1/*
2 * Copyright (c) 1991, 1993
3 * The Regents of the University of California. All rights reserved.
4 *
5 * This code is derived from software contributed to Berkeley by
6 * The Mach Operating System project at Carnegie-Mellon University.
7 *
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
10 * are met:
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
16 * 3. All advertising materials mentioning features or use of this software
17 * must display the following acknowledgement:
18 * This product includes software developed by the University of
19 * California, Berkeley and its contributors.
20 * 4. Neither the name of the University nor the names of its contributors
21 * may be used to endorse or promote products derived from this software
22 * without specific prior written permission.
23 *
24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34 * SUCH DAMAGE.
35 *
36 * from: @(#)vm_glue.c 8.6 (Berkeley) 1/5/94
37 *
38 *
39 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
40 * All rights reserved.
41 *
42 * Permission to use, copy, modify and distribute this software and
43 * its documentation is hereby granted, provided that both the copyright
44 * notice and this permission notice appear in all copies of the
45 * software, derivative works or modified versions, and any portions
46 * thereof, and that both notices appear in supporting documentation.
47 *
48 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
49 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
50 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
51 *
52 * Carnegie Mellon requests users of this software to return to
53 *
54 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
55 * School of Computer Science
56 * Carnegie Mellon University
57 * Pittsburgh PA 15213-3890
58 *
59 * any improvements or extensions that they make and grant Carnegie the
60 * rights to redistribute these changes.
61 *
62 * $FreeBSD: head/sys/vm/vm_glue.c 103216 2002-09-11 08:13:56Z julian $
63 */
64
65#include "opt_vm.h"
66
67#include <sys/param.h>
68#include <sys/systm.h>
69#include <sys/lock.h>
70#include <sys/mutex.h>
71#include <sys/proc.h>
72#include <sys/resourcevar.h>
73#include <sys/shm.h>
74#include <sys/vmmeter.h>
75#include <sys/sx.h>
76#include <sys/sysctl.h>
77
78#include <sys/kernel.h>
79#include <sys/ktr.h>
80#include <sys/unistd.h>
81
82#include <machine/limits.h>
83
84#include <vm/vm.h>
85#include <vm/vm_param.h>
86#include <vm/pmap.h>
87#include <vm/vm_map.h>
88#include <vm/vm_page.h>
89#include <vm/vm_pageout.h>
90#include <vm/vm_object.h>
91#include <vm/vm_kern.h>
92#include <vm/vm_extern.h>
93#include <vm/vm_pager.h>
94
95#include <sys/user.h>
96
97extern int maxslp;
98
99/*
100 * System initialization
101 *
102 * Note: proc0 from proc.h
103 */
104static void vm_init_limits(void *);
105SYSINIT(vm_limits, SI_SUB_VM_CONF, SI_ORDER_FIRST, vm_init_limits, &proc0)
106
107/*
108 * THIS MUST BE THE LAST INITIALIZATION ITEM!!!
109 *
110 * Note: run scheduling should be divorced from the vm system.
111 */
112static void scheduler(void *);
113SYSINIT(scheduler, SI_SUB_RUN_SCHEDULER, SI_ORDER_FIRST, scheduler, NULL)
114
115#ifndef NO_SWAPPING
116static void swapout(struct proc *);
117static void vm_proc_swapin(struct proc *p);
118static void vm_proc_swapout(struct proc *p);
119#endif
120
121/*
122 * MPSAFE
123 */
124int
125kernacc(addr, len, rw)
126 caddr_t addr;
127 int len, rw;
128{
129 boolean_t rv;
130 vm_offset_t saddr, eaddr;
131 vm_prot_t prot;
132
133 KASSERT((rw & ~VM_PROT_ALL) == 0,
134 ("illegal ``rw'' argument to kernacc (%x)\n", rw));
135 prot = rw;
136 saddr = trunc_page((vm_offset_t)addr);
137 eaddr = round_page((vm_offset_t)addr + len);
138 rv = vm_map_check_protection(kernel_map, saddr, eaddr, prot);
139 return (rv == TRUE);
140}
141
142/*
143 * MPSAFE
144 */
145int
146useracc(addr, len, rw)
147 caddr_t addr;
148 int len, rw;
149{
150 boolean_t rv;
151 vm_prot_t prot;
152
153 KASSERT((rw & ~VM_PROT_ALL) == 0,
154 ("illegal ``rw'' argument to useracc (%x)\n", rw));
155 prot = rw;
156 /*
157 * XXX - check separately to disallow access to user area and user
158 * page tables - they are in the map.
