xref: /freebsd/sys/vm/vm_glue.c (revision 145992504973bd16cf3518af9ba5ce185fefa82a)
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/racct.h>
73 #include <sys/resourcevar.h>
74 #include <sys/sched.h>
75 #include <sys/sf_buf.h>
76 #include <sys/shm.h>
77 #include <sys/vmmeter.h>
78 #include <sys/sx.h>
79 #include <sys/sysctl.h>
80 #include <sys/_kstack_cache.h>
81 #include <sys/eventhandler.h>
82 #include <sys/kernel.h>
83 #include <sys/ktr.h>
84 #include <sys/unistd.h>
85 
86 #include <vm/vm.h>
87 #include <vm/vm_param.h>
88 #include <vm/pmap.h>
89 #include <vm/vm_map.h>
90 #include <vm/vm_page.h>
91 #include <vm/vm_pageout.h>
92 #include <vm/vm_object.h>
93 #include <vm/vm_kern.h>
94 #include <vm/vm_extern.h>
95 #include <vm/vm_pager.h>
96 #include <vm/swap_pager.h>
97 
98 /*
99  * System initialization
100  *
101  * THIS MUST BE THE LAST INITIALIZATION ITEM!!!
102  *
103  * Note: run scheduling should be divorced from the vm system.
104  */
105 static void scheduler(void *);
106 SYSINIT(scheduler, SI_SUB_RUN_SCHEDULER, SI_ORDER_ANY, scheduler, NULL);
107 
108 #ifndef NO_SWAPPING
109 static int swapout(struct proc *);
110 static void swapclear(struct proc *);
111 static void vm_thread_swapin(struct thread *td);
112 static void vm_thread_swapout(struct thread *td);
113 #endif
114 
115 /*
116  * MPSAFE
117  *
118  * WARNING!  This code calls vm_map_check_protection() which only checks
119  * the associated vm_map_entry range.  It does not determine whether the
120  * contents of the memory is actually readable or writable.  In most cases
121  * just checking the vm_map_entry is sufficient within the kernel's address
122  * space.
123  */
124 int
125 kernacc(addr, len, rw)
126 	void *addr;
127 	int len, rw;
128 {
129 	boolean_t rv;
130 	vm_offset_t saddr, eaddr;
131 	vm_prot_t prot;
132 
133 	KASSERT((rw & ~VM_PROT_ALL) == 0,
134 	    ("illegal ``rw'' argument to kernacc (%x)\n", rw));
135 
136 	if ((vm_offset_t)addr + len > kernel_map->max_offset ||
137 	    (vm_offset_t)addr + len < (vm_offset_t)addr)
138 		return (FALSE);
139 
140 	prot = rw;
141 	saddr = trunc_page((vm_offset_t)addr);
142 	eaddr = round_page((vm_offset_t)addr + len);
143 	vm_map_lock_read(kernel_map);
144 	rv = vm_map_check_protection(kernel_map, saddr, eaddr, prot);
145 	vm_map_unlock_read(kernel_map);
146 	return (rv == TRUE);
147 }
148 
149 /*
150  * MPSAFE
151  *
152  * WARNING!  This code calls vm_map_check_protection() which only checks
153  * the associated vm_map_entry range.  It does not determine whether the
154  * contents of the memory is actually readable or writable.  vmapbuf(),
155  * vm_fault_quick(), or copyin()/copout()/su*()/fu*() functions should be
156  * used in conjuction with this call.
157  */
158 int
159 useracc(addr, len, rw)
160 	void *addr;
161 	int len, rw;
162 {
163 	boolean_t rv;
164 	vm_prot_t prot;
165 	vm_map_t map;
166 
167 	KASSERT((rw & ~VM_PROT_ALL) == 0,
168 	    ("illegal ``rw'' argument to useracc (%x)\n", rw));
169 	prot = rw;
170 	map = &curproc->p_vmspace->vm_map;
171 	if ((vm_offset_t)addr + len > vm_map_max(map) ||
172 	    (vm_offset_t)addr + len < (vm_offset_t)addr) {
173 		return (FALSE);
174 	}
175 	vm_map_lock_read(map);
176 	rv = vm_map_check_protection(map, trunc_page((vm_offset_t)addr),
177 	    round_page((vm_offset_t)addr + len), prot);
178 	vm_map_unlock_read(map);
179 	return (rv == TRUE);
180 }
181 
182 int
183 vslock(void *addr, size_t len)
184 {
185 	vm_offset_t end, last, start;
186 	vm_size_t npages;
187 	int error;
188 
189 	last = (vm_offset_t)addr + len;
190 	start = trunc_page((vm_offset_t)addr);
191 	end = round_page(last);
192 	if (last < (vm_offset_t)addr || end < (vm_offset_t)addr)
193 		return (EINVAL);
194 	npages = atop(end - start);
195 	if (npages > vm_page_max_wired)
196 		return (ENOMEM);
197 #if 0
198 	/*
199 	 * XXX - not yet
200 	 *
201 	 * The limit for transient usage of wired pages should be
202 	 * larger than for "permanent" wired pages (mlock()).
