xref: /freebsd/sys/vm/vm_glue.c (revision 5dae51da3da0cc94d17bd67b308fad304ebec7e0)
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 #include "opt_kstack_usage_prof.h"
66 
67 #include <sys/param.h>
68 #include <sys/systm.h>
69 #include <sys/limits.h>
70 #include <sys/lock.h>
71 #include <sys/malloc.h>
72 #include <sys/mutex.h>
73 #include <sys/proc.h>
74 #include <sys/racct.h>
75 #include <sys/resourcevar.h>
76 #include <sys/rwlock.h>
77 #include <sys/sched.h>
78 #include <sys/sf_buf.h>
79 #include <sys/shm.h>
80 #include <sys/vmmeter.h>
81 #include <sys/vmem.h>
82 #include <sys/sx.h>
83 #include <sys/sysctl.h>
84 #include <sys/_kstack_cache.h>
85 #include <sys/eventhandler.h>
86 #include <sys/kernel.h>
87 #include <sys/ktr.h>
88 #include <sys/unistd.h>
89 
90 #include <vm/vm.h>
91 #include <vm/vm_param.h>
92 #include <vm/pmap.h>
93 #include <vm/vm_map.h>
94 #include <vm/vm_page.h>
95 #include <vm/vm_pageout.h>
96 #include <vm/vm_object.h>
97 #include <vm/vm_kern.h>
98 #include <vm/vm_extern.h>
99 #include <vm/vm_pager.h>
100 #include <vm/swap_pager.h>
101 
102 #include <machine/cpu.h>
103 
104 #ifndef NO_SWAPPING
105 static int swapout(struct proc *);
106 static void swapclear(struct proc *);
107 static void vm_thread_swapin(struct thread *td);
108 static void vm_thread_swapout(struct thread *td);
109 #endif
110 
111 /*
112  * MPSAFE
113  *
114  * WARNING!  This code calls vm_map_check_protection() which only checks
115  * the associated vm_map_entry range.  It does not determine whether the
116  * contents of the memory is actually readable or writable.  In most cases
117  * just checking the vm_map_entry is sufficient within the kernel's address
118  * space.
119  */
120 int
121 kernacc(addr, len, rw)
122 	void *addr;
123 	int len, rw;
124 {
125 	boolean_t rv;
126 	vm_offset_t saddr, eaddr;
127 	vm_prot_t prot;
128 
129 	KASSERT((rw & ~VM_PROT_ALL) == 0,
130 	    ("illegal ``rw'' argument to kernacc (%x)\n", rw));
131 
132 	if ((vm_offset_t)addr + len > kernel_map->max_offset ||
133 	    (vm_offset_t)addr + len < (vm_offset_t)addr)
134 		return (FALSE);
135 
136 	prot = rw;
137 	saddr = trunc_page((vm_offset_t)addr);
138 	eaddr = round_page((vm_offset_t)addr + len);
139 	vm_map_lock_read(kernel_map);
140 	rv = vm_map_check_protection(kernel_map, saddr, eaddr, prot);
141 	vm_map_unlock_read(kernel_map);
142 	return (rv == TRUE);
143 }
144 
145 /*
146  * MPSAFE
147  *
148  * WARNING!  This code calls vm_map_check_protection() which only checks
149  * the associated vm_map_entry range.  It does not determine whether the
150  * contents of the memory is actually readable or writable.  vmapbuf(),
151  * vm_fault_quick(), or copyin()/copout()/su*()/fu*() functions should be
152  * used in conjunction with this call.
153  */
154 int
155 useracc(addr, len, rw)
156 	void *addr;
157 	int len, rw;
158 {
159 	boolean_t rv;
160 	vm_prot_t prot;
161 	vm_map_t map;
162 
163 	KASSERT((rw & ~VM_PROT_ALL) == 0,
164 	    ("illegal ``rw'' argument to useracc (%x)\n", rw));
165 	prot = rw;
166 	map = &curproc->p_vmspace->vm_map;
167 	if ((vm_offset_t)addr + len > vm_map_max(map) ||
168 	    (vm_offset_t)addr + len < (vm_offset_t)addr) {
169 		return (FALSE);
170 	}
171 	vm_map_lock_read(map);
172 	rv = vm_map_check_protection(map, trunc_page((vm_offset_t)addr),
173 	    round_page((vm_offset_t)addr + len), prot);
174 	vm_map_unlock_read(map);
175 	return (rv == TRUE);
176 }
177 
178 int
179 vslock(void *addr, size_t len)
180 {
181 	vm_offset_t end, last, start;
182 	vm_size_t npages;
183 	int error;
184 
185 	last = (vm_offset_t)addr + len;
186 	start = trunc_page((vm_offset_t)addr);
187 	end = round_page(last);
188 	if (last < (vm_offset_t)addr || end < (vm_offset_t)addr)
189 		return (EINVAL);
190 	npages = atop(end - start);
191 	if (npages > vm_page_max_wired)
192 		return (ENOMEM);
193 #if 0
194 	/*
195 	 * XXX - not yet
196 	 *
197 	 * The limit for transient usage of wired pages should be
198 	 * larger than for "permanent" wired pages (mlock()).
