xref: /titanic_41/usr/src/uts/common/os/sched.c (revision 5203bc321053fb87d7073c7640548fab73634793)
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 2009 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 /*	Copyright (c) 1984, 1986, 1987, 1988, 1989 AT&T	*/
28 /*	  All Rights Reserved	*/
29 
30 #include <sys/param.h>
31 #include <sys/types.h>
32 #include <sys/sysmacros.h>
33 #include <sys/systm.h>
34 #include <sys/proc.h>
35 #include <sys/cpuvar.h>
36 #include <sys/var.h>
37 #include <sys/tuneable.h>
38 #include <sys/cmn_err.h>
39 #include <sys/buf.h>
40 #include <sys/disp.h>
41 #include <sys/vmsystm.h>
42 #include <sys/vmparam.h>
43 #include <sys/class.h>
44 #include <sys/vtrace.h>
45 #include <sys/modctl.h>
46 #include <sys/debug.h>
47 #include <sys/tnf_probe.h>
48 #include <sys/procfs.h>
49 
50 #include <vm/seg.h>
51 #include <vm/seg_kp.h>
52 #include <vm/as.h>
53 #include <vm/rm.h>
54 #include <vm/seg_kmem.h>
55 #include <sys/callb.h>
56 
57 /*
58  * The swapper sleeps on runout when there is no one to swap in.
59  * It sleeps on runin when it could not find space to swap someone
60  * in or after swapping someone in.
61  */
62 char	runout;
63 char	runin;
64 char	wake_sched;	/* flag tells clock to wake swapper on next tick */
65 char	wake_sched_sec;	/* flag tells clock to wake swapper after a second */
66 
67 /*
68  * The swapper swaps processes to reduce memory demand and runs
69  * when avefree < desfree.  The swapper resorts to SOFTSWAP when
70  * avefree < desfree which results in swapping out all processes
71  * sleeping for more than maxslp seconds.  HARDSWAP occurs when the
72  * system is on the verge of thrashing and this results in swapping
73  * out runnable threads or threads sleeping for less than maxslp secs.
74  *
75  * The swapper runs through all the active processes in the system
76  * and invokes the scheduling class specific swapin/swapout routine
77  * for every thread in the process to obtain an effective priority
78  * for the process.  A priority of -1 implies that the thread isn't
79  * swappable.  This effective priority is used to find the most
80  * eligible process to swapout or swapin.
81  *
82  * NOTE:  Threads which have been swapped are not linked on any
83  *	  queue and their dispatcher lock points at the "swapped_lock".
84  *
85  * Processes containing threads with the TS_DONT_SWAP flag set cannot be
86  * swapped out immediately by the swapper.  This is due to the fact that
87  * such threads may be holding locks which may be needed by the swapper
88  * to push its pages out.  The TS_SWAPENQ flag is set on such threads
89  * to prevent them running in user mode.  When such threads reach a
90  * safe point (i.e., are not holding any locks - CL_TRAPRET), they
91  * queue themseleves onto the swap queue which is processed by the
92  * swapper.  This results in reducing memory demand when the system
93  * is desparate for memory as the thread can't run in user mode.
94  *
95  * The swap queue consists of threads, linked via t_link, which are
96  * haven't been swapped, are runnable but not on the run queue.  The
97  * swap queue is protected by the "swapped_lock".  The dispatcher
98  * lock (t_lockp) of all threads on the swap queue points at the
99  * "swapped_lock".  Thus, the entire queue and/or threads on the
100  * queue can be locked by acquiring "swapped_lock".
101  */
102 static kthread_t *tswap_queue;
103 extern disp_lock_t swapped_lock; /* protects swap queue and threads on it */
104 
105 int	maxslp = 0;
106 pgcnt_t	avefree;	/* 5 sec moving average of free memory */
107 pgcnt_t	avefree30;	/* 30 sec moving average of free memory */
108 
109 /*
110  * Minimum size used to decide if sufficient memory is available
111  * before a process is swapped in.  This is necessary since in most
112  * cases the actual size of a process (p_swrss) being swapped in
113  * is usually 2 pages (kernel stack pages).  This is due to the fact
114  * almost all user pages of a process are stolen by pageout before
115  * the swapper decides to swapout it out.
116  */
117 int	min_procsize = 12;
118 
119 static int	swapin(proc_t *);
120 static int	swapout(proc_t *, uint_t *, int);
121 static void	process_swap_queue();
122 
123 #ifdef __sparc
124 extern void lwp_swapin(kthread_t *);
125 #endif /* __sparc */
126 
127 /*
128  * Counters to keep track of the number of swapins or swapouts.
