xref: /freebsd/sys/vm/vm_glue.c (revision 6574b8ed19b093f0af09501d2c9676c28993cb97)
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_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_INACTIVE);
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 #ifndef NO_SWAPPING
490 /*
491  * Allow a thread's kernel stack to be paged out.
492  */
493 static void
494 vm_thread_swapout(struct thread *td)
495 {
496 	vm_object_t ksobj;
497 	vm_page_t m;
498 	int i, pages;
499 
500 	cpu_thread_swapout(td);
501 	pages = td->td_kstack_pages;
502 	ksobj = td->td_kstack_obj;
503 	pmap_qremove(td->td_kstack, pages);
504 	VM_OBJECT_WLOCK(ksobj);
505 	for (i = 0; i < pages; i++) {
506 		m = vm_page_lookup(ksobj, i);
507 		if (m == NULL)
508 			panic("vm_thread_swapout: kstack already missing?");
509 		vm_page_dirty(m);
510 		vm_page_lock(m);
511 		vm_page_unwire(m, PQ_INACTIVE);
512 		vm_page_unlock(m);
513 	}
514 	VM_OBJECT_WUNLOCK(ksobj);
515 }
516 
517 /*
518  * Bring the kernel stack for a specified thread back in.
519  */
520 static void
521 vm_thread_swapin(struct thread *td)
522 {
523 	vm_object_t ksobj;
524 	vm_page_t ma[KSTACK_MAX_PAGES];
525 	int i, j, k, pages, rv;
526 
527 	pages = td->td_kstack_pages;
528 	ksobj = td->td_kstack_obj;
529 	VM_OBJECT_WLOCK(ksobj);
530 	for (i = 0; i < pages; i++)
531 		ma[i] = vm_page_grab(ksobj, i, VM_ALLOC_NORMAL |
532 		    VM_ALLOC_WIRED);
533 	for (i = 0; i < pages; i++) {
534 		if (ma[i]->valid != VM_PAGE_BITS_ALL) {
535 			vm_page_assert_xbusied(ma[i]);
536 			vm_object_pip_add(ksobj, 1);
537 			for (j = i + 1; j < pages; j++) {
538 				if (ma[j]->valid != VM_PAGE_BITS_ALL)
539 					vm_page_assert_xbusied(ma[j]);
540 				if (ma[j]->valid == VM_PAGE_BITS_ALL)
541 					break;
542 			}
543 			rv = vm_pager_get_pages(ksobj, ma + i, j - i, 0);
544 			if (rv != VM_PAGER_OK)
545 	panic("vm_thread_swapin: cannot get kstack for proc: %d",
546 				    td->td_proc->p_pid);
547 			vm_object_pip_wakeup(ksobj);
548 			for (k = i; k < j; k++)
549 				ma[k] = vm_page_lookup(ksobj, k);
550 			vm_page_xunbusy(ma[i]);
551 		} else if (vm_page_xbusied(ma[i]))
552 			vm_page_xunbusy(ma[i]);
553 	}
554 	VM_OBJECT_WUNLOCK(ksobj);
555 	pmap_qenter(td->td_kstack, ma, pages);
556 	cpu_thread_swapin(td);
557 }
558 #endif /* !NO_SWAPPING */
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 void
633 faultin(p)
634 	struct proc *p;
635 {
636 #ifdef NO_SWAPPING
637 
638 	PROC_LOCK_ASSERT(p, MA_OWNED);
639 	if ((p->p_flag & P_INMEM) == 0)
640 		panic("faultin: proc swapped out with NO_SWAPPING!");
641 #else /* !NO_SWAPPING */
642 	struct thread *td;
643 
644 	PROC_LOCK_ASSERT(p, MA_OWNED);
645 	/*
646 	 * If another process is swapping in this process,
647 	 * just wait until it finishes.
648 	 */
649 	if (p->p_flag & P_SWAPPINGIN) {
650 		while (p->p_flag & P_SWAPPINGIN)
651 			msleep(&p->p_flag, &p->p_mtx, PVM, "faultin", 0);
652 		return;
653 	}
654 	if ((p->p_flag & P_INMEM) == 0) {
655 		/*
656 		 * Don't let another thread swap process p out while we are
657 		 * busy swapping it in.
