xref: /netbsd-src/sys/uvm/uvm_page.c (revision 37b34d511dea595d3ba03a661cf3b775038ea5f8)
1 /*	$NetBSD: uvm_page.c,v 1.79 2002/09/27 15:38:09 provos Exp $	*/
2 
3 /*
4  * Copyright (c) 1997 Charles D. Cranor and Washington University.
5  * Copyright (c) 1991, 1993, The Regents of the University of California.
6  *
7  * All rights reserved.
8  *
9  * This code is derived from software contributed to Berkeley by
10  * The Mach Operating System project at Carnegie-Mellon University.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  * 3. All advertising materials mentioning features or use of this software
21  *    must display the following acknowledgement:
22  *	This product includes software developed by Charles D. Cranor,
23  *      Washington University, the University of California, Berkeley and
24  *      its contributors.
25  * 4. Neither the name of the University nor the names of its contributors
26  *    may be used to endorse or promote products derived from this software
27  *    without specific prior written permission.
28  *
29  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
30  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
31  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
32  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
33  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
34  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
35  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
36  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
37  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
38  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
39  * SUCH DAMAGE.
40  *
41  *	@(#)vm_page.c   8.3 (Berkeley) 3/21/94
42  * from: Id: uvm_page.c,v 1.1.2.18 1998/02/06 05:24:42 chs Exp
43  *
44  *
45  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
46  * All rights reserved.
47  *
48  * Permission to use, copy, modify and distribute this software and
49  * its documentation is hereby granted, provided that both the copyright
50  * notice and this permission notice appear in all copies of the
51  * software, derivative works or modified versions, and any portions
52  * thereof, and that both notices appear in supporting documentation.
53  *
54  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
55  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
56  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
57  *
58  * Carnegie Mellon requests users of this software to return to
59  *
60  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
61  *  School of Computer Science
62  *  Carnegie Mellon University
63  *  Pittsburgh PA 15213-3890
64  *
65  * any improvements or extensions that they make and grant Carnegie the
66  * rights to redistribute these changes.
67  */
68 
69 /*
70  * uvm_page.c: page ops.
71  */
72 
73 #include <sys/cdefs.h>
74 __KERNEL_RCSID(0, "$NetBSD: uvm_page.c,v 1.79 2002/09/27 15:38:09 provos Exp $");
75 
76 #include "opt_uvmhist.h"
77 
78 #include <sys/param.h>
79 #include <sys/systm.h>
80 #include <sys/malloc.h>
81 #include <sys/sched.h>
82 #include <sys/kernel.h>
83 #include <sys/vnode.h>
84 #include <sys/proc.h>
85 
86 #define UVM_PAGE                /* pull in uvm_page.h functions */
87 #include <uvm/uvm.h>
88 
89 /*
90  * global vars... XXXCDC: move to uvm. structure.
91  */
92 
93 /*
94  * physical memory config is stored in vm_physmem.
95  */
96 
97 struct vm_physseg vm_physmem[VM_PHYSSEG_MAX];	/* XXXCDC: uvm.physmem */
98 int vm_nphysseg = 0;				/* XXXCDC: uvm.nphysseg */
99 
100 /*
101  * Some supported CPUs in a given architecture don't support all
102  * of the things necessary to do idle page zero'ing efficiently.
103  * We therefore provide a way to disable it from machdep code here.
104  */
105 /*
106  * XXX disabled until we can find a way to do this without causing
107  * problems for either cpu caches or DMA latency.
108  */
109 boolean_t vm_page_zero_enable = FALSE;
110 
111 /*
112  * local variables
113  */
114 
115 /*
116  * these variables record the values returned by vm_page_bootstrap,
117  * for debugging purposes.  The implementation of uvm_pageboot_alloc
118  * and pmap_startup here also uses them internally.
119  */
120 
121 static vaddr_t      virtual_space_start;
122 static vaddr_t      virtual_space_end;
123 
124 /*
125  * we use a hash table with only one bucket during bootup.  we will
126  * later rehash (resize) the hash table once the allocator is ready.
127  * we static allocate the one bootstrap bucket below...
128  */
129 
130 static struct pglist uvm_bootbucket;
131 
132 /*
133  * we allocate an initial number of page colors in uvm_page_init(),
134  * and remember them.  We may re-color pages as cache sizes are
135  * discovered during the autoconfiguration phase.  But we can never
136  * free the initial set of buckets, since they are allocated using
137  * uvm_pageboot_alloc().
138  */
139 
140 static boolean_t have_recolored_pages /* = FALSE */;
141 
142 /*
143  * local prototypes
144  */
145 
146 static void uvm_pageinsert __P((struct vm_page *));
147 static void uvm_pageremove __P((struct vm_page *));
148 
149 /*
150  * inline functions
151  */
152 
153 /*
154  * uvm_pageinsert: insert a page in the object and the hash table
155  *
156  * => caller must lock object
157  * => caller must lock page queues
158  * => call should have already set pg's object and offset pointers
159  *    and bumped the version counter
160  */
161 
162 __inline static void
163 uvm_pageinsert(pg)
164 	struct vm_page *pg;
165 {
166 	struct pglist *buck;
167 	struct uvm_object *uobj = pg->uobject;
168 
169 	KASSERT((pg->flags & PG_TABLED) == 0);
170 	buck = &uvm.page_hash[uvm_pagehash(uobj, pg->offset)];
171 	simple_lock(&uvm.hashlock);
172 	TAILQ_INSERT_TAIL(buck, pg, hashq);
173 	simple_unlock(&uvm.hashlock);
174 
175 	if (UVM_OBJ_IS_AOBJ(uobj)) {
176 		uvmexp.anonpages++;
177 	}
178 
179 	TAILQ_INSERT_TAIL(&uobj->memq, pg, listq);
180 	pg->flags |= PG_TABLED;
181 	uobj->uo_npages++;
182 }
183 
184 /*
185  * uvm_page_remove: remove page from object and hash
186  *
187  * => caller must lock object
188  * => caller must lock page queues
189  */
190 
191 static __inline void
192 uvm_pageremove(pg)
193 	struct vm_page *pg;
194 {
195 	struct pglist *buck;
196 	struct uvm_object *uobj = pg->uobject;
197 
198 	KASSERT(pg->flags & PG_TABLED);
199 	buck = &uvm.page_hash[uvm_pagehash(uobj ,pg->offset)];
200 	simple_lock(&uvm.hashlock);
201 	TAILQ_REMOVE(buck, pg, hashq);
202 	simple_unlock(&uvm.hashlock);
203 
204 	if (UVM_OBJ_IS_VTEXT(uobj)) {
205 		uvmexp.execpages--;
206 	} else if (UVM_OBJ_IS_VNODE(uobj)) {
207 		uvmexp.filepages--;
208 	} else if (UVM_OBJ_IS_AOBJ(uobj)) {
209 		uvmexp.anonpages--;
210 	}
211 
212 	/* object should be locked */
213 	uobj->uo_npages--;
214 	TAILQ_REMOVE(&uobj->memq, pg, listq);
215 	pg->flags &= ~PG_TABLED;
216 	pg->uobject = NULL;
217 }
218 
219 static void
220 uvm_page_init_buckets(struct pgfreelist *pgfl)
221 {
222 	int color, i;
223 
224 	for (color = 0; color < uvmexp.ncolors; color++) {
225 		for (i = 0; i < PGFL_NQUEUES; i++) {
226 			TAILQ_INIT(&pgfl->pgfl_buckets[
227 			    color].pgfl_queues[i]);
228 		}
229 	}
230 }
231 
232 /*
233  * uvm_page_init: init the page system.   called from uvm_init().
