xref: /freebsd/sys/vm/vm_glue.c (revision d8b878873e7aa8df1972cc6a642804b17eb61087)
1 /*-
2  * Copyright (c) 1991, 1993
3  *	The Regents of the University of California.  All rights reserved.
4  *
5  * This code is derived from software contributed to Berkeley by
6  * The Mach Operating System project at Carnegie-Mellon University.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  * 4. Neither the name of the University nor the names of its contributors
17  *    may be used to endorse or promote products derived from this software
18  *    without specific prior written permission.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
21  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
24  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30  * SUCH DAMAGE.
31  *
32  *	from: @(#)vm_glue.c	8.6 (Berkeley) 1/5/94
33  *
34  *
35  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
36  * All rights reserved.
37  *
38  * Permission to use, copy, modify and distribute this software and
39  * its documentation is hereby granted, provided that both the copyright
40  * notice and this permission notice appear in all copies of the
41  * software, derivative works or modified versions, and any portions
42  * thereof, and that both notices appear in supporting documentation.
43  *
44  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
45  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
46  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
47  *
48  * Carnegie Mellon requests users of this software to return to
49  *
50  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
51  *  School of Computer Science
52  *  Carnegie Mellon University
53  *  Pittsburgh PA 15213-3890
54  *
55  * any improvements or extensions that they make and grant Carnegie the
56  * rights to redistribute these changes.
57  */
58 
59 #include <sys/cdefs.h>
60 __FBSDID("$FreeBSD$");
61 
62 #include "opt_vm.h"
63 #include "opt_kstack_pages.h"
64 #include "opt_kstack_max_pages.h"
65 
66 #include <sys/param.h>
67 #include <sys/systm.h>
68 #include <sys/limits.h>
69 #include <sys/lock.h>
70 #include <sys/mutex.h>
71 #include <sys/proc.h>
72 #include <sys/resourcevar.h>
73 #include <sys/sched.h>
74 #include <sys/sf_buf.h>
75 #include <sys/shm.h>
76 #include <sys/vmmeter.h>
77 #include <sys/sx.h>
78 #include <sys/sysctl.h>
79 
80 #include <sys/eventhandler.h>
81 #include <sys/kernel.h>
82 #include <sys/ktr.h>
83 #include <sys/unistd.h>
84 
85 #include <vm/vm.h>
86 #include <vm/vm_param.h>
87 #include <vm/pmap.h>
88 #include <vm/vm_map.h>
89 #include <vm/vm_page.h>
90 #include <vm/vm_pageout.h>
91 #include <vm/vm_object.h>
92 #include <vm/vm_kern.h>
93 #include <vm/vm_extern.h>
94 #include <vm/vm_pager.h>
95 #include <vm/swap_pager.h>
96 
97 extern int maxslp;
98 
99 /*
100  * System initialization
101  *
102  * THIS MUST BE THE LAST INITIALIZATION ITEM!!!
103  *
104  * Note: run scheduling should be divorced from the vm system.
105  */
106 static void scheduler(void *);
107 SYSINIT(scheduler, SI_SUB_RUN_SCHEDULER, SI_ORDER_ANY, scheduler, NULL);
108 
109 #ifndef NO_SWAPPING
110 static int swapout(struct proc *);
111 static void swapclear(struct proc *);
112 static void vm_thread_swapin(struct thread *td);
113 static void vm_thread_swapout(struct thread *td);
114 #endif
115 
116 /*
117  * MPSAFE
118  *
119  * WARNING!  This code calls vm_map_check_protection() which only checks
120  * the associated vm_map_entry range.  It does not determine whether the
121  * contents of the memory is actually readable or writable.  In most cases
122  * just checking the vm_map_entry is sufficient within the kernel's address
123  * space.
