xref: /freebsd/sys/vm/vm_glue.c (revision 9fd69f37d28cfd7438cac3eeb45fe9dd46b4d7dd)
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/mutex.h>
71 #include <sys/proc.h>
72 #include <sys/resourcevar.h>
73 #include <sys/sched.h>
74 #include <sys/sf_buf.h>
75 #include <sys/shm.h>
76 #include <sys/vmmeter.h>
77 #include <sys/sx.h>
78 #include <sys/sysctl.h>
79 
80 #include <sys/eventhandler.h>
81 #include <sys/kernel.h>
82 #include <sys/ktr.h>
83 #include <sys/unistd.h>
84 
85 #include <vm/vm.h>
86 #include <vm/vm_param.h>
87 #include <vm/pmap.h>
88 #include <vm/vm_map.h>
89 #include <vm/vm_page.h>
90 #include <vm/vm_pageout.h>
91 #include <vm/vm_object.h>
92 #include <vm/vm_kern.h>
93 #include <vm/vm_extern.h>
94 #include <vm/vm_pager.h>
95 #include <vm/swap_pager.h>
96 
97 extern int maxslp;
98 
99 /*
100  * System initialization
101  *
102  * Note: proc0 from proc.h
103  */
104 static void vm_init_limits(void *);
105 SYSINIT(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  */
112 static void scheduler(void *);
113 SYSINIT(scheduler, SI_SUB_RUN_SCHEDULER, SI_ORDER_ANY, scheduler, NULL);
114 
115 #ifndef NO_SWAPPING
116 static int swapout(struct proc *);
117 static void swapclear(struct proc *);
118 #endif
119 
120 /*
121  * MPSAFE
122  *
123  * WARNING!  This code calls vm_map_check_protection() which only checks
124  * the associated vm_map_entry range.  It does not determine whether the
125  * contents of the memory is actually readable or writable.  In most cases
126  * just checking the vm_map_entry is sufficient within the kernel's address
127  * space.
128  */
129 int
130 kernacc(addr, len, rw)
131 	void *addr;
132 	int len, rw;
133 {
134 	boolean_t rv;
135 	vm_offset_t saddr, eaddr;
136 	vm_prot_t prot;
137 
138 	KASSERT((rw & ~VM_PROT_ALL) == 0,
139 	    ("illegal ``rw'' argument to kernacc (%x)\n", rw));
140 
141 	if ((vm_offset_t)addr + len > kernel_map->max_offset ||
142 	    (vm_offset_t)addr + len < (vm_offset_t)addr)
143 		return (FALSE);
144 
145 	prot = rw;
146 	saddr = trunc_page((vm_offset_t)addr);
147 	eaddr = round_page((vm_offset_t)addr + len);
148 	vm_map_lock_read(kernel_map);
149 	rv = vm_map_check_protection(kernel_map, saddr, eaddr, prot);
150 	vm_map_unlock_read(kernel_map);
151 	return (rv == TRUE);
152 }
153 
154 /*
155  * MPSAFE
156  *
157  * WARNING!  This code calls vm_map_check_protection() which only checks
158  * the associated vm_map_entry range.  It does not determine whether the
159  * contents of the memory is actually readable or writable.  vmapbuf(),
160  * vm_fault_quick(), or copyin()/copout()/su*()/fu*() functions should be
161  * used in conjuction with this call.
162  */
163 int
164 useracc(addr, len, rw)
165 	void *addr;
166 	int len, rw;
167 {
168 	boolean_t rv;
169 	vm_prot_t prot;
170 	vm_map_t map;
171 
172 	KASSERT((rw & ~VM_PROT_ALL) == 0,
173 	    ("illegal ``rw'' argument to useracc (%x)\n", rw));
174 	prot = rw;
175 	map = &curproc->p_vmspace->vm_map;
176 	if ((vm_offset_t)addr + len > vm_map_max(map) ||
177 	    (vm_offset_t)addr + len < (vm_offset_t)addr) {
178 		return (FALSE);
179 	}
180 	vm_map_lock_read(map);
181 	rv = vm_map_check_protection(map, trunc_page((vm_offset_t)addr),
182 	    round_page((vm_offset_t)addr + len), prot);
183 	vm_map_unlock_read(map);
184 	return (rv == TRUE);
185 }
186 
187 int
188 vslock(void *addr, size_t len)
189 {
190 	vm_offset_t end, last, start;
191 	vm_size_t npages;
192 	int error;
193 
194 	last = (vm_offset_t)addr + len;
195 	start = trunc_page((vm_offset_t)addr);
196 	end = round_page(last);
197 	if (last < (vm_offset_t)addr || end < (vm_offset_t)addr)
198 		return (EINVAL);
199 	npages = atop(end - start);
200 	if (npages > vm_page_max_wired)
201 		return (ENOMEM);
202 	PROC_LOCK(curproc);
203 	if (ptoa(npages +
204 	    pmap_wired_count(vm_map_pmap(&curproc->p_vmspace->vm_map))) >
205 	    lim_cur(curproc, RLIMIT_MEMLOCK)) {
206 		PROC_UNLOCK(curproc);
207 		return (ENOMEM);
208 	}
209 	PROC_UNLOCK(curproc);
210 #if 0
211 	/*
212 	 * XXX - not yet
213 	 *
214 	 * The limit for transient usage of wired pages should be
215 	 * larger than for "permanent" wired pages (mlock()).
