xref: /freebsd/sys/vm/vm_glue.c (revision c6ec7d31830ab1c80edae95ad5e4b9dba10c47ac)
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 MTX_SYSINIT(kstack_cache, &kstack_cache_mtx, "kstkch", MTX_DEF);
311 
312 SYSCTL_INT(_vm, OID_AUTO, kstack_cache_size, CTLFLAG_RW, &kstack_cache_size, 0,
313     "");
314 SYSCTL_INT(_vm, OID_AUTO, kstacks, CTLFLAG_RD, &kstacks, 0,
315     "");
316 
317 #ifndef KSTACK_MAX_PAGES
318 #define KSTACK_MAX_PAGES 32
319 #endif
320 
321 /*
322  * Create the kernel stack (including pcb for i386) for a new thread.
323  * This routine directly affects the fork perf for a process and
324  * create performance for a thread.
325  */
326 int
327 vm_thread_new(struct thread *td, int pages)
328 {
329 	vm_object_t ksobj;
330 	vm_offset_t ks;
331 	vm_page_t m, ma[KSTACK_MAX_PAGES];
332 	struct kstack_cache_entry *ks_ce;
333 	int i;
334 
335 	/* Bounds check */
336 	if (pages <= 1)
337 		pages = KSTACK_PAGES;
338 	else if (pages > KSTACK_MAX_PAGES)
339 		pages = KSTACK_MAX_PAGES;
340 
341 	if (pages == KSTACK_PAGES) {
342 		mtx_lock(&kstack_cache_mtx);
343 		if (kstack_cache != NULL) {
344 			ks_ce = kstack_cache;
345 			kstack_cache = ks_ce->next_ks_entry;
346 			mtx_unlock(&kstack_cache_mtx);
347 
348 			td->td_kstack_obj = ks_ce->ksobj;
349 			td->td_kstack = (vm_offset_t)ks_ce;
350 			td->td_kstack_pages = KSTACK_PAGES;
351 			return (1);
352 		}
353 		mtx_unlock(&kstack_cache_mtx);
354 	}
355 
356 	/*
357 	 * Allocate an object for the kstack.
358 	 */
359 	ksobj = vm_object_allocate(OBJT_DEFAULT, pages);
360 
361 	/*
362 	 * Get a kernel virtual address for this thread's kstack.
363 	 */
364 #if defined(__mips__)
365 	/*
366 	 * We need to align the kstack's mapped address to fit within
367 	 * a single TLB entry.
368 	 */
369 	ks = kmem_alloc_nofault_space(kernel_map,
370 	    (pages + KSTACK_GUARD_PAGES) * PAGE_SIZE, VMFS_TLB_ALIGNED_SPACE);
371 #else
372 	ks = kmem_alloc_nofault(kernel_map,
373 	   (pages + KSTACK_GUARD_PAGES) * PAGE_SIZE);
374 #endif
375 	if (ks == 0) {
376 		printf("vm_thread_new: kstack allocation failed\n");
377 		vm_object_deallocate(ksobj);
378 		return (0);
379 	}
380 
381 	atomic_add_int(&kstacks, 1);
382 	if (KSTACK_GUARD_PAGES != 0) {
383 		pmap_qremove(ks, KSTACK_GUARD_PAGES);
384 		ks += KSTACK_GUARD_PAGES * PAGE_SIZE;
385 	}
386 	td->td_kstack_obj = ksobj;
387 	td->td_kstack = ks;
388 	/*
389 	 * Knowing the number of pages allocated is useful when you
390 	 * want to deallocate them.
391 	 */
392 	td->td_kstack_pages = pages;
393 	/*
394 	 * For the length of the stack, link in a real page of ram for each
395 	 * page of stack.
396 	 */
397 	VM_OBJECT_LOCK(ksobj);
398 	for (i = 0; i < pages; i++) {
399 		/*
400 		 * Get a kernel stack page.
