xref: /freebsd/sys/vm/vm_glue.c (revision 38f0b757fd84d17d0fc24739a7cda160c4516d81)
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  * 4. Neither the name of the University nor the names of its contributors
17  *    may be used to endorse or promote products derived from this software
18  *    without specific prior written permission.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
21  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
24  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30  * SUCH DAMAGE.
31  *
32  *	from: @(#)vm_glue.c	8.6 (Berkeley) 1/5/94
33  *
34  *
35  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
36  * All rights reserved.
37  *
38  * Permission to use, copy, modify and distribute this software and
39  * its documentation is hereby granted, provided that both the copyright
40  * notice and this permission notice appear in all copies of the
41  * software, derivative works or modified versions, and any portions
42  * thereof, and that both notices appear in supporting documentation.
43  *
44  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
45  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
46  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
47  *
48  * Carnegie Mellon requests users of this software to return to
49  *
50  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
51  *  School of Computer Science
52  *  Carnegie Mellon University
53  *  Pittsburgh PA 15213-3890
54  *
55  * any improvements or extensions that they make and grant Carnegie the
56  * rights to redistribute these changes.
57  */
58 
59 #include <sys/cdefs.h>
60 __FBSDID("$FreeBSD$");
61 
62 #include "opt_vm.h"
63 #include "opt_kstack_pages.h"
64 #include "opt_kstack_max_pages.h"
65 
66 #include <sys/param.h>
67 #include <sys/systm.h>
68 #include <sys/limits.h>
69 #include <sys/lock.h>
70 #include <sys/malloc.h>
71 #include <sys/mutex.h>
72 #include <sys/proc.h>
73 #include <sys/racct.h>
74 #include <sys/resourcevar.h>
75 #include <sys/rwlock.h>
76 #include <sys/sched.h>
77 #include <sys/sf_buf.h>
78 #include <sys/shm.h>
79 #include <sys/vmmeter.h>
80 #include <sys/vmem.h>
81 #include <sys/sx.h>
82 #include <sys/sysctl.h>
83 #include <sys/_kstack_cache.h>
84 #include <sys/eventhandler.h>
85 #include <sys/kernel.h>
86 #include <sys/ktr.h>
87 #include <sys/unistd.h>
88 
89 #include <vm/vm.h>
90 #include <vm/vm_param.h>
91 #include <vm/pmap.h>
92 #include <vm/vm_map.h>
93 #include <vm/vm_page.h>
94 #include <vm/vm_pageout.h>
95 #include <vm/vm_object.h>
96 #include <vm/vm_kern.h>
97 #include <vm/vm_extern.h>
98 #include <vm/vm_pager.h>
99 #include <vm/swap_pager.h>
100 
101 #ifndef NO_SWAPPING
102 static int swapout(struct proc *);
103 static void swapclear(struct proc *);
104 static void vm_thread_swapin(struct thread *td);
105 static void vm_thread_swapout(struct thread *td);
106 #endif
107 
108 /*
109  * MPSAFE
110  *
111  * WARNING!  This code calls vm_map_check_protection() which only checks
112  * the associated vm_map_entry range.  It does not determine whether the
113  * contents of the memory is actually readable or writable.  In most cases
114  * just checking the vm_map_entry is sufficient within the kernel's address
115  * space.
116  */
117 int
118 kernacc(addr, len, rw)
119 	void *addr;
120 	int len, rw;
121 {
122 	boolean_t rv;
123 	vm_offset_t saddr, eaddr;
124 	vm_prot_t prot;
125 
126 	KASSERT((rw & ~VM_PROT_ALL) == 0,
127 	    ("illegal ``rw'' argument to kernacc (%x)\n", rw));
128 
129 	if ((vm_offset_t)addr + len > kernel_map->max_offset ||
130 	    (vm_offset_t)addr + len < (vm_offset_t)addr)
131 		return (FALSE);
132 
133 	prot = rw;
134 	saddr = trunc_page((vm_offset_t)addr);
135 	eaddr = round_page((vm_offset_t)addr + len);
136 	vm_map_lock_read(kernel_map);
137 	rv = vm_map_check_protection(kernel_map, saddr, eaddr, prot);
138 	vm_map_unlock_read(kernel_map);
139 	return (rv == TRUE);
140 }
141 
142 /*
143  * MPSAFE
144  *
145  * WARNING!  This code calls vm_map_check_protection() which only checks
146  * the associated vm_map_entry range.  It does not determine whether the
147  * contents of the memory is actually readable or writable.  vmapbuf(),
148  * vm_fault_quick(), or copyin()/copout()/su*()/fu*() functions should be
149  * used in conjuction with this call.