159 *
160 * XXX - VM_MAXUSER_ADDRESS is an end address, not a max. It was once
161 * only used (as an end address) in trap.c. Use it as an end address
162 * here too. This bogusness has spread. I just fixed where it was
163 * used as a max in vm_mmap.c.
164 */
165 if ((vm_offset_t) addr + len > /* XXX */ VM_MAXUSER_ADDRESS
166 || (vm_offset_t) addr + len < (vm_offset_t) addr) {
167 return (FALSE);
168 }
169 rv = vm_map_check_protection(&curproc->p_vmspace->vm_map,
170 trunc_page((vm_offset_t)addr), round_page((vm_offset_t)addr + len),
171 prot);
172 return (rv == TRUE);
173}
174
175/*
176 * MPSAFE
177 */
178void
179vslock(addr, len)
180 caddr_t addr;
181 u_int len;
182{
183
184 vm_map_wire(&curproc->p_vmspace->vm_map, trunc_page((vm_offset_t)addr),
185 round_page((vm_offset_t)addr + len), FALSE);
186}
187
188/*
189 * MPSAFE
190 */
191void
192vsunlock(addr, len)
193 caddr_t addr;
194 u_int len;
195{
196
197 vm_map_unwire(&curproc->p_vmspace->vm_map,
198 trunc_page((vm_offset_t)addr),
199 round_page((vm_offset_t)addr + len), FALSE);
200}
201
202/*
203 * Create the U area for a new process.
204 * This routine directly affects the fork perf for a process.
205 */
206void
207vm_proc_new(struct proc *p)
208{
209 vm_page_t ma[UAREA_PAGES];
210 vm_object_t upobj;
211 vm_offset_t up;
212 vm_page_t m;
213 u_int i;
214
215 /*
216 * Allocate object for the upage.
217 */
218 upobj = vm_object_allocate(OBJT_DEFAULT, UAREA_PAGES);
219 p->p_upages_obj = upobj;
220
221 /*
222 * Get a kernel virtual address for the U area for this process.
223 */
224 up = kmem_alloc_nofault(kernel_map, UAREA_PAGES * PAGE_SIZE);
225 if (up == 0)
226 panic("vm_proc_new: upage allocation failed");
227 p->p_uarea = (struct user *)up;
228
229 for (i = 0; i < UAREA_PAGES; i++) {
230 /*
231 * Get a uarea page.
232 */
233 m = vm_page_grab(upobj, i,
234 VM_ALLOC_NORMAL | VM_ALLOC_RETRY | VM_ALLOC_WIRED);
235 ma[i] = m;
236
237 vm_page_wakeup(m);
238 vm_page_flag_clear(m, PG_ZERO);
239 m->valid = VM_PAGE_BITS_ALL;
240 }
241
242 /*
243 * Enter the pages into the kernel address space.
244 */
245 pmap_qenter(up, ma, UAREA_PAGES);
246}
247
248/*
249 * Dispose the U area for a process that has exited.
250 * This routine directly impacts the exit perf of a process.
251 * XXX proc_zone is marked UMA_ZONE_NOFREE, so this should never be called.
252 */
253void
254vm_proc_dispose(struct proc *p)
255{
256 vm_object_t upobj;
257 vm_offset_t up;
258 vm_page_t m;
259
260 upobj = p->p_upages_obj;
261 if (upobj->resident_page_count != UAREA_PAGES)
262 panic("vm_proc_dispose: incorrect number of pages in upobj");
263 vm_page_lock_queues();
264 while ((m = TAILQ_FIRST(&upobj->memq)) != NULL) {
265 vm_page_busy(m);
266 vm_page_unwire(m, 0);
267 vm_page_free(m);
268 }
269 vm_page_unlock_queues();
270 up = (vm_offset_t)p->p_uarea;
271 pmap_qremove(up, UAREA_PAGES);
272 kmem_free(kernel_map, up, UAREA_PAGES * PAGE_SIZE);
273 vm_object_deallocate(upobj);
274}
275
276#ifndef NO_SWAPPING
277/*
278 * Allow the U area for a process to be prejudicially paged out.