401 		 */
402 		m = vm_page_grab(ksobj, i, VM_ALLOC_NOBUSY |
403 		    VM_ALLOC_NORMAL | VM_ALLOC_RETRY | VM_ALLOC_WIRED);
404 		ma[i] = m;
405 		m->valid = VM_PAGE_BITS_ALL;
406 	}
407 	VM_OBJECT_UNLOCK(ksobj);
408 	pmap_qenter(ks, ma, pages);
409 	return (1);
410 }
411 
412 static void
413 vm_thread_stack_dispose(vm_object_t ksobj, vm_offset_t ks, int pages)
414 {
415 	vm_page_t m;
416 	int i;
417 
418 	atomic_add_int(&kstacks, -1);
419 	pmap_qremove(ks, pages);
420 	VM_OBJECT_LOCK(ksobj);
421 	for (i = 0; i < pages; i++) {
422 		m = vm_page_lookup(ksobj, i);
423 		if (m == NULL)
424 			panic("vm_thread_dispose: kstack already missing?");
425 		vm_page_lock(m);
426 		vm_page_unwire(m, 0);
427 		vm_page_free(m);
428 		vm_page_unlock(m);
429 	}
430 	VM_OBJECT_UNLOCK(ksobj);
431 	vm_object_deallocate(ksobj);
432 	kmem_free(kernel_map, ks - (KSTACK_GUARD_PAGES * PAGE_SIZE),
433 	    (pages + KSTACK_GUARD_PAGES) * PAGE_SIZE);
434 }
435 
436 /*
437  * Dispose of a thread's kernel stack.
438  */
439 void
440 vm_thread_dispose(struct thread *td)
441 {
442 	vm_object_t ksobj;
443 	vm_offset_t ks;
444 	struct kstack_cache_entry *ks_ce;
445 	int pages;
446 
447 	pages = td->td_kstack_pages;
448 	ksobj = td->td_kstack_obj;
449 	ks = td->td_kstack;
450 	td->td_kstack = 0;
451 	td->td_kstack_pages = 0;
452 	if (pages == KSTACK_PAGES && kstacks <= kstack_cache_size) {
453 		ks_ce = (struct kstack_cache_entry *)ks;
454 		ks_ce->ksobj = ksobj;
455 		mtx_lock(&kstack_cache_mtx);
456 		ks_ce->next_ks_entry = kstack_cache;
457 		kstack_cache = ks_ce;
458 		mtx_unlock(&kstack_cache_mtx);
459 		return;
460 	}
461 	vm_thread_stack_dispose(ksobj, ks, pages);
462 }
463 
464 static void
465 vm_thread_stack_lowmem(void *nulll)
466 {
467 	struct kstack_cache_entry *ks_ce, *ks_ce1;
468 
469 	mtx_lock(&kstack_cache_mtx);
470 	ks_ce = kstack_cache;
471 	kstack_cache = NULL;
472 	mtx_unlock(&kstack_cache_mtx);
473 
474 	while (ks_ce != NULL) {
475 		ks_ce1 = ks_ce;
476 		ks_ce = ks_ce->next_ks_entry;
477 
478 		vm_thread_stack_dispose(ks_ce1->ksobj, (vm_offset_t)ks_ce1,
479 		    KSTACK_PAGES);
480 	}
481 }
482 
483 static void
484 kstack_cache_init(void *nulll)
485 {
486 
487 	EVENTHANDLER_REGISTER(vm_lowmem, vm_thread_stack_lowmem, NULL,
488 	    EVENTHANDLER_PRI_ANY);
489 }
490 
491 SYSINIT(vm_kstacks, SI_SUB_KTHREAD_INIT, SI_ORDER_ANY, kstack_cache_init, NULL);
492 
493 #ifndef NO_SWAPPING
494 /*
495  * Allow a thread's kernel stack to be paged out.