150  */
151 int
152 useracc(addr, len, rw)
153 	void *addr;
154 	int len, rw;
155 {
156 	boolean_t rv;
157 	vm_prot_t prot;
158 	vm_map_t map;
159 
160 	KASSERT((rw & ~VM_PROT_ALL) == 0,
161 	    ("illegal ``rw'' argument to useracc (%x)\n", rw));
162 	prot = rw;
163 	map = &curproc->p_vmspace->vm_map;
164 	if ((vm_offset_t)addr + len > vm_map_max(map) ||
165 	    (vm_offset_t)addr + len < (vm_offset_t)addr) {
166 		return (FALSE);
167 	}
168 	vm_map_lock_read(map);
169 	rv = vm_map_check_protection(map, trunc_page((vm_offset_t)addr),
170 	    round_page((vm_offset_t)addr + len), prot);
171 	vm_map_unlock_read(map);
172 	return (rv == TRUE);
173 }
174 
175 int
176 vslock(void *addr, size_t len)
177 {
178 	vm_offset_t end, last, start;
179 	vm_size_t npages;
180 	int error;
181 
182 	last = (vm_offset_t)addr + len;
183 	start = trunc_page((vm_offset_t)addr);
184 	end = round_page(last);
185 	if (last < (vm_offset_t)addr || end < (vm_offset_t)addr)
186 		return (EINVAL);
187 	npages = atop(end - start);
188 	if (npages > vm_page_max_wired)
189 		return (ENOMEM);
190 #if 0
191 	/*
192 	 * XXX - not yet
193 	 *
194 	 * The limit for transient usage of wired pages should be
195 	 * larger than for "permanent" wired pages (mlock()).
196 	 *
197 	 * Also, the sysctl code, which is the only present user
198 	 * of vslock(), does a hard loop on EAGAIN.
199 	 */
200 	if (npages + vm_cnt.v_wire_count > vm_page_max_wired)
201 		return (EAGAIN);
202 #endif
203 	error = vm_map_wire(&curproc->p_vmspace->vm_map, start, end,
204 	    VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES);
205 	/*
206 	 * Return EFAULT on error to match copy{in,out}() behaviour
207 	 * rather than returning ENOMEM like mlock() would.
208 	 */
209 	return (error == KERN_SUCCESS ? 0 : EFAULT);
210 }
211 
212 void
213 vsunlock(void *addr, size_t len)
214 {
215 
216 	/* Rely on the parameter sanity checks performed by vslock(). */
217 	(void)vm_map_unwire(&curproc->p_vmspace->vm_map,
218 	    trunc_page((vm_offset_t)addr), round_page((vm_offset_t)addr + len),
219 	    VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES);
220 }
221 
222 /*
223  * Pin the page contained within the given object at the given offset.  If the
224  * page is not resident, allocate and load it using the given object's pager.
225  * Return the pinned page if successful; otherwise, return NULL.
226  */
227 static vm_page_t
228 vm_imgact_hold_page(vm_object_t object, vm_ooffset_t offset)
229 {
230 	vm_page_t m, ma[1];
231 	vm_pindex_t pindex;
232 	int rv;
233 
234 	VM_OBJECT_WLOCK(object);
235 	pindex = OFF_TO_IDX(offset);
236 	m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL);
237 	if (m->valid != VM_PAGE_BITS_ALL) {
238 		ma[0] = m;
239 		rv = vm_pager_get_pages(object, ma, 1, 0);
240 		m = vm_page_lookup(object, pindex);
241 		if (m == NULL)
242 			goto out;
243 		if (rv != VM_PAGER_OK) {
244 			vm_page_lock(m);
245 			vm_page_free(m);
246 			vm_page_unlock(m);
247 			m = NULL;
248 			goto out;
249 		}
250 	}
251 	vm_page_xunbusy(m);
252 	vm_page_lock(m);
253 	vm_page_hold(m);
254 	vm_page_unlock(m);
255 out:
256 	VM_OBJECT_WUNLOCK(object);
257 	return (m);
258 }
259 
260 /*
261  * Return a CPU private mapping to the page at the given offset within the
262  * given object.  The page is pinned before it is mapped.
263  */
264 struct sf_buf *
265 vm_imgact_map_page(vm_object_t object, vm_ooffset_t offset)
266 {
267 	vm_page_t m;
268 
269 	m = vm_imgact_hold_page(object, offset);
270 	if (m == NULL)
271 		return (NULL);
272 	sched_pin();
273 	return (sf_buf_alloc(m, SFB_CPUPRIVATE));
274 }
275 
276 /*
277  * Destroy the given CPU private mapping and unpin the page that it mapped.