203 	 *
204 	 * Also, the sysctl code, which is the only present user
205 	 * of vslock(), does a hard loop on EAGAIN.
206 	 */
207 	if (npages + cnt.v_wire_count > vm_page_max_wired)
208 		return (EAGAIN);
209 #endif
210 	error = vm_map_wire(&curproc->p_vmspace->vm_map, start, end,
211 	    VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES);
212 	/*
213 	 * Return EFAULT on error to match copy{in,out}() behaviour
214 	 * rather than returning ENOMEM like mlock() would.
215 	 */
216 	return (error == KERN_SUCCESS ? 0 : EFAULT);
217 }
218 
219 void
220 vsunlock(void *addr, size_t len)
221 {
222 
223 	/* Rely on the parameter sanity checks performed by vslock(). */
224 	(void)vm_map_unwire(&curproc->p_vmspace->vm_map,
225 	    trunc_page((vm_offset_t)addr), round_page((vm_offset_t)addr + len),
226 	    VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES);
227 }
228 
229 /*
230  * Pin the page contained within the given object at the given offset.  If the
231  * page is not resident, allocate and load it using the given object's pager.
232  * Return the pinned page if successful; otherwise, return NULL.
233  */
234 static vm_page_t
235 vm_imgact_hold_page(vm_object_t object, vm_ooffset_t offset)
236 {
237 	vm_page_t m, ma[1];
238 	vm_pindex_t pindex;
239 	int rv;
240 
241 	VM_OBJECT_LOCK(object);
242 	pindex = OFF_TO_IDX(offset);
243 	m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL | VM_ALLOC_RETRY);
244 	if (m->valid != VM_PAGE_BITS_ALL) {
245 		ma[0] = m;
246 		rv = vm_pager_get_pages(object, ma, 1, 0);
247 		m = vm_page_lookup(object, pindex);
248 		if (m == NULL)
249 			goto out;
250 		if (rv != VM_PAGER_OK) {
251 			vm_page_lock(m);
252 			vm_page_free(m);
253 			vm_page_unlock(m);
254 			m = NULL;
255 			goto out;
256 		}
257 	}
258 	vm_page_lock(m);
259 	vm_page_hold(m);
260 	vm_page_unlock(m);
261 	vm_page_wakeup(m);
262 out:
263 	VM_OBJECT_UNLOCK(object);
264 	return (m);
265 }
266 
267 /*
268  * Return a CPU private mapping to the page at the given offset within the
269  * given object.  The page is pinned before it is mapped.
270  */
271 struct sf_buf *
272 vm_imgact_map_page(vm_object_t object, vm_ooffset_t offset)
273 {
274 	vm_page_t m;
275 
276 	m = vm_imgact_hold_page(object, offset);
277 	if (m == NULL)
278 		return (NULL);
279 	sched_pin();
280 	return (sf_buf_alloc(m, SFB_CPUPRIVATE));
281 }
282 
283 /*
284  * Destroy the given CPU private mapping and unpin the page that it mapped.
285  */
286 void
287 vm_imgact_unmap_page(struct sf_buf *sf)
288 {
289 	vm_page_t m;
290 
291 	m = sf_buf_page(sf);
292 	sf_buf_free(sf);
293 	sched_unpin();
294 	vm_page_lock(m);
295 	vm_page_unhold(m);
296 	vm_page_unlock(m);
297 }
298 
299 void
300 vm_sync_icache(vm_map_t map, vm_offset_t va, vm_offset_t sz)
301 {
302 
303 	pmap_sync_icache(map->pmap, va, sz);
304 }
305 
306 struct kstack_cache_entry *kstack_cache;
307 static int kstack_cache_size = 128;
308 static int kstacks;
309 static struct mtx kstack_cache_mtx;
310 SYSCTL_INT(_vm, OID_AUTO, kstack_cache_size, CTLFLAG_RW, &kstack_cache_size, 0,
311     "");
312 SYSCTL_INT(_vm, OID_AUTO, kstacks, CTLFLAG_RD, &kstacks, 0,
313     "");
314 
315 #ifndef KSTACK_MAX_PAGES
316 #define KSTACK_MAX_PAGES 32
317 #endif
318 
319 /*
320  * Create the kernel stack (including pcb for i386) for a new thread.
