();
1093 	lgrp_move_thread(t, NULL, 1);
1094 	if (t->t_tid == 1) {
1095 		p->p_t1_lgrpid = LGRP_NONE;
1096 	}
1097 	kpreempt_enable();
1098 
1099 	lwp_ctmpl_clear(ttolwp(t));
1100 }
1101 
1102 int
lwp_suspend(kthread_t * t)1103 lwp_suspend(kthread_t *t)
1104 {
1105 	int tid;
1106 	proc_t *p = ttoproc(t);
1107 
1108 	ASSERT(MUTEX_HELD(&p->p_lock));
1109 
1110 	/*
1111 	 * Set the thread's TP_HOLDLWP flag so it will stop in holdlwp().
1112 	 * If an lwp is stopping itself, there is no need to wait.
1113 	 */
1114 top:
1115 	t->t_proc_flag |= TP_HOLDLWP;
1116 	if (t == curthread) {
1117 		t->t_sig_check = 1;
1118 	} else {
1119 		/*
1120 		 * Make sure the lwp stops promptly.
1121 		 */
1122 		thread_lock(t);
1123 		t->t_sig_check = 1;
1124 		/*
1125 		 * XXX Should use virtual stop like /proc does instead of
1126 		 * XXX waking the thread to get it to stop.
1127 		 */
1128 		if (ISWAKEABLE(t) || ISWAITING(t)) {
1129 			setrun_locked(t);
1130 		} else if (t->t_state == TS_ONPROC && t->t_cpu != CPU) {
1131 			poke_cpu(t->t_cpu->cpu_id);
1132 		}
1133 
1134 		tid = t->t_tid;	 /* remember thread ID */
1135 		/*
1136 		 * Wait for lwp to stop
1137 		 */
1138 		while (!SUSPENDED(t)) {
1139 			/*
1140 			 * Drop the thread lock before waiting and reacquire it
1141 			 * afterwards, so the thread can change its t_state
1142 			 * field.
1143 			 */
1144 			thread_unlock(t);
1145 
1146 			/*
1147 			 * Check if aborted by exitlwps().
1148 			 */
1149 			if (p->p_flag & SEXITLWPS)
1150 				lwp_exit();
1151 
1152 			/*
1153 			 * Cooperate with jobcontrol signals and /proc stopping
1154 			 * by calling cv_wait_sig() to wait for the target
1155 			 * lwp to stop.  Just using cv_wait() can lead to
1156 			 * deadlock because, if some other lwp has stopped
1157 			 * by either of these mechanisms, then p_lwprcnt will
1158 			 * never become zero if we do a cv_wait().
1159 			 */
1160 			if (!cv_wait_sig(&p->p_holdlwps, &p->p_lock))
1161 				return (EINTR);
1162 
1163 			/*
1164 			 * Check to see if thread died while we were
1165 			 * waiting for it to suspend.
1166 			 */
1167 			if (idtot(p, tid) == NULL)
1168 				return (ESRCH);
1169 
1170 			thread_lock(t);
1171 			/*
1172 			 * If the TP_HOLDLWP flag went away, lwp_continue()
1173 			 * or vfork() must have been called while we were
1174 			 * waiting, so start over again.
1175 			 */
1176 			if ((t->t_proc_flag & TP_HOLDLWP) == 0) {
1177 				thread_unlock(t);
1178 				goto top;
1179 			}
1180 		}
1181 		thread_unlock(t);
1182 	}
1183 	return (0);
1184 }
1185 
1186 /*
1187  * continue a lwp that's been stopped by lwp_suspend().
1188  */
1189 void
lwp_continue(kthread_t * t)1190 lwp_continue(kthread_t *t)
1191 {
1192 	proc_t *p = ttoproc(t);
1193 	int was_suspended = t->t_proc_flag & TP_HOLDLWP;
1194 
1195 	ASSERT(MUTEX_HELD(&p->p_lock));
1196 
1197 	t->t_proc_flag &= ~TP_HOLDLWP;
1198 	thread_lock(t);
1199 	if (SUSPENDED(t) &&
1200 	    !(p->p_flag & (SHOLDFORK | SHOLDFORK1 | SHOLDWATCH))) {
1201 		p->p_lwprcnt++;
1202 		t->t_schedflag |= TS_CSTART;
1203 		setrun_locked(t);
1204 	}
1205 	thread_unlock(t);
1206 	/*
1207 	 * Wakeup anyone waiting for this thread to be suspended
1208 	 */
1209 	if (was_suspended)
1210 		cv_broadcast(&p->p_holdlwps);
1211 }
1212 
1213 /*
1214  * ********************************
1215  *  Miscellaneous lwp routines	  *
1216  * ********************************
1217  */
1218 /*
1219  * When a process is undergoing a forkall(), its p_flag is set to SHOLDFORK.
1220  * This will cause the process's lwps to stop at a hold point.  A hold
1221  * point is where a kernel thread has a flat stack.  This is at the
1222  * return from a system call and at the return from a user level trap.
1223  *
1224  * When a process is undergoing a fork1() or vfork(), its p_flag is set to
1225  * SHOLDFORK1.  This will cause the process's lwps to stop at a modified
1226  * hold point.  The lwps in the process are not being cloned, so they
1227  * are held at the usual hold points and also within issig_forreal().