658 		 */
659 		++p->p_lock;
660 		p->p_flag |= P_SWAPPINGIN;
661 		PROC_UNLOCK(p);
662 
663 		/*
664 		 * We hold no lock here because the list of threads
665 		 * can not change while all threads in the process are
666 		 * swapped out.
667 		 */
668 		FOREACH_THREAD_IN_PROC(p, td)
669 			vm_thread_swapin(td);
670 		PROC_LOCK(p);
671 		swapclear(p);
672 		p->p_swtick = ticks;
673 
674 		wakeup(&p->p_flag);
675 
676 		/* Allow other threads to swap p out now. */
677 		--p->p_lock;
678 	}
679 #endif /* NO_SWAPPING */
680 }
681 
682 /*
683  * This swapin algorithm attempts to swap-in processes only if there
684  * is enough space for them.  Of course, if a process waits for a long
685  * time, it will be swapped in anyway.
686  *
687  * Giant is held on entry.
688  */
689 void
690 swapper(void)
691 {
692 	struct proc *p;
693 	struct thread *td;
694 	struct proc *pp;
695 	int slptime;
696 	int swtime;
697 	int ppri;
698 	int pri;
699 
700 loop:
701 	if (vm_page_count_min()) {
702 		VM_WAIT;
703 		goto loop;
704 	}
705 
706 	pp = NULL;
707 	ppri = INT_MIN;
708 	sx_slock(&allproc_lock);
709 	FOREACH_PROC_IN_SYSTEM(p) {
710 		PROC_LOCK(p);
711 		if (p->p_state == PRS_NEW ||
712 		    p->p_flag & (P_SWAPPINGOUT | P_SWAPPINGIN | P_INMEM)) {
713 			PROC_UNLOCK(p);
714 			continue;
715 		}
716 		swtime = (ticks - p->p_swtick) / hz;
717 		FOREACH_THREAD_IN_PROC(p, td) {
718 			/*
719 			 * An otherwise runnable thread of a process
720 			 * swapped out has only the TDI_SWAPPED bit set.
721 			 *
722 			 */
723 			thread_lock(td);
724 			if (td->td_inhibitors == TDI_SWAPPED) {
725 				slptime = (ticks - td->td_slptick) / hz;
726 				pri = swtime + slptime;
727 				if ((td->td_flags & TDF_SWAPINREQ) == 0)
728 					pri -= p->p_nice * 8;
729 				/*
730 				 * if this thread is higher priority
731 				 * and there is enough space, then select
732 				 * this process instead of the previous
733 				 * selection.
734 				 */
735 				if (pri > ppri) {
736 					pp = p;
737 					ppri = pri;
738 				}
739 			}
740 			thread_unlock(td);
741 		}
742 		PROC_UNLOCK(p);
743 	}
744 	sx_sunlock(&allproc_lock);
745 
746 	/*
747 	 * Nothing to do, back to sleep.
748 	 */
749 	if ((p = pp) == NULL) {
750 		tsleep(&proc0, PVM, "swapin", MAXSLP * hz / 2);
751 		goto loop;
752 	}
753 	PROC_LOCK(p);
754 
755 	/*
756 	 * Another process may be bringing or may have already
757 	 * brought this process in while we traverse all threads.
758 	 * Or, this process may even be being swapped out again.
759 	 */
760 	if (p->p_flag & (P_INMEM | P_SWAPPINGOUT | P_SWAPPINGIN)) {
761 		PROC_UNLOCK(p);
762 		goto loop;
763 	}
764 
765 	/*
766 	 * We would like to bring someone in. (only if there is space).
767 	 * [What checks the space? ]
768 	 */
769 	faultin(p);
770 	PROC_UNLOCK(p);
771 	goto loop;
772 }
773 
774 void
775 kick_proc0(void)
776 {
777 
778 	wakeup(&proc0);
779 }
780 
781 #ifndef NO_SWAPPING
782 
783 /*
784  * Swap_idle_threshold1 is the guaranteed swapped in time for a process
785  */
786 static int swap_idle_threshold1 = 2;
787 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold1, CTLFLAG_RW,
788     &swap_idle_threshold1, 0, "Guaranteed swapped in time for a process");
789 
790 /*
791  * Swap_idle_threshold2 is the time that a process can be idle before
792  * it will be swapped out, if idle swapping is enabled.