321  * This routine directly affects the fork perf for a process and
322  * create performance for a thread.
323  */
324 int
325 vm_thread_new(struct thread *td, int pages)
326 {
327 	vm_object_t ksobj;
328 	vm_offset_t ks;
329 	vm_page_t m, ma[KSTACK_MAX_PAGES];
330 	struct kstack_cache_entry *ks_ce;
331 	int i;
332 
333 	/* Bounds check */
334 	if (pages <= 1)
335 		pages = KSTACK_PAGES;
336 	else if (pages > KSTACK_MAX_PAGES)
337 		pages = KSTACK_MAX_PAGES;
338 
339 	if (pages == KSTACK_PAGES) {
340 		mtx_lock(&kstack_cache_mtx);
341 		if (kstack_cache != NULL) {
342 			ks_ce = kstack_cache;
343 			kstack_cache = ks_ce->next_ks_entry;
344 			mtx_unlock(&kstack_cache_mtx);
345 
346 			td->td_kstack_obj = ks_ce->ksobj;
347 			td->td_kstack = (vm_offset_t)ks_ce;
348 			td->td_kstack_pages = KSTACK_PAGES;
349 			return (1);
350 		}
351 		mtx_unlock(&kstack_cache_mtx);
352 	}
353 
354 	/*
355 	 * Allocate an object for the kstack.
356 	 */
357 	ksobj = vm_object_allocate(OBJT_DEFAULT, pages);
358 
359 	/*
360 	 * Get a kernel virtual address for this thread's kstack.
361 	 */
362 #if defined(__mips__)
363 	/*
364 	 * We need to align the kstack's mapped address to fit within
365 	 * a single TLB entry.
366 	 */
367 	ks = kmem_alloc_nofault_space(kernel_map,
368 	    (pages + KSTACK_GUARD_PAGES) * PAGE_SIZE, VMFS_TLB_ALIGNED_SPACE);
369 #else
370 	ks = kmem_alloc_nofault(kernel_map,
371 	   (pages + KSTACK_GUARD_PAGES) * PAGE_SIZE);
372 #endif
373 	if (ks == 0) {
374 		printf("vm_thread_new: kstack allocation failed\n");
375 		vm_object_deallocate(ksobj);
376 		return (0);
377 	}
378 
379 	atomic_add_int(&kstacks, 1);
380 	if (KSTACK_GUARD_PAGES != 0) {
381 		pmap_qremove(ks, KSTACK_GUARD_PAGES);
382 		ks += KSTACK_GUARD_PAGES * PAGE_SIZE;
383 	}
384 	td->td_kstack_obj = ksobj;
385 	td->td_kstack = ks;
386 	/*
387 	 * Knowing the number of pages allocated is useful when you
388 	 * want to deallocate them.
389 	 */
390 	td->td_kstack_pages = pages;
391 	/*
392 	 * For the length of the stack, link in a real page of ram for each
393 	 * page of stack.
394 	 */
395 	VM_OBJECT_LOCK(ksobj);
396 	for (i = 0; i < pages; i++) {
397 		/*
398 		 * Get a kernel stack page.
399 		 */
400 		m = vm_page_grab(ksobj, i, VM_ALLOC_NOBUSY |
401 		    VM_ALLOC_NORMAL | VM_ALLOC_RETRY | VM_ALLOC_WIRED);
402 		ma[i] = m;
403 		m->valid = VM_PAGE_BITS_ALL;
404 	}
405 	VM_OBJECT_UNLOCK(ksobj);
406 	pmap_qenter(ks, ma, pages);
407 	return (1);
408 }
409 
410 static void
411 vm_thread_stack_dispose(vm_object_t ksobj, vm_offset_t ks, int pages)
412 {
413 	vm_page_t m;
414 	int i;
415 
416 	atomic_add_int(&kstacks, -1);
417 	pmap_qremove(ks, pages);
418 	VM_OBJECT_LOCK(ksobj);
419 	for (i = 0; i < pages; i++) {
420 		m = vm_page_lookup(ksobj, i);
421 		if (m == NULL)
422 			panic("vm_thread_dispose: kstack already missing?");
423 		vm_page_lock(m);
424 		vm_page_unwire(m, 0);
425 		vm_page_free(m);
426 		vm_page_unlock(m);
427 	}
428 	VM_OBJECT_UNLOCK(ksobj);
429 	vm_object_deallocate(ksobj);
430 	kmem_free(kernel_map, ks - (KSTACK_GUARD_PAGES * PAGE_SIZE),
431 	    (pages + KSTACK_GUARD_PAGES) * PAGE_SIZE);
432 }
433 
434 /*
435  * Dispose of a thread's kernel stack.
