xref: /titanic_41/usr/src/uts/common/os/sched.c (revision 6f45ec7b0b964c3be967c4880e8867ac1e7763a5)
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 2006 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 
31 #pragma ident	"%Z%%M%	%I%	%E% SMI"
32 
33 #include <sys/param.h>
34 #include <sys/types.h>
35 #include <sys/sysmacros.h>
36 #include <sys/systm.h>
37 #include <sys/proc.h>
38 #include <sys/cpuvar.h>
39 #include <sys/var.h>
40 #include <sys/tuneable.h>
41 #include <sys/cmn_err.h>
42 #include <sys/buf.h>
43 #include <sys/disp.h>
44 #include <sys/vmsystm.h>
45 #include <sys/vmparam.h>
46 #include <sys/class.h>
47 #include <sys/vtrace.h>
48 #include <sys/modctl.h>
49 #include <sys/debug.h>
50 #include <sys/tnf_probe.h>
51 #include <sys/procfs.h>
52 
53 #include <vm/seg.h>
54 #include <vm/seg_kp.h>
55 #include <vm/as.h>
56 #include <vm/rm.h>
57 #include <vm/seg_kmem.h>
58 #include <sys/callb.h>
59 
60 /*
61  * The swapper sleeps on runout when there is no one to swap in.
62  * It sleeps on runin when it could not find space to swap someone
63  * in or after swapping someone in.
64  */
65 char	runout;
66 char	runin;
67 char	wake_sched;	/* flag tells clock to wake swapper on next tick */
68 char	wake_sched_sec;	/* flag tells clock to wake swapper after a second */
69 
70 /*
71  * The swapper swaps processes to reduce memory demand and runs
72  * when avefree < desfree.  The swapper resorts to SOFTSWAP when
73  * avefree < desfree which results in swapping out all processes
74  * sleeping for more than maxslp seconds.  HARDSWAP occurs when the
75  * system is on the verge of thrashing and this results in swapping
76  * out runnable threads or threads sleeping for less than maxslp secs.
77  *
78  * The swapper runs through all the active processes in the system
79  * and invokes the scheduling class specific swapin/swapout routine
80  * for every thread in the process to obtain an effective priority
81  * for the process.  A priority of -1 implies that the thread isn't
82  * swappable.  This effective priority is used to find the most
83  * eligible process to swapout or swapin.
84  *
85  * NOTE:  Threads which have been swapped are not linked on any
86  *	  queue and their dispatcher lock points at the "swapped_lock".
87  *
88  * Processes containing threads with the TS_DONT_SWAP flag set cannot be
89  * swapped out immediately by the swapper.  This is due to the fact that
90  * such threads may be holding locks which may be needed by the swapper
91  * to push its pages out.  The TS_SWAPENQ flag is set on such threads
92  * to prevent them running in user mode.  When such threads reach a
93  * safe point (i.e., are not holding any locks - CL_TRAPRET), they
94  * queue themseleves onto the swap queue which is processed by the
95  * swapper.  This results in reducing memory demand when the system
96  * is desparate for memory as the thread can't run in user mode.
97  *
98  * The swap queue consists of threads, linked via t_link, which are
99  * haven't been swapped, are runnable but not on the run queue.  The
100  * swap queue is protected by the "swapped_lock".  The dispatcher
101  * lock (t_lockp) of all threads on the swap queue points at the
102  * "swapped_lock".  Thus, the entire queue and/or threads on the
103  * queue can be locked by acquiring "swapped_lock".
104  */
105 static kthread_t *tswap_queue;
106 extern disp_lock_t swapped_lock; /* protects swap queue and threads on it */
107 
108 int	maxslp = 0;
109 pgcnt_t	avefree;	/* 5 sec moving average of free memory */
110 pgcnt_t	avefree30;	/* 30 sec moving average of free memory */
111 
112 /*
113  * Minimum size used to decide if sufficient memory is available
114  * before a process is swapped in.  This is necessary since in most
115  * cases the actual size of a process (p_swrss) being swapped in
116  * is usually 2 pages (kernel stack pages).  This is due to the fact
117  * almost all user pages of a process are stolen by pageout before
118  * the swapper decides to swapout it out.
119  */
120 int	min_procsize = 12;
121 
122 static int	swapin(proc_t *);
123 static int	swapout(proc_t *, uint_t *, int);
124 static void	process_swap_queue();
125 
126 #ifdef __sparc
127 extern void lwp_swapin(kthread_t *);
128 #endif /* __sparc */
129 
130 /*
131  * Counters to keep track of the number of swapins or swapouts.