124  */
125 int
126 kernacc(addr, len, rw)
127 	void *addr;
128 	int len, rw;
129 {
130 	boolean_t rv;
131 	vm_offset_t saddr, eaddr;
132 	vm_prot_t prot;
133 
134 	KASSERT((rw & ~VM_PROT_ALL) == 0,
135 	    ("illegal ``rw'' argument to kernacc (%x)\n", rw));
136 
137 	if ((vm_offset_t)addr + len > kernel_map->max_offset ||
138 	    (vm_offset_t)addr + len < (vm_offset_t)addr)
139 		return (FALSE);
140 
141 	prot = rw;
142 	saddr = trunc_page((vm_offset_t)addr);
143 	eaddr = round_page((vm_offset_t)addr + len);
144 	vm_map_lock_read(kernel_map);
145 	rv = vm_map_check_protection(kernel_map, saddr, eaddr, prot);
146 	vm_map_unlock_read(kernel_map);
147 	return (rv == TRUE);
148 }
149 
150 /*
151  * MPSAFE
152  *
153  * WARNING!  This code calls vm_map_check_protection() which only checks
154  * the associated vm_map_entry range.  It does not determine whether the
155  * contents of the memory is actually readable or writable.  vmapbuf(),
156  * vm_fault_quick(), or copyin()/copout()/su*()/fu*() functions should be
157  * used in conjuction with this call.
158  */
159 int
160 useracc(addr, len, rw)
161 	void *addr;
162 	int len, rw;
163 {
164 	boolean_t rv;
165 	vm_prot_t prot;
166 	vm_map_t map;
167 
168 	KASSERT((rw & ~VM_PROT_ALL) == 0,
169 	    ("illegal ``rw'' argument to useracc (%x)\n", rw));
170 	prot = rw;
171 	map = &curproc->p_vmspace->vm_map;
172 	if ((vm_offset_t)addr + len > vm_map_max(map) ||
173 	    (vm_offset_t)addr + len < (vm_offset_t)addr) {
174 		return (FALSE);
175 	}
176 	vm_map_lock_read(map);
177 	rv = vm_map_check_protection(map, trunc_page((vm_offset_t)addr),
178 	    round_page((vm_offset_t)addr + len), prot);
179 	vm_map_unlock_read(map);
180 	return (rv == TRUE);
181 }
182 
183 int
184 vslock(void *addr, size_t len)
185 {
186 	vm_offset_t end, last, start;
187 	vm_size_t npages;
188 	int error;
189 
190 	last = (vm_offset_t)addr + len;
191 	start = trunc_page((vm_offset_t)addr);
192 	end = round_page(last);
193 	if (last < (vm_offset_t)addr || end < (vm_offset_t)addr)
194 		return (EINVAL);
195 	npages = atop(end - start);
196 	if (npages > vm_page_max_wired)
197 		return (ENOMEM);
198 	PROC_LOCK(curproc);
199 	if (ptoa(npages +
200 	    pmap_wired_count(vm_map_pmap(&curproc->p_vmspace->vm_map))) >
201 	    lim_cur(curproc, RLIMIT_MEMLOCK)) {
202 		PROC_UNLOCK(curproc);
203 		return (ENOMEM);
204 	}
205 	PROC_UNLOCK(curproc);
206 #if 0
207 	/*
208 	 * XXX - not yet
209 	 *
210 	 * The limit for transient usage of wired pages should be
211 	 * larger than for "permanent" wired pages (mlock()).
212 	 *
213 	 * Also, the sysctl code, which is the only present user
214 	 * of vslock(), does a hard loop on EAGAIN.
215 	 */
216 	if (npages + cnt.v_wire_count > vm_page_max_wired)
217 		return (EAGAIN);
218 #endif
219 	error = vm_map_wire(&curproc->p_vmspace->vm_map, start, end,
220 	    VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES);
221 	/*
222 	 * Return EFAULT on error to match copy{in,out}() behaviour
223 	 * rather than returning ENOMEM like mlock() would.
224 	 */
225 	return (error == KERN_SUCCESS ? 0 : EFAULT);
226 }
227 
228 void
229 vsunlock(void *addr, size_t len)
230 {
231 
232 	/* Rely on the parameter sanity checks performed by vslock(). */
233 	(void)vm_map_unwire(&curproc->p_vmspace->vm_map,
234 	    trunc_page((vm_offset_t)addr), round_page((vm_offset_t)addr + len),
235 	    VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES);
236 }
237 
238 /*
239  * Pin the page contained within the given object at the given offset.  If the
240  * page is not resident, allocate and load it using the given object's pager.