278  */
279 void
280 vm_imgact_unmap_page(struct sf_buf *sf)
281 {
282 	vm_page_t m;
283 
284 	m = sf_buf_page(sf);
285 	sf_buf_free(sf);
286 	sched_unpin();
287 	vm_page_lock(m);
288 	vm_page_unhold(m);
289 	vm_page_unlock(m);
290 }
291 
292 void
293 vm_sync_icache(vm_map_t map, vm_offset_t va, vm_offset_t sz)
294 {
295 
296 	pmap_sync_icache(map->pmap, va, sz);
297 }
298 
299 struct kstack_cache_entry *kstack_cache;
300 static int kstack_cache_size = 128;
301 static int kstacks;
302 static struct mtx kstack_cache_mtx;
303 MTX_SYSINIT(kstack_cache, &kstack_cache_mtx, "kstkch", MTX_DEF);
304 
305 SYSCTL_INT(_vm, OID_AUTO, kstack_cache_size, CTLFLAG_RW, &kstack_cache_size, 0,
306     "");
307 SYSCTL_INT(_vm, OID_AUTO, kstacks, CTLFLAG_RD, &kstacks, 0,
308     "");
309 
310 #ifndef KSTACK_MAX_PAGES
311 #define KSTACK_MAX_PAGES 32
312 #endif
313 
314 /*
315  * Create the kernel stack (including pcb for i386) for a new thread.
316  * This routine directly affects the fork perf for a process and
317  * create performance for a thread.
318  */
319 int
320 vm_thread_new(struct thread *td, int pages)
321 {
322 	vm_object_t ksobj;
323 	vm_offset_t ks;
324 	vm_page_t m, ma[KSTACK_MAX_PAGES];
325 	struct kstack_cache_entry *ks_ce;
326 	int i;
327 
328 	/* Bounds check */
329 	if (pages <= 1)
330 		pages = KSTACK_PAGES;
331 	else if (pages > KSTACK_MAX_PAGES)
332 		pages = KSTACK_MAX_PAGES;
333 
334 	if (pages == KSTACK_PAGES) {
335 		mtx_lock(&kstack_cache_mtx);
336 		if (kstack_cache != NULL) {
337 			ks_ce = kstack_cache;
338 			kstack_cache = ks_ce->next_ks_entry;
339 			mtx_unlock(&kstack_cache_mtx);
340 
341 			td->td_kstack_obj = ks_ce->ksobj;
342 			td->td_kstack = (vm_offset_t)ks_ce;
343 			td->td_kstack_pages = KSTACK_PAGES;
344 			return (1);
345 		}
346 		mtx_unlock(&kstack_cache_mtx);
347 	}
348 
349 	/*
350 	 * Allocate an object for the kstack.
351 	 */
352 	ksobj = vm_object_allocate(OBJT_DEFAULT, pages);
353 
354 	/*
355 	 * Get a kernel virtual address for this thread's kstack.
356 	 */
357 #if defined(__mips__)
358 	/*
359 	 * We need to align the kstack's mapped address to fit within
360 	 * a single TLB entry.
361 	 */
362 	if (vmem_xalloc(kernel_arena, (pages + KSTACK_GUARD_PAGES) * PAGE_SIZE,
363 	    PAGE_SIZE * 2, 0, 0, VMEM_ADDR_MIN, VMEM_ADDR_MAX,
364 	    M_BESTFIT | M_NOWAIT, &ks)) {
365 		ks = 0;
366 	}
367 #else
368 	ks = kva_alloc((pages + KSTACK_GUARD_PAGES) * PAGE_SIZE);
369 #endif
370 	if (ks == 0) {
371 		printf("vm_thread_new: kstack allocation failed\n");
372 		vm_object_deallocate(ksobj);
373 		return (0);
374 	}
375 
376 	atomic_add_int(&kstacks, 1);
377 	if (KSTACK_GUARD_PAGES != 0) {
378 		pmap_qremove(ks, KSTACK_GUARD_PAGES);
379 		ks += KSTACK_GUARD_PAGES * PAGE_SIZE;
380 	}
381 	td->td_kstack_obj = ksobj;
382 	td->td_kstack = ks;
383 	/*
384 	 * Knowing the number of pages allocated is useful when you
385 	 * want to deallocate them.
386 	 */
387 	td->td_kstack_pages = pages;
388 	/*
389 	 * For the length of the stack, link in a real page of ram for each
390 	 * page of stack.
391 	 */
392 	VM_OBJECT_WLOCK(ksobj);
393 	for (i = 0; i < pages; i++) {
394 		/*
395 		 * Get a kernel stack page.