199 	 *
200 	 * Also, the sysctl code, which is the only present user
201 	 * of vslock(), does a hard loop on EAGAIN.
202 	 */
203 	if (npages + vm_cnt.v_wire_count > vm_page_max_wired)
204 		return (EAGAIN);
205 #endif
206 	error = vm_map_wire(&curproc->p_vmspace->vm_map, start, end,
207 	    VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES);
208 	/*
209 	 * Return EFAULT on error to match copy{in,out}() behaviour
210 	 * rather than returning ENOMEM like mlock() would.
211 	 */
212 	return (error == KERN_SUCCESS ? 0 : EFAULT);
213 }
214 
215 void
216 vsunlock(void *addr, size_t len)
217 {
218 
219 	/* Rely on the parameter sanity checks performed by vslock(). */
220 	(void)vm_map_unwire(&curproc->p_vmspace->vm_map,
221 	    trunc_page((vm_offset_t)addr), round_page((vm_offset_t)addr + len),
222 	    VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES);
223 }
224 
225 /*
226  * Pin the page contained within the given object at the given offset.  If the
227  * page is not resident, allocate and load it using the given object's pager.
228  * Return the pinned page if successful; otherwise, return NULL.
229  */
230 static vm_page_t
231 vm_imgact_hold_page(vm_object_t object, vm_ooffset_t offset)
232 {
233 	vm_page_t m;
234 	vm_pindex_t pindex;
235 	int rv;
236 
237 	VM_OBJECT_WLOCK(object);
238 	pindex = OFF_TO_IDX(offset);
239 	m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL | VM_ALLOC_NOBUSY);
240 	if (m->valid != VM_PAGE_BITS_ALL) {
241 		vm_page_xbusy(m);
242 		rv = vm_pager_get_pages(object, &m, 1, NULL, NULL);
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 		vm_page_xunbusy(m);
251 	}
252 	vm_page_lock(m);
253 	vm_page_hold(m);
254 	vm_page_activate(m);
255 	vm_page_unlock(m);
256 out:
257 	VM_OBJECT_WUNLOCK(object);
258 	return (m);
259 }
260 
261 /*
262  * Return a CPU private mapping to the page at the given offset within the
263  * given object.  The page is pinned before it is mapped.
264  */
265 struct sf_buf *
266 vm_imgact_map_page(vm_object_t object, vm_ooffset_t offset)
267 {
268 	vm_page_t m;
269 
270 	m = vm_imgact_hold_page(object, offset);
271 	if (m == NULL)
272 		return (NULL);
273 	sched_pin();
274 	return (sf_buf_alloc(m, SFB_CPUPRIVATE));
275 }
276 
277 /*
278  * Destroy the given CPU private mapping and unpin the page that it mapped.
279  */
280 void
281 vm_imgact_unmap_page(struct sf_buf *sf)
282 {
283 	vm_page_t m;
284 
285 	m = sf_buf_page(sf);
286 	sf_buf_free(sf);
287 	sched_unpin();
288 	vm_page_lock(m);
289 	vm_page_unhold(m);
290 	vm_page_unlock(m);
291 }
292 
293 void
294 vm_sync_icache(vm_map_t map, vm_offset_t va, vm_offset_t sz)
295 {
296 
297 	pmap_sync_icache(map->pmap, va, sz);
298 }
299 
300 struct kstack_cache_entry *kstack_cache;
301 static int kstack_cache_size = 128;
302 static int kstacks;
303 static struct mtx kstack_cache_mtx;
304 MTX_SYSINIT(kstack_cache, &kstack_cache_mtx, "kstkch", MTX_DEF);
305 
306 SYSCTL_INT(_vm, OID_AUTO, kstack_cache_size, CTLFLAG_RW, &kstack_cache_size, 0,
307     "");
308 SYSCTL_INT(_vm, OID_AUTO, kstacks, CTLFLAG_RD, &kstacks, 0,
309     "");
310 
311 #ifndef KSTACK_MAX_PAGES
312 #define KSTACK_MAX_PAGES 32
313 #endif
314 
315 /*
316  * Create the kernel stack (including pcb for i386) for a new thread.
317  * This routine directly affects the fork perf for a process and
318  * create performance for a thread.
