xref: /freebsd/sys/vm/vm_glue.c (revision 9517e866259191fcd39434a97ad849a9b59b9b9f)
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/kernel.h>
81 #include <sys/ktr.h>
82 #include <sys/unistd.h>
83 
84 #include <vm/vm.h>
85 #include <vm/vm_param.h>
86 #include <vm/pmap.h>
87 #include <vm/vm_map.h>
88 #include <vm/vm_page.h>
89 #include <vm/vm_pageout.h>
90 #include <vm/vm_object.h>
91 #include <vm/vm_kern.h>
92 #include <vm/vm_extern.h>
93 #include <vm/vm_pager.h>
94 #include <vm/swap_pager.h>
95 
96 extern int maxslp;
97 
98 /*
99  * System initialization
100  *
101  * Note: proc0 from proc.h
102  */
103 static void vm_init_limits(void *);
104 SYSINIT(vm_limits, SI_SUB_VM_CONF, SI_ORDER_FIRST, vm_init_limits, &proc0);
105 
106 /*
107  * THIS MUST BE THE LAST INITIALIZATION ITEM!!!
108  *
109  * Note: run scheduling should be divorced from the vm system.
110  */
111 static void scheduler(void *);
112 SYSINIT(scheduler, SI_SUB_RUN_SCHEDULER, SI_ORDER_ANY, scheduler, NULL);
113 
114 #ifndef NO_SWAPPING
115 static int swapout(struct proc *);
116 static void swapclear(struct proc *);
117 #endif
118 
119 /*
120  * MPSAFE
121  *
122  * WARNING!  This code calls vm_map_check_protection() which only checks
123  * the associated vm_map_entry range.  It does not determine whether the
124  * contents of the memory is actually readable or writable.  In most cases
125  * just checking the vm_map_entry is sufficient within the kernel's address
126  * space.
127  */
128 int
129 kernacc(addr, len, rw)
130 	void *addr;
131 	int len, rw;
132 {
133 	boolean_t rv;
134 	vm_offset_t saddr, eaddr;
135 	vm_prot_t prot;
136 
137 	KASSERT((rw & ~VM_PROT_ALL) == 0,
138 	    ("illegal ``rw'' argument to kernacc (%x)\n", rw));
139 
140 	if ((vm_offset_t)addr + len > kernel_map->max_offset ||
141 	    (vm_offset_t)addr + len < (vm_offset_t)addr)
142 		return (FALSE);
143 
144 	prot = rw;
145 	saddr = trunc_page((vm_offset_t)addr);
146 	eaddr = round_page((vm_offset_t)addr + len);
147 	vm_map_lock_read(kernel_map);
148 	rv = vm_map_check_protection(kernel_map, saddr, eaddr, prot);
149 	vm_map_unlock_read(kernel_map);
150 	return (rv == TRUE);
151 }
152 
153 /*
154  * MPSAFE
155  *
156  * WARNING!  This code calls vm_map_check_protection() which only checks
157  * the associated vm_map_entry range.  It does not determine whether the
158  * contents of the memory is actually readable or writable.  vmapbuf(),
159  * vm_fault_quick(), or copyin()/copout()/su*()/fu*() functions should be
160  * used in conjuction with this call.