496  */
497 static void
498 vm_thread_swapout(struct thread *td)
499 {
500 	vm_object_t ksobj;
501 	vm_page_t m;
502 	int i, pages;
503 
504 	cpu_thread_swapout(td);
505 	pages = td->td_kstack_pages;
506 	ksobj = td->td_kstack_obj;
507 	pmap_qremove(td->td_kstack, pages);
508 	VM_OBJECT_LOCK(ksobj);
509 	for (i = 0; i < pages; i++) {
510 		m = vm_page_lookup(ksobj, i);
511 		if (m == NULL)
512 			panic("vm_thread_swapout: kstack already missing?");
513 		vm_page_dirty(m);
514 		vm_page_lock(m);
515 		vm_page_unwire(m, 0);
516 		vm_page_unlock(m);
517 	}
518 	VM_OBJECT_UNLOCK(ksobj);
519 }
520 
521 /*
522  * Bring the kernel stack for a specified thread back in.
523  */
524 static void
525 vm_thread_swapin(struct thread *td)
526 {
527 	vm_object_t ksobj;
528 	vm_page_t ma[KSTACK_MAX_PAGES];
529 	int i, j, k, pages, rv;
530 
531 	pages = td->td_kstack_pages;
532 	ksobj = td->td_kstack_obj;
533 	VM_OBJECT_LOCK(ksobj);
534 	for (i = 0; i < pages; i++)
535 		ma[i] = vm_page_grab(ksobj, i, VM_ALLOC_NORMAL | VM_ALLOC_RETRY |
536 		    VM_ALLOC_WIRED);
537 	for (i = 0; i < pages; i++) {
538 		if (ma[i]->valid != VM_PAGE_BITS_ALL) {
539 			KASSERT(ma[i]->oflags & VPO_BUSY,
540 			    ("lost busy 1"));
541 			vm_object_pip_add(ksobj, 1);
542 			for (j = i + 1; j < pages; j++) {
543 				KASSERT(ma[j]->valid == VM_PAGE_BITS_ALL ||
544 				    (ma[j]->oflags & VPO_BUSY),
545 				    ("lost busy 2"));
546 				if (ma[j]->valid == VM_PAGE_BITS_ALL)
547 					break;
548 			}
549 			rv = vm_pager_get_pages(ksobj, ma + i, j - i, 0);
550 			if (rv != VM_PAGER_OK)
551 	panic("vm_thread_swapin: cannot get kstack for proc: %d",
552 				    td->td_proc->p_pid);
553 			vm_object_pip_wakeup(ksobj);
554 			for (k = i; k < j; k++)
555 				ma[k] = vm_page_lookup(ksobj, k);
556 			vm_page_wakeup(ma[i]);
557 		} else if (ma[i]->oflags & VPO_BUSY)
558 			vm_page_wakeup(ma[i]);
559 	}
560 	VM_OBJECT_UNLOCK(ksobj);
561 	pmap_qenter(td->td_kstack, ma, pages);
562 	cpu_thread_swapin(td);
563 }
564 #endif /* !NO_SWAPPING */
565 
566 /*
567  * Implement fork's actions on an address space.
568  * Here we arrange for the address space to be copied or referenced,
569  * allocate a user struct (pcb and kernel stack), then call the
570  * machine-dependent layer to fill those in and make the new process
571  * ready to run.  The new process is set up so that it returns directly
572  * to user mode to avoid stack copying and relocation problems.
573  */
574 int
575 vm_forkproc(td, p2, td2, vm2, flags)
576 	struct thread *td;
577 	struct proc *p2;
578 	struct thread *td2;
579 	struct vmspace *vm2;
580 	int flags;
581 {
582 	struct proc *p1 = td->td_proc;
583 	int error;
584 
585 	if ((flags & RFPROC) == 0) {
586 		/*
587 		 * Divorce the memory, if it is shared, essentially
588 		 * this changes shared memory amongst threads, into
589 		 * COW locally.