129  */
130 uint_t tot_swapped_in, tot_swapped_out;
131 uint_t softswap, hardswap, swapqswap;
132 
133 /*
134  * Macro to determine if a process is eligble to be swapped.
135  */
136 #define	not_swappable(p)					\
137 	(((p)->p_flag & SSYS) || (p)->p_stat == SIDL ||		\
138 	    (p)->p_stat == SZOMB || (p)->p_as == NULL ||	\
139 	    (p)->p_as == &kas)
140 
141 /*
142  * Memory scheduler.
143  */
144 void
145 sched()
146 {
147 	kthread_id_t	t;
148 	pri_t		proc_pri;
149 	pri_t		thread_pri;
150 	pri_t		swapin_pri;
151 	int		desperate;
152 	pgcnt_t		needs;
153 	int		divisor;
154 	proc_t		*prp;
155 	proc_t		*swapout_prp;
156 	proc_t		*swapin_prp;
157 	spgcnt_t	avail;
158 	int		chosen_pri;
159 	time_t		swapout_time;
160 	time_t		swapin_proc_time;
161 	callb_cpr_t	cprinfo;
162 	kmutex_t	swap_cpr_lock;
163 
164 	mutex_init(&swap_cpr_lock, NULL, MUTEX_DEFAULT, NULL);
165 	CALLB_CPR_INIT(&cprinfo, &swap_cpr_lock, callb_generic_cpr, "sched");
166 	if (maxslp == 0)
167 		maxslp = MAXSLP;
168 loop:
169 	needs = 0;
170 	desperate = 0;
171 
172 	swapin_pri = v.v_nglobpris;
173 	swapin_prp = NULL;
174 	chosen_pri = -1;
175 
176 	process_swap_queue();
177 
178 	/*
179 	 * Set desperate if
180 	 * 	1.  At least 2 runnable processes (on average).
181 	 *	2.  Short (5 sec) and longer (30 sec) average is less
182 	 *	    than minfree and desfree respectively.
183 	 *	3.  Pagein + pageout rate is excessive.
184 	 */
185 	if (avenrun[0] >= 2 * FSCALE &&
186 	    (MAX(avefree, avefree30) < desfree) &&
187 	    (pginrate + pgoutrate > maxpgio || avefree < minfree)) {
188 		TRACE_4(TR_FAC_SCHED, TR_DESPERATE,
189 		    "desp:avefree: %d, avefree30: %d, freemem: %d"
190 		    " pginrate: %d\n", avefree, avefree30, freemem, pginrate);
191 		desperate = 1;
192 		goto unload;
193 	}
194 
195 	/*
196 	 * Search list of processes to swapin and swapout deadwood.
197 	 */
198 	swapin_proc_time = 0;
199 top:
200 	mutex_enter(&pidlock);
201 	for (prp = practive; prp != NULL; prp = prp->p_next) {
202 		if (not_swappable(prp))
203 			continue;
204 
205 		/*
206 		 * Look at processes with at least one swapped lwp.
207 		 */
208 		if (prp->p_swapcnt) {
209 			time_t proc_time;
210 
211 			/*
212 			 * Higher priority processes are good candidates
213 			 * to swapin.
214 			 */
215 			mutex_enter(&prp->p_lock);
216 			proc_pri = -1;
217 			t = prp->p_tlist;
218 			proc_time = 0;
219 			do {
220 				if (t->t_schedflag & TS_LOAD)
221 					continue;
222 
223 				thread_lock(t);
224 				thread_pri = CL_SWAPIN(t, 0);
225 				thread_unlock(t);
226 
227 				if (t->t_stime - proc_time > 0)
228 					proc_time = t->t_stime;
229 				if (thread_pri > proc_pri)
230 					proc_pri = thread_pri;
231 			} while ((t = t->t_forw) != prp->p_tlist);
232 			mutex_exit(&prp->p_lock);
233 
234 			if (proc_pri == -1)
235 				continue;
236 
237 			TRACE_3(TR_FAC_SCHED, TR_CHOOSE_SWAPIN,
238 			    "prp %p epri %d proc_time %d",
239 			    prp, proc_pri, proc_time);
240 
241 			/*
242 			 * Swapin processes with a high effective priority.