216 	 *
217 	 * Also, the sysctl code, which is the only present user
218 	 * of vslock(), does a hard loop on EAGAIN.
219 	 */
220 	if (npages + cnt.v_wire_count > vm_page_max_wired)
221 		return (EAGAIN);
222 #endif
223 	error = vm_map_wire(&curproc->p_vmspace->vm_map, start, end,
224 	    VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES);
225 	/*
226 	 * Return EFAULT on error to match copy{in,out}() behaviour
227 	 * rather than returning ENOMEM like mlock() would.
228 	 */
229 	return (error == KERN_SUCCESS ? 0 : EFAULT);
230 }
231 
232 void
233 vsunlock(void *addr, size_t len)
234 {
235 
236 	/* Rely on the parameter sanity checks performed by vslock(). */
237 	(void)vm_map_unwire(&curproc->p_vmspace->vm_map,
238 	    trunc_page((vm_offset_t)addr), round_page((vm_offset_t)addr + len),
239 	    VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES);
240 }
241 
242 /*
243  * Pin the page contained within the given object at the given offset.  If the
244  * page is not resident, allocate and load it using the given object's pager.
245  * Return the pinned page if successful; otherwise, return NULL.
246  */
247 static vm_page_t
248 vm_imgact_hold_page(vm_object_t object, vm_ooffset_t offset)
249 {
250 	vm_page_t m, ma[1];
251 	vm_pindex_t pindex;
252 	int rv;
253 
254 	VM_OBJECT_LOCK(object);
255 	pindex = OFF_TO_IDX(offset);
256 	m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL | VM_ALLOC_RETRY);
257 	if (m->valid != VM_PAGE_BITS_ALL) {
258 		ma[0] = m;
259 		rv = vm_pager_get_pages(object, ma, 1, 0);
260 		m = vm_page_lookup(object, pindex);
261 		if (m == NULL)
262 			goto out;
263 		if (rv != VM_PAGER_OK) {
264 			vm_page_lock_queues();
265 			vm_page_free(m);
266 			vm_page_unlock_queues();
267 			m = NULL;
268 			goto out;
269 		}
270 	}
271 	vm_page_lock_queues();
272 	vm_page_hold(m);
273 	vm_page_unlock_queues();
274 	vm_page_wakeup(m);
275 out:
276 	VM_OBJECT_UNLOCK(object);
277 	return (m);
278 }
279 
280 /*
281  * Return a CPU private mapping to the page at the given offset within the
282  * given object.  The page is pinned before it is mapped.
283  */
284 struct sf_buf *
285 vm_imgact_map_page(vm_object_t object, vm_ooffset_t offset)
286 {
287 	vm_page_t m;
288 
289 	m = vm_imgact_hold_page(object, offset);
290 	if (m == NULL)
291 		return (NULL);
292 	sched_pin();
293 	return (sf_buf_alloc(m, SFB_CPUPRIVATE));
294 }
295 
296 /*
297  * Destroy the given CPU private mapping and unpin the page that it mapped.
298  */
299 void
300 vm_imgact_unmap_page(struct sf_buf *sf)
301 {
302 	vm_page_t m;
303 
304 	m = sf_buf_page(sf);
305 	sf_buf_free(sf);
306 	sched_unpin();
307 	vm_page_lock_queues();
308 	vm_page_unhold(m);
309 	vm_page_unlock_queues();
310 }
311 
312 void
313 vm_sync_icache(vm_map_t map, vm_offset_t va, vm_offset_t sz)
314 {
315 
316 	pmap_sync_icache(map->pmap, va, sz);
317 }
318 
319 struct kstack_cache_entry {
320 	vm_object_t ksobj;
321 	struct kstack_cache_entry *next_ks_entry;
322 };
323 
324 static struct kstack_cache_entry *kstack_cache;
325 static int kstack_cache_size = 128;
326 static int kstacks;
327 static struct mtx kstack_cache_mtx;
328 SYSCTL_INT(_vm, OID_AUTO, kstack_cache_size, CTLFLAG_RW, &kstack_cache_size, 0,
329     "");
330 SYSCTL_INT(_vm, OID_AUTO, kstacks, CTLFLAG_RD, &kstacks, 0,
331     "");
332 
333 #ifndef KSTACK_MAX_PAGES
334 #define KSTACK_MAX_PAGES 32
335 #endif
336 
337 /*
338  * Create the kernel stack (including pcb for i386) for a new thread.