396 		 */
397 		m = vm_page_grab(ksobj, i, VM_ALLOC_NOBUSY |
398 		    VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
399 		ma[i] = m;
400 		m->valid = VM_PAGE_BITS_ALL;
401 	}
402 	VM_OBJECT_WUNLOCK(ksobj);
403 	pmap_qenter(ks, ma, pages);
404 	return (1);
405 }
406 
407 static void
408 vm_thread_stack_dispose(vm_object_t ksobj, vm_offset_t ks, int pages)
409 {
410 	vm_page_t m;
411 	int i;
412 
413 	atomic_add_int(&kstacks, -1);
414 	pmap_qremove(ks, pages);
415 	VM_OBJECT_WLOCK(ksobj);
416 	for (i = 0; i < pages; i++) {
417 		m = vm_page_lookup(ksobj, i);
418 		if (m == NULL)
419 			panic("vm_thread_dispose: kstack already missing?");
420 		vm_page_lock(m);
421 		vm_page_unwire(m, 0);
422 		vm_page_free(m);
423 		vm_page_unlock(m);
424 	}
425 	VM_OBJECT_WUNLOCK(ksobj);
426 	vm_object_deallocate(ksobj);
427 	kva_free(ks - (KSTACK_GUARD_PAGES * PAGE_SIZE),
428 	    (pages + KSTACK_GUARD_PAGES) * PAGE_SIZE);
429 }
430 
431 /*
432  * Dispose of a thread's kernel stack.
433  */
434 void
435 vm_thread_dispose(struct thread *td)
436 {
437 	vm_object_t ksobj;
438 	vm_offset_t ks;
439 	struct kstack_cache_entry *ks_ce;
440 	int pages;
441 
442 	pages = td->td_kstack_pages;
443 	ksobj = td->td_kstack_obj;
444 	ks = td->td_kstack;
445 	td->td_kstack = 0;
446 	td->td_kstack_pages = 0;
447 	if (pages == KSTACK_PAGES && kstacks <= kstack_cache_size) {
448 		ks_ce = (struct kstack_cache_entry *)ks;
449 		ks_ce->ksobj = ksobj;
450 		mtx_lock(&kstack_cache_mtx);
451 		ks_ce->next_ks_entry = kstack_cache;
452 		kstack_cache = ks_ce;
453 		mtx_unlock(&kstack_cache_mtx);
454 		return;
455 	}
456 	vm_thread_stack_dispose(ksobj, ks, pages);
457 }
458 
459 static void
460 vm_thread_stack_lowmem(void *nulll)
461 {
462 	struct kstack_cache_entry *ks_ce, *ks_ce1;
463 
464 	mtx_lock(&kstack_cache_mtx);
465 	ks_ce = kstack_cache;
466 	kstack_cache = NULL;
467 	mtx_unlock(&kstack_cache_mtx);
468 
469 	while (ks_ce != NULL) {
470 		ks_ce1 = ks_ce;
471 		ks_ce = ks_ce->next_ks_entry;
472 
473 		vm_thread_stack_dispose(ks_ce1->ksobj, (vm_offset_t)ks_ce1,
474 		    KSTACK_PAGES);
475 	}
476 }
477 
478 static void
479 kstack_cache_init(void *nulll)
480 {
481 
482 	EVENTHANDLER_REGISTER(vm_lowmem, vm_thread_stack_lowmem, NULL,
483 	    EVENTHANDLER_PRI_ANY);
484 }
485 
486 SYSINIT(vm_kstacks, SI_SUB_KTHREAD_INIT, SI_ORDER_ANY, kstack_cache_init, NULL);
487 
488 #ifndef NO_SWAPPING
489 /*
490  * Allow a thread's kernel stack to be paged out.
491  */
492 static void
493 vm_thread_swapout(struct thread *td)
494 {
495 	vm_object_t ksobj;
496 	vm_page_t m;
497 	int i, pages;
498 
499 	cpu_thread_swapout(td);
500 	pages = td->td_kstack_pages;
501 	ksobj = td->td_kstack_obj;
502 	pmap_qremove(td->td_kstack, pages);
503 	VM_OBJECT_WLOCK(ksobj);
504 	for (i = 0; i < pages; i++) {
505 		m = vm_page_lookup(ksobj, i);
506 		if (m == NULL)
507 			panic("vm_thread_swapout: kstack already missing?");
508 		vm_page_dirty(m);
509 		vm_page_lock(m);
510 		vm_page_unwire(m, 0);
511 		vm_page_unlock(m);
512 	}
513 	VM_OBJECT_WUNLOCK(ksobj);
514 }
515 
516 /*
517  * Bring the kernel stack for a specified thread back in.