590 		 */
591 		if ((flags & RFMEM) == 0) {
592 			if (p1->p_vmspace->vm_refcnt > 1) {
593 				error = vmspace_unshare(p1);
594 				if (error)
595 					return (error);
596 			}
597 		}
598 		cpu_fork(td, p2, td2, flags);
599 		return (0);
600 	}
601 
602 	if (flags & RFMEM) {
603 		p2->p_vmspace = p1->p_vmspace;
604 		atomic_add_int(&p1->p_vmspace->vm_refcnt, 1);
605 	}
606 
607 	while (vm_page_count_severe()) {
608 		VM_WAIT;
609 	}
610 
611 	if ((flags & RFMEM) == 0) {
612 		p2->p_vmspace = vm2;
613 		if (p1->p_vmspace->vm_shm)
614 			shmfork(p1, p2);
615 	}
616 
617 	/*
618 	 * cpu_fork will copy and update the pcb, set up the kernel stack,
619 	 * and make the child ready to run.
620 	 */
621 	cpu_fork(td, p2, td2, flags);
622 	return (0);
623 }
624 
625 /*
626  * Called after process has been wait(2)'ed apon and is being reaped.
627  * The idea is to reclaim resources that we could not reclaim while
628  * the process was still executing.
629  */
630 void
631 vm_waitproc(p)
632 	struct proc *p;
633 {
634 
635 	vmspace_exitfree(p);		/* and clean-out the vmspace */
636 }
637 
638 void
639 faultin(p)
640 	struct proc *p;
641 {
642 #ifdef NO_SWAPPING
643 
644 	PROC_LOCK_ASSERT(p, MA_OWNED);
645 	if ((p->p_flag & P_INMEM) == 0)
646 		panic("faultin: proc swapped out with NO_SWAPPING!");
647 #else /* !NO_SWAPPING */
648 	struct thread *td;
649 
650 	PROC_LOCK_ASSERT(p, MA_OWNED);
651 	/*
652 	 * If another process is swapping in this process,
653 	 * just wait until it finishes.
654 	 */
655 	if (p->p_flag & P_SWAPPINGIN) {
656 		while (p->p_flag & P_SWAPPINGIN)
657 			msleep(&p->p_flag, &p->p_mtx, PVM, "faultin", 0);
658 		return;
659 	}
660 	if ((p->p_flag & P_INMEM) == 0) {
661 		/*
662 		 * Don't let another thread swap process p out while we are
663 		 * busy swapping it in.
664 		 */
665 		++p->p_lock;
666 		p->p_flag |= P_SWAPPINGIN;
667 		PROC_UNLOCK(p);
668 
669 		/*
670 		 * We hold no lock here because the list of threads
671 		 * can not change while all threads in the process are
672 		 * swapped out.
673 		 */
674 		FOREACH_THREAD_IN_PROC(p, td)
675 			vm_thread_swapin(td);
676 		PROC_LOCK(p);
677 		swapclear(p);
678 		p->p_swtick = ticks;
679 
680 		wakeup(&p->p_flag);
681 
682 		/* Allow other threads to swap p out now. */
683 		--p->p_lock;
684 	}
685 #endif /* NO_SWAPPING */
686 }
687 
688 /*
689  * This swapin algorithm attempts to swap-in processes only if there
690  * is enough space for them.  Of course, if a process waits for a long
691  * time, it will be swapped in anyway.
692  *
693  * Giant is held on entry.
694  */
695 /* ARGSUSED*/
696 static void
697 scheduler(dummy)
698 	void *dummy;
699 {
700 	struct proc *p;
701 	struct thread *td;
702 	struct proc *pp;
703 	int slptime;
704 	int swtime;
705 	int ppri;
706 	int pri;
707 
708 	mtx_assert(&Giant, MA_OWNED | MA_NOTRECURSED);
709 	mtx_unlock(&Giant);
710 
711 loop:
712 	if (vm_page_count_min()) {
713 		VM_WAIT;
714 		goto loop;
715 	}
716 
717 	pp = NULL;
718 	ppri = INT_MIN;
719 	sx_slock(&allproc_lock);
720 	FOREACH_PROC_IN_SYSTEM(p) {
721 		PROC_LOCK(p);
722 		if (p->p_state == PRS_NEW ||
723 		    p->p_flag & (P_SWAPPINGOUT | P_SWAPPINGIN | P_INMEM)) {
724 			PROC_UNLOCK(p);
725 			continue;
726 		}
727 		swtime = (ticks - p->p_swtick) / hz;
728 		FOREACH_THREAD_IN_PROC(p, td) {
729 			/*
730 			 * An otherwise runnable thread of a process
731 			 * swapped out has only the TDI_SWAPPED bit set.