1228  * This has the side-effect that their system calls do not return
1229  * showing EINTR.
1230  *
1231  * An lwp can also be held.  This is identified by the TP_HOLDLWP flag on
1232  * the thread.  The TP_HOLDLWP flag is set in lwp_suspend(), where the active
1233  * lwp is waiting for the target lwp to be stopped.
1234  */
1235 void
holdlwp(void)1236 holdlwp(void)
1237 {
1238 	proc_t *p = curproc;
1239 	kthread_t *t = curthread;
1240 
1241 	mutex_enter(&p->p_lock);
1242 	/*
1243 	 * Don't terminate immediately if the process is dumping core.
1244 	 * Once the process has dumped core, all lwps are terminated.
1245 	 */
1246 	if (!(p->p_flag & SCOREDUMP)) {
1247 		if ((p->p_flag & SEXITLWPS) || (t->t_proc_flag & TP_EXITLWP))
1248 			lwp_exit();
1249 	}
1250 	if (!(ISHOLD(p)) && !(p->p_flag & (SHOLDFORK1 | SHOLDWATCH))) {
1251 		mutex_exit(&p->p_lock);
1252 		return;
1253 	}
1254 	/*
1255 	 * stop() decrements p->p_lwprcnt and cv_signal()s &p->p_holdlwps
1256 	 * when p->p_lwprcnt becomes zero.
1257 	 */
1258 	stop(PR_SUSPENDED, SUSPEND_NORMAL);
1259 	if (p->p_flag & SEXITLWPS)
1260 		lwp_exit();
1261 	mutex_exit(&p->p_lock);
1262 }
1263 
1264 /*
1265  * Have all lwps within the process hold at a point where they are
1266  * cloneable (SHOLDFORK) or just safe w.r.t. fork1 (SHOLDFORK1).
1267  */
1268 int
holdlwps(int holdflag)1269 holdlwps(int holdflag)
1270 {
1271 	proc_t *p = curproc;
1272 
1273 	ASSERT(holdflag == SHOLDFORK || holdflag == SHOLDFORK1);
1274 	mutex_enter(&p->p_lock);
1275 	schedctl_finish_sigblock(curthread);
1276 again:
1277 	while (p->p_flag & (SEXITLWPS | SHOLDFORK | SHOLDFORK1 | SHOLDWATCH)) {
1278 		/*
1279 		 * If another lwp is doing a forkall() or proc_exit(), bail out.
1280 		 */
1281 		if (p->p_flag & (SEXITLWPS | SHOLDFORK)) {
1282 			mutex_exit(&p->p_lock);
1283 			return (0);
1284 		}
1285 		/*
1286 		 * Another lwp is doing a fork1() or is undergoing
1287 		 * watchpoint activity.  We hold here for it to complete.
1288 		 */
1289 		stop(PR_SUSPENDED, SUSPEND_NORMAL);
1290 	}
1291 	p->p_flag |= holdflag;
1292 	pokelwps(p);
1293 	--p->p_lwprcnt;
1294 	/*
1295 	 * Wait for the process to become quiescent (p->p_lwprcnt == 0).
1296 	 */
1297 	while (p->p_lwprcnt > 0) {
1298 		/*
1299 		 * Check if aborted by exitlwps().
1300 		 * Also check if SHOLDWATCH is set; it takes precedence.
1301 		 */
1302 		if (p->p_flag & (SEXITLWPS | SHOLDWATCH)) {
1303 			p->p_lwprcnt++;
1304 			p->p_flag &= ~holdflag;
1305 			cv_broadcast(&p->p_holdlwps);
1306 			goto again;
1307 		}
1308 		/*
1309 		 * Cooperate with jobcontrol signals and /proc stopping.
1310 		 * If some other lwp has stopped by either of these
1311 		 * mechanisms, then p_lwprcnt will never become zero
1312 		 * and the process will appear deadlocked unless we
1313 		 * stop here in sympathy with the other lwp before
1314 		 * doing the cv_wait() below.
1315 		 *
1316 		 * If the other lwp stops after we do the cv_wait(), it
1317 		 * will wake us up to loop around and do the sympathy stop.
1318 		 *
1319 		 * Since stop() drops p->p_lock, we must start from
1320 		 * the top again on returning from stop().
1321 		 */
1322 		if (p->p_stopsig | (curthread->t_proc_flag & TP_PRSTOP)) {
1323 			int whystop = p->p_stopsig? PR_JOBCONTROL :
1324 			    PR_REQUESTED;
1325 			p->p_lwprcnt++;
1326 			p->p_flag &= ~holdflag;
1327 			stop(whystop, p->p_stopsig);
1328 			goto again;
1329 		}
1330 		cv_wait(&p->p_holdlwps, &p->p_lock);
1331 	}
1332 	p->p_lwprcnt++;
1333 	p->p_flag &= ~holdflag;
1334 	mutex_exit(&p->p_lock);
1335 	return (1);
1336 }
1337 
1338 /*
1339  * See comments for holdwatch(), below.