319  */
320 int
321 vm_thread_new(struct thread *td, int pages)
322 {
323 	vm_object_t ksobj;
324 	vm_offset_t ks;
325 	vm_page_t m, ma[KSTACK_MAX_PAGES];
326 	struct kstack_cache_entry *ks_ce;
327 	int i;
328 
329 	/* Bounds check */
330 	if (pages <= 1)
331 		pages = kstack_pages;
332 	else if (pages > KSTACK_MAX_PAGES)
333 		pages = KSTACK_MAX_PAGES;
334 
335 	if (pages == kstack_pages) {
336 		mtx_lock(&kstack_cache_mtx);
337 		if (kstack_cache != NULL) {
338 			ks_ce = kstack_cache;
339 			kstack_cache = ks_ce->next_ks_entry;
340 			mtx_unlock(&kstack_cache_mtx);
341 
342 			td->td_kstack_obj = ks_ce->ksobj;
343 			td->td_kstack = (vm_offset_t)ks_ce;
344 			td->td_kstack_pages = kstack_pages;
345 			return (1);
346 		}
347 		mtx_unlock(&kstack_cache_mtx);
348 	}
349 
350 	/*
351 	 * Allocate an object for the kstack.
352 	 */
353 	ksobj = vm_object_allocate(OBJT_DEFAULT, pages);
354 
355 	/*
356 	 * Get a kernel virtual address for this thread's kstack.
357 	 */
358 #if defined(__mips__)
359 	/*
360 	 * We need to align the kstack's mapped address to fit within
361 	 * a single TLB entry.
362 	 */
363 	if (vmem_xalloc(kernel_arena, (pages + KSTACK_GUARD_PAGES) * PAGE_SIZE,
364 	    PAGE_SIZE * 2, 0, 0, VMEM_ADDR_MIN, VMEM_ADDR_MAX,
365 	    M_BESTFIT | M_NOWAIT, &ks)) {
366 		ks = 0;
367 	}
368 #else
369 	ks = kva_alloc((pages + KSTACK_GUARD_PAGES) * PAGE_SIZE);
370 #endif
371 	if (ks == 0) {
372 		printf("vm_thread_new: kstack allocation failed\n");
373 		vm_object_deallocate(ksobj);
374 		return (0);
375 	}
376 
377 	atomic_add_int(&kstacks, 1);
378 	if (KSTACK_GUARD_PAGES != 0) {
379 		pmap_qremove(ks, KSTACK_GUARD_PAGES);
380 		ks += KSTACK_GUARD_PAGES * PAGE_SIZE;
381 	}
382 	td->td_kstack_obj = ksobj;
383 	td->td_kstack = ks;
384 	/*
385 	 * Knowing the number of pages allocated is useful when you
386 	 * want to deallocate them.
387 	 */
388 	td->td_kstack_pages = pages;
389 	/*
390 	 * For the length of the stack, link in a real page of ram for each
391 	 * page of stack.
392 	 */
393 	VM_OBJECT_WLOCK(ksobj);
394 	for (i = 0; i < pages; i++) {
395 		/*
396 		 * Get a kernel stack page.
397 		 */
398 		m = vm_page_grab(ksobj, i, VM_ALLOC_NOBUSY |
399 		    VM_ALLOC_NORMAL | VM_ALLOC_WIRED);
400 		ma[i] = m;
401 		m->valid = VM_PAGE_BITS_ALL;
402 	}
403 	VM_OBJECT_WUNLOCK(ksobj);
404 	pmap_qenter(ks, ma, pages);
405 	return (1);
406 }
407 
408 static void
409 vm_thread_stack_dispose(vm_object_t ksobj, vm_offset_t ks, int pages)
410 {
411 	vm_page_t m;
412 	int i;
413 
414 	atomic_add_int(&kstacks, -1);
415 	pmap_qremove(ks, pages);
416 	VM_OBJECT_WLOCK(ksobj);
417 	for (i = 0; i < pages; i++) {
418 		m = vm_page_lookup(ksobj, i);
419 		if (m == NULL)
420 			panic("vm_thread_dispose: kstack already missing?");
421 		vm_page_lock(m);
422 		vm_page_unwire(m, PQ_NONE);
423 		vm_page_free(m);
424 		vm_page_unlock(m);
425 	}
426 	VM_OBJECT_WUNLOCK(ksobj);
427 	vm_object_deallocate(ksobj);
428 	kva_free(ks - (KSTACK_GUARD_PAGES * PAGE_SIZE),
429 	    (pages + KSTACK_GUARD_PAGES) * PAGE_SIZE);
430 }
431 
432 /*
433  * Dispose of a thread's kernel stack.