234  *
235  * => we return the range of kernel virtual memory in kvm_startp/kvm_endp
236  */
237 
238 void
239 uvm_page_init(kvm_startp, kvm_endp)
240 	vaddr_t *kvm_startp, *kvm_endp;
241 {
242 	vsize_t freepages, pagecount, bucketcount, n;
243 	struct pgflbucket *bucketarray;
244 	struct vm_page *pagearray;
245 	int lcv, i;
246 	paddr_t paddr;
247 
248 	/*
249 	 * init the page queues and page queue locks, except the free
250 	 * list; we allocate that later (with the initial vm_page
251 	 * structures).
252 	 */
253 
254 	TAILQ_INIT(&uvm.page_active);
255 	TAILQ_INIT(&uvm.page_inactive);
256 	simple_lock_init(&uvm.pageqlock);
257 	simple_lock_init(&uvm.fpageqlock);
258 
259 	/*
260 	 * init the <obj,offset> => <page> hash table.  for now
261 	 * we just have one bucket (the bootstrap bucket).  later on we
262 	 * will allocate new buckets as we dynamically resize the hash table.
263 	 */
264 
265 	uvm.page_nhash = 1;			/* 1 bucket */
266 	uvm.page_hashmask = 0;			/* mask for hash function */
267 	uvm.page_hash = &uvm_bootbucket;	/* install bootstrap bucket */
268 	TAILQ_INIT(uvm.page_hash);		/* init hash table */
269 	simple_lock_init(&uvm.hashlock);	/* init hash table lock */
270 
271 	/*
272 	 * allocate vm_page structures.
273 	 */
274 
275 	/*
276 	 * sanity check:
277 	 * before calling this function the MD code is expected to register
278 	 * some free RAM with the uvm_page_physload() function.   our job
279 	 * now is to allocate vm_page structures for this memory.
280 	 */
281 
282 	if (vm_nphysseg == 0)
283 		panic("uvm_page_bootstrap: no memory pre-allocated");
284 
285 	/*
286 	 * first calculate the number of free pages...
287 	 *
288 	 * note that we use start/end rather than avail_start/avail_end.
289 	 * this allows us to allocate extra vm_page structures in case we
290 	 * want to return some memory to the pool after booting.
291 	 */
292 
293 	freepages = 0;
294 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
295 		freepages += (vm_physmem[lcv].end - vm_physmem[lcv].start);
296 
297 	/*
298 	 * Let MD code initialize the number of colors, or default
299 	 * to 1 color if MD code doesn't care.
300 	 */
301 	if (uvmexp.ncolors == 0)
302 		uvmexp.ncolors = 1;
303 	uvmexp.colormask = uvmexp.ncolors - 1;
304 
305 	/*
306 	 * we now know we have (PAGE_SIZE * freepages) bytes of memory we can
307 	 * use.   for each page of memory we use we need a vm_page structure.
308 	 * thus, the total number of pages we can use is the total size of
309 	 * the memory divided by the PAGE_SIZE plus the size of the vm_page
310 	 * structure.   we add one to freepages as a fudge factor to avoid
311 	 * truncation errors (since we can only allocate in terms of whole
312 	 * pages).
313 	 */
314 
315 	bucketcount = uvmexp.ncolors * VM_NFREELIST;
316 	pagecount = ((freepages + 1) << PAGE_SHIFT) /
317 	    (PAGE_SIZE + sizeof(struct vm_page));
318 
319 	bucketarray = (void *)uvm_pageboot_alloc((bucketcount *
320 	    sizeof(struct pgflbucket)) + (pagecount *
321 	    sizeof(struct vm_page)));
322 	pagearray = (struct vm_page *)(bucketarray + bucketcount);
323 
324 	for (lcv = 0; lcv < VM_NFREELIST; lcv++) {
325 		uvm.page_free[lcv].pgfl_buckets =
326 		    (bucketarray + (lcv * uvmexp.ncolors));
327 		uvm_page_init_buckets(&uvm.page_free[lcv]);
328 	}
329 	memset(pagearray, 0, pagecount * sizeof(struct vm_page));
330 
331 	/*
332 	 * init the vm_page structures and put them in the correct place.
333 	 */
334 
335 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++) {
336 		n = vm_physmem[lcv].end - vm_physmem[lcv].start;
337 
338 		/* set up page array pointers */
339 		vm_physmem[lcv].pgs = pagearray;
340 		pagearray += n;
341 		pagecount -= n;
342 		vm_physmem[lcv].lastpg = vm_physmem[lcv].pgs + (n - 1);
343 
344 		/* init and free vm_pages (we've already zeroed them) */
345 		paddr = ptoa(vm_physmem[lcv].start);
346 		for (i = 0 ; i < n ; i++, paddr += PAGE_SIZE) {
347 			vm_physmem[lcv].pgs[i].phys_addr = paddr;
348 #ifdef __HAVE_VM_PAGE_MD
349 			VM_MDPAGE_INIT(&vm_physmem[lcv].pgs[i]);
350 #endif
351 			if (atop(paddr) >= vm_physmem[lcv].avail_start &&
352 			    atop(paddr) <= vm_physmem[lcv].avail_end) {
353 				uvmexp.npages++;
354 				/* add page to free pool */
355 				uvm_pagefree(&vm_physmem[lcv].pgs[i]);
356 			}
357 		}
358 	}
359 
360 	/*
361 	 * pass up the values of virtual_space_start and
362 	 * virtual_space_end (obtained by uvm_pageboot_alloc) to the upper
363 	 * layers of the VM.
364 	 */
365 
366 	*kvm_startp = round_page(virtual_space_start);
367 	*kvm_endp = trunc_page(virtual_space_end);
368 
369 	/*
370 	 * init locks for kernel threads
371 	 */
372 
373 	simple_lock_init(&uvm.pagedaemon_lock);
374 	simple_lock_init(&uvm.aiodoned_lock);
375 
376 	/*
377 	 * init various thresholds.