339  * This routine directly affects the fork perf for a process and
340  * create performance for a thread.
341  */
342 int
343 vm_thread_new(struct thread *td, int pages)
344 {
345 	vm_object_t ksobj;
346 	vm_offset_t ks;
347 	vm_page_t m, ma[KSTACK_MAX_PAGES];
348 	struct kstack_cache_entry *ks_ce;
349 	int i;
350 
351 	/* Bounds check */
352 	if (pages <= 1)
353 		pages = KSTACK_PAGES;
354 	else if (pages > KSTACK_MAX_PAGES)
355 		pages = KSTACK_MAX_PAGES;
356 
357 	if (pages == KSTACK_PAGES) {
358 		mtx_lock(&kstack_cache_mtx);
359 		if (kstack_cache != NULL) {
360 			ks_ce = kstack_cache;
361 			kstack_cache = ks_ce->next_ks_entry;
362 			mtx_unlock(&kstack_cache_mtx);
363 
364 			td->td_kstack_obj = ks_ce->ksobj;
365 			td->td_kstack = (vm_offset_t)ks_ce;
366 			td->td_kstack_pages = KSTACK_PAGES;
367 			return (1);
368 		}
369 		mtx_unlock(&kstack_cache_mtx);
370 	}
371 
372 	/*
373 	 * Allocate an object for the kstack.
374 	 */
375 	ksobj = vm_object_allocate(OBJT_DEFAULT, pages);
376 
377 	/*
378 	 * Get a kernel virtual address for this thread's kstack.
379 	 */
380 	ks = kmem_alloc_nofault(kernel_map,
381 	   (pages + KSTACK_GUARD_PAGES) * PAGE_SIZE);
382 	if (ks == 0) {
383 		printf("vm_thread_new: kstack allocation failed\n");
384 		vm_object_deallocate(ksobj);
385 		return (0);
386 	}
387 
388 	atomic_add_int(&kstacks, 1);
389 	if (KSTACK_GUARD_PAGES != 0) {
390 		pmap_qremove(ks, KSTACK_GUARD_PAGES);
391 		ks += KSTACK_GUARD_PAGES * PAGE_SIZE;
392 	}
393 	td->td_kstack_obj = ksobj;
394 	td->td_kstack = ks;
395 	/*
396 	 * Knowing the number of pages allocated is useful when you
397 	 * want to deallocate them.
398 	 */
399 	td->td_kstack_pages = pages;
400 	/*
401 	 * For the length of the stack, link in a real page of ram for each
402 	 * page of stack.
403 	 */
404 	VM_OBJECT_LOCK(ksobj);
405 	for (i = 0; i < pages; i++) {
406 		/*
407 		 * Get a kernel stack page.
408 		 */
409 		m = vm_page_grab(ksobj, i, VM_ALLOC_NOBUSY |
410 		    VM_ALLOC_NORMAL | VM_ALLOC_RETRY | VM_ALLOC_WIRED);
411 		ma[i] = m;
412 		m->valid = VM_PAGE_BITS_ALL;
413 	}
414 	VM_OBJECT_UNLOCK(ksobj);
415 	pmap_qenter(ks, ma, pages);
416 	return (1);
417 }
418 
419 static void
420 vm_thread_stack_dispose(vm_object_t ksobj, vm_offset_t ks, int pages)
421 {
422 	vm_page_t m;
423 	int i;
424 
425 	atomic_add_int(&kstacks, -1);
426 	pmap_qremove(ks, pages);
427 	VM_OBJECT_LOCK(ksobj);
428 	for (i = 0; i < pages; i++) {
429 		m = vm_page_lookup(ksobj, i);
430 		if (m == NULL)
431 			panic("vm_thread_dispose: kstack already missing?");
432 		vm_page_lock_queues();
433 		vm_page_unwire(m, 0);
434 		vm_page_free(m);
435 		vm_page_unlock_queues();
436 	}
437 	VM_OBJECT_UNLOCK(ksobj);
438 	vm_object_deallocate(ksobj);
439 	kmem_free(kernel_map, ks - (KSTACK_GUARD_PAGES * PAGE_SIZE),
440 	    (pages + KSTACK_GUARD_PAGES) * PAGE_SIZE);
441 }
442 
443 /*
444  * Dispose of a thread's kernel stack.