1340  */
1341 static int
holdcheck(int clearflags)1342 holdcheck(int clearflags)
1343 {
1344 	proc_t *p = curproc;
1345 
1346 	/*
1347 	 * If we are trying to exit, that takes precedence over anything else.
1348 	 */
1349 	if (p->p_flag & SEXITLWPS) {
1350 		p->p_lwprcnt++;
1351 		p->p_flag &= ~clearflags;
1352 		lwp_exit();
1353 	}
1354 
1355 	/*
1356 	 * If another thread is calling fork1(), stop the current thread so the
1357 	 * other can complete.
1358 	 */
1359 	if (p->p_flag & SHOLDFORK1) {
1360 		p->p_lwprcnt++;
1361 		stop(PR_SUSPENDED, SUSPEND_NORMAL);
1362 		if (p->p_flag & SEXITLWPS) {
1363 			p->p_flag &= ~clearflags;
1364 			lwp_exit();
1365 		}
1366 		return (-1);
1367 	}
1368 
1369 	/*
1370 	 * If another thread is calling fork(), then indicate we are doing
1371 	 * watchpoint activity.  This will cause holdlwps() above to stop the
1372 	 * forking thread, at which point we can continue with watchpoint
1373 	 * activity.
1374 	 */
1375 	if (p->p_flag & SHOLDFORK) {
1376 		p->p_lwprcnt++;
1377 		while (p->p_flag & SHOLDFORK) {
1378 			p->p_flag |= SHOLDWATCH;
1379 			cv_broadcast(&p->p_holdlwps);
1380 			cv_wait(&p->p_holdlwps, &p->p_lock);
1381 			p->p_flag &= ~SHOLDWATCH;
1382 		}
1383 		return (-1);
1384 	}
1385 
1386 	return (0);
1387 }
1388 
1389 /*
1390  * Stop all lwps within the process, holding themselves in the kernel while the
1391  * active lwp undergoes watchpoint activity.  This is more complicated than
1392  * expected because stop() relies on calling holdwatch() in order to copyin data
1393  * from the user's address space.  A double barrier is used to prevent an
1394  * infinite loop.
1395  *
1396  * 	o The first thread into holdwatch() is the 'master' thread and does
1397  *        the following:
1398  *
1399  *              - Sets SHOLDWATCH on the current process
1400  *              - Sets TP_WATCHSTOP on the current thread
1401  *              - Waits for all threads to be either stopped or have
1402  *                TP_WATCHSTOP set.
1403  *              - Sets the SWATCHOK flag on the process
1404  *              - Unsets TP_WATCHSTOP
1405  *              - Waits for the other threads to completely stop
1406  *              - Unsets SWATCHOK
1407  *
1408  * 	o If SHOLDWATCH is already set when we enter this function, then another
1409  *        thread is already trying to stop this thread.  This 'slave' thread
1410  *        does the following:
1411  *
1412  *              - Sets TP_WATCHSTOP on the current thread
1413  *              - Waits for SWATCHOK flag to be set
1414  *              - Calls stop()
1415  *
1416  * 	o If SWATCHOK is set on the process, then this function immediately
1417  *        returns, as we must have been called via stop().
1418  *
1419  * In addition, there are other flags that take precedence over SHOLDWATCH:
1420  *
1421  * 	o If SEXITLWPS is set, exit immediately.
1422  *
1423  * 	o If SHOLDFORK1 is set, wait for fork1() to complete.
1424  *
1425  * 	o If SHOLDFORK is set, then watchpoint activity takes precedence In this
1426  *        case, set SHOLDWATCH, signalling the forking thread to stop first.
1427  *
1428  * 	o If the process is being stopped via /proc (TP_PRSTOP is set), then we
1429  *        stop the current thread.
1430  *
1431  * Returns 0 if all threads have been quiesced.  Returns non-zero if not all
1432  * threads were stopped, or the list of watched pages has changed.
1433  */
1434 int
holdwatch(void)1435 holdwatch(void)
1436 {
1437 	proc_t *p = curproc;
1438 	kthread_t *t = curthread;
1439 	int ret = 0;
1440 
1441 	mutex_enter(&p->p_lock);
1442 
1443 	p->p_lwprcnt--;
1444 
1445 	/*
1446 	 * Check for bail-out conditions as outlined above.
1447 	 */
1448 	if (holdcheck(0) != 0) {
1449 		mutex_exit(&p->p_lock);
1450 		return (-1);
1451 	}
1452 
1453 	if (!(p->p_flag & SHOLDWATCH)) {
1454 		/*
1455 		 * We are the master watchpoint thread.  Set SHOLDWATCH and poke
1456 		 * the other threads.
1457 		 */
1458 		p->p_flag |= SHOLDWATCH;
1459 		pokelwps(p);
1460 
1461 		/*
1462 		 * Wait for all threads to be stopped or have TP_WATCHSTOP set.
1463 		 */
1464 		while (pr_allstopped(p, 1) > 0) {
1465 			if (holdcheck(SHOLDWATCH) != 0) {
1466 				p->p_flag &= ~SHOLDWATCH;
1467 				mutex_exit(&p->p_lock);
1468 				return (-1);
1469 			}
1470 
1471 			cv_wait(&p->p_holdlwps, &p->p_lock);
1472 		}
1473 
1474 		/*
1475 		 * All threads are now stopped or in the process of stopping.