132  */
133 uint_t tot_swapped_in, tot_swapped_out;
134 uint_t softswap, hardswap, swapqswap;
135 
136 /*
137  * Macro to determine if a process is eligble to be swapped.
138  */
139 #define	not_swappable(p)					\
140 	(((p)->p_flag & SSYS) || (p)->p_stat == SIDL ||		\
141 	    (p)->p_stat == SZOMB || (p)->p_as == NULL ||	\
142 	    (p)->p_as == &kas)
143 
144 /*
145  * Memory scheduler.
146  */
147 void
148 sched()
149 {
150 	kthread_id_t	t;
151 	pri_t		proc_pri;
152 	pri_t		thread_pri;
153 	pri_t		swapin_pri;
154 	int		desperate;
155 	pgcnt_t		needs;
156 	int		divisor;
157 	proc_t		*prp;
158 	proc_t		*swapout_prp;
159 	proc_t		*swapin_prp;
160 	spgcnt_t	avail;
161 	int		chosen_pri;
162 	time_t		swapout_time;
163 	time_t		swapin_proc_time;
164 	callb_cpr_t	cprinfo;
165 	kmutex_t	swap_cpr_lock;
166 
167 	mutex_init(&swap_cpr_lock, NULL, MUTEX_DEFAULT, NULL);
168 	CALLB_CPR_INIT(&cprinfo, &swap_cpr_lock, callb_generic_cpr, "sched");
169 	if (maxslp == 0)
170 		maxslp = MAXSLP;
171 loop:
172 	needs = 0;
173 	desperate = 0;
174 
175 	swapin_pri = v.v_nglobpris;
176 	swapin_prp = NULL;
177 	chosen_pri = -1;
178 
179 	process_swap_queue();
180 
181 	/*
182 	 * Set desperate if
183 	 * 	1.  At least 2 runnable processes (on average).
184 	 *	2.  Short (5 sec) and longer (30 sec) average is less
185 	 *	    than minfree and desfree respectively.
186 	 *	3.  Pagein + pageout rate is excessive.
187 	 */
188 	if (avenrun[0] >= 2 * FSCALE &&
189 	    (MAX(avefree, avefree30) < desfree) &&
190 	    (pginrate + pgoutrate > maxpgio || avefree < minfree)) {
191 		TRACE_4(TR_FAC_SCHED, TR_DESPERATE,
192 		    "desp:avefree: %d, avefree30: %d, freemem: %d"
193 		    " pginrate: %d\n", avefree, avefree30, freemem, pginrate);
194 		desperate = 1;
195 		goto unload;
196 	}
197 
198 	/*
199 	 * Search list of processes to swapin and swapout deadwood.
200 	 */
201 	swapin_proc_time = 0;
202 top:
203 	mutex_enter(&pidlock);
204 	for (prp = practive; prp != NULL; prp = prp->p_next) {
205 		if (not_swappable(prp))
206 			continue;
207 
208 		/*
209 		 * Look at processes with at least one swapped lwp.
210 		 */
211 		if (prp->p_swapcnt) {
212 			time_t proc_time;
213 
214 			/*
215 			 * Higher priority processes are good candidates
216 			 * to swapin.
217 			 */
218 			mutex_enter(&prp->p_lock);
219 			proc_pri = -1;
220 			t = prp->p_tlist;
221 			proc_time = 0;
222 			do {
223 				if (t->t_schedflag & TS_LOAD)
224 					continue;
225 
226 				thread_lock(t);
227 				thread_pri = CL_SWAPIN(t, 0);
228 				thread_unlock(t);
229 
230 				if (t->t_stime - proc_time > 0)
231 					proc_time = t->t_stime;
232 				if (thread_pri > proc_pri)
233 					proc_pri = thread_pri;
234 			} while ((t = t->t_forw) != prp->p_tlist);
235 			mutex_exit(&prp->p_lock);
236 
237 			if (proc_pri == -1)
238 				continue;
239 
240 			TRACE_3(TR_FAC_SCHED, TR_CHOOSE_SWAPIN,
241 			    "prp %p epri %d proc_time %d",
242 			    prp, proc_pri, proc_time);
243 
244 			/*
245 			 * Swapin processes with a high effective priority.