445  */
446 void
447 vm_thread_dispose(struct thread *td)
448 {
449 	vm_object_t ksobj;
450 	vm_offset_t ks;
451 	struct kstack_cache_entry *ks_ce;
452 	int pages;
453 
454 	pages = td->td_kstack_pages;
455 	ksobj = td->td_kstack_obj;
456 	ks = td->td_kstack;
457 	td->td_kstack = 0;
458 	td->td_kstack_pages = 0;
459 	if (pages == KSTACK_PAGES && kstacks <= kstack_cache_size) {
460 		ks_ce = (struct kstack_cache_entry *)ks;
461 		ks_ce->ksobj = ksobj;
462 		mtx_lock(&kstack_cache_mtx);
463 		ks_ce->next_ks_entry = kstack_cache;
464 		kstack_cache = ks_ce;
465 		mtx_unlock(&kstack_cache_mtx);
466 		return;
467 	}
468 	vm_thread_stack_dispose(ksobj, ks, pages);
469 }
470 
471 static void
472 vm_thread_stack_lowmem(void *nulll)
473 {
474 	struct kstack_cache_entry *ks_ce, *ks_ce1;
475 
476 	mtx_lock(&kstack_cache_mtx);
477 	ks_ce = kstack_cache;
478 	kstack_cache = NULL;
479 	mtx_unlock(&kstack_cache_mtx);
480 
481 	while (ks_ce != NULL) {
482 		ks_ce1 = ks_ce;
483 		ks_ce = ks_ce->next_ks_entry;
484 
485 		vm_thread_stack_dispose(ks_ce1->ksobj, (vm_offset_t)ks_ce1,
486 		    KSTACK_PAGES);
487 	}
488 }
489 
490 static void
491 kstack_cache_init(void *nulll)
492 {
493 
494 	EVENTHANDLER_REGISTER(vm_lowmem, vm_thread_stack_lowmem, NULL,
495 	    EVENTHANDLER_PRI_ANY);
496 }
497 
498 MTX_SYSINIT(kstack_cache, &kstack_cache_mtx, "kstkch", MTX_DEF);
499 SYSINIT(vm_kstacks, SI_SUB_KTHREAD_INIT, SI_ORDER_ANY, kstack_cache_init, NULL);
500 
501 /*
502  * Allow a thread's kernel stack to be paged out.
503  */
504 void
505 vm_thread_swapout(struct thread *td)
506 {
507 	vm_object_t ksobj;
508 	vm_page_t m;
509 	int i, pages;
510 
511 	cpu_thread_swapout(td);
512 	pages = td->td_kstack_pages;
513 	ksobj = td->td_kstack_obj;
514 	pmap_qremove(td->td_kstack, pages);
515 	VM_OBJECT_LOCK(ksobj);
516 	for (i = 0; i < pages; i++) {
517 		m = vm_page_lookup(ksobj, i);
518 		if (m == NULL)
519 			panic("vm_thread_swapout: kstack already missing?");
520 		vm_page_lock_queues();
521 		vm_page_dirty(m);
522 		vm_page_unwire(m, 0);
523 		vm_page_unlock_queues();
524 	}
525 	VM_OBJECT_UNLOCK(ksobj);
526 }
527 
528 /*
529  * Bring the kernel stack for a specified thread back in.
530  */
531 void
532 vm_thread_swapin(struct thread *td)
533 {
534 	vm_object_t ksobj;
535 	vm_page_t m, ma[KSTACK_MAX_PAGES];
536 	int i, pages, rv;
537 
538 	pages = td->td_kstack_pages;
539 	ksobj = td->td_kstack_obj;
540 	VM_OBJECT_LOCK(ksobj);
541 	for (i = 0; i < pages; i++) {
542 		m = vm_page_grab(ksobj, i, VM_ALLOC_NORMAL | VM_ALLOC_RETRY);
543 		if (m->valid != VM_PAGE_BITS_ALL) {
544 			rv = vm_pager_get_pages(ksobj, &m, 1, 0);
545 			if (rv != VM_PAGER_OK)
546 				panic("vm_thread_swapin: cannot get kstack for proc: %d", td->td_proc->p_pid);
547 			m = vm_page_lookup(ksobj, i);
548 		}
549 		ma[i] = m;
550 		vm_page_lock_queues();
551 		vm_page_wire(m);
552 		vm_page_unlock_queues();
553 		vm_page_wakeup(m);
554 	}
555 	VM_OBJECT_UNLOCK(ksobj);
556 	pmap_qenter(td->td_kstack, ma, pages);
557 	cpu_thread_swapin(td);
558 }
559 
560 /*
561  * Implement fork's actions on an address space.