1476 		 * Set SWATCHOK and let them stop completely.
1477 		 */
1478 		p->p_flag |= SWATCHOK;
1479 		t->t_proc_flag &= ~TP_WATCHSTOP;
1480 		cv_broadcast(&p->p_holdlwps);
1481 
1482 		while (pr_allstopped(p, 0) > 0) {
1483 			/*
1484 			 * At first glance, it may appear that we don't need a
1485 			 * call to holdcheck() here.  But if the process gets a
1486 			 * SIGKILL signal, one of our stopped threads may have
1487 			 * been awakened and is waiting in exitlwps(), which
1488 			 * takes precedence over watchpoints.
1489 			 */
1490 			if (holdcheck(SHOLDWATCH | SWATCHOK) != 0) {
1491 				p->p_flag &= ~(SHOLDWATCH | SWATCHOK);
1492 				mutex_exit(&p->p_lock);
1493 				return (-1);
1494 			}
1495 
1496 			cv_wait(&p->p_holdlwps, &p->p_lock);
1497 		}
1498 
1499 		/*
1500 		 * All threads are now completely stopped.
1501 		 */
1502 		p->p_flag &= ~SWATCHOK;
1503 		p->p_flag &= ~SHOLDWATCH;
1504 		p->p_lwprcnt++;
1505 
1506 	} else if (!(p->p_flag & SWATCHOK)) {
1507 
1508 		/*
1509 		 * SHOLDWATCH is set, so another thread is trying to do
1510 		 * watchpoint activity.  Indicate this thread is stopping, and
1511 		 * wait for the OK from the master thread.
1512 		 */
1513 		t->t_proc_flag |= TP_WATCHSTOP;
1514 		cv_broadcast(&p->p_holdlwps);
1515 
1516 		while (!(p->p_flag & SWATCHOK)) {
1517 			if (holdcheck(0) != 0) {
1518 				t->t_proc_flag &= ~TP_WATCHSTOP;
1519 				mutex_exit(&p->p_lock);
1520 				return (-1);
1521 			}
1522 
1523 			cv_wait(&p->p_holdlwps, &p->p_lock);
1524 		}
1525 
1526 		/*
1527 		 * Once the master thread has given the OK, this thread can
1528 		 * actually call stop().
1529 		 */
1530 		t->t_proc_flag &= ~TP_WATCHSTOP;
1531 		p->p_lwprcnt++;
1532 
1533 		stop(PR_SUSPENDED, SUSPEND_NORMAL);
1534 
1535 		/*
1536 		 * It's not OK to do watchpoint activity, notify caller to
1537 		 * retry.
1538 		 */
1539 		ret = -1;
1540 
1541 	} else {
1542 
1543 		/*
1544 		 * The only way we can hit the case where SHOLDWATCH is set and
1545 		 * SWATCHOK is set is if we are triggering this from within a
1546 		 * stop() call.  Assert that this is the case.
1547 		 */
1548 
1549 		ASSERT(t->t_proc_flag & TP_STOPPING);
1550 		p->p_lwprcnt++;
1551 	}
1552 
1553 	mutex_exit(&p->p_lock);
1554 
1555 	return (ret);
1556 }
1557 
1558 /*
1559  * force all interruptible lwps to trap into the kernel.
1560  */
1561 void
pokelwps(proc_t * p)1562 pokelwps(proc_t *p)
1563 {
1564 	kthread_t *t;
1565 
1566 	ASSERT(MUTEX_HELD(&p->p_lock));
1567 
1568 	t = p->p_tlist;
1569 	do {
1570 		if (t == curthread)
1571 			continue;
1572 		thread_lock(t);
1573 		aston(t);	/* make thread trap or do post_syscall */
1574 		if (ISWAKEABLE(t) || ISWAITING(t)) {
1575 			setrun_locked(t);
1576 		} else if (t->t_state == TS_STOPPED) {
1577 			/*
1578 			 * Ensure that proc_exit() is not blocked by lwps
1579 			 * that were stopped via jobcontrol or /proc.
1580 			 */
1581 			if (p->p_flag & SEXITLWPS) {
1582 				p->p_stopsig = 0;
1583 				t->t_schedflag |= (TS_XSTART | TS_PSTART);
1584 				setrun_locked(t);
1585 			}
1586 			/*
1587 			 * If we are holding lwps for a forkall(),
1588 			 * force lwps that have been suspended via
1589 			 * lwp_suspend() and are suspended inside
1590 			 * of a system call to proceed to their
1591 			 * holdlwp() points where they are clonable.
1592 			 */
1593 			if ((p->p_flag & SHOLDFORK) && SUSPENDED(t)) {
1594 				if ((t->t_schedflag & TS_CSTART) == 0) {
1595 					p->p_lwprcnt++;
1596 					t->t_schedflag |= TS_CSTART;
1597 					setrun_locked(t);
1598 				}
1599 			}
1600 		} else if (t->t_state == TS_ONPROC) {
1601 			if (t->t_cpu != CPU)
1602 				poke_cpu(t->t_cpu->cpu_id);
1603 		}
1604 		thread_unlock(t);
1605 	} while ((t = t->t_forw) != p->p_tlist);
1606 }
1607 
1608 /*
1609  * undo the effects of holdlwps() or holdwatch().