279 */
280void
281vm_proc_swapout(struct proc *p)
282{
283 vm_object_t upobj;
284 vm_offset_t up;
285 vm_page_t m;
286
287 upobj = p->p_upages_obj;
288 if (upobj->resident_page_count != UAREA_PAGES)
289 panic("vm_proc_dispose: incorrect number of pages in upobj");
290 vm_page_lock_queues();
291 TAILQ_FOREACH(m, &upobj->memq, listq) {
292 vm_page_dirty(m);
293 vm_page_unwire(m, 0);
294 }
295 vm_page_unlock_queues();
296 up = (vm_offset_t)p->p_uarea;
297 pmap_qremove(up, UAREA_PAGES);
298}
299
300/*
301 * Bring the U area for a specified process back in.
302 */
303void
304vm_proc_swapin(struct proc *p)
305{
306 vm_page_t ma[UAREA_PAGES];
307 vm_object_t upobj;
308 vm_offset_t up;
309 vm_page_t m;
310 int rv;
311 int i;
312
313 upobj = p->p_upages_obj;
314 for (i = 0; i < UAREA_PAGES; i++) {
315 m = vm_page_grab(upobj, i, VM_ALLOC_NORMAL | VM_ALLOC_RETRY);
316 if (m->valid != VM_PAGE_BITS_ALL) {
317 rv = vm_pager_get_pages(upobj, &m, 1, 0);
318 if (rv != VM_PAGER_OK)
319 panic("vm_proc_swapin: cannot get upage");
320 }
321 ma[i] = m;
322 }
323 if (upobj->resident_page_count != UAREA_PAGES)
324 panic("vm_proc_swapin: lost pages from upobj");
325 vm_page_lock_queues();
326 TAILQ_FOREACH(m, &upobj->memq, listq) {
327 m->valid = VM_PAGE_BITS_ALL;
328 vm_page_wire(m);
329 vm_page_wakeup(m);
330 }
331 vm_page_unlock_queues();
332 up = (vm_offset_t)p->p_uarea;
333 pmap_qenter(up, ma, UAREA_PAGES);
334}
335#endif
336
337/*
338 * Implement fork's actions on an address space.
339 * Here we arrange for the address space to be copied or referenced,
340 * allocate a user struct (pcb and kernel stack), then call the
341 * machine-dependent layer to fill those in and make the new process
342 * ready to run. The new process is set up so that it returns directly
343 * to user mode to avoid stack copying and relocation problems.
344 */
345void
346vm_forkproc(td, p2, td2, flags)
347 struct thread *td;
348 struct proc *p2;
349 struct thread *td2;
350 int flags;
351{
352 struct proc *p1 = td->td_proc;
353 struct user *up;
354
355 GIANT_REQUIRED;
356
357 if ((flags & RFPROC) == 0) {
358 /*
359 * Divorce the memory, if it is shared, essentially
360 * this changes shared memory amongst threads, into
361 * COW locally.
362 */
363 if ((flags & RFMEM) == 0) {
364 if (p1->p_vmspace->vm_refcnt > 1) {
365 vmspace_unshare(p1);
366 }
367 }
368 cpu_fork(td, p2, td2, flags);
369 return;
370 }
371
372 if (flags & RFMEM) {
373 p2->p_vmspace = p1->p_vmspace;
374 p1->p_vmspace->vm_refcnt++;
375 }
376
377 while (vm_page_count_severe()) {
378 VM_WAIT;
379 }
380
381 if ((flags & RFMEM) == 0) {
382 p2->p_vmspace = vmspace_fork(p1->p_vmspace);
383
384 pmap_pinit2(vmspace_pmap(p2->p_vmspace));
385
386 if (p1->p_vmspace->vm_shm)
387 shmfork(p1, p2);
388 }
389
390 /* XXXKSE this is unsatisfactory but should be adequate */
391 up = p2->p_uarea;
392
393 /*
394 * p_stats currently points at fields in the user struct
395 * but not at &u, instead at p_addr. Copy parts of
396 * p_stats; zero the rest of p_stats (statistics).