161  */
162 int
163 useracc(addr, len, rw)
164 	void *addr;
165 	int len, rw;
166 {
167 	boolean_t rv;
168 	vm_prot_t prot;
169 	vm_map_t map;
170 
171 	KASSERT((rw & ~VM_PROT_ALL) == 0,
172 	    ("illegal ``rw'' argument to useracc (%x)\n", rw));
173 	prot = rw;
174 	map = &curproc->p_vmspace->vm_map;
175 	if ((vm_offset_t)addr + len > vm_map_max(map) ||
176 	    (vm_offset_t)addr + len < (vm_offset_t)addr) {
177 		return (FALSE);
178 	}
179 	vm_map_lock_read(map);
180 	rv = vm_map_check_protection(map, trunc_page((vm_offset_t)addr),
181 	    round_page((vm_offset_t)addr + len), prot);
182 	vm_map_unlock_read(map);
183 	return (rv == TRUE);
184 }
185 
186 int
187 vslock(void *addr, size_t len)
188 {
189 	vm_offset_t end, last, start;
190 	vm_size_t npages;
191 	int error;
192 
193 	last = (vm_offset_t)addr + len;
194 	start = trunc_page((vm_offset_t)addr);
195 	end = round_page(last);
196 	if (last < (vm_offset_t)addr || end < (vm_offset_t)addr)
197 		return (EINVAL);
198 	npages = atop(end - start);
199 	if (npages > vm_page_max_wired)
200 		return (ENOMEM);
201 	PROC_LOCK(curproc);
202 	if (ptoa(npages +
203 	    pmap_wired_count(vm_map_pmap(&curproc->p_vmspace->vm_map))) >
204 	    lim_cur(curproc, RLIMIT_MEMLOCK)) {
205 		PROC_UNLOCK(curproc);
206 		return (ENOMEM);
207 	}
208 	PROC_UNLOCK(curproc);
209 #if 0
210 	/*
211 	 * XXX - not yet
212 	 *
213 	 * The limit for transient usage of wired pages should be
214 	 * larger than for "permanent" wired pages (mlock()).
215 	 *
216 	 * Also, the sysctl code, which is the only present user
217 	 * of vslock(), does a hard loop on EAGAIN.
218 	 */
219 	if (npages + cnt.v_wire_count > vm_page_max_wired)
220 		return (EAGAIN);
221 #endif
222 	error = vm_map_wire(&curproc->p_vmspace->vm_map, start, end,
223 	    VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES);
224 	/*
225 	 * Return EFAULT on error to match copy{in,out}() behaviour
226 	 * rather than returning ENOMEM like mlock() would.
227 	 */
228 	return (error == KERN_SUCCESS ? 0 : EFAULT);
229 }
230 
231 void
232 vsunlock(void *addr, size_t len)
233 {
234 
235 	/* Rely on the parameter sanity checks performed by vslock(). */
236 	(void)vm_map_unwire(&curproc->p_vmspace->vm_map,
237 	    trunc_page((vm_offset_t)addr), round_page((vm_offset_t)addr + len),
238 	    VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES);
239 }
240 
241 /*
242  * Pin the page contained within the given object at the given offset.  If the
243  * page is not resident, allocate and load it using the given object's pager.
244  * Return the pinned page if successful; otherwise, return NULL.
245  */
246 static vm_page_t
247 vm_imgact_hold_page(vm_object_t object, vm_ooffset_t offset)
248 {
249 	vm_page_t m, ma[1];
250 	vm_pindex_t pindex;
251 	int rv;
252 
253 	VM_OBJECT_LOCK(object);
254 	pindex = OFF_TO_IDX(offset);
255 	m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL | VM_ALLOC_RETRY);
256 	if (m->valid != VM_PAGE_BITS_ALL) {
257 		ma[0] = m;
258 		rv = vm_pager_get_pages(object, ma, 1, 0);
259 		m = vm_page_lookup(object, pindex);
260 		if (m == NULL)
261 			goto out;
262 		if (rv != VM_PAGER_OK) {
263 			vm_page_lock_queues();
264 			vm_page_free(m);
265 			vm_page_unlock_queues();
266 			m = NULL;
267 			goto out;
268 		}
269 	}
270 	vm_page_lock_queues();
271 	vm_page_hold(m);
272 	vm_page_unlock_queues();
273 	vm_page_wakeup(m);
274 out:
275 	VM_OBJECT_UNLOCK(object);
276 	return (m);
277 }
278 
279 /*
280  * Return a CPU private mapping to the page at the given offset within the
281  * given object.  The page is pinned before it is mapped.
282  */
283 struct sf_buf *
284 vm_imgact_map_page(vm_object_t object, vm_ooffset_t offset)
285 {
286 	vm_page_t m;
287 
288 	m = vm_imgact_hold_page(object, offset);
289 	if (m == NULL)
290 		return (NULL);
291 	sched_pin();
292 	return (sf_buf_alloc(m, SFB_CPUPRIVATE));
293 }
294 
295 /*
296  * Destroy the given CPU private mapping and unpin the page that it mapped.