436  */
437 void
438 vm_thread_dispose(struct thread *td)
439 {
440 	vm_object_t ksobj;
441 	vm_offset_t ks;
442 	struct kstack_cache_entry *ks_ce;
443 	int pages;
444 
445 	pages = td->td_kstack_pages;
446 	ksobj = td->td_kstack_obj;
447 	ks = td->td_kstack;
448 	td->td_kstack = 0;
449 	td->td_kstack_pages = 0;
450 	if (pages == KSTACK_PAGES && kstacks <= kstack_cache_size) {
451 		ks_ce = (struct kstack_cache_entry *)ks;
452 		ks_ce->ksobj = ksobj;
453 		mtx_lock(&kstack_cache_mtx);
454 		ks_ce->next_ks_entry = kstack_cache;
455 		kstack_cache = ks_ce;
456 		mtx_unlock(&kstack_cache_mtx);
457 		return;
458 	}
459 	vm_thread_stack_dispose(ksobj, ks, pages);
460 }
461 
462 static void
463 vm_thread_stack_lowmem(void *nulll)
464 {
465 	struct kstack_cache_entry *ks_ce, *ks_ce1;
466 
467 	mtx_lock(&kstack_cache_mtx);
468 	ks_ce = kstack_cache;
469 	kstack_cache = NULL;
470 	mtx_unlock(&kstack_cache_mtx);
471 
472 	while (ks_ce != NULL) {
473 		ks_ce1 = ks_ce;
474 		ks_ce = ks_ce->next_ks_entry;
475 
476 		vm_thread_stack_dispose(ks_ce1->ksobj, (vm_offset_t)ks_ce1,
477 		    KSTACK_PAGES);
478 	}
479 }
480 
481 static void
482 kstack_cache_init(void *nulll)
483 {
484 
485 	EVENTHANDLER_REGISTER(vm_lowmem, vm_thread_stack_lowmem, NULL,
486 	    EVENTHANDLER_PRI_ANY);
487 }
488 
489 MTX_SYSINIT(kstack_cache, &kstack_cache_mtx, "kstkch", MTX_DEF);
490 SYSINIT(vm_kstacks, SI_SUB_KTHREAD_INIT, SI_ORDER_ANY, kstack_cache_init, NULL);
491 
492 #ifndef NO_SWAPPING
493 /*
494  * Allow a thread's kernel stack to be paged out.
495  */
496 static void
497 vm_thread_swapout(struct thread *td)
498 {
499 	vm_object_t ksobj;
500 	vm_page_t m;
501 	int i, pages;
502 
503 	cpu_thread_swapout(td);
504 	pages = td->td_kstack_pages;
505 	ksobj = td->td_kstack_obj;
506 	pmap_qremove(td->td_kstack, pages);
507 	VM_OBJECT_LOCK(ksobj);
508 	for (i = 0; i < pages; i++) {
509 		m = vm_page_lookup(ksobj, i);
510 		if (m == NULL)
511 			panic("vm_thread_swapout: kstack already missing?");
512 		vm_page_dirty(m);
513 		vm_page_lock(m);
514 		vm_page_unwire(m, 0);
515 		vm_page_unlock(m);
516 	}
517 	VM_OBJECT_UNLOCK(ksobj);
518 }
519 
520 /*
521  * Bring the kernel stack for a specified thread back in.