378 	 */
379 
380 	uvmexp.reserve_pagedaemon = 1;
381 	uvmexp.reserve_kernel = 5;
382 	uvmexp.anonminpct = 10;
383 	uvmexp.fileminpct = 10;
384 	uvmexp.execminpct = 5;
385 	uvmexp.anonmaxpct = 80;
386 	uvmexp.filemaxpct = 50;
387 	uvmexp.execmaxpct = 30;
388 	uvmexp.anonmin = uvmexp.anonminpct * 256 / 100;
389 	uvmexp.filemin = uvmexp.fileminpct * 256 / 100;
390 	uvmexp.execmin = uvmexp.execminpct * 256 / 100;
391 	uvmexp.anonmax = uvmexp.anonmaxpct * 256 / 100;
392 	uvmexp.filemax = uvmexp.filemaxpct * 256 / 100;
393 	uvmexp.execmax = uvmexp.execmaxpct * 256 / 100;
394 
395 	/*
396 	 * determine if we should zero pages in the idle loop.
397 	 */
398 
399 	uvm.page_idle_zero = vm_page_zero_enable;
400 
401 	/*
402 	 * done!
403 	 */
404 
405 	uvm.page_init_done = TRUE;
406 }
407 
408 /*
409  * uvm_setpagesize: set the page size
410  *
411  * => sets page_shift and page_mask from uvmexp.pagesize.
412  */
413 
414 void
415 uvm_setpagesize()
416 {
417 	if (uvmexp.pagesize == 0)
418 		uvmexp.pagesize = DEFAULT_PAGE_SIZE;
419 	uvmexp.pagemask = uvmexp.pagesize - 1;
420 	if ((uvmexp.pagemask & uvmexp.pagesize) != 0)
421 		panic("uvm_setpagesize: page size not a power of two");
422 	for (uvmexp.pageshift = 0; ; uvmexp.pageshift++)
423 		if ((1 << uvmexp.pageshift) == uvmexp.pagesize)
424 			break;
425 }
426 
427 /*
428  * uvm_pageboot_alloc: steal memory from physmem for bootstrapping
429  */
430 
431 vaddr_t
432 uvm_pageboot_alloc(size)
433 	vsize_t size;
434 {
435 	static boolean_t initialized = FALSE;
436 	vaddr_t addr;
437 #if !defined(PMAP_STEAL_MEMORY)
438 	vaddr_t vaddr;
439 	paddr_t paddr;
440 #endif
441 
442 	/*
443 	 * on first call to this function, initialize ourselves.
444 	 */
445 	if (initialized == FALSE) {
446 		pmap_virtual_space(&virtual_space_start, &virtual_space_end);
447 
448 		/* round it the way we like it */
449 		virtual_space_start = round_page(virtual_space_start);
450 		virtual_space_end = trunc_page(virtual_space_end);
451 
452 		initialized = TRUE;
453 	}
454 
455 	/* round to page size */
456 	size = round_page(size);
457 
458 #if defined(PMAP_STEAL_MEMORY)
459 
460 	/*
461 	 * defer bootstrap allocation to MD code (it may want to allocate
462 	 * from a direct-mapped segment).  pmap_steal_memory should adjust
463 	 * virtual_space_start/virtual_space_end if necessary.
464 	 */
465 
466 	addr = pmap_steal_memory(size, &virtual_space_start,
467 	    &virtual_space_end);
468 
469 	return(addr);
470 
471 #else /* !PMAP_STEAL_MEMORY */
472 
473 	/*
474 	 * allocate virtual memory for this request
475 	 */
476 	if (virtual_space_start == virtual_space_end ||
477 	    (virtual_space_end - virtual_space_start) < size)
478 		panic("uvm_pageboot_alloc: out of virtual space");
479 
480 	addr = virtual_space_start;
481 
482 #ifdef PMAP_GROWKERNEL
483 	/*
484 	 * If the kernel pmap can't map the requested space,
485 	 * then allocate more resources for it.
486 	 */
487 	if (uvm_maxkaddr < (addr + size)) {
488 		uvm_maxkaddr = pmap_growkernel(addr + size);
489 		if (uvm_maxkaddr < (addr + size))
490 			panic("uvm_pageboot_alloc: pmap_growkernel() failed");
491 	}
492 #endif
493 
494 	virtual_space_start += size;
495 
496 	/*
497 	 * allocate and mapin physical pages to back new virtual pages
498 	 */
499 
500 	for (vaddr = round_page(addr) ; vaddr < addr + size ;
501 	    vaddr += PAGE_SIZE) {
502 
503 		if (!uvm_page_physget(&paddr))
504 			panic("uvm_pageboot_alloc: out of memory");
505 
506 		/*
507 		 * Note this memory is no longer managed, so using
508 		 * pmap_kenter is safe.
509 		 */
510 		pmap_kenter_pa(vaddr, paddr, VM_PROT_READ|VM_PROT_WRITE);
511 	}
512 	pmap_update(pmap_kernel());
513 	return(addr);
514 #endif	/* PMAP_STEAL_MEMORY */
515 }
516 
517 #if !defined(PMAP_STEAL_MEMORY)
518 /*
519  * uvm_page_physget: "steal" one page from the vm_physmem structure.
520  *
521  * => attempt to allocate it off the end of a segment in which the "avail"
522  *    values match the start/end values.   if we can't do that, then we
523  *    will advance both values (making them equal, and removing some
524  *    vm_page structures from the non-avail area).
525  * => return false if out of memory.