732 			 *
733 			 */
734 			thread_lock(td);
735 			if (td->td_inhibitors == TDI_SWAPPED) {
736 				slptime = (ticks - td->td_slptick) / hz;
737 				pri = swtime + slptime;
738 				if ((td->td_flags & TDF_SWAPINREQ) == 0)
739 					pri -= p->p_nice * 8;
740 				/*
741 				 * if this thread is higher priority
742 				 * and there is enough space, then select
743 				 * this process instead of the previous
744 				 * selection.
745 				 */
746 				if (pri > ppri) {
747 					pp = p;
748 					ppri = pri;
749 				}
750 			}
751 			thread_unlock(td);
752 		}
753 		PROC_UNLOCK(p);
754 	}
755 	sx_sunlock(&allproc_lock);
756 
757 	/*
758 	 * Nothing to do, back to sleep.
759 	 */
760 	if ((p = pp) == NULL) {
761 		tsleep(&proc0, PVM, "sched", MAXSLP * hz / 2);
762 		goto loop;
763 	}
764 	PROC_LOCK(p);
765 
766 	/*
767 	 * Another process may be bringing or may have already
768 	 * brought this process in while we traverse all threads.
769 	 * Or, this process may even be being swapped out again.
770 	 */
771 	if (p->p_flag & (P_INMEM | P_SWAPPINGOUT | P_SWAPPINGIN)) {
772 		PROC_UNLOCK(p);
773 		goto loop;
774 	}
775 
776 	/*
777 	 * We would like to bring someone in. (only if there is space).
778 	 * [What checks the space? ]
779 	 */
780 	faultin(p);
781 	PROC_UNLOCK(p);
782 	goto loop;
783 }
784 
785 void
786 kick_proc0(void)
787 {
788 
789 	wakeup(&proc0);
790 }
791 
792 #ifndef NO_SWAPPING
793 
794 /*
795  * Swap_idle_threshold1 is the guaranteed swapped in time for a process
796  */
797 static int swap_idle_threshold1 = 2;
798 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold1, CTLFLAG_RW,
799     &swap_idle_threshold1, 0, "Guaranteed swapped in time for a process");
800 
801 /*
802  * Swap_idle_threshold2 is the time that a process can be idle before
803  * it will be swapped out, if idle swapping is enabled.
804  */
805 static int swap_idle_threshold2 = 10;
806 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold2, CTLFLAG_RW,
807     &swap_idle_threshold2, 0, "Time before a process will be swapped out");
808 
809 /*
810  * First, if any processes have been sleeping or stopped for at least
811  * "swap_idle_threshold1" seconds, they are swapped out.  If, however,
812  * no such processes exist, then the longest-sleeping or stopped
813  * process is swapped out.  Finally, and only as a last resort, if
814  * there are no sleeping or stopped processes, the longest-resident
815  * process is swapped out.
816  */
817 void
818 swapout_procs(action)
819 int action;
820 {
821 	struct proc *p;
822 	struct thread *td;
823 	int didswap = 0;
824 
825 retry:
826 	sx_slock(&allproc_lock);
827 	FOREACH_PROC_IN_SYSTEM(p) {
828 		struct vmspace *vm;
829 		int minslptime = 100000;
830 		int slptime;
831 
832 		/*
833 		 * Watch out for a process in
834 		 * creation.  It may have no
835 		 * address space or lock yet.