241  * Return the pinned page if successful; otherwise, return NULL.
242  */
243 static vm_page_t
244 vm_imgact_hold_page(vm_object_t object, vm_ooffset_t offset)
245 {
246 	vm_page_t m, ma[1];
247 	vm_pindex_t pindex;
248 	int rv;
249 
250 	VM_OBJECT_LOCK(object);
251 	pindex = OFF_TO_IDX(offset);
252 	m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL | VM_ALLOC_RETRY);
253 	if (m->valid != VM_PAGE_BITS_ALL) {
254 		ma[0] = m;
255 		rv = vm_pager_get_pages(object, ma, 1, 0);
256 		m = vm_page_lookup(object, pindex);
257 		if (m == NULL)
258 			goto out;
259 		if (rv != VM_PAGER_OK) {
260 			vm_page_lock_queues();
261 			vm_page_free(m);
262 			vm_page_unlock_queues();
263 			m = NULL;
264 			goto out;
265 		}
266 	}
267 	vm_page_lock_queues();
268 	vm_page_hold(m);
269 	vm_page_unlock_queues();
270 	vm_page_wakeup(m);
271 out:
272 	VM_OBJECT_UNLOCK(object);
273 	return (m);
274 }
275 
276 /*
277  * Return a CPU private mapping to the page at the given offset within the
278  * given object.  The page is pinned before it is mapped.
279  */
280 struct sf_buf *
281 vm_imgact_map_page(vm_object_t object, vm_ooffset_t offset)
282 {
283 	vm_page_t m;
284 
285 	m = vm_imgact_hold_page(object, offset);
286 	if (m == NULL)
287 		return (NULL);
288 	sched_pin();
289 	return (sf_buf_alloc(m, SFB_CPUPRIVATE));
290 }
291 
292 /*
293  * Destroy the given CPU private mapping and unpin the page that it mapped.
294  */
295 void
296 vm_imgact_unmap_page(struct sf_buf *sf)
297 {
298 	vm_page_t m;
299 
300 	m = sf_buf_page(sf);
301 	sf_buf_free(sf);
302 	sched_unpin();
303 	vm_page_lock_queues();
304 	vm_page_unhold(m);
305 	vm_page_unlock_queues();
306 }
307 
308 void
309 vm_sync_icache(vm_map_t map, vm_offset_t va, vm_offset_t sz)
310 {
311 
312 	pmap_sync_icache(map->pmap, va, sz);
313 }
314 
315 struct kstack_cache_entry {
316 	vm_object_t ksobj;
317 	struct kstack_cache_entry *next_ks_entry;
318 };
319 
320 static struct kstack_cache_entry *kstack_cache;
321 static int kstack_cache_size = 128;
322 static int kstacks;
323 static struct mtx kstack_cache_mtx;
324 SYSCTL_INT(_vm, OID_AUTO, kstack_cache_size, CTLFLAG_RW, &kstack_cache_size, 0,
325     "");
326 SYSCTL_INT(_vm, OID_AUTO, kstacks, CTLFLAG_RD, &kstacks, 0,
327     "");
328 
329 #ifndef KSTACK_MAX_PAGES
330 #define KSTACK_MAX_PAGES 32
331 #endif
332 
333 /*
334  * Create the kernel stack (including pcb for i386) for a new thread.
335  * This routine directly affects the fork perf for a process and
336  * create performance for a thread.