526  */
527 
528 /* subroutine: try to allocate from memory chunks on the specified freelist */
529 static boolean_t uvm_page_physget_freelist __P((paddr_t *, int));
530 
531 static boolean_t
532 uvm_page_physget_freelist(paddrp, freelist)
533 	paddr_t *paddrp;
534 	int freelist;
535 {
536 	int lcv, x;
537 
538 	/* pass 1: try allocating from a matching end */
539 #if (VM_PHYSSEG_STRAT == VM_PSTRAT_BIGFIRST)
540 	for (lcv = vm_nphysseg - 1 ; lcv >= 0 ; lcv--)
541 #else
542 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
543 #endif
544 	{
545 
546 		if (uvm.page_init_done == TRUE)
547 			panic("uvm_page_physget: called _after_ bootstrap");
548 
549 		if (vm_physmem[lcv].free_list != freelist)
550 			continue;
551 
552 		/* try from front */
553 		if (vm_physmem[lcv].avail_start == vm_physmem[lcv].start &&
554 		    vm_physmem[lcv].avail_start < vm_physmem[lcv].avail_end) {
555 			*paddrp = ptoa(vm_physmem[lcv].avail_start);
556 			vm_physmem[lcv].avail_start++;
557 			vm_physmem[lcv].start++;
558 			/* nothing left?   nuke it */
559 			if (vm_physmem[lcv].avail_start ==
560 			    vm_physmem[lcv].end) {
561 				if (vm_nphysseg == 1)
562 				    panic("vum_page_physget: out of memory!");
563 				vm_nphysseg--;
564 				for (x = lcv ; x < vm_nphysseg ; x++)
565 					/* structure copy */
566 					vm_physmem[x] = vm_physmem[x+1];
567 			}
568 			return (TRUE);
569 		}
570 
571 		/* try from rear */
572 		if (vm_physmem[lcv].avail_end == vm_physmem[lcv].end &&
573 		    vm_physmem[lcv].avail_start < vm_physmem[lcv].avail_end) {
574 			*paddrp = ptoa(vm_physmem[lcv].avail_end - 1);
575 			vm_physmem[lcv].avail_end--;
576 			vm_physmem[lcv].end--;
577 			/* nothing left?   nuke it */
578 			if (vm_physmem[lcv].avail_end ==
579 			    vm_physmem[lcv].start) {
580 				if (vm_nphysseg == 1)
581 				    panic("uvm_page_physget: out of memory!");
582 				vm_nphysseg--;
583 				for (x = lcv ; x < vm_nphysseg ; x++)
584 					/* structure copy */
585 					vm_physmem[x] = vm_physmem[x+1];
586 			}
587 			return (TRUE);
588 		}
589 	}
590 
591 	/* pass2: forget about matching ends, just allocate something */
592 #if (VM_PHYSSEG_STRAT == VM_PSTRAT_BIGFIRST)
593 	for (lcv = vm_nphysseg - 1 ; lcv >= 0 ; lcv--)
594 #else
595 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
596 #endif
597 	{
598 
599 		/* any room in this bank? */
600 		if (vm_physmem[lcv].avail_start >= vm_physmem[lcv].avail_end)
601 			continue;  /* nope */
602 
603 		*paddrp = ptoa(vm_physmem[lcv].avail_start);
604 		vm_physmem[lcv].avail_start++;
605 		/* truncate! */
606 		vm_physmem[lcv].start = vm_physmem[lcv].avail_start;
607 
608 		/* nothing left?   nuke it */
609 		if (vm_physmem[lcv].avail_start == vm_physmem[lcv].end) {
610 			if (vm_nphysseg == 1)
611 				panic("uvm_page_physget: out of memory!");
612 			vm_nphysseg--;
613 			for (x = lcv ; x < vm_nphysseg ; x++)
614 				/* structure copy */
615 				vm_physmem[x] = vm_physmem[x+1];
616 		}
617 		return (TRUE);
618 	}
619 
620 	return (FALSE);        /* whoops! */
621 }
622 
623 boolean_t
624 uvm_page_physget(paddrp)
625 	paddr_t *paddrp;
626 {
627 	int i;
628 
629 	/* try in the order of freelist preference */
630 	for (i = 0; i < VM_NFREELIST; i++)
631 		if (uvm_page_physget_freelist(paddrp, i) == TRUE)
632 			return (TRUE);
633 	return (FALSE);
634 }
635 #endif /* PMAP_STEAL_MEMORY */
636 
637 /*
638  * uvm_page_physload: load physical memory into VM system
639  *
640  * => all args are PFs
641  * => all pages in start/end get vm_page structures
642  * => areas marked by avail_start/avail_end get added to the free page pool
643  * => we are limited to VM_PHYSSEG_MAX physical memory segments
644  */
645 
646 void
647 uvm_page_physload(start, end, avail_start, avail_end, free_list)
648 	paddr_t start, end, avail_start, avail_end;
649 	int free_list;
650 {
651 	int preload, lcv;
652 	psize_t npages;
653 	struct vm_page *pgs;
654 	struct vm_physseg *ps;
655 
656 	if (uvmexp.pagesize == 0)
657 		panic("uvm_page_physload: page size not set!");
658 	if (free_list >= VM_NFREELIST || free_list < VM_FREELIST_DEFAULT)
659 		panic("uvm_page_physload: bad free list %d", free_list);
660 	if (start >= end)
661 		panic("uvm_page_physload: start >= end");
662 
663 	/*
664 	 * do we have room?
665 	 */
666 
667 	if (vm_nphysseg == VM_PHYSSEG_MAX) {
668 		printf("uvm_page_physload: unable to load physical memory "
669 		    "segment\n");
670 		printf("\t%d segments allocated, ignoring 0x%llx -> 0x%llx\n",
671 		    VM_PHYSSEG_MAX, (long long)start, (long long)end);
672 		printf("\tincrease VM_PHYSSEG_MAX\n");
673 		return;
674 	}
675 
676 	/*
677 	 * check to see if this is a "preload" (i.e. uvm_mem_init hasn't been
678 	 * called yet, so malloc is not available).
679 	 */
680 
681 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++) {
682 		if (vm_physmem[lcv].pgs)
683 			break;
684 	}
685 	preload = (lcv == vm_nphysseg);
686 
687 	/*
688 	 * if VM is already running, attempt to malloc() vm_page structures
689 	 */
690 
691 	if (!preload) {
692 #if defined(VM_PHYSSEG_NOADD)
693 		panic("uvm_page_physload: tried to add RAM after vm_mem_init");
694 #else
695 		/* XXXCDC: need some sort of lockout for this case */
696 		paddr_t paddr;
697 		npages = end - start;  /* # of pages */
698 		pgs = malloc(sizeof(struct vm_page) * npages,
699 		    M_VMPAGE, M_NOWAIT);
700 		if (pgs == NULL) {
701 			printf("uvm_page_physload: can not malloc vm_page "
702 			    "structs for segment\n");
703 			printf("\tignoring 0x%lx -> 0x%lx\n", start, end);
704 			return;
705 		}
706 		/* zero data, init phys_addr and free_list, and free pages */
707 		memset(pgs, 0, sizeof(struct vm_page) * npages);
708 		for (lcv = 0, paddr = ptoa(start) ;
709 				 lcv < npages ; lcv++, paddr += PAGE_SIZE) {
710 			pgs[lcv].phys_addr = paddr;
711 			pgs[lcv].free_list = free_list;
712 			if (atop(paddr) >= avail_start &&
713 			    atop(paddr) <= avail_end)
714 				uvm_pagefree(&pgs[lcv]);
715 		}
716 		/* XXXCDC: incomplete: need to update uvmexp.free, what else? */
717 		/* XXXCDC: need hook to tell pmap to rebuild pv_list, etc... */
718 #endif
719 	} else {
720 		pgs = NULL;
721 		npages = 0;
722 	}
723 
724 	/*
725 	 * now insert us in the proper place in vm_physmem[]
726 	 */
727 
728 #if (VM_PHYSSEG_STRAT == VM_PSTRAT_RANDOM)
729 	/* random: put it at the end (easy!) */
730 	ps = &vm_physmem[vm_nphysseg];
731 #elif (VM_PHYSSEG_STRAT == VM_PSTRAT_BSEARCH)
732 	{
733 		int x;
734 		/* sort by address for binary search */
735 		for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
736 			if (start < vm_physmem[lcv].start)
737 				break;
738 		ps = &vm_physmem[lcv];
739 		/* move back other entries, if necessary ... */
740 		for (x = vm_nphysseg ; x > lcv ; x--)
741 			/* structure copy */
742 			vm_physmem[x] = vm_physmem[x - 1];
743 	}
744 #elif (VM_PHYSSEG_STRAT == VM_PSTRAT_BIGFIRST)
745 	{
746 		int x;
747 		/* sort by largest segment first */
748 		for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
749 			if ((end - start) >
750 			    (vm_physmem[lcv].end - vm_physmem[lcv].start))
751 				break;
752 		ps = &vm_physmem[lcv];
753 		/* move back other entries, if necessary ... */
754 		for (x = vm_nphysseg ; x > lcv ; x--)
755 			/* structure copy */
756 			vm_physmem[x] = vm_physmem[x - 1];
757 	}
758 #else
759 	panic("uvm_page_physload: unknown physseg strategy selected!");
760 #endif
761 
762 	ps->start = start;
763 	ps->end = end;
764 	ps->avail_start = avail_start;
765 	ps->avail_end = avail_end;
766 	if (preload) {
767 		ps->pgs = NULL;
768 	} else {
769 		ps->pgs = pgs;
770 		ps->lastpg = pgs + npages - 1;
771 	}
772 	ps->free_list = free_list;
773 	vm_nphysseg++;
774 
775 	if (!preload)
776 		uvm_page_rehash();
777 }
778 
779 /*
780  * uvm_page_rehash: reallocate hash table based on number of free pages.