1610  */
1611 void
continuelwps(proc_t * p)1612 continuelwps(proc_t *p)
1613 {
1614 	kthread_t *t;
1615 
1616 	/*
1617 	 * If this flag is set, then the original holdwatch() didn't actually
1618 	 * stop the process.  See comments for holdwatch().
1619 	 */
1620 	if (p->p_flag & SWATCHOK) {
1621 		ASSERT(curthread->t_proc_flag & TP_STOPPING);
1622 		return;
1623 	}
1624 
1625 	ASSERT(MUTEX_HELD(&p->p_lock));
1626 	ASSERT((p->p_flag & (SHOLDFORK | SHOLDFORK1 | SHOLDWATCH)) == 0);
1627 
1628 	t = p->p_tlist;
1629 	do {
1630 		thread_lock(t);		/* SUSPENDED looks at t_schedflag */
1631 		if (SUSPENDED(t) && !(t->t_proc_flag & TP_HOLDLWP)) {
1632 			p->p_lwprcnt++;
1633 			t->t_schedflag |= TS_CSTART;
1634 			setrun_locked(t);
1635 		}
1636 		thread_unlock(t);
1637 	} while ((t = t->t_forw) != p->p_tlist);
1638 }
1639 
1640 /*
1641  * Force all other LWPs in the current process other than the caller to exit,
1642  * and then cv_wait() on p_holdlwps for them to exit.  The exitlwps() function
1643  * is typically used in these situations:
1644  *
1645  *   (a) prior to an exec() system call
1646  *   (b) prior to dumping a core file
1647  *   (c) prior to a uadmin() shutdown
1648  *
1649  * If the 'coredump' flag is set, other LWPs are quiesced but not destroyed.
1650  * Multiple threads in the process can call this function at one time by
1651  * triggering execs or core dumps simultaneously, so the SEXITLWPS bit is used
1652  * to declare one particular thread the winner who gets to kill the others.
1653  * If a thread wins the exitlwps() dance, zero is returned; otherwise an
1654  * appropriate errno value is returned to caller for its system call to return.
1655  */
1656 int
exitlwps(int coredump)1657 exitlwps(int coredump)
1658 {
1659 	proc_t *p = curproc;
1660 	int heldcnt;
1661 
1662 	if (curthread->t_door)
1663 		door_slam();
1664 	if (p->p_door_list)
1665 		door_revoke_all();
1666 	if (curthread->t_schedctl != NULL)
1667 		schedctl_lwp_cleanup(curthread);
1668 
1669 	/*
1670 	 * Ensure that before starting to wait for other lwps to exit,
1671 	 * cleanup all upimutexes held by curthread. Otherwise, some other
1672 	 * lwp could be waiting (uninterruptibly) for a upimutex held by
1673 	 * curthread, and the call to pokelwps() below would deadlock.
1674 	 * Even if a blocked upimutex_lock is made interruptible,
1675 	 * curthread's upimutexes need to be unlocked: do it here.
1676 	 */
1677 	if (curthread->t_upimutex != NULL)
1678 		upimutex_cleanup();
1679 
1680 	/*
1681 	 * Grab p_lock in order to check and set SEXITLWPS to declare a winner.
1682 	 * We must also block any further /proc access from this point forward.
1683 	 */
1684 	mutex_enter(&p->p_lock);
1685 	prbarrier(p);
1686 
1687 	if (p->p_flag & SEXITLWPS) {
1688 		mutex_exit(&p->p_lock);
1689 		aston(curthread);	/* force a trip through post_syscall */
1690 		return (set_errno(EINTR));
1691 	}
1692 
1693 	p->p_flag |= SEXITLWPS;
1694 	if (coredump)		/* tell other lwps to stop, not exit */
1695 		p->p_flag |= SCOREDUMP;
1696 
1697 	/*
1698 	 * Give precedence to exitlwps() if a holdlwps() is
1699 	 * in progress. The lwp doing the holdlwps() operation
1700 	 * is aborted when it is awakened.
1701 	 */
1702 	while (p->p_flag & (SHOLDFORK | SHOLDFORK1 | SHOLDWATCH)) {
1703 		cv_broadcast(&p->p_holdlwps);
1704 		cv_wait(&p->p_holdlwps, &p->p_lock);
1705 		prbarrier(p);
1706 	}
1707 	p->p_flag |= SHOLDFORK;
1708 	pokelwps(p);
1709 
1710 	/*
1711 	 * Wait for process to become quiescent.
1712 	 */
1713 	--p->p_lwprcnt;
1714 	while (p->p_lwprcnt > 0) {
1715 		cv_wait(&p->p_holdlwps, &p->p_lock);
1716 		prbarrier(p);
1717 	}
1718 	p->p_lwprcnt++;
1719 	ASSERT(p->p_lwprcnt == 1);
1720 
1721 	/*
1722 	 * The SCOREDUMP flag puts the process into a quiescent
1723 	 * state.  The process's lwps remain attached to this
1724 	 * process until exitlwps() is called again without the
1725 	 * 'coredump' flag set, then the lwps are terminated
1726 	 * and the process can exit.