434  */
435 void
436 vm_thread_dispose(struct thread *td)
437 {
438 	vm_object_t ksobj;
439 	vm_offset_t ks;
440 	struct kstack_cache_entry *ks_ce;
441 	int pages;
442 
443 	pages = td->td_kstack_pages;
444 	ksobj = td->td_kstack_obj;
445 	ks = td->td_kstack;
446 	td->td_kstack = 0;
447 	td->td_kstack_pages = 0;
448 	if (pages == kstack_pages && kstacks <= kstack_cache_size) {
449 		ks_ce = (struct kstack_cache_entry *)ks;
450 		ks_ce->ksobj = ksobj;
451 		mtx_lock(&kstack_cache_mtx);
452 		ks_ce->next_ks_entry = kstack_cache;
453 		kstack_cache = ks_ce;
454 		mtx_unlock(&kstack_cache_mtx);
455 		return;
456 	}
457 	vm_thread_stack_dispose(ksobj, ks, pages);
458 }
459 
460 static void
461 vm_thread_stack_lowmem(void *nulll)
462 {
463 	struct kstack_cache_entry *ks_ce, *ks_ce1;
464 
465 	mtx_lock(&kstack_cache_mtx);
466 	ks_ce = kstack_cache;
467 	kstack_cache = NULL;
468 	mtx_unlock(&kstack_cache_mtx);
469 
470 	while (ks_ce != NULL) {
471 		ks_ce1 = ks_ce;
472 		ks_ce = ks_ce->next_ks_entry;
473 
474 		vm_thread_stack_dispose(ks_ce1->ksobj, (vm_offset_t)ks_ce1,
475 		    kstack_pages);
476 	}
477 }
478 
479 static void
480 kstack_cache_init(void *nulll)
481 {
482 
483 	EVENTHANDLER_REGISTER(vm_lowmem, vm_thread_stack_lowmem, NULL,
484 	    EVENTHANDLER_PRI_ANY);
485 }
486 
487 SYSINIT(vm_kstacks, SI_SUB_KTHREAD_INIT, SI_ORDER_ANY, kstack_cache_init, NULL);
488 
489 #ifdef KSTACK_USAGE_PROF
490 /*
491  * Track maximum stack used by a thread in kernel.
492  */
493 static int max_kstack_used;
494 
495 SYSCTL_INT(_debug, OID_AUTO, max_kstack_used, CTLFLAG_RD,
496     &max_kstack_used, 0,
497     "Maxiumum stack depth used by a thread in kernel");
498 
499 void
500 intr_prof_stack_use(struct thread *td, struct trapframe *frame)
501 {
502 	vm_offset_t stack_top;
503 	vm_offset_t current;
504 	int used, prev_used;
505 
506 	/*
507 	 * Testing for interrupted kernel mode isn't strictly
508 	 * needed. It optimizes the execution, since interrupts from
509 	 * usermode will have only the trap frame on the stack.
510 	 */
511 	if (TRAPF_USERMODE(frame))
512 		return;
513 
514 	stack_top = td->td_kstack + td->td_kstack_pages * PAGE_SIZE;
515 	current = (vm_offset_t)(uintptr_t)&stack_top;
516 
517 	/*
518 	 * Try to detect if interrupt is using kernel thread stack.
519 	 * Hardware could use a dedicated stack for interrupt handling.
520 	 */
521 	if (stack_top <= current || current < td->td_kstack)
522 		return;
523 
524 	used = stack_top - current;
525 	for (;;) {
526 		prev_used = max_kstack_used;
527 		if (prev_used >= used)
528 			break;
529 		if (atomic_cmpset_int(&max_kstack_used, prev_used, used))
530 			break;
531 	}
532 }
533 #endif /* KSTACK_USAGE_PROF */
534 
535 #ifndef NO_SWAPPING
536 /*
537  * Allow a thread's kernel stack to be paged out.
538  */
539 static void
540 vm_thread_swapout(struct thread *td)
541 {
542 	vm_object_t ksobj;
543 	vm_page_t m;
544 	int i, pages;
545 
546 	cpu_thread_swapout(td);
547 	pages = td->td_kstack_pages;
548 	ksobj = td->td_kstack_obj;
549 	pmap_qremove(td->td_kstack, pages);
550 	VM_OBJECT_WLOCK(ksobj);
551 	for (i = 0; i < pages; i++) {
552 		m = vm_page_lookup(ksobj, i);
553 		if (m == NULL)
554 			panic("vm_thread_swapout: kstack already missing?");
555 		vm_page_dirty(m);
556 		vm_page_lock(m);
557 		vm_page_unwire(m, PQ_INACTIVE);
558 		vm_page_unlock(m);
559 	}
560 	VM_OBJECT_WUNLOCK(ksobj);
561 }
562 
563 /*
564  * Bring the kernel stack for a specified thread back in.