781  */
782 
783 void
784 uvm_page_rehash()
785 {
786 	int freepages, lcv, bucketcount, oldcount;
787 	struct pglist *newbuckets, *oldbuckets;
788 	struct vm_page *pg;
789 	size_t newsize, oldsize;
790 
791 	/*
792 	 * compute number of pages that can go in the free pool
793 	 */
794 
795 	freepages = 0;
796 	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
797 		freepages +=
798 		    (vm_physmem[lcv].avail_end - vm_physmem[lcv].avail_start);
799 
800 	/*
801 	 * compute number of buckets needed for this number of pages
802 	 */
803 
804 	bucketcount = 1;
805 	while (bucketcount < freepages)
806 		bucketcount = bucketcount * 2;
807 
808 	/*
809 	 * compute the size of the current table and new table.
810 	 */
811 
812 	oldbuckets = uvm.page_hash;
813 	oldcount = uvm.page_nhash;
814 	oldsize = round_page(sizeof(struct pglist) * oldcount);
815 	newsize = round_page(sizeof(struct pglist) * bucketcount);
816 
817 	/*
818 	 * allocate the new buckets
819 	 */
820 
821 	newbuckets = (struct pglist *) uvm_km_alloc(kernel_map, newsize);
822 	if (newbuckets == NULL) {
823 		printf("uvm_page_physrehash: WARNING: could not grow page "
824 		    "hash table\n");
825 		return;
826 	}
827 	for (lcv = 0 ; lcv < bucketcount ; lcv++)
828 		TAILQ_INIT(&newbuckets[lcv]);
829 
830 	/*
831 	 * now replace the old buckets with the new ones and rehash everything
832 	 */
833 
834 	simple_lock(&uvm.hashlock);
835 	uvm.page_hash = newbuckets;
836 	uvm.page_nhash = bucketcount;
837 	uvm.page_hashmask = bucketcount - 1;  /* power of 2 */
838 
839 	/* ... and rehash */
840 	for (lcv = 0 ; lcv < oldcount ; lcv++) {
841 		while ((pg = oldbuckets[lcv].tqh_first) != NULL) {
842 			TAILQ_REMOVE(&oldbuckets[lcv], pg, hashq);
843 			TAILQ_INSERT_TAIL(
844 			  &uvm.page_hash[uvm_pagehash(pg->uobject, pg->offset)],
845 			  pg, hashq);
846 		}
847 	}
848 	simple_unlock(&uvm.hashlock);
849 
850 	/*
851 	 * free old bucket array if is not the boot-time table
852 	 */
853 
854 	if (oldbuckets != &uvm_bootbucket)
855 		uvm_km_free(kernel_map, (vaddr_t) oldbuckets, oldsize);
856 }
857 
858 /*
859  * uvm_page_recolor: Recolor the pages if the new bucket count is
860  * larger than the old one.
861  */
862 
863 void
864 uvm_page_recolor(int newncolors)
865 {
866 	struct pgflbucket *bucketarray, *oldbucketarray;
867 	struct pgfreelist pgfl;
868 	struct vm_page *pg;
869 	vsize_t bucketcount;
870 	int s, lcv, color, i, ocolors;
871 
872 	if (newncolors <= uvmexp.ncolors)
873 		return;
874 
875 	if (uvm.page_init_done == FALSE) {
876 		uvmexp.ncolors = newncolors;
877 		return;
878 	}
879 
880 	bucketcount = newncolors * VM_NFREELIST;
881 	bucketarray = malloc(bucketcount * sizeof(struct pgflbucket),
882 	    M_VMPAGE, M_NOWAIT);
883 	if (bucketarray == NULL) {
884 		printf("WARNING: unable to allocate %ld page color buckets\n",
885 		    (long) bucketcount);
886 		return;
887 	}
888 
889 	s = uvm_lock_fpageq();
890 
891 	/* Make sure we should still do this. */
892 	if (newncolors <= uvmexp.ncolors) {
893 		uvm_unlock_fpageq(s);
894 		free(bucketarray, M_VMPAGE);
895 		return;
896 	}
897 
898 	oldbucketarray = uvm.page_free[0].pgfl_buckets;
899 	ocolors = uvmexp.ncolors;
900 
901 	uvmexp.ncolors = newncolors;
902 	uvmexp.colormask = uvmexp.ncolors - 1;
903 
904 	for (lcv = 0; lcv < VM_NFREELIST; lcv++) {
905 		pgfl.pgfl_buckets = (bucketarray + (lcv * newncolors));
906 		uvm_page_init_buckets(&pgfl);
907 		for (color = 0; color < ocolors; color++) {
908 			for (i = 0; i < PGFL_NQUEUES; i++) {
909 				while ((pg = TAILQ_FIRST(&uvm.page_free[
910 				    lcv].pgfl_buckets[color].pgfl_queues[i]))
911 				    != NULL) {
912 					TAILQ_REMOVE(&uvm.page_free[
913 					    lcv].pgfl_buckets[
914 					    color].pgfl_queues[i], pg, pageq);
915 					TAILQ_INSERT_TAIL(&pgfl.pgfl_buckets[
916 					    VM_PGCOLOR_BUCKET(pg)].pgfl_queues[
917 					    i], pg, pageq);
918 				}
919 			}
920 		}
921 		uvm.page_free[lcv].pgfl_buckets = pgfl.pgfl_buckets;
922 	}
923 
924 	if (have_recolored_pages) {
925 		uvm_unlock_fpageq(s);
926 		free(oldbucketarray, M_VMPAGE);
927 		return;
928 	}
929 
930 	have_recolored_pages = TRUE;
931 	uvm_unlock_fpageq(s);
932 }
933 
934 /*
935  * uvm_pagealloc_pgfl: helper routine for uvm_pagealloc_strat
936  */
937 
938 static __inline struct vm_page *
939 uvm_pagealloc_pgfl(struct pgfreelist *pgfl, int try1, int try2,
940     int *trycolorp)
941 {
942 	struct pglist *freeq;
943 	struct vm_page *pg;
944 	int color, trycolor = *trycolorp;
945 
946 	color = trycolor;
947 	do {
948 		if ((pg = TAILQ_FIRST((freeq =
949 		    &pgfl->pgfl_buckets[color].pgfl_queues[try1]))) != NULL)
950 			goto gotit;
951 		if ((pg = TAILQ_FIRST((freeq =
952 		    &pgfl->pgfl_buckets[color].pgfl_queues[try2]))) != NULL)
953 			goto gotit;
954 		color = (color + 1) & uvmexp.colormask;
955 	} while (color != trycolor);
956 
957 	return (NULL);
958 
959  gotit:
960 	TAILQ_REMOVE(freeq, pg, pageq);
961 	uvmexp.free--;
962 
963 	/* update zero'd page count */
964 	if (pg->flags & PG_ZERO)
965 		uvmexp.zeropages--;
966 
967 	if (color == trycolor)
968 		uvmexp.colorhit++;
969 	else {
970 		uvmexp.colormiss++;
971 		*trycolorp = color;
972 	}
973 
974 	return (pg);
975 }
976 
977 /*
978  * uvm_pagealloc_strat: allocate vm_page from a particular free list.