297  */
298 void
299 vm_imgact_unmap_page(struct sf_buf *sf)
300 {
301 	vm_page_t m;
302 
303 	m = sf_buf_page(sf);
304 	sf_buf_free(sf);
305 	sched_unpin();
306 	vm_page_lock_queues();
307 	vm_page_unhold(m);
308 	vm_page_unlock_queues();
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 	int i;
327 
328 	/* Bounds check */
329 	if (pages <= 1)
330 		pages = KSTACK_PAGES;
331 	else if (pages > KSTACK_MAX_PAGES)
332 		pages = KSTACK_MAX_PAGES;
333 	/*
334 	 * Allocate an object for the kstack.
335 	 */
336 	ksobj = vm_object_allocate(OBJT_DEFAULT, pages);
337 
338 	/*
339 	 * Get a kernel virtual address for this thread's kstack.
340 	 */
341 	ks = kmem_alloc_nofault(kernel_map,
342 	   (pages + KSTACK_GUARD_PAGES) * PAGE_SIZE);
343 	if (ks == 0) {
344 		printf("vm_thread_new: kstack allocation failed\n");
345 		vm_object_deallocate(ksobj);
346 		return (0);
347 	}
348 
349 	if (KSTACK_GUARD_PAGES != 0) {
350 		pmap_qremove(ks, KSTACK_GUARD_PAGES);
351 		ks += KSTACK_GUARD_PAGES * PAGE_SIZE;
352 	}
353 	td->td_kstack_obj = ksobj;
354 	td->td_kstack = ks;
355 	/*
356 	 * Knowing the number of pages allocated is useful when you
357 	 * want to deallocate them.
358 	 */
359 	td->td_kstack_pages = pages;
360 	/*
361 	 * For the length of the stack, link in a real page of ram for each
362 	 * page of stack.
363 	 */
364 	VM_OBJECT_LOCK(ksobj);
365 	for (i = 0; i < pages; i++) {
366 		/*
367 		 * Get a kernel stack page.
368 		 */
369 		m = vm_page_grab(ksobj, i, VM_ALLOC_NOBUSY |
370 		    VM_ALLOC_NORMAL | VM_ALLOC_RETRY | VM_ALLOC_WIRED);
371 		ma[i] = m;
372 		m->valid = VM_PAGE_BITS_ALL;
373 	}
374 	VM_OBJECT_UNLOCK(ksobj);
375 	pmap_qenter(ks, ma, pages);
376 	return (1);
377 }
378 
379 /*
380  * Dispose of a thread's kernel stack.
381  */
382 void
383 vm_thread_dispose(struct thread *td)
384 {
385 	vm_object_t ksobj;
386 	vm_offset_t ks;
387 	vm_page_t m;
388 	int i, pages;
389 
390 	pages = td->td_kstack_pages;
391 	ksobj = td->td_kstack_obj;
392 	ks = td->td_kstack;
393 	pmap_qremove(ks, pages);
394 	VM_OBJECT_LOCK(ksobj);
395 	for (i = 0; i < pages; i++) {
396 		m = vm_page_lookup(ksobj, i);
397 		if (m == NULL)
398 			panic("vm_thread_dispose: kstack already missing?");
399 		vm_page_lock_queues();
400 		vm_page_unwire(m, 0);
401 		vm_page_free(m);
402 		vm_page_unlock_queues();
403 	}
404 	VM_OBJECT_UNLOCK(ksobj);
405 	vm_object_deallocate(ksobj);
406 	kmem_free(kernel_map, ks - (KSTACK_GUARD_PAGES * PAGE_SIZE),
407 	    (pages + KSTACK_GUARD_PAGES) * PAGE_SIZE);
408 	td->td_kstack = 0;
409 }
410 
411 /*
412  * Allow a thread's kernel stack to be paged out.