522  */
523 static void
524 vm_thread_swapin(struct thread *td)
525 {
526 	vm_object_t ksobj;
527 	vm_page_t ma[KSTACK_MAX_PAGES];
528 	int i, j, k, pages, rv;
529 
530 	pages = td->td_kstack_pages;
531 	ksobj = td->td_kstack_obj;
532 	VM_OBJECT_LOCK(ksobj);
533 	for (i = 0; i < pages; i++)
534 		ma[i] = vm_page_grab(ksobj, i, VM_ALLOC_NORMAL | VM_ALLOC_RETRY |
535 		    VM_ALLOC_WIRED);
536 	for (i = 0; i < pages; i++) {
537 		if (ma[i]->valid != VM_PAGE_BITS_ALL) {
538 			KASSERT(ma[i]->oflags & VPO_BUSY,
539 			    ("lost busy 1"));
540 			vm_object_pip_add(ksobj, 1);
541 			for (j = i + 1; j < pages; j++) {
542 				KASSERT(ma[j]->valid == VM_PAGE_BITS_ALL ||
543 				    (ma[j]->oflags & VPO_BUSY),
544 				    ("lost busy 2"));
545 				if (ma[j]->valid == VM_PAGE_BITS_ALL)
546 					break;
547 			}
548 			rv = vm_pager_get_pages(ksobj, ma + i, j - i, 0);
549 			if (rv != VM_PAGER_OK)
550 	panic("vm_thread_swapin: cannot get kstack for proc: %d",
551 				    td->td_proc->p_pid);
552 			vm_object_pip_wakeup(ksobj);
553 			for (k = i; k < j; k++)
554 				ma[k] = vm_page_lookup(ksobj, k);
555 			vm_page_wakeup(ma[i]);
556 		} else if (ma[i]->oflags & VPO_BUSY)
557 			vm_page_wakeup(ma[i]);
558 	}
559 	VM_OBJECT_UNLOCK(ksobj);
560 	pmap_qenter(td->td_kstack, ma, pages);
561 	cpu_thread_swapin(td);
562 }
563 #endif /* !NO_SWAPPING */
564 
565 /*
566  * Implement fork's actions on an address space.
567  * Here we arrange for the address space to be copied or referenced,
568  * allocate a user struct (pcb and kernel stack), then call the
569  * machine-dependent layer to fill those in and make the new process
570  * ready to run.  The new process is set up so that it returns directly
571  * to user mode to avoid stack copying and relocation problems.
572  */
573 int
574 vm_forkproc(td, p2, td2, vm2, flags)
575 	struct thread *td;
576 	struct proc *p2;
577 	struct thread *td2;
578 	struct vmspace *vm2;
579 	int flags;
580 {
581 	struct proc *p1 = td->td_proc;
582 	int error;
583 
584 	if ((flags & RFPROC) == 0) {
585 		/*
586 		 * Divorce the memory, if it is shared, essentially
587 		 * this changes shared memory amongst threads, into
588 		 * COW locally.
589 		 */
590 		if ((flags & RFMEM) == 0) {
591 			if (p1->p_vmspace->vm_refcnt > 1) {
592 				error = vmspace_unshare(p1);
593 				if (error)
594 					return (error);
595 			}
596 		}
597 		cpu_fork(td, p2, td2, flags);
598 		return (0);
599 	}
600 
601 	if (flags & RFMEM) {
602 		p2->p_vmspace = p1->p_vmspace;
603 		atomic_add_int(&p1->p_vmspace->vm_refcnt, 1);
604 	}
605 
606 	while (vm_page_count_severe()) {
607 		VM_WAIT;
608 	}
609 
610 	if ((flags & RFMEM) == 0) {
611 		p2->p_vmspace = vm2;
612 		if (p1->p_vmspace->vm_shm)
613 			shmfork(p1, p2);
614 	}
615 
616 	/*
617 	 * cpu_fork will copy and update the pcb, set up the kernel stack,
618 	 * and make the child ready to run.
619 	 */
620 	cpu_fork(td, p2, td2, flags);
621 	return (0);
622 }
623 
624 /*
625  * Called after process has been wait(2)'ed apon and is being reaped.
626  * The idea is to reclaim resources that we could not reclaim while
627  * the process was still executing.
628  */
629 void
630 vm_waitproc(p)
631 	struct proc *p;
632 {
633 
634 	vmspace_exitfree(p);		/* and clean-out the vmspace */
635 }
636 
637 void
638 faultin(p)
639 	struct proc *p;
640 {
641 #ifdef NO_SWAPPING
642 
643 	PROC_LOCK_ASSERT(p, MA_OWNED);
644 	if ((p->p_flag & P_INMEM) == 0)
645 		panic("faultin: proc swapped out with NO_SWAPPING!");
646 #else /* !NO_SWAPPING */
647 	struct thread *td;
648 
649 	PROC_LOCK_ASSERT(p, MA_OWNED);
650 	/*
651 	 * If another process is swapping in this process,
652 	 * just wait until it finishes.
653 	 */
654 	if (p->p_flag & P_SWAPPINGIN) {
655 		while (p->p_flag & P_SWAPPINGIN)
656 			msleep(&p->p_flag, &p->p_mtx, PVM, "faultin", 0);
657 		return;
658 	}
659 	if ((p->p_flag & P_INMEM) == 0) {
660 		/*
661 		 * Don't let another thread swap process p out while we are
662 		 * busy swapping it in.