397 *
398 * If procsig->ps_refcnt is 1 and p2->p_sigacts is NULL we dont' need
399 * to share sigacts, so we use the up->u_sigacts.
400 */
401 p2->p_stats = &up->u_stats;
402 if (p2->p_sigacts == NULL) {
403 if (p2->p_procsig->ps_refcnt != 1)
404 printf ("PID:%d NULL sigacts with refcnt not 1!\n",p2->p_pid);
405 p2->p_sigacts = &up->u_sigacts;
406 up->u_sigacts = *p1->p_sigacts;
407 }
408
409 bzero(&up->u_stats.pstat_startzero,
410 (unsigned) ((caddr_t) &up->u_stats.pstat_endzero -
411 (caddr_t) &up->u_stats.pstat_startzero));
412 bcopy(&p1->p_stats->pstat_startcopy, &up->u_stats.pstat_startcopy,
413 ((caddr_t) &up->u_stats.pstat_endcopy -
414 (caddr_t) &up->u_stats.pstat_startcopy));
415
416
417 /*
418 * cpu_fork will copy and update the pcb, set up the kernel stack,
419 * and make the child ready to run.
420 */
421 cpu_fork(td, p2, td2, flags);
422}
423
424/*
425 * Called after process has been wait(2)'ed apon and is being reaped.
426 * The idea is to reclaim resources that we could not reclaim while
427 * the process was still executing.
428 */
429void
430vm_waitproc(p)
431 struct proc *p;
432{
433
434 GIANT_REQUIRED;
435 cpu_wait(p);
436 vmspace_exitfree(p); /* and clean-out the vmspace */
437}
438
439/*
440 * Set default limits for VM system.
441 * Called for proc 0, and then inherited by all others.
442 *
443 * XXX should probably act directly on proc0.
444 */
445static void
446vm_init_limits(udata)
447 void *udata;
448{
449 struct proc *p = udata;
450 int rss_limit;
451
452 /*
453 * Set up the initial limits on process VM. Set the maximum resident
454 * set size to be half of (reasonably) available memory. Since this
455 * is a soft limit, it comes into effect only when the system is out
456 * of memory - half of main memory helps to favor smaller processes,
457 * and reduces thrashing of the object cache.
458 */
459 p->p_rlimit[RLIMIT_STACK].rlim_cur = dflssiz;
460 p->p_rlimit[RLIMIT_STACK].rlim_max = maxssiz;
461 p->p_rlimit[RLIMIT_DATA].rlim_cur = dfldsiz;
462 p->p_rlimit[RLIMIT_DATA].rlim_max = maxdsiz;
463 /* limit the limit to no less than 2MB */
464 rss_limit = max(cnt.v_free_count, 512);
465 p->p_rlimit[RLIMIT_RSS].rlim_cur = ptoa(rss_limit);
466 p->p_rlimit[RLIMIT_RSS].rlim_max = RLIM_INFINITY;
467}
468
469void
470faultin(p)
471 struct proc *p;
472{
473
474 GIANT_REQUIRED;
475 PROC_LOCK_ASSERT(p, MA_OWNED);
476 mtx_assert(&sched_lock, MA_OWNED);
477#ifdef NO_SWAPPING
478 if ((p->p_sflag & PS_INMEM) == 0)
479 panic("faultin: proc swapped out with NO_SWAPPING!");
480#else
481 if ((p->p_sflag & PS_INMEM) == 0) {
482 struct thread *td;
483
484 ++p->p_lock;
485 /*
486 * If another process is swapping in this process,
487 * just wait until it finishes.