246 			 */
247 			if (swapin_prp == NULL || proc_pri > chosen_pri) {
248 				swapin_prp = prp;
249 				chosen_pri = proc_pri;
250 				swapin_pri = proc_pri;
251 				swapin_proc_time = proc_time;
252 			}
253 		} else {
254 			/*
255 			 * No need to soft swap if we have sufficient
256 			 * memory.
257 			 */
258 			if (avefree > desfree ||
259 			    avefree < desfree && freemem > desfree)
260 				continue;
261 
262 			/*
263 			 * Skip processes that are exiting
264 			 * or whose address spaces are locked.
265 			 */
266 			mutex_enter(&prp->p_lock);
267 			if ((prp->p_flag & SEXITING) ||
268 			    (prp->p_as != NULL && AS_ISPGLCK(prp->p_as))) {
269 				mutex_exit(&prp->p_lock);
270 				continue;
271 			}
272 
273 			/*
274 			 * Softswapping to kick out deadwood.
275 			 */
276 			proc_pri = -1;
277 			t = prp->p_tlist;
278 			do {
279 				if ((t->t_schedflag & (TS_SWAPENQ |
280 				    TS_ON_SWAPQ | TS_LOAD)) != TS_LOAD)
281 					continue;
282 
283 				thread_lock(t);
284 				thread_pri = CL_SWAPOUT(t, SOFTSWAP);
285 				thread_unlock(t);
286 				if (thread_pri > proc_pri)
287 					proc_pri = thread_pri;
288 			} while ((t = t->t_forw) != prp->p_tlist);
289 
290 			if (proc_pri != -1) {
291 				uint_t swrss;
292 
293 				mutex_exit(&pidlock);
294 
295 				TRACE_1(TR_FAC_SCHED, TR_SOFTSWAP,
296 				    "softswap:prp %p", prp);
297 
298 				(void) swapout(prp, &swrss, SOFTSWAP);
299 				softswap++;
300 				prp->p_swrss += swrss;
301 				mutex_exit(&prp->p_lock);
302 				goto top;
303 			}
304 			mutex_exit(&prp->p_lock);
305 		}
306 	}
307 	if (swapin_prp != NULL)
308 		mutex_enter(&swapin_prp->p_lock);
309 	mutex_exit(&pidlock);
310 
311 	if (swapin_prp == NULL) {
312 		TRACE_3(TR_FAC_SCHED, TR_RUNOUT,
313 		"schedrunout:runout nswapped: %d, avefree: %ld freemem: %ld",
314 		    nswapped, avefree, freemem);
315 
316 		t = curthread;
317 		thread_lock(t);
318 		runout++;
319 		t->t_schedflag |= (TS_ALLSTART & ~TS_CSTART);
320 		t->t_whystop = PR_SUSPENDED;
321 		t->t_whatstop = SUSPEND_NORMAL;
322 		(void) new_mstate(t, LMS_SLEEP);
323 		mutex_enter(&swap_cpr_lock);
324 		CALLB_CPR_SAFE_BEGIN(&cprinfo);
325 		mutex_exit(&swap_cpr_lock);
326 		thread_stop(t);		/* change state and drop lock */
327 		swtch();
328 		mutex_enter(&swap_cpr_lock);
329 		CALLB_CPR_SAFE_END(&cprinfo, &swap_cpr_lock);
330 		mutex_exit(&swap_cpr_lock);
331 		goto loop;
332 	}
333 
334 	/*
335 	 * Decide how deserving this process is to be brought in.
336 	 * Needs is an estimate of how much core the process will
337 	 * need.  If the process has been out for a while, then we
338 	 * will bring it in with 1/2 the core needed, otherwise
339 	 * we are conservative.
340 	 */
341 	divisor = 1;
342 	swapout_time = (lbolt - swapin_proc_time) / hz;
343 	if (swapout_time > maxslp / 2)
344 		divisor = 2;
345 
346 	needs = MIN(swapin_prp->p_swrss, lotsfree);
347 	needs = MAX(needs, min_procsize);
348 	needs = needs / divisor;
349 
350 	/*
351 	 * Use freemem, since we want processes to be swapped
352 	 * in quickly.