243 			 */
244 			if (swapin_prp == NULL || proc_pri > chosen_pri) {
245 				swapin_prp = prp;
246 				chosen_pri = proc_pri;
247 				swapin_pri = proc_pri;
248 				swapin_proc_time = proc_time;
249 			}
250 		} else {
251 			/*
252 			 * No need to soft swap if we have sufficient
253 			 * memory.
254 			 */
255 			if (avefree > desfree ||
256 			    avefree < desfree && freemem > desfree)
257 				continue;
258 
259 			/*
260 			 * Skip processes that are exiting
261 			 * or whose address spaces are locked.
262 			 */
263 			mutex_enter(&prp->p_lock);
264 			if ((prp->p_flag & SEXITING) ||
265 			    (prp->p_as != NULL && AS_ISPGLCK(prp->p_as))) {
266 				mutex_exit(&prp->p_lock);
267 				continue;
268 			}
269 
270 			/*
271 			 * Softswapping to kick out deadwood.
272 			 */
273 			proc_pri = -1;
274 			t = prp->p_tlist;
275 			do {
276 				if ((t->t_schedflag & (TS_SWAPENQ |
277 				    TS_ON_SWAPQ | TS_LOAD)) != TS_LOAD)
278 					continue;
279 
280 				thread_lock(t);
281 				thread_pri = CL_SWAPOUT(t, SOFTSWAP);
282 				thread_unlock(t);
283 				if (thread_pri > proc_pri)
284 					proc_pri = thread_pri;
285 			} while ((t = t->t_forw) != prp->p_tlist);
286 
287 			if (proc_pri != -1) {
288 				uint_t swrss;
289 
290 				mutex_exit(&pidlock);
291 
292 				TRACE_1(TR_FAC_SCHED, TR_SOFTSWAP,
293 				    "softswap:prp %p", prp);
294 
295 				(void) swapout(prp, &swrss, SOFTSWAP);
296 				softswap++;
297 				prp->p_swrss += swrss;
298 				mutex_exit(&prp->p_lock);
299 				goto top;
300 			}
301 			mutex_exit(&prp->p_lock);
302 		}
303 	}
304 	if (swapin_prp != NULL)
305 		mutex_enter(&swapin_prp->p_lock);
306 	mutex_exit(&pidlock);
307 
308 	if (swapin_prp == NULL) {
309 		TRACE_3(TR_FAC_SCHED, TR_RUNOUT,
310 		"schedrunout:runout nswapped: %d, avefree: %ld freemem: %ld",
311 		    nswapped, avefree, freemem);
312 
313 		t = curthread;
314 		thread_lock(t);
315 		runout++;
316 		t->t_schedflag |= (TS_ALLSTART & ~TS_CSTART);
317 		t->t_whystop = PR_SUSPENDED;
318 		t->t_whatstop = SUSPEND_NORMAL;
319 		(void) new_mstate(t, LMS_SLEEP);
320 		mutex_enter(&swap_cpr_lock);
321 		CALLB_CPR_SAFE_BEGIN(&cprinfo);
322 		mutex_exit(&swap_cpr_lock);
323 		thread_stop(t);		/* change state and drop lock */
324 		swtch();
325 		mutex_enter(&swap_cpr_lock);
326 		CALLB_CPR_SAFE_END(&cprinfo, &swap_cpr_lock);
327 		mutex_exit(&swap_cpr_lock);
328 		goto loop;
329 	}
330 
331 	/*
332 	 * Decide how deserving this process is to be brought in.
333 	 * Needs is an estimate of how much core the process will
334 	 * need.  If the process has been out for a while, then we
335 	 * will bring it in with 1/2 the core needed, otherwise
336 	 * we are conservative.
337 	 */
338 	divisor = 1;
339 	swapout_time = (ddi_get_lbolt() - swapin_proc_time) / hz;
340 	if (swapout_time > maxslp / 2)
341 		divisor = 2;
342 
343 	needs = MIN(swapin_prp->p_swrss, lotsfree);
344 	needs = MAX(needs, min_procsize);
345 	needs = needs / divisor;
346 
347 	/*
348 	 * Use freemem, since we want processes to be swapped
349 	 * in quickly.