562  * Here we arrange for the address space to be copied or referenced,
563  * allocate a user struct (pcb and kernel stack), then call the
564  * machine-dependent layer to fill those in and make the new process
565  * ready to run.  The new process is set up so that it returns directly
566  * to user mode to avoid stack copying and relocation problems.
567  */
568 int
569 vm_forkproc(td, p2, td2, vm2, flags)
570 	struct thread *td;
571 	struct proc *p2;
572 	struct thread *td2;
573 	struct vmspace *vm2;
574 	int flags;
575 {
576 	struct proc *p1 = td->td_proc;
577 	int error;
578 
579 	if ((flags & RFPROC) == 0) {
580 		/*
581 		 * Divorce the memory, if it is shared, essentially
582 		 * this changes shared memory amongst threads, into
583 		 * COW locally.
584 		 */
585 		if ((flags & RFMEM) == 0) {
586 			if (p1->p_vmspace->vm_refcnt > 1) {
587 				error = vmspace_unshare(p1);
588 				if (error)
589 					return (error);
590 			}
591 		}
592 		cpu_fork(td, p2, td2, flags);
593 		return (0);
594 	}
595 
596 	if (flags & RFMEM) {
597 		p2->p_vmspace = p1->p_vmspace;
598 		atomic_add_int(&p1->p_vmspace->vm_refcnt, 1);
599 	}
600 
601 	while (vm_page_count_severe()) {
602 		VM_WAIT;
603 	}
604 
605 	if ((flags & RFMEM) == 0) {
606 		p2->p_vmspace = vm2;
607 		if (p1->p_vmspace->vm_shm)
608 			shmfork(p1, p2);
609 	}
610 
611 	/*
612 	 * cpu_fork will copy and update the pcb, set up the kernel stack,
613 	 * and make the child ready to run.
614 	 */
615 	cpu_fork(td, p2, td2, flags);
616 	return (0);
617 }
618 
619 /*
620  * Called after process has been wait(2)'ed apon and is being reaped.
621  * The idea is to reclaim resources that we could not reclaim while
622  * the process was still executing.
623  */
624 void
625 vm_waitproc(p)
626 	struct proc *p;
627 {
628 
629 	vmspace_exitfree(p);		/* and clean-out the vmspace */
630 }
631 
632 /*
633  * Set default limits for VM system.
634  * Called for proc 0, and then inherited by all others.
635  *
636  * XXX should probably act directly on proc0.
637  */
638 static void
639 vm_init_limits(udata)
640 	void *udata;
641 {
642 	struct proc *p = udata;
643 	struct plimit *limp;
644 	int rss_limit;
645 
646 	/*
647 	 * Set up the initial limits on process VM. Set the maximum resident
648 	 * set size to be half of (reasonably) available memory.  Since this
649 	 * is a soft limit, it comes into effect only when the system is out
650 	 * of memory - half of main memory helps to favor smaller processes,
651 	 * and reduces thrashing of the object cache.
652 	 */
653 	limp = p->p_limit;
654 	limp->pl_rlimit[RLIMIT_STACK].rlim_cur = dflssiz;
655 	limp->pl_rlimit[RLIMIT_STACK].rlim_max = maxssiz;
656 	limp->pl_rlimit[RLIMIT_DATA].rlim_cur = dfldsiz;
657 	limp->pl_rlimit[RLIMIT_DATA].rlim_max = maxdsiz;
658 	/* limit the limit to no less than 2MB */
659 	rss_limit = max(cnt.v_free_count, 512);
660 	limp->pl_rlimit[RLIMIT_RSS].rlim_cur = ptoa(rss_limit);
661 	limp->pl_rlimit[RLIMIT_RSS].rlim_max = RLIM_INFINITY;
662 }
663 
664 void
665 faultin(p)
666 	struct proc *p;
667 {
668 #ifdef NO_SWAPPING
669 
670 	PROC_LOCK_ASSERT(p, MA_OWNED);
671 	if ((p->p_flag & P_INMEM) == 0)
672 		panic("faultin: proc swapped out with NO_SWAPPING!");
673 #else /* !NO_SWAPPING */
674 	struct thread *td;
675 
676 	PROC_LOCK_ASSERT(p, MA_OWNED);
677 	/*
678 	 * If another process is swapping in this process,
679 	 * just wait until it finishes.