518  */
519 static void
520 vm_thread_swapin(struct thread *td)
521 {
522 	vm_object_t ksobj;
523 	vm_page_t ma[KSTACK_MAX_PAGES];
524 	int i, j, k, pages, rv;
525 
526 	pages = td->td_kstack_pages;
527 	ksobj = td->td_kstack_obj;
528 	VM_OBJECT_WLOCK(ksobj);
529 	for (i = 0; i < pages; i++)
530 		ma[i] = vm_page_grab(ksobj, i, VM_ALLOC_NORMAL |
531 		    VM_ALLOC_WIRED);
532 	for (i = 0; i < pages; i++) {
533 		if (ma[i]->valid != VM_PAGE_BITS_ALL) {
534 			vm_page_assert_xbusied(ma[i]);
535 			vm_object_pip_add(ksobj, 1);
536 			for (j = i + 1; j < pages; j++) {
537 				if (ma[j]->valid != VM_PAGE_BITS_ALL)
538 					vm_page_assert_xbusied(ma[j]);
539 				if (ma[j]->valid == VM_PAGE_BITS_ALL)
540 					break;
541 			}
542 			rv = vm_pager_get_pages(ksobj, ma + i, j - i, 0);
543 			if (rv != VM_PAGER_OK)
544 	panic("vm_thread_swapin: cannot get kstack for proc: %d",
545 				    td->td_proc->p_pid);
546 			vm_object_pip_wakeup(ksobj);
547 			for (k = i; k < j; k++)
548 				ma[k] = vm_page_lookup(ksobj, k);
549 			vm_page_xunbusy(ma[i]);
550 		} else if (vm_page_xbusied(ma[i]))
551 			vm_page_xunbusy(ma[i]);
552 	}
553 	VM_OBJECT_WUNLOCK(ksobj);
554 	pmap_qenter(td->td_kstack, ma, pages);
555 	cpu_thread_swapin(td);
556 }
557 #endif /* !NO_SWAPPING */
558 
559 /*
560  * Implement fork's actions on an address space.
561  * Here we arrange for the address space to be copied or referenced,
562  * allocate a user struct (pcb and kernel stack), then call the
563  * machine-dependent layer to fill those in and make the new process
564  * ready to run.  The new process is set up so that it returns directly
565  * to user mode to avoid stack copying and relocation problems.
566  */
567 int
568 vm_forkproc(td, p2, td2, vm2, flags)
569 	struct thread *td;
570 	struct proc *p2;
571 	struct thread *td2;
572 	struct vmspace *vm2;
573 	int flags;
574 {
575 	struct proc *p1 = td->td_proc;
576 	int error;
577 
578 	if ((flags & RFPROC) == 0) {
579 		/*
580 		 * Divorce the memory, if it is shared, essentially
581 		 * this changes shared memory amongst threads, into
582 		 * COW locally.
583 		 */
584 		if ((flags & RFMEM) == 0) {
585 			if (p1->p_vmspace->vm_refcnt > 1) {
586 				error = vmspace_unshare(p1);
587 				if (error)
588 					return (error);
589 			}
590 		}
591 		cpu_fork(td, p2, td2, flags);
592 		return (0);
593 	}
594 
595 	if (flags & RFMEM) {
596 		p2->p_vmspace = p1->p_vmspace;
597 		atomic_add_int(&p1->p_vmspace->vm_refcnt, 1);
598 	}
599 
600 	while (vm_page_count_severe()) {
601 		VM_WAIT;
602 	}
603 
604 	if ((flags & RFMEM) == 0) {
605 		p2->p_vmspace = vm2;
606 		if (p1->p_vmspace->vm_shm)
607 			shmfork(p1, p2);
608 	}
609 
610 	/*
611 	 * cpu_fork will copy and update the pcb, set up the kernel stack,
612 	 * and make the child ready to run.
613 	 */
614 	cpu_fork(td, p2, td2, flags);
615 	return (0);
616 }
617 
618 /*
619  * Called after process has been wait(2)'ed apon and is being reaped.
620  * The idea is to reclaim resources that we could not reclaim while
621  * the process was still executing.
622  */
623 void
624 vm_waitproc(p)
625 	struct proc *p;
626 {
627 
628 	vmspace_exitfree(p);		/* and clean-out the vmspace */
629 }
630 
631 void
632 faultin(p)
633 	struct proc *p;
634 {
635 #ifdef NO_SWAPPING
636 
637 	PROC_LOCK_ASSERT(p, MA_OWNED);
638 	if ((p->p_flag & P_INMEM) == 0)
639 		panic("faultin: proc swapped out with NO_SWAPPING!");
640 #else /* !NO_SWAPPING */
641 	struct thread *td;
642 
643 	PROC_LOCK_ASSERT(p, MA_OWNED);
644 	/*
645 	 * If another process is swapping in this process,
646 	 * just wait until it finishes.