488 */
489 if (p->p_sflag & PS_SWAPPINGIN) {
490 mtx_unlock_spin(&sched_lock);
491 msleep(&p->p_sflag, &p->p_mtx, PVM, "faultin", 0);
492 mtx_lock_spin(&sched_lock);
493 --p->p_lock;
494 return;
495 }
496
497 p->p_sflag |= PS_SWAPPINGIN;
498 mtx_unlock_spin(&sched_lock);
499 PROC_UNLOCK(p);
500
501 vm_proc_swapin(p);
502 FOREACH_THREAD_IN_PROC (p, td) {
503 pmap_swapin_thread(td);
504 TD_CLR_SWAPPED(td);
505 }
506
507 PROC_LOCK(p);
508 mtx_lock_spin(&sched_lock);
509 p->p_sflag &= ~PS_SWAPPINGIN;
510 p->p_sflag |= PS_INMEM;
511 FOREACH_THREAD_IN_PROC (p, td)
512 if (TD_CAN_RUN(td))
513 setrunnable(td);
514
515 wakeup(&p->p_sflag);
516
517 /* undo the effect of setting SLOCK above */
518 --p->p_lock;
519 }
520#endif
521}
522
523/*
524 * This swapin algorithm attempts to swap-in processes only if there
525 * is enough space for them. Of course, if a process waits for a long
526 * time, it will be swapped in anyway.
527 *
528 * XXXKSE - process with the thread with highest priority counts..
529 *
530 * Giant is still held at this point, to be released in tsleep.
531 */
532/* ARGSUSED*/
533static void
534scheduler(dummy)
535 void *dummy;
536{
537 struct proc *p;
538 struct thread *td;
539 int pri;
540 struct proc *pp;
541 int ppri;
542
543 mtx_assert(&Giant, MA_OWNED | MA_NOTRECURSED);
544 /* GIANT_REQUIRED */
545
546loop:
547 if (vm_page_count_min()) {
548 VM_WAIT;
549 goto loop;
550 }
551
552 pp = NULL;
553 ppri = INT_MIN;
554 sx_slock(&allproc_lock);
555 FOREACH_PROC_IN_SYSTEM(p) {
556 struct ksegrp *kg;
557 if (p->p_sflag & (PS_INMEM | PS_SWAPPING | PS_SWAPPINGIN)) {
558 continue;
559 }
560 mtx_lock_spin(&sched_lock);
561 FOREACH_THREAD_IN_PROC(p, td) {
562 /*
563 * An otherwise runnable thread of a process
564 * swapped out has only the TDI_SWAPPED bit set.
565 *
566 */
567 if (td->td_inhibitors == TDI_SWAPPED) {
568 kg = td->td_ksegrp;
569 pri = p->p_swtime + kg->kg_slptime;
570 if ((p->p_sflag & PS_SWAPINREQ) == 0) {
571 pri -= kg->kg_nice * 8;
572 }
573
574 /*
575 * if this ksegrp is higher priority
576 * and there is enough space, then select
577 * this process instead of the previous
578 * selection.
579 */
580 if (pri > ppri) {
581 pp = p;
582 ppri = pri;
583 }
584 }
585 }
586 mtx_unlock_spin(&sched_lock);
587 }
588 sx_sunlock(&allproc_lock);
589
590 /*
591 * Nothing to do, back to sleep.
592 */
593 if ((p = pp) == NULL) {
594 tsleep(&proc0, PVM, "sched", maxslp * hz / 2);
595 goto loop;
596 }
597 PROC_LOCK(p);
598 mtx_lock_spin(&sched_lock);
599
600 /*
601 * Another process may be bringing or may have already
602 * brought this process in while we traverse all threads.
603 * Or, this process may even be being swapped out again.
604 */
605 if (p->p_sflag & (PS_INMEM|PS_SWAPPING|PS_SWAPPINGIN)) {
606 mtx_unlock_spin(&sched_lock);
607 PROC_UNLOCK(p);
608 goto loop;
609 }
610
611 p->p_sflag &= ~PS_SWAPINREQ;
612
613 /*
614 * We would like to bring someone in. (only if there is space).
615 * [What checks the space? ]
616 */
617 faultin(p);
618 PROC_UNLOCK(p);
619 p->p_swtime = 0;
620 mtx_unlock_spin(&sched_lock);
621 goto loop;
622}
623
624#ifndef NO_SWAPPING
625
626/*
627 * Swap_idle_threshold1 is the guaranteed swapped in time for a process
628 */
629static int swap_idle_threshold1 = 2;
630SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold1,
631 CTLFLAG_RW, &swap_idle_threshold1, 0, "");
632
633/*
634 * Swap_idle_threshold2 is the time that a process can be idle before
635 * it will be swapped out, if idle swapping is enabled.