350 	 */
351 	avail = freemem - deficit;
352 	if (avail > (spgcnt_t)needs) {
353 		deficit += needs;
354 
355 		TRACE_2(TR_FAC_SCHED, TR_SWAPIN_VALUES,
356 		    "swapin_values: prp %p needs %lu", swapin_prp, needs);
357 
358 		if (swapin(swapin_prp)) {
359 			mutex_exit(&swapin_prp->p_lock);
360 			goto loop;
361 		}
362 		deficit -= MIN(needs, deficit);
363 		mutex_exit(&swapin_prp->p_lock);
364 	} else {
365 		mutex_exit(&swapin_prp->p_lock);
366 		/*
367 		 * If deficit is high, too many processes have been
368 		 * swapped in so wait a sec before attempting to
369 		 * swapin more.
370 		 */
371 		if (freemem > needs) {
372 			TRACE_2(TR_FAC_SCHED, TR_HIGH_DEFICIT,
373 			    "deficit: prp %p needs %lu", swapin_prp, needs);
374 			goto block;
375 		}
376 	}
377 
378 	TRACE_2(TR_FAC_SCHED, TR_UNLOAD,
379 	    "unload: prp %p needs %lu", swapin_prp, needs);
380 
381 unload:
382 	/*
383 	 * Unload all unloadable modules, free all other memory
384 	 * resources we can find, then look for a thread to hardswap.
385 	 */
386 	modreap();
387 	segkp_cache_free();
388 
389 	swapout_prp = NULL;
390 	mutex_enter(&pidlock);
391 	for (prp = practive; prp != NULL; prp = prp->p_next) {
392 
393 		/*
394 		 * No need to soft swap if we have sufficient
395 		 * memory.
396 		 */
397 		if (not_swappable(prp))
398 			continue;
399 
400 		if (avefree > minfree ||
401 		    avefree < minfree && freemem > desfree) {
402 			swapout_prp = NULL;
403 			break;
404 		}
405 
406 		/*
407 		 * Skip processes that are exiting
408 		 * or whose address spaces are locked.
409 		 */
410 		mutex_enter(&prp->p_lock);
411 		if ((prp->p_flag & SEXITING) ||
412 		    (prp->p_as != NULL && AS_ISPGLCK(prp->p_as))) {
413 			mutex_exit(&prp->p_lock);
414 			continue;
415 		}
416 
417 		proc_pri = -1;
418 		t = prp->p_tlist;
419 		do {
420 			if ((t->t_schedflag & (TS_SWAPENQ |
421 			    TS_ON_SWAPQ | TS_LOAD)) != TS_LOAD)
422 				continue;
423 
424 			thread_lock(t);
425 			thread_pri = CL_SWAPOUT(t, HARDSWAP);
426 			thread_unlock(t);
427 			if (thread_pri > proc_pri)
428 				proc_pri = thread_pri;
429 		} while ((t = t->t_forw) != prp->p_tlist);
430 
431 		mutex_exit(&prp->p_lock);
432 		if (proc_pri == -1)
433 			continue;
434 
435 		/*
436 		 * Swapout processes sleeping with a lower priority
437 		 * than the one currently being swapped in, if any.
438 		 */
439 		if (swapin_prp == NULL || swapin_pri > proc_pri) {
440 			TRACE_2(TR_FAC_SCHED, TR_CHOOSE_SWAPOUT,
441 			    "hardswap: prp %p needs %lu", prp, needs);
442 
443 			if (swapout_prp == NULL || proc_pri < chosen_pri) {
444 				swapout_prp = prp;
445 				chosen_pri = proc_pri;
446 			}
447 		}
448 	}
449 
450 	/*
451 	 * Acquire the "p_lock" before dropping "pidlock"
452 	 * to prevent the proc structure from being freed
453 	 * if the process exits before swapout completes.
454 	 */
455 	if (swapout_prp != NULL)
456 		mutex_enter(&swapout_prp->p_lock);
457 	mutex_exit(&pidlock);
458 
459 	if ((prp = swapout_prp) != NULL) {
460 		uint_t swrss = 0;
461 		int swapped;
462 
463 		swapped = swapout(prp, &swrss, HARDSWAP);
464 		if (swapped) {
465 			/*
466 			 * If desperate, we want to give the space obtained
467 			 * by swapping this process out to processes in core,
468 			 * so we give them a chance by increasing deficit.
469 			 */
470 			prp->p_swrss += swrss;
471 			if (desperate)
472 				deficit += MIN(prp->p_swrss, lotsfree);
473 			hardswap++;
474 		}
475 		mutex_exit(&swapout_prp->p_lock);
476 
477 		if (swapped)
478 			goto loop;
479 	}
480 
481 	/*
482 	 * Delay for 1 second and look again later.