647 	 */
648 	if (p->p_flag & P_SWAPPINGIN) {
649 		while (p->p_flag & P_SWAPPINGIN)
650 			msleep(&p->p_flag, &p->p_mtx, PVM, "faultin", 0);
651 		return;
652 	}
653 	if ((p->p_flag & P_INMEM) == 0) {
654 		/*
655 		 * Don't let another thread swap process p out while we are
656 		 * busy swapping it in.
657 		 */
658 		++p->p_lock;
659 		p->p_flag |= P_SWAPPINGIN;
660 		PROC_UNLOCK(p);
661 
662 		/*
663 		 * We hold no lock here because the list of threads
664 		 * can not change while all threads in the process are
665 		 * swapped out.
666 		 */
667 		FOREACH_THREAD_IN_PROC(p, td)
668 			vm_thread_swapin(td);
669 		PROC_LOCK(p);
670 		swapclear(p);
671 		p->p_swtick = ticks;
672 
673 		wakeup(&p->p_flag);
674 
675 		/* Allow other threads to swap p out now. */
676 		--p->p_lock;
677 	}
678 #endif /* NO_SWAPPING */
679 }
680 
681 /*
682  * This swapin algorithm attempts to swap-in processes only if there
683  * is enough space for them.  Of course, if a process waits for a long
684  * time, it will be swapped in anyway.
685  *
686  * Giant is held on entry.
687  */
688 void
689 swapper(void)
690 {
691 	struct proc *p;
692 	struct thread *td;
693 	struct proc *pp;
694 	int slptime;
695 	int swtime;
696 	int ppri;
697 	int pri;
698 
699 loop:
700 	if (vm_page_count_min()) {
701 		VM_WAIT;
702 		goto loop;
703 	}
704 
705 	pp = NULL;
706 	ppri = INT_MIN;
707 	sx_slock(&allproc_lock);
708 	FOREACH_PROC_IN_SYSTEM(p) {
709 		PROC_LOCK(p);
710 		if (p->p_state == PRS_NEW ||
711 		    p->p_flag & (P_SWAPPINGOUT | P_SWAPPINGIN | P_INMEM)) {
712 			PROC_UNLOCK(p);
713 			continue;
714 		}
715 		swtime = (ticks - p->p_swtick) / hz;
716 		FOREACH_THREAD_IN_PROC(p, td) {
717 			/*
718 			 * An otherwise runnable thread of a process
719 			 * swapped out has only the TDI_SWAPPED bit set.
720 			 *
721 			 */
722 			thread_lock(td);
723 			if (td->td_inhibitors == TDI_SWAPPED) {
724 				slptime = (ticks - td->td_slptick) / hz;
725 				pri = swtime + slptime;
726 				if ((td->td_flags & TDF_SWAPINREQ) == 0)
727 					pri -= p->p_nice * 8;
728 				/*
729 				 * if this thread is higher priority
730 				 * and there is enough space, then select
731 				 * this process instead of the previous
732 				 * selection.
733 				 */
734 				if (pri > ppri) {
735 					pp = p;
736 					ppri = pri;
737 				}
738 			}
739 			thread_unlock(td);
740 		}
741 		PROC_UNLOCK(p);
742 	}
743 	sx_sunlock(&allproc_lock);
744 
745 	/*
746 	 * Nothing to do, back to sleep.
747 	 */
748 	if ((p = pp) == NULL) {
749 		tsleep(&proc0, PVM, "swapin", MAXSLP * hz / 2);
750 		goto loop;
751 	}
752 	PROC_LOCK(p);
753 
754 	/*
755 	 * Another process may be bringing or may have already
756 	 * brought this process in while we traverse all threads.
757 	 * Or, this process may even be being swapped out again.
758 	 */
759 	if (p->p_flag & (P_INMEM | P_SWAPPINGOUT | P_SWAPPINGIN)) {
760 		PROC_UNLOCK(p);
761 		goto loop;
762 	}
763 
764 	/*
765 	 * We would like to bring someone in. (only if there is space).