979  *
980  * => return null if no pages free
981  * => wake up pagedaemon if number of free pages drops below low water mark
982  * => if obj != NULL, obj must be locked (to put in hash)
983  * => if anon != NULL, anon must be locked (to put in anon)
984  * => only one of obj or anon can be non-null
985  * => caller must activate/deactivate page if it is not wired.
986  * => free_list is ignored if strat == UVM_PGA_STRAT_NORMAL.
987  * => policy decision: it is more important to pull a page off of the
988  *	appropriate priority free list than it is to get a zero'd or
989  *	unknown contents page.  This is because we live with the
990  *	consequences of a bad free list decision for the entire
991  *	lifetime of the page, e.g. if the page comes from memory that
992  *	is slower to access.
993  */
994 
995 struct vm_page *
996 uvm_pagealloc_strat(obj, off, anon, flags, strat, free_list)
997 	struct uvm_object *obj;
998 	voff_t off;
999 	int flags;
1000 	struct vm_anon *anon;
1001 	int strat, free_list;
1002 {
1003 	int lcv, try1, try2, s, zeroit = 0, color;
1004 	struct vm_page *pg;
1005 	boolean_t use_reserve;
1006 
1007 	KASSERT(obj == NULL || anon == NULL);
1008 	KASSERT(off == trunc_page(off));
1009 	LOCK_ASSERT(obj == NULL || simple_lock_held(&obj->vmobjlock));
1010 	LOCK_ASSERT(anon == NULL || simple_lock_held(&anon->an_lock));
1011 
1012 	s = uvm_lock_fpageq();
1013 
1014 	/*
1015 	 * This implements a global round-robin page coloring
1016 	 * algorithm.
1017 	 *
1018 	 * XXXJRT: Should we make the `nextcolor' per-cpu?
1019 	 * XXXJRT: What about virtually-indexed caches?
1020 	 */
1021 
1022 	color = uvm.page_free_nextcolor;
1023 
1024 	/*
1025 	 * check to see if we need to generate some free pages waking
1026 	 * the pagedaemon.
1027 	 */
1028 
1029 	UVM_KICK_PDAEMON();
1030 
1031 	/*
1032 	 * fail if any of these conditions is true:
1033 	 * [1]  there really are no free pages, or
1034 	 * [2]  only kernel "reserved" pages remain and
1035 	 *        the page isn't being allocated to a kernel object.
1036 	 * [3]  only pagedaemon "reserved" pages remain and
1037 	 *        the requestor isn't the pagedaemon.
1038 	 */
1039 
1040 	use_reserve = (flags & UVM_PGA_USERESERVE) ||
1041 		(obj && UVM_OBJ_IS_KERN_OBJECT(obj));
1042 	if ((uvmexp.free <= uvmexp.reserve_kernel && !use_reserve) ||
1043 	    (uvmexp.free <= uvmexp.reserve_pagedaemon &&
1044 	     !(use_reserve && curproc == uvm.pagedaemon_proc)))
1045 		goto fail;
1046 
1047 #if PGFL_NQUEUES != 2
1048 #error uvm_pagealloc_strat needs to be updated
1049 #endif
1050 
1051 	/*
1052 	 * If we want a zero'd page, try the ZEROS queue first, otherwise
1053 	 * we try the UNKNOWN queue first.
1054 	 */
1055 	if (flags & UVM_PGA_ZERO) {
1056 		try1 = PGFL_ZEROS;
1057 		try2 = PGFL_UNKNOWN;
1058 	} else {
1059 		try1 = PGFL_UNKNOWN;
1060 		try2 = PGFL_ZEROS;
1061 	}
1062 
1063  again:
1064 	switch (strat) {
1065 	case UVM_PGA_STRAT_NORMAL:
1066 		/* Check all freelists in descending priority order. */
1067 		for (lcv = 0; lcv < VM_NFREELIST; lcv++) {
1068 			pg = uvm_pagealloc_pgfl(&uvm.page_free[lcv],
1069 			    try1, try2, &color);
1070 			if (pg != NULL)
1071 				goto gotit;
1072 		}
1073 
1074 		/* No pages free! */
1075 		goto fail;
1076 
1077 	case UVM_PGA_STRAT_ONLY:
1078 	case UVM_PGA_STRAT_FALLBACK:
1079 		/* Attempt to allocate from the specified free list. */
1080 		KASSERT(free_list >= 0 && free_list < VM_NFREELIST);
1081 		pg = uvm_pagealloc_pgfl(&uvm.page_free[free_list],
1082 		    try1, try2, &color);
1083 		if (pg != NULL)
1084 			goto gotit;
1085 
1086 		/* Fall back, if possible. */
1087 		if (strat == UVM_PGA_STRAT_FALLBACK) {
1088 			strat = UVM_PGA_STRAT_NORMAL;
1089 			goto again;
1090 		}
1091 
1092 		/* No pages free! */
1093 		goto fail;
1094 
1095 	default:
1096 		panic("uvm_pagealloc_strat: bad strat %d", strat);
1097 		/* NOTREACHED */
1098 	}
1099 
1100  gotit:
1101 	/*
1102 	 * We now know which color we actually allocated from; set
1103 	 * the next color accordingly.