1727 	 */
1728 	if (coredump) {
1729 		p->p_flag &= ~(SCOREDUMP | SHOLDFORK | SEXITLWPS);
1730 		goto out;
1731 	}
1732 
1733 	/*
1734 	 * Determine if there are any lwps left dangling in
1735 	 * the stopped state.  This happens when exitlwps()
1736 	 * aborts a holdlwps() operation.
1737 	 */
1738 	p->p_flag &= ~SHOLDFORK;
1739 	if ((heldcnt = p->p_lwpcnt) > 1) {
1740 		kthread_t *t;
1741 		for (t = curthread->t_forw; --heldcnt > 0; t = t->t_forw) {
1742 			t->t_proc_flag &= ~TP_TWAIT;
1743 			lwp_continue(t);
1744 		}
1745 	}
1746 
1747 	/*
1748 	 * Wait for all other lwps to exit.
1749 	 */
1750 	--p->p_lwprcnt;
1751 	while (p->p_lwpcnt > 1) {
1752 		cv_wait(&p->p_holdlwps, &p->p_lock);
1753 		prbarrier(p);
1754 	}
1755 	++p->p_lwprcnt;
1756 	ASSERT(p->p_lwpcnt == 1 && p->p_lwprcnt == 1);
1757 
1758 	p->p_flag &= ~SEXITLWPS;
1759 	curthread->t_proc_flag &= ~TP_TWAIT;
1760 
1761 out:
1762 	if (!coredump && p->p_zombcnt) {	/* cleanup the zombie lwps */
1763 		lwpdir_t *ldp;
1764 		lwpent_t *lep;
1765 		int i;
1766 
1767 		for (ldp = p->p_lwpdir, i = 0; i < p->p_lwpdir_sz; i++, ldp++) {
1768 			lep = ldp->ld_entry;
1769 			if (lep != NULL && lep->le_thread != curthread) {
1770 				ASSERT(lep->le_thread == NULL);
1771 				p->p_zombcnt--;
1772 				lwp_hash_out(p, lep->le_lwpid);
1773 			}
1774 		}
1775 		ASSERT(p->p_zombcnt == 0);
1776 	}
1777 
1778 	/*
1779 	 * If some other LWP in the process wanted us to suspend ourself,
1780 	 * then we will not do it.  The other LWP is now terminated and
1781 	 * no one will ever continue us again if we suspend ourself.
1782 	 */
1783 	curthread->t_proc_flag &= ~TP_HOLDLWP;
1784 	p->p_flag &= ~(SHOLDFORK | SHOLDFORK1 | SHOLDWATCH | SLWPWRAP);
1785 	mutex_exit(&p->p_lock);
1786 	return (0);
1787 }
1788 
1789 /*
1790  * duplicate a lwp.
1791  */
1792 klwp_t *
forklwp(klwp_t * lwp,proc_t * cp,id_t lwpid)1793 forklwp(klwp_t *lwp, proc_t *cp, id_t lwpid)
1794 {
1795 	klwp_t *clwp;
1796 	void *tregs, *tfpu;
1797 	kthread_t *t = lwptot(lwp);
1798 	kthread_t *ct;
1799 	proc_t *p = lwptoproc(lwp);
1800 	int cid;
1801 	void *bufp;
1802 	void *brand_data;
1803 	int val;
1804 
1805 	ASSERT(p == curproc);
1806 	ASSERT(t == curthread || (SUSPENDED(t) && lwp->lwp_asleep == 0));
1807 
1808 #if defined(__sparc)
1809 	if (t == curthread)
1810 		(void) flush_user_windows_to_stack(NULL);
1811 #endif
1812 
1813 	if (t == curthread)
1814 		/* copy args out of registers first */
1815 		(void) save_syscall_args();
1816 
1817 	clwp = lwp_create(cp->p_lwpcnt == 0 ? lwp_rtt_initial : lwp_rtt,
1818 	    NULL, 0, cp, TS_STOPPED, t->t_pri, &t->t_hold, NOCLASS, lwpid);
1819 	if (clwp == NULL)
1820 		return (NULL);
1821 
1822 	/*
1823 	 * most of the parent's lwp can be copied to its duplicate,
1824 	 * except for the fields that are unique to each lwp, like
1825 	 * lwp_thread, lwp_procp, lwp_regs, and lwp_ap.
1826 	 */
1827 	ct = clwp->lwp_thread;
1828 	tregs = clwp->lwp_regs;
1829 	tfpu = clwp->lwp_fpu;
1830 	brand_data = clwp->lwp_brand;
1831 
1832 	/*
1833 	 * Copy parent lwp to child lwp.  Hold child's p_lock to prevent
1834 	 * mstate_aggr_state() from reading stale mstate entries copied
1835 	 * from lwp to clwp.