766 	 * [What checks the space? ]
767 	 */
768 	faultin(p);
769 	PROC_UNLOCK(p);
770 	goto loop;
771 }
772 
773 void
774 kick_proc0(void)
775 {
776 
777 	wakeup(&proc0);
778 }
779 
780 #ifndef NO_SWAPPING
781 
782 /*
783  * Swap_idle_threshold1 is the guaranteed swapped in time for a process
784  */
785 static int swap_idle_threshold1 = 2;
786 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold1, CTLFLAG_RW,
787     &swap_idle_threshold1, 0, "Guaranteed swapped in time for a process");
788 
789 /*
790  * Swap_idle_threshold2 is the time that a process can be idle before
791  * it will be swapped out, if idle swapping is enabled.
792  */
793 static int swap_idle_threshold2 = 10;
794 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold2, CTLFLAG_RW,
795     &swap_idle_threshold2, 0, "Time before a process will be swapped out");
796 
797 /*
798  * First, if any processes have been sleeping or stopped for at least
799  * "swap_idle_threshold1" seconds, they are swapped out.  If, however,
800  * no such processes exist, then the longest-sleeping or stopped
801  * process is swapped out.  Finally, and only as a last resort, if
802  * there are no sleeping or stopped processes, the longest-resident
803  * process is swapped out.
804  */
805 void
806 swapout_procs(action)
807 int action;
808 {
809 	struct proc *p;
810 	struct thread *td;
811 	int didswap = 0;
812 
813 retry:
814 	sx_slock(&allproc_lock);
815 	FOREACH_PROC_IN_SYSTEM(p) {
816 		struct vmspace *vm;
817 		int minslptime = 100000;
818 		int slptime;
819 
820 		/*
821 		 * Watch out for a process in
822 		 * creation.  It may have no
823 		 * address space or lock yet.
824 		 */
825 		if (p->p_state == PRS_NEW)
826 			continue;
827 		/*
828 		 * An aio daemon switches its
829 		 * address space while running.
830 		 * Perform a quick check whether
831 		 * a process has P_SYSTEM.
832 		 */
833 		if ((p->p_flag & P_SYSTEM) != 0)
834 			continue;
835 		/*
836 		 * Do not swapout a process that
837 		 * is waiting for VM data
838 		 * structures as there is a possible
839 		 * deadlock.  Test this first as
840 		 * this may block.
841 		 *
842 		 * Lock the map until swapout
843 		 * finishes, or a thread of this
844 		 * process may attempt to alter
845 		 * the map.
846 		 */
847 		vm = vmspace_acquire_ref(p);
848 		if (vm == NULL)
849 			continue;
850 		if (!vm_map_trylock(&vm->vm_map))
851 			goto nextproc1;
852 
853 		PROC_LOCK(p);
854 		if (p->p_lock != 0 ||
855 		    (p->p_flag & (P_STOPPED_SINGLE|P_TRACED|P_SYSTEM|P_WEXIT)
856 		    ) != 0) {
857 			goto nextproc;
858 		}
859 		/*
860 		 * only aiod changes vmspace, however it will be
861 		 * skipped because of the if statement above checking
862 		 * for P_SYSTEM
863 		 */
864 		if ((p->p_flag & (P_INMEM|P_SWAPPINGOUT|P_SWAPPINGIN)) != P_INMEM)
865 			goto nextproc;
866 
867 		switch (p->p_state) {
868 		default:
869 			/* Don't swap out processes in any sort
870 			 * of 'special' state. */
871 			break;
872 
873 		case PRS_NORMAL:
874 			/*
875 			 * do not swapout a realtime process
876 			 * Check all the thread groups..
877 			 */
878 			FOREACH_THREAD_IN_PROC(p, td) {
879 				thread_lock(td);
880 				if (PRI_IS_REALTIME(td->td_pri_class)) {
881 					thread_unlock(td);
882 					goto nextproc;
883 				}
884 				slptime = (ticks - td->td_slptick) / hz;
885 				/*
886 				 * Guarantee swap_idle_threshold1
887 				 * time in memory.
888 				 */
889 				if (slptime < swap_idle_threshold1) {
890 					thread_unlock(td);
891 					goto nextproc;
892 				}
893 
894 				/*
895 				 * Do not swapout a process if it is
896 				 * waiting on a critical event of some
897 				 * kind or there is a thread whose
898 				 * pageable memory may be accessed.
899 				 *
900 				 * This could be refined to support
901 				 * swapping out a thread.
902 				 */
903 				if (!thread_safetoswapout(td)) {
904 					thread_unlock(td);
905 					goto nextproc;
906 				}
907 				/*
908 				 * If the system is under memory stress,
909 				 * or if we are swapping
910 				 * idle processes >= swap_idle_threshold2,
911 				 * then swap the process out.