337  */
338 int
339 vm_thread_new(struct thread *td, int pages)
340 {
341 	vm_object_t ksobj;
342 	vm_offset_t ks;
343 	vm_page_t m, ma[KSTACK_MAX_PAGES];
344 	struct kstack_cache_entry *ks_ce;
345 	int i;
346 
347 	/* Bounds check */
348 	if (pages <= 1)
349 		pages = KSTACK_PAGES;
350 	else if (pages > KSTACK_MAX_PAGES)
351 		pages = KSTACK_MAX_PAGES;
352 
353 	if (pages == KSTACK_PAGES) {
354 		mtx_lock(&kstack_cache_mtx);
355 		if (kstack_cache != NULL) {
356 			ks_ce = kstack_cache;
357 			kstack_cache = ks_ce->next_ks_entry;
358 			mtx_unlock(&kstack_cache_mtx);
359 
360 			td->td_kstack_obj = ks_ce->ksobj;
361 			td->td_kstack = (vm_offset_t)ks_ce;
362 			td->td_kstack_pages = KSTACK_PAGES;
363 			return (1);
364 		}
365 		mtx_unlock(&kstack_cache_mtx);
366 	}
367 
368 	/*
369 	 * Allocate an object for the kstack.
370 	 */
371 	ksobj = vm_object_allocate(OBJT_DEFAULT, pages);
372 
373 	/*
374 	 * Get a kernel virtual address for this thread's kstack.
375 	 */
376 #if defined(__mips__)
377 	/*
378 	 * We need to align the kstack's mapped address to fit within
379 	 * a single TLB entry.
380 	 */
381 	ks = kmem_alloc_nofault_space(kernel_map,
382 	    (pages + KSTACK_GUARD_PAGES) * PAGE_SIZE, VMFS_TLB_ALIGNED_SPACE);
383 #else
384 	ks = kmem_alloc_nofault(kernel_map,
385 	   (pages + KSTACK_GUARD_PAGES) * PAGE_SIZE);
386 #endif
387 	if (ks == 0) {
388 		printf("vm_thread_new: kstack allocation failed\n");
389 		vm_object_deallocate(ksobj);
390 		return (0);
391 	}
392 
393 	atomic_add_int(&kstacks, 1);
394 	if (KSTACK_GUARD_PAGES != 0) {
395 		pmap_qremove(ks, KSTACK_GUARD_PAGES);
396 		ks += KSTACK_GUARD_PAGES * PAGE_SIZE;
397 	}
398 	td->td_kstack_obj = ksobj;
399 	td->td_kstack = ks;
400 	/*
401 	 * Knowing the number of pages allocated is useful when you
402 	 * want to deallocate them.
403 	 */
404 	td->td_kstack_pages = pages;
405 	/*
406 	 * For the length of the stack, link in a real page of ram for each
407 	 * page of stack.
408 	 */
409 	VM_OBJECT_LOCK(ksobj);
410 	for (i = 0; i < pages; i++) {
411 		/*
412 		 * Get a kernel stack page.
413 		 */
414 		m = vm_page_grab(ksobj, i, VM_ALLOC_NOBUSY |
415 		    VM_ALLOC_NORMAL | VM_ALLOC_RETRY | VM_ALLOC_WIRED);
416 		ma[i] = m;
417 		m->valid = VM_PAGE_BITS_ALL;
418 	}
419 	VM_OBJECT_UNLOCK(ksobj);
420 	pmap_qenter(ks, ma, pages);
421 	return (1);
422 }
423 
424 static void
425 vm_thread_stack_dispose(vm_object_t ksobj, vm_offset_t ks, int pages)
426 {
427 	vm_page_t m;
428 	int i;
429 
430 	atomic_add_int(&kstacks, -1);
431 	pmap_qremove(ks, pages);
432 	VM_OBJECT_LOCK(ksobj);
433 	for (i = 0; i < pages; i++) {
434 		m = vm_page_lookup(ksobj, i);
435 		if (m == NULL)
436 			panic("vm_thread_dispose: kstack already missing?");
437 		vm_page_lock_queues();
438 		vm_page_unwire(m, 0);
439 		vm_page_free(m);
440 		vm_page_unlock_queues();
441 	}
442 	VM_OBJECT_UNLOCK(ksobj);
443 	vm_object_deallocate(ksobj);
444 	kmem_free(kernel_map, ks - (KSTACK_GUARD_PAGES * PAGE_SIZE),
445 	    (pages + KSTACK_GUARD_PAGES) * PAGE_SIZE);
446 }
447 
448 /*
449  * Dispose of a thread's kernel stack.