483 	 */
484 	TRACE_3(TR_FAC_SCHED, TR_RUNIN,
485 	    "schedrunin:runin nswapped: %d, avefree: %ld freemem: %ld",
486 	    nswapped, avefree, freemem);
487 
488 block:
489 	t = curthread;
490 	thread_lock(t);
491 	runin++;
492 	t->t_schedflag |= (TS_ALLSTART & ~TS_CSTART);
493 	t->t_whystop = PR_SUSPENDED;
494 	t->t_whatstop = SUSPEND_NORMAL;
495 	(void) new_mstate(t, LMS_SLEEP);
496 	mutex_enter(&swap_cpr_lock);
497 	CALLB_CPR_SAFE_BEGIN(&cprinfo);
498 	mutex_exit(&swap_cpr_lock);
499 	thread_stop(t);		/* change to stop state and drop lock */
500 	swtch();
501 	mutex_enter(&swap_cpr_lock);
502 	CALLB_CPR_SAFE_END(&cprinfo, &swap_cpr_lock);
503 	mutex_exit(&swap_cpr_lock);
504 	goto loop;
505 }
506 
507 /*
508  * Remove the specified thread from the swap queue.
509  */
510 static void
511 swapdeq(kthread_id_t tp)
512 {
513 	kthread_id_t *tpp;
514 
515 	ASSERT(THREAD_LOCK_HELD(tp));
516 	ASSERT(tp->t_schedflag & TS_ON_SWAPQ);
517 
518 	tpp = &tswap_queue;
519 	for (;;) {
520 		ASSERT(*tpp != NULL);
521 		if (*tpp == tp)
522 			break;
523 		tpp = &(*tpp)->t_link;
524 	}
525 	*tpp = tp->t_link;
526 	tp->t_schedflag &= ~TS_ON_SWAPQ;
527 }
528 
529 /*
530  * Swap in lwps.  Returns nonzero on success (i.e., if at least one lwp is
531  * swapped in) and 0 on failure.
532  */
533 static int
534 swapin(proc_t *pp)
535 {
536 	kthread_id_t tp;
537 	int err;
538 	int num_swapped_in = 0;
539 	struct cpu *cpup = CPU;
540 	pri_t thread_pri;
541 
542 	ASSERT(MUTEX_HELD(&pp->p_lock));
543 	ASSERT(pp->p_swapcnt);
544 
545 top:
546 	tp = pp->p_tlist;
547 	do {
548 		/*
549 		 * Only swapin eligible lwps (specified by the scheduling
550 		 * class) which are unloaded and ready to run.
551 		 */
552 		thread_lock(tp);
553 		thread_pri = CL_SWAPIN(tp, 0);
554 		if (thread_pri != -1 && tp->t_state == TS_RUN &&
555 		    (tp->t_schedflag & TS_LOAD) == 0) {
556 			size_t stack_size;
557 			pgcnt_t stack_pages;
558 
559 			ASSERT((tp->t_schedflag & TS_ON_SWAPQ) == 0);
560 
561 			thread_unlock(tp);
562 			/*
563 			 * Now drop the p_lock since the stack needs
564 			 * to brought in.
565 			 */
566 			mutex_exit(&pp->p_lock);
567 
568 			stack_size = swapsize(tp->t_swap);
569 			stack_pages = btopr(stack_size);
570 			/* Kernel probe */
571 			TNF_PROBE_4(swapin_lwp, "vm swap swapin", /* CSTYLED */,
572 			    tnf_pid,		pid,		pp->p_pid,
573 			    tnf_lwpid,		lwpid,		tp->t_tid,
574 			    tnf_kthread_id,	tid,		tp,
575 			    tnf_ulong,		page_count,	stack_pages);
576 
577 			rw_enter(&kas.a_lock, RW_READER);
578 			err = segkp_fault(segkp->s_as->a_hat, segkp,
579 			    tp->t_swap, stack_size, F_SOFTLOCK, S_OTHER);
580 			rw_exit(&kas.a_lock);
581 
582 			/*
583 			 * Re-acquire the p_lock.