680 	 */
681 	if (p->p_flag & P_SWAPPINGIN) {
682 		while (p->p_flag & P_SWAPPINGIN)
683 			msleep(&p->p_flag, &p->p_mtx, PVM, "faultin", 0);
684 		return;
685 	}
686 	if ((p->p_flag & P_INMEM) == 0) {
687 		/*
688 		 * Don't let another thread swap process p out while we are
689 		 * busy swapping it in.
690 		 */
691 		++p->p_lock;
692 		p->p_flag |= P_SWAPPINGIN;
693 		PROC_UNLOCK(p);
694 
695 		/*
696 		 * We hold no lock here because the list of threads
697 		 * can not change while all threads in the process are
698 		 * swapped out.
699 		 */
700 		FOREACH_THREAD_IN_PROC(p, td)
701 			vm_thread_swapin(td);
702 		PROC_LOCK(p);
703 		swapclear(p);
704 		p->p_swtick = ticks;
705 
706 		wakeup(&p->p_flag);
707 
708 		/* Allow other threads to swap p out now. */
709 		--p->p_lock;
710 	}
711 #endif /* NO_SWAPPING */
712 }
713 
714 /*
715  * This swapin algorithm attempts to swap-in processes only if there
716  * is enough space for them.  Of course, if a process waits for a long
717  * time, it will be swapped in anyway.
718  *
719  * Giant is held on entry.
720  */
721 /* ARGSUSED*/
722 static void
723 scheduler(dummy)
724 	void *dummy;
725 {
726 	struct proc *p;
727 	struct thread *td;
728 	struct proc *pp;
729 	int slptime;
730 	int swtime;
731 	int ppri;
732 	int pri;
733 
734 	mtx_assert(&Giant, MA_OWNED | MA_NOTRECURSED);
735 	mtx_unlock(&Giant);
736 
737 loop:
738 	if (vm_page_count_min()) {
739 		VM_WAIT;
740 		goto loop;
741 	}
742 
743 	pp = NULL;
744 	ppri = INT_MIN;
745 	sx_slock(&allproc_lock);
746 	FOREACH_PROC_IN_SYSTEM(p) {
747 		PROC_LOCK(p);
748 		if (p->p_flag & (P_SWAPPINGOUT | P_SWAPPINGIN | P_INMEM)) {
749 			PROC_UNLOCK(p);
750 			continue;
751 		}
752 		swtime = (ticks - p->p_swtick) / hz;
753 		FOREACH_THREAD_IN_PROC(p, td) {
754 			/*
755 			 * An otherwise runnable thread of a process
756 			 * swapped out has only the TDI_SWAPPED bit set.
757 			 *
758 			 */
759 			thread_lock(td);
760 			if (td->td_inhibitors == TDI_SWAPPED) {
761 				slptime = (ticks - td->td_slptick) / hz;
762 				pri = swtime + slptime;
763 				if ((td->td_flags & TDF_SWAPINREQ) == 0)
764 					pri -= p->p_nice * 8;
765 				/*
766 				 * if this thread is higher priority
767 				 * and there is enough space, then select
768 				 * this process instead of the previous
769 				 * selection.
770 				 */
771 				if (pri > ppri) {
772 					pp = p;
773 					ppri = pri;
774 				}
775 			}
776 			thread_unlock(td);
777 		}
778 		PROC_UNLOCK(p);
779 	}
780 	sx_sunlock(&allproc_lock);
781 
782 	/*
783 	 * Nothing to do, back to sleep.
784 	 */
785 	if ((p = pp) == NULL) {
786 		tsleep(&proc0, PVM, "sched", maxslp * hz / 2);
787 		goto loop;
788 	}
789 	PROC_LOCK(p);
790 
791 	/*
792 	 * Another process may be bringing or may have already
793 	 * brought this process in while we traverse all threads.
794 	 * Or, this process may even be being swapped out again.
795 	 */
796 	if (p->p_flag & (P_INMEM | P_SWAPPINGOUT | P_SWAPPINGIN)) {
797 		PROC_UNLOCK(p);
798 		goto loop;
799 	}
800 
801 	/*
802 	 * We would like to bring someone in. (only if there is space).
803 	 * [What checks the space? ]
804 	 */
805 	faultin(p);
806 	PROC_UNLOCK(p);
807 	goto loop;
808 }
809 
810 void
811 kick_proc0(void)
812 {
813 
814 	wakeup(&proc0);
815 }
816 
817 #ifndef NO_SWAPPING
818 
819 /*
820  * Swap_idle_threshold1 is the guaranteed swapped in time for a process
821  */
822 static int swap_idle_threshold1 = 2;
823 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold1, CTLFLAG_RW,
824     &swap_idle_threshold1, 0, "Guaranteed swapped in time for a process");
825 
826 /*
827  * Swap_idle_threshold2 is the time that a process can be idle before
828  * it will be swapped out, if idle swapping is enabled.