663 		 */
664 		++p->p_lock;
665 		p->p_flag |= P_SWAPPINGIN;
666 		PROC_UNLOCK(p);
667 
668 		/*
669 		 * We hold no lock here because the list of threads
670 		 * can not change while all threads in the process are
671 		 * swapped out.
672 		 */
673 		FOREACH_THREAD_IN_PROC(p, td)
674 			vm_thread_swapin(td);
675 		PROC_LOCK(p);
676 		swapclear(p);
677 		p->p_swtick = ticks;
678 
679 		wakeup(&p->p_flag);
680 
681 		/* Allow other threads to swap p out now. */
682 		--p->p_lock;
683 	}
684 #endif /* NO_SWAPPING */
685 }
686 
687 /*
688  * This swapin algorithm attempts to swap-in processes only if there
689  * is enough space for them.  Of course, if a process waits for a long
690  * time, it will be swapped in anyway.
691  *
692  * Giant is held on entry.
693  */
694 /* ARGSUSED*/
695 static void
696 scheduler(dummy)
697 	void *dummy;
698 {
699 	struct proc *p;
700 	struct thread *td;
701 	struct proc *pp;
702 	int slptime;
703 	int swtime;
704 	int ppri;
705 	int pri;
706 
707 	mtx_assert(&Giant, MA_OWNED | MA_NOTRECURSED);
708 	mtx_unlock(&Giant);
709 
710 loop:
711 	if (vm_page_count_min()) {
712 		VM_WAIT;
713 		goto loop;
714 	}
715 
716 	pp = NULL;
717 	ppri = INT_MIN;
718 	sx_slock(&allproc_lock);
719 	FOREACH_PROC_IN_SYSTEM(p) {
720 		PROC_LOCK(p);
721 		if (p->p_state == PRS_NEW ||
722 		    p->p_flag & (P_SWAPPINGOUT | P_SWAPPINGIN | P_INMEM)) {
723 			PROC_UNLOCK(p);
724 			continue;
725 		}
726 		swtime = (ticks - p->p_swtick) / hz;
727 		FOREACH_THREAD_IN_PROC(p, td) {
728 			/*
729 			 * An otherwise runnable thread of a process
730 			 * swapped out has only the TDI_SWAPPED bit set.
731 			 *
732 			 */
733 			thread_lock(td);
734 			if (td->td_inhibitors == TDI_SWAPPED) {
735 				slptime = (ticks - td->td_slptick) / hz;
736 				pri = swtime + slptime;
737 				if ((td->td_flags & TDF_SWAPINREQ) == 0)
738 					pri -= p->p_nice * 8;
739 				/*
740 				 * if this thread is higher priority
741 				 * and there is enough space, then select
742 				 * this process instead of the previous
743 				 * selection.
744 				 */
745 				if (pri > ppri) {
746 					pp = p;
747 					ppri = pri;
748 				}
749 			}
750 			thread_unlock(td);
751 		}
752 		PROC_UNLOCK(p);
753 	}
754 	sx_sunlock(&allproc_lock);
755 
756 	/*
757 	 * Nothing to do, back to sleep.
758 	 */
759 	if ((p = pp) == NULL) {
760 		tsleep(&proc0, PVM, "sched", MAXSLP * hz / 2);
761 		goto loop;
762 	}
763 	PROC_LOCK(p);
764 
765 	/*
766 	 * Another process may be bringing or may have already
767 	 * brought this process in while we traverse all threads.
768 	 * Or, this process may even be being swapped out again.
769 	 */
770 	if (p->p_flag & (P_INMEM | P_SWAPPINGOUT | P_SWAPPINGIN)) {
771 		PROC_UNLOCK(p);
772 		goto loop;
773 	}
774 
775 	/*
776 	 * We would like to bring someone in. (only if there is space).
777 	 * [What checks the space? ]
778 	 */
779 	faultin(p);
780 	PROC_UNLOCK(p);
781 	goto loop;
782 }
783 
784 void
785 kick_proc0(void)
786 {
787 
788 	wakeup(&proc0);
789 }
790 
791 #ifndef NO_SWAPPING
792 
793 /*
794  * Swap_idle_threshold1 is the guaranteed swapped in time for a process
795  */
796 static int swap_idle_threshold1 = 2;
797 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold1, CTLFLAG_RW,
798     &swap_idle_threshold1, 0, "Guaranteed swapped in time for a process");
799 
800 /*
801  * Swap_idle_threshold2 is the time that a process can be idle before
802  * it will be swapped out, if idle swapping is enabled.