584 			 */
585 			mutex_enter(&pp->p_lock);
586 			if (err) {
587 				num_swapped_in = 0;
588 				break;
589 			} else {
590 #ifdef __sparc
591 				lwp_swapin(tp);
592 #endif /* __sparc */
593 				CPU_STATS_ADDQ(cpup, vm, swapin, 1);
594 				CPU_STATS_ADDQ(cpup, vm, pgswapin,
595 				    stack_pages);
596 
597 				pp->p_swapcnt--;
598 				pp->p_swrss -= stack_pages;
599 
600 				thread_lock(tp);
601 				tp->t_schedflag |= TS_LOAD;
602 				dq_sruninc(tp);
603 
604 				/* set swapin time */
605 				tp->t_stime = ddi_get_lbolt();
606 				thread_unlock(tp);
607 
608 				nswapped--;
609 				tot_swapped_in++;
610 				num_swapped_in++;
611 
612 				TRACE_2(TR_FAC_SCHED, TR_SWAPIN,
613 				    "swapin: pp %p stack_pages %lu",
614 				    pp, stack_pages);
615 				goto top;
616 			}
617 		}
618 		thread_unlock(tp);
619 	} while ((tp = tp->t_forw) != pp->p_tlist);
620 	return (num_swapped_in);
621 }
622 
623 /*
624  * Swap out lwps.  Returns nonzero on success (i.e., if at least one lwp is
625  * swapped out) and 0 on failure.
626  */
627 static int
628 swapout(proc_t *pp, uint_t *swrss, int swapflags)
629 {
630 	kthread_id_t tp;
631 	pgcnt_t ws_pages = 0;
632 	int err;
633 	int swapped_lwps = 0;
634 	struct as *as = pp->p_as;
635 	struct cpu *cpup = CPU;
636 	pri_t thread_pri;
637 
638 	ASSERT(MUTEX_HELD(&pp->p_lock));
639 
640 	if (pp->p_flag & SEXITING)
641 		return (0);
642 
643 top:
644 	tp = pp->p_tlist;
645 	do {
646 		klwp_t *lwp = ttolwp(tp);
647 
648 		/*
649 		 * Swapout eligible lwps (specified by the scheduling
650 		 * class) which don't have TS_DONT_SWAP set.  Set the
651 		 * "intent to swap" flag (TS_SWAPENQ) on threads
652 		 * which have TS_DONT_SWAP set so that they can be
653 		 * swapped if and when they reach a safe point.
654 		 */
655 		thread_lock(tp);
656 		thread_pri = CL_SWAPOUT(tp, swapflags);
657 		if (thread_pri != -1) {
658 			if (tp->t_schedflag & TS_DONT_SWAP) {
659 				tp->t_schedflag |= TS_SWAPENQ;
660 				tp->t_trapret = 1;
661 				aston(tp);
662 			} else {
663 				pgcnt_t stack_pages;
664 				size_t stack_size;
665 
666 				ASSERT((tp->t_schedflag &
667 				    (TS_DONT_SWAP | TS_LOAD)) == TS_LOAD);
668 
669 				if (lock_try(&tp->t_lock)) {
670 					/*
671 					 * Remove thread from the swap_queue.
672 					 */
673 					if (tp->t_schedflag & TS_ON_SWAPQ) {
674 						ASSERT(!(tp->t_schedflag &
675 						    TS_SWAPENQ));
676 						swapdeq(tp);
677 					} else if (tp->t_state == TS_RUN)
678 						dq_srundec(tp);
679 
680 					tp->t_schedflag &=
681 					    ~(TS_LOAD | TS_SWAPENQ);
682 					lock_clear(&tp->t_lock);
683 
684 					/*
685 					 * Set swapout time if the thread isn't
686 					 * sleeping.
687 					 */
688 					if (tp->t_state != TS_SLEEP)
689 						tp->t_stime = ddi_get_lbolt();
690 					thread_unlock(tp);
691 
692 					nswapped++;
693 					tot_swapped_out++;
694 
695 					lwp->lwp_ru.nswap++;
696 
697 					/*
698 					 * Now drop the p_lock since the
699 					 * stack needs to pushed out.