829  */
830 static int swap_idle_threshold2 = 10;
831 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold2, CTLFLAG_RW,
832     &swap_idle_threshold2, 0, "Time before a process will be swapped out");
833 
834 /*
835  * Swapout is driven by the pageout daemon.  Very simple, we find eligible
836  * procs and swap out their stacks.  We try to always "swap" at least one
837  * process in case we need the room for a swapin.
838  * If any procs have been sleeping/stopped for at least maxslp seconds,
839  * they are swapped.  Else, we swap the longest-sleeping or stopped process,
840  * if any, otherwise the longest-resident process.
841  */
842 void
843 swapout_procs(action)
844 int action;
845 {
846 	struct proc *p;
847 	struct thread *td;
848 	int didswap = 0;
849 
850 retry:
851 	sx_slock(&allproc_lock);
852 	FOREACH_PROC_IN_SYSTEM(p) {
853 		struct vmspace *vm;
854 		int minslptime = 100000;
855 		int slptime;
856 
857 		/*
858 		 * Watch out for a process in
859 		 * creation.  It may have no
860 		 * address space or lock yet.
861 		 */
862 		if (p->p_state == PRS_NEW)
863 			continue;
864 		/*
865 		 * An aio daemon switches its
866 		 * address space while running.
867 		 * Perform a quick check whether
868 		 * a process has P_SYSTEM.
869 		 */
870 		if ((p->p_flag & P_SYSTEM) != 0)
871 			continue;
872 		/*
873 		 * Do not swapout a process that
874 		 * is waiting for VM data
875 		 * structures as there is a possible
876 		 * deadlock.  Test this first as
877 		 * this may block.
878 		 *
879 		 * Lock the map until swapout
880 		 * finishes, or a thread of this
881 		 * process may attempt to alter
882 		 * the map.
883 		 */
884 		vm = vmspace_acquire_ref(p);
885 		if (vm == NULL)
886 			continue;
887 		if (!vm_map_trylock(&vm->vm_map))
888 			goto nextproc1;
889 
890 		PROC_LOCK(p);
891 		if (p->p_lock != 0 ||
892 		    (p->p_flag & (P_STOPPED_SINGLE|P_TRACED|P_SYSTEM|P_WEXIT)
893 		    ) != 0) {
894 			goto nextproc;
895 		}
896 		/*
897 		 * only aiod changes vmspace, however it will be
898 		 * skipped because of the if statement above checking
899 		 * for P_SYSTEM
900 		 */
901 		if ((p->p_flag & (P_INMEM|P_SWAPPINGOUT|P_SWAPPINGIN)) != P_INMEM)
902 			goto nextproc;
903 
904 		switch (p->p_state) {
905 		default:
906 			/* Don't swap out processes in any sort
907 			 * of 'special' state. */
908 			break;
909 
910 		case PRS_NORMAL:
911 			/*
912 			 * do not swapout a realtime process
913 			 * Check all the thread groups..
914 			 */
915 			FOREACH_THREAD_IN_PROC(p, td) {
916 				thread_lock(td);
917 				if (PRI_IS_REALTIME(td->td_pri_class)) {
918 					thread_unlock(td);
919 					goto nextproc;
920 				}
921 				slptime = (ticks - td->td_slptick) / hz;
922 				/*
923 				 * Guarantee swap_idle_threshold1
924 				 * time in memory.
925 				 */
926 				if (slptime < swap_idle_threshold1) {
927 					thread_unlock(td);
928 					goto nextproc;
929 				}
930 
931 				/*
932 				 * Do not swapout a process if it is
933 				 * waiting on a critical event of some
934 				 * kind or there is a thread whose
935 				 * pageable memory may be accessed.
936 				 *
937 				 * This could be refined to support
938 				 * swapping out a thread.
939 				 */
940 				if (!thread_safetoswapout(td)) {
941 					thread_unlock(td);
942 					goto nextproc;
943 				}
944 				/*
945 				 * If the system is under memory stress,
946 				 * or if we are swapping
947 				 * idle processes >= swap_idle_threshold2,
948 				 * then swap the process out.