353 	 */
354 	avail = freemem - deficit;
355 	if (avail > (spgcnt_t)needs) {
356 		deficit += needs;
357 
358 		TRACE_2(TR_FAC_SCHED, TR_SWAPIN_VALUES,
359 		    "swapin_values: prp %p needs %lu", swapin_prp, needs);
360 
361 		if (swapin(swapin_prp)) {
362 			mutex_exit(&swapin_prp->p_lock);
363 			goto loop;
364 		}
365 		deficit -= MIN(needs, deficit);
366 		mutex_exit(&swapin_prp->p_lock);
367 	} else {
368 		mutex_exit(&swapin_prp->p_lock);
369 		/*
370 		 * If deficit is high, too many processes have been
371 		 * swapped in so wait a sec before attempting to
372 		 * swapin more.
373 		 */
374 		if (freemem > needs) {
375 			TRACE_2(TR_FAC_SCHED, TR_HIGH_DEFICIT,
376 			    "deficit: prp %p needs %lu", swapin_prp, needs);
377 			goto block;
378 		}
379 	}
380 
381 	TRACE_2(TR_FAC_SCHED, TR_UNLOAD,
382 	    "unload: prp %p needs %lu", swapin_prp, needs);
383 
384 unload:
385 	/*
386 	 * Unload all unloadable modules, free all other memory
387 	 * resources we can find, then look for a thread to hardswap.
388 	 */
389 	modreap();
390 	segkp_cache_free();
391 
392 	swapout_prp = NULL;
393 	mutex_enter(&pidlock);
394 	for (prp = practive; prp != NULL; prp = prp->p_next) {
395 
396 		/*
397 		 * No need to soft swap if we have sufficient
398 		 * memory.
399 		 */
400 		if (not_swappable(prp))
401 			continue;
402 
403 		if (avefree > minfree ||
404 		    avefree < minfree && freemem > desfree) {
405 			swapout_prp = NULL;
406 			break;
407 		}
408 
409 		/*
410 		 * Skip processes that are exiting
411 		 * or whose address spaces are locked.
412 		 */
413 		mutex_enter(&prp->p_lock);
414 		if ((prp->p_flag & SEXITING) ||
415 		    (prp->p_as != NULL && AS_ISPGLCK(prp->p_as))) {
416 			mutex_exit(&prp->p_lock);
417 			continue;
418 		}
419 
420 		proc_pri = -1;
421 		t = prp->p_tlist;
422 		do {
423 			if ((t->t_schedflag & (TS_SWAPENQ |
424 			    TS_ON_SWAPQ | TS_LOAD)) != TS_LOAD)
425 				continue;
426 
427 			thread_lock(t);
428 			thread_pri = CL_SWAPOUT(t, HARDSWAP);
429 			thread_unlock(t);
430 			if (thread_pri > proc_pri)
431 				proc_pri = thread_pri;
432 		} while ((t = t->t_forw) != prp->p_tlist);
433 
434 		mutex_exit(&prp->p_lock);
435 		if (proc_pri == -1)
436 			continue;
437 
438 		/*
439 		 * Swapout processes sleeping with a lower priority
440 		 * than the one currently being swapped in, if any.
441 		 */
442 		if (swapin_prp == NULL || swapin_pri > proc_pri) {
443 			TRACE_2(TR_FAC_SCHED, TR_CHOOSE_SWAPOUT,
444 			    "hardswap: prp %p needs %lu", prp, needs);
445 
446 			if (swapout_prp == NULL || proc_pri < chosen_pri) {
447 				swapout_prp = prp;
448 				chosen_pri = proc_pri;
449 			}
450 		}
451 	}
452 
453 	/*
454 	 * Acquire the "p_lock" before dropping "pidlock"
455 	 * to prevent the proc structure from being freed
456 	 * if the process exits before swapout completes.
457 	 */
458 	if (swapout_prp != NULL)
459 		mutex_enter(&swapout_prp->p_lock);
460 	mutex_exit(&pidlock);
461 
462 	if ((prp = swapout_prp) != NULL) {
463 		uint_t swrss = 0;
464 		int swapped;
465 
466 		swapped = swapout(prp, &swrss, HARDSWAP);
467 		if (swapped) {
468 			/*
469 			 * If desperate, we want to give the space obtained
470 			 * by swapping this process out to processes in core,
471 			 * so we give them a chance by increasing deficit.
472 			 */
473 			prp->p_swrss += swrss;
474 			if (desperate)
475 				deficit += MIN(prp->p_swrss, lotsfree);
476 			hardswap++;
477 		}
478 		mutex_exit(&swapout_prp->p_lock);
479 
480 		if (swapped)
481 			goto loop;
482 	}
483 
484 	/*
485 	 * Delay for 1 second and look again later.