636 */
637static int swap_idle_threshold2 = 10;
638SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold2,
639 CTLFLAG_RW, &swap_idle_threshold2, 0, "");
640
641/*
642 * Swapout is driven by the pageout daemon. Very simple, we find eligible
643 * procs and unwire their u-areas. We try to always "swap" at least one
644 * process in case we need the room for a swapin.
645 * If any procs have been sleeping/stopped for at least maxslp seconds,
646 * they are swapped. Else, we swap the longest-sleeping or stopped process,
647 * if any, otherwise the longest-resident process.
648 */
649void
650swapout_procs(action)
651int action;
652{
653 struct proc *p;
654 struct thread *td;
655 struct ksegrp *kg;
656 struct proc *outp, *outp2;
657 int outpri, outpri2;
658 int didswap = 0;
659
660 GIANT_REQUIRED;
661
662 outp = outp2 = NULL;
663 outpri = outpri2 = INT_MIN;
664retry:
665 sx_slock(&allproc_lock);
666 FOREACH_PROC_IN_SYSTEM(p) {
667 struct vmspace *vm;
668 int minslptime = 100000;
669
670 /*
671 * Do not swapout a process that
672 * is waiting for VM data
673 * structures there is a possible
674 * deadlock. Test this first as
675 * this may block.
676 *
677 * Lock the map until swapout
678 * finishes, or a thread of this
679 * process may attempt to alter
680 * the map.
681 *
682 * Watch out for a process in
683 * creation. It may have no
684 * address space yet.
685 *
686 * An aio daemon switches its
687 * address space while running.
688 * Perform a quick check whether
689 * a process has P_SYSTEM.
690 */
691 PROC_LOCK(p);
692 if ((p->p_flag & P_SYSTEM) != 0) {
693 PROC_UNLOCK(p);
694 continue;
695 }
696 mtx_lock_spin(&sched_lock);
697 if (p->p_state == PRS_NEW) {
698 mtx_unlock_spin(&sched_lock);
699 PROC_UNLOCK(p);
700 continue;
701 }
702 vm = p->p_vmspace;
703 KASSERT(vm != NULL,
704 ("swapout_procs: a process has no address space"));
705 ++vm->vm_refcnt;
706 mtx_unlock_spin(&sched_lock);
707 PROC_UNLOCK(p);
708 if (!vm_map_trylock(&vm->vm_map))
709 goto nextproc1;
710
711 PROC_LOCK(p);
712 if (p->p_lock != 0 ||
713 (p->p_flag & (P_STOPPED_SINGLE|P_TRACED|P_SYSTEM|P_WEXIT)
714 ) != 0) {
715 goto nextproc2;
716 }
717 /*
718 * only aiod changes vmspace, however it will be
719 * skipped because of the if statement above checking
720 * for P_SYSTEM
721 */
722 mtx_lock_spin(&sched_lock);
723 if ((p->p_sflag & (PS_INMEM|PS_SWAPPING|PS_SWAPPINGIN)) != PS_INMEM)
724 goto nextproc;
725
726 switch (p->p_state) {
727 default:
728 /* Don't swap out processes in any sort
729 * of 'special' state. */
730 goto nextproc;
731
732 case PRS_NORMAL:
733 /*
734 * do not swapout a realtime process
735 * Check all the thread groups..
736 */
737 FOREACH_KSEGRP_IN_PROC(p, kg) {
738 if (PRI_IS_REALTIME(kg->kg_pri_class))
739 goto nextproc;
740
741 /*
742 * Guarantee swap_idle_threshold1
743 * time in memory.
744 */
745 if (kg->kg_slptime < swap_idle_threshold1)
746 goto nextproc;
747
748 /*
749 * Do not swapout a process if it is
750 * waiting on a critical event of some
751 * kind or there is a thread whose
752 * pageable memory may be accessed.
753 *
754 * This could be refined to support
755 * swapping out a thread.