949 				 */
950 				if (((action & VM_SWAP_NORMAL) == 0) &&
951 				    (((action & VM_SWAP_IDLE) == 0) ||
952 				    (slptime < swap_idle_threshold2))) {
953 					thread_unlock(td);
954 					goto nextproc;
955 				}
956 
957 				if (minslptime > slptime)
958 					minslptime = slptime;
959 				thread_unlock(td);
960 			}
961 
962 			/*
963 			 * If the pageout daemon didn't free enough pages,
964 			 * or if this process is idle and the system is
965 			 * configured to swap proactively, swap it out.
966 			 */
967 			if ((action & VM_SWAP_NORMAL) ||
968 				((action & VM_SWAP_IDLE) &&
969 				 (minslptime > swap_idle_threshold2))) {
970 				if (swapout(p) == 0)
971 					didswap++;
972 				PROC_UNLOCK(p);
973 				vm_map_unlock(&vm->vm_map);
974 				vmspace_free(vm);
975 				sx_sunlock(&allproc_lock);
976 				goto retry;
977 			}
978 		}
979 nextproc:
980 		PROC_UNLOCK(p);
981 		vm_map_unlock(&vm->vm_map);
982 nextproc1:
983 		vmspace_free(vm);
984 		continue;
985 	}
986 	sx_sunlock(&allproc_lock);
987 	/*
988 	 * If we swapped something out, and another process needed memory,
989 	 * then wakeup the sched process.
990 	 */
991 	if (didswap)
992 		wakeup(&proc0);
993 }
994 
995 static void
996 swapclear(p)
997 	struct proc *p;
998 {
999 	struct thread *td;
1000 
1001 	PROC_LOCK_ASSERT(p, MA_OWNED);
1002 
1003 	FOREACH_THREAD_IN_PROC(p, td) {
1004 		thread_lock(td);
1005 		td->td_flags |= TDF_INMEM;
1006 		td->td_flags &= ~TDF_SWAPINREQ;
1007 		TD_CLR_SWAPPED(td);
1008 		if (TD_CAN_RUN(td))
1009 			if (setrunnable(td)) {
1010 #ifdef INVARIANTS
1011 				/*
1012 				 * XXX: We just cleared TDI_SWAPPED
1013 				 * above and set TDF_INMEM, so this
1014 				 * should never happen.
1015 				 */
1016 				panic("not waking up swapper");
1017 #endif
1018 			}
1019 		thread_unlock(td);
1020 	}
1021 	p->p_flag &= ~(P_SWAPPINGIN|P_SWAPPINGOUT);
1022 	p->p_flag |= P_INMEM;
1023 }
1024 
1025 static int
1026 swapout(p)
1027 	struct proc *p;
1028 {
1029 	struct thread *td;
1030 
1031 	PROC_LOCK_ASSERT(p, MA_OWNED);
1032 #if defined(SWAP_DEBUG)
1033 	printf("swapping out %d\n", p->p_pid);
1034 #endif
1035 
1036 	/*
1037 	 * The states of this process and its threads may have changed
1038 	 * by now.  Assuming that there is only one pageout daemon thread,
1039 	 * this process should still be in memory.
1040 	 */
1041 	KASSERT((p->p_flag & (P_INMEM|P_SWAPPINGOUT|P_SWAPPINGIN)) == P_INMEM,
1042 		("swapout: lost a swapout race?"));
1043 
1044 	/*
1045 	 * remember the process resident count
1046 	 */
1047 	p->p_vmspace->vm_swrss = vmspace_resident_count(p->p_vmspace);
1048 	/*
1049 	 * Check and mark all threads before we proceed.
1050 	 */
1051 	p->p_flag &= ~P_INMEM;
1052 	p->p_flag |= P_SWAPPINGOUT;
1053 	FOREACH_THREAD_IN_PROC(p, td) {
1054 		thread_lock(td);
1055 		if (!thread_safetoswapout(td)) {
1056 			thread_unlock(td);
1057 			swapclear(p);
1058 			return (EBUSY);
1059 		}
1060 		td->td_flags &= ~TDF_INMEM;
1061 		TD_SET_SWAPPED(td);
1062 		thread_unlock(td);
1063 	}
1064 	td = FIRST_THREAD_IN_PROC(p);
1065 	++td->td_ru.ru_nswap;
1066 	PROC_UNLOCK(p);
1067 
1068 	/*
1069 	 * This list is stable because all threads are now prevented from
1070 	 * running.  The list is only modified in the context of a running
1071 	 * thread in this process.
1072 	 */
1073 	FOREACH_THREAD_IN_PROC(p, td)
1074 		vm_thread_swapout(td);
1075 
1076 	PROC_LOCK(p);
1077 	p->p_flag &= ~P_SWAPPINGOUT;
1078 	p->p_swtick = ticks;
1079 	return (0);
1080 }
1081 #endif /* !NO_SWAPPING */
1082