565  */
566 static void
567 vm_thread_swapin(struct thread *td)
568 {
569 	vm_object_t ksobj;
570 	vm_page_t ma[KSTACK_MAX_PAGES];
571 	int pages;
572 
573 	pages = td->td_kstack_pages;
574 	ksobj = td->td_kstack_obj;
575 	VM_OBJECT_WLOCK(ksobj);
576 	for (int i = 0; i < pages; i++)
577 		ma[i] = vm_page_grab(ksobj, i, VM_ALLOC_NORMAL |
578 		    VM_ALLOC_WIRED);
579 	for (int i = 0; i < pages;) {
580 		int j, a, count, rv;
581 
582 		vm_page_assert_xbusied(ma[i]);
583 		if (ma[i]->valid == VM_PAGE_BITS_ALL) {
584 			vm_page_xunbusy(ma[i]);
585 			i++;
586 			continue;
587 		}
588 		vm_object_pip_add(ksobj, 1);
589 		for (j = i + 1; j < pages; j++)
590 			if (ma[j]->valid == VM_PAGE_BITS_ALL)
591 				break;
592 		rv = vm_pager_has_page(ksobj, ma[i]->pindex, NULL, &a);
593 		KASSERT(rv == 1, ("%s: missing page %p", __func__, ma[i]));
594 		count = min(a + 1, j - i);
595 		rv = vm_pager_get_pages(ksobj, ma + i, count, NULL, NULL);
596 		KASSERT(rv == VM_PAGER_OK, ("%s: cannot get kstack for proc %d",
597 		    __func__, td->td_proc->p_pid));
598 		vm_object_pip_wakeup(ksobj);
599 		for (j = i; j < i + count; j++)
600 			vm_page_xunbusy(ma[j]);
601 		i += count;
602 	}
603 	VM_OBJECT_WUNLOCK(ksobj);
604 	pmap_qenter(td->td_kstack, ma, pages);
605 	cpu_thread_swapin(td);
606 }
607 #endif /* !NO_SWAPPING */
608 
609 /*
610  * Implement fork's actions on an address space.
611  * Here we arrange for the address space to be copied or referenced,
612  * allocate a user struct (pcb and kernel stack), then call the
613  * machine-dependent layer to fill those in and make the new process
614  * ready to run.  The new process is set up so that it returns directly
615  * to user mode to avoid stack copying and relocation problems.
616  */
617 int
618 vm_forkproc(td, p2, td2, vm2, flags)
619 	struct thread *td;
620 	struct proc *p2;
621 	struct thread *td2;
622 	struct vmspace *vm2;
623 	int flags;
624 {
625 	struct proc *p1 = td->td_proc;
626 	int error;
627 
628 	if ((flags & RFPROC) == 0) {
629 		/*
630 		 * Divorce the memory, if it is shared, essentially
631 		 * this changes shared memory amongst threads, into
632 		 * COW locally.
633 		 */
634 		if ((flags & RFMEM) == 0) {
635 			if (p1->p_vmspace->vm_refcnt > 1) {
636 				error = vmspace_unshare(p1);
637 				if (error)
638 					return (error);
639 			}
640 		}
641 		cpu_fork(td, p2, td2, flags);
642 		return (0);
643 	}
644 
645 	if (flags & RFMEM) {
646 		p2->p_vmspace = p1->p_vmspace;
647 		atomic_add_int(&p1->p_vmspace->vm_refcnt, 1);
648 	}
649 
650 	while (vm_page_count_severe()) {
651 		VM_WAIT;
652 	}
653 
654 	if ((flags & RFMEM) == 0) {
655 		p2->p_vmspace = vm2;
656 		if (p1->p_vmspace->vm_shm)
657 			shmfork(p1, p2);
658 	}
659 
660 	/*
661 	 * cpu_fork will copy and update the pcb, set up the kernel stack,
662 	 * and make the child ready to run.
663 	 */
664 	cpu_fork(td, p2, td2, flags);
665 	return (0);
666 }
667 
668 /*
669  * Called after process has been wait(2)'ed upon and is being reaped.
670  * The idea is to reclaim resources that we could not reclaim while
671  * the process was still executing.
672  */
673 void
674 vm_waitproc(p)
675 	struct proc *p;
676 {
677 
678 	vmspace_exitfree(p);		/* and clean-out the vmspace */
679 }
680 
681 void
682 faultin(p)
683 	struct proc *p;
684 {
685 #ifdef NO_SWAPPING
686 
687 	PROC_LOCK_ASSERT(p, MA_OWNED);
688 	if ((p->p_flag & P_INMEM) == 0)
689 		panic("faultin: proc swapped out with NO_SWAPPING!");
690 #else /* !NO_SWAPPING */
691 	struct thread *td;
692 
693 	PROC_LOCK_ASSERT(p, MA_OWNED);
694 	/*
695 	 * If another process is swapping in this process,
696 	 * just wait until it finishes.