756 */
757 FOREACH_THREAD_IN_GROUP(kg, td) {
758 if ((td->td_priority) < PSOCK ||
759 !thread_safetoswapout(td))
760 goto nextproc;
761 }
762 /*
763 * If the system is under memory stress,
764 * or if we are swapping
765 * idle processes >= swap_idle_threshold2,
766 * then swap the process out.
767 */
768 if (((action & VM_SWAP_NORMAL) == 0) &&
769 (((action & VM_SWAP_IDLE) == 0) ||
770 (kg->kg_slptime < swap_idle_threshold2)))
771 goto nextproc;
772
773 if (minslptime > kg->kg_slptime)
774 minslptime = kg->kg_slptime;
775 }
776
777 /*
778 * If the process has been asleep for awhile and had
779 * most of its pages taken away already, swap it out.
780 */
781 if ((action & VM_SWAP_NORMAL) ||
782 ((action & VM_SWAP_IDLE) &&
783 (minslptime > swap_idle_threshold2))) {
784 swapout(p);
785 didswap++;
786
787 /*
788 * swapout() unlocks a proc lock. This is
789 * ugly, but avoids superfluous lock.
790 */
791 mtx_unlock_spin(&sched_lock);
792 vm_map_unlock(&vm->vm_map);
793 vmspace_free(vm);
794 sx_sunlock(&allproc_lock);
795 goto retry;
796 }
797 }
798nextproc:
799 mtx_unlock_spin(&sched_lock);
800nextproc2:
801 PROC_UNLOCK(p);
802 vm_map_unlock(&vm->vm_map);
803nextproc1:
804 vmspace_free(vm);
805 continue;
806 }
807 sx_sunlock(&allproc_lock);
808 /*
809 * If we swapped something out, and another process needed memory,
810 * then wakeup the sched process.
811 */
812 if (didswap)
813 wakeup(&proc0);
814}
815
816static void
817swapout(p)
818 struct proc *p;
819{
820 struct thread *td;
821
822 PROC_LOCK_ASSERT(p, MA_OWNED);
823 mtx_assert(&sched_lock, MA_OWNED | MA_NOTRECURSED);
824#if defined(SWAP_DEBUG)
825 printf("swapping out %d\n", p->p_pid);
826#endif
827
828 /*
829 * The states of this process and its threads may have changed
830 * by now. Assuming that there is only one pageout daemon thread,
831 * this process should still be in memory.
832 */
833 KASSERT((p->p_sflag & (PS_INMEM|PS_SWAPPING|PS_SWAPPINGIN)) == PS_INMEM,
834 ("swapout: lost a swapout race?"));
835
836#if defined(INVARIANTS)
837 /*
838 * Make sure that all threads are safe to be swapped out.
839 *
840 * Alternatively, we could swap out only safe threads.
841 */
842 FOREACH_THREAD_IN_PROC(p, td) {
843 KASSERT(thread_safetoswapout(td),
844 ("swapout: there is a thread not safe for swapout"));
845 }
846#endif /* INVARIANTS */
847
848 ++p->p_stats->p_ru.ru_nswap;
849 /*
850 * remember the process resident count
851 */
852 p->p_vmspace->vm_swrss = vmspace_resident_count(p->p_vmspace);
853
854 PROC_UNLOCK(p);
855 FOREACH_THREAD_IN_PROC (p, td) /* shouldn't be possible, but..... */
856 if (TD_ON_RUNQ(td)) { /* XXXKSE */
857 panic("swapping out runnable process");
858 remrunqueue(td); /* XXXKSE */
859 }
860 p->p_sflag &= ~PS_INMEM;
861 p->p_sflag |= PS_SWAPPING;
862 mtx_unlock_spin(&sched_lock);
863
864 vm_proc_swapout(p);
865 FOREACH_THREAD_IN_PROC(p, td) {
866 pmap_swapout_thread(td);
867 TD_SET_SWAPPED(td);
868 }
869 mtx_lock_spin(&sched_lock);
870 p->p_sflag &= ~PS_SWAPPING;
871 p->p_swtime = 0;
872}
873#endif /* !NO_SWAPPING */