413  */
414 void
415 vm_thread_swapout(struct thread *td)
416 {
417 	vm_object_t ksobj;
418 	vm_page_t m;
419 	int i, pages;
420 
421 	cpu_thread_swapout(td);
422 	pages = td->td_kstack_pages;
423 	ksobj = td->td_kstack_obj;
424 	pmap_qremove(td->td_kstack, pages);
425 	VM_OBJECT_LOCK(ksobj);
426 	for (i = 0; i < pages; i++) {
427 		m = vm_page_lookup(ksobj, i);
428 		if (m == NULL)
429 			panic("vm_thread_swapout: kstack already missing?");
430 		vm_page_lock_queues();
431 		vm_page_dirty(m);
432 		vm_page_unwire(m, 0);
433 		vm_page_unlock_queues();
434 	}
435 	VM_OBJECT_UNLOCK(ksobj);
436 }
437 
438 /*
439  * Bring the kernel stack for a specified thread back in.
440  */
441 void
442 vm_thread_swapin(struct thread *td)
443 {
444 	vm_object_t ksobj;
445 	vm_page_t m, ma[KSTACK_MAX_PAGES];
446 	int i, pages, rv;
447 
448 	pages = td->td_kstack_pages;
449 	ksobj = td->td_kstack_obj;
450 	VM_OBJECT_LOCK(ksobj);
451 	for (i = 0; i < pages; i++) {
452 		m = vm_page_grab(ksobj, i, VM_ALLOC_NORMAL | VM_ALLOC_RETRY);
453 		if (m->valid != VM_PAGE_BITS_ALL) {
454 			rv = vm_pager_get_pages(ksobj, &m, 1, 0);
455 			if (rv != VM_PAGER_OK)
456 				panic("vm_thread_swapin: cannot get kstack for proc: %d", td->td_proc->p_pid);
457 			m = vm_page_lookup(ksobj, i);
458 		}
459 		ma[i] = m;
460 		vm_page_lock_queues();
461 		vm_page_wire(m);
462 		vm_page_unlock_queues();
463 		vm_page_wakeup(m);
464 	}
465 	VM_OBJECT_UNLOCK(ksobj);
466 	pmap_qenter(td->td_kstack, ma, pages);
467 	cpu_thread_swapin(td);
468 }
469 
470 /*
471  * Set up a variable-sized alternate kstack.
472  */
473 int
474 vm_thread_new_altkstack(struct thread *td, int pages)
475 {
476 
477 	td->td_altkstack = td->td_kstack;
478 	td->td_altkstack_obj = td->td_kstack_obj;
479 	td->td_altkstack_pages = td->td_kstack_pages;
480 
481 	return (vm_thread_new(td, pages));
482 }
483 
484 /*
485  * Restore the original kstack.
486  */
487 void
488 vm_thread_dispose_altkstack(struct thread *td)
489 {
490 
491 	vm_thread_dispose(td);
492 
493 	td->td_kstack = td->td_altkstack;
494 	td->td_kstack_obj = td->td_altkstack_obj;
495 	td->td_kstack_pages = td->td_altkstack_pages;
496 	td->td_altkstack = 0;
497 	td->td_altkstack_obj = NULL;
498 	td->td_altkstack_pages = 0;
499 }
500 
501 /*
502  * Implement fork's actions on an address space.
503  * Here we arrange for the address space to be copied or referenced,
504  * allocate a user struct (pcb and kernel stack), then call the
505  * machine-dependent layer to fill those in and make the new process
506  * ready to run.  The new process is set up so that it returns directly
507  * to user mode to avoid stack copying and relocation problems.
508  */
509 int
510 vm_forkproc(td, p2, td2, vm2, flags)
511 	struct thread *td;
512 	struct proc *p2;
513 	struct thread *td2;
514 	struct vmspace *vm2;
515 	int flags;
516 {
517 	struct proc *p1 = td->td_proc;
518 	int error;
519 
520 	if ((flags & RFPROC) == 0) {
521 		/*
522 		 * Divorce the memory, if it is shared, essentially
523 		 * this changes shared memory amongst threads, into
524 		 * COW locally.
525 		 */
526 		if ((flags & RFMEM) == 0) {
527 			if (p1->p_vmspace->vm_refcnt > 1) {
528 				error = vmspace_unshare(p1);
529 				if (error)
530 					return (error);
531 			}
532 		}
533 		cpu_fork(td, p2, td2, flags);
534 		return (0);
535 	}
536 
537 	if (flags & RFMEM) {
538 		p2->p_vmspace = p1->p_vmspace;
539 		atomic_add_int(&p1->p_vmspace->vm_refcnt, 1);
540 	}
541 
542 	while (vm_page_count_severe()) {
543 		VM_WAIT;
544 	}
545 
546 	if ((flags & RFMEM) == 0) {
547 		p2->p_vmspace = vm2;
548 		if (p1->p_vmspace->vm_shm)
549 			shmfork(p1, p2);
550 	}
551 
552 	/*
553 	 * cpu_fork will copy and update the pcb, set up the kernel stack,
554 	 * and make the child ready to run.