700 					 */
701 					mutex_exit(&pp->p_lock);
702 
703 					stack_size = swapsize(tp->t_swap);
704 					stack_pages = btopr(stack_size);
705 					ws_pages += stack_pages;
706 					/* Kernel probe */
707 					TNF_PROBE_4(swapout_lwp,
708 					    "vm swap swapout",
709 					    /* CSTYLED */,
710 					    tnf_pid, pid, pp->p_pid,
711 					    tnf_lwpid, lwpid, tp->t_tid,
712 					    tnf_kthread_id, tid, tp,
713 					    tnf_ulong, page_count,
714 					    stack_pages);
715 
716 					rw_enter(&kas.a_lock, RW_READER);
717 					err = segkp_fault(segkp->s_as->a_hat,
718 					    segkp, tp->t_swap, stack_size,
719 					    F_SOFTUNLOCK, S_WRITE);
720 					rw_exit(&kas.a_lock);
721 
722 					if (err) {
723 						cmn_err(CE_PANIC,
724 						    "swapout: segkp_fault "
725 						    "failed err: %d", err);
726 					}
727 					CPU_STATS_ADDQ(cpup,
728 					    vm, pgswapout, stack_pages);
729 
730 					mutex_enter(&pp->p_lock);
731 					pp->p_swapcnt++;
732 					swapped_lwps++;
733 					goto top;
734 				}
735 			}
736 		}
737 		thread_unlock(tp);
738 	} while ((tp = tp->t_forw) != pp->p_tlist);
739 
740 	/*
741 	 * Unload address space when all lwps are swapped out.
742 	 */
743 	if (pp->p_swapcnt == pp->p_lwpcnt) {
744 		size_t as_size = 0;
745 
746 		/*
747 		 * Avoid invoking as_swapout() if the process has
748 		 * no MMU resources since pageout will eventually
749 		 * steal pages belonging to this address space.  This
750 		 * saves CPU cycles as the number of pages that are
751 		 * potentially freed or pushed out by the segment
752 		 * swapout operation is very small.
753 		 */
754 		if (rm_asrss(pp->p_as) != 0)
755 			as_size = as_swapout(as);
756 
757 		CPU_STATS_ADDQ(cpup, vm, pgswapout, btop(as_size));
758 		CPU_STATS_ADDQ(cpup, vm, swapout, 1);
759 		ws_pages += btop(as_size);
760 
761 		TRACE_2(TR_FAC_SCHED, TR_SWAPOUT,
762 		    "swapout: pp %p pages_pushed %lu", pp, ws_pages);
763 		/* Kernel probe */
764 		TNF_PROBE_2(swapout_process, "vm swap swapout", /* CSTYLED */,
765 		    tnf_pid,	pid,		pp->p_pid,
766 		    tnf_ulong,	page_count,	ws_pages);
767 	}
768 	*swrss = ws_pages;
769 	return (swapped_lwps);
770 }
771 
772 void
773 swapout_lwp(klwp_t *lwp)
774 {
775 	kthread_id_t tp = curthread;
776 
777 	ASSERT(curthread == lwptot(lwp));
778 
779 	/*
780 	 * Don't insert the thread onto the swap queue if
781 	 * sufficient memory is available.
782 	 */
783 	if (avefree > desfree || avefree < desfree && freemem > desfree) {
784 		thread_lock(tp);
785 		tp->t_schedflag &= ~TS_SWAPENQ;
786 		thread_unlock(tp);
787 		return;
788 	}
789 
790 	/*
791 	 * Lock the thread, then move it to the swapped queue from the
792 	 * onproc queue and set its state to be TS_RUN.
793 	 */
794 	thread_lock(tp);
795 	ASSERT(tp->t_state == TS_ONPROC);
796 	if (tp->t_schedflag & TS_SWAPENQ) {
797 		tp->t_schedflag &= ~TS_SWAPENQ;
798 
799 		/*
800 		 * Set the state of this thread to be runnable
801 		 * and move it from the onproc queue to the swap queue.
802 		 */
803 		disp_swapped_enq(tp);
804 
805 		/*
806 		 * Insert the thread onto the swap queue.
807 		 */
808 		tp->t_link = tswap_queue;
809 		tswap_queue = tp;
810 		tp->t_schedflag |= TS_ON_SWAPQ;
811 
812 		thread_unlock_nopreempt(tp);
813 
814 		TRACE_1(TR_FAC_SCHED, TR_SWAPOUT_LWP, "swapout_lwp:%x", lwp);
815 
816 		swtch();
817 	} else {
818 		thread_unlock(tp);
819 	}
820 }
821 
822 /*
823  * Swap all threads on the swap queue.
824  */
825 static void
826 process_swap_queue(void)
827 {
828 	kthread_id_t tp;
829 	uint_t ws_pages;
830 	proc_t *pp;
831 	struct cpu *cpup = CPU;
832 	klwp_t *lwp;
833 	int err;
834 
835 	if (tswap_queue == NULL)
836 		return;
837 
838 	/*
839 	 * Acquire the "swapped_lock" which locks the swap queue,
840 	 * and unload the stacks of all threads on it.