486 	 */
487 	TRACE_3(TR_FAC_SCHED, TR_RUNIN,
488 	    "schedrunin:runin nswapped: %d, avefree: %ld freemem: %ld",
489 	    nswapped, avefree, freemem);
490 
491 block:
492 	t = curthread;
493 	thread_lock(t);
494 	runin++;
495 	t->t_schedflag |= (TS_ALLSTART & ~TS_CSTART);
496 	t->t_whystop = PR_SUSPENDED;
497 	t->t_whatstop = SUSPEND_NORMAL;
498 	(void) new_mstate(t, LMS_SLEEP);
499 	mutex_enter(&swap_cpr_lock);
500 	CALLB_CPR_SAFE_BEGIN(&cprinfo);
501 	mutex_exit(&swap_cpr_lock);
502 	thread_stop(t);		/* change to stop state and drop lock */
503 	swtch();
504 	mutex_enter(&swap_cpr_lock);
505 	CALLB_CPR_SAFE_END(&cprinfo, &swap_cpr_lock);
506 	mutex_exit(&swap_cpr_lock);
507 	goto loop;
508 }
509 
510 /*
511  * Remove the specified thread from the swap queue.
512  */
513 static void
514 swapdeq(kthread_id_t tp)
515 {
516 	kthread_id_t *tpp;
517 
518 	ASSERT(THREAD_LOCK_HELD(tp));
519 	ASSERT(tp->t_schedflag & TS_ON_SWAPQ);
520 
521 	tpp = &tswap_queue;
522 	for (;;) {
523 		ASSERT(*tpp != NULL);
524 		if (*tpp == tp)
525 			break;
526 		tpp = &(*tpp)->t_link;
527 	}
528 	*tpp = tp->t_link;
529 	tp->t_schedflag &= ~TS_ON_SWAPQ;
530 }
531 
532 /*
533  * Swap in lwps.  Returns nonzero on success (i.e., if at least one lwp is
534  * swapped in) and 0 on failure.
535  */
536 static int
537 swapin(proc_t *pp)
538 {
539 	kthread_id_t tp;
540 	int err;
541 	int num_swapped_in = 0;
542 	struct cpu *cpup = CPU;
543 	pri_t thread_pri;
544 
545 	ASSERT(MUTEX_HELD(&pp->p_lock));
546 	ASSERT(pp->p_swapcnt);
547 
548 top:
549 	tp = pp->p_tlist;
550 	do {
551 		/*
552 		 * Only swapin eligible lwps (specified by the scheduling
553 		 * class) which are unloaded and ready to run.
554 		 */
555 		thread_lock(tp);
556 		thread_pri = CL_SWAPIN(tp, 0);
557 		if (thread_pri != -1 && tp->t_state == TS_RUN &&
558 		    (tp->t_schedflag & TS_LOAD) == 0) {
559 			size_t stack_size;
560 			pgcnt_t stack_pages;
561 
562 			ASSERT((tp->t_schedflag & TS_ON_SWAPQ) == 0);
563 
564 			thread_unlock(tp);
565 			/*
566 			 * Now drop the p_lock since the stack needs
567 			 * to brought in.
568 			 */
569 			mutex_exit(&pp->p_lock);
570 
571 			stack_size = swapsize(tp->t_swap);
572 			stack_pages = btopr(stack_size);
573 			/* Kernel probe */
574 			TNF_PROBE_4(swapin_lwp, "vm swap swapin", /* CSTYLED */,
575 				tnf_pid,	pid,		pp->p_pid,
576 				tnf_lwpid,	lwpid,		tp->t_tid,
577 				tnf_kthread_id,	tid,		tp,
578 				tnf_ulong,	page_count,	stack_pages);
579 
580 			rw_enter(&kas.a_lock, RW_READER);
581 			err = segkp_fault(segkp->s_as->a_hat, segkp,
582 			    tp->t_swap, stack_size, F_SOFTLOCK, S_OTHER);
583 			rw_exit(&kas.a_lock);
584 
585 			/*
586 			 * Re-acquire the p_lock.