555 	 */
556 	cpu_fork(td, p2, td2, flags);
557 	return (0);
558 }
559 
560 /*
561  * Called after process has been wait(2)'ed apon and is being reaped.
562  * The idea is to reclaim resources that we could not reclaim while
563  * the process was still executing.
564  */
565 void
566 vm_waitproc(p)
567 	struct proc *p;
568 {
569 
570 	vmspace_exitfree(p);		/* and clean-out the vmspace */
571 }
572 
573 /*
574  * Set default limits for VM system.
575  * Called for proc 0, and then inherited by all others.
576  *
577  * XXX should probably act directly on proc0.
578  */
579 static void
580 vm_init_limits(udata)
581 	void *udata;
582 {
583 	struct proc *p = udata;
584 	struct plimit *limp;
585 	int rss_limit;
586 
587 	/*
588 	 * Set up the initial limits on process VM. Set the maximum resident
589 	 * set size to be half of (reasonably) available memory.  Since this
590 	 * is a soft limit, it comes into effect only when the system is out
591 	 * of memory - half of main memory helps to favor smaller processes,
592 	 * and reduces thrashing of the object cache.
593 	 */
594 	limp = p->p_limit;
595 	limp->pl_rlimit[RLIMIT_STACK].rlim_cur = dflssiz;
596 	limp->pl_rlimit[RLIMIT_STACK].rlim_max = maxssiz;
597 	limp->pl_rlimit[RLIMIT_DATA].rlim_cur = dfldsiz;
598 	limp->pl_rlimit[RLIMIT_DATA].rlim_max = maxdsiz;
599 	/* limit the limit to no less than 2MB */
600 	rss_limit = max(cnt.v_free_count, 512);
601 	limp->pl_rlimit[RLIMIT_RSS].rlim_cur = ptoa(rss_limit);
602 	limp->pl_rlimit[RLIMIT_RSS].rlim_max = RLIM_INFINITY;
603 }
604 
605 void
606 faultin(p)
607 	struct proc *p;
608 {
609 #ifdef NO_SWAPPING
610 
611 	PROC_LOCK_ASSERT(p, MA_OWNED);
612 	if ((p->p_flag & P_INMEM) == 0)
613 		panic("faultin: proc swapped out with NO_SWAPPING!");
614 #else /* !NO_SWAPPING */
615 	struct thread *td;
616 
617 	PROC_LOCK_ASSERT(p, MA_OWNED);
618 	/*
619 	 * If another process is swapping in this process,
620 	 * just wait until it finishes.
621 	 */
622 	if (p->p_flag & P_SWAPPINGIN) {
623 		while (p->p_flag & P_SWAPPINGIN)
624 			msleep(&p->p_flag, &p->p_mtx, PVM, "faultin", 0);
625 		return;
626 	}
627 	if ((p->p_flag & P_INMEM) == 0) {
628 		/*
629 		 * Don't let another thread swap process p out while we are
630 		 * busy swapping it in.
631 		 */
632 		++p->p_lock;
633 		p->p_flag |= P_SWAPPINGIN;
634 		PROC_UNLOCK(p);
635 
636 		/*
637 		 * We hold no lock here because the list of threads
638 		 * can not change while all threads in the process are
639 		 * swapped out.
640 		 */
641 		FOREACH_THREAD_IN_PROC(p, td)
642 			vm_thread_swapin(td);
643 		PROC_LOCK(p);
644 		swapclear(p);
645 		p->p_swtick = ticks;
646 
647 		wakeup(&p->p_flag);
648 
649 		/* Allow other threads to swap p out now. */
650 		--p->p_lock;
651 	}
652 #endif /* NO_SWAPPING */
653 }
654 
655 /*
656  * This swapin algorithm attempts to swap-in processes only if there
657  * is enough space for them.  Of course, if a process waits for a long
658  * time, it will be swapped in anyway.