697 	 */
698 	if (p->p_flag & P_SWAPPINGIN) {
699 		while (p->p_flag & P_SWAPPINGIN)
700 			msleep(&p->p_flag, &p->p_mtx, PVM, "faultin", 0);
701 		return;
702 	}
703 	if ((p->p_flag & P_INMEM) == 0) {
704 		/*
705 		 * Don't let another thread swap process p out while we are
706 		 * busy swapping it in.
707 		 */
708 		++p->p_lock;
709 		p->p_flag |= P_SWAPPINGIN;
710 		PROC_UNLOCK(p);
711 
712 		/*
713 		 * We hold no lock here because the list of threads
714 		 * can not change while all threads in the process are
715 		 * swapped out.
716 		 */
717 		FOREACH_THREAD_IN_PROC(p, td)
718 			vm_thread_swapin(td);
719 		PROC_LOCK(p);
720 		swapclear(p);
721 		p->p_swtick = ticks;
722 
723 		wakeup(&p->p_flag);
724 
725 		/* Allow other threads to swap p out now. */
726 		--p->p_lock;
727 	}
728 #endif /* NO_SWAPPING */
729 }
730 
731 /*
732  * This swapin algorithm attempts to swap-in processes only if there
733  * is enough space for them.  Of course, if a process waits for a long
734  * time, it will be swapped in anyway.
735  */
736 void
737 swapper(void)
738 {
739 	struct proc *p;
740 	struct thread *td;
741 	struct proc *pp;
742 	int slptime;
743 	int swtime;
744 	int ppri;
745 	int pri;
746 
747 loop:
748 	if (vm_page_count_min()) {
749 		VM_WAIT;
750 		goto loop;
751 	}
752 
753 	pp = NULL;
754 	ppri = INT_MIN;
755 	sx_slock(&allproc_lock);
756 	FOREACH_PROC_IN_SYSTEM(p) {
757 		PROC_LOCK(p);
758 		if (p->p_state == PRS_NEW ||
759 		    p->p_flag & (P_SWAPPINGOUT | P_SWAPPINGIN | P_INMEM)) {
760 			PROC_UNLOCK(p);
761 			continue;
762 		}
763 		swtime = (ticks - p->p_swtick) / hz;
764 		FOREACH_THREAD_IN_PROC(p, td) {
765 			/*
766 			 * An otherwise runnable thread of a process
767 			 * swapped out has only the TDI_SWAPPED bit set.
768 			 *
769 			 */
770 			thread_lock(td);
771 			if (td->td_inhibitors == TDI_SWAPPED) {
772 				slptime = (ticks - td->td_slptick) / hz;
773 				pri = swtime + slptime;
774 				if ((td->td_flags & TDF_SWAPINREQ) == 0)
775 					pri -= p->p_nice * 8;
776 				/*
777 				 * if this thread is higher priority
778 				 * and there is enough space, then select
779 				 * this process instead of the previous
780 				 * selection.
781 				 */
782 				if (pri > ppri) {
783 					pp = p;
784 					ppri = pri;
785 				}
786 			}
787 			thread_unlock(td);
788 		}
789 		PROC_UNLOCK(p);
790 	}
791 	sx_sunlock(&allproc_lock);
792 
793 	/*
794 	 * Nothing to do, back to sleep.
795 	 */
796 	if ((p = pp) == NULL) {
797 		tsleep(&proc0, PVM, "swapin", MAXSLP * hz / 2);
798 		goto loop;
799 	}
800 	PROC_LOCK(p);
801 
802 	/*
803 	 * Another process may be bringing or may have already
804 	 * brought this process in while we traverse all threads.
805 	 * Or, this process may even be being swapped out again.
806 	 */
807 	if (p->p_flag & (P_INMEM | P_SWAPPINGOUT | P_SWAPPINGIN)) {
808 		PROC_UNLOCK(p);
809 		goto loop;
810 	}
811 
812 	/*
813 	 * We would like to bring someone in. (only if there is space).
814 	 * [What checks the space? ]
815 	 */
816 	faultin(p);
817 	PROC_UNLOCK(p);
818 	goto loop;
819 }
820 
821 void
822 kick_proc0(void)
823 {
824 
825 	wakeup(&proc0);
826 }
827 
828 #ifndef NO_SWAPPING
829 
830 /*
831  * Swap_idle_threshold1 is the guaranteed swapped in time for a process
832  */
833 static int swap_idle_threshold1 = 2;
834 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold1, CTLFLAG_RW,
835     &swap_idle_threshold1, 0, "Guaranteed swapped in time for a process");
836 
837 /*
838  * Swap_idle_threshold2 is the time that a process can be idle before
839  * it will be swapped out, if idle swapping is enabled.