587 			 */
588 			mutex_enter(&pp->p_lock);
589 			if (err) {
590 				num_swapped_in = 0;
591 				break;
592 			} else {
593 #ifdef __sparc
594 				lwp_swapin(tp);
595 #endif /* __sparc */
596 				CPU_STATS_ADDQ(cpup, vm, swapin, 1);
597 				CPU_STATS_ADDQ(cpup, vm, pgswapin,
598 				    stack_pages);
599 
600 				pp->p_swapcnt--;
601 				pp->p_swrss -= stack_pages;
602 
603 				thread_lock(tp);
604 				tp->t_schedflag |= TS_LOAD;
605 				dq_sruninc(tp);
606 
607 				tp->t_stime = lbolt;	/* set swapin time */
608 				thread_unlock(tp);
609 
610 				nswapped--;
611 				tot_swapped_in++;
612 				num_swapped_in++;
613 
614 				TRACE_2(TR_FAC_SCHED, TR_SWAPIN,
615 				    "swapin: pp %p stack_pages %lu",
616 				    pp, stack_pages);
617 				goto top;
618 			}
619 		}
620 		thread_unlock(tp);
621 	} while ((tp = tp->t_forw) != pp->p_tlist);
622 	return (num_swapped_in);
623 }
624 
625 /*
626  * Swap out lwps.  Returns nonzero on success (i.e., if at least one lwp is
627  * swapped out) and 0 on failure.
628  */
629 static int
630 swapout(proc_t *pp, uint_t *swrss, int swapflags)
631 {
632 	kthread_id_t tp;
633 	pgcnt_t ws_pages = 0;
634 	int err;
635 	int swapped_lwps = 0;
636 	struct as *as = pp->p_as;
637 	struct cpu *cpup = CPU;
638 	pri_t thread_pri;
639 
640 	ASSERT(MUTEX_HELD(&pp->p_lock));
641 
642 	if (pp->p_flag & SEXITING)
643 		return (0);
644 
645 top:
646 	tp = pp->p_tlist;
647 	do {
648 		klwp_t *lwp = ttolwp(tp);
649 
650 		/*
651 		 * Swapout eligible lwps (specified by the scheduling
652 		 * class) which don't have TS_DONT_SWAP set.  Set the
653 		 * "intent to swap" flag (TS_SWAPENQ) on threads
654 		 * which have TS_DONT_SWAP set so that they can be
655 		 * swapped if and when they reach a safe point.
656 		 */
657 		thread_lock(tp);
658 		thread_pri = CL_SWAPOUT(tp, swapflags);
659 		if (thread_pri != -1) {
660 			if (tp->t_schedflag & TS_DONT_SWAP) {
661 				tp->t_schedflag |= TS_SWAPENQ;
662 				tp->t_trapret = 1;
663 				aston(tp);
664 			} else {
665 				pgcnt_t stack_pages;
666 				size_t stack_size;
667 
668 				ASSERT((tp->t_schedflag &
669 				    (TS_DONT_SWAP | TS_LOAD)) == TS_LOAD);
670 
671 				if (lock_try(&tp->t_lock)) {
672 					/*
673 					 * Remove thread from the swap_queue.
674 					 */
675 					if (tp->t_schedflag & TS_ON_SWAPQ) {
676 						ASSERT(!(tp->t_schedflag &
677 						    TS_SWAPENQ));
678 						swapdeq(tp);
679 					} else if (tp->t_state == TS_RUN)
680 						dq_srundec(tp);
681 
682 					tp->t_schedflag &=
683 					    ~(TS_LOAD | TS_SWAPENQ);
684 					lock_clear(&tp->t_lock);
685 
686 					/*
687 					 * Set swapout time if the thread isn't
688 					 * sleeping.
689 					 */
690 					if (tp->t_state != TS_SLEEP)
691 						tp->t_stime = lbolt;
692 					thread_unlock(tp);
693 
694 					nswapped++;
695 					tot_swapped_out++;
696 
697 					lwp->lwp_ru.nswap++;
698 
699 					/*
700 					 * Now drop the p_lock since the
701 					 * stack needs to pushed out.