659  *
660  * Giant is held on entry.
661  */
662 /* ARGSUSED*/
663 static void
664 scheduler(dummy)
665 	void *dummy;
666 {
667 	struct proc *p;
668 	struct thread *td;
669 	struct proc *pp;
670 	int slptime;
671 	int swtime;
672 	int ppri;
673 	int pri;
674 
675 	mtx_assert(&Giant, MA_OWNED | MA_NOTRECURSED);
676 	mtx_unlock(&Giant);
677 
678 loop:
679 	if (vm_page_count_min()) {
680 		VM_WAIT;
681 		goto loop;
682 	}
683 
684 	pp = NULL;
685 	ppri = INT_MIN;
686 	sx_slock(&allproc_lock);
687 	FOREACH_PROC_IN_SYSTEM(p) {
688 		PROC_LOCK(p);
689 		if (p->p_flag & (P_SWAPPINGOUT | P_SWAPPINGIN | P_INMEM)) {
690 			PROC_UNLOCK(p);
691 			continue;
692 		}
693 		swtime = (ticks - p->p_swtick) / hz;
694 		FOREACH_THREAD_IN_PROC(p, td) {
695 			/*
696 			 * An otherwise runnable thread of a process
697 			 * swapped out has only the TDI_SWAPPED bit set.
698 			 *
699 			 */
700 			thread_lock(td);
701 			if (td->td_inhibitors == TDI_SWAPPED) {
702 				slptime = (ticks - td->td_slptick) / hz;
703 				pri = swtime + slptime;
704 				if ((td->td_flags & TDF_SWAPINREQ) == 0)
705 					pri -= p->p_nice * 8;
706 				/*
707 				 * if this thread is higher priority
708 				 * and there is enough space, then select
709 				 * this process instead of the previous
710 				 * selection.
711 				 */
712 				if (pri > ppri) {
713 					pp = p;
714 					ppri = pri;
715 				}
716 			}
717 			thread_unlock(td);
718 		}
719 		PROC_UNLOCK(p);
720 	}
721 	sx_sunlock(&allproc_lock);
722 
723 	/*
724 	 * Nothing to do, back to sleep.
725 	 */
726 	if ((p = pp) == NULL) {
727 		tsleep(&proc0, PVM, "sched", maxslp * hz / 2);
728 		goto loop;
729 	}
730 	PROC_LOCK(p);
731 
732 	/*
733 	 * Another process may be bringing or may have already
734 	 * brought this process in while we traverse all threads.
735 	 * Or, this process may even be being swapped out again.
736 	 */
737 	if (p->p_flag & (P_INMEM | P_SWAPPINGOUT | P_SWAPPINGIN)) {
738 		PROC_UNLOCK(p);
739 		goto loop;
740 	}
741 
742 	/*
743 	 * We would like to bring someone in. (only if there is space).
744 	 * [What checks the space? ]
745 	 */
746 	faultin(p);
747 	PROC_UNLOCK(p);
748 	goto loop;
749 }
750 
751 void
752 kick_proc0(void)
753 {
754 
755 	wakeup(&proc0);
756 }
757 
758 #ifndef NO_SWAPPING
759 
760 /*
761  * Swap_idle_threshold1 is the guaranteed swapped in time for a process
762  */
763 static int swap_idle_threshold1 = 2;
764 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold1, CTLFLAG_RW,
765     &swap_idle_threshold1, 0, "Guaranteed swapped in time for a process");
766 
767 /*
768  * Swap_idle_threshold2 is the time that a process can be idle before
769  * it will be swapped out, if idle swapping is enabled.
770  */
771 static int swap_idle_threshold2 = 10;
772 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold2, CTLFLAG_RW,
773     &swap_idle_threshold2, 0, "Time before a process will be swapped out");
774 
775 /*
776  * Swapout is driven by the pageout daemon.  Very simple, we find eligible
777  * procs and swap out their stacks.  We try to always "swap" at least one
778  * process in case we need the room for a swapin.