803  */
804 static int swap_idle_threshold2 = 10;
805 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold2, CTLFLAG_RW,
806     &swap_idle_threshold2, 0, "Time before a process will be swapped out");
807 
808 /*
809  * First, if any processes have been sleeping or stopped for at least
810  * "swap_idle_threshold1" seconds, they are swapped out.  If, however,
811  * no such processes exist, then the longest-sleeping or stopped
812  * process is swapped out.  Finally, and only as a last resort, if
813  * there are no sleeping or stopped processes, the longest-resident
814  * process is swapped out.
815  */
816 void
817 swapout_procs(action)
818 int action;
819 {
820 	struct proc *p;
821 	struct thread *td;
822 	int didswap = 0;
823 
824 retry:
825 	sx_slock(&allproc_lock);
826 	FOREACH_PROC_IN_SYSTEM(p) {
827 		struct vmspace *vm;
828 		int minslptime = 100000;
829 		int slptime;
830 
831 		/*
832 		 * Watch out for a process in
833 		 * creation.  It may have no
834 		 * address space or lock yet.
835 		 */
836 		if (p->p_state == PRS_NEW)
837 			continue;
838 		/*
839 		 * An aio daemon switches its
840 		 * address space while running.
841 		 * Perform a quick check whether
842 		 * a process has P_SYSTEM.
843 		 */
844 		if ((p->p_flag & P_SYSTEM) != 0)
845 			continue;
846 		/*
847 		 * Do not swapout a process that
848 		 * is waiting for VM data
849 		 * structures as there is a possible
850 		 * deadlock.  Test this first as
851 		 * this may block.
852 		 *
853 		 * Lock the map until swapout
854 		 * finishes, or a thread of this
855 		 * process may attempt to alter
856 		 * the map.
857 		 */
858 		vm = vmspace_acquire_ref(p);
859 		if (vm == NULL)
860 			continue;
861 		if (!vm_map_trylock(&vm->vm_map))
862 			goto nextproc1;
863 
864 		PROC_LOCK(p);
865 		if (p->p_lock != 0 ||
866 		    (p->p_flag & (P_STOPPED_SINGLE|P_TRACED|P_SYSTEM|P_WEXIT)
867 		    ) != 0) {
868 			goto nextproc;
869 		}
870 		/*
871 		 * only aiod changes vmspace, however it will be
872 		 * skipped because of the if statement above checking
873 		 * for P_SYSTEM
874 		 */
875 		if ((p->p_flag & (P_INMEM|P_SWAPPINGOUT|P_SWAPPINGIN)) != P_INMEM)
876 			goto nextproc;
877 
878 		switch (p->p_state) {
879 		default:
880 			/* Don't swap out processes in any sort
881 			 * of 'special' state. */
882 			break;
883 
884 		case PRS_NORMAL:
885 			/*
886 			 * do not swapout a realtime process
887 			 * Check all the thread groups..
888 			 */
889 			FOREACH_THREAD_IN_PROC(p, td) {
890 				thread_lock(td);
891 				if (PRI_IS_REALTIME(td->td_pri_class)) {
892 					thread_unlock(td);
893 					goto nextproc;
894 				}
895 				slptime = (ticks - td->td_slptick) / hz;
896 				/*
897 				 * Guarantee swap_idle_threshold1
898 				 * time in memory.
899 				 */
900 				if (slptime < swap_idle_threshold1) {
901 					thread_unlock(td);
902 					goto nextproc;
903 				}
904 
905 				/*
906 				 * Do not swapout a process if it is
907 				 * waiting on a critical event of some
908 				 * kind or there is a thread whose
909 				 * pageable memory may be accessed.
910 				 *
911 				 * This could be refined to support
912 				 * swapping out a thread.
913 				 */
914 				if (!thread_safetoswapout(td)) {
915 					thread_unlock(td);
916 					goto nextproc;
917 				}
918 				/*
919 				 * If the system is under memory stress,
920 				 * or if we are swapping
921 				 * idle processes >= swap_idle_threshold2,
922 				 * then swap the process out.
923 				 */
924 				if (((action & VM_SWAP_NORMAL) == 0) &&
925 				    (((action & VM_SWAP_IDLE) == 0) ||
926 				    (slptime < swap_idle_threshold2))) {
927 					thread_unlock(td);
928 					goto nextproc;
929 				}
930 
931 				if (minslptime > slptime)
932 					minslptime = slptime;
933 				thread_unlock(td);
934 			}
935 
936 			/*
937 			 * If the pageout daemon didn't free enough pages,
938 			 * or if this process is idle and the system is
939 			 * configured to swap proactively, swap it out.