702 					 */
703 					mutex_exit(&pp->p_lock);
704 
705 					stack_size = swapsize(tp->t_swap);
706 					stack_pages = btopr(stack_size);
707 					ws_pages += stack_pages;
708 					/* Kernel probe */
709 					TNF_PROBE_4(swapout_lwp,
710 						"vm swap swapout",
711 						/* CSTYLED */,
712 						tnf_pid, pid, pp->p_pid,
713 						tnf_lwpid, lwpid, tp->t_tid,
714 						tnf_kthread_id, tid, tp,
715 						tnf_ulong, page_count,
716 							stack_pages);
717 
718 					rw_enter(&kas.a_lock, RW_READER);
719 					err = segkp_fault(segkp->s_as->a_hat,
720 					    segkp, tp->t_swap, stack_size,
721 					    F_SOFTUNLOCK, S_WRITE);
722 					rw_exit(&kas.a_lock);
723 
724 					if (err) {
725 						cmn_err(CE_PANIC,
726 						    "swapout: segkp_fault "
727 						    "failed err: %d", err);
728 					}
729 					CPU_STATS_ADDQ(cpup,
730 					    vm, pgswapout, stack_pages);
731 
732 					mutex_enter(&pp->p_lock);
733 					pp->p_swapcnt++;
734 					swapped_lwps++;
735 					goto top;
736 				}
737 			}
738 		}
739 		thread_unlock(tp);
740 	} while ((tp = tp->t_forw) != pp->p_tlist);
741 
742 	/*
743 	 * Unload address space when all lwps are swapped out.
744 	 */
745 	if (pp->p_swapcnt == pp->p_lwpcnt) {
746 		size_t as_size = 0;
747 
748 		/*
749 		 * Avoid invoking as_swapout() if the process has
750 		 * no MMU resources since pageout will eventually
751 		 * steal pages belonging to this address space.  This
752 		 * saves CPU cycles as the number of pages that are
753 		 * potentially freed or pushed out by the segment
754 		 * swapout operation is very small.
755 		 */
756 		if (rm_asrss(pp->p_as) != 0)
757 			as_size = as_swapout(as);
758 
759 		CPU_STATS_ADDQ(cpup, vm, pgswapout, btop(as_size));
760 		CPU_STATS_ADDQ(cpup, vm, swapout, 1);
761 		ws_pages += btop(as_size);
762 
763 		TRACE_2(TR_FAC_SCHED, TR_SWAPOUT,
764 		    "swapout: pp %p pages_pushed %lu", pp, ws_pages);
765 		/* Kernel probe */
766 		TNF_PROBE_2(swapout_process, "vm swap swapout", /* CSTYLED */,
767 			tnf_pid,	pid,		pp->p_pid,
768 			tnf_ulong,	page_count,	ws_pages);
769 	}
770 	*swrss = ws_pages;
771 	return (swapped_lwps);
772 }
773 
774 void
775 swapout_lwp(klwp_t *lwp)
776 {
777 	kthread_id_t tp = curthread;
778 
779 	ASSERT(curthread == lwptot(lwp));
780 
781 	/*
782 	 * Don't insert the thread onto the swap queue if
783 	 * sufficient memory is available.
784 	 */
785 	if (avefree > desfree || avefree < desfree && freemem > desfree) {
786 		thread_lock(tp);
787 		tp->t_schedflag &= ~TS_SWAPENQ;
788 		thread_unlock(tp);
789 		return;
790 	}
791 
792 	/*
793 	 * Lock the thread, then move it to the swapped queue from the
794 	 * onproc queue and set its state to be TS_RUN.
795 	 */
796 	thread_lock(tp);
797 	ASSERT(tp->t_state == TS_ONPROC);
798 	if (tp->t_schedflag & TS_SWAPENQ) {
799 		tp->t_schedflag &= ~TS_SWAPENQ;
800 
801 		/*
802 		 * Set the state of this thread to be runnable
803 		 * and move it from the onproc queue to the swap queue.
804 		 */
805 		disp_swapped_enq(tp);
806 
807 		/*
808 		 * Insert the thread onto the swap queue.
809 		 */
810 		tp->t_link = tswap_queue;
811 		tswap_queue = tp;
812 		tp->t_schedflag |= TS_ON_SWAPQ;
813 
814 		thread_unlock_nopreempt(tp);
815 
816 		TRACE_1(TR_FAC_SCHED, TR_SWAPOUT_LWP, "swapout_lwp:%x", lwp);
817 
818 		swtch();
819 	} else {
820 		thread_unlock(tp);
821 	}
822 }
823 
824 /*
825  * Swap all threads on the swap queue.
826  */
827 static void
828 process_swap_queue(void)
829 {
830 	kthread_id_t tp;
831 	uint_t ws_pages;
832 	proc_t *pp;
833 	struct cpu *cpup = CPU;
834 	klwp_t *lwp;
835 	int err;
836 
837 	if (tswap_queue == NULL)
838 		return;
839 
840 	/*
841 	 * Acquire the "swapped_lock" which locks the swap queue,
842 	 * and unload the stacks of all threads on it.