779  * If any procs have been sleeping/stopped for at least maxslp seconds,
780  * they are swapped.  Else, we swap the longest-sleeping or stopped process,
781  * if any, otherwise the longest-resident process.
782  */
783 void
784 swapout_procs(action)
785 int action;
786 {
787 	struct proc *p;
788 	struct thread *td;
789 	int didswap = 0;
790 
791 retry:
792 	sx_slock(&allproc_lock);
793 	FOREACH_PROC_IN_SYSTEM(p) {
794 		struct vmspace *vm;
795 		int minslptime = 100000;
796 		int slptime;
797 
798 		/*
799 		 * Watch out for a process in
800 		 * creation.  It may have no
801 		 * address space or lock yet.
802 		 */
803 		if (p->p_state == PRS_NEW)
804 			continue;
805 		/*
806 		 * An aio daemon switches its
807 		 * address space while running.
808 		 * Perform a quick check whether
809 		 * a process has P_SYSTEM.
810 		 */
811 		if ((p->p_flag & P_SYSTEM) != 0)
812 			continue;
813 		/*
814 		 * Do not swapout a process that
815 		 * is waiting for VM data
816 		 * structures as there is a possible
817 		 * deadlock.  Test this first as
818 		 * this may block.
819 		 *
820 		 * Lock the map until swapout
821 		 * finishes, or a thread of this
822 		 * process may attempt to alter
823 		 * the map.
824 		 */
825 		vm = vmspace_acquire_ref(p);
826 		if (vm == NULL)
827 			continue;
828 		if (!vm_map_trylock(&vm->vm_map))
829 			goto nextproc1;
830 
831 		PROC_LOCK(p);
832 		if (p->p_lock != 0 ||
833 		    (p->p_flag & (P_STOPPED_SINGLE|P_TRACED|P_SYSTEM|P_WEXIT)
834 		    ) != 0) {
835 			goto nextproc;
836 		}
837 		/*
838 		 * only aiod changes vmspace, however it will be
839 		 * skipped because of the if statement above checking
840 		 * for P_SYSTEM
841 		 */
842 		if ((p->p_flag & (P_INMEM|P_SWAPPINGOUT|P_SWAPPINGIN)) != P_INMEM)
843 			goto nextproc;
844 
845 		switch (p->p_state) {
846 		default:
847 			/* Don't swap out processes in any sort
848 			 * of 'special' state. */
849 			break;
850 
851 		case PRS_NORMAL:
852 			/*
853 			 * do not swapout a realtime process
854 			 * Check all the thread groups..
855 			 */
856 			FOREACH_THREAD_IN_PROC(p, td) {
857 				thread_lock(td);
858 				if (PRI_IS_REALTIME(td->td_pri_class)) {
859 					thread_unlock(td);
860 					goto nextproc;
861 				}
862 				slptime = (ticks - td->td_slptick) / hz;
863 				/*
864 				 * Guarantee swap_idle_threshold1
865 				 * time in memory.
866 				 */
867 				if (slptime < swap_idle_threshold1) {
868 					thread_unlock(td);
869 					goto nextproc;
870 				}
871 
872 				/*
873 				 * Do not swapout a process if it is
874 				 * waiting on a critical event of some
875 				 * kind or there is a thread whose
876 				 * pageable memory may be accessed.
877 				 *
878 				 * This could be refined to support
879 				 * swapping out a thread.
880 				 */
881 				if (!thread_safetoswapout(td)) {
882 					thread_unlock(td);
883 					goto nextproc;
884 				}
885 				/*
886 				 * If the system is under memory stress,
887 				 * or if we are swapping
888 				 * idle processes >= swap_idle_threshold2,
889 				 * then swap the process out.
890 				 */
891 				if (((action & VM_SWAP_NORMAL) == 0) &&
892 				    (((action & VM_SWAP_IDLE) == 0) ||
893 				    (slptime < swap_idle_threshold2))) {
894 					thread_unlock(td);
895 					goto nextproc;
896 				}
897 
898 				if (minslptime > slptime)
899 					minslptime = slptime;
900 				thread_unlock(td);
901 			}
902 
903 			/*
904 			 * If the pageout daemon didn't free enough pages,
905 			 * or if this process is idle and the system is
906 			 * configured to swap proactively, swap it out.