793  */
794 static int swap_idle_threshold2 = 10;
795 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold2, CTLFLAG_RW,
796     &swap_idle_threshold2, 0, "Time before a process will be swapped out");
797 
798 /*
799  * First, if any processes have been sleeping or stopped for at least
800  * "swap_idle_threshold1" seconds, they are swapped out.  If, however,
801  * no such processes exist, then the longest-sleeping or stopped
802  * process is swapped out.  Finally, and only as a last resort, if
803  * there are no sleeping or stopped processes, the longest-resident
804  * process is swapped out.
805  */
806 void
807 swapout_procs(action)
808 int action;
809 {
810 	struct proc *p;
811 	struct thread *td;
812 	int didswap = 0;
813 
814 retry:
815 	sx_slock(&allproc_lock);
816 	FOREACH_PROC_IN_SYSTEM(p) {
817 		struct vmspace *vm;
818 		int minslptime = 100000;
819 		int slptime;
820 
821 		/*
822 		 * Watch out for a process in
823 		 * creation.  It may have no
824 		 * address space or lock yet.
825 		 */
826 		if (p->p_state == PRS_NEW)
827 			continue;
828 		/*
829 		 * An aio daemon switches its
830 		 * address space while running.
831 		 * Perform a quick check whether
832 		 * a process has P_SYSTEM.
833 		 */
834 		if ((p->p_flag & P_SYSTEM) != 0)
835 			continue;
836 		/*
837 		 * Do not swapout a process that
838 		 * is waiting for VM data
839 		 * structures as there is a possible
840 		 * deadlock.  Test this first as
841 		 * this may block.
842 		 *
843 		 * Lock the map until swapout
844 		 * finishes, or a thread of this
845 		 * process may attempt to alter
846 		 * the map.
847 		 */
848 		vm = vmspace_acquire_ref(p);
849 		if (vm == NULL)
850 			continue;
851 		if (!vm_map_trylock(&vm->vm_map))
852 			goto nextproc1;
853 
854 		PROC_LOCK(p);
855 		if (p->p_lock != 0 ||
856 		    (p->p_flag & (P_STOPPED_SINGLE|P_TRACED|P_SYSTEM|P_WEXIT)
857 		    ) != 0) {
858 			goto nextproc;
859 		}
860 		/*
861 		 * only aiod changes vmspace, however it will be
862 		 * skipped because of the if statement above checking
863 		 * for P_SYSTEM
864 		 */
865 		if ((p->p_flag & (P_INMEM|P_SWAPPINGOUT|P_SWAPPINGIN)) != P_INMEM)
866 			goto nextproc;
867 
868 		switch (p->p_state) {
869 		default:
870 			/* Don't swap out processes in any sort
871 			 * of 'special' state. */
872 			break;
873 
874 		case PRS_NORMAL:
875 			/*
876 			 * do not swapout a realtime process
877 			 * Check all the thread groups..
878 			 */
879 			FOREACH_THREAD_IN_PROC(p, td) {
880 				thread_lock(td);
881 				if (PRI_IS_REALTIME(td->td_pri_class)) {
882 					thread_unlock(td);
883 					goto nextproc;
884 				}
885 				slptime = (ticks - td->td_slptick) / hz;
886 				/*
887 				 * Guarantee swap_idle_threshold1
888 				 * time in memory.
889 				 */
890 				if (slptime < swap_idle_threshold1) {
891 					thread_unlock(td);
892 					goto nextproc;
893 				}
894 
895 				/*
896 				 * Do not swapout a process if it is
897 				 * waiting on a critical event of some
898 				 * kind or there is a thread whose
899 				 * pageable memory may be accessed.
900 				 *
901 				 * This could be refined to support
902 				 * swapping out a thread.
903 				 */
904 				if (!thread_safetoswapout(td)) {
905 					thread_unlock(td);
906 					goto nextproc;
907 				}
908 				/*
909 				 * If the system is under memory stress,
910 				 * or if we are swapping
911 				 * idle processes >= swap_idle_threshold2,
912 				 * then swap the process out.
913 				 */
914 				if (((action & VM_SWAP_NORMAL) == 0) &&
915 				    (((action & VM_SWAP_IDLE) == 0) ||
916 				    (slptime < swap_idle_threshold2))) {
917 					thread_unlock(td);
918 					goto nextproc;
919 				}
920 
921 				if (minslptime > slptime)
922 					minslptime = slptime;
923 				thread_unlock(td);
924 			}
925 
926 			/*
927 			 * If the pageout daemon didn't free enough pages,
928 			 * or if this process is idle and the system is
929 			 * configured to swap proactively, swap it out.