1104 	 */
1105 
1106 	uvm.page_free_nextcolor = (color + 1) & uvmexp.colormask;
1107 
1108 	/*
1109 	 * update allocation statistics and remember if we have to
1110 	 * zero the page
1111 	 */
1112 
1113 	if (flags & UVM_PGA_ZERO) {
1114 		if (pg->flags & PG_ZERO) {
1115 			uvmexp.pga_zerohit++;
1116 			zeroit = 0;
1117 		} else {
1118 			uvmexp.pga_zeromiss++;
1119 			zeroit = 1;
1120 		}
1121 	}
1122 	uvm_unlock_fpageq(s);
1123 
1124 	pg->offset = off;
1125 	pg->uobject = obj;
1126 	pg->uanon = anon;
1127 	pg->flags = PG_BUSY|PG_CLEAN|PG_FAKE;
1128 	if (anon) {
1129 		anon->u.an_page = pg;
1130 		pg->pqflags = PQ_ANON;
1131 		uvmexp.anonpages++;
1132 	} else {
1133 		if (obj) {
1134 			uvm_pageinsert(pg);
1135 		}
1136 		pg->pqflags = 0;
1137 	}
1138 #if defined(UVM_PAGE_TRKOWN)
1139 	pg->owner_tag = NULL;
1140 #endif
1141 	UVM_PAGE_OWN(pg, "new alloc");
1142 
1143 	if (flags & UVM_PGA_ZERO) {
1144 		/*
1145 		 * A zero'd page is not clean.  If we got a page not already
1146 		 * zero'd, then we have to zero it ourselves.
1147 		 */
1148 		pg->flags &= ~PG_CLEAN;
1149 		if (zeroit)
1150 			pmap_zero_page(VM_PAGE_TO_PHYS(pg));
1151 	}
1152 
1153 	return(pg);
1154 
1155  fail:
1156 	uvm_unlock_fpageq(s);
1157 	return (NULL);
1158 }
1159 
1160 /*
1161  * uvm_pagerealloc: reallocate a page from one object to another
1162  *
1163  * => both objects must be locked
1164  */
1165 
1166 void
1167 uvm_pagerealloc(pg, newobj, newoff)
1168 	struct vm_page *pg;
1169 	struct uvm_object *newobj;
1170 	voff_t newoff;
1171 {
1172 	/*
1173 	 * remove it from the old object
1174 	 */
1175 
1176 	if (pg->uobject) {
1177 		uvm_pageremove(pg);
1178 	}
1179 
1180 	/*
1181 	 * put it in the new object
1182 	 */
1183 
1184 	if (newobj) {
1185 		pg->uobject = newobj;
1186 		pg->offset = newoff;
1187 		uvm_pageinsert(pg);
1188 	}
1189 }
1190 
1191 /*
1192  * uvm_pagefree: free page
1193  *
1194  * => erase page's identity (i.e. remove from hash/object)
1195  * => put page on free list
1196  * => caller must lock owning object (either anon or uvm_object)
1197  * => caller must lock page queues
1198  * => assumes all valid mappings of pg are gone
1199  */
1200 
1201 void
1202 uvm_pagefree(pg)
1203 	struct vm_page *pg;
1204 {
1205 	int s;
1206 
1207 	KASSERT((pg->flags & PG_PAGEOUT) == 0);
1208 	LOCK_ASSERT(simple_lock_held(&uvm.pageqlock) ||
1209 		    (pg->pqflags & (PQ_ACTIVE|PQ_INACTIVE)) == 0);
1210 	LOCK_ASSERT(pg->uobject == NULL ||
1211 		    simple_lock_held(&pg->uobject->vmobjlock));
1212 	LOCK_ASSERT(pg->uobject != NULL || pg->uanon == NULL ||
1213 		    simple_lock_held(&pg->uanon->an_lock));
1214 
1215 #ifdef DEBUG
1216 	if (pg->uobject == (void *)0xdeadbeef &&
1217 	    pg->uanon == (void *)0xdeadbeef) {
1218 		panic("uvm_pagefree: freeing free page %p", pg);
1219 	}
1220 #endif
1221 
1222 	/*
1223 	 * if the page is loaned, resolve the loan instead of freeing.
1224 	 */
1225 
1226 	if (pg->loan_count) {
1227 		KASSERT(pg->wire_count == 0);
1228 
1229 		/*
1230 		 * if the page is owned by an anon then we just want to
1231 		 * drop anon ownership.  the kernel will free the page when
1232 		 * it is done with it.  if the page is owned by an object,
1233 		 * remove it from the object and mark it dirty for the benefit
1234 		 * of possible anon owners.
1235 		 *
1236 		 * regardless of previous ownership, wakeup any waiters,
1237 		 * unbusy the page, and we're done.
1238 		 */
1239 
1240 		if (pg->uobject != NULL) {
1241 			uvm_pageremove(pg);
1242 			pg->flags &= ~PG_CLEAN;
1243 		} else if (pg->uanon != NULL) {
1244 			if ((pg->pqflags & PQ_ANON) == 0) {
1245 				pg->loan_count--;
1246 			} else {
1247 				pg->pqflags &= ~PQ_ANON;
1248 			}
1249 			pg->uanon = NULL;
1250 		}
1251 		if (pg->flags & PG_WANTED) {
1252 			wakeup(pg);
1253 		}
1254 		pg->flags &= ~(PG_WANTED|PG_BUSY|PG_RELEASED);
1255 #ifdef UVM_PAGE_TRKOWN
1256 		pg->owner_tag = NULL;
1257 #endif
1258 		if (pg->loan_count) {
1259 			uvm_pagedequeue(pg);
1260 			return;
1261 		}
1262 	}
1263 
1264 	/*
1265 	 * remove page from its object or anon.
1266 	 */
1267 
1268 	if (pg->uobject != NULL) {
1269 		uvm_pageremove(pg);
1270 	} else if (pg->uanon != NULL) {
1271 		pg->uanon->u.an_page = NULL;
1272 		uvmexp.anonpages--;
1273 	}
1274 
1275 	/*
1276 	 * now remove the page from the queues.
1277 	 */
1278 
1279 	uvm_pagedequeue(pg);
1280 
1281 	/*
1282 	 * if the page was wired, unwire it now.
1283 	 */
1284 
1285 	if (pg->wire_count) {
1286 		pg->wire_count = 0;
1287 		uvmexp.wired--;
1288 	}
1289 
1290 	/*
1291 	 * and put on free queue
1292 	 */
1293 
1294 	pg->flags &= ~PG_ZERO;
1295 
1296 	s = uvm_lock_fpageq();
1297 	TAILQ_INSERT_TAIL(&uvm.page_free[
1298 	    uvm_page_lookup_freelist(pg)].pgfl_buckets[
1299 	    VM_PGCOLOR_BUCKET(pg)].pgfl_queues[PGFL_UNKNOWN], pg, pageq);
1300 	pg->pqflags = PQ_FREE;
1301 #ifdef DEBUG
1302 	pg->uobject = (void *)0xdeadbeef;
1303 	pg->offset = 0xdeadbeef;
1304 	pg->uanon = (void *)0xdeadbeef;
1305 #endif
1306 	uvmexp.free++;
1307 
1308 	if (uvmexp.zeropages < UVM_PAGEZERO_TARGET)
1309 		uvm.page_idle_zero = vm_page_zero_enable;
1310 
1311 	uvm_unlock_fpageq(s);
1312 }
1313 
1314 /*
1315  * uvm_page_unbusy: unbusy an array of pages.