1836 	 */
1837 	mutex_enter(&cp->p_lock);
1838 	*clwp = *lwp;
1839 
1840 	/* clear microstate and resource usage data in new lwp */
1841 	init_mstate(ct, LMS_STOPPED);
1842 	bzero(&clwp->lwp_ru, sizeof (clwp->lwp_ru));
1843 	mutex_exit(&cp->p_lock);
1844 
1845 	/* fix up child's lwp */
1846 
1847 	clwp->lwp_pcb.pcb_flags = 0;
1848 #if defined(__sparc)
1849 	clwp->lwp_pcb.pcb_step = STEP_NONE;
1850 #endif
1851 	clwp->lwp_cursig = 0;
1852 	clwp->lwp_extsig = 0;
1853 	clwp->lwp_curinfo = (struct sigqueue *)0;
1854 	clwp->lwp_thread = ct;
1855 	ct->t_sysnum = t->t_sysnum;
1856 	clwp->lwp_regs = tregs;
1857 	clwp->lwp_fpu = tfpu;
1858 	clwp->lwp_brand = brand_data;
1859 	clwp->lwp_ap = clwp->lwp_arg;
1860 	clwp->lwp_procp = cp;
1861 	bzero(clwp->lwp_timer, sizeof (clwp->lwp_timer));
1862 	clwp->lwp_lastfault = 0;
1863 	clwp->lwp_lastfaddr = 0;
1864 
1865 	/* copy parent's struct regs to child. */
1866 	lwp_forkregs(lwp, clwp);
1867 
1868 	/*
1869 	 * Fork thread context ops, if any.
1870 	 */
1871 	if (t->t_ctx)
1872 		forkctx(t, ct);
1873 
1874 	/* fix door state in the child */
1875 	if (t->t_door)
1876 		door_fork(t, ct);
1877 
1878 	/* copy current contract templates, clear latest contracts */
1879 	lwp_ctmpl_copy(clwp, lwp);
1880 
1881 	mutex_enter(&cp->p_lock);
1882 	/* lwp_create() set the TP_HOLDLWP flag */
1883 	if (!(t->t_proc_flag & TP_HOLDLWP))
1884 		ct->t_proc_flag &= ~TP_HOLDLWP;
1885 	if (cp->p_flag & SMSACCT)
1886 		ct->t_proc_flag |= TP_MSACCT;
1887 	mutex_exit(&cp->p_lock);
1888 
1889 	/* Allow brand to propagate brand-specific state */
1890 	if (PROC_IS_BRANDED(p))
1891 		BROP(p)->b_forklwp(lwp, clwp);
1892 
1893 retry:
1894 	cid = t->t_cid;
1895 
1896 	val = CL_ALLOC(&bufp, cid, KM_SLEEP);
1897 	ASSERT(val == 0);
1898 
1899 	mutex_enter(&p->p_lock);
1900 	if (cid != t->t_cid) {
1901 		/*
1902 		 * Someone just changed this thread's scheduling class,
1903 		 * so try pre-allocating the buffer again.  Hopefully we
1904 		 * don't hit this often.
1905 		 */
1906 		mutex_exit(&p->p_lock);
1907 		CL_FREE(cid, bufp);
1908 		goto retry;
1909 	}
1910 
1911 	ct->t_unpark = t->t_unpark;
1912 	ct->t_clfuncs = t->t_clfuncs;
1913 	CL_FORK(t, ct, bufp);
1914 	ct->t_cid = t->t_cid;	/* after data allocated so prgetpsinfo works */
1915 	mutex_exit(&p->p_lock);
1916 
1917 	return (clwp);
1918 }
1919 
1920 /*
1921  * Add a new lwp entry to the lwp directory and to the lwpid hash table.
1922  */
1923 void
lwp_hash_in(proc_t * p,lwpent_t * lep,tidhash_t * tidhash,uint_t tidhash_sz,int do_lock)1924 lwp_hash_in(proc_t *p, lwpent_t *lep, tidhash_t *tidhash, uint_t tidhash_sz,
1925     int do_lock)
1926 {
1927 	tidhash_t *thp = &tidhash[TIDHASH(lep->le_lwpid, tidhash_sz)];
1928 	lwpdir_t **ldpp;
1929 	lwpdir_t *ldp;
1930 	kthread_t *t;
1931 
1932 	/*
1933 	 * Allocate a directory element from the free list.
1934 	 * Code elsewhere guarantees a free slot.
1935 	 */
1936 	ldp = p->p_lwpfree;
1937 	p->p_lwpfree = ldp->ld_next;
1938 	ASSERT(ldp->ld_entry == NULL);
1939 	ldp->ld_entry = lep;
1940 
1941 	if (do_lock)
1942 		mutex_enter(&thp->th_lock);
1943 
1944 	/*
1945 	 * Insert it into the lwpid hash table.
1946 	 */
1947 	ldpp = &thp->th_list;
1948 	ldp->ld_next = *ldpp;
1949 	*ldpp = ldp;
1950 
1951 	/*
1952 	 * Set the active thread's directory slot entry.
1953 	 */
1954 	if ((t = lep->le_thread) != NULL) {
1955 		ASSERT(lep->le_lwpid == t->t_tid);
1956 		t->t_dslot = (int)(ldp - p->p_lwpdir);
1957 	}
1958 
1959 	if (do_lock)
1960 		mutex_exit(&thp->th_lock);
1961 }
1962 
1963 /*
1964  * Remove an lwp from the lwpid hash table and free its directory entry.
1965  * This is done when a detached lwp exits in lwp_exit() or
1966  * when a non-detached lwp is waited for in lwp_wait() or
1967  * when a zombie lwp is detached in lwp_detach().