843 	 */
844 	disp_lock_enter(&swapped_lock);
845 	while ((tp = tswap_queue) != NULL) {
846 		pgcnt_t stack_pages;
847 		size_t stack_size;
848 
849 		tswap_queue = tp->t_link;
850 		tp->t_link = NULL;
851 
852 		/*
853 		 * Drop the "dispatcher lock" before acquiring "t_lock"
854 		 * to avoid spinning on it since the thread at the front
855 		 * of the swap queue could be pinned before giving up
856 		 * its "t_lock" in resume.
857 		 */
858 		disp_lock_exit(&swapped_lock);
859 		lock_set(&tp->t_lock);
860 
861 		/*
862 		 * Now, re-acquire the "swapped_lock".  Acquiring this lock
863 		 * results in locking the thread since its dispatcher lock
864 		 * (t_lockp) is the "swapped_lock".
865 		 */
866 		disp_lock_enter(&swapped_lock);
867 		ASSERT(tp->t_state == TS_RUN);
868 		ASSERT(tp->t_schedflag & (TS_LOAD | TS_ON_SWAPQ));
869 
870 		tp->t_schedflag &= ~(TS_LOAD | TS_ON_SWAPQ);
871 		tp->t_stime = lbolt;		/* swapout time */
872 		disp_lock_exit(&swapped_lock);
873 		lock_clear(&tp->t_lock);
874 
875 		lwp = ttolwp(tp);
876 		lwp->lwp_ru.nswap++;
877 
878 		pp = ttoproc(tp);
879 		stack_size = swapsize(tp->t_swap);
880 		stack_pages = btopr(stack_size);
881 
882 		/* Kernel probe */
883 		TNF_PROBE_4(swapout_lwp, "vm swap swapout", /* CSTYLED */,
884 			tnf_pid,	pid,		pp->p_pid,
885 			tnf_lwpid,	lwpid,		tp->t_tid,
886 			tnf_kthread_id,	tid,		tp,
887 			tnf_ulong,	page_count,	stack_pages);
888 
889 		rw_enter(&kas.a_lock, RW_READER);
890 		err = segkp_fault(segkp->s_as->a_hat, segkp, tp->t_swap,
891 		    stack_size, F_SOFTUNLOCK, S_WRITE);
892 		rw_exit(&kas.a_lock);
893 
894 		if (err) {
895 			cmn_err(CE_PANIC,
896 			"process_swap_list: segkp_fault failed err: %d", err);
897 		}
898 		CPU_STATS_ADDQ(cpup, vm, pgswapout, stack_pages);
899 
900 		nswapped++;
901 		tot_swapped_out++;
902 		swapqswap++;
903 
904 		/*
905 		 * Don't need p_lock since the swapper is the only
906 		 * thread which increments/decrements p_swapcnt and p_swrss.
907 		 */
908 		ws_pages = stack_pages;
909 		pp->p_swapcnt++;
910 
911 		TRACE_1(TR_FAC_SCHED, TR_SWAPQ_LWP, "swaplist: pp %p", pp);
912 
913 		/*
914 		 * Unload address space when all lwps are swapped out.
915 		 */
916 		if (pp->p_swapcnt == pp->p_lwpcnt) {
917 			size_t as_size = 0;
918 
919 			if (rm_asrss(pp->p_as) != 0)
920 				as_size = as_swapout(pp->p_as);
921 
922 			CPU_STATS_ADDQ(cpup, vm, pgswapout,
923 			    btop(as_size));
924 			CPU_STATS_ADDQ(cpup, vm, swapout, 1);
925 
926 			ws_pages += btop(as_size);
927 
928 			TRACE_2(TR_FAC_SCHED, TR_SWAPQ_PROC,
929 			    "swaplist_proc: pp %p pages_pushed: %lu",
930 			    pp, ws_pages);
931 			/* Kernel probe */
932 			TNF_PROBE_2(swapout_process, "vm swap swapout",
933 				/* CSTYLED */,
934 				tnf_pid,	pid,		pp->p_pid,
935 				tnf_ulong,	page_count,	ws_pages);
936 		}
937 		pp->p_swrss += ws_pages;
938 		disp_lock_enter(&swapped_lock);
939 	}
940 	disp_lock_exit(&swapped_lock);
941 }
942