930 			 */
931 			if ((action & VM_SWAP_NORMAL) ||
932 				((action & VM_SWAP_IDLE) &&
933 				 (minslptime > swap_idle_threshold2))) {
934 				if (swapout(p) == 0)
935 					didswap++;
936 				PROC_UNLOCK(p);
937 				vm_map_unlock(&vm->vm_map);
938 				vmspace_free(vm);
939 				sx_sunlock(&allproc_lock);
940 				goto retry;
941 			}
942 		}
943 nextproc:
944 		PROC_UNLOCK(p);
945 		vm_map_unlock(&vm->vm_map);
946 nextproc1:
947 		vmspace_free(vm);
948 		continue;
949 	}
950 	sx_sunlock(&allproc_lock);
951 	/*
952 	 * If we swapped something out, and another process needed memory,
953 	 * then wakeup the sched process.
954 	 */
955 	if (didswap)
956 		wakeup(&proc0);
957 }
958 
959 static void
960 swapclear(p)
961 	struct proc *p;
962 {
963 	struct thread *td;
964 
965 	PROC_LOCK_ASSERT(p, MA_OWNED);
966 
967 	FOREACH_THREAD_IN_PROC(p, td) {
968 		thread_lock(td);
969 		td->td_flags |= TDF_INMEM;
970 		td->td_flags &= ~TDF_SWAPINREQ;
971 		TD_CLR_SWAPPED(td);
972 		if (TD_CAN_RUN(td))
973 			if (setrunnable(td)) {
974 #ifdef INVARIANTS
975 				/*
976 				 * XXX: We just cleared TDI_SWAPPED
977 				 * above and set TDF_INMEM, so this
978 				 * should never happen.
979 				 */
980 				panic("not waking up swapper");
981 #endif
982 			}
983 		thread_unlock(td);
984 	}
985 	p->p_flag &= ~(P_SWAPPINGIN|P_SWAPPINGOUT);
986 	p->p_flag |= P_INMEM;
987 }
988 
989 static int
990 swapout(p)
991 	struct proc *p;
992 {
993 	struct thread *td;
994 
995 	PROC_LOCK_ASSERT(p, MA_OWNED);
996 #if defined(SWAP_DEBUG)
997 	printf("swapping out %d\n", p->p_pid);
998 #endif
999 
1000 	/*
1001 	 * The states of this process and its threads may have changed
1002 	 * by now.  Assuming that there is only one pageout daemon thread,
1003 	 * this process should still be in memory.
1004 	 */
1005 	KASSERT((p->p_flag & (P_INMEM|P_SWAPPINGOUT|P_SWAPPINGIN)) == P_INMEM,
1006 		("swapout: lost a swapout race?"));
1007 
1008 	/*
1009 	 * remember the process resident count
1010 	 */
1011 	p->p_vmspace->vm_swrss = vmspace_resident_count(p->p_vmspace);
1012 	/*
1013 	 * Check and mark all threads before we proceed.
1014 	 */
1015 	p->p_flag &= ~P_INMEM;
1016 	p->p_flag |= P_SWAPPINGOUT;
1017 	FOREACH_THREAD_IN_PROC(p, td) {
1018 		thread_lock(td);
1019 		if (!thread_safetoswapout(td)) {
1020 			thread_unlock(td);
1021 			swapclear(p);
1022 			return (EBUSY);
1023 		}
1024 		td->td_flags &= ~TDF_INMEM;
1025 		TD_SET_SWAPPED(td);
1026 		thread_unlock(td);
1027 	}
1028 	td = FIRST_THREAD_IN_PROC(p);
1029 	++td->td_ru.ru_nswap;
1030 	PROC_UNLOCK(p);
1031 
1032 	/*
1033 	 * This list is stable because all threads are now prevented from
1034 	 * running.  The list is only modified in the context of a running
1035 	 * thread in this process.
1036 	 */
1037 	FOREACH_THREAD_IN_PROC(p, td)
1038 		vm_thread_swapout(td);
1039 
1040 	PROC_LOCK(p);
1041 	p->p_flag &= ~P_SWAPPINGOUT;
1042 	p->p_swtick = ticks;
1043 	return (0);
1044 }
1045 #endif /* !NO_SWAPPING */
1046