907 			 */
908 			if ((action & VM_SWAP_NORMAL) ||
909 				((action & VM_SWAP_IDLE) &&
910 				 (minslptime > swap_idle_threshold2))) {
911 				if (swapout(p) == 0)
912 					didswap++;
913 				PROC_UNLOCK(p);
914 				vm_map_unlock(&vm->vm_map);
915 				vmspace_free(vm);
916 				sx_sunlock(&allproc_lock);
917 				goto retry;
918 			}
919 		}
920 nextproc:
921 		PROC_UNLOCK(p);
922 		vm_map_unlock(&vm->vm_map);
923 nextproc1:
924 		vmspace_free(vm);
925 		continue;
926 	}
927 	sx_sunlock(&allproc_lock);
928 	/*
929 	 * If we swapped something out, and another process needed memory,
930 	 * then wakeup the sched process.
931 	 */
932 	if (didswap)
933 		wakeup(&proc0);
934 }
935 
936 static void
937 swapclear(p)
938 	struct proc *p;
939 {
940 	struct thread *td;
941 
942 	PROC_LOCK_ASSERT(p, MA_OWNED);
943 
944 	FOREACH_THREAD_IN_PROC(p, td) {
945 		thread_lock(td);
946 		td->td_flags |= TDF_INMEM;
947 		td->td_flags &= ~TDF_SWAPINREQ;
948 		TD_CLR_SWAPPED(td);
949 		if (TD_CAN_RUN(td))
950 			if (setrunnable(td)) {
951 #ifdef INVARIANTS
952 				/*
953 				 * XXX: We just cleared TDI_SWAPPED
954 				 * above and set TDF_INMEM, so this
955 				 * should never happen.
956 				 */
957 				panic("not waking up swapper");
958 #endif
959 			}
960 		thread_unlock(td);
961 	}
962 	p->p_flag &= ~(P_SWAPPINGIN|P_SWAPPINGOUT);
963 	p->p_flag |= P_INMEM;
964 }
965 
966 static int
967 swapout(p)
968 	struct proc *p;
969 {
970 	struct thread *td;
971 
972 	PROC_LOCK_ASSERT(p, MA_OWNED);
973 #if defined(SWAP_DEBUG)
974 	printf("swapping out %d\n", p->p_pid);
975 #endif
976 
977 	/*
978 	 * The states of this process and its threads may have changed
979 	 * by now.  Assuming that there is only one pageout daemon thread,
980 	 * this process should still be in memory.
981 	 */
982 	KASSERT((p->p_flag & (P_INMEM|P_SWAPPINGOUT|P_SWAPPINGIN)) == P_INMEM,
983 		("swapout: lost a swapout race?"));
984 
985 	/*
986 	 * remember the process resident count
987 	 */
988 	p->p_vmspace->vm_swrss = vmspace_resident_count(p->p_vmspace);
989 	/*
990 	 * Check and mark all threads before we proceed.
991 	 */
992 	p->p_flag &= ~P_INMEM;
993 	p->p_flag |= P_SWAPPINGOUT;
994 	FOREACH_THREAD_IN_PROC(p, td) {
995 		thread_lock(td);
996 		if (!thread_safetoswapout(td)) {
997 			thread_unlock(td);
998 			swapclear(p);
999 			return (EBUSY);
1000 		}
1001 		td->td_flags &= ~TDF_INMEM;
1002 		TD_SET_SWAPPED(td);
1003 		thread_unlock(td);
1004 	}
1005 	td = FIRST_THREAD_IN_PROC(p);
1006 	++td->td_ru.ru_nswap;
1007 	PROC_UNLOCK(p);
1008 
1009 	/*
1010 	 * This list is stable because all threads are now prevented from
1011 	 * running.  The list is only modified in the context of a running
1012 	 * thread in this process.
1013 	 */
1014 	FOREACH_THREAD_IN_PROC(p, td)
1015 		vm_thread_swapout(td);
1016 
1017 	PROC_LOCK(p);
1018 	p->p_flag &= ~P_SWAPPINGOUT;
1019 	p->p_swtick = ticks;
1020 	return (0);
1021 }
1022 #endif /* !NO_SWAPPING */
1023