940 			 */
941 			if ((action & VM_SWAP_NORMAL) ||
942 				((action & VM_SWAP_IDLE) &&
943 				 (minslptime > swap_idle_threshold2))) {
944 				if (swapout(p) == 0)
945 					didswap++;
946 				PROC_UNLOCK(p);
947 				vm_map_unlock(&vm->vm_map);
948 				vmspace_free(vm);
949 				sx_sunlock(&allproc_lock);
950 				goto retry;
951 			}
952 		}
953 nextproc:
954 		PROC_UNLOCK(p);
955 		vm_map_unlock(&vm->vm_map);
956 nextproc1:
957 		vmspace_free(vm);
958 		continue;
959 	}
960 	sx_sunlock(&allproc_lock);
961 	/*
962 	 * If we swapped something out, and another process needed memory,
963 	 * then wakeup the sched process.
964 	 */
965 	if (didswap)
966 		wakeup(&proc0);
967 }
968 
969 static void
970 swapclear(p)
971 	struct proc *p;
972 {
973 	struct thread *td;
974 
975 	PROC_LOCK_ASSERT(p, MA_OWNED);
976 
977 	FOREACH_THREAD_IN_PROC(p, td) {
978 		thread_lock(td);
979 		td->td_flags |= TDF_INMEM;
980 		td->td_flags &= ~TDF_SWAPINREQ;
981 		TD_CLR_SWAPPED(td);
982 		if (TD_CAN_RUN(td))
983 			if (setrunnable(td)) {
984 #ifdef INVARIANTS
985 				/*
986 				 * XXX: We just cleared TDI_SWAPPED
987 				 * above and set TDF_INMEM, so this
988 				 * should never happen.
989 				 */
990 				panic("not waking up swapper");
991 #endif
992 			}
993 		thread_unlock(td);
994 	}
995 	p->p_flag &= ~(P_SWAPPINGIN|P_SWAPPINGOUT);
996 	p->p_flag |= P_INMEM;
997 }
998 
999 static int
1000 swapout(p)
1001 	struct proc *p;
1002 {
1003 	struct thread *td;
1004 
1005 	PROC_LOCK_ASSERT(p, MA_OWNED);
1006 #if defined(SWAP_DEBUG)
1007 	printf("swapping out %d\n", p->p_pid);
1008 #endif
1009 
1010 	/*
1011 	 * The states of this process and its threads may have changed
1012 	 * by now.  Assuming that there is only one pageout daemon thread,
1013 	 * this process should still be in memory.
1014 	 */
1015 	KASSERT((p->p_flag & (P_INMEM|P_SWAPPINGOUT|P_SWAPPINGIN)) == P_INMEM,
1016 		("swapout: lost a swapout race?"));
1017 
1018 	/*
1019 	 * remember the process resident count
1020 	 */
1021 	p->p_vmspace->vm_swrss = vmspace_resident_count(p->p_vmspace);
1022 	/*
1023 	 * Check and mark all threads before we proceed.
1024 	 */
1025 	p->p_flag &= ~P_INMEM;
1026 	p->p_flag |= P_SWAPPINGOUT;
1027 	FOREACH_THREAD_IN_PROC(p, td) {
1028 		thread_lock(td);
1029 		if (!thread_safetoswapout(td)) {
1030 			thread_unlock(td);
1031 			swapclear(p);
1032 			return (EBUSY);
1033 		}
1034 		td->td_flags &= ~TDF_INMEM;
1035 		TD_SET_SWAPPED(td);
1036 		thread_unlock(td);
1037 	}
1038 	td = FIRST_THREAD_IN_PROC(p);
1039 	++td->td_ru.ru_nswap;
1040 	PROC_UNLOCK(p);
1041 
1042 	/*
1043 	 * This list is stable because all threads are now prevented from
1044 	 * running.  The list is only modified in the context of a running
1045 	 * thread in this process.
1046 	 */
1047 	FOREACH_THREAD_IN_PROC(p, td)
1048 		vm_thread_swapout(td);
1049 
1050 	PROC_LOCK(p);
1051 	p->p_flag &= ~P_SWAPPINGOUT;
1052 	p->p_swtick = ticks;
1053 	return (0);
1054 }
1055 #endif /* !NO_SWAPPING */
1056