841 	 */
842 	disp_lock_enter(&swapped_lock);
843 	while ((tp = tswap_queue) != NULL) {
844 		pgcnt_t stack_pages;
845 		size_t stack_size;
846 
847 		tswap_queue = tp->t_link;
848 		tp->t_link = NULL;
849 
850 		/*
851 		 * Drop the "dispatcher lock" before acquiring "t_lock"
852 		 * to avoid spinning on it since the thread at the front
853 		 * of the swap queue could be pinned before giving up
854 		 * its "t_lock" in resume.
855 		 */
856 		disp_lock_exit(&swapped_lock);
857 		lock_set(&tp->t_lock);
858 
859 		/*
860 		 * Now, re-acquire the "swapped_lock".  Acquiring this lock
861 		 * results in locking the thread since its dispatcher lock
862 		 * (t_lockp) is the "swapped_lock".
863 		 */
864 		disp_lock_enter(&swapped_lock);
865 		ASSERT(tp->t_state == TS_RUN);
866 		ASSERT(tp->t_schedflag & (TS_LOAD | TS_ON_SWAPQ));
867 
868 		tp->t_schedflag &= ~(TS_LOAD | TS_ON_SWAPQ);
869 		tp->t_stime = ddi_get_lbolt();		/* swapout time */
870 		disp_lock_exit(&swapped_lock);
871 		lock_clear(&tp->t_lock);
872 
873 		lwp = ttolwp(tp);
874 		lwp->lwp_ru.nswap++;
875 
876 		pp = ttoproc(tp);
877 		stack_size = swapsize(tp->t_swap);
878 		stack_pages = btopr(stack_size);
879 
880 		/* Kernel probe */
881 		TNF_PROBE_4(swapout_lwp, "vm swap swapout", /* CSTYLED */,
882 		    tnf_pid,		pid,		pp->p_pid,
883 		    tnf_lwpid,		lwpid,		tp->t_tid,
884 		    tnf_kthread_id,	tid,		tp,
885 		    tnf_ulong,		page_count,	stack_pages);
886 
887 		rw_enter(&kas.a_lock, RW_READER);
888 		err = segkp_fault(segkp->s_as->a_hat, segkp, tp->t_swap,
889 		    stack_size, F_SOFTUNLOCK, S_WRITE);
890 		rw_exit(&kas.a_lock);
891 
892 		if (err) {
893 			cmn_err(CE_PANIC,
894 			"process_swap_list: segkp_fault failed err: %d", err);
895 		}
896 		CPU_STATS_ADDQ(cpup, vm, pgswapout, stack_pages);
897 
898 		nswapped++;
899 		tot_swapped_out++;
900 		swapqswap++;
901 
902 		/*
903 		 * Don't need p_lock since the swapper is the only
904 		 * thread which increments/decrements p_swapcnt and p_swrss.
905 		 */
906 		ws_pages = stack_pages;
907 		pp->p_swapcnt++;
908 
909 		TRACE_1(TR_FAC_SCHED, TR_SWAPQ_LWP, "swaplist: pp %p", pp);
910 
911 		/*
912 		 * Unload address space when all lwps are swapped out.
913 		 */
914 		if (pp->p_swapcnt == pp->p_lwpcnt) {
915 			size_t as_size = 0;
916 
917 			if (rm_asrss(pp->p_as) != 0)
918 				as_size = as_swapout(pp->p_as);
919 
920 			CPU_STATS_ADDQ(cpup, vm, pgswapout,
921 			    btop(as_size));
922 			CPU_STATS_ADDQ(cpup, vm, swapout, 1);
923 
924 			ws_pages += btop(as_size);
925 
926 			TRACE_2(TR_FAC_SCHED, TR_SWAPQ_PROC,
927 			    "swaplist_proc: pp %p pages_pushed: %lu",
928 			    pp, ws_pages);
929 			/* Kernel probe */
930 			TNF_PROBE_2(swapout_process, "vm swap swapout",
931 			    /* CSTYLED */,
932 			    tnf_pid,	pid,		pp->p_pid,
933 			    tnf_ulong,	page_count,	ws_pages);
934 		}
935 		pp->p_swrss += ws_pages;
936 		disp_lock_enter(&swapped_lock);
937 	}
938 	disp_lock_exit(&swapped_lock);
939 }
940