450  */
451 void
452 vm_thread_dispose(struct thread *td)
453 {
454 	vm_object_t ksobj;
455 	vm_offset_t ks;
456 	struct kstack_cache_entry *ks_ce;
457 	int pages;
458 
459 	pages = td->td_kstack_pages;
460 	ksobj = td->td_kstack_obj;
461 	ks = td->td_kstack;
462 	td->td_kstack = 0;
463 	td->td_kstack_pages = 0;
464 	if (pages == KSTACK_PAGES && kstacks <= kstack_cache_size) {
465 		ks_ce = (struct kstack_cache_entry *)ks;
466 		ks_ce->ksobj = ksobj;
467 		mtx_lock(&kstack_cache_mtx);
468 		ks_ce->next_ks_entry = kstack_cache;
469 		kstack_cache = ks_ce;
470 		mtx_unlock(&kstack_cache_mtx);
471 		return;
472 	}
473 	vm_thread_stack_dispose(ksobj, ks, pages);
474 }
475 
476 static void
477 vm_thread_stack_lowmem(void *nulll)
478 {
479 	struct kstack_cache_entry *ks_ce, *ks_ce1;
480 
481 	mtx_lock(&kstack_cache_mtx);
482 	ks_ce = kstack_cache;
483 	kstack_cache = NULL;
484 	mtx_unlock(&kstack_cache_mtx);
485 
486 	while (ks_ce != NULL) {
487 		ks_ce1 = ks_ce;
488 		ks_ce = ks_ce->next_ks_entry;
489 
490 		vm_thread_stack_dispose(ks_ce1->ksobj, (vm_offset_t)ks_ce1,
491 		    KSTACK_PAGES);
492 	}
493 }
494 
495 static void
496 kstack_cache_init(void *nulll)
497 {
498 
499 	EVENTHANDLER_REGISTER(vm_lowmem, vm_thread_stack_lowmem, NULL,
500 	    EVENTHANDLER_PRI_ANY);
501 }
502 
503 MTX_SYSINIT(kstack_cache, &kstack_cache_mtx, "kstkch", MTX_DEF);
504 SYSINIT(vm_kstacks, SI_SUB_KTHREAD_INIT, SI_ORDER_ANY, kstack_cache_init, NULL);
505 
506 #ifndef NO_SWAPPING
507 /*
508  * Allow a thread's kernel stack to be paged out.
509  */
510 static void
511 vm_thread_swapout(struct thread *td)
512 {
513 	vm_object_t ksobj;
514 	vm_page_t m;
515 	int i, pages;
516 
517 	cpu_thread_swapout(td);
518 	pages = td->td_kstack_pages;
519 	ksobj = td->td_kstack_obj;
520 	pmap_qremove(td->td_kstack, pages);
521 	VM_OBJECT_LOCK(ksobj);
522 	for (i = 0; i < pages; i++) {
523 		m = vm_page_lookup(ksobj, i);
524 		if (m == NULL)
525 			panic("vm_thread_swapout: kstack already missing?");
526 		vm_page_dirty(m);
527 		vm_page_lock_queues();
528 		vm_page_unwire(m, 0);
529 		vm_page_unlock_queues();
530 	}
531 	VM_OBJECT_UNLOCK(ksobj);
532 }
533 
534 /*
535  * Bring the kernel stack for a specified thread back in.