1968  */
1969 void
lwp_hash_out(proc_t * p,id_t lwpid)1970 lwp_hash_out(proc_t *p, id_t lwpid)
1971 {
1972 	tidhash_t *thp = &p->p_tidhash[TIDHASH(lwpid, p->p_tidhash_sz)];
1973 	lwpdir_t **ldpp;
1974 	lwpdir_t *ldp;
1975 	lwpent_t *lep;
1976 
1977 	mutex_enter(&thp->th_lock);
1978 	for (ldpp = &thp->th_list;
1979 	    (ldp = *ldpp) != NULL; ldpp = &ldp->ld_next) {
1980 		lep = ldp->ld_entry;
1981 		if (lep->le_lwpid == lwpid) {
1982 			prlwpfree(p, lep);	/* /proc deals with le_trace */
1983 			*ldpp = ldp->ld_next;
1984 			ldp->ld_entry = NULL;
1985 			ldp->ld_next = p->p_lwpfree;
1986 			p->p_lwpfree = ldp;
1987 			kmem_free(lep, sizeof (*lep));
1988 			break;
1989 		}
1990 	}
1991 	mutex_exit(&thp->th_lock);
1992 }
1993 
1994 /*
1995  * Lookup an lwp in the lwpid hash table by lwpid.
1996  */
1997 lwpdir_t *
lwp_hash_lookup(proc_t * p,id_t lwpid)1998 lwp_hash_lookup(proc_t *p, id_t lwpid)
1999 {
2000 	tidhash_t *thp;
2001 	lwpdir_t *ldp;
2002 
2003 	/*
2004 	 * The process may be exiting, after p_tidhash has been set to NULL in
2005 	 * proc_exit() but before prfee() has been called.  Return failure in
2006 	 * this case.
2007 	 */
2008 	if (p->p_tidhash == NULL)
2009 		return (NULL);
2010 
2011 	thp = &p->p_tidhash[TIDHASH(lwpid, p->p_tidhash_sz)];
2012 	for (ldp = thp->th_list; ldp != NULL; ldp = ldp->ld_next) {
2013 		if (ldp->ld_entry->le_lwpid == lwpid)
2014 			return (ldp);
2015 	}
2016 
2017 	return (NULL);
2018 }
2019 
2020 /*
2021  * Same as lwp_hash_lookup(), but acquire and return
2022  * the tid hash table entry lock on success.
2023  */
2024 lwpdir_t *
lwp_hash_lookup_and_lock(proc_t * p,id_t lwpid,kmutex_t ** mpp)2025 lwp_hash_lookup_and_lock(proc_t *p, id_t lwpid, kmutex_t **mpp)
2026 {
2027 	tidhash_t *tidhash;
2028 	uint_t tidhash_sz;
2029 	tidhash_t *thp;
2030 	lwpdir_t *ldp;
2031 
2032 top:
2033 	tidhash_sz = p->p_tidhash_sz;
2034 	membar_consumer();
2035 	if ((tidhash = p->p_tidhash) == NULL)
2036 		return (NULL);
2037 
2038 	thp = &tidhash[TIDHASH(lwpid, tidhash_sz)];
2039 	mutex_enter(&thp->th_lock);
2040 
2041 	/*
2042 	 * Since we are not holding p->p_lock, the tid hash table
2043 	 * may have changed.  If so, start over.  If not, then
2044 	 * it cannot change until after we drop &thp->th_lock;
2045 	 */
2046 	if (tidhash != p->p_tidhash || tidhash_sz != p->p_tidhash_sz) {
2047 		mutex_exit(&thp->th_lock);
2048 		goto top;
2049 	}
2050 
2051 	for (ldp = thp->th_list; ldp != NULL; ldp = ldp->ld_next) {
2052 		if (ldp->ld_entry->le_lwpid == lwpid) {
2053 			*mpp = &thp->th_lock;
2054 			return (ldp);
2055 		}
2056 	}
2057 
2058 	mutex_exit(&thp->th_lock);
2059 	return (NULL);
2060 }
2061 
2062 /*
2063  * Update the indicated LWP usage statistic for the current LWP.
2064  */
2065 void
lwp_stat_update(lwp_stat_id_t lwp_stat_id,long inc)2066 lwp_stat_update(lwp_stat_id_t lwp_stat_id, long inc)
2067 {
2068 	klwp_t *lwp = ttolwp(curthread);
2069 
2070 	if (lwp == NULL)
2071 		return;
2072 
2073 	switch (lwp_stat_id) {
2074 	case LWP_STAT_INBLK:
2075 		lwp->lwp_ru.inblock += inc;
2076 		break;
2077 	case LWP_STAT_OUBLK:
2078 		lwp->lwp_ru.oublock += inc;
2079 		break;
2080 	case LWP_STAT_MSGRCV:
2081 		lwp->lwp_ru.msgrcv += inc;
2082 		break;
2083 	case LWP_STAT_MSGSND:
2084 		lwp->lwp_ru.msgsnd += inc;
2085 		break;
2086 	default:
2087 		panic("lwp_stat_update: invalid lwp_stat_id 0x%x", lwp_stat_id);
2088 	}
2089 }
2090