836 		 */
837 		if (p->p_state == PRS_NEW)
838 			continue;
839 		/*
840 		 * An aio daemon switches its
841 		 * address space while running.
842 		 * Perform a quick check whether
843 		 * a process has P_SYSTEM.
844 		 */
845 		if ((p->p_flag & P_SYSTEM) != 0)
846 			continue;
847 		/*
848 		 * Do not swapout a process that
849 		 * is waiting for VM data
850 		 * structures as there is a possible
851 		 * deadlock.  Test this first as
852 		 * this may block.
853 		 *
854 		 * Lock the map until swapout
855 		 * finishes, or a thread of this
856 		 * process may attempt to alter
857 		 * the map.
858 		 */
859 		vm = vmspace_acquire_ref(p);
860 		if (vm == NULL)
861 			continue;
862 		if (!vm_map_trylock(&vm->vm_map))
863 			goto nextproc1;
864 
865 		PROC_LOCK(p);
866 		if (p->p_lock != 0 ||
867 		    (p->p_flag & (P_STOPPED_SINGLE|P_TRACED|P_SYSTEM|P_WEXIT)
868 		    ) != 0) {
869 			goto nextproc;
870 		}
871 		/*
872 		 * only aiod changes vmspace, however it will be
873 		 * skipped because of the if statement above checking
874 		 * for P_SYSTEM
875 		 */
876 		if ((p->p_flag & (P_INMEM|P_SWAPPINGOUT|P_SWAPPINGIN)) != P_INMEM)
877 			goto nextproc;
878 
879 		switch (p->p_state) {
880 		default:
881 			/* Don't swap out processes in any sort
882 			 * of 'special' state. */
883 			break;
884 
885 		case PRS_NORMAL:
886 			/*
887 			 * do not swapout a realtime process
888 			 * Check all the thread groups..
889 			 */
890 			FOREACH_THREAD_IN_PROC(p, td) {
891 				thread_lock(td);
892 				if (PRI_IS_REALTIME(td->td_pri_class)) {
893 					thread_unlock(td);
894 					goto nextproc;
895 				}
896 				slptime = (ticks - td->td_slptick) / hz;
897 				/*
898 				 * Guarantee swap_idle_threshold1
899 				 * time in memory.
900 				 */
901 				if (slptime < swap_idle_threshold1) {
902 					thread_unlock(td);
903 					goto nextproc;
904 				}
905 
906 				/*
907 				 * Do not swapout a process if it is
908 				 * waiting on a critical event of some
909 				 * kind or there is a thread whose
910 				 * pageable memory may be accessed.
911 				 *
912 				 * This could be refined to support
913 				 * swapping out a thread.
914 				 */
915 				if (!thread_safetoswapout(td)) {
916 					thread_unlock(td);
917 					goto nextproc;
918 				}
919 				/*
920 				 * If the system is under memory stress,
921 				 * or if we are swapping
922 				 * idle processes >= swap_idle_threshold2,
923 				 * then swap the process out.
924 				 */
925 				if (((action & VM_SWAP_NORMAL) == 0) &&
926 				    (((action & VM_SWAP_IDLE) == 0) ||
927 				    (slptime < swap_idle_threshold2))) {
928 					thread_unlock(td);
929 					goto nextproc;
930 				}
931 
932 				if (minslptime > slptime)
933 					minslptime = slptime;
934 				thread_unlock(td);
935 			}
936 
937 			/*
938 			 * If the pageout daemon didn't free enough pages,
939 			 * or if this process is idle and the system is
940 			 * configured to swap proactively, swap it out.