536  */
537 static void
538 vm_thread_swapin(struct thread *td)
539 {
540 	vm_object_t ksobj;
541 	vm_page_t m, ma[KSTACK_MAX_PAGES];
542 	int i, pages, rv;
543 
544 	pages = td->td_kstack_pages;
545 	ksobj = td->td_kstack_obj;
546 	VM_OBJECT_LOCK(ksobj);
547 	for (i = 0; i < pages; i++) {
548 		m = vm_page_grab(ksobj, i, VM_ALLOC_NORMAL | VM_ALLOC_RETRY |
549 		    VM_ALLOC_WIRED);
550 		if (m->valid != VM_PAGE_BITS_ALL) {
551 			rv = vm_pager_get_pages(ksobj, &m, 1, 0);
552 			if (rv != VM_PAGER_OK)
553 				panic("vm_thread_swapin: cannot get kstack for proc: %d", td->td_proc->p_pid);
554 			m = vm_page_lookup(ksobj, i);
555 		}
556 		ma[i] = m;
557 		vm_page_wakeup(m);
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_flag & (P_SWAPPINGOUT | P_SWAPPINGIN | P_INMEM)) {
722 			PROC_UNLOCK(p);
723 			continue;
724 		}
725 		swtime = (ticks - p->p_swtick) / hz;
726 		FOREACH_THREAD_IN_PROC(p, td) {
727 			/*
728 			 * An otherwise runnable thread of a process
729 			 * swapped out has only the TDI_SWAPPED bit set.
730 			 *
731 			 */
732 			thread_lock(td);
733 			if (td->td_inhibitors == TDI_SWAPPED) {
734 				slptime = (ticks - td->td_slptick) / hz;
735 				pri = swtime + slptime;
736 				if ((td->td_flags & TDF_SWAPINREQ) == 0)
737 					pri -= p->p_nice * 8;
738 				/*
739 				 * if this thread is higher priority
740 				 * and there is enough space, then select
741 				 * this process instead of the previous
742 				 * selection.
743 				 */
744 				if (pri > ppri) {
745 					pp = p;
746 					ppri = pri;
747 				}
748 			}
749 			thread_unlock(td);
750 		}
751 		PROC_UNLOCK(p);
752 	}
753 	sx_sunlock(&allproc_lock);
754 
755 	/*
756 	 * Nothing to do, back to sleep.
757 	 */
758 	if ((p = pp) == NULL) {
759 		tsleep(&proc0, PVM, "sched", maxslp * hz / 2);
760 		goto loop;
761 	}
762 	PROC_LOCK(p);
763 
764 	/*
765 	 * Another process may be bringing or may have already
766 	 * brought this process in while we traverse all threads.
767 	 * Or, this process may even be being swapped out again.
768 	 */
769 	if (p->p_flag & (P_INMEM | P_SWAPPINGOUT | P_SWAPPINGIN)) {
770 		PROC_UNLOCK(p);
771 		goto loop;
772 	}
773 
774 	/*
775 	 * We would like to bring someone in. (only if there is space).
776 	 * [What checks the space? ]
777 	 */
778 	faultin(p);
779 	PROC_UNLOCK(p);
780 	goto loop;
781 }
782 
783 void
784 kick_proc0(void)
785 {
786 
787 	wakeup(&proc0);
788 }
789 
790 #ifndef NO_SWAPPING
791 
792 /*
793  * Swap_idle_threshold1 is the guaranteed swapped in time for a process
794  */
795 static int swap_idle_threshold1 = 2;
796 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold1, CTLFLAG_RW,
797     &swap_idle_threshold1, 0, "Guaranteed swapped in time for a process");
798 
799 /*
800  * Swap_idle_threshold2 is the time that a process can be idle before
801  * it will be swapped out, if idle swapping is enabled.
802  */
803 static int swap_idle_threshold2 = 10;
804 SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold2, CTLFLAG_RW,
805     &swap_idle_threshold2, 0, "Time before a process will be swapped out");
806 
807 /*
808  * Swapout is driven by the pageout daemon.  Very simple, we find eligible
809  * procs and swap out their stacks.  We try to always "swap" at least one
810  * process in case we need the room for a swapin.
811  * If any procs have been sleeping/stopped for at least maxslp seconds,
812  * they are swapped.  Else, we swap the longest-sleeping or stopped process,
813  * if any, otherwise the longest-resident process.