840  */
841 static int swap_idle_threshold2 = 10;
842 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold2, CTLFLAG_RW,
843     &swap_idle_threshold2, 0, "Time before a process will be swapped out");
844 
845 /*
846  * First, if any processes have been sleeping or stopped for at least
847  * "swap_idle_threshold1" seconds, they are swapped out.  If, however,
848  * no such processes exist, then the longest-sleeping or stopped
849  * process is swapped out.  Finally, and only as a last resort, if
850  * there are no sleeping or stopped processes, the longest-resident
851  * process is swapped out.
852  */
853 void
854 swapout_procs(action)
855 int action;
856 {
857 	struct proc *p;
858 	struct thread *td;
859 	int didswap = 0;
860 
861 retry:
862 	sx_slock(&allproc_lock);
863 	FOREACH_PROC_IN_SYSTEM(p) {
864 		struct vmspace *vm;
865 		int minslptime = 100000;
866 		int slptime;
867 
868 		PROC_LOCK(p);
869 		/*
870 		 * Watch out for a process in
871 		 * creation.  It may have no
872 		 * address space or lock yet.
873 		 */
874 		if (p->p_state == PRS_NEW) {
875 			PROC_UNLOCK(p);
876 			continue;
877 		}
878 		/*
879 		 * An aio daemon switches its
880 		 * address space while running.
881 		 * Perform a quick check whether
882 		 * a process has P_SYSTEM.
883 		 * Filter out exiting processes.
884 		 */
885 		if ((p->p_flag & (P_SYSTEM | P_WEXIT)) != 0) {
886 			PROC_UNLOCK(p);
887 			continue;
888 		}
889 		_PHOLD_LITE(p);
890 		PROC_UNLOCK(p);
891 		sx_sunlock(&allproc_lock);
892 
893 		/*
894 		 * Do not swapout a process that
895 		 * is waiting for VM data
896 		 * structures as there is a possible
897 		 * deadlock.  Test this first as
898 		 * this may block.
899 		 *
900 		 * Lock the map until swapout
901 		 * finishes, or a thread of this
902 		 * process may attempt to alter
903 		 * the map.
904 		 */
905 		vm = vmspace_acquire_ref(p);
906 		if (vm == NULL)
907 			goto nextproc2;
908 		if (!vm_map_trylock(&vm->vm_map))
909 			goto nextproc1;
910 
911 		PROC_LOCK(p);
912 		if (p->p_lock != 1 || (p->p_flag & (P_STOPPED_SINGLE |
913 		    P_TRACED | P_SYSTEM)) != 0)
914 			goto nextproc;
915 
916 		/*
917 		 * only aiod changes vmspace, however it will be
918 		 * skipped because of the if statement above checking
919 		 * for P_SYSTEM
920 		 */
921 		if ((p->p_flag & (P_INMEM|P_SWAPPINGOUT|P_SWAPPINGIN)) != P_INMEM)
922 			goto nextproc;
923 
924 		switch (p->p_state) {
925 		default:
926 			/* Don't swap out processes in any sort
927 			 * of 'special' state. */
928 			break;
929 
930 		case PRS_NORMAL:
931 			/*
932 			 * do not swapout a realtime process
933 			 * Check all the thread groups..
934 			 */
935 			FOREACH_THREAD_IN_PROC(p, td) {
936 				thread_lock(td);
937 				if (PRI_IS_REALTIME(td->td_pri_class)) {
938 					thread_unlock(td);
939 					goto nextproc;
940 				}
941 				slptime = (ticks - td->td_slptick) / hz;
942 				/*
943 				 * Guarantee swap_idle_threshold1
944 				 * time in memory.
945 				 */
946 				if (slptime < swap_idle_threshold1) {
947 					thread_unlock(td);
948 					goto nextproc;
949 				}
950 
951 				/*
952 				 * Do not swapout a process if it is
953 				 * waiting on a critical event of some
954 				 * kind or there is a thread whose
955 				 * pageable memory may be accessed.
956 				 *
957 				 * This could be refined to support
958 				 * swapping out a thread.
959 				 */
960 				if (!thread_safetoswapout(td)) {
961 					thread_unlock(td);
962 					goto nextproc;
963 				}
964 				/*
965 				 * If the system is under memory stress,
966 				 * or if we are swapping
967 				 * idle processes >= swap_idle_threshold2,
968 				 * then swap the process out.