941 			 */
942 			if ((action & VM_SWAP_NORMAL) ||
943 				((action & VM_SWAP_IDLE) &&
944 				 (minslptime > swap_idle_threshold2))) {
945 				if (swapout(p) == 0)
946 					didswap++;
947 				PROC_UNLOCK(p);
948 				vm_map_unlock(&vm->vm_map);
949 				vmspace_free(vm);
950 				sx_sunlock(&allproc_lock);
951 				goto retry;
952 			}
953 		}
954 nextproc:
955 		PROC_UNLOCK(p);
956 		vm_map_unlock(&vm->vm_map);
957 nextproc1:
958 		vmspace_free(vm);
959 		continue;
960 	}
961 	sx_sunlock(&allproc_lock);
962 	/*
963 	 * If we swapped something out, and another process needed memory,
964 	 * then wakeup the sched process.
965 	 */
966 	if (didswap)
967 		wakeup(&proc0);
968 }
969 
970 static void
971 swapclear(p)
972 	struct proc *p;
973 {
974 	struct thread *td;
975 
976 	PROC_LOCK_ASSERT(p, MA_OWNED);
977 
978 	FOREACH_THREAD_IN_PROC(p, td) {
979 		thread_lock(td);
980 		td->td_flags |= TDF_INMEM;
981 		td->td_flags &= ~TDF_SWAPINREQ;
982 		TD_CLR_SWAPPED(td);
983 		if (TD_CAN_RUN(td))
984 			if (setrunnable(td)) {
985 #ifdef INVARIANTS
986 				/*
987 				 * XXX: We just cleared TDI_SWAPPED
988 				 * above and set TDF_INMEM, so this
989 				 * should never happen.
990 				 */
991 				panic("not waking up swapper");
992 #endif
993 			}
994 		thread_unlock(td);
995 	}
996 	p->p_flag &= ~(P_SWAPPINGIN|P_SWAPPINGOUT);
997 	p->p_flag |= P_INMEM;
998 }
999 
1000 static int
1001 swapout(p)
1002 	struct proc *p;
1003 {
1004 	struct thread *td;
1005 
1006 	PROC_LOCK_ASSERT(p, MA_OWNED);
1007 #if defined(SWAP_DEBUG)
1008 	printf("swapping out %d\n", p->p_pid);
1009 #endif
1010 
1011 	/*
1012 	 * The states of this process and its threads may have changed
1013 	 * by now.  Assuming that there is only one pageout daemon thread,
1014 	 * this process should still be in memory.
1015 	 */
1016 	KASSERT((p->p_flag & (P_INMEM|P_SWAPPINGOUT|P_SWAPPINGIN)) == P_INMEM,
1017 		("swapout: lost a swapout race?"));
1018 
1019 	/*
1020 	 * remember the process resident count
1021 	 */
1022 	p->p_vmspace->vm_swrss = vmspace_resident_count(p->p_vmspace);
1023 	/*
1024 	 * Check and mark all threads before we proceed.
1025 	 */
1026 	p->p_flag &= ~P_INMEM;
1027 	p->p_flag |= P_SWAPPINGOUT;
1028 	FOREACH_THREAD_IN_PROC(p, td) {
1029 		thread_lock(td);
1030 		if (!thread_safetoswapout(td)) {
1031 			thread_unlock(td);
1032 			swapclear(p);
1033 			return (EBUSY);
1034 		}
1035 		td->td_flags &= ~TDF_INMEM;
1036 		TD_SET_SWAPPED(td);
1037 		thread_unlock(td);
1038 	}
1039 	td = FIRST_THREAD_IN_PROC(p);
1040 	++td->td_ru.ru_nswap;
1041 	PROC_UNLOCK(p);
1042 
1043 	/*
1044 	 * This list is stable because all threads are now prevented from
1045 	 * running.  The list is only modified in the context of a running
1046 	 * thread in this process.
1047 	 */
1048 	FOREACH_THREAD_IN_PROC(p, td)
1049 		vm_thread_swapout(td);
1050 
1051 	PROC_LOCK(p);
1052 	p->p_flag &= ~P_SWAPPINGOUT;
1053 	p->p_swtick = ticks;
1054 	return (0);
1055 }
1056 #endif /* !NO_SWAPPING */
1057