814  */
815 void
816 swapout_procs(action)
817 int action;
818 {
819 	struct proc *p;
820 	struct thread *td;
821 	int didswap = 0;
822 
823 retry:
824 	sx_slock(&allproc_lock);
825 	FOREACH_PROC_IN_SYSTEM(p) {
826 		struct vmspace *vm;
827 		int minslptime = 100000;
828 		int slptime;
829 
830 		/*
831 		 * Watch out for a process in
832 		 * creation.  It may have no
833 		 * address space or lock yet.
834 		 */
835 		if (p->p_state == PRS_NEW)
836 			continue;
837 		/*
838 		 * An aio daemon switches its
839 		 * address space while running.
840 		 * Perform a quick check whether
841 		 * a process has P_SYSTEM.
842 		 */
843 		if ((p->p_flag & P_SYSTEM) != 0)
844 			continue;
845 		/*
846 		 * Do not swapout a process that
847 		 * is waiting for VM data
848 		 * structures as there is a possible
849 		 * deadlock.  Test this first as
850 		 * this may block.
851 		 *
852 		 * Lock the map until swapout
853 		 * finishes, or a thread of this
854 		 * process may attempt to alter
855 		 * the map.
856 		 */
857 		vm = vmspace_acquire_ref(p);
858 		if (vm == NULL)
859 			continue;
860 		if (!vm_map_trylock(&vm->vm_map))
861 			goto nextproc1;
862 
863 		PROC_LOCK(p);
864 		if (p->p_lock != 0 ||
865 		    (p->p_flag & (P_STOPPED_SINGLE|P_TRACED|P_SYSTEM|P_WEXIT)
866 		    ) != 0) {
867 			goto nextproc;
868 		}
869 		/*
870 		 * only aiod changes vmspace, however it will be
871 		 * skipped because of the if statement above checking
872 		 * for P_SYSTEM
873 		 */
874 		if ((p->p_flag & (P_INMEM|P_SWAPPINGOUT|P_SWAPPINGIN)) != P_INMEM)
875 			goto nextproc;
876 
877 		switch (p->p_state) {
878 		default:
879 			/* Don't swap out processes in any sort
880 			 * of 'special' state. */
881 			break;
882 
883 		case PRS_NORMAL:
884 			/*
885 			 * do not swapout a realtime process
886 			 * Check all the thread groups..
887 			 */
888 			FOREACH_THREAD_IN_PROC(p, td) {
889 				thread_lock(td);
890 				if (PRI_IS_REALTIME(td->td_pri_class)) {
891 					thread_unlock(td);
892 					goto nextproc;
893 				}
894 				slptime = (ticks - td->td_slptick) / hz;
895 				/*
896 				 * Guarantee swap_idle_threshold1
897 				 * time in memory.
898 				 */
899 				if (slptime < swap_idle_threshold1) {
900 					thread_unlock(td);
901 					goto nextproc;
902 				}
903 
904 				/*
905 				 * Do not swapout a process if it is
906 				 * waiting on a critical event of some
907 				 * kind or there is a thread whose
908 				 * pageable memory may be accessed.
909 				 *
910 				 * This could be refined to support
911 				 * swapping out a thread.
912 				 */
913 				if (!thread_safetoswapout(td)) {
914 					thread_unlock(td);
915 					goto nextproc;
916 				}
917 				/*
918 				 * If the system is under memory stress,
919 				 * or if we are swapping
920 				 * idle processes >= swap_idle_threshold2,
921 				 * then swap the process out.
922 				 */
923 				if (((action & VM_SWAP_NORMAL) == 0) &&
924 				    (((action & VM_SWAP_IDLE) == 0) ||
925 				    (slptime < swap_idle_threshold2))) {
926 					thread_unlock(td);
927 					goto nextproc;
928 				}
929 
930 				if (minslptime > slptime)
931 					minslptime = slptime;
932 				thread_unlock(td);
933 			}
934 
935 			/*
936 			 * If the pageout daemon didn't free enough pages,
937 			 * or if this process is idle and the system is
938 			 * configured to swap proactively, swap it out.