969 				 */
970 				if (((action & VM_SWAP_NORMAL) == 0) &&
971 				    (((action & VM_SWAP_IDLE) == 0) ||
972 				    (slptime < swap_idle_threshold2))) {
973 					thread_unlock(td);
974 					goto nextproc;
975 				}
976 
977 				if (minslptime > slptime)
978 					minslptime = slptime;
979 				thread_unlock(td);
980 			}
981 
982 			/*
983 			 * If the pageout daemon didn't free enough pages,
984 			 * or if this process is idle and the system is
985 			 * configured to swap proactively, swap it out.
986 			 */
987 			if ((action & VM_SWAP_NORMAL) ||
988 				((action & VM_SWAP_IDLE) &&
989 				 (minslptime > swap_idle_threshold2))) {
990 				_PRELE(p);
991 				if (swapout(p) == 0)
992 					didswap++;
993 				PROC_UNLOCK(p);
994 				vm_map_unlock(&vm->vm_map);
995 				vmspace_free(vm);
996 				goto retry;
997 			}
998 		}
999 nextproc:
1000 		PROC_UNLOCK(p);
1001 		vm_map_unlock(&vm->vm_map);
1002 nextproc1:
1003 		vmspace_free(vm);
1004 nextproc2:
1005 		sx_slock(&allproc_lock);
1006 		PRELE(p);
1007 	}
1008 	sx_sunlock(&allproc_lock);
1009 	/*
1010 	 * If we swapped something out, and another process needed memory,
1011 	 * then wakeup the sched process.
1012 	 */
1013 	if (didswap)
1014 		wakeup(&proc0);
1015 }
1016 
1017 static void
1018 swapclear(p)
1019 	struct proc *p;
1020 {
1021 	struct thread *td;
1022 
1023 	PROC_LOCK_ASSERT(p, MA_OWNED);
1024 
1025 	FOREACH_THREAD_IN_PROC(p, td) {
1026 		thread_lock(td);
1027 		td->td_flags |= TDF_INMEM;
1028 		td->td_flags &= ~TDF_SWAPINREQ;
1029 		TD_CLR_SWAPPED(td);
1030 		if (TD_CAN_RUN(td))
1031 			if (setrunnable(td)) {
1032 #ifdef INVARIANTS
1033 				/*
1034 				 * XXX: We just cleared TDI_SWAPPED
1035 				 * above and set TDF_INMEM, so this
1036 				 * should never happen.
1037 				 */
1038 				panic("not waking up swapper");
1039 #endif
1040 			}
1041 		thread_unlock(td);
1042 	}
1043 	p->p_flag &= ~(P_SWAPPINGIN|P_SWAPPINGOUT);
1044 	p->p_flag |= P_INMEM;
1045 }
1046 
1047 static int
1048 swapout(p)
1049 	struct proc *p;
1050 {
1051 	struct thread *td;
1052 
1053 	PROC_LOCK_ASSERT(p, MA_OWNED);
1054 #if defined(SWAP_DEBUG)
1055 	printf("swapping out %d\n", p->p_pid);
1056 #endif
1057 
1058 	/*
1059 	 * The states of this process and its threads may have changed
1060 	 * by now.  Assuming that there is only one pageout daemon thread,
1061 	 * this process should still be in memory.
1062 	 */
1063 	KASSERT((p->p_flag & (P_INMEM|P_SWAPPINGOUT|P_SWAPPINGIN)) == P_INMEM,
1064 		("swapout: lost a swapout race?"));
1065 
1066 	/*
1067 	 * remember the process resident count
1068 	 */
1069 	p->p_vmspace->vm_swrss = vmspace_resident_count(p->p_vmspace);
1070 	/*
1071 	 * Check and mark all threads before we proceed.
1072 	 */
1073 	p->p_flag &= ~P_INMEM;
1074 	p->p_flag |= P_SWAPPINGOUT;
1075 	FOREACH_THREAD_IN_PROC(p, td) {
1076 		thread_lock(td);
1077 		if (!thread_safetoswapout(td)) {
1078 			thread_unlock(td);
1079 			swapclear(p);
1080 			return (EBUSY);
1081 		}
1082 		td->td_flags &= ~TDF_INMEM;
1083 		TD_SET_SWAPPED(td);
1084 		thread_unlock(td);
1085 	}
1086 	td = FIRST_THREAD_IN_PROC(p);
1087 	++td->td_ru.ru_nswap;
1088 	PROC_UNLOCK(p);
1089 
1090 	/*
1091 	 * This list is stable because all threads are now prevented from
1092 	 * running.  The list is only modified in the context of a running
1093 	 * thread in this process.
1094 	 */
1095 	FOREACH_THREAD_IN_PROC(p, td)
1096 		vm_thread_swapout(td);
1097 
1098 	PROC_LOCK(p);
1099 	p->p_flag &= ~P_SWAPPINGOUT;
1100 	p->p_swtick = ticks;
1101 	return (0);
1102 }
1103 #endif /* !NO_SWAPPING */
1104