912 				 */
913 				if (((action & VM_SWAP_NORMAL) == 0) &&
914 				    (((action & VM_SWAP_IDLE) == 0) ||
915 				    (slptime < swap_idle_threshold2))) {
916 					thread_unlock(td);
917 					goto nextproc;
918 				}
919 
920 				if (minslptime > slptime)
921 					minslptime = slptime;
922 				thread_unlock(td);
923 			}
924 
925 			/*
926 			 * If the pageout daemon didn't free enough pages,
927 			 * or if this process is idle and the system is
928 			 * configured to swap proactively, swap it out.
929 			 */
930 			if ((action & VM_SWAP_NORMAL) ||
931 				((action & VM_SWAP_IDLE) &&
932 				 (minslptime > swap_idle_threshold2))) {
933 				if (swapout(p) == 0)
934 					didswap++;
935 				PROC_UNLOCK(p);
936 				vm_map_unlock(&vm->vm_map);
937 				vmspace_free(vm);
938 				sx_sunlock(&allproc_lock);
939 				goto retry;
940 			}
941 		}
942 nextproc:
943 		PROC_UNLOCK(p);
944 		vm_map_unlock(&vm->vm_map);
945 nextproc1:
946 		vmspace_free(vm);
947 		continue;
948 	}
949 	sx_sunlock(&allproc_lock);
950 	/*
951 	 * If we swapped something out, and another process needed memory,
952 	 * then wakeup the sched process.
953 	 */
954 	if (didswap)
955 		wakeup(&proc0);
956 }
957 
958 static void
959 swapclear(p)
960 	struct proc *p;
961 {
962 	struct thread *td;
963 
964 	PROC_LOCK_ASSERT(p, MA_OWNED);
965 
966 	FOREACH_THREAD_IN_PROC(p, td) {
967 		thread_lock(td);
968 		td->td_flags |= TDF_INMEM;
969 		td->td_flags &= ~TDF_SWAPINREQ;
970 		TD_CLR_SWAPPED(td);
971 		if (TD_CAN_RUN(td))
972 			if (setrunnable(td)) {
973 #ifdef INVARIANTS
974 				/*
975 				 * XXX: We just cleared TDI_SWAPPED
976 				 * above and set TDF_INMEM, so this
977 				 * should never happen.
978 				 */
979 				panic("not waking up swapper");
980 #endif
981 			}
982 		thread_unlock(td);
983 	}
984 	p->p_flag &= ~(P_SWAPPINGIN|P_SWAPPINGOUT);
985 	p->p_flag |= P_INMEM;
986 }
987 
988 static int
989 swapout(p)
990 	struct proc *p;
991 {
992 	struct thread *td;
993 
994 	PROC_LOCK_ASSERT(p, MA_OWNED);
995 #if defined(SWAP_DEBUG)
996 	printf("swapping out %d\n", p->p_pid);
997 #endif
998 
999 	/*
1000 	 * The states of this process and its threads may have changed
1001 	 * by now.  Assuming that there is only one pageout daemon thread,
1002 	 * this process should still be in memory.
1003 	 */
1004 	KASSERT((p->p_flag & (P_INMEM|P_SWAPPINGOUT|P_SWAPPINGIN)) == P_INMEM,
1005 		("swapout: lost a swapout race?"));
1006 
1007 	/*
1008 	 * remember the process resident count
1009 	 */
1010 	p->p_vmspace->vm_swrss = vmspace_resident_count(p->p_vmspace);
1011 	/*
1012 	 * Check and mark all threads before we proceed.
1013 	 */
1014 	p->p_flag &= ~P_INMEM;
1015 	p->p_flag |= P_SWAPPINGOUT;
1016 	FOREACH_THREAD_IN_PROC(p, td) {
1017 		thread_lock(td);
1018 		if (!thread_safetoswapout(td)) {
1019 			thread_unlock(td);
1020 			swapclear(p);
1021 			return (EBUSY);
1022 		}
1023 		td->td_flags &= ~TDF_INMEM;
1024 		TD_SET_SWAPPED(td);
1025 		thread_unlock(td);
1026 	}
1027 	td = FIRST_THREAD_IN_PROC(p);
1028 	++td->td_ru.ru_nswap;
1029 	PROC_UNLOCK(p);
1030 
1031 	/*
1032 	 * This list is stable because all threads are now prevented from
1033 	 * running.  The list is only modified in the context of a running
1034 	 * thread in this process.
1035 	 */
1036 	FOREACH_THREAD_IN_PROC(p, td)
1037 		vm_thread_swapout(td);
1038 
1039 	PROC_LOCK(p);
1040 	p->p_flag &= ~P_SWAPPINGOUT;
1041 	p->p_swtick = ticks;
1042 	return (0);
1043 }
1044 #endif /* !NO_SWAPPING */
1045