939 			 */
940 			if ((action & VM_SWAP_NORMAL) ||
941 				((action & VM_SWAP_IDLE) &&
942 				 (minslptime > swap_idle_threshold2))) {
943 				if (swapout(p) == 0)
944 					didswap++;
945 				PROC_UNLOCK(p);
946 				vm_map_unlock(&vm->vm_map);
947 				vmspace_free(vm);
948 				sx_sunlock(&allproc_lock);
949 				goto retry;
950 			}
951 		}
952 nextproc:
953 		PROC_UNLOCK(p);
954 		vm_map_unlock(&vm->vm_map);
955 nextproc1:
956 		vmspace_free(vm);
957 		continue;
958 	}
959 	sx_sunlock(&allproc_lock);
960 	/*
961 	 * If we swapped something out, and another process needed memory,
962 	 * then wakeup the sched process.
963 	 */
964 	if (didswap)
965 		wakeup(&proc0);
966 }
967 
968 static void
969 swapclear(p)
970 	struct proc *p;
971 {
972 	struct thread *td;
973 
974 	PROC_LOCK_ASSERT(p, MA_OWNED);
975 
976 	FOREACH_THREAD_IN_PROC(p, td) {
977 		thread_lock(td);
978 		td->td_flags |= TDF_INMEM;
979 		td->td_flags &= ~TDF_SWAPINREQ;
980 		TD_CLR_SWAPPED(td);
981 		if (TD_CAN_RUN(td))
982 			if (setrunnable(td)) {
983 #ifdef INVARIANTS
984 				/*
985 				 * XXX: We just cleared TDI_SWAPPED
986 				 * above and set TDF_INMEM, so this
987 				 * should never happen.
988 				 */
989 				panic("not waking up swapper");
990 #endif
991 			}
992 		thread_unlock(td);
993 	}
994 	p->p_flag &= ~(P_SWAPPINGIN|P_SWAPPINGOUT);
995 	p->p_flag |= P_INMEM;
996 }
997 
998 static int
999 swapout(p)
1000 	struct proc *p;
1001 {
1002 	struct thread *td;
1003 
1004 	PROC_LOCK_ASSERT(p, MA_OWNED);
1005 #if defined(SWAP_DEBUG)
1006 	printf("swapping out %d\n", p->p_pid);
1007 #endif
1008 
1009 	/*
1010 	 * The states of this process and its threads may have changed
1011 	 * by now.  Assuming that there is only one pageout daemon thread,
1012 	 * this process should still be in memory.
1013 	 */
1014 	KASSERT((p->p_flag & (P_INMEM|P_SWAPPINGOUT|P_SWAPPINGIN)) == P_INMEM,
1015 		("swapout: lost a swapout race?"));
1016 
1017 	/*
1018 	 * remember the process resident count
1019 	 */
1020 	p->p_vmspace->vm_swrss = vmspace_resident_count(p->p_vmspace);
1021 	/*
1022 	 * Check and mark all threads before we proceed.
1023 	 */
1024 	p->p_flag &= ~P_INMEM;
1025 	p->p_flag |= P_SWAPPINGOUT;
1026 	FOREACH_THREAD_IN_PROC(p, td) {
1027 		thread_lock(td);
1028 		if (!thread_safetoswapout(td)) {
1029 			thread_unlock(td);
1030 			swapclear(p);
1031 			return (EBUSY);
1032 		}
1033 		td->td_flags &= ~TDF_INMEM;
1034 		TD_SET_SWAPPED(td);
1035 		thread_unlock(td);
1036 	}
1037 	td = FIRST_THREAD_IN_PROC(p);
1038 	++td->td_ru.ru_nswap;
1039 	PROC_UNLOCK(p);
1040 
1041 	/*
1042 	 * This list is stable because all threads are now prevented from
1043 	 * running.  The list is only modified in the context of a running
1044 	 * thread in this process.
1045 	 */
1046 	FOREACH_THREAD_IN_PROC(p, td)
1047 		vm_thread_swapout(td);
1048 
1049 	PROC_LOCK(p);
1050 	p->p_flag &= ~P_SWAPPINGOUT;
1051 	p->p_swtick = ticks;
1052 	return (0);
1053 }
1054 #endif /* !NO_SWAPPING */
1055