1316  *
1317  * => pages must either all belong to the same object, or all belong to anons.
1318  * => if pages are object-owned, object must be locked.
1319  * => if pages are anon-owned, anons must be locked.
1320  * => caller must lock page queues if pages may be released.
1321  */
1322 
1323 void
1324 uvm_page_unbusy(pgs, npgs)
1325 	struct vm_page **pgs;
1326 	int npgs;
1327 {
1328 	struct vm_page *pg;
1329 	int i;
1330 	UVMHIST_FUNC("uvm_page_unbusy"); UVMHIST_CALLED(ubchist);
1331 
1332 	for (i = 0; i < npgs; i++) {
1333 		pg = pgs[i];
1334 		if (pg == NULL) {
1335 			continue;
1336 		}
1337 		if (pg->flags & PG_WANTED) {
1338 			wakeup(pg);
1339 		}
1340 		if (pg->flags & PG_RELEASED) {
1341 			UVMHIST_LOG(ubchist, "releasing pg %p", pg,0,0,0);
1342 			pg->flags &= ~PG_RELEASED;
1343 			uvm_pagefree(pg);
1344 		} else {
1345 			UVMHIST_LOG(ubchist, "unbusying pg %p", pg,0,0,0);
1346 			pg->flags &= ~(PG_WANTED|PG_BUSY);
1347 			UVM_PAGE_OWN(pg, NULL);
1348 		}
1349 	}
1350 }
1351 
1352 #if defined(UVM_PAGE_TRKOWN)
1353 /*
1354  * uvm_page_own: set or release page ownership
1355  *
1356  * => this is a debugging function that keeps track of who sets PG_BUSY
1357  *	and where they do it.   it can be used to track down problems
1358  *	such a process setting "PG_BUSY" and never releasing it.
1359  * => page's object [if any] must be locked
1360  * => if "tag" is NULL then we are releasing page ownership
1361  */
1362 void
1363 uvm_page_own(pg, tag)
1364 	struct vm_page *pg;
1365 	char *tag;
1366 {
1367 	KASSERT((pg->flags & (PG_PAGEOUT|PG_RELEASED)) == 0);
1368 
1369 	/* gain ownership? */
1370 	if (tag) {
1371 		if (pg->owner_tag) {
1372 			printf("uvm_page_own: page %p already owned "
1373 			    "by proc %d [%s]\n", pg,
1374 			    pg->owner, pg->owner_tag);
1375 			panic("uvm_page_own");
1376 		}
1377 		pg->owner = (curproc) ? curproc->p_pid :  (pid_t) -1;
1378 		pg->owner_tag = tag;
1379 		return;
1380 	}
1381 
1382 	/* drop ownership */
1383 	if (pg->owner_tag == NULL) {
1384 		printf("uvm_page_own: dropping ownership of an non-owned "
1385 		    "page (%p)\n", pg);
1386 		panic("uvm_page_own");
1387 	}
1388 	KASSERT((pg->pqflags & (PQ_ACTIVE|PQ_INACTIVE)) ||
1389 	    (pg->uanon == NULL && pg->uobject == NULL) ||
1390 	    pg->uobject == uvm.kernel_object ||
1391 	    pg->wire_count > 0 ||
1392 	    (pg->loan_count == 1 && pg->uanon == NULL) ||
1393 	    pg->loan_count > 1);
1394 	pg->owner_tag = NULL;
1395 }
1396 #endif
1397 
1398 /*
1399  * uvm_pageidlezero: zero free pages while the system is idle.
1400  *
1401  * => try to complete one color bucket at a time, to reduce our impact
1402  *	on the CPU cache.
1403  * => we loop until we either reach the target or whichqs indicates that
1404  *	there is a process ready to run.
1405  */
1406 void
1407 uvm_pageidlezero()
1408 {
1409 	struct vm_page *pg;
1410 	struct pgfreelist *pgfl;
1411 	int free_list, s, firstbucket;
1412 	static int nextbucket;
1413 
1414 	s = uvm_lock_fpageq();
1415 	firstbucket = nextbucket;
1416 	do {
1417 		if (sched_whichqs != 0) {
1418 			uvm_unlock_fpageq(s);
1419 			return;
1420 		}
1421 		if (uvmexp.zeropages >= UVM_PAGEZERO_TARGET) {
1422 			uvm.page_idle_zero = FALSE;
1423 			uvm_unlock_fpageq(s);
1424 			return;
1425 		}
1426 		for (free_list = 0; free_list < VM_NFREELIST; free_list++) {
1427 			pgfl = &uvm.page_free[free_list];
1428 			while ((pg = TAILQ_FIRST(&pgfl->pgfl_buckets[
1429 			    nextbucket].pgfl_queues[PGFL_UNKNOWN])) != NULL) {
1430 				if (sched_whichqs != 0) {
1431 					uvm_unlock_fpageq(s);
1432 					return;
1433 				}
1434 
1435 				TAILQ_REMOVE(&pgfl->pgfl_buckets[
1436 				    nextbucket].pgfl_queues[PGFL_UNKNOWN],
1437 				    pg, pageq);
1438 				uvmexp.free--;
1439 				uvm_unlock_fpageq(s);
1440 #ifdef PMAP_PAGEIDLEZERO
1441 				if (!PMAP_PAGEIDLEZERO(VM_PAGE_TO_PHYS(pg))) {
1442 
1443 					/*
1444 					 * The machine-dependent code detected
1445 					 * some reason for us to abort zeroing
1446 					 * pages, probably because there is a
1447 					 * process now ready to run.
1448 					 */
1449 
1450 					s = uvm_lock_fpageq();
1451 					TAILQ_INSERT_HEAD(&pgfl->pgfl_buckets[
1452 					    nextbucket].pgfl_queues[
1453 					    PGFL_UNKNOWN], pg, pageq);
1454 					uvmexp.free++;
1455 					uvmexp.zeroaborts++;
1456 					uvm_unlock_fpageq(s);
1457 					return;
1458 				}
1459 #else
1460 				pmap_zero_page(VM_PAGE_TO_PHYS(pg));
1461 #endif /* PMAP_PAGEIDLEZERO */
1462 				pg->flags |= PG_ZERO;
1463 
1464 				s = uvm_lock_fpageq();
1465 				TAILQ_INSERT_HEAD(&pgfl->pgfl_buckets[
1466 				    nextbucket].pgfl_queues[PGFL_ZEROS],
1467 				    pg, pageq);
1468 				uvmexp.free++;
1469 				uvmexp.zeropages++;
1470 			}
1471 		}
1472 		nextbucket = (nextbucket + 1) & uvmexp.colormask;
1473 	} while (nextbucket != firstbucket);
1474 	uvm_unlock_fpageq(s);
1475 }
1476